Mention branches and keyring.
[releases.git] / netfilter / nf_tables_api.c
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 #define NFT_SET_MAX_ANONLEN 16
28
29 unsigned int nf_tables_net_id __read_mostly;
30
31 static LIST_HEAD(nf_tables_expressions);
32 static LIST_HEAD(nf_tables_objects);
33 static LIST_HEAD(nf_tables_flowtables);
34 static LIST_HEAD(nf_tables_destroy_list);
35 static LIST_HEAD(nf_tables_gc_list);
36 static DEFINE_SPINLOCK(nf_tables_destroy_list_lock);
37 static DEFINE_SPINLOCK(nf_tables_gc_list_lock);
38
39 enum {
40         NFT_VALIDATE_SKIP       = 0,
41         NFT_VALIDATE_NEED,
42         NFT_VALIDATE_DO,
43 };
44
45 static struct rhltable nft_objname_ht;
46
47 static u32 nft_chain_hash(const void *data, u32 len, u32 seed);
48 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed);
49 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *, const void *);
50
51 static u32 nft_objname_hash(const void *data, u32 len, u32 seed);
52 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed);
53 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *, const void *);
54
55 static const struct rhashtable_params nft_chain_ht_params = {
56         .head_offset            = offsetof(struct nft_chain, rhlhead),
57         .key_offset             = offsetof(struct nft_chain, name),
58         .hashfn                 = nft_chain_hash,
59         .obj_hashfn             = nft_chain_hash_obj,
60         .obj_cmpfn              = nft_chain_hash_cmp,
61         .automatic_shrinking    = true,
62 };
63
64 static const struct rhashtable_params nft_objname_ht_params = {
65         .head_offset            = offsetof(struct nft_object, rhlhead),
66         .key_offset             = offsetof(struct nft_object, key),
67         .hashfn                 = nft_objname_hash,
68         .obj_hashfn             = nft_objname_hash_obj,
69         .obj_cmpfn              = nft_objname_hash_cmp,
70         .automatic_shrinking    = true,
71 };
72
73 struct nft_audit_data {
74         struct nft_table *table;
75         int entries;
76         int op;
77         struct list_head list;
78 };
79
80 static const u8 nft2audit_op[NFT_MSG_MAX] = { // enum nf_tables_msg_types
81         [NFT_MSG_NEWTABLE]      = AUDIT_NFT_OP_TABLE_REGISTER,
82         [NFT_MSG_GETTABLE]      = AUDIT_NFT_OP_INVALID,
83         [NFT_MSG_DELTABLE]      = AUDIT_NFT_OP_TABLE_UNREGISTER,
84         [NFT_MSG_NEWCHAIN]      = AUDIT_NFT_OP_CHAIN_REGISTER,
85         [NFT_MSG_GETCHAIN]      = AUDIT_NFT_OP_INVALID,
86         [NFT_MSG_DELCHAIN]      = AUDIT_NFT_OP_CHAIN_UNREGISTER,
87         [NFT_MSG_NEWRULE]       = AUDIT_NFT_OP_RULE_REGISTER,
88         [NFT_MSG_GETRULE]       = AUDIT_NFT_OP_INVALID,
89         [NFT_MSG_DELRULE]       = AUDIT_NFT_OP_RULE_UNREGISTER,
90         [NFT_MSG_NEWSET]        = AUDIT_NFT_OP_SET_REGISTER,
91         [NFT_MSG_GETSET]        = AUDIT_NFT_OP_INVALID,
92         [NFT_MSG_DELSET]        = AUDIT_NFT_OP_SET_UNREGISTER,
93         [NFT_MSG_NEWSETELEM]    = AUDIT_NFT_OP_SETELEM_REGISTER,
94         [NFT_MSG_GETSETELEM]    = AUDIT_NFT_OP_INVALID,
95         [NFT_MSG_DELSETELEM]    = AUDIT_NFT_OP_SETELEM_UNREGISTER,
96         [NFT_MSG_NEWGEN]        = AUDIT_NFT_OP_GEN_REGISTER,
97         [NFT_MSG_GETGEN]        = AUDIT_NFT_OP_INVALID,
98         [NFT_MSG_TRACE]         = AUDIT_NFT_OP_INVALID,
99         [NFT_MSG_NEWOBJ]        = AUDIT_NFT_OP_OBJ_REGISTER,
100         [NFT_MSG_GETOBJ]        = AUDIT_NFT_OP_INVALID,
101         [NFT_MSG_DELOBJ]        = AUDIT_NFT_OP_OBJ_UNREGISTER,
102         [NFT_MSG_GETOBJ_RESET]  = AUDIT_NFT_OP_OBJ_RESET,
103         [NFT_MSG_NEWFLOWTABLE]  = AUDIT_NFT_OP_FLOWTABLE_REGISTER,
104         [NFT_MSG_GETFLOWTABLE]  = AUDIT_NFT_OP_INVALID,
105         [NFT_MSG_DELFLOWTABLE]  = AUDIT_NFT_OP_FLOWTABLE_UNREGISTER,
106 };
107
108 static void nft_validate_state_update(struct net *net, u8 new_validate_state)
109 {
110         struct nftables_pernet *nft_net = nft_pernet(net);
111
112         switch (nft_net->validate_state) {
113         case NFT_VALIDATE_SKIP:
114                 WARN_ON_ONCE(new_validate_state == NFT_VALIDATE_DO);
115                 break;
116         case NFT_VALIDATE_NEED:
117                 break;
118         case NFT_VALIDATE_DO:
119                 if (new_validate_state == NFT_VALIDATE_NEED)
120                         return;
121         }
122
123         nft_net->validate_state = new_validate_state;
124 }
125 static void nf_tables_trans_destroy_work(struct work_struct *w);
126 static DECLARE_WORK(trans_destroy_work, nf_tables_trans_destroy_work);
127
128 static void nft_trans_gc_work(struct work_struct *work);
129 static DECLARE_WORK(trans_gc_work, nft_trans_gc_work);
130
131 static void nft_ctx_init(struct nft_ctx *ctx,
132                          struct net *net,
133                          const struct sk_buff *skb,
134                          const struct nlmsghdr *nlh,
135                          u8 family,
136                          struct nft_table *table,
137                          struct nft_chain *chain,
138                          const struct nlattr * const *nla)
139 {
140         ctx->net        = net;
141         ctx->family     = family;
142         ctx->level      = 0;
143         ctx->table      = table;
144         ctx->chain      = chain;
145         ctx->nla        = nla;
146         ctx->portid     = NETLINK_CB(skb).portid;
147         ctx->report     = nlmsg_report(nlh);
148         ctx->flags      = nlh->nlmsg_flags;
149         ctx->seq        = nlh->nlmsg_seq;
150 }
151
152 static struct nft_trans *nft_trans_alloc_gfp(const struct nft_ctx *ctx,
153                                              int msg_type, u32 size, gfp_t gfp)
154 {
155         struct nft_trans *trans;
156
157         trans = kzalloc(sizeof(struct nft_trans) + size, gfp);
158         if (trans == NULL)
159                 return NULL;
160
161         INIT_LIST_HEAD(&trans->list);
162         INIT_LIST_HEAD(&trans->binding_list);
163         trans->msg_type = msg_type;
164         trans->ctx      = *ctx;
165
166         return trans;
167 }
168
169 static struct nft_trans *nft_trans_alloc(const struct nft_ctx *ctx,
170                                          int msg_type, u32 size)
171 {
172         return nft_trans_alloc_gfp(ctx, msg_type, size, GFP_KERNEL);
173 }
174
175 static void nft_trans_list_del(struct nft_trans *trans)
176 {
177         list_del(&trans->list);
178         list_del(&trans->binding_list);
179 }
180
181 static void nft_trans_destroy(struct nft_trans *trans)
182 {
183         nft_trans_list_del(trans);
184         kfree(trans);
185 }
186
187 static void __nft_set_trans_bind(const struct nft_ctx *ctx, struct nft_set *set,
188                                  bool bind)
189 {
190         struct nftables_pernet *nft_net;
191         struct net *net = ctx->net;
192         struct nft_trans *trans;
193
194         if (!nft_set_is_anonymous(set))
195                 return;
196
197         nft_net = nft_pernet(net);
198         list_for_each_entry_reverse(trans, &nft_net->commit_list, list) {
199                 switch (trans->msg_type) {
200                 case NFT_MSG_NEWSET:
201                         if (nft_trans_set(trans) == set)
202                                 nft_trans_set_bound(trans) = bind;
203                         break;
204                 case NFT_MSG_NEWSETELEM:
205                         if (nft_trans_elem_set(trans) == set)
206                                 nft_trans_elem_set_bound(trans) = bind;
207                         break;
208                 }
209         }
210 }
211
212 static void nft_set_trans_bind(const struct nft_ctx *ctx, struct nft_set *set)
213 {
214         return __nft_set_trans_bind(ctx, set, true);
215 }
216
217 static void nft_set_trans_unbind(const struct nft_ctx *ctx, struct nft_set *set)
218 {
219         return __nft_set_trans_bind(ctx, set, false);
220 }
221
222 static void __nft_chain_trans_bind(const struct nft_ctx *ctx,
223                                    struct nft_chain *chain, bool bind)
224 {
225         struct nftables_pernet *nft_net;
226         struct net *net = ctx->net;
227         struct nft_trans *trans;
228
229         if (!nft_chain_binding(chain))
230                 return;
231
232         nft_net = nft_pernet(net);
233         list_for_each_entry_reverse(trans, &nft_net->commit_list, list) {
234                 switch (trans->msg_type) {
235                 case NFT_MSG_NEWCHAIN:
236                         if (nft_trans_chain(trans) == chain)
237                                 nft_trans_chain_bound(trans) = bind;
238                         break;
239                 case NFT_MSG_NEWRULE:
240                         if (trans->ctx.chain == chain)
241                                 nft_trans_rule_bound(trans) = bind;
242                         break;
243                 }
244         }
245 }
246
247 static void nft_chain_trans_bind(const struct nft_ctx *ctx,
248                                  struct nft_chain *chain)
249 {
250         __nft_chain_trans_bind(ctx, chain, true);
251 }
252
253 int nf_tables_bind_chain(const struct nft_ctx *ctx, struct nft_chain *chain)
254 {
255         if (!nft_chain_binding(chain))
256                 return 0;
257
258         if (nft_chain_binding(ctx->chain))
259                 return -EOPNOTSUPP;
260
261         if (chain->bound)
262                 return -EBUSY;
263
264         if (!nft_use_inc(&chain->use))
265                 return -EMFILE;
266
267         chain->bound = true;
268         nft_chain_trans_bind(ctx, chain);
269
270         return 0;
271 }
272
273 void nf_tables_unbind_chain(const struct nft_ctx *ctx, struct nft_chain *chain)
274 {
275         __nft_chain_trans_bind(ctx, chain, false);
276 }
277
278 static int nft_netdev_register_hooks(struct net *net,
279                                      struct list_head *hook_list)
280 {
281         struct nft_hook *hook;
282         int err, j;
283
284         j = 0;
285         list_for_each_entry(hook, hook_list, list) {
286                 err = nf_register_net_hook(net, &hook->ops);
287                 if (err < 0)
288                         goto err_register;
289
290                 j++;
291         }
292         return 0;
293
294 err_register:
295         list_for_each_entry(hook, hook_list, list) {
296                 if (j-- <= 0)
297                         break;
298
299                 nf_unregister_net_hook(net, &hook->ops);
300         }
301         return err;
302 }
303
304 static void nft_netdev_unregister_hooks(struct net *net,
305                                         struct list_head *hook_list,
306                                         bool release_netdev)
307 {
308         struct nft_hook *hook, *next;
309
310         list_for_each_entry_safe(hook, next, hook_list, list) {
311                 nf_unregister_net_hook(net, &hook->ops);
312                 if (release_netdev) {
313                         list_del(&hook->list);
314                         kfree_rcu(hook, rcu);
315                 }
316         }
317 }
318
319 static int nf_tables_register_hook(struct net *net,
320                                    const struct nft_table *table,
321                                    struct nft_chain *chain)
322 {
323         struct nft_base_chain *basechain;
324         const struct nf_hook_ops *ops;
325
326         if (table->flags & NFT_TABLE_F_DORMANT ||
327             !nft_is_base_chain(chain))
328                 return 0;
329
330         basechain = nft_base_chain(chain);
331         ops = &basechain->ops;
332
333         if (basechain->type->ops_register)
334                 return basechain->type->ops_register(net, ops);
335
336         if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
337                 return nft_netdev_register_hooks(net, &basechain->hook_list);
338
339         return nf_register_net_hook(net, &basechain->ops);
340 }
341
342 static void __nf_tables_unregister_hook(struct net *net,
343                                         const struct nft_table *table,
344                                         struct nft_chain *chain,
345                                         bool release_netdev)
346 {
347         struct nft_base_chain *basechain;
348         const struct nf_hook_ops *ops;
349
350         if (table->flags & NFT_TABLE_F_DORMANT ||
351             !nft_is_base_chain(chain))
352                 return;
353         basechain = nft_base_chain(chain);
354         ops = &basechain->ops;
355
356         if (basechain->type->ops_unregister)
357                 return basechain->type->ops_unregister(net, ops);
358
359         if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
360                 nft_netdev_unregister_hooks(net, &basechain->hook_list,
361                                             release_netdev);
362         else
363                 nf_unregister_net_hook(net, &basechain->ops);
364 }
365
366 static void nf_tables_unregister_hook(struct net *net,
367                                       const struct nft_table *table,
368                                       struct nft_chain *chain)
369 {
370         return __nf_tables_unregister_hook(net, table, chain, false);
371 }
372
373 static void nft_trans_commit_list_add_tail(struct net *net, struct nft_trans *trans)
374 {
375         struct nftables_pernet *nft_net = nft_pernet(net);
376
377         switch (trans->msg_type) {
378         case NFT_MSG_NEWSET:
379                 if (!nft_trans_set_update(trans) &&
380                     nft_set_is_anonymous(nft_trans_set(trans)))
381                         list_add_tail(&trans->binding_list, &nft_net->binding_list);
382                 break;
383         case NFT_MSG_NEWCHAIN:
384                 if (!nft_trans_chain_update(trans) &&
385                     nft_chain_binding(nft_trans_chain(trans)))
386                         list_add_tail(&trans->binding_list, &nft_net->binding_list);
387                 break;
388         }
389
390         list_add_tail(&trans->list, &nft_net->commit_list);
391 }
392
393 static int nft_trans_table_add(struct nft_ctx *ctx, int msg_type)
394 {
395         struct nft_trans *trans;
396
397         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_table));
398         if (trans == NULL)
399                 return -ENOMEM;
400
401         if (msg_type == NFT_MSG_NEWTABLE)
402                 nft_activate_next(ctx->net, ctx->table);
403
404         nft_trans_commit_list_add_tail(ctx->net, trans);
405         return 0;
406 }
407
408 static int nft_deltable(struct nft_ctx *ctx)
409 {
410         int err;
411
412         err = nft_trans_table_add(ctx, NFT_MSG_DELTABLE);
413         if (err < 0)
414                 return err;
415
416         nft_deactivate_next(ctx->net, ctx->table);
417         return err;
418 }
419
420 static struct nft_trans *nft_trans_chain_add(struct nft_ctx *ctx, int msg_type)
421 {
422         struct nft_trans *trans;
423
424         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_chain));
425         if (trans == NULL)
426                 return ERR_PTR(-ENOMEM);
427
428         if (msg_type == NFT_MSG_NEWCHAIN) {
429                 nft_activate_next(ctx->net, ctx->chain);
430
431                 if (ctx->nla[NFTA_CHAIN_ID]) {
432                         nft_trans_chain_id(trans) =
433                                 ntohl(nla_get_be32(ctx->nla[NFTA_CHAIN_ID]));
434                 }
435         }
436         nft_trans_chain(trans) = ctx->chain;
437         nft_trans_commit_list_add_tail(ctx->net, trans);
438
439         return trans;
440 }
441
442 static int nft_delchain(struct nft_ctx *ctx)
443 {
444         struct nft_trans *trans;
445
446         trans = nft_trans_chain_add(ctx, NFT_MSG_DELCHAIN);
447         if (IS_ERR(trans))
448                 return PTR_ERR(trans);
449
450         nft_use_dec(&ctx->table->use);
451         nft_deactivate_next(ctx->net, ctx->chain);
452
453         return 0;
454 }
455
456 void nft_rule_expr_activate(const struct nft_ctx *ctx, struct nft_rule *rule)
457 {
458         struct nft_expr *expr;
459
460         expr = nft_expr_first(rule);
461         while (nft_expr_more(rule, expr)) {
462                 if (expr->ops->activate)
463                         expr->ops->activate(ctx, expr);
464
465                 expr = nft_expr_next(expr);
466         }
467 }
468
469 void nft_rule_expr_deactivate(const struct nft_ctx *ctx, struct nft_rule *rule,
470                               enum nft_trans_phase phase)
471 {
472         struct nft_expr *expr;
473
474         expr = nft_expr_first(rule);
475         while (nft_expr_more(rule, expr)) {
476                 if (expr->ops->deactivate)
477                         expr->ops->deactivate(ctx, expr, phase);
478
479                 expr = nft_expr_next(expr);
480         }
481 }
482
483 static int
484 nf_tables_delrule_deactivate(struct nft_ctx *ctx, struct nft_rule *rule)
485 {
486         /* You cannot delete the same rule twice */
487         if (nft_is_active_next(ctx->net, rule)) {
488                 nft_deactivate_next(ctx->net, rule);
489                 nft_use_dec(&ctx->chain->use);
490                 return 0;
491         }
492         return -ENOENT;
493 }
494
495 static struct nft_trans *nft_trans_rule_add(struct nft_ctx *ctx, int msg_type,
496                                             struct nft_rule *rule)
497 {
498         struct nft_trans *trans;
499
500         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_rule));
501         if (trans == NULL)
502                 return NULL;
503
504         if (msg_type == NFT_MSG_NEWRULE && ctx->nla[NFTA_RULE_ID] != NULL) {
505                 nft_trans_rule_id(trans) =
506                         ntohl(nla_get_be32(ctx->nla[NFTA_RULE_ID]));
507         }
508         nft_trans_rule(trans) = rule;
509         nft_trans_commit_list_add_tail(ctx->net, trans);
510
511         return trans;
512 }
513
514 static int nft_delrule(struct nft_ctx *ctx, struct nft_rule *rule)
515 {
516         struct nft_flow_rule *flow;
517         struct nft_trans *trans;
518         int err;
519
520         trans = nft_trans_rule_add(ctx, NFT_MSG_DELRULE, rule);
521         if (trans == NULL)
522                 return -ENOMEM;
523
524         if (ctx->chain->flags & NFT_CHAIN_HW_OFFLOAD) {
525                 flow = nft_flow_rule_create(ctx->net, rule);
526                 if (IS_ERR(flow)) {
527                         nft_trans_destroy(trans);
528                         return PTR_ERR(flow);
529                 }
530
531                 nft_trans_flow_rule(trans) = flow;
532         }
533
534         err = nf_tables_delrule_deactivate(ctx, rule);
535         if (err < 0) {
536                 nft_trans_destroy(trans);
537                 return err;
538         }
539         nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_PREPARE);
540
541         return 0;
542 }
543
544 static int nft_delrule_by_chain(struct nft_ctx *ctx)
545 {
546         struct nft_rule *rule;
547         int err;
548
549         list_for_each_entry(rule, &ctx->chain->rules, list) {
550                 if (!nft_is_active_next(ctx->net, rule))
551                         continue;
552
553                 err = nft_delrule(ctx, rule);
554                 if (err < 0)
555                         return err;
556         }
557         return 0;
558 }
559
560 static int __nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
561                                struct nft_set *set,
562                                const struct nft_set_desc *desc)
563 {
564         struct nft_trans *trans;
565
566         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_set));
567         if (trans == NULL)
568                 return -ENOMEM;
569
570         if (msg_type == NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID] && !desc) {
571                 nft_trans_set_id(trans) =
572                         ntohl(nla_get_be32(ctx->nla[NFTA_SET_ID]));
573                 nft_activate_next(ctx->net, set);
574         }
575         nft_trans_set(trans) = set;
576         if (desc) {
577                 nft_trans_set_update(trans) = true;
578                 nft_trans_set_gc_int(trans) = desc->gc_int;
579                 nft_trans_set_timeout(trans) = desc->timeout;
580         }
581         nft_trans_commit_list_add_tail(ctx->net, trans);
582
583         return 0;
584 }
585
586 static int nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
587                              struct nft_set *set)
588 {
589         return __nft_trans_set_add(ctx, msg_type, set, NULL);
590 }
591
592 static int nft_mapelem_deactivate(const struct nft_ctx *ctx,
593                                   struct nft_set *set,
594                                   const struct nft_set_iter *iter,
595                                   struct nft_set_elem *elem)
596 {
597         nft_setelem_data_deactivate(ctx->net, set, elem);
598
599         return 0;
600 }
601
602 struct nft_set_elem_catchall {
603         struct list_head        list;
604         struct rcu_head         rcu;
605         void                    *elem;
606 };
607
608 static void nft_map_catchall_deactivate(const struct nft_ctx *ctx,
609                                         struct nft_set *set)
610 {
611         u8 genmask = nft_genmask_next(ctx->net);
612         struct nft_set_elem_catchall *catchall;
613         struct nft_set_elem elem;
614         struct nft_set_ext *ext;
615
616         list_for_each_entry(catchall, &set->catchall_list, list) {
617                 ext = nft_set_elem_ext(set, catchall->elem);
618                 if (!nft_set_elem_active(ext, genmask))
619                         continue;
620
621                 elem.priv = catchall->elem;
622                 nft_setelem_data_deactivate(ctx->net, set, &elem);
623                 break;
624         }
625 }
626
627 static void nft_map_deactivate(const struct nft_ctx *ctx, struct nft_set *set)
628 {
629         struct nft_set_iter iter = {
630                 .genmask        = nft_genmask_next(ctx->net),
631                 .fn             = nft_mapelem_deactivate,
632         };
633
634         set->ops->walk(ctx, set, &iter);
635         WARN_ON_ONCE(iter.err);
636
637         nft_map_catchall_deactivate(ctx, set);
638 }
639
640 static int nft_delset(const struct nft_ctx *ctx, struct nft_set *set)
641 {
642         int err;
643
644         err = nft_trans_set_add(ctx, NFT_MSG_DELSET, set);
645         if (err < 0)
646                 return err;
647
648         if (set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
649                 nft_map_deactivate(ctx, set);
650
651         nft_deactivate_next(ctx->net, set);
652         nft_use_dec(&ctx->table->use);
653
654         return err;
655 }
656
657 static int nft_trans_obj_add(struct nft_ctx *ctx, int msg_type,
658                              struct nft_object *obj)
659 {
660         struct nft_trans *trans;
661
662         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_obj));
663         if (trans == NULL)
664                 return -ENOMEM;
665
666         if (msg_type == NFT_MSG_NEWOBJ)
667                 nft_activate_next(ctx->net, obj);
668
669         nft_trans_obj(trans) = obj;
670         nft_trans_commit_list_add_tail(ctx->net, trans);
671
672         return 0;
673 }
674
675 static int nft_delobj(struct nft_ctx *ctx, struct nft_object *obj)
676 {
677         int err;
678
679         err = nft_trans_obj_add(ctx, NFT_MSG_DELOBJ, obj);
680         if (err < 0)
681                 return err;
682
683         nft_deactivate_next(ctx->net, obj);
684         nft_use_dec(&ctx->table->use);
685
686         return err;
687 }
688
689 static struct nft_trans *
690 nft_trans_flowtable_add(struct nft_ctx *ctx, int msg_type,
691                         struct nft_flowtable *flowtable)
692 {
693         struct nft_trans *trans;
694
695         trans = nft_trans_alloc(ctx, msg_type,
696                                 sizeof(struct nft_trans_flowtable));
697         if (trans == NULL)
698                 return ERR_PTR(-ENOMEM);
699
700         if (msg_type == NFT_MSG_NEWFLOWTABLE)
701                 nft_activate_next(ctx->net, flowtable);
702
703         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
704         nft_trans_flowtable(trans) = flowtable;
705         nft_trans_commit_list_add_tail(ctx->net, trans);
706
707         return trans;
708 }
709
710 static int nft_delflowtable(struct nft_ctx *ctx,
711                             struct nft_flowtable *flowtable)
712 {
713         struct nft_trans *trans;
714
715         trans = nft_trans_flowtable_add(ctx, NFT_MSG_DELFLOWTABLE, flowtable);
716         if (IS_ERR(trans))
717                 return PTR_ERR(trans);
718
719         nft_deactivate_next(ctx->net, flowtable);
720         nft_use_dec(&ctx->table->use);
721
722         return 0;
723 }
724
725 static void __nft_reg_track_clobber(struct nft_regs_track *track, u8 dreg)
726 {
727         int i;
728
729         for (i = track->regs[dreg].num_reg; i > 0; i--)
730                 __nft_reg_track_cancel(track, dreg - i);
731 }
732
733 static void __nft_reg_track_update(struct nft_regs_track *track,
734                                    const struct nft_expr *expr,
735                                    u8 dreg, u8 num_reg)
736 {
737         track->regs[dreg].selector = expr;
738         track->regs[dreg].bitwise = NULL;
739         track->regs[dreg].num_reg = num_reg;
740 }
741
742 void nft_reg_track_update(struct nft_regs_track *track,
743                           const struct nft_expr *expr, u8 dreg, u8 len)
744 {
745         unsigned int regcount;
746         int i;
747
748         __nft_reg_track_clobber(track, dreg);
749
750         regcount = DIV_ROUND_UP(len, NFT_REG32_SIZE);
751         for (i = 0; i < regcount; i++, dreg++)
752                 __nft_reg_track_update(track, expr, dreg, i);
753 }
754 EXPORT_SYMBOL_GPL(nft_reg_track_update);
755
756 void nft_reg_track_cancel(struct nft_regs_track *track, u8 dreg, u8 len)
757 {
758         unsigned int regcount;
759         int i;
760
761         __nft_reg_track_clobber(track, dreg);
762
763         regcount = DIV_ROUND_UP(len, NFT_REG32_SIZE);
764         for (i = 0; i < regcount; i++, dreg++)
765                 __nft_reg_track_cancel(track, dreg);
766 }
767 EXPORT_SYMBOL_GPL(nft_reg_track_cancel);
768
769 void __nft_reg_track_cancel(struct nft_regs_track *track, u8 dreg)
770 {
771         track->regs[dreg].selector = NULL;
772         track->regs[dreg].bitwise = NULL;
773         track->regs[dreg].num_reg = 0;
774 }
775 EXPORT_SYMBOL_GPL(__nft_reg_track_cancel);
776
777 /*
778  * Tables
779  */
780
781 static struct nft_table *nft_table_lookup(const struct net *net,
782                                           const struct nlattr *nla,
783                                           u8 family, u8 genmask, u32 nlpid)
784 {
785         struct nftables_pernet *nft_net;
786         struct nft_table *table;
787
788         if (nla == NULL)
789                 return ERR_PTR(-EINVAL);
790
791         nft_net = nft_pernet(net);
792         list_for_each_entry_rcu(table, &nft_net->tables, list,
793                                 lockdep_is_held(&nft_net->commit_mutex)) {
794                 if (!nla_strcmp(nla, table->name) &&
795                     table->family == family &&
796                     nft_active_genmask(table, genmask)) {
797                         if (nft_table_has_owner(table) &&
798                             nlpid && table->nlpid != nlpid)
799                                 return ERR_PTR(-EPERM);
800
801                         return table;
802                 }
803         }
804
805         return ERR_PTR(-ENOENT);
806 }
807
808 static struct nft_table *nft_table_lookup_byhandle(const struct net *net,
809                                                    const struct nlattr *nla,
810                                                    int family, u8 genmask, u32 nlpid)
811 {
812         struct nftables_pernet *nft_net;
813         struct nft_table *table;
814
815         nft_net = nft_pernet(net);
816         list_for_each_entry(table, &nft_net->tables, list) {
817                 if (be64_to_cpu(nla_get_be64(nla)) == table->handle &&
818                     table->family == family &&
819                     nft_active_genmask(table, genmask)) {
820                         if (nft_table_has_owner(table) &&
821                             nlpid && table->nlpid != nlpid)
822                                 return ERR_PTR(-EPERM);
823
824                         return table;
825                 }
826         }
827
828         return ERR_PTR(-ENOENT);
829 }
830
831 static inline u64 nf_tables_alloc_handle(struct nft_table *table)
832 {
833         return ++table->hgenerator;
834 }
835
836 static const struct nft_chain_type *chain_type[NFPROTO_NUMPROTO][NFT_CHAIN_T_MAX];
837
838 static const struct nft_chain_type *
839 __nft_chain_type_get(u8 family, enum nft_chain_types type)
840 {
841         if (family >= NFPROTO_NUMPROTO ||
842             type >= NFT_CHAIN_T_MAX)
843                 return NULL;
844
845         return chain_type[family][type];
846 }
847
848 static const struct nft_chain_type *
849 __nf_tables_chain_type_lookup(const struct nlattr *nla, u8 family)
850 {
851         const struct nft_chain_type *type;
852         int i;
853
854         for (i = 0; i < NFT_CHAIN_T_MAX; i++) {
855                 type = __nft_chain_type_get(family, i);
856                 if (!type)
857                         continue;
858                 if (!nla_strcmp(nla, type->name))
859                         return type;
860         }
861         return NULL;
862 }
863
864 struct nft_module_request {
865         struct list_head        list;
866         char                    module[MODULE_NAME_LEN];
867         bool                    done;
868 };
869
870 #ifdef CONFIG_MODULES
871 __printf(2, 3) int nft_request_module(struct net *net, const char *fmt,
872                                       ...)
873 {
874         char module_name[MODULE_NAME_LEN];
875         struct nftables_pernet *nft_net;
876         struct nft_module_request *req;
877         va_list args;
878         int ret;
879
880         va_start(args, fmt);
881         ret = vsnprintf(module_name, MODULE_NAME_LEN, fmt, args);
882         va_end(args);
883         if (ret >= MODULE_NAME_LEN)
884                 return 0;
885
886         nft_net = nft_pernet(net);
887         list_for_each_entry(req, &nft_net->module_list, list) {
888                 if (!strcmp(req->module, module_name)) {
889                         if (req->done)
890                                 return 0;
891
892                         /* A request to load this module already exists. */
893                         return -EAGAIN;
894                 }
895         }
896
897         req = kmalloc(sizeof(*req), GFP_KERNEL);
898         if (!req)
899                 return -ENOMEM;
900
901         req->done = false;
902         strscpy(req->module, module_name, MODULE_NAME_LEN);
903         list_add_tail(&req->list, &nft_net->module_list);
904
905         return -EAGAIN;
906 }
907 EXPORT_SYMBOL_GPL(nft_request_module);
908 #endif
909
910 static void lockdep_nfnl_nft_mutex_not_held(void)
911 {
912 #ifdef CONFIG_PROVE_LOCKING
913         if (debug_locks)
914                 WARN_ON_ONCE(lockdep_nfnl_is_held(NFNL_SUBSYS_NFTABLES));
915 #endif
916 }
917
918 static const struct nft_chain_type *
919 nf_tables_chain_type_lookup(struct net *net, const struct nlattr *nla,
920                             u8 family, bool autoload)
921 {
922         const struct nft_chain_type *type;
923
924         type = __nf_tables_chain_type_lookup(nla, family);
925         if (type != NULL)
926                 return type;
927
928         lockdep_nfnl_nft_mutex_not_held();
929 #ifdef CONFIG_MODULES
930         if (autoload) {
931                 if (nft_request_module(net, "nft-chain-%u-%.*s", family,
932                                        nla_len(nla),
933                                        (const char *)nla_data(nla)) == -EAGAIN)
934                         return ERR_PTR(-EAGAIN);
935         }
936 #endif
937         return ERR_PTR(-ENOENT);
938 }
939
940 static __be16 nft_base_seq(const struct net *net)
941 {
942         struct nftables_pernet *nft_net = nft_pernet(net);
943
944         return htons(nft_net->base_seq & 0xffff);
945 }
946
947 static const struct nla_policy nft_table_policy[NFTA_TABLE_MAX + 1] = {
948         [NFTA_TABLE_NAME]       = { .type = NLA_STRING,
949                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
950         [NFTA_TABLE_FLAGS]      = { .type = NLA_U32 },
951         [NFTA_TABLE_HANDLE]     = { .type = NLA_U64 },
952         [NFTA_TABLE_USERDATA]   = { .type = NLA_BINARY,
953                                     .len = NFT_USERDATA_MAXLEN }
954 };
955
956 static int nf_tables_fill_table_info(struct sk_buff *skb, struct net *net,
957                                      u32 portid, u32 seq, int event, u32 flags,
958                                      int family, const struct nft_table *table)
959 {
960         struct nlmsghdr *nlh;
961
962         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
963         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
964                            NFNETLINK_V0, nft_base_seq(net));
965         if (!nlh)
966                 goto nla_put_failure;
967
968         if (nla_put_string(skb, NFTA_TABLE_NAME, table->name) ||
969             nla_put_be32(skb, NFTA_TABLE_FLAGS,
970                          htonl(table->flags & NFT_TABLE_F_MASK)) ||
971             nla_put_be32(skb, NFTA_TABLE_USE, htonl(table->use)) ||
972             nla_put_be64(skb, NFTA_TABLE_HANDLE, cpu_to_be64(table->handle),
973                          NFTA_TABLE_PAD))
974                 goto nla_put_failure;
975         if (nft_table_has_owner(table) &&
976             nla_put_be32(skb, NFTA_TABLE_OWNER, htonl(table->nlpid)))
977                 goto nla_put_failure;
978
979         if (table->udata) {
980                 if (nla_put(skb, NFTA_TABLE_USERDATA, table->udlen, table->udata))
981                         goto nla_put_failure;
982         }
983
984         nlmsg_end(skb, nlh);
985         return 0;
986
987 nla_put_failure:
988         nlmsg_trim(skb, nlh);
989         return -1;
990 }
991
992 struct nftnl_skb_parms {
993         bool report;
994 };
995 #define NFT_CB(skb)     (*(struct nftnl_skb_parms*)&((skb)->cb))
996
997 static void nft_notify_enqueue(struct sk_buff *skb, bool report,
998                                struct list_head *notify_list)
999 {
1000         NFT_CB(skb).report = report;
1001         list_add_tail(&skb->list, notify_list);
1002 }
1003
1004 static void nf_tables_table_notify(const struct nft_ctx *ctx, int event)
1005 {
1006         struct nftables_pernet *nft_net;
1007         struct sk_buff *skb;
1008         u16 flags = 0;
1009         int err;
1010
1011         if (!ctx->report &&
1012             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1013                 return;
1014
1015         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1016         if (skb == NULL)
1017                 goto err;
1018
1019         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
1020                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
1021
1022         err = nf_tables_fill_table_info(skb, ctx->net, ctx->portid, ctx->seq,
1023                                         event, flags, ctx->family, ctx->table);
1024         if (err < 0) {
1025                 kfree_skb(skb);
1026                 goto err;
1027         }
1028
1029         nft_net = nft_pernet(ctx->net);
1030         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
1031         return;
1032 err:
1033         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1034 }
1035
1036 static int nf_tables_dump_tables(struct sk_buff *skb,
1037                                  struct netlink_callback *cb)
1038 {
1039         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1040         struct nftables_pernet *nft_net;
1041         const struct nft_table *table;
1042         unsigned int idx = 0, s_idx = cb->args[0];
1043         struct net *net = sock_net(skb->sk);
1044         int family = nfmsg->nfgen_family;
1045
1046         rcu_read_lock();
1047         nft_net = nft_pernet(net);
1048         cb->seq = READ_ONCE(nft_net->base_seq);
1049
1050         list_for_each_entry_rcu(table, &nft_net->tables, list) {
1051                 if (family != NFPROTO_UNSPEC && family != table->family)
1052                         continue;
1053
1054                 if (idx < s_idx)
1055                         goto cont;
1056                 if (idx > s_idx)
1057                         memset(&cb->args[1], 0,
1058                                sizeof(cb->args) - sizeof(cb->args[0]));
1059                 if (!nft_is_active(net, table))
1060                         continue;
1061                 if (nf_tables_fill_table_info(skb, net,
1062                                               NETLINK_CB(cb->skb).portid,
1063                                               cb->nlh->nlmsg_seq,
1064                                               NFT_MSG_NEWTABLE, NLM_F_MULTI,
1065                                               table->family, table) < 0)
1066                         goto done;
1067
1068                 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1069 cont:
1070                 idx++;
1071         }
1072 done:
1073         rcu_read_unlock();
1074         cb->args[0] = idx;
1075         return skb->len;
1076 }
1077
1078 static int nft_netlink_dump_start_rcu(struct sock *nlsk, struct sk_buff *skb,
1079                                       const struct nlmsghdr *nlh,
1080                                       struct netlink_dump_control *c)
1081 {
1082         int err;
1083
1084         if (!try_module_get(THIS_MODULE))
1085                 return -EINVAL;
1086
1087         rcu_read_unlock();
1088         err = netlink_dump_start(nlsk, skb, nlh, c);
1089         rcu_read_lock();
1090         module_put(THIS_MODULE);
1091
1092         return err;
1093 }
1094
1095 /* called with rcu_read_lock held */
1096 static int nf_tables_gettable(struct sk_buff *skb, const struct nfnl_info *info,
1097                               const struct nlattr * const nla[])
1098 {
1099         struct netlink_ext_ack *extack = info->extack;
1100         u8 genmask = nft_genmask_cur(info->net);
1101         u8 family = info->nfmsg->nfgen_family;
1102         const struct nft_table *table;
1103         struct net *net = info->net;
1104         struct sk_buff *skb2;
1105         int err;
1106
1107         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1108                 struct netlink_dump_control c = {
1109                         .dump = nf_tables_dump_tables,
1110                         .module = THIS_MODULE,
1111                 };
1112
1113                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
1114         }
1115
1116         table = nft_table_lookup(net, nla[NFTA_TABLE_NAME], family, genmask, 0);
1117         if (IS_ERR(table)) {
1118                 NL_SET_BAD_ATTR(extack, nla[NFTA_TABLE_NAME]);
1119                 return PTR_ERR(table);
1120         }
1121
1122         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1123         if (!skb2)
1124                 return -ENOMEM;
1125
1126         err = nf_tables_fill_table_info(skb2, net, NETLINK_CB(skb).portid,
1127                                         info->nlh->nlmsg_seq, NFT_MSG_NEWTABLE,
1128                                         0, family, table);
1129         if (err < 0)
1130                 goto err_fill_table_info;
1131
1132         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
1133
1134 err_fill_table_info:
1135         kfree_skb(skb2);
1136         return err;
1137 }
1138
1139 static void nft_table_disable(struct net *net, struct nft_table *table, u32 cnt)
1140 {
1141         struct nft_chain *chain;
1142         u32 i = 0;
1143
1144         list_for_each_entry(chain, &table->chains, list) {
1145                 if (!nft_is_active_next(net, chain))
1146                         continue;
1147                 if (!nft_is_base_chain(chain))
1148                         continue;
1149
1150                 if (cnt && i++ == cnt)
1151                         break;
1152
1153                 nf_tables_unregister_hook(net, table, chain);
1154         }
1155 }
1156
1157 static int nf_tables_table_enable(struct net *net, struct nft_table *table)
1158 {
1159         struct nft_chain *chain;
1160         int err, i = 0;
1161
1162         list_for_each_entry(chain, &table->chains, list) {
1163                 if (!nft_is_active_next(net, chain))
1164                         continue;
1165                 if (!nft_is_base_chain(chain))
1166                         continue;
1167
1168                 err = nf_tables_register_hook(net, table, chain);
1169                 if (err < 0)
1170                         goto err_register_hooks;
1171
1172                 i++;
1173         }
1174         return 0;
1175
1176 err_register_hooks:
1177         if (i)
1178                 nft_table_disable(net, table, i);
1179         return err;
1180 }
1181
1182 static void nf_tables_table_disable(struct net *net, struct nft_table *table)
1183 {
1184         table->flags &= ~NFT_TABLE_F_DORMANT;
1185         nft_table_disable(net, table, 0);
1186         table->flags |= NFT_TABLE_F_DORMANT;
1187 }
1188
1189 #define __NFT_TABLE_F_INTERNAL          (NFT_TABLE_F_MASK + 1)
1190 #define __NFT_TABLE_F_WAS_DORMANT       (__NFT_TABLE_F_INTERNAL << 0)
1191 #define __NFT_TABLE_F_WAS_AWAKEN        (__NFT_TABLE_F_INTERNAL << 1)
1192 #define __NFT_TABLE_F_UPDATE            (__NFT_TABLE_F_WAS_DORMANT | \
1193                                          __NFT_TABLE_F_WAS_AWAKEN)
1194
1195 static bool nft_table_pending_update(const struct nft_ctx *ctx)
1196 {
1197         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
1198         struct nft_trans *trans;
1199
1200         if (ctx->table->flags & __NFT_TABLE_F_UPDATE)
1201                 return true;
1202
1203         list_for_each_entry(trans, &nft_net->commit_list, list) {
1204                 if (trans->ctx.table == ctx->table &&
1205                     trans->msg_type == NFT_MSG_DELCHAIN &&
1206                     nft_is_base_chain(trans->ctx.chain))
1207                         return true;
1208         }
1209
1210         return false;
1211 }
1212
1213 static int nf_tables_updtable(struct nft_ctx *ctx)
1214 {
1215         struct nft_trans *trans;
1216         u32 flags;
1217         int ret;
1218
1219         if (!ctx->nla[NFTA_TABLE_FLAGS])
1220                 return 0;
1221
1222         flags = ntohl(nla_get_be32(ctx->nla[NFTA_TABLE_FLAGS]));
1223         if (flags & ~NFT_TABLE_F_MASK)
1224                 return -EOPNOTSUPP;
1225
1226         if (flags == (ctx->table->flags & NFT_TABLE_F_MASK))
1227                 return 0;
1228
1229         if ((nft_table_has_owner(ctx->table) &&
1230              !(flags & NFT_TABLE_F_OWNER)) ||
1231             (!nft_table_has_owner(ctx->table) &&
1232              flags & NFT_TABLE_F_OWNER))
1233                 return -EOPNOTSUPP;
1234
1235         /* No dormant off/on/off/on games in single transaction */
1236         if (nft_table_pending_update(ctx))
1237                 return -EINVAL;
1238
1239         trans = nft_trans_alloc(ctx, NFT_MSG_NEWTABLE,
1240                                 sizeof(struct nft_trans_table));
1241         if (trans == NULL)
1242                 return -ENOMEM;
1243
1244         if ((flags & NFT_TABLE_F_DORMANT) &&
1245             !(ctx->table->flags & NFT_TABLE_F_DORMANT)) {
1246                 ctx->table->flags |= NFT_TABLE_F_DORMANT;
1247                 if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE))
1248                         ctx->table->flags |= __NFT_TABLE_F_WAS_AWAKEN;
1249         } else if (!(flags & NFT_TABLE_F_DORMANT) &&
1250                    ctx->table->flags & NFT_TABLE_F_DORMANT) {
1251                 ctx->table->flags &= ~NFT_TABLE_F_DORMANT;
1252                 if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE)) {
1253                         ret = nf_tables_table_enable(ctx->net, ctx->table);
1254                         if (ret < 0)
1255                                 goto err_register_hooks;
1256
1257                         ctx->table->flags |= __NFT_TABLE_F_WAS_DORMANT;
1258                 }
1259         }
1260
1261         nft_trans_table_update(trans) = true;
1262         nft_trans_commit_list_add_tail(ctx->net, trans);
1263
1264         return 0;
1265
1266 err_register_hooks:
1267         ctx->table->flags |= NFT_TABLE_F_DORMANT;
1268         nft_trans_destroy(trans);
1269         return ret;
1270 }
1271
1272 static u32 nft_chain_hash(const void *data, u32 len, u32 seed)
1273 {
1274         const char *name = data;
1275
1276         return jhash(name, strlen(name), seed);
1277 }
1278
1279 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed)
1280 {
1281         const struct nft_chain *chain = data;
1282
1283         return nft_chain_hash(chain->name, 0, seed);
1284 }
1285
1286 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *arg,
1287                               const void *ptr)
1288 {
1289         const struct nft_chain *chain = ptr;
1290         const char *name = arg->key;
1291
1292         return strcmp(chain->name, name);
1293 }
1294
1295 static u32 nft_objname_hash(const void *data, u32 len, u32 seed)
1296 {
1297         const struct nft_object_hash_key *k = data;
1298
1299         seed ^= hash_ptr(k->table, 32);
1300
1301         return jhash(k->name, strlen(k->name), seed);
1302 }
1303
1304 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed)
1305 {
1306         const struct nft_object *obj = data;
1307
1308         return nft_objname_hash(&obj->key, 0, seed);
1309 }
1310
1311 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *arg,
1312                                 const void *ptr)
1313 {
1314         const struct nft_object_hash_key *k = arg->key;
1315         const struct nft_object *obj = ptr;
1316
1317         if (obj->key.table != k->table)
1318                 return -1;
1319
1320         return strcmp(obj->key.name, k->name);
1321 }
1322
1323 static bool nft_supported_family(u8 family)
1324 {
1325         return false
1326 #ifdef CONFIG_NF_TABLES_INET
1327                 || family == NFPROTO_INET
1328 #endif
1329 #ifdef CONFIG_NF_TABLES_IPV4
1330                 || family == NFPROTO_IPV4
1331 #endif
1332 #ifdef CONFIG_NF_TABLES_ARP
1333                 || family == NFPROTO_ARP
1334 #endif
1335 #ifdef CONFIG_NF_TABLES_NETDEV
1336                 || family == NFPROTO_NETDEV
1337 #endif
1338 #if IS_ENABLED(CONFIG_NF_TABLES_BRIDGE)
1339                 || family == NFPROTO_BRIDGE
1340 #endif
1341 #ifdef CONFIG_NF_TABLES_IPV6
1342                 || family == NFPROTO_IPV6
1343 #endif
1344                 ;
1345 }
1346
1347 static int nf_tables_newtable(struct sk_buff *skb, const struct nfnl_info *info,
1348                               const struct nlattr * const nla[])
1349 {
1350         struct nftables_pernet *nft_net = nft_pernet(info->net);
1351         struct netlink_ext_ack *extack = info->extack;
1352         u8 genmask = nft_genmask_next(info->net);
1353         u8 family = info->nfmsg->nfgen_family;
1354         struct net *net = info->net;
1355         const struct nlattr *attr;
1356         struct nft_table *table;
1357         struct nft_ctx ctx;
1358         u32 flags = 0;
1359         int err;
1360
1361         if (!nft_supported_family(family))
1362                 return -EOPNOTSUPP;
1363
1364         lockdep_assert_held(&nft_net->commit_mutex);
1365         attr = nla[NFTA_TABLE_NAME];
1366         table = nft_table_lookup(net, attr, family, genmask,
1367                                  NETLINK_CB(skb).portid);
1368         if (IS_ERR(table)) {
1369                 if (PTR_ERR(table) != -ENOENT)
1370                         return PTR_ERR(table);
1371         } else {
1372                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
1373                         NL_SET_BAD_ATTR(extack, attr);
1374                         return -EEXIST;
1375                 }
1376                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
1377                         return -EOPNOTSUPP;
1378
1379                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1380
1381                 return nf_tables_updtable(&ctx);
1382         }
1383
1384         if (nla[NFTA_TABLE_FLAGS]) {
1385                 flags = ntohl(nla_get_be32(nla[NFTA_TABLE_FLAGS]));
1386                 if (flags & ~NFT_TABLE_F_MASK)
1387                         return -EOPNOTSUPP;
1388         }
1389
1390         err = -ENOMEM;
1391         table = kzalloc(sizeof(*table), GFP_KERNEL_ACCOUNT);
1392         if (table == NULL)
1393                 goto err_kzalloc;
1394
1395         table->name = nla_strdup(attr, GFP_KERNEL_ACCOUNT);
1396         if (table->name == NULL)
1397                 goto err_strdup;
1398
1399         if (nla[NFTA_TABLE_USERDATA]) {
1400                 table->udata = nla_memdup(nla[NFTA_TABLE_USERDATA], GFP_KERNEL_ACCOUNT);
1401                 if (table->udata == NULL)
1402                         goto err_table_udata;
1403
1404                 table->udlen = nla_len(nla[NFTA_TABLE_USERDATA]);
1405         }
1406
1407         err = rhltable_init(&table->chains_ht, &nft_chain_ht_params);
1408         if (err)
1409                 goto err_chain_ht;
1410
1411         INIT_LIST_HEAD(&table->chains);
1412         INIT_LIST_HEAD(&table->sets);
1413         INIT_LIST_HEAD(&table->objects);
1414         INIT_LIST_HEAD(&table->flowtables);
1415         table->family = family;
1416         table->flags = flags;
1417         table->handle = ++nft_net->table_handle;
1418         if (table->flags & NFT_TABLE_F_OWNER)
1419                 table->nlpid = NETLINK_CB(skb).portid;
1420
1421         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1422         err = nft_trans_table_add(&ctx, NFT_MSG_NEWTABLE);
1423         if (err < 0)
1424                 goto err_trans;
1425
1426         list_add_tail_rcu(&table->list, &nft_net->tables);
1427         return 0;
1428 err_trans:
1429         rhltable_destroy(&table->chains_ht);
1430 err_chain_ht:
1431         kfree(table->udata);
1432 err_table_udata:
1433         kfree(table->name);
1434 err_strdup:
1435         kfree(table);
1436 err_kzalloc:
1437         return err;
1438 }
1439
1440 static int nft_flush_table(struct nft_ctx *ctx)
1441 {
1442         struct nft_flowtable *flowtable, *nft;
1443         struct nft_chain *chain, *nc;
1444         struct nft_object *obj, *ne;
1445         struct nft_set *set, *ns;
1446         int err;
1447
1448         list_for_each_entry(chain, &ctx->table->chains, list) {
1449                 if (!nft_is_active_next(ctx->net, chain))
1450                         continue;
1451
1452                 if (nft_chain_binding(chain))
1453                         continue;
1454
1455                 ctx->chain = chain;
1456
1457                 err = nft_delrule_by_chain(ctx);
1458                 if (err < 0)
1459                         goto out;
1460         }
1461
1462         list_for_each_entry_safe(set, ns, &ctx->table->sets, list) {
1463                 if (!nft_is_active_next(ctx->net, set))
1464                         continue;
1465
1466                 if (nft_set_is_anonymous(set))
1467                         continue;
1468
1469                 err = nft_delset(ctx, set);
1470                 if (err < 0)
1471                         goto out;
1472         }
1473
1474         list_for_each_entry_safe(flowtable, nft, &ctx->table->flowtables, list) {
1475                 if (!nft_is_active_next(ctx->net, flowtable))
1476                         continue;
1477
1478                 err = nft_delflowtable(ctx, flowtable);
1479                 if (err < 0)
1480                         goto out;
1481         }
1482
1483         list_for_each_entry_safe(obj, ne, &ctx->table->objects, list) {
1484                 if (!nft_is_active_next(ctx->net, obj))
1485                         continue;
1486
1487                 err = nft_delobj(ctx, obj);
1488                 if (err < 0)
1489                         goto out;
1490         }
1491
1492         list_for_each_entry_safe(chain, nc, &ctx->table->chains, list) {
1493                 if (!nft_is_active_next(ctx->net, chain))
1494                         continue;
1495
1496                 if (nft_chain_binding(chain))
1497                         continue;
1498
1499                 ctx->chain = chain;
1500
1501                 err = nft_delchain(ctx);
1502                 if (err < 0)
1503                         goto out;
1504         }
1505
1506         err = nft_deltable(ctx);
1507 out:
1508         return err;
1509 }
1510
1511 static int nft_flush(struct nft_ctx *ctx, int family)
1512 {
1513         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
1514         const struct nlattr * const *nla = ctx->nla;
1515         struct nft_table *table, *nt;
1516         int err = 0;
1517
1518         list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
1519                 if (family != AF_UNSPEC && table->family != family)
1520                         continue;
1521
1522                 ctx->family = table->family;
1523
1524                 if (!nft_is_active_next(ctx->net, table))
1525                         continue;
1526
1527                 if (nft_table_has_owner(table) && table->nlpid != ctx->portid)
1528                         continue;
1529
1530                 if (nla[NFTA_TABLE_NAME] &&
1531                     nla_strcmp(nla[NFTA_TABLE_NAME], table->name) != 0)
1532                         continue;
1533
1534                 ctx->table = table;
1535
1536                 err = nft_flush_table(ctx);
1537                 if (err < 0)
1538                         goto out;
1539         }
1540 out:
1541         return err;
1542 }
1543
1544 static int nf_tables_deltable(struct sk_buff *skb, const struct nfnl_info *info,
1545                               const struct nlattr * const nla[])
1546 {
1547         struct netlink_ext_ack *extack = info->extack;
1548         u8 genmask = nft_genmask_next(info->net);
1549         u8 family = info->nfmsg->nfgen_family;
1550         struct net *net = info->net;
1551         const struct nlattr *attr;
1552         struct nft_table *table;
1553         struct nft_ctx ctx;
1554
1555         nft_ctx_init(&ctx, net, skb, info->nlh, 0, NULL, NULL, nla);
1556         if (family == AF_UNSPEC ||
1557             (!nla[NFTA_TABLE_NAME] && !nla[NFTA_TABLE_HANDLE]))
1558                 return nft_flush(&ctx, family);
1559
1560         if (nla[NFTA_TABLE_HANDLE]) {
1561                 attr = nla[NFTA_TABLE_HANDLE];
1562                 table = nft_table_lookup_byhandle(net, attr, family, genmask,
1563                                                   NETLINK_CB(skb).portid);
1564         } else {
1565                 attr = nla[NFTA_TABLE_NAME];
1566                 table = nft_table_lookup(net, attr, family, genmask,
1567                                          NETLINK_CB(skb).portid);
1568         }
1569
1570         if (IS_ERR(table)) {
1571                 NL_SET_BAD_ATTR(extack, attr);
1572                 return PTR_ERR(table);
1573         }
1574
1575         if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
1576             table->use > 0)
1577                 return -EBUSY;
1578
1579         ctx.family = family;
1580         ctx.table = table;
1581
1582         return nft_flush_table(&ctx);
1583 }
1584
1585 static void nf_tables_table_destroy(struct nft_ctx *ctx)
1586 {
1587         if (WARN_ON(ctx->table->use > 0))
1588                 return;
1589
1590         rhltable_destroy(&ctx->table->chains_ht);
1591         kfree(ctx->table->name);
1592         kfree(ctx->table->udata);
1593         kfree(ctx->table);
1594 }
1595
1596 void nft_register_chain_type(const struct nft_chain_type *ctype)
1597 {
1598         nfnl_lock(NFNL_SUBSYS_NFTABLES);
1599         if (WARN_ON(__nft_chain_type_get(ctype->family, ctype->type))) {
1600                 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1601                 return;
1602         }
1603         chain_type[ctype->family][ctype->type] = ctype;
1604         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1605 }
1606 EXPORT_SYMBOL_GPL(nft_register_chain_type);
1607
1608 void nft_unregister_chain_type(const struct nft_chain_type *ctype)
1609 {
1610         nfnl_lock(NFNL_SUBSYS_NFTABLES);
1611         chain_type[ctype->family][ctype->type] = NULL;
1612         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1613 }
1614 EXPORT_SYMBOL_GPL(nft_unregister_chain_type);
1615
1616 /*
1617  * Chains
1618  */
1619
1620 static struct nft_chain *
1621 nft_chain_lookup_byhandle(const struct nft_table *table, u64 handle, u8 genmask)
1622 {
1623         struct nft_chain *chain;
1624
1625         list_for_each_entry(chain, &table->chains, list) {
1626                 if (chain->handle == handle &&
1627                     nft_active_genmask(chain, genmask))
1628                         return chain;
1629         }
1630
1631         return ERR_PTR(-ENOENT);
1632 }
1633
1634 static bool lockdep_commit_lock_is_held(const struct net *net)
1635 {
1636 #ifdef CONFIG_PROVE_LOCKING
1637         struct nftables_pernet *nft_net = nft_pernet(net);
1638
1639         return lockdep_is_held(&nft_net->commit_mutex);
1640 #else
1641         return true;
1642 #endif
1643 }
1644
1645 static struct nft_chain *nft_chain_lookup(struct net *net,
1646                                           struct nft_table *table,
1647                                           const struct nlattr *nla, u8 genmask)
1648 {
1649         char search[NFT_CHAIN_MAXNAMELEN + 1];
1650         struct rhlist_head *tmp, *list;
1651         struct nft_chain *chain;
1652
1653         if (nla == NULL)
1654                 return ERR_PTR(-EINVAL);
1655
1656         nla_strscpy(search, nla, sizeof(search));
1657
1658         WARN_ON(!rcu_read_lock_held() &&
1659                 !lockdep_commit_lock_is_held(net));
1660
1661         chain = ERR_PTR(-ENOENT);
1662         rcu_read_lock();
1663         list = rhltable_lookup(&table->chains_ht, search, nft_chain_ht_params);
1664         if (!list)
1665                 goto out_unlock;
1666
1667         rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
1668                 if (nft_active_genmask(chain, genmask))
1669                         goto out_unlock;
1670         }
1671         chain = ERR_PTR(-ENOENT);
1672 out_unlock:
1673         rcu_read_unlock();
1674         return chain;
1675 }
1676
1677 static const struct nla_policy nft_chain_policy[NFTA_CHAIN_MAX + 1] = {
1678         [NFTA_CHAIN_TABLE]      = { .type = NLA_STRING,
1679                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
1680         [NFTA_CHAIN_HANDLE]     = { .type = NLA_U64 },
1681         [NFTA_CHAIN_NAME]       = { .type = NLA_STRING,
1682                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
1683         [NFTA_CHAIN_HOOK]       = { .type = NLA_NESTED },
1684         [NFTA_CHAIN_POLICY]     = { .type = NLA_U32 },
1685         [NFTA_CHAIN_TYPE]       = { .type = NLA_STRING,
1686                                     .len = NFT_MODULE_AUTOLOAD_LIMIT },
1687         [NFTA_CHAIN_COUNTERS]   = { .type = NLA_NESTED },
1688         [NFTA_CHAIN_FLAGS]      = { .type = NLA_U32 },
1689         [NFTA_CHAIN_ID]         = { .type = NLA_U32 },
1690         [NFTA_CHAIN_USERDATA]   = { .type = NLA_BINARY,
1691                                     .len = NFT_USERDATA_MAXLEN },
1692 };
1693
1694 static const struct nla_policy nft_hook_policy[NFTA_HOOK_MAX + 1] = {
1695         [NFTA_HOOK_HOOKNUM]     = { .type = NLA_U32 },
1696         [NFTA_HOOK_PRIORITY]    = { .type = NLA_U32 },
1697         [NFTA_HOOK_DEV]         = { .type = NLA_STRING,
1698                                     .len = IFNAMSIZ - 1 },
1699 };
1700
1701 static int nft_dump_stats(struct sk_buff *skb, struct nft_stats __percpu *stats)
1702 {
1703         struct nft_stats *cpu_stats, total;
1704         struct nlattr *nest;
1705         unsigned int seq;
1706         u64 pkts, bytes;
1707         int cpu;
1708
1709         if (!stats)
1710                 return 0;
1711
1712         memset(&total, 0, sizeof(total));
1713         for_each_possible_cpu(cpu) {
1714                 cpu_stats = per_cpu_ptr(stats, cpu);
1715                 do {
1716                         seq = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1717                         pkts = cpu_stats->pkts;
1718                         bytes = cpu_stats->bytes;
1719                 } while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, seq));
1720                 total.pkts += pkts;
1721                 total.bytes += bytes;
1722         }
1723         nest = nla_nest_start_noflag(skb, NFTA_CHAIN_COUNTERS);
1724         if (nest == NULL)
1725                 goto nla_put_failure;
1726
1727         if (nla_put_be64(skb, NFTA_COUNTER_PACKETS, cpu_to_be64(total.pkts),
1728                          NFTA_COUNTER_PAD) ||
1729             nla_put_be64(skb, NFTA_COUNTER_BYTES, cpu_to_be64(total.bytes),
1730                          NFTA_COUNTER_PAD))
1731                 goto nla_put_failure;
1732
1733         nla_nest_end(skb, nest);
1734         return 0;
1735
1736 nla_put_failure:
1737         return -ENOSPC;
1738 }
1739
1740 static int nft_dump_basechain_hook(struct sk_buff *skb, int family,
1741                                    const struct nft_base_chain *basechain)
1742 {
1743         const struct nf_hook_ops *ops = &basechain->ops;
1744         struct nft_hook *hook, *first = NULL;
1745         struct nlattr *nest, *nest_devs;
1746         int n = 0;
1747
1748         nest = nla_nest_start_noflag(skb, NFTA_CHAIN_HOOK);
1749         if (nest == NULL)
1750                 goto nla_put_failure;
1751         if (nla_put_be32(skb, NFTA_HOOK_HOOKNUM, htonl(ops->hooknum)))
1752                 goto nla_put_failure;
1753         if (nla_put_be32(skb, NFTA_HOOK_PRIORITY, htonl(ops->priority)))
1754                 goto nla_put_failure;
1755
1756         if (nft_base_chain_netdev(family, ops->hooknum)) {
1757                 nest_devs = nla_nest_start_noflag(skb, NFTA_HOOK_DEVS);
1758                 list_for_each_entry(hook, &basechain->hook_list, list) {
1759                         if (!first)
1760                                 first = hook;
1761
1762                         if (nla_put_string(skb, NFTA_DEVICE_NAME,
1763                                            hook->ops.dev->name))
1764                                 goto nla_put_failure;
1765                         n++;
1766                 }
1767                 nla_nest_end(skb, nest_devs);
1768
1769                 if (n == 1 &&
1770                     nla_put_string(skb, NFTA_HOOK_DEV, first->ops.dev->name))
1771                         goto nla_put_failure;
1772         }
1773         nla_nest_end(skb, nest);
1774
1775         return 0;
1776 nla_put_failure:
1777         return -1;
1778 }
1779
1780 static int nf_tables_fill_chain_info(struct sk_buff *skb, struct net *net,
1781                                      u32 portid, u32 seq, int event, u32 flags,
1782                                      int family, const struct nft_table *table,
1783                                      const struct nft_chain *chain)
1784 {
1785         struct nlmsghdr *nlh;
1786
1787         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
1788         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
1789                            NFNETLINK_V0, nft_base_seq(net));
1790         if (!nlh)
1791                 goto nla_put_failure;
1792
1793         if (nla_put_string(skb, NFTA_CHAIN_TABLE, table->name))
1794                 goto nla_put_failure;
1795         if (nla_put_be64(skb, NFTA_CHAIN_HANDLE, cpu_to_be64(chain->handle),
1796                          NFTA_CHAIN_PAD))
1797                 goto nla_put_failure;
1798         if (nla_put_string(skb, NFTA_CHAIN_NAME, chain->name))
1799                 goto nla_put_failure;
1800
1801         if (nft_is_base_chain(chain)) {
1802                 const struct nft_base_chain *basechain = nft_base_chain(chain);
1803                 struct nft_stats __percpu *stats;
1804
1805                 if (nft_dump_basechain_hook(skb, family, basechain))
1806                         goto nla_put_failure;
1807
1808                 if (nla_put_be32(skb, NFTA_CHAIN_POLICY,
1809                                  htonl(basechain->policy)))
1810                         goto nla_put_failure;
1811
1812                 if (nla_put_string(skb, NFTA_CHAIN_TYPE, basechain->type->name))
1813                         goto nla_put_failure;
1814
1815                 stats = rcu_dereference_check(basechain->stats,
1816                                               lockdep_commit_lock_is_held(net));
1817                 if (nft_dump_stats(skb, stats))
1818                         goto nla_put_failure;
1819         }
1820
1821         if (chain->flags &&
1822             nla_put_be32(skb, NFTA_CHAIN_FLAGS, htonl(chain->flags)))
1823                 goto nla_put_failure;
1824
1825         if (nla_put_be32(skb, NFTA_CHAIN_USE, htonl(chain->use)))
1826                 goto nla_put_failure;
1827
1828         if (chain->udata &&
1829             nla_put(skb, NFTA_CHAIN_USERDATA, chain->udlen, chain->udata))
1830                 goto nla_put_failure;
1831
1832         nlmsg_end(skb, nlh);
1833         return 0;
1834
1835 nla_put_failure:
1836         nlmsg_trim(skb, nlh);
1837         return -1;
1838 }
1839
1840 static void nf_tables_chain_notify(const struct nft_ctx *ctx, int event)
1841 {
1842         struct nftables_pernet *nft_net;
1843         struct sk_buff *skb;
1844         u16 flags = 0;
1845         int err;
1846
1847         if (!ctx->report &&
1848             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1849                 return;
1850
1851         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1852         if (skb == NULL)
1853                 goto err;
1854
1855         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
1856                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
1857
1858         err = nf_tables_fill_chain_info(skb, ctx->net, ctx->portid, ctx->seq,
1859                                         event, flags, ctx->family, ctx->table,
1860                                         ctx->chain);
1861         if (err < 0) {
1862                 kfree_skb(skb);
1863                 goto err;
1864         }
1865
1866         nft_net = nft_pernet(ctx->net);
1867         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
1868         return;
1869 err:
1870         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1871 }
1872
1873 static int nf_tables_dump_chains(struct sk_buff *skb,
1874                                  struct netlink_callback *cb)
1875 {
1876         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1877         unsigned int idx = 0, s_idx = cb->args[0];
1878         struct net *net = sock_net(skb->sk);
1879         int family = nfmsg->nfgen_family;
1880         struct nftables_pernet *nft_net;
1881         const struct nft_table *table;
1882         const struct nft_chain *chain;
1883
1884         rcu_read_lock();
1885         nft_net = nft_pernet(net);
1886         cb->seq = READ_ONCE(nft_net->base_seq);
1887
1888         list_for_each_entry_rcu(table, &nft_net->tables, list) {
1889                 if (family != NFPROTO_UNSPEC && family != table->family)
1890                         continue;
1891
1892                 list_for_each_entry_rcu(chain, &table->chains, list) {
1893                         if (idx < s_idx)
1894                                 goto cont;
1895                         if (idx > s_idx)
1896                                 memset(&cb->args[1], 0,
1897                                        sizeof(cb->args) - sizeof(cb->args[0]));
1898                         if (!nft_is_active(net, chain))
1899                                 continue;
1900                         if (nf_tables_fill_chain_info(skb, net,
1901                                                       NETLINK_CB(cb->skb).portid,
1902                                                       cb->nlh->nlmsg_seq,
1903                                                       NFT_MSG_NEWCHAIN,
1904                                                       NLM_F_MULTI,
1905                                                       table->family, table,
1906                                                       chain) < 0)
1907                                 goto done;
1908
1909                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1910 cont:
1911                         idx++;
1912                 }
1913         }
1914 done:
1915         rcu_read_unlock();
1916         cb->args[0] = idx;
1917         return skb->len;
1918 }
1919
1920 /* called with rcu_read_lock held */
1921 static int nf_tables_getchain(struct sk_buff *skb, const struct nfnl_info *info,
1922                               const struct nlattr * const nla[])
1923 {
1924         struct netlink_ext_ack *extack = info->extack;
1925         u8 genmask = nft_genmask_cur(info->net);
1926         u8 family = info->nfmsg->nfgen_family;
1927         const struct nft_chain *chain;
1928         struct net *net = info->net;
1929         struct nft_table *table;
1930         struct sk_buff *skb2;
1931         int err;
1932
1933         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1934                 struct netlink_dump_control c = {
1935                         .dump = nf_tables_dump_chains,
1936                         .module = THIS_MODULE,
1937                 };
1938
1939                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
1940         }
1941
1942         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask, 0);
1943         if (IS_ERR(table)) {
1944                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
1945                 return PTR_ERR(table);
1946         }
1947
1948         chain = nft_chain_lookup(net, table, nla[NFTA_CHAIN_NAME], genmask);
1949         if (IS_ERR(chain)) {
1950                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
1951                 return PTR_ERR(chain);
1952         }
1953
1954         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1955         if (!skb2)
1956                 return -ENOMEM;
1957
1958         err = nf_tables_fill_chain_info(skb2, net, NETLINK_CB(skb).portid,
1959                                         info->nlh->nlmsg_seq, NFT_MSG_NEWCHAIN,
1960                                         0, family, table, chain);
1961         if (err < 0)
1962                 goto err_fill_chain_info;
1963
1964         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
1965
1966 err_fill_chain_info:
1967         kfree_skb(skb2);
1968         return err;
1969 }
1970
1971 static const struct nla_policy nft_counter_policy[NFTA_COUNTER_MAX + 1] = {
1972         [NFTA_COUNTER_PACKETS]  = { .type = NLA_U64 },
1973         [NFTA_COUNTER_BYTES]    = { .type = NLA_U64 },
1974 };
1975
1976 static struct nft_stats __percpu *nft_stats_alloc(const struct nlattr *attr)
1977 {
1978         struct nlattr *tb[NFTA_COUNTER_MAX+1];
1979         struct nft_stats __percpu *newstats;
1980         struct nft_stats *stats;
1981         int err;
1982
1983         err = nla_parse_nested_deprecated(tb, NFTA_COUNTER_MAX, attr,
1984                                           nft_counter_policy, NULL);
1985         if (err < 0)
1986                 return ERR_PTR(err);
1987
1988         if (!tb[NFTA_COUNTER_BYTES] || !tb[NFTA_COUNTER_PACKETS])
1989                 return ERR_PTR(-EINVAL);
1990
1991         newstats = netdev_alloc_pcpu_stats(struct nft_stats);
1992         if (newstats == NULL)
1993                 return ERR_PTR(-ENOMEM);
1994
1995         /* Restore old counters on this cpu, no problem. Per-cpu statistics
1996          * are not exposed to userspace.
1997          */
1998         preempt_disable();
1999         stats = this_cpu_ptr(newstats);
2000         stats->bytes = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_BYTES]));
2001         stats->pkts = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_PACKETS]));
2002         preempt_enable();
2003
2004         return newstats;
2005 }
2006
2007 static void nft_chain_stats_replace(struct nft_trans *trans)
2008 {
2009         struct nft_base_chain *chain = nft_base_chain(trans->ctx.chain);
2010
2011         if (!nft_trans_chain_stats(trans))
2012                 return;
2013
2014         nft_trans_chain_stats(trans) =
2015                 rcu_replace_pointer(chain->stats, nft_trans_chain_stats(trans),
2016                                     lockdep_commit_lock_is_held(trans->ctx.net));
2017
2018         if (!nft_trans_chain_stats(trans))
2019                 static_branch_inc(&nft_counters_enabled);
2020 }
2021
2022 static void nf_tables_chain_free_chain_rules(struct nft_chain *chain)
2023 {
2024         struct nft_rule_blob *g0 = rcu_dereference_raw(chain->blob_gen_0);
2025         struct nft_rule_blob *g1 = rcu_dereference_raw(chain->blob_gen_1);
2026
2027         if (g0 != g1)
2028                 kvfree(g1);
2029         kvfree(g0);
2030
2031         /* should be NULL either via abort or via successful commit */
2032         WARN_ON_ONCE(chain->blob_next);
2033         kvfree(chain->blob_next);
2034 }
2035
2036 void nf_tables_chain_destroy(struct nft_ctx *ctx)
2037 {
2038         struct nft_chain *chain = ctx->chain;
2039         struct nft_hook *hook, *next;
2040
2041         if (WARN_ON(chain->use > 0))
2042                 return;
2043
2044         /* no concurrent access possible anymore */
2045         nf_tables_chain_free_chain_rules(chain);
2046
2047         if (nft_is_base_chain(chain)) {
2048                 struct nft_base_chain *basechain = nft_base_chain(chain);
2049
2050                 if (nft_base_chain_netdev(ctx->family, basechain->ops.hooknum)) {
2051                         list_for_each_entry_safe(hook, next,
2052                                                  &basechain->hook_list, list) {
2053                                 list_del_rcu(&hook->list);
2054                                 kfree_rcu(hook, rcu);
2055                         }
2056                 }
2057                 module_put(basechain->type->owner);
2058                 if (rcu_access_pointer(basechain->stats)) {
2059                         static_branch_dec(&nft_counters_enabled);
2060                         free_percpu(rcu_dereference_raw(basechain->stats));
2061                 }
2062                 kfree(chain->name);
2063                 kfree(chain->udata);
2064                 kfree(basechain);
2065         } else {
2066                 kfree(chain->name);
2067                 kfree(chain->udata);
2068                 kfree(chain);
2069         }
2070 }
2071
2072 static struct nft_hook *nft_netdev_hook_alloc(struct net *net,
2073                                               const struct nlattr *attr)
2074 {
2075         struct net_device *dev;
2076         char ifname[IFNAMSIZ];
2077         struct nft_hook *hook;
2078         int err;
2079
2080         hook = kzalloc(sizeof(struct nft_hook), GFP_KERNEL_ACCOUNT);
2081         if (!hook) {
2082                 err = -ENOMEM;
2083                 goto err_hook_alloc;
2084         }
2085
2086         nla_strscpy(ifname, attr, IFNAMSIZ);
2087         /* nf_tables_netdev_event() is called under rtnl_mutex, this is
2088          * indirectly serializing all the other holders of the commit_mutex with
2089          * the rtnl_mutex.
2090          */
2091         dev = __dev_get_by_name(net, ifname);
2092         if (!dev) {
2093                 err = -ENOENT;
2094                 goto err_hook_dev;
2095         }
2096         hook->ops.dev = dev;
2097
2098         return hook;
2099
2100 err_hook_dev:
2101         kfree(hook);
2102 err_hook_alloc:
2103         return ERR_PTR(err);
2104 }
2105
2106 static struct nft_hook *nft_hook_list_find(struct list_head *hook_list,
2107                                            const struct nft_hook *this)
2108 {
2109         struct nft_hook *hook;
2110
2111         list_for_each_entry(hook, hook_list, list) {
2112                 if (this->ops.dev == hook->ops.dev)
2113                         return hook;
2114         }
2115
2116         return NULL;
2117 }
2118
2119 static int nf_tables_parse_netdev_hooks(struct net *net,
2120                                         const struct nlattr *attr,
2121                                         struct list_head *hook_list)
2122 {
2123         struct nft_hook *hook, *next;
2124         const struct nlattr *tmp;
2125         int rem, n = 0, err;
2126
2127         nla_for_each_nested(tmp, attr, rem) {
2128                 if (nla_type(tmp) != NFTA_DEVICE_NAME) {
2129                         err = -EINVAL;
2130                         goto err_hook;
2131                 }
2132
2133                 hook = nft_netdev_hook_alloc(net, tmp);
2134                 if (IS_ERR(hook)) {
2135                         err = PTR_ERR(hook);
2136                         goto err_hook;
2137                 }
2138                 if (nft_hook_list_find(hook_list, hook)) {
2139                         kfree(hook);
2140                         err = -EEXIST;
2141                         goto err_hook;
2142                 }
2143                 list_add_tail(&hook->list, hook_list);
2144                 n++;
2145
2146                 if (n == NFT_NETDEVICE_MAX) {
2147                         err = -EFBIG;
2148                         goto err_hook;
2149                 }
2150         }
2151
2152         return 0;
2153
2154 err_hook:
2155         list_for_each_entry_safe(hook, next, hook_list, list) {
2156                 list_del(&hook->list);
2157                 kfree(hook);
2158         }
2159         return err;
2160 }
2161
2162 struct nft_chain_hook {
2163         u32                             num;
2164         s32                             priority;
2165         const struct nft_chain_type     *type;
2166         struct list_head                list;
2167 };
2168
2169 static int nft_chain_parse_netdev(struct net *net,
2170                                   struct nlattr *tb[],
2171                                   struct list_head *hook_list)
2172 {
2173         struct nft_hook *hook;
2174         int err;
2175
2176         if (tb[NFTA_HOOK_DEV]) {
2177                 hook = nft_netdev_hook_alloc(net, tb[NFTA_HOOK_DEV]);
2178                 if (IS_ERR(hook))
2179                         return PTR_ERR(hook);
2180
2181                 list_add_tail(&hook->list, hook_list);
2182         } else if (tb[NFTA_HOOK_DEVS]) {
2183                 err = nf_tables_parse_netdev_hooks(net, tb[NFTA_HOOK_DEVS],
2184                                                    hook_list);
2185                 if (err < 0)
2186                         return err;
2187
2188                 if (list_empty(hook_list))
2189                         return -EINVAL;
2190         } else {
2191                 return -EINVAL;
2192         }
2193
2194         return 0;
2195 }
2196
2197 static int nft_chain_parse_hook(struct net *net,
2198                                 const struct nlattr * const nla[],
2199                                 struct nft_chain_hook *hook, u8 family,
2200                                 struct netlink_ext_ack *extack, bool autoload)
2201 {
2202         struct nftables_pernet *nft_net = nft_pernet(net);
2203         struct nlattr *ha[NFTA_HOOK_MAX + 1];
2204         const struct nft_chain_type *type;
2205         int err;
2206
2207         lockdep_assert_held(&nft_net->commit_mutex);
2208         lockdep_nfnl_nft_mutex_not_held();
2209
2210         err = nla_parse_nested_deprecated(ha, NFTA_HOOK_MAX,
2211                                           nla[NFTA_CHAIN_HOOK],
2212                                           nft_hook_policy, NULL);
2213         if (err < 0)
2214                 return err;
2215
2216         if (ha[NFTA_HOOK_HOOKNUM] == NULL ||
2217             ha[NFTA_HOOK_PRIORITY] == NULL)
2218                 return -EINVAL;
2219
2220         hook->num = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
2221         hook->priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
2222
2223         type = __nft_chain_type_get(family, NFT_CHAIN_T_DEFAULT);
2224         if (!type)
2225                 return -EOPNOTSUPP;
2226
2227         if (nla[NFTA_CHAIN_TYPE]) {
2228                 type = nf_tables_chain_type_lookup(net, nla[NFTA_CHAIN_TYPE],
2229                                                    family, autoload);
2230                 if (IS_ERR(type)) {
2231                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
2232                         return PTR_ERR(type);
2233                 }
2234         }
2235         if (hook->num >= NFT_MAX_HOOKS || !(type->hook_mask & (1 << hook->num)))
2236                 return -EOPNOTSUPP;
2237
2238         if (type->type == NFT_CHAIN_T_NAT &&
2239             hook->priority <= NF_IP_PRI_CONNTRACK)
2240                 return -EOPNOTSUPP;
2241
2242         if (!try_module_get(type->owner)) {
2243                 if (nla[NFTA_CHAIN_TYPE])
2244                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
2245                 return -ENOENT;
2246         }
2247
2248         hook->type = type;
2249
2250         INIT_LIST_HEAD(&hook->list);
2251         if (nft_base_chain_netdev(family, hook->num)) {
2252                 err = nft_chain_parse_netdev(net, ha, &hook->list);
2253                 if (err < 0) {
2254                         module_put(type->owner);
2255                         return err;
2256                 }
2257         } else if (ha[NFTA_HOOK_DEV] || ha[NFTA_HOOK_DEVS]) {
2258                 module_put(type->owner);
2259                 return -EOPNOTSUPP;
2260         }
2261
2262         return 0;
2263 }
2264
2265 static void nft_chain_release_hook(struct nft_chain_hook *hook)
2266 {
2267         struct nft_hook *h, *next;
2268
2269         list_for_each_entry_safe(h, next, &hook->list, list) {
2270                 list_del(&h->list);
2271                 kfree(h);
2272         }
2273         module_put(hook->type->owner);
2274 }
2275
2276 struct nft_rules_old {
2277         struct rcu_head h;
2278         struct nft_rule_blob *blob;
2279 };
2280
2281 static void nft_last_rule(struct nft_rule_blob *blob, const void *ptr)
2282 {
2283         struct nft_rule_dp *prule;
2284
2285         prule = (struct nft_rule_dp *)ptr;
2286         prule->is_last = 1;
2287         /* blob size does not include the trailer rule */
2288 }
2289
2290 static struct nft_rule_blob *nf_tables_chain_alloc_rules(unsigned int size)
2291 {
2292         struct nft_rule_blob *blob;
2293
2294         /* size must include room for the last rule */
2295         if (size < offsetof(struct nft_rule_dp, data))
2296                 return NULL;
2297
2298         size += sizeof(struct nft_rule_blob) + sizeof(struct nft_rules_old);
2299         if (size > INT_MAX)
2300                 return NULL;
2301
2302         blob = kvmalloc(size, GFP_KERNEL_ACCOUNT);
2303         if (!blob)
2304                 return NULL;
2305
2306         blob->size = 0;
2307         nft_last_rule(blob, blob->data);
2308
2309         return blob;
2310 }
2311
2312 static void nft_basechain_hook_init(struct nf_hook_ops *ops, u8 family,
2313                                     const struct nft_chain_hook *hook,
2314                                     struct nft_chain *chain)
2315 {
2316         ops->pf                 = family;
2317         ops->hooknum            = hook->num;
2318         ops->priority           = hook->priority;
2319         ops->priv               = chain;
2320         ops->hook               = hook->type->hooks[ops->hooknum];
2321         ops->hook_ops_type      = NF_HOOK_OP_NF_TABLES;
2322 }
2323
2324 static int nft_basechain_init(struct nft_base_chain *basechain, u8 family,
2325                               struct nft_chain_hook *hook, u32 flags)
2326 {
2327         struct nft_chain *chain;
2328         struct nft_hook *h;
2329
2330         basechain->type = hook->type;
2331         INIT_LIST_HEAD(&basechain->hook_list);
2332         chain = &basechain->chain;
2333
2334         if (nft_base_chain_netdev(family, hook->num)) {
2335                 list_splice_init(&hook->list, &basechain->hook_list);
2336                 list_for_each_entry(h, &basechain->hook_list, list)
2337                         nft_basechain_hook_init(&h->ops, family, hook, chain);
2338
2339                 basechain->ops.hooknum  = hook->num;
2340                 basechain->ops.priority = hook->priority;
2341         } else {
2342                 nft_basechain_hook_init(&basechain->ops, family, hook, chain);
2343         }
2344
2345         chain->flags |= NFT_CHAIN_BASE | flags;
2346         basechain->policy = NF_ACCEPT;
2347         if (chain->flags & NFT_CHAIN_HW_OFFLOAD &&
2348             !nft_chain_offload_support(basechain)) {
2349                 list_splice_init(&basechain->hook_list, &hook->list);
2350                 return -EOPNOTSUPP;
2351         }
2352
2353         flow_block_init(&basechain->flow_block);
2354
2355         return 0;
2356 }
2357
2358 int nft_chain_add(struct nft_table *table, struct nft_chain *chain)
2359 {
2360         int err;
2361
2362         err = rhltable_insert_key(&table->chains_ht, chain->name,
2363                                   &chain->rhlhead, nft_chain_ht_params);
2364         if (err)
2365                 return err;
2366
2367         list_add_tail_rcu(&chain->list, &table->chains);
2368
2369         return 0;
2370 }
2371
2372 static u64 chain_id;
2373
2374 static int nf_tables_addchain(struct nft_ctx *ctx, u8 family, u8 genmask,
2375                               u8 policy, u32 flags,
2376                               struct netlink_ext_ack *extack)
2377 {
2378         const struct nlattr * const *nla = ctx->nla;
2379         struct nft_table *table = ctx->table;
2380         struct nft_base_chain *basechain;
2381         struct net *net = ctx->net;
2382         char name[NFT_NAME_MAXLEN];
2383         struct nft_rule_blob *blob;
2384         struct nft_trans *trans;
2385         struct nft_chain *chain;
2386         unsigned int data_size;
2387         int err;
2388
2389         if (nla[NFTA_CHAIN_HOOK]) {
2390                 struct nft_stats __percpu *stats = NULL;
2391                 struct nft_chain_hook hook;
2392
2393                 if (table->flags & __NFT_TABLE_F_UPDATE)
2394                         return -EINVAL;
2395
2396                 if (flags & NFT_CHAIN_BINDING)
2397                         return -EOPNOTSUPP;
2398
2399                 err = nft_chain_parse_hook(net, nla, &hook, family, extack,
2400                                            true);
2401                 if (err < 0)
2402                         return err;
2403
2404                 basechain = kzalloc(sizeof(*basechain), GFP_KERNEL_ACCOUNT);
2405                 if (basechain == NULL) {
2406                         nft_chain_release_hook(&hook);
2407                         return -ENOMEM;
2408                 }
2409                 chain = &basechain->chain;
2410
2411                 if (nla[NFTA_CHAIN_COUNTERS]) {
2412                         stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2413                         if (IS_ERR(stats)) {
2414                                 nft_chain_release_hook(&hook);
2415                                 kfree(basechain);
2416                                 return PTR_ERR(stats);
2417                         }
2418                         rcu_assign_pointer(basechain->stats, stats);
2419                 }
2420
2421                 err = nft_basechain_init(basechain, family, &hook, flags);
2422                 if (err < 0) {
2423                         nft_chain_release_hook(&hook);
2424                         kfree(basechain);
2425                         free_percpu(stats);
2426                         return err;
2427                 }
2428                 if (stats)
2429                         static_branch_inc(&nft_counters_enabled);
2430         } else {
2431                 if (flags & NFT_CHAIN_BASE)
2432                         return -EINVAL;
2433                 if (flags & NFT_CHAIN_HW_OFFLOAD)
2434                         return -EOPNOTSUPP;
2435
2436                 chain = kzalloc(sizeof(*chain), GFP_KERNEL_ACCOUNT);
2437                 if (chain == NULL)
2438                         return -ENOMEM;
2439
2440                 chain->flags = flags;
2441         }
2442         ctx->chain = chain;
2443
2444         INIT_LIST_HEAD(&chain->rules);
2445         chain->handle = nf_tables_alloc_handle(table);
2446         chain->table = table;
2447
2448         if (nla[NFTA_CHAIN_NAME]) {
2449                 chain->name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL_ACCOUNT);
2450         } else {
2451                 if (!(flags & NFT_CHAIN_BINDING)) {
2452                         err = -EINVAL;
2453                         goto err_destroy_chain;
2454                 }
2455
2456                 snprintf(name, sizeof(name), "__chain%llu", ++chain_id);
2457                 chain->name = kstrdup(name, GFP_KERNEL_ACCOUNT);
2458         }
2459
2460         if (!chain->name) {
2461                 err = -ENOMEM;
2462                 goto err_destroy_chain;
2463         }
2464
2465         if (nla[NFTA_CHAIN_USERDATA]) {
2466                 chain->udata = nla_memdup(nla[NFTA_CHAIN_USERDATA], GFP_KERNEL_ACCOUNT);
2467                 if (chain->udata == NULL) {
2468                         err = -ENOMEM;
2469                         goto err_destroy_chain;
2470                 }
2471                 chain->udlen = nla_len(nla[NFTA_CHAIN_USERDATA]);
2472         }
2473
2474         data_size = offsetof(struct nft_rule_dp, data); /* last rule */
2475         blob = nf_tables_chain_alloc_rules(data_size);
2476         if (!blob) {
2477                 err = -ENOMEM;
2478                 goto err_destroy_chain;
2479         }
2480
2481         RCU_INIT_POINTER(chain->blob_gen_0, blob);
2482         RCU_INIT_POINTER(chain->blob_gen_1, blob);
2483
2484         if (!nft_use_inc(&table->use)) {
2485                 err = -EMFILE;
2486                 goto err_destroy_chain;
2487         }
2488
2489         trans = nft_trans_chain_add(ctx, NFT_MSG_NEWCHAIN);
2490         if (IS_ERR(trans)) {
2491                 err = PTR_ERR(trans);
2492                 goto err_trans;
2493         }
2494
2495         nft_trans_chain_policy(trans) = NFT_CHAIN_POLICY_UNSET;
2496         if (nft_is_base_chain(chain))
2497                 nft_trans_chain_policy(trans) = policy;
2498
2499         err = nft_chain_add(table, chain);
2500         if (err < 0)
2501                 goto err_chain_add;
2502
2503         /* This must be LAST to ensure no packets are walking over this chain. */
2504         err = nf_tables_register_hook(net, table, chain);
2505         if (err < 0)
2506                 goto err_register_hook;
2507
2508         return 0;
2509
2510 err_register_hook:
2511         nft_chain_del(chain);
2512 err_chain_add:
2513         nft_trans_destroy(trans);
2514 err_trans:
2515         nft_use_dec_restore(&table->use);
2516 err_destroy_chain:
2517         nf_tables_chain_destroy(ctx);
2518
2519         return err;
2520 }
2521
2522 static bool nft_hook_list_equal(struct list_head *hook_list1,
2523                                 struct list_head *hook_list2)
2524 {
2525         struct nft_hook *hook;
2526         int n = 0, m = 0;
2527
2528         n = 0;
2529         list_for_each_entry(hook, hook_list2, list) {
2530                 if (!nft_hook_list_find(hook_list1, hook))
2531                         return false;
2532
2533                 n++;
2534         }
2535         list_for_each_entry(hook, hook_list1, list)
2536                 m++;
2537
2538         return n == m;
2539 }
2540
2541 static int nf_tables_updchain(struct nft_ctx *ctx, u8 genmask, u8 policy,
2542                               u32 flags, const struct nlattr *attr,
2543                               struct netlink_ext_ack *extack)
2544 {
2545         const struct nlattr * const *nla = ctx->nla;
2546         struct nft_table *table = ctx->table;
2547         struct nft_chain *chain = ctx->chain;
2548         struct nft_base_chain *basechain;
2549         struct nft_stats *stats = NULL;
2550         struct nft_chain_hook hook;
2551         struct nf_hook_ops *ops;
2552         struct nft_trans *trans;
2553         int err;
2554
2555         if (chain->flags ^ flags)
2556                 return -EOPNOTSUPP;
2557
2558         if (nla[NFTA_CHAIN_HOOK]) {
2559                 if (!nft_is_base_chain(chain)) {
2560                         NL_SET_BAD_ATTR(extack, attr);
2561                         return -EEXIST;
2562                 }
2563                 err = nft_chain_parse_hook(ctx->net, nla, &hook, ctx->family,
2564                                            extack, false);
2565                 if (err < 0)
2566                         return err;
2567
2568                 basechain = nft_base_chain(chain);
2569                 if (basechain->type != hook.type) {
2570                         nft_chain_release_hook(&hook);
2571                         NL_SET_BAD_ATTR(extack, attr);
2572                         return -EEXIST;
2573                 }
2574
2575                 if (nft_base_chain_netdev(ctx->family, hook.num)) {
2576                         if (!nft_hook_list_equal(&basechain->hook_list,
2577                                                  &hook.list)) {
2578                                 nft_chain_release_hook(&hook);
2579                                 NL_SET_BAD_ATTR(extack, attr);
2580                                 return -EEXIST;
2581                         }
2582                 } else {
2583                         ops = &basechain->ops;
2584                         if (ops->hooknum != hook.num ||
2585                             ops->priority != hook.priority) {
2586                                 nft_chain_release_hook(&hook);
2587                                 NL_SET_BAD_ATTR(extack, attr);
2588                                 return -EEXIST;
2589                         }
2590                 }
2591                 nft_chain_release_hook(&hook);
2592         }
2593
2594         if (nla[NFTA_CHAIN_HANDLE] &&
2595             nla[NFTA_CHAIN_NAME]) {
2596                 struct nft_chain *chain2;
2597
2598                 chain2 = nft_chain_lookup(ctx->net, table,
2599                                           nla[NFTA_CHAIN_NAME], genmask);
2600                 if (!IS_ERR(chain2)) {
2601                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2602                         return -EEXIST;
2603                 }
2604         }
2605
2606         if (nla[NFTA_CHAIN_COUNTERS]) {
2607                 if (!nft_is_base_chain(chain))
2608                         return -EOPNOTSUPP;
2609
2610                 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2611                 if (IS_ERR(stats))
2612                         return PTR_ERR(stats);
2613         }
2614
2615         err = -ENOMEM;
2616         trans = nft_trans_alloc(ctx, NFT_MSG_NEWCHAIN,
2617                                 sizeof(struct nft_trans_chain));
2618         if (trans == NULL)
2619                 goto err;
2620
2621         nft_trans_chain_stats(trans) = stats;
2622         nft_trans_chain_update(trans) = true;
2623
2624         if (nla[NFTA_CHAIN_POLICY])
2625                 nft_trans_chain_policy(trans) = policy;
2626         else
2627                 nft_trans_chain_policy(trans) = -1;
2628
2629         if (nla[NFTA_CHAIN_HANDLE] &&
2630             nla[NFTA_CHAIN_NAME]) {
2631                 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
2632                 struct nft_trans *tmp;
2633                 char *name;
2634
2635                 err = -ENOMEM;
2636                 name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL_ACCOUNT);
2637                 if (!name)
2638                         goto err;
2639
2640                 err = -EEXIST;
2641                 list_for_each_entry(tmp, &nft_net->commit_list, list) {
2642                         if (tmp->msg_type == NFT_MSG_NEWCHAIN &&
2643                             tmp->ctx.table == table &&
2644                             nft_trans_chain_update(tmp) &&
2645                             nft_trans_chain_name(tmp) &&
2646                             strcmp(name, nft_trans_chain_name(tmp)) == 0) {
2647                                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2648                                 kfree(name);
2649                                 goto err;
2650                         }
2651                 }
2652
2653                 nft_trans_chain_name(trans) = name;
2654         }
2655         nft_trans_commit_list_add_tail(ctx->net, trans);
2656
2657         return 0;
2658 err:
2659         free_percpu(stats);
2660         kfree(trans);
2661         return err;
2662 }
2663
2664 static struct nft_chain *nft_chain_lookup_byid(const struct net *net,
2665                                                const struct nft_table *table,
2666                                                const struct nlattr *nla, u8 genmask)
2667 {
2668         struct nftables_pernet *nft_net = nft_pernet(net);
2669         u32 id = ntohl(nla_get_be32(nla));
2670         struct nft_trans *trans;
2671
2672         list_for_each_entry(trans, &nft_net->commit_list, list) {
2673                 struct nft_chain *chain = trans->ctx.chain;
2674
2675                 if (trans->msg_type == NFT_MSG_NEWCHAIN &&
2676                     chain->table == table &&
2677                     id == nft_trans_chain_id(trans) &&
2678                     nft_active_genmask(chain, genmask))
2679                         return chain;
2680         }
2681         return ERR_PTR(-ENOENT);
2682 }
2683
2684 static int nf_tables_newchain(struct sk_buff *skb, const struct nfnl_info *info,
2685                               const struct nlattr * const nla[])
2686 {
2687         struct nftables_pernet *nft_net = nft_pernet(info->net);
2688         struct netlink_ext_ack *extack = info->extack;
2689         u8 genmask = nft_genmask_next(info->net);
2690         u8 family = info->nfmsg->nfgen_family;
2691         struct nft_chain *chain = NULL;
2692         struct net *net = info->net;
2693         const struct nlattr *attr;
2694         struct nft_table *table;
2695         u8 policy = NF_ACCEPT;
2696         struct nft_ctx ctx;
2697         u64 handle = 0;
2698         u32 flags = 0;
2699
2700         lockdep_assert_held(&nft_net->commit_mutex);
2701
2702         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2703                                  NETLINK_CB(skb).portid);
2704         if (IS_ERR(table)) {
2705                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2706                 return PTR_ERR(table);
2707         }
2708
2709         chain = NULL;
2710         attr = nla[NFTA_CHAIN_NAME];
2711
2712         if (nla[NFTA_CHAIN_HANDLE]) {
2713                 handle = be64_to_cpu(nla_get_be64(nla[NFTA_CHAIN_HANDLE]));
2714                 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2715                 if (IS_ERR(chain)) {
2716                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_HANDLE]);
2717                         return PTR_ERR(chain);
2718                 }
2719                 attr = nla[NFTA_CHAIN_HANDLE];
2720         } else if (nla[NFTA_CHAIN_NAME]) {
2721                 chain = nft_chain_lookup(net, table, attr, genmask);
2722                 if (IS_ERR(chain)) {
2723                         if (PTR_ERR(chain) != -ENOENT) {
2724                                 NL_SET_BAD_ATTR(extack, attr);
2725                                 return PTR_ERR(chain);
2726                         }
2727                         chain = NULL;
2728                 }
2729         } else if (!nla[NFTA_CHAIN_ID]) {
2730                 return -EINVAL;
2731         }
2732
2733         if (nla[NFTA_CHAIN_POLICY]) {
2734                 if (chain != NULL &&
2735                     !nft_is_base_chain(chain)) {
2736                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2737                         return -EOPNOTSUPP;
2738                 }
2739
2740                 if (chain == NULL &&
2741                     nla[NFTA_CHAIN_HOOK] == NULL) {
2742                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2743                         return -EOPNOTSUPP;
2744                 }
2745
2746                 policy = ntohl(nla_get_be32(nla[NFTA_CHAIN_POLICY]));
2747                 switch (policy) {
2748                 case NF_DROP:
2749                 case NF_ACCEPT:
2750                         break;
2751                 default:
2752                         return -EINVAL;
2753                 }
2754         }
2755
2756         if (nla[NFTA_CHAIN_FLAGS])
2757                 flags = ntohl(nla_get_be32(nla[NFTA_CHAIN_FLAGS]));
2758         else if (chain)
2759                 flags = chain->flags;
2760
2761         if (flags & ~NFT_CHAIN_FLAGS)
2762                 return -EOPNOTSUPP;
2763
2764         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2765
2766         if (chain != NULL) {
2767                 if (chain->flags & NFT_CHAIN_BINDING)
2768                         return -EINVAL;
2769
2770                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
2771                         NL_SET_BAD_ATTR(extack, attr);
2772                         return -EEXIST;
2773                 }
2774                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
2775                         return -EOPNOTSUPP;
2776
2777                 flags |= chain->flags & NFT_CHAIN_BASE;
2778                 return nf_tables_updchain(&ctx, genmask, policy, flags, attr,
2779                                           extack);
2780         }
2781
2782         return nf_tables_addchain(&ctx, family, genmask, policy, flags, extack);
2783 }
2784
2785 static int nf_tables_delchain(struct sk_buff *skb, const struct nfnl_info *info,
2786                               const struct nlattr * const nla[])
2787 {
2788         struct netlink_ext_ack *extack = info->extack;
2789         u8 genmask = nft_genmask_next(info->net);
2790         u8 family = info->nfmsg->nfgen_family;
2791         struct net *net = info->net;
2792         const struct nlattr *attr;
2793         struct nft_table *table;
2794         struct nft_chain *chain;
2795         struct nft_rule *rule;
2796         struct nft_ctx ctx;
2797         u64 handle;
2798         u32 use;
2799         int err;
2800
2801         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2802                                  NETLINK_CB(skb).portid);
2803         if (IS_ERR(table)) {
2804                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2805                 return PTR_ERR(table);
2806         }
2807
2808         if (nla[NFTA_CHAIN_HANDLE]) {
2809                 attr = nla[NFTA_CHAIN_HANDLE];
2810                 handle = be64_to_cpu(nla_get_be64(attr));
2811                 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2812         } else {
2813                 attr = nla[NFTA_CHAIN_NAME];
2814                 chain = nft_chain_lookup(net, table, attr, genmask);
2815         }
2816         if (IS_ERR(chain)) {
2817                 NL_SET_BAD_ATTR(extack, attr);
2818                 return PTR_ERR(chain);
2819         }
2820
2821         if (nft_chain_binding(chain))
2822                 return -EOPNOTSUPP;
2823
2824         if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
2825             chain->use > 0)
2826                 return -EBUSY;
2827
2828         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2829
2830         use = chain->use;
2831         list_for_each_entry(rule, &chain->rules, list) {
2832                 if (!nft_is_active_next(net, rule))
2833                         continue;
2834                 use--;
2835
2836                 err = nft_delrule(&ctx, rule);
2837                 if (err < 0)
2838                         return err;
2839         }
2840
2841         /* There are rules and elements that are still holding references to us,
2842          * we cannot do a recursive removal in this case.
2843          */
2844         if (use > 0) {
2845                 NL_SET_BAD_ATTR(extack, attr);
2846                 return -EBUSY;
2847         }
2848
2849         return nft_delchain(&ctx);
2850 }
2851
2852 /*
2853  * Expressions
2854  */
2855
2856 /**
2857  *      nft_register_expr - register nf_tables expr type
2858  *      @type: expr type
2859  *
2860  *      Registers the expr type for use with nf_tables. Returns zero on
2861  *      success or a negative errno code otherwise.
2862  */
2863 int nft_register_expr(struct nft_expr_type *type)
2864 {
2865         nfnl_lock(NFNL_SUBSYS_NFTABLES);
2866         if (type->family == NFPROTO_UNSPEC)
2867                 list_add_tail_rcu(&type->list, &nf_tables_expressions);
2868         else
2869                 list_add_rcu(&type->list, &nf_tables_expressions);
2870         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2871         return 0;
2872 }
2873 EXPORT_SYMBOL_GPL(nft_register_expr);
2874
2875 /**
2876  *      nft_unregister_expr - unregister nf_tables expr type
2877  *      @type: expr type
2878  *
2879  *      Unregisters the expr typefor use with nf_tables.
2880  */
2881 void nft_unregister_expr(struct nft_expr_type *type)
2882 {
2883         nfnl_lock(NFNL_SUBSYS_NFTABLES);
2884         list_del_rcu(&type->list);
2885         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2886 }
2887 EXPORT_SYMBOL_GPL(nft_unregister_expr);
2888
2889 static const struct nft_expr_type *__nft_expr_type_get(u8 family,
2890                                                        struct nlattr *nla)
2891 {
2892         const struct nft_expr_type *type, *candidate = NULL;
2893
2894         list_for_each_entry_rcu(type, &nf_tables_expressions, list) {
2895                 if (!nla_strcmp(nla, type->name)) {
2896                         if (!type->family && !candidate)
2897                                 candidate = type;
2898                         else if (type->family == family)
2899                                 candidate = type;
2900                 }
2901         }
2902         return candidate;
2903 }
2904
2905 #ifdef CONFIG_MODULES
2906 static int nft_expr_type_request_module(struct net *net, u8 family,
2907                                         struct nlattr *nla)
2908 {
2909         if (nft_request_module(net, "nft-expr-%u-%.*s", family,
2910                                nla_len(nla), (char *)nla_data(nla)) == -EAGAIN)
2911                 return -EAGAIN;
2912
2913         return 0;
2914 }
2915 #endif
2916
2917 static const struct nft_expr_type *nft_expr_type_get(struct net *net,
2918                                                      u8 family,
2919                                                      struct nlattr *nla)
2920 {
2921         const struct nft_expr_type *type;
2922
2923         if (nla == NULL)
2924                 return ERR_PTR(-EINVAL);
2925
2926         rcu_read_lock();
2927         type = __nft_expr_type_get(family, nla);
2928         if (type != NULL && try_module_get(type->owner)) {
2929                 rcu_read_unlock();
2930                 return type;
2931         }
2932         rcu_read_unlock();
2933
2934         lockdep_nfnl_nft_mutex_not_held();
2935 #ifdef CONFIG_MODULES
2936         if (type == NULL) {
2937                 if (nft_expr_type_request_module(net, family, nla) == -EAGAIN)
2938                         return ERR_PTR(-EAGAIN);
2939
2940                 if (nft_request_module(net, "nft-expr-%.*s",
2941                                        nla_len(nla),
2942                                        (char *)nla_data(nla)) == -EAGAIN)
2943                         return ERR_PTR(-EAGAIN);
2944         }
2945 #endif
2946         return ERR_PTR(-ENOENT);
2947 }
2948
2949 static const struct nla_policy nft_expr_policy[NFTA_EXPR_MAX + 1] = {
2950         [NFTA_EXPR_NAME]        = { .type = NLA_STRING,
2951                                     .len = NFT_MODULE_AUTOLOAD_LIMIT },
2952         [NFTA_EXPR_DATA]        = { .type = NLA_NESTED },
2953 };
2954
2955 static int nf_tables_fill_expr_info(struct sk_buff *skb,
2956                                     const struct nft_expr *expr)
2957 {
2958         if (nla_put_string(skb, NFTA_EXPR_NAME, expr->ops->type->name))
2959                 goto nla_put_failure;
2960
2961         if (expr->ops->dump) {
2962                 struct nlattr *data = nla_nest_start_noflag(skb,
2963                                                             NFTA_EXPR_DATA);
2964                 if (data == NULL)
2965                         goto nla_put_failure;
2966                 if (expr->ops->dump(skb, expr) < 0)
2967                         goto nla_put_failure;
2968                 nla_nest_end(skb, data);
2969         }
2970
2971         return skb->len;
2972
2973 nla_put_failure:
2974         return -1;
2975 };
2976
2977 int nft_expr_dump(struct sk_buff *skb, unsigned int attr,
2978                   const struct nft_expr *expr)
2979 {
2980         struct nlattr *nest;
2981
2982         nest = nla_nest_start_noflag(skb, attr);
2983         if (!nest)
2984                 goto nla_put_failure;
2985         if (nf_tables_fill_expr_info(skb, expr) < 0)
2986                 goto nla_put_failure;
2987         nla_nest_end(skb, nest);
2988         return 0;
2989
2990 nla_put_failure:
2991         return -1;
2992 }
2993
2994 struct nft_expr_info {
2995         const struct nft_expr_ops       *ops;
2996         const struct nlattr             *attr;
2997         struct nlattr                   *tb[NFT_EXPR_MAXATTR + 1];
2998 };
2999
3000 static int nf_tables_expr_parse(const struct nft_ctx *ctx,
3001                                 const struct nlattr *nla,
3002                                 struct nft_expr_info *info)
3003 {
3004         const struct nft_expr_type *type;
3005         const struct nft_expr_ops *ops;
3006         struct nlattr *tb[NFTA_EXPR_MAX + 1];
3007         int err;
3008
3009         err = nla_parse_nested_deprecated(tb, NFTA_EXPR_MAX, nla,
3010                                           nft_expr_policy, NULL);
3011         if (err < 0)
3012                 return err;
3013
3014         type = nft_expr_type_get(ctx->net, ctx->family, tb[NFTA_EXPR_NAME]);
3015         if (IS_ERR(type))
3016                 return PTR_ERR(type);
3017
3018         if (tb[NFTA_EXPR_DATA]) {
3019                 err = nla_parse_nested_deprecated(info->tb, type->maxattr,
3020                                                   tb[NFTA_EXPR_DATA],
3021                                                   type->policy, NULL);
3022                 if (err < 0)
3023                         goto err1;
3024         } else
3025                 memset(info->tb, 0, sizeof(info->tb[0]) * (type->maxattr + 1));
3026
3027         if (type->select_ops != NULL) {
3028                 ops = type->select_ops(ctx,
3029                                        (const struct nlattr * const *)info->tb);
3030                 if (IS_ERR(ops)) {
3031                         err = PTR_ERR(ops);
3032 #ifdef CONFIG_MODULES
3033                         if (err == -EAGAIN)
3034                                 if (nft_expr_type_request_module(ctx->net,
3035                                                                  ctx->family,
3036                                                                  tb[NFTA_EXPR_NAME]) != -EAGAIN)
3037                                         err = -ENOENT;
3038 #endif
3039                         goto err1;
3040                 }
3041         } else
3042                 ops = type->ops;
3043
3044         info->attr = nla;
3045         info->ops = ops;
3046
3047         return 0;
3048
3049 err1:
3050         module_put(type->owner);
3051         return err;
3052 }
3053
3054 static int nf_tables_newexpr(const struct nft_ctx *ctx,
3055                              const struct nft_expr_info *expr_info,
3056                              struct nft_expr *expr)
3057 {
3058         const struct nft_expr_ops *ops = expr_info->ops;
3059         int err;
3060
3061         expr->ops = ops;
3062         if (ops->init) {
3063                 err = ops->init(ctx, expr, (const struct nlattr **)expr_info->tb);
3064                 if (err < 0)
3065                         goto err1;
3066         }
3067
3068         return 0;
3069 err1:
3070         expr->ops = NULL;
3071         return err;
3072 }
3073
3074 static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
3075                                    struct nft_expr *expr)
3076 {
3077         const struct nft_expr_type *type = expr->ops->type;
3078
3079         if (expr->ops->destroy)
3080                 expr->ops->destroy(ctx, expr);
3081         module_put(type->owner);
3082 }
3083
3084 static struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
3085                                       const struct nlattr *nla)
3086 {
3087         struct nft_expr_info expr_info;
3088         struct nft_expr *expr;
3089         struct module *owner;
3090         int err;
3091
3092         err = nf_tables_expr_parse(ctx, nla, &expr_info);
3093         if (err < 0)
3094                 goto err_expr_parse;
3095
3096         err = -EOPNOTSUPP;
3097         if (!(expr_info.ops->type->flags & NFT_EXPR_STATEFUL))
3098                 goto err_expr_stateful;
3099
3100         err = -ENOMEM;
3101         expr = kzalloc(expr_info.ops->size, GFP_KERNEL_ACCOUNT);
3102         if (expr == NULL)
3103                 goto err_expr_stateful;
3104
3105         err = nf_tables_newexpr(ctx, &expr_info, expr);
3106         if (err < 0)
3107                 goto err_expr_new;
3108
3109         return expr;
3110 err_expr_new:
3111         kfree(expr);
3112 err_expr_stateful:
3113         owner = expr_info.ops->type->owner;
3114         if (expr_info.ops->type->release_ops)
3115                 expr_info.ops->type->release_ops(expr_info.ops);
3116
3117         module_put(owner);
3118 err_expr_parse:
3119         return ERR_PTR(err);
3120 }
3121
3122 int nft_expr_clone(struct nft_expr *dst, struct nft_expr *src)
3123 {
3124         int err;
3125
3126         if (src->ops->clone) {
3127                 dst->ops = src->ops;
3128                 err = src->ops->clone(dst, src);
3129                 if (err < 0)
3130                         return err;
3131         } else {
3132                 memcpy(dst, src, src->ops->size);
3133         }
3134
3135         __module_get(src->ops->type->owner);
3136
3137         return 0;
3138 }
3139
3140 void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
3141 {
3142         nf_tables_expr_destroy(ctx, expr);
3143         kfree(expr);
3144 }
3145
3146 /*
3147  * Rules
3148  */
3149
3150 static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
3151                                           u64 handle)
3152 {
3153         struct nft_rule *rule;
3154
3155         // FIXME: this sucks
3156         list_for_each_entry_rcu(rule, &chain->rules, list) {
3157                 if (handle == rule->handle)
3158                         return rule;
3159         }
3160
3161         return ERR_PTR(-ENOENT);
3162 }
3163
3164 static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
3165                                         const struct nlattr *nla)
3166 {
3167         if (nla == NULL)
3168                 return ERR_PTR(-EINVAL);
3169
3170         return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
3171 }
3172
3173 static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
3174         [NFTA_RULE_TABLE]       = { .type = NLA_STRING,
3175                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
3176         [NFTA_RULE_CHAIN]       = { .type = NLA_STRING,
3177                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
3178         [NFTA_RULE_HANDLE]      = { .type = NLA_U64 },
3179         [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
3180         [NFTA_RULE_COMPAT]      = { .type = NLA_NESTED },
3181         [NFTA_RULE_POSITION]    = { .type = NLA_U64 },
3182         [NFTA_RULE_USERDATA]    = { .type = NLA_BINARY,
3183                                     .len = NFT_USERDATA_MAXLEN },
3184         [NFTA_RULE_ID]          = { .type = NLA_U32 },
3185         [NFTA_RULE_POSITION_ID] = { .type = NLA_U32 },
3186         [NFTA_RULE_CHAIN_ID]    = { .type = NLA_U32 },
3187 };
3188
3189 static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
3190                                     u32 portid, u32 seq, int event,
3191                                     u32 flags, int family,
3192                                     const struct nft_table *table,
3193                                     const struct nft_chain *chain,
3194                                     const struct nft_rule *rule, u64 handle)
3195 {
3196         struct nlmsghdr *nlh;
3197         const struct nft_expr *expr, *next;
3198         struct nlattr *list;
3199         u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3200
3201         nlh = nfnl_msg_put(skb, portid, seq, type, flags, family, NFNETLINK_V0,
3202                            nft_base_seq(net));
3203         if (!nlh)
3204                 goto nla_put_failure;
3205
3206         if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
3207                 goto nla_put_failure;
3208         if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
3209                 goto nla_put_failure;
3210         if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
3211                          NFTA_RULE_PAD))
3212                 goto nla_put_failure;
3213
3214         if (event != NFT_MSG_DELRULE && handle) {
3215                 if (nla_put_be64(skb, NFTA_RULE_POSITION, cpu_to_be64(handle),
3216                                  NFTA_RULE_PAD))
3217                         goto nla_put_failure;
3218         }
3219
3220         if (chain->flags & NFT_CHAIN_HW_OFFLOAD)
3221                 nft_flow_rule_stats(chain, rule);
3222
3223         list = nla_nest_start_noflag(skb, NFTA_RULE_EXPRESSIONS);
3224         if (list == NULL)
3225                 goto nla_put_failure;
3226         nft_rule_for_each_expr(expr, next, rule) {
3227                 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
3228                         goto nla_put_failure;
3229         }
3230         nla_nest_end(skb, list);
3231
3232         if (rule->udata) {
3233                 struct nft_userdata *udata = nft_userdata(rule);
3234                 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
3235                             udata->data) < 0)
3236                         goto nla_put_failure;
3237         }
3238
3239         nlmsg_end(skb, nlh);
3240         return 0;
3241
3242 nla_put_failure:
3243         nlmsg_trim(skb, nlh);
3244         return -1;
3245 }
3246
3247 static void nf_tables_rule_notify(const struct nft_ctx *ctx,
3248                                   const struct nft_rule *rule, int event)
3249 {
3250         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3251         const struct nft_rule *prule;
3252         struct sk_buff *skb;
3253         u64 handle = 0;
3254         u16 flags = 0;
3255         int err;
3256
3257         if (!ctx->report &&
3258             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3259                 return;
3260
3261         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
3262         if (skb == NULL)
3263                 goto err;
3264
3265         if (event == NFT_MSG_NEWRULE &&
3266             !list_is_first(&rule->list, &ctx->chain->rules) &&
3267             !list_is_last(&rule->list, &ctx->chain->rules)) {
3268                 prule = list_prev_entry(rule, list);
3269                 handle = prule->handle;
3270         }
3271         if (ctx->flags & (NLM_F_APPEND | NLM_F_REPLACE))
3272                 flags |= NLM_F_APPEND;
3273         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
3274                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
3275
3276         err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
3277                                        event, flags, ctx->family, ctx->table,
3278                                        ctx->chain, rule, handle);
3279         if (err < 0) {
3280                 kfree_skb(skb);
3281                 goto err;
3282         }
3283
3284         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
3285         return;
3286 err:
3287         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
3288 }
3289
3290 struct nft_rule_dump_ctx {
3291         char *table;
3292         char *chain;
3293 };
3294
3295 static int __nf_tables_dump_rules(struct sk_buff *skb,
3296                                   unsigned int *idx,
3297                                   struct netlink_callback *cb,
3298                                   const struct nft_table *table,
3299                                   const struct nft_chain *chain)
3300 {
3301         struct net *net = sock_net(skb->sk);
3302         const struct nft_rule *rule, *prule;
3303         unsigned int s_idx = cb->args[0];
3304         u64 handle;
3305
3306         prule = NULL;
3307         list_for_each_entry_rcu(rule, &chain->rules, list) {
3308                 if (!nft_is_active(net, rule))
3309                         goto cont_skip;
3310                 if (*idx < s_idx)
3311                         goto cont;
3312                 if (prule)
3313                         handle = prule->handle;
3314                 else
3315                         handle = 0;
3316
3317                 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
3318                                         cb->nlh->nlmsg_seq,
3319                                         NFT_MSG_NEWRULE,
3320                                         NLM_F_MULTI | NLM_F_APPEND,
3321                                         table->family,
3322                                         table, chain, rule, handle) < 0)
3323                         return 1;
3324
3325                 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
3326 cont:
3327                 prule = rule;
3328 cont_skip:
3329                 (*idx)++;
3330         }
3331         return 0;
3332 }
3333
3334 static int nf_tables_dump_rules(struct sk_buff *skb,
3335                                 struct netlink_callback *cb)
3336 {
3337         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3338         const struct nft_rule_dump_ctx *ctx = cb->data;
3339         struct nft_table *table;
3340         const struct nft_chain *chain;
3341         unsigned int idx = 0;
3342         struct net *net = sock_net(skb->sk);
3343         int family = nfmsg->nfgen_family;
3344         struct nftables_pernet *nft_net;
3345
3346         rcu_read_lock();
3347         nft_net = nft_pernet(net);
3348         cb->seq = READ_ONCE(nft_net->base_seq);
3349
3350         list_for_each_entry_rcu(table, &nft_net->tables, list) {
3351                 if (family != NFPROTO_UNSPEC && family != table->family)
3352                         continue;
3353
3354                 if (ctx && ctx->table && strcmp(ctx->table, table->name) != 0)
3355                         continue;
3356
3357                 if (ctx && ctx->table && ctx->chain) {
3358                         struct rhlist_head *list, *tmp;
3359
3360                         list = rhltable_lookup(&table->chains_ht, ctx->chain,
3361                                                nft_chain_ht_params);
3362                         if (!list)
3363                                 goto done;
3364
3365                         rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
3366                                 if (!nft_is_active(net, chain))
3367                                         continue;
3368                                 __nf_tables_dump_rules(skb, &idx,
3369                                                        cb, table, chain);
3370                                 break;
3371                         }
3372                         goto done;
3373                 }
3374
3375                 list_for_each_entry_rcu(chain, &table->chains, list) {
3376                         if (__nf_tables_dump_rules(skb, &idx, cb, table, chain))
3377                                 goto done;
3378                 }
3379
3380                 if (ctx && ctx->table)
3381                         break;
3382         }
3383 done:
3384         rcu_read_unlock();
3385
3386         cb->args[0] = idx;
3387         return skb->len;
3388 }
3389
3390 static int nf_tables_dump_rules_start(struct netlink_callback *cb)
3391 {
3392         const struct nlattr * const *nla = cb->data;
3393         struct nft_rule_dump_ctx *ctx = NULL;
3394
3395         if (nla[NFTA_RULE_TABLE] || nla[NFTA_RULE_CHAIN]) {
3396                 ctx = kzalloc(sizeof(*ctx), GFP_ATOMIC);
3397                 if (!ctx)
3398                         return -ENOMEM;
3399
3400                 if (nla[NFTA_RULE_TABLE]) {
3401                         ctx->table = nla_strdup(nla[NFTA_RULE_TABLE],
3402                                                         GFP_ATOMIC);
3403                         if (!ctx->table) {
3404                                 kfree(ctx);
3405                                 return -ENOMEM;
3406                         }
3407                 }
3408                 if (nla[NFTA_RULE_CHAIN]) {
3409                         ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN],
3410                                                 GFP_ATOMIC);
3411                         if (!ctx->chain) {
3412                                 kfree(ctx->table);
3413                                 kfree(ctx);
3414                                 return -ENOMEM;
3415                         }
3416                 }
3417         }
3418
3419         cb->data = ctx;
3420         return 0;
3421 }
3422
3423 static int nf_tables_dump_rules_done(struct netlink_callback *cb)
3424 {
3425         struct nft_rule_dump_ctx *ctx = cb->data;
3426
3427         if (ctx) {
3428                 kfree(ctx->table);
3429                 kfree(ctx->chain);
3430                 kfree(ctx);
3431         }
3432         return 0;
3433 }
3434
3435 /* called with rcu_read_lock held */
3436 static int nf_tables_getrule(struct sk_buff *skb, const struct nfnl_info *info,
3437                              const struct nlattr * const nla[])
3438 {
3439         struct netlink_ext_ack *extack = info->extack;
3440         u8 genmask = nft_genmask_cur(info->net);
3441         u8 family = info->nfmsg->nfgen_family;
3442         const struct nft_chain *chain;
3443         const struct nft_rule *rule;
3444         struct net *net = info->net;
3445         struct nft_table *table;
3446         struct sk_buff *skb2;
3447         int err;
3448
3449         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3450                 struct netlink_dump_control c = {
3451                         .start= nf_tables_dump_rules_start,
3452                         .dump = nf_tables_dump_rules,
3453                         .done = nf_tables_dump_rules_done,
3454                         .module = THIS_MODULE,
3455                         .data = (void *)nla,
3456                 };
3457
3458                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
3459         }
3460
3461         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask, 0);
3462         if (IS_ERR(table)) {
3463                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3464                 return PTR_ERR(table);
3465         }
3466
3467         chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
3468         if (IS_ERR(chain)) {
3469                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3470                 return PTR_ERR(chain);
3471         }
3472
3473         rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3474         if (IS_ERR(rule)) {
3475                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3476                 return PTR_ERR(rule);
3477         }
3478
3479         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3480         if (!skb2)
3481                 return -ENOMEM;
3482
3483         err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
3484                                        info->nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
3485                                        family, table, chain, rule, 0);
3486         if (err < 0)
3487                 goto err_fill_rule_info;
3488
3489         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
3490
3491 err_fill_rule_info:
3492         kfree_skb(skb2);
3493         return err;
3494 }
3495
3496 void nf_tables_rule_destroy(const struct nft_ctx *ctx, struct nft_rule *rule)
3497 {
3498         struct nft_expr *expr, *next;
3499
3500         /*
3501          * Careful: some expressions might not be initialized in case this
3502          * is called on error from nf_tables_newrule().
3503          */
3504         expr = nft_expr_first(rule);
3505         while (nft_expr_more(rule, expr)) {
3506                 next = nft_expr_next(expr);
3507                 nf_tables_expr_destroy(ctx, expr);
3508                 expr = next;
3509         }
3510         kfree(rule);
3511 }
3512
3513 static void nf_tables_rule_release(const struct nft_ctx *ctx, struct nft_rule *rule)
3514 {
3515         nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_RELEASE);
3516         nf_tables_rule_destroy(ctx, rule);
3517 }
3518
3519 int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
3520 {
3521         struct nft_expr *expr, *last;
3522         const struct nft_data *data;
3523         struct nft_rule *rule;
3524         int err;
3525
3526         if (ctx->level == NFT_JUMP_STACK_SIZE)
3527                 return -EMLINK;
3528
3529         list_for_each_entry(rule, &chain->rules, list) {
3530                 if (!nft_is_active_next(ctx->net, rule))
3531                         continue;
3532
3533                 nft_rule_for_each_expr(expr, last, rule) {
3534                         if (!expr->ops->validate)
3535                                 continue;
3536
3537                         err = expr->ops->validate(ctx, expr, &data);
3538                         if (err < 0)
3539                                 return err;
3540                 }
3541         }
3542
3543         return 0;
3544 }
3545 EXPORT_SYMBOL_GPL(nft_chain_validate);
3546
3547 static int nft_table_validate(struct net *net, const struct nft_table *table)
3548 {
3549         struct nft_chain *chain;
3550         struct nft_ctx ctx = {
3551                 .net    = net,
3552                 .family = table->family,
3553         };
3554         int err;
3555
3556         list_for_each_entry(chain, &table->chains, list) {
3557                 if (!nft_is_base_chain(chain))
3558                         continue;
3559
3560                 ctx.chain = chain;
3561                 err = nft_chain_validate(&ctx, chain);
3562                 if (err < 0)
3563                         return err;
3564
3565                 cond_resched();
3566         }
3567
3568         return 0;
3569 }
3570
3571 int nft_setelem_validate(const struct nft_ctx *ctx, struct nft_set *set,
3572                          const struct nft_set_iter *iter,
3573                          struct nft_set_elem *elem)
3574 {
3575         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
3576         struct nft_ctx *pctx = (struct nft_ctx *)ctx;
3577         const struct nft_data *data;
3578         int err;
3579
3580         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
3581             *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
3582                 return 0;
3583
3584         data = nft_set_ext_data(ext);
3585         switch (data->verdict.code) {
3586         case NFT_JUMP:
3587         case NFT_GOTO:
3588                 pctx->level++;
3589                 err = nft_chain_validate(ctx, data->verdict.chain);
3590                 if (err < 0)
3591                         return err;
3592                 pctx->level--;
3593                 break;
3594         default:
3595                 break;
3596         }
3597
3598         return 0;
3599 }
3600
3601 int nft_set_catchall_validate(const struct nft_ctx *ctx, struct nft_set *set)
3602 {
3603         u8 genmask = nft_genmask_next(ctx->net);
3604         struct nft_set_elem_catchall *catchall;
3605         struct nft_set_elem elem;
3606         struct nft_set_ext *ext;
3607         int ret = 0;
3608
3609         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
3610                 ext = nft_set_elem_ext(set, catchall->elem);
3611                 if (!nft_set_elem_active(ext, genmask))
3612                         continue;
3613
3614                 elem.priv = catchall->elem;
3615                 ret = nft_setelem_validate(ctx, set, NULL, &elem);
3616                 if (ret < 0)
3617                         return ret;
3618         }
3619
3620         return ret;
3621 }
3622
3623 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3624                                              const struct nft_chain *chain,
3625                                              const struct nlattr *nla);
3626
3627 #define NFT_RULE_MAXEXPRS       128
3628
3629 static int nf_tables_newrule(struct sk_buff *skb, const struct nfnl_info *info,
3630                              const struct nlattr * const nla[])
3631 {
3632         struct nftables_pernet *nft_net = nft_pernet(info->net);
3633         struct netlink_ext_ack *extack = info->extack;
3634         unsigned int size, i, n, ulen = 0, usize = 0;
3635         u8 genmask = nft_genmask_next(info->net);
3636         struct nft_rule *rule, *old_rule = NULL;
3637         struct nft_expr_info *expr_info = NULL;
3638         u8 family = info->nfmsg->nfgen_family;
3639         struct nft_flow_rule *flow = NULL;
3640         struct net *net = info->net;
3641         struct nft_userdata *udata;
3642         struct nft_table *table;
3643         struct nft_chain *chain;
3644         struct nft_trans *trans;
3645         u64 handle, pos_handle;
3646         struct nft_expr *expr;
3647         struct nft_ctx ctx;
3648         struct nlattr *tmp;
3649         int err, rem;
3650
3651         lockdep_assert_held(&nft_net->commit_mutex);
3652
3653         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3654                                  NETLINK_CB(skb).portid);
3655         if (IS_ERR(table)) {
3656                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3657                 return PTR_ERR(table);
3658         }
3659
3660         if (nla[NFTA_RULE_CHAIN]) {
3661                 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3662                                          genmask);
3663                 if (IS_ERR(chain)) {
3664                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3665                         return PTR_ERR(chain);
3666                 }
3667
3668         } else if (nla[NFTA_RULE_CHAIN_ID]) {
3669                 chain = nft_chain_lookup_byid(net, table, nla[NFTA_RULE_CHAIN_ID],
3670                                               genmask);
3671                 if (IS_ERR(chain)) {
3672                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN_ID]);
3673                         return PTR_ERR(chain);
3674                 }
3675         } else {
3676                 return -EINVAL;
3677         }
3678
3679         if (nft_chain_is_bound(chain))
3680                 return -EOPNOTSUPP;
3681
3682         if (nla[NFTA_RULE_HANDLE]) {
3683                 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
3684                 rule = __nft_rule_lookup(chain, handle);
3685                 if (IS_ERR(rule)) {
3686                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3687                         return PTR_ERR(rule);
3688                 }
3689
3690                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
3691                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3692                         return -EEXIST;
3693                 }
3694                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
3695                         old_rule = rule;
3696                 else
3697                         return -EOPNOTSUPP;
3698         } else {
3699                 if (!(info->nlh->nlmsg_flags & NLM_F_CREATE) ||
3700                     info->nlh->nlmsg_flags & NLM_F_REPLACE)
3701                         return -EINVAL;
3702                 handle = nf_tables_alloc_handle(table);
3703
3704                 if (nla[NFTA_RULE_POSITION]) {
3705                         pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
3706                         old_rule = __nft_rule_lookup(chain, pos_handle);
3707                         if (IS_ERR(old_rule)) {
3708                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
3709                                 return PTR_ERR(old_rule);
3710                         }
3711                 } else if (nla[NFTA_RULE_POSITION_ID]) {
3712                         old_rule = nft_rule_lookup_byid(net, chain, nla[NFTA_RULE_POSITION_ID]);
3713                         if (IS_ERR(old_rule)) {
3714                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION_ID]);
3715                                 return PTR_ERR(old_rule);
3716                         }
3717                 }
3718         }
3719
3720         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3721
3722         n = 0;
3723         size = 0;
3724         if (nla[NFTA_RULE_EXPRESSIONS]) {
3725                 expr_info = kvmalloc_array(NFT_RULE_MAXEXPRS,
3726                                            sizeof(struct nft_expr_info),
3727                                            GFP_KERNEL);
3728                 if (!expr_info)
3729                         return -ENOMEM;
3730
3731                 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
3732                         err = -EINVAL;
3733                         if (nla_type(tmp) != NFTA_LIST_ELEM)
3734                                 goto err_release_expr;
3735                         if (n == NFT_RULE_MAXEXPRS)
3736                                 goto err_release_expr;
3737                         err = nf_tables_expr_parse(&ctx, tmp, &expr_info[n]);
3738                         if (err < 0) {
3739                                 NL_SET_BAD_ATTR(extack, tmp);
3740                                 goto err_release_expr;
3741                         }
3742                         size += expr_info[n].ops->size;
3743                         n++;
3744                 }
3745         }
3746         /* Check for overflow of dlen field */
3747         err = -EFBIG;
3748         if (size >= 1 << 12)
3749                 goto err_release_expr;
3750
3751         if (nla[NFTA_RULE_USERDATA]) {
3752                 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
3753                 if (ulen > 0)
3754                         usize = sizeof(struct nft_userdata) + ulen;
3755         }
3756
3757         err = -ENOMEM;
3758         rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL_ACCOUNT);
3759         if (rule == NULL)
3760                 goto err_release_expr;
3761
3762         nft_activate_next(net, rule);
3763
3764         rule->handle = handle;
3765         rule->dlen   = size;
3766         rule->udata  = ulen ? 1 : 0;
3767
3768         if (ulen) {
3769                 udata = nft_userdata(rule);
3770                 udata->len = ulen - 1;
3771                 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
3772         }
3773
3774         expr = nft_expr_first(rule);
3775         for (i = 0; i < n; i++) {
3776                 err = nf_tables_newexpr(&ctx, &expr_info[i], expr);
3777                 if (err < 0) {
3778                         NL_SET_BAD_ATTR(extack, expr_info[i].attr);
3779                         goto err_release_rule;
3780                 }
3781
3782                 if (expr_info[i].ops->validate)
3783                         nft_validate_state_update(net, NFT_VALIDATE_NEED);
3784
3785                 expr_info[i].ops = NULL;
3786                 expr = nft_expr_next(expr);
3787         }
3788
3789         if (chain->flags & NFT_CHAIN_HW_OFFLOAD) {
3790                 flow = nft_flow_rule_create(net, rule);
3791                 if (IS_ERR(flow)) {
3792                         err = PTR_ERR(flow);
3793                         goto err_release_rule;
3794                 }
3795         }
3796
3797         if (!nft_use_inc(&chain->use)) {
3798                 err = -EMFILE;
3799                 goto err_release_rule;
3800         }
3801
3802         if (info->nlh->nlmsg_flags & NLM_F_REPLACE) {
3803                 if (nft_chain_binding(chain)) {
3804                         err = -EOPNOTSUPP;
3805                         goto err_destroy_flow_rule;
3806                 }
3807
3808                 err = nft_delrule(&ctx, old_rule);
3809                 if (err < 0)
3810                         goto err_destroy_flow_rule;
3811
3812                 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3813                 if (trans == NULL) {
3814                         err = -ENOMEM;
3815                         goto err_destroy_flow_rule;
3816                 }
3817                 list_add_tail_rcu(&rule->list, &old_rule->list);
3818         } else {
3819                 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3820                 if (!trans) {
3821                         err = -ENOMEM;
3822                         goto err_destroy_flow_rule;
3823                 }
3824
3825                 if (info->nlh->nlmsg_flags & NLM_F_APPEND) {
3826                         if (old_rule)
3827                                 list_add_rcu(&rule->list, &old_rule->list);
3828                         else
3829                                 list_add_tail_rcu(&rule->list, &chain->rules);
3830                  } else {
3831                         if (old_rule)
3832                                 list_add_tail_rcu(&rule->list, &old_rule->list);
3833                         else
3834                                 list_add_rcu(&rule->list, &chain->rules);
3835                 }
3836         }
3837         kvfree(expr_info);
3838
3839         if (flow)
3840                 nft_trans_flow_rule(trans) = flow;
3841
3842         if (nft_net->validate_state == NFT_VALIDATE_DO)
3843                 return nft_table_validate(net, table);
3844
3845         return 0;
3846
3847 err_destroy_flow_rule:
3848         nft_use_dec_restore(&chain->use);
3849         if (flow)
3850                 nft_flow_rule_destroy(flow);
3851 err_release_rule:
3852         nft_rule_expr_deactivate(&ctx, rule, NFT_TRANS_PREPARE_ERROR);
3853         nf_tables_rule_destroy(&ctx, rule);
3854 err_release_expr:
3855         for (i = 0; i < n; i++) {
3856                 if (expr_info[i].ops) {
3857                         module_put(expr_info[i].ops->type->owner);
3858                         if (expr_info[i].ops->type->release_ops)
3859                                 expr_info[i].ops->type->release_ops(expr_info[i].ops);
3860                 }
3861         }
3862         kvfree(expr_info);
3863
3864         return err;
3865 }
3866
3867 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3868                                              const struct nft_chain *chain,
3869                                              const struct nlattr *nla)
3870 {
3871         struct nftables_pernet *nft_net = nft_pernet(net);
3872         u32 id = ntohl(nla_get_be32(nla));
3873         struct nft_trans *trans;
3874
3875         list_for_each_entry(trans, &nft_net->commit_list, list) {
3876                 if (trans->msg_type == NFT_MSG_NEWRULE &&
3877                     trans->ctx.chain == chain &&
3878                     id == nft_trans_rule_id(trans))
3879                         return nft_trans_rule(trans);
3880         }
3881         return ERR_PTR(-ENOENT);
3882 }
3883
3884 static int nf_tables_delrule(struct sk_buff *skb, const struct nfnl_info *info,
3885                              const struct nlattr * const nla[])
3886 {
3887         struct netlink_ext_ack *extack = info->extack;
3888         u8 genmask = nft_genmask_next(info->net);
3889         u8 family = info->nfmsg->nfgen_family;
3890         struct nft_chain *chain = NULL;
3891         struct net *net = info->net;
3892         struct nft_table *table;
3893         struct nft_rule *rule;
3894         struct nft_ctx ctx;
3895         int err = 0;
3896
3897         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3898                                  NETLINK_CB(skb).portid);
3899         if (IS_ERR(table)) {
3900                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3901                 return PTR_ERR(table);
3902         }
3903
3904         if (nla[NFTA_RULE_CHAIN]) {
3905                 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3906                                          genmask);
3907                 if (IS_ERR(chain)) {
3908                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3909                         return PTR_ERR(chain);
3910                 }
3911                 if (nft_chain_binding(chain))
3912                         return -EOPNOTSUPP;
3913         }
3914
3915         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3916
3917         if (chain) {
3918                 if (nla[NFTA_RULE_HANDLE]) {
3919                         rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3920                         if (IS_ERR(rule)) {
3921                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3922                                 return PTR_ERR(rule);
3923                         }
3924
3925                         err = nft_delrule(&ctx, rule);
3926                 } else if (nla[NFTA_RULE_ID]) {
3927                         rule = nft_rule_lookup_byid(net, chain, nla[NFTA_RULE_ID]);
3928                         if (IS_ERR(rule)) {
3929                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
3930                                 return PTR_ERR(rule);
3931                         }
3932
3933                         err = nft_delrule(&ctx, rule);
3934                 } else {
3935                         err = nft_delrule_by_chain(&ctx);
3936                 }
3937         } else {
3938                 list_for_each_entry(chain, &table->chains, list) {
3939                         if (!nft_is_active_next(net, chain))
3940                                 continue;
3941                         if (nft_chain_binding(chain))
3942                                 continue;
3943
3944                         ctx.chain = chain;
3945                         err = nft_delrule_by_chain(&ctx);
3946                         if (err < 0)
3947                                 break;
3948                 }
3949         }
3950
3951         return err;
3952 }
3953
3954 /*
3955  * Sets
3956  */
3957 static const struct nft_set_type *nft_set_types[] = {
3958         &nft_set_hash_fast_type,
3959         &nft_set_hash_type,
3960         &nft_set_rhash_type,
3961         &nft_set_bitmap_type,
3962         &nft_set_rbtree_type,
3963 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
3964         &nft_set_pipapo_avx2_type,
3965 #endif
3966         &nft_set_pipapo_type,
3967 };
3968
3969 #define NFT_SET_FEATURES        (NFT_SET_INTERVAL | NFT_SET_MAP | \
3970                                  NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
3971                                  NFT_SET_EVAL)
3972
3973 static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
3974 {
3975         return (flags & type->features) == (flags & NFT_SET_FEATURES);
3976 }
3977
3978 /*
3979  * Select a set implementation based on the data characteristics and the
3980  * given policy. The total memory use might not be known if no size is
3981  * given, in that case the amount of memory per element is used.
3982  */
3983 static const struct nft_set_ops *
3984 nft_select_set_ops(const struct nft_ctx *ctx,
3985                    const struct nlattr * const nla[],
3986                    const struct nft_set_desc *desc)
3987 {
3988         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3989         const struct nft_set_ops *ops, *bops;
3990         struct nft_set_estimate est, best;
3991         const struct nft_set_type *type;
3992         u32 flags = 0;
3993         int i;
3994
3995         lockdep_assert_held(&nft_net->commit_mutex);
3996         lockdep_nfnl_nft_mutex_not_held();
3997
3998         if (nla[NFTA_SET_FLAGS] != NULL)
3999                 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
4000
4001         bops        = NULL;
4002         best.size   = ~0;
4003         best.lookup = ~0;
4004         best.space  = ~0;
4005
4006         for (i = 0; i < ARRAY_SIZE(nft_set_types); i++) {
4007                 type = nft_set_types[i];
4008                 ops = &type->ops;
4009
4010                 if (!nft_set_ops_candidate(type, flags))
4011                         continue;
4012                 if (!ops->estimate(desc, flags, &est))
4013                         continue;
4014
4015                 switch (desc->policy) {
4016                 case NFT_SET_POL_PERFORMANCE:
4017                         if (est.lookup < best.lookup)
4018                                 break;
4019                         if (est.lookup == best.lookup &&
4020                             est.space < best.space)
4021                                 break;
4022                         continue;
4023                 case NFT_SET_POL_MEMORY:
4024                         if (!desc->size) {
4025                                 if (est.space < best.space)
4026                                         break;
4027                                 if (est.space == best.space &&
4028                                     est.lookup < best.lookup)
4029                                         break;
4030                         } else if (est.size < best.size || !bops) {
4031                                 break;
4032                         }
4033                         continue;
4034                 default:
4035                         break;
4036                 }
4037
4038                 bops = ops;
4039                 best = est;
4040         }
4041
4042         if (bops != NULL)
4043                 return bops;
4044
4045         return ERR_PTR(-EOPNOTSUPP);
4046 }
4047
4048 static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
4049         [NFTA_SET_TABLE]                = { .type = NLA_STRING,
4050                                             .len = NFT_TABLE_MAXNAMELEN - 1 },
4051         [NFTA_SET_NAME]                 = { .type = NLA_STRING,
4052                                             .len = NFT_SET_MAXNAMELEN - 1 },
4053         [NFTA_SET_FLAGS]                = { .type = NLA_U32 },
4054         [NFTA_SET_KEY_TYPE]             = { .type = NLA_U32 },
4055         [NFTA_SET_KEY_LEN]              = { .type = NLA_U32 },
4056         [NFTA_SET_DATA_TYPE]            = { .type = NLA_U32 },
4057         [NFTA_SET_DATA_LEN]             = { .type = NLA_U32 },
4058         [NFTA_SET_POLICY]               = { .type = NLA_U32 },
4059         [NFTA_SET_DESC]                 = { .type = NLA_NESTED },
4060         [NFTA_SET_ID]                   = { .type = NLA_U32 },
4061         [NFTA_SET_TIMEOUT]              = { .type = NLA_U64 },
4062         [NFTA_SET_GC_INTERVAL]          = { .type = NLA_U32 },
4063         [NFTA_SET_USERDATA]             = { .type = NLA_BINARY,
4064                                             .len  = NFT_USERDATA_MAXLEN },
4065         [NFTA_SET_OBJ_TYPE]             = { .type = NLA_U32 },
4066         [NFTA_SET_HANDLE]               = { .type = NLA_U64 },
4067         [NFTA_SET_EXPR]                 = { .type = NLA_NESTED },
4068         [NFTA_SET_EXPRESSIONS]          = { .type = NLA_NESTED },
4069 };
4070
4071 static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
4072         [NFTA_SET_DESC_SIZE]            = { .type = NLA_U32 },
4073         [NFTA_SET_DESC_CONCAT]          = { .type = NLA_NESTED },
4074 };
4075
4076 static struct nft_set *nft_set_lookup(const struct nft_table *table,
4077                                       const struct nlattr *nla, u8 genmask)
4078 {
4079         struct nft_set *set;
4080
4081         if (nla == NULL)
4082                 return ERR_PTR(-EINVAL);
4083
4084         list_for_each_entry_rcu(set, &table->sets, list) {
4085                 if (!nla_strcmp(nla, set->name) &&
4086                     nft_active_genmask(set, genmask))
4087                         return set;
4088         }
4089         return ERR_PTR(-ENOENT);
4090 }
4091
4092 static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
4093                                                const struct nlattr *nla,
4094                                                u8 genmask)
4095 {
4096         struct nft_set *set;
4097
4098         list_for_each_entry(set, &table->sets, list) {
4099                 if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
4100                     nft_active_genmask(set, genmask))
4101                         return set;
4102         }
4103         return ERR_PTR(-ENOENT);
4104 }
4105
4106 static struct nft_set *nft_set_lookup_byid(const struct net *net,
4107                                            const struct nft_table *table,
4108                                            const struct nlattr *nla, u8 genmask)
4109 {
4110         struct nftables_pernet *nft_net = nft_pernet(net);
4111         u32 id = ntohl(nla_get_be32(nla));
4112         struct nft_trans *trans;
4113
4114         list_for_each_entry(trans, &nft_net->commit_list, list) {
4115                 if (trans->msg_type == NFT_MSG_NEWSET) {
4116                         struct nft_set *set = nft_trans_set(trans);
4117
4118                         if (id == nft_trans_set_id(trans) &&
4119                             set->table == table &&
4120                             nft_active_genmask(set, genmask))
4121                                 return set;
4122                 }
4123         }
4124         return ERR_PTR(-ENOENT);
4125 }
4126
4127 struct nft_set *nft_set_lookup_global(const struct net *net,
4128                                       const struct nft_table *table,
4129                                       const struct nlattr *nla_set_name,
4130                                       const struct nlattr *nla_set_id,
4131                                       u8 genmask)
4132 {
4133         struct nft_set *set;
4134
4135         set = nft_set_lookup(table, nla_set_name, genmask);
4136         if (IS_ERR(set)) {
4137                 if (!nla_set_id)
4138                         return set;
4139
4140                 set = nft_set_lookup_byid(net, table, nla_set_id, genmask);
4141         }
4142         return set;
4143 }
4144 EXPORT_SYMBOL_GPL(nft_set_lookup_global);
4145
4146 static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
4147                                     const char *name)
4148 {
4149         const struct nft_set *i;
4150         const char *p;
4151         unsigned long *inuse;
4152         unsigned int n = 0, min = 0;
4153
4154         p = strchr(name, '%');
4155         if (p != NULL) {
4156                 if (p[1] != 'd' || strchr(p + 2, '%'))
4157                         return -EINVAL;
4158
4159                 if (strnlen(name, NFT_SET_MAX_ANONLEN) >= NFT_SET_MAX_ANONLEN)
4160                         return -EINVAL;
4161
4162                 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
4163                 if (inuse == NULL)
4164                         return -ENOMEM;
4165 cont:
4166                 list_for_each_entry(i, &ctx->table->sets, list) {
4167                         int tmp;
4168
4169                         if (!nft_is_active_next(ctx->net, i))
4170                                 continue;
4171                         if (!sscanf(i->name, name, &tmp))
4172                                 continue;
4173                         if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
4174                                 continue;
4175
4176                         set_bit(tmp - min, inuse);
4177                 }
4178
4179                 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
4180                 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
4181                         min += BITS_PER_BYTE * PAGE_SIZE;
4182                         memset(inuse, 0, PAGE_SIZE);
4183                         goto cont;
4184                 }
4185                 free_page((unsigned long)inuse);
4186         }
4187
4188         set->name = kasprintf(GFP_KERNEL_ACCOUNT, name, min + n);
4189         if (!set->name)
4190                 return -ENOMEM;
4191
4192         list_for_each_entry(i, &ctx->table->sets, list) {
4193                 if (!nft_is_active_next(ctx->net, i))
4194                         continue;
4195                 if (!strcmp(set->name, i->name)) {
4196                         kfree(set->name);
4197                         set->name = NULL;
4198                         return -ENFILE;
4199                 }
4200         }
4201         return 0;
4202 }
4203
4204 int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
4205 {
4206         u64 ms = be64_to_cpu(nla_get_be64(nla));
4207         u64 max = (u64)(~((u64)0));
4208
4209         max = div_u64(max, NSEC_PER_MSEC);
4210         if (ms >= max)
4211                 return -ERANGE;
4212
4213         ms *= NSEC_PER_MSEC;
4214         *result = nsecs_to_jiffies64(ms);
4215         return 0;
4216 }
4217
4218 __be64 nf_jiffies64_to_msecs(u64 input)
4219 {
4220         return cpu_to_be64(jiffies64_to_msecs(input));
4221 }
4222
4223 static int nf_tables_fill_set_concat(struct sk_buff *skb,
4224                                      const struct nft_set *set)
4225 {
4226         struct nlattr *concat, *field;
4227         int i;
4228
4229         concat = nla_nest_start_noflag(skb, NFTA_SET_DESC_CONCAT);
4230         if (!concat)
4231                 return -ENOMEM;
4232
4233         for (i = 0; i < set->field_count; i++) {
4234                 field = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4235                 if (!field)
4236                         return -ENOMEM;
4237
4238                 if (nla_put_be32(skb, NFTA_SET_FIELD_LEN,
4239                                  htonl(set->field_len[i])))
4240                         return -ENOMEM;
4241
4242                 nla_nest_end(skb, field);
4243         }
4244
4245         nla_nest_end(skb, concat);
4246
4247         return 0;
4248 }
4249
4250 static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
4251                               const struct nft_set *set, u16 event, u16 flags)
4252 {
4253         u64 timeout = READ_ONCE(set->timeout);
4254         u32 gc_int = READ_ONCE(set->gc_int);
4255         u32 portid = ctx->portid;
4256         struct nlmsghdr *nlh;
4257         struct nlattr *nest;
4258         u32 seq = ctx->seq;
4259         int i;
4260
4261         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
4262         nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
4263                            NFNETLINK_V0, nft_base_seq(ctx->net));
4264         if (!nlh)
4265                 goto nla_put_failure;
4266
4267         if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
4268                 goto nla_put_failure;
4269         if (nla_put_string(skb, NFTA_SET_NAME, set->name))
4270                 goto nla_put_failure;
4271         if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
4272                          NFTA_SET_PAD))
4273                 goto nla_put_failure;
4274         if (set->flags != 0)
4275                 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
4276                         goto nla_put_failure;
4277
4278         if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
4279                 goto nla_put_failure;
4280         if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
4281                 goto nla_put_failure;
4282         if (set->flags & NFT_SET_MAP) {
4283                 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
4284                         goto nla_put_failure;
4285                 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
4286                         goto nla_put_failure;
4287         }
4288         if (set->flags & NFT_SET_OBJECT &&
4289             nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
4290                 goto nla_put_failure;
4291
4292         if (timeout &&
4293             nla_put_be64(skb, NFTA_SET_TIMEOUT,
4294                          nf_jiffies64_to_msecs(timeout),
4295                          NFTA_SET_PAD))
4296                 goto nla_put_failure;
4297         if (gc_int &&
4298             nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(gc_int)))
4299                 goto nla_put_failure;
4300
4301         if (set->policy != NFT_SET_POL_PERFORMANCE) {
4302                 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
4303                         goto nla_put_failure;
4304         }
4305
4306         if (set->udata &&
4307             nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
4308                 goto nla_put_failure;
4309
4310         nest = nla_nest_start_noflag(skb, NFTA_SET_DESC);
4311         if (!nest)
4312                 goto nla_put_failure;
4313         if (set->size &&
4314             nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
4315                 goto nla_put_failure;
4316
4317         if (set->field_count > 1 &&
4318             nf_tables_fill_set_concat(skb, set))
4319                 goto nla_put_failure;
4320
4321         nla_nest_end(skb, nest);
4322
4323         if (set->num_exprs == 1) {
4324                 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPR);
4325                 if (nf_tables_fill_expr_info(skb, set->exprs[0]) < 0)
4326                         goto nla_put_failure;
4327
4328                 nla_nest_end(skb, nest);
4329         } else if (set->num_exprs > 1) {
4330                 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPRESSIONS);
4331                 if (nest == NULL)
4332                         goto nla_put_failure;
4333
4334                 for (i = 0; i < set->num_exprs; i++) {
4335                         if (nft_expr_dump(skb, NFTA_LIST_ELEM,
4336                                           set->exprs[i]) < 0)
4337                                 goto nla_put_failure;
4338                 }
4339                 nla_nest_end(skb, nest);
4340         }
4341
4342         nlmsg_end(skb, nlh);
4343         return 0;
4344
4345 nla_put_failure:
4346         nlmsg_trim(skb, nlh);
4347         return -1;
4348 }
4349
4350 static void nf_tables_set_notify(const struct nft_ctx *ctx,
4351                                  const struct nft_set *set, int event,
4352                                  gfp_t gfp_flags)
4353 {
4354         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
4355         u32 portid = ctx->portid;
4356         struct sk_buff *skb;
4357         u16 flags = 0;
4358         int err;
4359
4360         if (!ctx->report &&
4361             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
4362                 return;
4363
4364         skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
4365         if (skb == NULL)
4366                 goto err;
4367
4368         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
4369                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
4370
4371         err = nf_tables_fill_set(skb, ctx, set, event, flags);
4372         if (err < 0) {
4373                 kfree_skb(skb);
4374                 goto err;
4375         }
4376
4377         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
4378         return;
4379 err:
4380         nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
4381 }
4382
4383 static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
4384 {
4385         const struct nft_set *set;
4386         unsigned int idx, s_idx = cb->args[0];
4387         struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
4388         struct net *net = sock_net(skb->sk);
4389         struct nft_ctx *ctx = cb->data, ctx_set;
4390         struct nftables_pernet *nft_net;
4391
4392         if (cb->args[1])
4393                 return skb->len;
4394
4395         rcu_read_lock();
4396         nft_net = nft_pernet(net);
4397         cb->seq = READ_ONCE(nft_net->base_seq);
4398
4399         list_for_each_entry_rcu(table, &nft_net->tables, list) {
4400                 if (ctx->family != NFPROTO_UNSPEC &&
4401                     ctx->family != table->family)
4402                         continue;
4403
4404                 if (ctx->table && ctx->table != table)
4405                         continue;
4406
4407                 if (cur_table) {
4408                         if (cur_table != table)
4409                                 continue;
4410
4411                         cur_table = NULL;
4412                 }
4413                 idx = 0;
4414                 list_for_each_entry_rcu(set, &table->sets, list) {
4415                         if (idx < s_idx)
4416                                 goto cont;
4417                         if (!nft_is_active(net, set))
4418                                 goto cont;
4419
4420                         ctx_set = *ctx;
4421                         ctx_set.table = table;
4422                         ctx_set.family = table->family;
4423
4424                         if (nf_tables_fill_set(skb, &ctx_set, set,
4425                                                NFT_MSG_NEWSET,
4426                                                NLM_F_MULTI) < 0) {
4427                                 cb->args[0] = idx;
4428                                 cb->args[2] = (unsigned long) table;
4429                                 goto done;
4430                         }
4431                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
4432 cont:
4433                         idx++;
4434                 }
4435                 if (s_idx)
4436                         s_idx = 0;
4437         }
4438         cb->args[1] = 1;
4439 done:
4440         rcu_read_unlock();
4441         return skb->len;
4442 }
4443
4444 static int nf_tables_dump_sets_start(struct netlink_callback *cb)
4445 {
4446         struct nft_ctx *ctx_dump = NULL;
4447
4448         ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
4449         if (ctx_dump == NULL)
4450                 return -ENOMEM;
4451
4452         cb->data = ctx_dump;
4453         return 0;
4454 }
4455
4456 static int nf_tables_dump_sets_done(struct netlink_callback *cb)
4457 {
4458         kfree(cb->data);
4459         return 0;
4460 }
4461
4462 /* called with rcu_read_lock held */
4463 static int nf_tables_getset(struct sk_buff *skb, const struct nfnl_info *info,
4464                             const struct nlattr * const nla[])
4465 {
4466         struct netlink_ext_ack *extack = info->extack;
4467         u8 genmask = nft_genmask_cur(info->net);
4468         u8 family = info->nfmsg->nfgen_family;
4469         struct nft_table *table = NULL;
4470         struct net *net = info->net;
4471         const struct nft_set *set;
4472         struct sk_buff *skb2;
4473         struct nft_ctx ctx;
4474         int err;
4475
4476         if (nla[NFTA_SET_TABLE]) {
4477                 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4478                                          genmask, 0);
4479                 if (IS_ERR(table)) {
4480                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4481                         return PTR_ERR(table);
4482                 }
4483         }
4484
4485         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4486
4487         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
4488                 struct netlink_dump_control c = {
4489                         .start = nf_tables_dump_sets_start,
4490                         .dump = nf_tables_dump_sets,
4491                         .done = nf_tables_dump_sets_done,
4492                         .data = &ctx,
4493                         .module = THIS_MODULE,
4494                 };
4495
4496                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
4497         }
4498
4499         /* Only accept unspec with dump */
4500         if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4501                 return -EAFNOSUPPORT;
4502         if (!nla[NFTA_SET_TABLE])
4503                 return -EINVAL;
4504
4505         set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4506         if (IS_ERR(set))
4507                 return PTR_ERR(set);
4508
4509         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
4510         if (skb2 == NULL)
4511                 return -ENOMEM;
4512
4513         err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
4514         if (err < 0)
4515                 goto err_fill_set_info;
4516
4517         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
4518
4519 err_fill_set_info:
4520         kfree_skb(skb2);
4521         return err;
4522 }
4523
4524 static const struct nla_policy nft_concat_policy[NFTA_SET_FIELD_MAX + 1] = {
4525         [NFTA_SET_FIELD_LEN]    = { .type = NLA_U32 },
4526 };
4527
4528 static int nft_set_desc_concat_parse(const struct nlattr *attr,
4529                                      struct nft_set_desc *desc)
4530 {
4531         struct nlattr *tb[NFTA_SET_FIELD_MAX + 1];
4532         u32 len;
4533         int err;
4534
4535         if (desc->field_count >= ARRAY_SIZE(desc->field_len))
4536                 return -E2BIG;
4537
4538         err = nla_parse_nested_deprecated(tb, NFTA_SET_FIELD_MAX, attr,
4539                                           nft_concat_policy, NULL);
4540         if (err < 0)
4541                 return err;
4542
4543         if (!tb[NFTA_SET_FIELD_LEN])
4544                 return -EINVAL;
4545
4546         len = ntohl(nla_get_be32(tb[NFTA_SET_FIELD_LEN]));
4547         if (!len || len > U8_MAX)
4548                 return -EINVAL;
4549
4550         desc->field_len[desc->field_count++] = len;
4551
4552         return 0;
4553 }
4554
4555 static int nft_set_desc_concat(struct nft_set_desc *desc,
4556                                const struct nlattr *nla)
4557 {
4558         u32 num_regs = 0, key_num_regs = 0;
4559         struct nlattr *attr;
4560         int rem, err, i;
4561
4562         nla_for_each_nested(attr, nla, rem) {
4563                 if (nla_type(attr) != NFTA_LIST_ELEM)
4564                         return -EINVAL;
4565
4566                 err = nft_set_desc_concat_parse(attr, desc);
4567                 if (err < 0)
4568                         return err;
4569         }
4570
4571         for (i = 0; i < desc->field_count; i++)
4572                 num_regs += DIV_ROUND_UP(desc->field_len[i], sizeof(u32));
4573
4574         key_num_regs = DIV_ROUND_UP(desc->klen, sizeof(u32));
4575         if (key_num_regs != num_regs)
4576                 return -EINVAL;
4577
4578         if (num_regs > NFT_REG32_COUNT)
4579                 return -E2BIG;
4580
4581         return 0;
4582 }
4583
4584 static int nf_tables_set_desc_parse(struct nft_set_desc *desc,
4585                                     const struct nlattr *nla)
4586 {
4587         struct nlattr *da[NFTA_SET_DESC_MAX + 1];
4588         int err;
4589
4590         err = nla_parse_nested_deprecated(da, NFTA_SET_DESC_MAX, nla,
4591                                           nft_set_desc_policy, NULL);
4592         if (err < 0)
4593                 return err;
4594
4595         if (da[NFTA_SET_DESC_SIZE] != NULL)
4596                 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
4597         if (da[NFTA_SET_DESC_CONCAT])
4598                 err = nft_set_desc_concat(desc, da[NFTA_SET_DESC_CONCAT]);
4599
4600         return err;
4601 }
4602
4603 static int nft_set_expr_alloc(struct nft_ctx *ctx, struct nft_set *set,
4604                               const struct nlattr * const *nla,
4605                               struct nft_expr **exprs, int *num_exprs,
4606                               u32 flags)
4607 {
4608         struct nft_expr *expr;
4609         int err, i;
4610
4611         if (nla[NFTA_SET_EXPR]) {
4612                 expr = nft_set_elem_expr_alloc(ctx, set, nla[NFTA_SET_EXPR]);
4613                 if (IS_ERR(expr)) {
4614                         err = PTR_ERR(expr);
4615                         goto err_set_expr_alloc;
4616                 }
4617                 exprs[0] = expr;
4618                 (*num_exprs)++;
4619         } else if (nla[NFTA_SET_EXPRESSIONS]) {
4620                 struct nlattr *tmp;
4621                 int left;
4622
4623                 if (!(flags & NFT_SET_EXPR)) {
4624                         err = -EINVAL;
4625                         goto err_set_expr_alloc;
4626                 }
4627                 i = 0;
4628                 nla_for_each_nested(tmp, nla[NFTA_SET_EXPRESSIONS], left) {
4629                         if (i == NFT_SET_EXPR_MAX) {
4630                                 err = -E2BIG;
4631                                 goto err_set_expr_alloc;
4632                         }
4633                         if (nla_type(tmp) != NFTA_LIST_ELEM) {
4634                                 err = -EINVAL;
4635                                 goto err_set_expr_alloc;
4636                         }
4637                         expr = nft_set_elem_expr_alloc(ctx, set, tmp);
4638                         if (IS_ERR(expr)) {
4639                                 err = PTR_ERR(expr);
4640                                 goto err_set_expr_alloc;
4641                         }
4642                         exprs[i++] = expr;
4643                         (*num_exprs)++;
4644                 }
4645         }
4646
4647         return 0;
4648
4649 err_set_expr_alloc:
4650         for (i = 0; i < *num_exprs; i++)
4651                 nft_expr_destroy(ctx, exprs[i]);
4652
4653         return err;
4654 }
4655
4656 static bool nft_set_is_same(const struct nft_set *set,
4657                             const struct nft_set_desc *desc,
4658                             struct nft_expr *exprs[], u32 num_exprs, u32 flags)
4659 {
4660         int i;
4661
4662         if (set->ktype != desc->ktype ||
4663             set->dtype != desc->dtype ||
4664             set->flags != flags ||
4665             set->klen != desc->klen ||
4666             set->dlen != desc->dlen ||
4667             set->field_count != desc->field_count ||
4668             set->num_exprs != num_exprs)
4669                 return false;
4670
4671         for (i = 0; i < desc->field_count; i++) {
4672                 if (set->field_len[i] != desc->field_len[i])
4673                         return false;
4674         }
4675
4676         for (i = 0; i < num_exprs; i++) {
4677                 if (set->exprs[i]->ops != exprs[i]->ops)
4678                         return false;
4679         }
4680
4681         return true;
4682 }
4683
4684 static int nf_tables_newset(struct sk_buff *skb, const struct nfnl_info *info,
4685                             const struct nlattr * const nla[])
4686 {
4687         struct netlink_ext_ack *extack = info->extack;
4688         u8 genmask = nft_genmask_next(info->net);
4689         u8 family = info->nfmsg->nfgen_family;
4690         const struct nft_set_ops *ops;
4691         struct net *net = info->net;
4692         struct nft_set_desc desc;
4693         struct nft_table *table;
4694         unsigned char *udata;
4695         struct nft_set *set;
4696         struct nft_ctx ctx;
4697         size_t alloc_size;
4698         int num_exprs = 0;
4699         char *name;
4700         int err, i;
4701         u16 udlen;
4702         u32 flags;
4703         u64 size;
4704
4705         if (nla[NFTA_SET_TABLE] == NULL ||
4706             nla[NFTA_SET_NAME] == NULL ||
4707             nla[NFTA_SET_KEY_LEN] == NULL ||
4708             nla[NFTA_SET_ID] == NULL)
4709                 return -EINVAL;
4710
4711         memset(&desc, 0, sizeof(desc));
4712
4713         desc.ktype = NFT_DATA_VALUE;
4714         if (nla[NFTA_SET_KEY_TYPE] != NULL) {
4715                 desc.ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
4716                 if ((desc.ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
4717                         return -EINVAL;
4718         }
4719
4720         desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
4721         if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
4722                 return -EINVAL;
4723
4724         flags = 0;
4725         if (nla[NFTA_SET_FLAGS] != NULL) {
4726                 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
4727                 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
4728                               NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
4729                               NFT_SET_MAP | NFT_SET_EVAL |
4730                               NFT_SET_OBJECT | NFT_SET_CONCAT | NFT_SET_EXPR))
4731                         return -EOPNOTSUPP;
4732                 /* Only one of these operations is supported */
4733                 if ((flags & (NFT_SET_MAP | NFT_SET_OBJECT)) ==
4734                              (NFT_SET_MAP | NFT_SET_OBJECT))
4735                         return -EOPNOTSUPP;
4736                 if ((flags & (NFT_SET_EVAL | NFT_SET_OBJECT)) ==
4737                              (NFT_SET_EVAL | NFT_SET_OBJECT))
4738                         return -EOPNOTSUPP;
4739                 if ((flags & (NFT_SET_ANONYMOUS | NFT_SET_TIMEOUT | NFT_SET_EVAL)) ==
4740                              (NFT_SET_ANONYMOUS | NFT_SET_TIMEOUT))
4741                         return -EOPNOTSUPP;
4742                 if ((flags & (NFT_SET_CONSTANT | NFT_SET_TIMEOUT)) ==
4743                              (NFT_SET_CONSTANT | NFT_SET_TIMEOUT))
4744                         return -EOPNOTSUPP;
4745         }
4746
4747         desc.dtype = 0;
4748         if (nla[NFTA_SET_DATA_TYPE] != NULL) {
4749                 if (!(flags & NFT_SET_MAP))
4750                         return -EINVAL;
4751
4752                 desc.dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
4753                 if ((desc.dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
4754                     desc.dtype != NFT_DATA_VERDICT)
4755                         return -EINVAL;
4756
4757                 if (desc.dtype != NFT_DATA_VERDICT) {
4758                         if (nla[NFTA_SET_DATA_LEN] == NULL)
4759                                 return -EINVAL;
4760                         desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
4761                         if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
4762                                 return -EINVAL;
4763                 } else
4764                         desc.dlen = sizeof(struct nft_verdict);
4765         } else if (flags & NFT_SET_MAP)
4766                 return -EINVAL;
4767
4768         if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
4769                 if (!(flags & NFT_SET_OBJECT))
4770                         return -EINVAL;
4771
4772                 desc.objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
4773                 if (desc.objtype == NFT_OBJECT_UNSPEC ||
4774                     desc.objtype > NFT_OBJECT_MAX)
4775                         return -EOPNOTSUPP;
4776         } else if (flags & NFT_SET_OBJECT)
4777                 return -EINVAL;
4778         else
4779                 desc.objtype = NFT_OBJECT_UNSPEC;
4780
4781         desc.timeout = 0;
4782         if (nla[NFTA_SET_TIMEOUT] != NULL) {
4783                 if (!(flags & NFT_SET_TIMEOUT))
4784                         return -EINVAL;
4785
4786                 if (flags & NFT_SET_ANONYMOUS)
4787                         return -EOPNOTSUPP;
4788
4789                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &desc.timeout);
4790                 if (err)
4791                         return err;
4792         }
4793         desc.gc_int = 0;
4794         if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
4795                 if (!(flags & NFT_SET_TIMEOUT))
4796                         return -EINVAL;
4797
4798                 if (flags & NFT_SET_ANONYMOUS)
4799                         return -EOPNOTSUPP;
4800
4801                 desc.gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
4802         }
4803
4804         desc.policy = NFT_SET_POL_PERFORMANCE;
4805         if (nla[NFTA_SET_POLICY] != NULL) {
4806                 desc.policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
4807                 switch (desc.policy) {
4808                 case NFT_SET_POL_PERFORMANCE:
4809                 case NFT_SET_POL_MEMORY:
4810                         break;
4811                 default:
4812                         return -EOPNOTSUPP;
4813                 }
4814         }
4815
4816         if (nla[NFTA_SET_DESC] != NULL) {
4817                 err = nf_tables_set_desc_parse(&desc, nla[NFTA_SET_DESC]);
4818                 if (err < 0)
4819                         return err;
4820
4821                 if (desc.field_count > 1) {
4822                         if (!(flags & NFT_SET_CONCAT))
4823                                 return -EINVAL;
4824                 } else if (flags & NFT_SET_CONCAT) {
4825                         return -EINVAL;
4826                 }
4827         } else if (flags & NFT_SET_CONCAT) {
4828                 return -EINVAL;
4829         }
4830
4831         if (nla[NFTA_SET_EXPR] || nla[NFTA_SET_EXPRESSIONS])
4832                 desc.expr = true;
4833
4834         table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask,
4835                                  NETLINK_CB(skb).portid);
4836         if (IS_ERR(table)) {
4837                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4838                 return PTR_ERR(table);
4839         }
4840
4841         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4842
4843         set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4844         if (IS_ERR(set)) {
4845                 if (PTR_ERR(set) != -ENOENT) {
4846                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4847                         return PTR_ERR(set);
4848                 }
4849         } else {
4850                 struct nft_expr *exprs[NFT_SET_EXPR_MAX] = {};
4851
4852                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
4853                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4854                         return -EEXIST;
4855                 }
4856                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
4857                         return -EOPNOTSUPP;
4858
4859                 if (nft_set_is_anonymous(set))
4860                         return -EOPNOTSUPP;
4861
4862                 err = nft_set_expr_alloc(&ctx, set, nla, exprs, &num_exprs, flags);
4863                 if (err < 0)
4864                         return err;
4865
4866                 err = 0;
4867                 if (!nft_set_is_same(set, &desc, exprs, num_exprs, flags)) {
4868                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4869                         err = -EEXIST;
4870                 }
4871
4872                 for (i = 0; i < num_exprs; i++)
4873                         nft_expr_destroy(&ctx, exprs[i]);
4874
4875                 if (err < 0)
4876                         return err;
4877
4878                 return __nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set, &desc);
4879         }
4880
4881         if (!(info->nlh->nlmsg_flags & NLM_F_CREATE))
4882                 return -ENOENT;
4883
4884         ops = nft_select_set_ops(&ctx, nla, &desc);
4885         if (IS_ERR(ops))
4886                 return PTR_ERR(ops);
4887
4888         udlen = 0;
4889         if (nla[NFTA_SET_USERDATA])
4890                 udlen = nla_len(nla[NFTA_SET_USERDATA]);
4891
4892         size = 0;
4893         if (ops->privsize != NULL)
4894                 size = ops->privsize(nla, &desc);
4895         alloc_size = sizeof(*set) + size + udlen;
4896         if (alloc_size < size || alloc_size > INT_MAX)
4897                 return -ENOMEM;
4898
4899         if (!nft_use_inc(&table->use))
4900                 return -EMFILE;
4901
4902         set = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT);
4903         if (!set) {
4904                 err = -ENOMEM;
4905                 goto err_alloc;
4906         }
4907
4908         name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL_ACCOUNT);
4909         if (!name) {
4910                 err = -ENOMEM;
4911                 goto err_set_name;
4912         }
4913
4914         err = nf_tables_set_alloc_name(&ctx, set, name);
4915         kfree(name);
4916         if (err < 0)
4917                 goto err_set_name;
4918
4919         udata = NULL;
4920         if (udlen) {
4921                 udata = set->data + size;
4922                 nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
4923         }
4924
4925         INIT_LIST_HEAD(&set->bindings);
4926         INIT_LIST_HEAD(&set->catchall_list);
4927         refcount_set(&set->refs, 1);
4928         set->table = table;
4929         write_pnet(&set->net, net);
4930         set->ops = ops;
4931         set->ktype = desc.ktype;
4932         set->klen = desc.klen;
4933         set->dtype = desc.dtype;
4934         set->objtype = desc.objtype;
4935         set->dlen = desc.dlen;
4936         set->flags = flags;
4937         set->size = desc.size;
4938         set->policy = desc.policy;
4939         set->udlen = udlen;
4940         set->udata = udata;
4941         set->timeout = desc.timeout;
4942         set->gc_int = desc.gc_int;
4943
4944         set->field_count = desc.field_count;
4945         for (i = 0; i < desc.field_count; i++)
4946                 set->field_len[i] = desc.field_len[i];
4947
4948         err = ops->init(set, &desc, nla);
4949         if (err < 0)
4950                 goto err_set_init;
4951
4952         err = nft_set_expr_alloc(&ctx, set, nla, set->exprs, &num_exprs, flags);
4953         if (err < 0)
4954                 goto err_set_destroy;
4955
4956         set->num_exprs = num_exprs;
4957         set->handle = nf_tables_alloc_handle(table);
4958         INIT_LIST_HEAD(&set->pending_update);
4959
4960         err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
4961         if (err < 0)
4962                 goto err_set_expr_alloc;
4963
4964         list_add_tail_rcu(&set->list, &table->sets);
4965
4966         return 0;
4967
4968 err_set_expr_alloc:
4969         for (i = 0; i < set->num_exprs; i++)
4970                 nft_expr_destroy(&ctx, set->exprs[i]);
4971 err_set_destroy:
4972         ops->destroy(&ctx, set);
4973 err_set_init:
4974         kfree(set->name);
4975 err_set_name:
4976         kvfree(set);
4977 err_alloc:
4978         nft_use_dec_restore(&table->use);
4979
4980         return err;
4981 }
4982
4983 static void nft_set_catchall_destroy(const struct nft_ctx *ctx,
4984                                      struct nft_set *set)
4985 {
4986         struct nft_set_elem_catchall *next, *catchall;
4987
4988         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
4989                 list_del_rcu(&catchall->list);
4990                 nf_tables_set_elem_destroy(ctx, set, catchall->elem);
4991                 kfree_rcu(catchall, rcu);
4992         }
4993 }
4994
4995 static void nft_set_put(struct nft_set *set)
4996 {
4997         if (refcount_dec_and_test(&set->refs)) {
4998                 kfree(set->name);
4999                 kvfree(set);
5000         }
5001 }
5002
5003 static void nft_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
5004 {
5005         int i;
5006
5007         if (WARN_ON(set->use > 0))
5008                 return;
5009
5010         for (i = 0; i < set->num_exprs; i++)
5011                 nft_expr_destroy(ctx, set->exprs[i]);
5012
5013         set->ops->destroy(ctx, set);
5014         nft_set_catchall_destroy(ctx, set);
5015         nft_set_put(set);
5016 }
5017
5018 static int nf_tables_delset(struct sk_buff *skb, const struct nfnl_info *info,
5019                             const struct nlattr * const nla[])
5020 {
5021         struct netlink_ext_ack *extack = info->extack;
5022         u8 genmask = nft_genmask_next(info->net);
5023         u8 family = info->nfmsg->nfgen_family;
5024         struct net *net = info->net;
5025         const struct nlattr *attr;
5026         struct nft_table *table;
5027         struct nft_set *set;
5028         struct nft_ctx ctx;
5029
5030         if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
5031                 return -EAFNOSUPPORT;
5032
5033         table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
5034                                  genmask, NETLINK_CB(skb).portid);
5035         if (IS_ERR(table)) {
5036                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
5037                 return PTR_ERR(table);
5038         }
5039
5040         if (nla[NFTA_SET_HANDLE]) {
5041                 attr = nla[NFTA_SET_HANDLE];
5042                 set = nft_set_lookup_byhandle(table, attr, genmask);
5043         } else {
5044                 attr = nla[NFTA_SET_NAME];
5045                 set = nft_set_lookup(table, attr, genmask);
5046         }
5047
5048         if (IS_ERR(set)) {
5049                 NL_SET_BAD_ATTR(extack, attr);
5050                 return PTR_ERR(set);
5051         }
5052         if (set->use ||
5053             (info->nlh->nlmsg_flags & NLM_F_NONREC &&
5054              atomic_read(&set->nelems) > 0)) {
5055                 NL_SET_BAD_ATTR(extack, attr);
5056                 return -EBUSY;
5057         }
5058
5059         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5060
5061         return nft_delset(&ctx, set);
5062 }
5063
5064 static int nft_validate_register_store(const struct nft_ctx *ctx,
5065                                        enum nft_registers reg,
5066                                        const struct nft_data *data,
5067                                        enum nft_data_types type,
5068                                        unsigned int len);
5069
5070 static int nft_setelem_data_validate(const struct nft_ctx *ctx,
5071                                      struct nft_set *set,
5072                                      struct nft_set_elem *elem)
5073 {
5074         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5075         enum nft_registers dreg;
5076
5077         dreg = nft_type_to_reg(set->dtype);
5078         return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
5079                                            set->dtype == NFT_DATA_VERDICT ?
5080                                            NFT_DATA_VERDICT : NFT_DATA_VALUE,
5081                                            set->dlen);
5082 }
5083
5084 static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
5085                                         struct nft_set *set,
5086                                         const struct nft_set_iter *iter,
5087                                         struct nft_set_elem *elem)
5088 {
5089         return nft_setelem_data_validate(ctx, set, elem);
5090 }
5091
5092 static int nft_set_catchall_bind_check(const struct nft_ctx *ctx,
5093                                        struct nft_set *set)
5094 {
5095         u8 genmask = nft_genmask_next(ctx->net);
5096         struct nft_set_elem_catchall *catchall;
5097         struct nft_set_elem elem;
5098         struct nft_set_ext *ext;
5099         int ret = 0;
5100
5101         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5102                 ext = nft_set_elem_ext(set, catchall->elem);
5103                 if (!nft_set_elem_active(ext, genmask))
5104                         continue;
5105
5106                 elem.priv = catchall->elem;
5107                 ret = nft_setelem_data_validate(ctx, set, &elem);
5108                 if (ret < 0)
5109                         break;
5110         }
5111
5112         return ret;
5113 }
5114
5115 int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
5116                        struct nft_set_binding *binding)
5117 {
5118         struct nft_set_binding *i;
5119         struct nft_set_iter iter;
5120
5121         if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
5122                 return -EBUSY;
5123
5124         if (binding->flags & NFT_SET_MAP) {
5125                 /* If the set is already bound to the same chain all
5126                  * jumps are already validated for that chain.
5127                  */
5128                 list_for_each_entry(i, &set->bindings, list) {
5129                         if (i->flags & NFT_SET_MAP &&
5130                             i->chain == binding->chain)
5131                                 goto bind;
5132                 }
5133
5134                 iter.genmask    = nft_genmask_next(ctx->net);
5135                 iter.skip       = 0;
5136                 iter.count      = 0;
5137                 iter.err        = 0;
5138                 iter.fn         = nf_tables_bind_check_setelem;
5139
5140                 set->ops->walk(ctx, set, &iter);
5141                 if (!iter.err)
5142                         iter.err = nft_set_catchall_bind_check(ctx, set);
5143
5144                 if (iter.err < 0)
5145                         return iter.err;
5146         }
5147 bind:
5148         if (!nft_use_inc(&set->use))
5149                 return -EMFILE;
5150
5151         binding->chain = ctx->chain;
5152         list_add_tail_rcu(&binding->list, &set->bindings);
5153         nft_set_trans_bind(ctx, set);
5154
5155         return 0;
5156 }
5157 EXPORT_SYMBOL_GPL(nf_tables_bind_set);
5158
5159 static void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
5160                                  struct nft_set_binding *binding, bool event)
5161 {
5162         list_del_rcu(&binding->list);
5163
5164         if (list_empty(&set->bindings) && nft_set_is_anonymous(set)) {
5165                 list_del_rcu(&set->list);
5166                 set->dead = 1;
5167                 if (event)
5168                         nf_tables_set_notify(ctx, set, NFT_MSG_DELSET,
5169                                              GFP_KERNEL);
5170         }
5171 }
5172
5173 static void nft_setelem_data_activate(const struct net *net,
5174                                       const struct nft_set *set,
5175                                       struct nft_set_elem *elem);
5176
5177 static int nft_mapelem_activate(const struct nft_ctx *ctx,
5178                                 struct nft_set *set,
5179                                 const struct nft_set_iter *iter,
5180                                 struct nft_set_elem *elem)
5181 {
5182         nft_setelem_data_activate(ctx->net, set, elem);
5183
5184         return 0;
5185 }
5186
5187 static void nft_map_catchall_activate(const struct nft_ctx *ctx,
5188                                       struct nft_set *set)
5189 {
5190         u8 genmask = nft_genmask_next(ctx->net);
5191         struct nft_set_elem_catchall *catchall;
5192         struct nft_set_elem elem;
5193         struct nft_set_ext *ext;
5194
5195         list_for_each_entry(catchall, &set->catchall_list, list) {
5196                 ext = nft_set_elem_ext(set, catchall->elem);
5197                 if (!nft_set_elem_active(ext, genmask))
5198                         continue;
5199
5200                 elem.priv = catchall->elem;
5201                 nft_setelem_data_activate(ctx->net, set, &elem);
5202                 break;
5203         }
5204 }
5205
5206 static void nft_map_activate(const struct nft_ctx *ctx, struct nft_set *set)
5207 {
5208         struct nft_set_iter iter = {
5209                 .genmask        = nft_genmask_next(ctx->net),
5210                 .fn             = nft_mapelem_activate,
5211         };
5212
5213         set->ops->walk(ctx, set, &iter);
5214         WARN_ON_ONCE(iter.err);
5215
5216         nft_map_catchall_activate(ctx, set);
5217 }
5218
5219 void nf_tables_activate_set(const struct nft_ctx *ctx, struct nft_set *set)
5220 {
5221         if (nft_set_is_anonymous(set)) {
5222                 if (set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
5223                         nft_map_activate(ctx, set);
5224
5225                 nft_clear(ctx->net, set);
5226         }
5227
5228         nft_use_inc_restore(&set->use);
5229 }
5230 EXPORT_SYMBOL_GPL(nf_tables_activate_set);
5231
5232 void nf_tables_deactivate_set(const struct nft_ctx *ctx, struct nft_set *set,
5233                               struct nft_set_binding *binding,
5234                               enum nft_trans_phase phase)
5235 {
5236         switch (phase) {
5237         case NFT_TRANS_PREPARE_ERROR:
5238                 nft_set_trans_unbind(ctx, set);
5239                 if (nft_set_is_anonymous(set))
5240                         nft_deactivate_next(ctx->net, set);
5241                 else
5242                         list_del_rcu(&binding->list);
5243
5244                 nft_use_dec(&set->use);
5245                 break;
5246         case NFT_TRANS_PREPARE:
5247                 if (nft_set_is_anonymous(set)) {
5248                         if (set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
5249                                 nft_map_deactivate(ctx, set);
5250
5251                         nft_deactivate_next(ctx->net, set);
5252                 }
5253                 nft_use_dec(&set->use);
5254                 return;
5255         case NFT_TRANS_ABORT:
5256         case NFT_TRANS_RELEASE:
5257                 if (nft_set_is_anonymous(set) &&
5258                     set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
5259                         nft_map_deactivate(ctx, set);
5260
5261                 nft_use_dec(&set->use);
5262                 fallthrough;
5263         default:
5264                 nf_tables_unbind_set(ctx, set, binding,
5265                                      phase == NFT_TRANS_COMMIT);
5266         }
5267 }
5268 EXPORT_SYMBOL_GPL(nf_tables_deactivate_set);
5269
5270 void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
5271 {
5272         if (list_empty(&set->bindings) && nft_set_is_anonymous(set))
5273                 nft_set_destroy(ctx, set);
5274 }
5275 EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
5276
5277 const struct nft_set_ext_type nft_set_ext_types[] = {
5278         [NFT_SET_EXT_KEY]               = {
5279                 .align  = __alignof__(u32),
5280         },
5281         [NFT_SET_EXT_DATA]              = {
5282                 .align  = __alignof__(u32),
5283         },
5284         [NFT_SET_EXT_EXPRESSIONS]       = {
5285                 .align  = __alignof__(struct nft_set_elem_expr),
5286         },
5287         [NFT_SET_EXT_OBJREF]            = {
5288                 .len    = sizeof(struct nft_object *),
5289                 .align  = __alignof__(struct nft_object *),
5290         },
5291         [NFT_SET_EXT_FLAGS]             = {
5292                 .len    = sizeof(u8),
5293                 .align  = __alignof__(u8),
5294         },
5295         [NFT_SET_EXT_TIMEOUT]           = {
5296                 .len    = sizeof(u64),
5297                 .align  = __alignof__(u64),
5298         },
5299         [NFT_SET_EXT_EXPIRATION]        = {
5300                 .len    = sizeof(u64),
5301                 .align  = __alignof__(u64),
5302         },
5303         [NFT_SET_EXT_USERDATA]          = {
5304                 .len    = sizeof(struct nft_userdata),
5305                 .align  = __alignof__(struct nft_userdata),
5306         },
5307         [NFT_SET_EXT_KEY_END]           = {
5308                 .align  = __alignof__(u32),
5309         },
5310 };
5311
5312 /*
5313  * Set elements
5314  */
5315
5316 static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
5317         [NFTA_SET_ELEM_KEY]             = { .type = NLA_NESTED },
5318         [NFTA_SET_ELEM_DATA]            = { .type = NLA_NESTED },
5319         [NFTA_SET_ELEM_FLAGS]           = { .type = NLA_U32 },
5320         [NFTA_SET_ELEM_TIMEOUT]         = { .type = NLA_U64 },
5321         [NFTA_SET_ELEM_EXPIRATION]      = { .type = NLA_U64 },
5322         [NFTA_SET_ELEM_USERDATA]        = { .type = NLA_BINARY,
5323                                             .len = NFT_USERDATA_MAXLEN },
5324         [NFTA_SET_ELEM_EXPR]            = { .type = NLA_NESTED },
5325         [NFTA_SET_ELEM_OBJREF]          = { .type = NLA_STRING,
5326                                             .len = NFT_OBJ_MAXNAMELEN - 1 },
5327         [NFTA_SET_ELEM_KEY_END]         = { .type = NLA_NESTED },
5328         [NFTA_SET_ELEM_EXPRESSIONS]     = { .type = NLA_NESTED },
5329 };
5330
5331 static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
5332         [NFTA_SET_ELEM_LIST_TABLE]      = { .type = NLA_STRING,
5333                                             .len = NFT_TABLE_MAXNAMELEN - 1 },
5334         [NFTA_SET_ELEM_LIST_SET]        = { .type = NLA_STRING,
5335                                             .len = NFT_SET_MAXNAMELEN - 1 },
5336         [NFTA_SET_ELEM_LIST_ELEMENTS]   = { .type = NLA_NESTED },
5337         [NFTA_SET_ELEM_LIST_SET_ID]     = { .type = NLA_U32 },
5338 };
5339
5340 static int nft_set_elem_expr_dump(struct sk_buff *skb,
5341                                   const struct nft_set *set,
5342                                   const struct nft_set_ext *ext)
5343 {
5344         struct nft_set_elem_expr *elem_expr;
5345         u32 size, num_exprs = 0;
5346         struct nft_expr *expr;
5347         struct nlattr *nest;
5348
5349         elem_expr = nft_set_ext_expr(ext);
5350         nft_setelem_expr_foreach(expr, elem_expr, size)
5351                 num_exprs++;
5352
5353         if (num_exprs == 1) {
5354                 expr = nft_setelem_expr_at(elem_expr, 0);
5355                 if (nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, expr) < 0)
5356                         return -1;
5357
5358                 return 0;
5359         } else if (num_exprs > 1) {
5360                 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_EXPRESSIONS);
5361                 if (nest == NULL)
5362                         goto nla_put_failure;
5363
5364                 nft_setelem_expr_foreach(expr, elem_expr, size) {
5365                         expr = nft_setelem_expr_at(elem_expr, size);
5366                         if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
5367                                 goto nla_put_failure;
5368                 }
5369                 nla_nest_end(skb, nest);
5370         }
5371         return 0;
5372
5373 nla_put_failure:
5374         return -1;
5375 }
5376
5377 static int nf_tables_fill_setelem(struct sk_buff *skb,
5378                                   const struct nft_set *set,
5379                                   const struct nft_set_elem *elem)
5380 {
5381         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5382         unsigned char *b = skb_tail_pointer(skb);
5383         struct nlattr *nest;
5384
5385         nest = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
5386         if (nest == NULL)
5387                 goto nla_put_failure;
5388
5389         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
5390             nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
5391                           NFT_DATA_VALUE, set->klen) < 0)
5392                 goto nla_put_failure;
5393
5394         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
5395             nft_data_dump(skb, NFTA_SET_ELEM_KEY_END, nft_set_ext_key_end(ext),
5396                           NFT_DATA_VALUE, set->klen) < 0)
5397                 goto nla_put_failure;
5398
5399         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5400             nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
5401                           set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
5402                           set->dlen) < 0)
5403                 goto nla_put_failure;
5404
5405         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS) &&
5406             nft_set_elem_expr_dump(skb, set, ext))
5407                 goto nla_put_failure;
5408
5409         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
5410             nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
5411                            (*nft_set_ext_obj(ext))->key.name) < 0)
5412                 goto nla_put_failure;
5413
5414         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
5415             nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
5416                          htonl(*nft_set_ext_flags(ext))))
5417                 goto nla_put_failure;
5418
5419         if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
5420             nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
5421                          nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
5422                          NFTA_SET_ELEM_PAD))
5423                 goto nla_put_failure;
5424
5425         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5426                 u64 expires, now = get_jiffies_64();
5427
5428                 expires = *nft_set_ext_expiration(ext);
5429                 if (time_before64(now, expires))
5430                         expires -= now;
5431                 else
5432                         expires = 0;
5433
5434                 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
5435                                  nf_jiffies64_to_msecs(expires),
5436                                  NFTA_SET_ELEM_PAD))
5437                         goto nla_put_failure;
5438         }
5439
5440         if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
5441                 struct nft_userdata *udata;
5442
5443                 udata = nft_set_ext_userdata(ext);
5444                 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
5445                             udata->len + 1, udata->data))
5446                         goto nla_put_failure;
5447         }
5448
5449         nla_nest_end(skb, nest);
5450         return 0;
5451
5452 nla_put_failure:
5453         nlmsg_trim(skb, b);
5454         return -EMSGSIZE;
5455 }
5456
5457 struct nft_set_dump_args {
5458         const struct netlink_callback   *cb;
5459         struct nft_set_iter             iter;
5460         struct sk_buff                  *skb;
5461 };
5462
5463 static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
5464                                   struct nft_set *set,
5465                                   const struct nft_set_iter *iter,
5466                                   struct nft_set_elem *elem)
5467 {
5468         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5469         struct nft_set_dump_args *args;
5470
5471         if (nft_set_elem_expired(ext) || nft_set_elem_is_dead(ext))
5472                 return 0;
5473
5474         args = container_of(iter, struct nft_set_dump_args, iter);
5475         return nf_tables_fill_setelem(args->skb, set, elem);
5476 }
5477
5478 struct nft_set_dump_ctx {
5479         const struct nft_set    *set;
5480         struct nft_ctx          ctx;
5481 };
5482
5483 static int nft_set_catchall_dump(struct net *net, struct sk_buff *skb,
5484                                  const struct nft_set *set)
5485 {
5486         struct nft_set_elem_catchall *catchall;
5487         u8 genmask = nft_genmask_cur(net);
5488         struct nft_set_elem elem;
5489         struct nft_set_ext *ext;
5490         int ret = 0;
5491
5492         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5493                 ext = nft_set_elem_ext(set, catchall->elem);
5494                 if (!nft_set_elem_active(ext, genmask) ||
5495                     nft_set_elem_expired(ext))
5496                         continue;
5497
5498                 elem.priv = catchall->elem;
5499                 ret = nf_tables_fill_setelem(skb, set, &elem);
5500                 break;
5501         }
5502
5503         return ret;
5504 }
5505
5506 static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
5507 {
5508         struct nft_set_dump_ctx *dump_ctx = cb->data;
5509         struct net *net = sock_net(skb->sk);
5510         struct nftables_pernet *nft_net;
5511         struct nft_table *table;
5512         struct nft_set *set;
5513         struct nft_set_dump_args args;
5514         bool set_found = false;
5515         struct nlmsghdr *nlh;
5516         struct nlattr *nest;
5517         u32 portid, seq;
5518         int event;
5519
5520         rcu_read_lock();
5521         nft_net = nft_pernet(net);
5522         cb->seq = READ_ONCE(nft_net->base_seq);
5523
5524         list_for_each_entry_rcu(table, &nft_net->tables, list) {
5525                 if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
5526                     dump_ctx->ctx.family != table->family)
5527                         continue;
5528
5529                 if (table != dump_ctx->ctx.table)
5530                         continue;
5531
5532                 list_for_each_entry_rcu(set, &table->sets, list) {
5533                         if (set == dump_ctx->set) {
5534                                 set_found = true;
5535                                 break;
5536                         }
5537                 }
5538                 break;
5539         }
5540
5541         if (!set_found) {
5542                 rcu_read_unlock();
5543                 return -ENOENT;
5544         }
5545
5546         event  = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
5547         portid = NETLINK_CB(cb->skb).portid;
5548         seq    = cb->nlh->nlmsg_seq;
5549
5550         nlh = nfnl_msg_put(skb, portid, seq, event, NLM_F_MULTI,
5551                            table->family, NFNETLINK_V0, nft_base_seq(net));
5552         if (!nlh)
5553                 goto nla_put_failure;
5554
5555         if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
5556                 goto nla_put_failure;
5557         if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
5558                 goto nla_put_failure;
5559
5560         nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5561         if (nest == NULL)
5562                 goto nla_put_failure;
5563
5564         args.cb                 = cb;
5565         args.skb                = skb;
5566         args.iter.genmask       = nft_genmask_cur(net);
5567         args.iter.skip          = cb->args[0];
5568         args.iter.count         = 0;
5569         args.iter.err           = 0;
5570         args.iter.fn            = nf_tables_dump_setelem;
5571         set->ops->walk(&dump_ctx->ctx, set, &args.iter);
5572
5573         if (!args.iter.err && args.iter.count == cb->args[0])
5574                 args.iter.err = nft_set_catchall_dump(net, skb, set);
5575         rcu_read_unlock();
5576
5577         nla_nest_end(skb, nest);
5578         nlmsg_end(skb, nlh);
5579
5580         if (args.iter.err && args.iter.err != -EMSGSIZE)
5581                 return args.iter.err;
5582         if (args.iter.count == cb->args[0])
5583                 return 0;
5584
5585         cb->args[0] = args.iter.count;
5586         return skb->len;
5587
5588 nla_put_failure:
5589         rcu_read_unlock();
5590         return -ENOSPC;
5591 }
5592
5593 static int nf_tables_dump_set_start(struct netlink_callback *cb)
5594 {
5595         struct nft_set_dump_ctx *dump_ctx = cb->data;
5596
5597         cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
5598
5599         return cb->data ? 0 : -ENOMEM;
5600 }
5601
5602 static int nf_tables_dump_set_done(struct netlink_callback *cb)
5603 {
5604         kfree(cb->data);
5605         return 0;
5606 }
5607
5608 static int nf_tables_fill_setelem_info(struct sk_buff *skb,
5609                                        const struct nft_ctx *ctx, u32 seq,
5610                                        u32 portid, int event, u16 flags,
5611                                        const struct nft_set *set,
5612                                        const struct nft_set_elem *elem)
5613 {
5614         struct nlmsghdr *nlh;
5615         struct nlattr *nest;
5616         int err;
5617
5618         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
5619         nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
5620                            NFNETLINK_V0, nft_base_seq(ctx->net));
5621         if (!nlh)
5622                 goto nla_put_failure;
5623
5624         if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
5625                 goto nla_put_failure;
5626         if (nla_put_string(skb, NFTA_SET_NAME, set->name))
5627                 goto nla_put_failure;
5628
5629         nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5630         if (nest == NULL)
5631                 goto nla_put_failure;
5632
5633         err = nf_tables_fill_setelem(skb, set, elem);
5634         if (err < 0)
5635                 goto nla_put_failure;
5636
5637         nla_nest_end(skb, nest);
5638
5639         nlmsg_end(skb, nlh);
5640         return 0;
5641
5642 nla_put_failure:
5643         nlmsg_trim(skb, nlh);
5644         return -1;
5645 }
5646
5647 static int nft_setelem_parse_flags(const struct nft_set *set,
5648                                    const struct nlattr *attr, u32 *flags)
5649 {
5650         if (attr == NULL)
5651                 return 0;
5652
5653         *flags = ntohl(nla_get_be32(attr));
5654         if (*flags & ~(NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5655                 return -EOPNOTSUPP;
5656         if (!(set->flags & NFT_SET_INTERVAL) &&
5657             *flags & NFT_SET_ELEM_INTERVAL_END)
5658                 return -EINVAL;
5659         if ((*flags & (NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL)) ==
5660             (NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5661                 return -EINVAL;
5662
5663         return 0;
5664 }
5665
5666 static int nft_setelem_parse_key(struct nft_ctx *ctx, struct nft_set *set,
5667                                  struct nft_data *key, struct nlattr *attr)
5668 {
5669         struct nft_data_desc desc = {
5670                 .type   = NFT_DATA_VALUE,
5671                 .size   = NFT_DATA_VALUE_MAXLEN,
5672                 .len    = set->klen,
5673         };
5674
5675         return nft_data_init(ctx, key, &desc, attr);
5676 }
5677
5678 static int nft_setelem_parse_data(struct nft_ctx *ctx, struct nft_set *set,
5679                                   struct nft_data_desc *desc,
5680                                   struct nft_data *data,
5681                                   struct nlattr *attr)
5682 {
5683         u32 dtype;
5684
5685         if (set->dtype == NFT_DATA_VERDICT)
5686                 dtype = NFT_DATA_VERDICT;
5687         else
5688                 dtype = NFT_DATA_VALUE;
5689
5690         desc->type = dtype;
5691         desc->size = NFT_DATA_VALUE_MAXLEN;
5692         desc->len = set->dlen;
5693         desc->flags = NFT_DATA_DESC_SETELEM;
5694
5695         return nft_data_init(ctx, data, desc, attr);
5696 }
5697
5698 static void *nft_setelem_catchall_get(const struct net *net,
5699                                       const struct nft_set *set)
5700 {
5701         struct nft_set_elem_catchall *catchall;
5702         u8 genmask = nft_genmask_cur(net);
5703         struct nft_set_ext *ext;
5704         void *priv = NULL;
5705
5706         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5707                 ext = nft_set_elem_ext(set, catchall->elem);
5708                 if (!nft_set_elem_active(ext, genmask) ||
5709                     nft_set_elem_expired(ext))
5710                         continue;
5711
5712                 priv = catchall->elem;
5713                 break;
5714         }
5715
5716         return priv;
5717 }
5718
5719 static int nft_setelem_get(struct nft_ctx *ctx, struct nft_set *set,
5720                            struct nft_set_elem *elem, u32 flags)
5721 {
5722         void *priv;
5723
5724         if (!(flags & NFT_SET_ELEM_CATCHALL)) {
5725                 priv = set->ops->get(ctx->net, set, elem, flags);
5726                 if (IS_ERR(priv))
5727                         return PTR_ERR(priv);
5728         } else {
5729                 priv = nft_setelem_catchall_get(ctx->net, set);
5730                 if (!priv)
5731                         return -ENOENT;
5732         }
5733         elem->priv = priv;
5734
5735         return 0;
5736 }
5737
5738 static int nft_get_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5739                             const struct nlattr *attr)
5740 {
5741         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5742         struct nft_set_elem elem;
5743         struct sk_buff *skb;
5744         uint32_t flags = 0;
5745         int err;
5746
5747         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5748                                           nft_set_elem_policy, NULL);
5749         if (err < 0)
5750                 return err;
5751
5752         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5753         if (err < 0)
5754                 return err;
5755
5756         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5757                 return -EINVAL;
5758
5759         if (nla[NFTA_SET_ELEM_KEY]) {
5760                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5761                                             nla[NFTA_SET_ELEM_KEY]);
5762                 if (err < 0)
5763                         return err;
5764         }
5765
5766         if (nla[NFTA_SET_ELEM_KEY_END]) {
5767                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5768                                             nla[NFTA_SET_ELEM_KEY_END]);
5769                 if (err < 0)
5770                         return err;
5771         }
5772
5773         err = nft_setelem_get(ctx, set, &elem, flags);
5774         if (err < 0)
5775                 return err;
5776
5777         err = -ENOMEM;
5778         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
5779         if (skb == NULL)
5780                 return err;
5781
5782         err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
5783                                           NFT_MSG_NEWSETELEM, 0, set, &elem);
5784         if (err < 0)
5785                 goto err_fill_setelem;
5786
5787         return nfnetlink_unicast(skb, ctx->net, ctx->portid);
5788
5789 err_fill_setelem:
5790         kfree_skb(skb);
5791         return err;
5792 }
5793
5794 /* called with rcu_read_lock held */
5795 static int nf_tables_getsetelem(struct sk_buff *skb,
5796                                 const struct nfnl_info *info,
5797                                 const struct nlattr * const nla[])
5798 {
5799         struct netlink_ext_ack *extack = info->extack;
5800         u8 genmask = nft_genmask_cur(info->net);
5801         u8 family = info->nfmsg->nfgen_family;
5802         struct net *net = info->net;
5803         struct nft_table *table;
5804         struct nft_set *set;
5805         struct nlattr *attr;
5806         struct nft_ctx ctx;
5807         int rem, err = 0;
5808
5809         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
5810                                  genmask, 0);
5811         if (IS_ERR(table)) {
5812                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
5813                 return PTR_ERR(table);
5814         }
5815
5816         set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
5817         if (IS_ERR(set))
5818                 return PTR_ERR(set);
5819
5820         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5821
5822         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
5823                 struct netlink_dump_control c = {
5824                         .start = nf_tables_dump_set_start,
5825                         .dump = nf_tables_dump_set,
5826                         .done = nf_tables_dump_set_done,
5827                         .module = THIS_MODULE,
5828                 };
5829                 struct nft_set_dump_ctx dump_ctx = {
5830                         .set = set,
5831                         .ctx = ctx,
5832                 };
5833
5834                 c.data = &dump_ctx;
5835                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
5836         }
5837
5838         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
5839                 return -EINVAL;
5840
5841         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
5842                 err = nft_get_set_elem(&ctx, set, attr);
5843                 if (err < 0) {
5844                         NL_SET_BAD_ATTR(extack, attr);
5845                         break;
5846                 }
5847         }
5848
5849         return err;
5850 }
5851
5852 static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
5853                                      const struct nft_set *set,
5854                                      const struct nft_set_elem *elem,
5855                                      int event)
5856 {
5857         struct nftables_pernet *nft_net;
5858         struct net *net = ctx->net;
5859         u32 portid = ctx->portid;
5860         struct sk_buff *skb;
5861         u16 flags = 0;
5862         int err;
5863
5864         if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5865                 return;
5866
5867         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5868         if (skb == NULL)
5869                 goto err;
5870
5871         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
5872                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
5873
5874         err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
5875                                           set, elem);
5876         if (err < 0) {
5877                 kfree_skb(skb);
5878                 goto err;
5879         }
5880
5881         nft_net = nft_pernet(net);
5882         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
5883         return;
5884 err:
5885         nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
5886 }
5887
5888 static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
5889                                               int msg_type,
5890                                               struct nft_set *set)
5891 {
5892         struct nft_trans *trans;
5893
5894         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
5895         if (trans == NULL)
5896                 return NULL;
5897
5898         nft_trans_elem_set(trans) = set;
5899         return trans;
5900 }
5901
5902 struct nft_expr *nft_set_elem_expr_alloc(const struct nft_ctx *ctx,
5903                                          const struct nft_set *set,
5904                                          const struct nlattr *attr)
5905 {
5906         struct nft_expr *expr;
5907         int err;
5908
5909         expr = nft_expr_init(ctx, attr);
5910         if (IS_ERR(expr))
5911                 return expr;
5912
5913         err = -EOPNOTSUPP;
5914         if (expr->ops->type->flags & NFT_EXPR_GC) {
5915                 if (set->flags & NFT_SET_TIMEOUT)
5916                         goto err_set_elem_expr;
5917                 if (!set->ops->gc_init)
5918                         goto err_set_elem_expr;
5919                 set->ops->gc_init(set);
5920         }
5921
5922         return expr;
5923
5924 err_set_elem_expr:
5925         nft_expr_destroy(ctx, expr);
5926         return ERR_PTR(err);
5927 }
5928
5929 static int nft_set_ext_check(const struct nft_set_ext_tmpl *tmpl, u8 id, u32 len)
5930 {
5931         len += nft_set_ext_types[id].len;
5932         if (len > tmpl->ext_len[id] ||
5933             len > U8_MAX)
5934                 return -1;
5935
5936         return 0;
5937 }
5938
5939 static int nft_set_ext_memcpy(const struct nft_set_ext_tmpl *tmpl, u8 id,
5940                               void *to, const void *from, u32 len)
5941 {
5942         if (nft_set_ext_check(tmpl, id, len) < 0)
5943                 return -1;
5944
5945         memcpy(to, from, len);
5946
5947         return 0;
5948 }
5949
5950 void *nft_set_elem_init(const struct nft_set *set,
5951                         const struct nft_set_ext_tmpl *tmpl,
5952                         const u32 *key, const u32 *key_end,
5953                         const u32 *data, u64 timeout, u64 expiration, gfp_t gfp)
5954 {
5955         struct nft_set_ext *ext;
5956         void *elem;
5957
5958         elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
5959         if (elem == NULL)
5960                 return ERR_PTR(-ENOMEM);
5961
5962         ext = nft_set_elem_ext(set, elem);
5963         nft_set_ext_init(ext, tmpl);
5964
5965         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
5966             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_KEY,
5967                                nft_set_ext_key(ext), key, set->klen) < 0)
5968                 goto err_ext_check;
5969
5970         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
5971             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_KEY_END,
5972                                nft_set_ext_key_end(ext), key_end, set->klen) < 0)
5973                 goto err_ext_check;
5974
5975         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5976             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_DATA,
5977                                nft_set_ext_data(ext), data, set->dlen) < 0)
5978                 goto err_ext_check;
5979
5980         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5981                 *nft_set_ext_expiration(ext) = get_jiffies_64() + expiration;
5982                 if (expiration == 0)
5983                         *nft_set_ext_expiration(ext) += timeout;
5984         }
5985         if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
5986                 *nft_set_ext_timeout(ext) = timeout;
5987
5988         return elem;
5989
5990 err_ext_check:
5991         kfree(elem);
5992
5993         return ERR_PTR(-EINVAL);
5994 }
5995
5996 static void __nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5997                                         struct nft_expr *expr)
5998 {
5999         if (expr->ops->destroy_clone) {
6000                 expr->ops->destroy_clone(ctx, expr);
6001                 module_put(expr->ops->type->owner);
6002         } else {
6003                 nf_tables_expr_destroy(ctx, expr);
6004         }
6005 }
6006
6007 static void nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
6008                                       struct nft_set_elem_expr *elem_expr)
6009 {
6010         struct nft_expr *expr;
6011         u32 size;
6012
6013         nft_setelem_expr_foreach(expr, elem_expr, size)
6014                 __nft_set_elem_expr_destroy(ctx, expr);
6015 }
6016
6017 /* Drop references and destroy. Called from gc, dynset and abort path. */
6018 void nft_set_elem_destroy(const struct nft_set *set, void *elem,
6019                           bool destroy_expr)
6020 {
6021         struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
6022         struct nft_ctx ctx = {
6023                 .net    = read_pnet(&set->net),
6024                 .family = set->table->family,
6025         };
6026
6027         nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
6028         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6029                 nft_data_release(nft_set_ext_data(ext), set->dtype);
6030         if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
6031                 nft_set_elem_expr_destroy(&ctx, nft_set_ext_expr(ext));
6032
6033         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6034                 nft_use_dec(&(*nft_set_ext_obj(ext))->use);
6035         kfree(elem);
6036 }
6037 EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
6038
6039 /* Destroy element. References have been already dropped in the preparation
6040  * path via nft_setelem_data_deactivate().
6041  */
6042 void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
6043                                 const struct nft_set *set, void *elem)
6044 {
6045         struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
6046
6047         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
6048                 nft_set_elem_expr_destroy(ctx, nft_set_ext_expr(ext));
6049
6050         kfree(elem);
6051 }
6052
6053 int nft_set_elem_expr_clone(const struct nft_ctx *ctx, struct nft_set *set,
6054                             struct nft_expr *expr_array[])
6055 {
6056         struct nft_expr *expr;
6057         int err, i, k;
6058
6059         for (i = 0; i < set->num_exprs; i++) {
6060                 expr = kzalloc(set->exprs[i]->ops->size, GFP_KERNEL_ACCOUNT);
6061                 if (!expr)
6062                         goto err_expr;
6063
6064                 err = nft_expr_clone(expr, set->exprs[i]);
6065                 if (err < 0) {
6066                         kfree(expr);
6067                         goto err_expr;
6068                 }
6069                 expr_array[i] = expr;
6070         }
6071
6072         return 0;
6073
6074 err_expr:
6075         for (k = i - 1; k >= 0; k--)
6076                 nft_expr_destroy(ctx, expr_array[k]);
6077
6078         return -ENOMEM;
6079 }
6080
6081 static int nft_set_elem_expr_setup(struct nft_ctx *ctx,
6082                                    const struct nft_set_ext_tmpl *tmpl,
6083                                    const struct nft_set_ext *ext,
6084                                    struct nft_expr *expr_array[],
6085                                    u32 num_exprs)
6086 {
6087         struct nft_set_elem_expr *elem_expr = nft_set_ext_expr(ext);
6088         u32 len = sizeof(struct nft_set_elem_expr);
6089         struct nft_expr *expr;
6090         int i, err;
6091
6092         if (num_exprs == 0)
6093                 return 0;
6094
6095         for (i = 0; i < num_exprs; i++)
6096                 len += expr_array[i]->ops->size;
6097
6098         if (nft_set_ext_check(tmpl, NFT_SET_EXT_EXPRESSIONS, len) < 0)
6099                 return -EINVAL;
6100
6101         for (i = 0; i < num_exprs; i++) {
6102                 expr = nft_setelem_expr_at(elem_expr, elem_expr->size);
6103                 err = nft_expr_clone(expr, expr_array[i]);
6104                 if (err < 0)
6105                         goto err_elem_expr_setup;
6106
6107                 elem_expr->size += expr_array[i]->ops->size;
6108                 nft_expr_destroy(ctx, expr_array[i]);
6109                 expr_array[i] = NULL;
6110         }
6111
6112         return 0;
6113
6114 err_elem_expr_setup:
6115         for (; i < num_exprs; i++) {
6116                 nft_expr_destroy(ctx, expr_array[i]);
6117                 expr_array[i] = NULL;
6118         }
6119
6120         return -ENOMEM;
6121 }
6122
6123 struct nft_set_ext *nft_set_catchall_lookup(const struct net *net,
6124                                             const struct nft_set *set)
6125 {
6126         struct nft_set_elem_catchall *catchall;
6127         u8 genmask = nft_genmask_cur(net);
6128         struct nft_set_ext *ext;
6129
6130         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6131                 ext = nft_set_elem_ext(set, catchall->elem);
6132                 if (nft_set_elem_active(ext, genmask) &&
6133                     !nft_set_elem_expired(ext) &&
6134                     !nft_set_elem_is_dead(ext))
6135                         return ext;
6136         }
6137
6138         return NULL;
6139 }
6140 EXPORT_SYMBOL_GPL(nft_set_catchall_lookup);
6141
6142 static int nft_setelem_catchall_insert(const struct net *net,
6143                                        struct nft_set *set,
6144                                        const struct nft_set_elem *elem,
6145                                        struct nft_set_ext **pext)
6146 {
6147         struct nft_set_elem_catchall *catchall;
6148         u8 genmask = nft_genmask_next(net);
6149         struct nft_set_ext *ext;
6150
6151         list_for_each_entry(catchall, &set->catchall_list, list) {
6152                 ext = nft_set_elem_ext(set, catchall->elem);
6153                 if (nft_set_elem_active(ext, genmask)) {
6154                         *pext = ext;
6155                         return -EEXIST;
6156                 }
6157         }
6158
6159         catchall = kmalloc(sizeof(*catchall), GFP_KERNEL);
6160         if (!catchall)
6161                 return -ENOMEM;
6162
6163         catchall->elem = elem->priv;
6164         list_add_tail_rcu(&catchall->list, &set->catchall_list);
6165
6166         return 0;
6167 }
6168
6169 static int nft_setelem_insert(const struct net *net,
6170                               struct nft_set *set,
6171                               const struct nft_set_elem *elem,
6172                               struct nft_set_ext **ext, unsigned int flags)
6173 {
6174         int ret;
6175
6176         if (flags & NFT_SET_ELEM_CATCHALL)
6177                 ret = nft_setelem_catchall_insert(net, set, elem, ext);
6178         else
6179                 ret = set->ops->insert(net, set, elem, ext);
6180
6181         return ret;
6182 }
6183
6184 static bool nft_setelem_is_catchall(const struct nft_set *set,
6185                                     const struct nft_set_elem *elem)
6186 {
6187         struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6188
6189         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
6190             *nft_set_ext_flags(ext) & NFT_SET_ELEM_CATCHALL)
6191                 return true;
6192
6193         return false;
6194 }
6195
6196 static void nft_setelem_activate(struct net *net, struct nft_set *set,
6197                                  struct nft_set_elem *elem)
6198 {
6199         struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6200
6201         if (nft_setelem_is_catchall(set, elem)) {
6202                 nft_set_elem_change_active(net, set, ext);
6203         } else {
6204                 set->ops->activate(net, set, elem);
6205         }
6206 }
6207
6208 static int nft_setelem_catchall_deactivate(const struct net *net,
6209                                            struct nft_set *set,
6210                                            struct nft_set_elem *elem)
6211 {
6212         struct nft_set_elem_catchall *catchall;
6213         struct nft_set_ext *ext;
6214
6215         list_for_each_entry(catchall, &set->catchall_list, list) {
6216                 ext = nft_set_elem_ext(set, catchall->elem);
6217                 if (!nft_is_active_next(net, ext))
6218                         continue;
6219
6220                 kfree(elem->priv);
6221                 elem->priv = catchall->elem;
6222                 nft_set_elem_change_active(net, set, ext);
6223                 return 0;
6224         }
6225
6226         return -ENOENT;
6227 }
6228
6229 static int __nft_setelem_deactivate(const struct net *net,
6230                                     struct nft_set *set,
6231                                     struct nft_set_elem *elem)
6232 {
6233         void *priv;
6234
6235         priv = set->ops->deactivate(net, set, elem);
6236         if (!priv)
6237                 return -ENOENT;
6238
6239         kfree(elem->priv);
6240         elem->priv = priv;
6241         set->ndeact++;
6242
6243         return 0;
6244 }
6245
6246 static int nft_setelem_deactivate(const struct net *net,
6247                                   struct nft_set *set,
6248                                   struct nft_set_elem *elem, u32 flags)
6249 {
6250         int ret;
6251
6252         if (flags & NFT_SET_ELEM_CATCHALL)
6253                 ret = nft_setelem_catchall_deactivate(net, set, elem);
6254         else
6255                 ret = __nft_setelem_deactivate(net, set, elem);
6256
6257         return ret;
6258 }
6259
6260 static void nft_setelem_catchall_destroy(struct nft_set_elem_catchall *catchall)
6261 {
6262         list_del_rcu(&catchall->list);
6263         kfree_rcu(catchall, rcu);
6264 }
6265
6266 static void nft_setelem_catchall_remove(const struct net *net,
6267                                         const struct nft_set *set,
6268                                         const struct nft_set_elem *elem)
6269 {
6270         struct nft_set_elem_catchall *catchall, *next;
6271
6272         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
6273                 if (catchall->elem == elem->priv) {
6274                         nft_setelem_catchall_destroy(catchall);
6275                         break;
6276                 }
6277         }
6278 }
6279
6280 static void nft_setelem_remove(const struct net *net,
6281                                const struct nft_set *set,
6282                                const struct nft_set_elem *elem)
6283 {
6284         if (nft_setelem_is_catchall(set, elem))
6285                 nft_setelem_catchall_remove(net, set, elem);
6286         else
6287                 set->ops->remove(net, set, elem);
6288 }
6289
6290 static bool nft_setelem_valid_key_end(const struct nft_set *set,
6291                                       struct nlattr **nla, u32 flags)
6292 {
6293         if ((set->flags & (NFT_SET_CONCAT | NFT_SET_INTERVAL)) ==
6294                           (NFT_SET_CONCAT | NFT_SET_INTERVAL)) {
6295                 if (flags & NFT_SET_ELEM_INTERVAL_END)
6296                         return false;
6297
6298                 if (nla[NFTA_SET_ELEM_KEY_END] &&
6299                     flags & NFT_SET_ELEM_CATCHALL)
6300                         return false;
6301         } else {
6302                 if (nla[NFTA_SET_ELEM_KEY_END])
6303                         return false;
6304         }
6305
6306         return true;
6307 }
6308
6309 static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
6310                             const struct nlattr *attr, u32 nlmsg_flags)
6311 {
6312         struct nft_expr *expr_array[NFT_SET_EXPR_MAX] = {};
6313         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6314         u8 genmask = nft_genmask_next(ctx->net);
6315         u32 flags = 0, size = 0, num_exprs = 0;
6316         struct nft_set_ext_tmpl tmpl;
6317         struct nft_set_ext *ext, *ext2;
6318         struct nft_set_elem elem;
6319         struct nft_set_binding *binding;
6320         struct nft_object *obj = NULL;
6321         struct nft_userdata *udata;
6322         struct nft_data_desc desc;
6323         enum nft_registers dreg;
6324         struct nft_trans *trans;
6325         u64 timeout;
6326         u64 expiration;
6327         int err, i;
6328         u8 ulen;
6329
6330         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6331                                           nft_set_elem_policy, NULL);
6332         if (err < 0)
6333                 return err;
6334
6335         nft_set_ext_prepare(&tmpl);
6336
6337         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6338         if (err < 0)
6339                 return err;
6340
6341         if (((flags & NFT_SET_ELEM_CATCHALL) && nla[NFTA_SET_ELEM_KEY]) ||
6342             (!(flags & NFT_SET_ELEM_CATCHALL) && !nla[NFTA_SET_ELEM_KEY]))
6343                 return -EINVAL;
6344
6345         if (flags != 0) {
6346                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
6347                 if (err < 0)
6348                         return err;
6349         }
6350
6351         if (set->flags & NFT_SET_MAP) {
6352                 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
6353                     !(flags & NFT_SET_ELEM_INTERVAL_END))
6354                         return -EINVAL;
6355         } else {
6356                 if (nla[NFTA_SET_ELEM_DATA] != NULL)
6357                         return -EINVAL;
6358         }
6359
6360         if (set->flags & NFT_SET_OBJECT) {
6361                 if (!nla[NFTA_SET_ELEM_OBJREF] &&
6362                     !(flags & NFT_SET_ELEM_INTERVAL_END))
6363                         return -EINVAL;
6364         } else {
6365                 if (nla[NFTA_SET_ELEM_OBJREF])
6366                         return -EINVAL;
6367         }
6368
6369         if (!nft_setelem_valid_key_end(set, nla, flags))
6370                 return -EINVAL;
6371
6372         if ((flags & NFT_SET_ELEM_INTERVAL_END) &&
6373              (nla[NFTA_SET_ELEM_DATA] ||
6374               nla[NFTA_SET_ELEM_OBJREF] ||
6375               nla[NFTA_SET_ELEM_TIMEOUT] ||
6376               nla[NFTA_SET_ELEM_EXPIRATION] ||
6377               nla[NFTA_SET_ELEM_USERDATA] ||
6378               nla[NFTA_SET_ELEM_EXPR] ||
6379               nla[NFTA_SET_ELEM_KEY_END] ||
6380               nla[NFTA_SET_ELEM_EXPRESSIONS]))
6381                 return -EINVAL;
6382
6383         timeout = 0;
6384         if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
6385                 if (!(set->flags & NFT_SET_TIMEOUT))
6386                         return -EINVAL;
6387                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
6388                                             &timeout);
6389                 if (err)
6390                         return err;
6391         } else if (set->flags & NFT_SET_TIMEOUT &&
6392                    !(flags & NFT_SET_ELEM_INTERVAL_END)) {
6393                 timeout = READ_ONCE(set->timeout);
6394         }
6395
6396         expiration = 0;
6397         if (nla[NFTA_SET_ELEM_EXPIRATION] != NULL) {
6398                 if (!(set->flags & NFT_SET_TIMEOUT))
6399                         return -EINVAL;
6400                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_EXPIRATION],
6401                                             &expiration);
6402                 if (err)
6403                         return err;
6404         }
6405
6406         if (nla[NFTA_SET_ELEM_EXPR]) {
6407                 struct nft_expr *expr;
6408
6409                 if (set->num_exprs && set->num_exprs != 1)
6410                         return -EOPNOTSUPP;
6411
6412                 expr = nft_set_elem_expr_alloc(ctx, set,
6413                                                nla[NFTA_SET_ELEM_EXPR]);
6414                 if (IS_ERR(expr))
6415                         return PTR_ERR(expr);
6416
6417                 expr_array[0] = expr;
6418                 num_exprs = 1;
6419
6420                 if (set->num_exprs && set->exprs[0]->ops != expr->ops) {
6421                         err = -EOPNOTSUPP;
6422                         goto err_set_elem_expr;
6423                 }
6424         } else if (nla[NFTA_SET_ELEM_EXPRESSIONS]) {
6425                 struct nft_expr *expr;
6426                 struct nlattr *tmp;
6427                 int left;
6428
6429                 i = 0;
6430                 nla_for_each_nested(tmp, nla[NFTA_SET_ELEM_EXPRESSIONS], left) {
6431                         if (i == NFT_SET_EXPR_MAX ||
6432                             (set->num_exprs && set->num_exprs == i)) {
6433                                 err = -E2BIG;
6434                                 goto err_set_elem_expr;
6435                         }
6436                         if (nla_type(tmp) != NFTA_LIST_ELEM) {
6437                                 err = -EINVAL;
6438                                 goto err_set_elem_expr;
6439                         }
6440                         expr = nft_set_elem_expr_alloc(ctx, set, tmp);
6441                         if (IS_ERR(expr)) {
6442                                 err = PTR_ERR(expr);
6443                                 goto err_set_elem_expr;
6444                         }
6445                         expr_array[i] = expr;
6446                         num_exprs++;
6447
6448                         if (set->num_exprs && expr->ops != set->exprs[i]->ops) {
6449                                 err = -EOPNOTSUPP;
6450                                 goto err_set_elem_expr;
6451                         }
6452                         i++;
6453                 }
6454                 if (set->num_exprs && set->num_exprs != i) {
6455                         err = -EOPNOTSUPP;
6456                         goto err_set_elem_expr;
6457                 }
6458         } else if (set->num_exprs > 0 &&
6459                    !(flags & NFT_SET_ELEM_INTERVAL_END)) {
6460                 err = nft_set_elem_expr_clone(ctx, set, expr_array);
6461                 if (err < 0)
6462                         goto err_set_elem_expr_clone;
6463
6464                 num_exprs = set->num_exprs;
6465         }
6466
6467         if (nla[NFTA_SET_ELEM_KEY]) {
6468                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6469                                             nla[NFTA_SET_ELEM_KEY]);
6470                 if (err < 0)
6471                         goto err_set_elem_expr;
6472
6473                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6474                 if (err < 0)
6475                         goto err_parse_key;
6476         }
6477
6478         if (nla[NFTA_SET_ELEM_KEY_END]) {
6479                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6480                                             nla[NFTA_SET_ELEM_KEY_END]);
6481                 if (err < 0)
6482                         goto err_parse_key;
6483
6484                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6485                 if (err < 0)
6486                         goto err_parse_key_end;
6487         }
6488
6489         if (timeout > 0) {
6490                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
6491                 if (err < 0)
6492                         goto err_parse_key_end;
6493
6494                 if (timeout != READ_ONCE(set->timeout)) {
6495                         err = nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
6496                         if (err < 0)
6497                                 goto err_parse_key_end;
6498                 }
6499         }
6500
6501         if (num_exprs) {
6502                 for (i = 0; i < num_exprs; i++)
6503                         size += expr_array[i]->ops->size;
6504
6505                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_EXPRESSIONS,
6506                                              sizeof(struct nft_set_elem_expr) + size);
6507                 if (err < 0)
6508                         goto err_parse_key_end;
6509         }
6510
6511         if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
6512                 obj = nft_obj_lookup(ctx->net, ctx->table,
6513                                      nla[NFTA_SET_ELEM_OBJREF],
6514                                      set->objtype, genmask);
6515                 if (IS_ERR(obj)) {
6516                         err = PTR_ERR(obj);
6517                         obj = NULL;
6518                         goto err_parse_key_end;
6519                 }
6520
6521                 if (!nft_use_inc(&obj->use)) {
6522                         err = -EMFILE;
6523                         obj = NULL;
6524                         goto err_parse_key_end;
6525                 }
6526
6527                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
6528                 if (err < 0)
6529                         goto err_parse_key_end;
6530         }
6531
6532         if (nla[NFTA_SET_ELEM_DATA] != NULL) {
6533                 err = nft_setelem_parse_data(ctx, set, &desc, &elem.data.val,
6534                                              nla[NFTA_SET_ELEM_DATA]);
6535                 if (err < 0)
6536                         goto err_parse_key_end;
6537
6538                 dreg = nft_type_to_reg(set->dtype);
6539                 list_for_each_entry(binding, &set->bindings, list) {
6540                         struct nft_ctx bind_ctx = {
6541                                 .net    = ctx->net,
6542                                 .family = ctx->family,
6543                                 .table  = ctx->table,
6544                                 .chain  = (struct nft_chain *)binding->chain,
6545                         };
6546
6547                         if (!(binding->flags & NFT_SET_MAP))
6548                                 continue;
6549
6550                         err = nft_validate_register_store(&bind_ctx, dreg,
6551                                                           &elem.data.val,
6552                                                           desc.type, desc.len);
6553                         if (err < 0)
6554                                 goto err_parse_data;
6555
6556                         if (desc.type == NFT_DATA_VERDICT &&
6557                             (elem.data.val.verdict.code == NFT_GOTO ||
6558                              elem.data.val.verdict.code == NFT_JUMP))
6559                                 nft_validate_state_update(ctx->net,
6560                                                           NFT_VALIDATE_NEED);
6561                 }
6562
6563                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, desc.len);
6564                 if (err < 0)
6565                         goto err_parse_data;
6566         }
6567
6568         /* The full maximum length of userdata can exceed the maximum
6569          * offset value (U8_MAX) for following extensions, therefor it
6570          * must be the last extension added.
6571          */
6572         ulen = 0;
6573         if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
6574                 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
6575                 if (ulen > 0) {
6576                         err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
6577                                                      ulen);
6578                         if (err < 0)
6579                                 goto err_parse_data;
6580                 }
6581         }
6582
6583         elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6584                                       elem.key_end.val.data, elem.data.val.data,
6585                                       timeout, expiration, GFP_KERNEL_ACCOUNT);
6586         if (IS_ERR(elem.priv)) {
6587                 err = PTR_ERR(elem.priv);
6588                 goto err_parse_data;
6589         }
6590
6591         ext = nft_set_elem_ext(set, elem.priv);
6592         if (flags)
6593                 *nft_set_ext_flags(ext) = flags;
6594
6595         if (obj)
6596                 *nft_set_ext_obj(ext) = obj;
6597
6598         if (ulen > 0) {
6599                 if (nft_set_ext_check(&tmpl, NFT_SET_EXT_USERDATA, ulen) < 0) {
6600                         err = -EINVAL;
6601                         goto err_elem_free;
6602                 }
6603                 udata = nft_set_ext_userdata(ext);
6604                 udata->len = ulen - 1;
6605                 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
6606         }
6607         err = nft_set_elem_expr_setup(ctx, &tmpl, ext, expr_array, num_exprs);
6608         if (err < 0)
6609                 goto err_elem_free;
6610
6611         trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
6612         if (trans == NULL) {
6613                 err = -ENOMEM;
6614                 goto err_elem_free;
6615         }
6616
6617         ext->genmask = nft_genmask_cur(ctx->net);
6618
6619         err = nft_setelem_insert(ctx->net, set, &elem, &ext2, flags);
6620         if (err) {
6621                 if (err == -EEXIST) {
6622                         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
6623                             nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
6624                             nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
6625                             nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF))
6626                                 goto err_element_clash;
6627                         if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
6628                              nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
6629                              memcmp(nft_set_ext_data(ext),
6630                                     nft_set_ext_data(ext2), set->dlen) != 0) ||
6631                             (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
6632                              nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
6633                              *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
6634                                 goto err_element_clash;
6635                         else if (!(nlmsg_flags & NLM_F_EXCL))
6636                                 err = 0;
6637                 } else if (err == -ENOTEMPTY) {
6638                         /* ENOTEMPTY reports overlapping between this element
6639                          * and an existing one.
6640                          */
6641                         err = -EEXIST;
6642                 }
6643                 goto err_element_clash;
6644         }
6645
6646         if (!(flags & NFT_SET_ELEM_CATCHALL) && set->size &&
6647             !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact)) {
6648                 err = -ENFILE;
6649                 goto err_set_full;
6650         }
6651
6652         nft_trans_elem(trans) = elem;
6653         nft_trans_commit_list_add_tail(ctx->net, trans);
6654         return 0;
6655
6656 err_set_full:
6657         nft_setelem_remove(ctx->net, set, &elem);
6658 err_element_clash:
6659         kfree(trans);
6660 err_elem_free:
6661         nf_tables_set_elem_destroy(ctx, set, elem.priv);
6662 err_parse_data:
6663         if (nla[NFTA_SET_ELEM_DATA] != NULL)
6664                 nft_data_release(&elem.data.val, desc.type);
6665 err_parse_key_end:
6666         if (obj)
6667                 nft_use_dec_restore(&obj->use);
6668
6669         nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
6670 err_parse_key:
6671         nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6672 err_set_elem_expr:
6673         for (i = 0; i < num_exprs && expr_array[i]; i++)
6674                 nft_expr_destroy(ctx, expr_array[i]);
6675 err_set_elem_expr_clone:
6676         return err;
6677 }
6678
6679 static int nf_tables_newsetelem(struct sk_buff *skb,
6680                                 const struct nfnl_info *info,
6681                                 const struct nlattr * const nla[])
6682 {
6683         struct nftables_pernet *nft_net = nft_pernet(info->net);
6684         struct netlink_ext_ack *extack = info->extack;
6685         u8 genmask = nft_genmask_next(info->net);
6686         u8 family = info->nfmsg->nfgen_family;
6687         struct net *net = info->net;
6688         const struct nlattr *attr;
6689         struct nft_table *table;
6690         struct nft_set *set;
6691         struct nft_ctx ctx;
6692         int rem, err;
6693
6694         if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
6695                 return -EINVAL;
6696
6697         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6698                                  genmask, NETLINK_CB(skb).portid);
6699         if (IS_ERR(table)) {
6700                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6701                 return PTR_ERR(table);
6702         }
6703
6704         set = nft_set_lookup_global(net, table, nla[NFTA_SET_ELEM_LIST_SET],
6705                                     nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
6706         if (IS_ERR(set))
6707                 return PTR_ERR(set);
6708
6709         if (!list_empty(&set->bindings) &&
6710             (set->flags & (NFT_SET_CONSTANT | NFT_SET_ANONYMOUS)))
6711                 return -EBUSY;
6712
6713         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6714
6715         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6716                 err = nft_add_set_elem(&ctx, set, attr, info->nlh->nlmsg_flags);
6717                 if (err < 0) {
6718                         NL_SET_BAD_ATTR(extack, attr);
6719                         return err;
6720                 }
6721         }
6722
6723         if (nft_net->validate_state == NFT_VALIDATE_DO)
6724                 return nft_table_validate(net, table);
6725
6726         return 0;
6727 }
6728
6729 /**
6730  *      nft_data_hold - hold a nft_data item
6731  *
6732  *      @data: struct nft_data to release
6733  *      @type: type of data
6734  *
6735  *      Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
6736  *      NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
6737  *      NFT_GOTO verdicts. This function must be called on active data objects
6738  *      from the second phase of the commit protocol.
6739  */
6740 void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
6741 {
6742         struct nft_chain *chain;
6743
6744         if (type == NFT_DATA_VERDICT) {
6745                 switch (data->verdict.code) {
6746                 case NFT_JUMP:
6747                 case NFT_GOTO:
6748                         chain = data->verdict.chain;
6749                         nft_use_inc_restore(&chain->use);
6750                         break;
6751                 }
6752         }
6753 }
6754
6755 static void nft_setelem_data_activate(const struct net *net,
6756                                       const struct nft_set *set,
6757                                       struct nft_set_elem *elem)
6758 {
6759         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6760
6761         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6762                 nft_data_hold(nft_set_ext_data(ext), set->dtype);
6763         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6764                 nft_use_inc_restore(&(*nft_set_ext_obj(ext))->use);
6765 }
6766
6767 void nft_setelem_data_deactivate(const struct net *net,
6768                                  const struct nft_set *set,
6769                                  struct nft_set_elem *elem)
6770 {
6771         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6772
6773         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6774                 nft_data_release(nft_set_ext_data(ext), set->dtype);
6775         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6776                 nft_use_dec(&(*nft_set_ext_obj(ext))->use);
6777 }
6778
6779 static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
6780                            const struct nlattr *attr)
6781 {
6782         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6783         struct nft_set_ext_tmpl tmpl;
6784         struct nft_set_elem elem;
6785         struct nft_set_ext *ext;
6786         struct nft_trans *trans;
6787         u32 flags = 0;
6788         int err;
6789
6790         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6791                                           nft_set_elem_policy, NULL);
6792         if (err < 0)
6793                 return err;
6794
6795         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6796         if (err < 0)
6797                 return err;
6798
6799         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
6800                 return -EINVAL;
6801
6802         if (!nft_setelem_valid_key_end(set, nla, flags))
6803                 return -EINVAL;
6804
6805         nft_set_ext_prepare(&tmpl);
6806
6807         if (flags != 0) {
6808                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
6809                 if (err < 0)
6810                         return err;
6811         }
6812
6813         if (nla[NFTA_SET_ELEM_KEY]) {
6814                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6815                                             nla[NFTA_SET_ELEM_KEY]);
6816                 if (err < 0)
6817                         return err;
6818
6819                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6820                 if (err < 0)
6821                         goto fail_elem;
6822         }
6823
6824         if (nla[NFTA_SET_ELEM_KEY_END]) {
6825                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6826                                             nla[NFTA_SET_ELEM_KEY_END]);
6827                 if (err < 0)
6828                         goto fail_elem;
6829
6830                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6831                 if (err < 0)
6832                         goto fail_elem_key_end;
6833         }
6834
6835         err = -ENOMEM;
6836         elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6837                                       elem.key_end.val.data, NULL, 0, 0,
6838                                       GFP_KERNEL_ACCOUNT);
6839         if (IS_ERR(elem.priv)) {
6840                 err = PTR_ERR(elem.priv);
6841                 goto fail_elem_key_end;
6842         }
6843
6844         ext = nft_set_elem_ext(set, elem.priv);
6845         if (flags)
6846                 *nft_set_ext_flags(ext) = flags;
6847
6848         trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
6849         if (trans == NULL)
6850                 goto fail_trans;
6851
6852         err = nft_setelem_deactivate(ctx->net, set, &elem, flags);
6853         if (err < 0)
6854                 goto fail_ops;
6855
6856         nft_setelem_data_deactivate(ctx->net, set, &elem);
6857
6858         nft_trans_elem(trans) = elem;
6859         nft_trans_commit_list_add_tail(ctx->net, trans);
6860         return 0;
6861
6862 fail_ops:
6863         kfree(trans);
6864 fail_trans:
6865         kfree(elem.priv);
6866 fail_elem_key_end:
6867         nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
6868 fail_elem:
6869         nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6870         return err;
6871 }
6872
6873 static int nft_setelem_flush(const struct nft_ctx *ctx,
6874                              struct nft_set *set,
6875                              const struct nft_set_iter *iter,
6876                              struct nft_set_elem *elem)
6877 {
6878         struct nft_trans *trans;
6879         int err;
6880
6881         trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6882                                     sizeof(struct nft_trans_elem), GFP_ATOMIC);
6883         if (!trans)
6884                 return -ENOMEM;
6885
6886         if (!set->ops->flush(ctx->net, set, elem->priv)) {
6887                 err = -ENOENT;
6888                 goto err1;
6889         }
6890         set->ndeact++;
6891
6892         nft_setelem_data_deactivate(ctx->net, set, elem);
6893         nft_trans_elem_set(trans) = set;
6894         nft_trans_elem(trans) = *elem;
6895         nft_trans_commit_list_add_tail(ctx->net, trans);
6896
6897         return 0;
6898 err1:
6899         kfree(trans);
6900         return err;
6901 }
6902
6903 static int __nft_set_catchall_flush(const struct nft_ctx *ctx,
6904                                     struct nft_set *set,
6905                                     struct nft_set_elem *elem)
6906 {
6907         struct nft_trans *trans;
6908
6909         trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6910                                     sizeof(struct nft_trans_elem), GFP_KERNEL);
6911         if (!trans)
6912                 return -ENOMEM;
6913
6914         nft_setelem_data_deactivate(ctx->net, set, elem);
6915         nft_trans_elem_set(trans) = set;
6916         nft_trans_elem(trans) = *elem;
6917         nft_trans_commit_list_add_tail(ctx->net, trans);
6918
6919         return 0;
6920 }
6921
6922 static int nft_set_catchall_flush(const struct nft_ctx *ctx,
6923                                   struct nft_set *set)
6924 {
6925         u8 genmask = nft_genmask_next(ctx->net);
6926         struct nft_set_elem_catchall *catchall;
6927         struct nft_set_elem elem;
6928         struct nft_set_ext *ext;
6929         int ret = 0;
6930
6931         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6932                 ext = nft_set_elem_ext(set, catchall->elem);
6933                 if (!nft_set_elem_active(ext, genmask))
6934                         continue;
6935
6936                 elem.priv = catchall->elem;
6937                 ret = __nft_set_catchall_flush(ctx, set, &elem);
6938                 if (ret < 0)
6939                         break;
6940                 nft_set_elem_change_active(ctx->net, set, ext);
6941         }
6942
6943         return ret;
6944 }
6945
6946 static int nft_set_flush(struct nft_ctx *ctx, struct nft_set *set, u8 genmask)
6947 {
6948         struct nft_set_iter iter = {
6949                 .genmask        = genmask,
6950                 .fn             = nft_setelem_flush,
6951         };
6952
6953         set->ops->walk(ctx, set, &iter);
6954         if (!iter.err)
6955                 iter.err = nft_set_catchall_flush(ctx, set);
6956
6957         return iter.err;
6958 }
6959
6960 static int nf_tables_delsetelem(struct sk_buff *skb,
6961                                 const struct nfnl_info *info,
6962                                 const struct nlattr * const nla[])
6963 {
6964         struct netlink_ext_ack *extack = info->extack;
6965         u8 genmask = nft_genmask_next(info->net);
6966         u8 family = info->nfmsg->nfgen_family;
6967         struct net *net = info->net;
6968         const struct nlattr *attr;
6969         struct nft_table *table;
6970         struct nft_set *set;
6971         struct nft_ctx ctx;
6972         int rem, err = 0;
6973
6974         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6975                                  genmask, NETLINK_CB(skb).portid);
6976         if (IS_ERR(table)) {
6977                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6978                 return PTR_ERR(table);
6979         }
6980
6981         set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
6982         if (IS_ERR(set))
6983                 return PTR_ERR(set);
6984
6985         if (nft_set_is_anonymous(set))
6986                 return -EOPNOTSUPP;
6987
6988         if (!list_empty(&set->bindings) && (set->flags & NFT_SET_CONSTANT))
6989                 return -EBUSY;
6990
6991         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6992
6993         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
6994                 return nft_set_flush(&ctx, set, genmask);
6995
6996         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6997                 err = nft_del_setelem(&ctx, set, attr);
6998                 if (err < 0) {
6999                         NL_SET_BAD_ATTR(extack, attr);
7000                         break;
7001                 }
7002         }
7003         return err;
7004 }
7005
7006 /*
7007  * Stateful objects
7008  */
7009
7010 /**
7011  *      nft_register_obj- register nf_tables stateful object type
7012  *      @obj_type: object type
7013  *
7014  *      Registers the object type for use with nf_tables. Returns zero on
7015  *      success or a negative errno code otherwise.
7016  */
7017 int nft_register_obj(struct nft_object_type *obj_type)
7018 {
7019         if (obj_type->type == NFT_OBJECT_UNSPEC)
7020                 return -EINVAL;
7021
7022         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7023         list_add_rcu(&obj_type->list, &nf_tables_objects);
7024         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7025         return 0;
7026 }
7027 EXPORT_SYMBOL_GPL(nft_register_obj);
7028
7029 /**
7030  *      nft_unregister_obj - unregister nf_tables object type
7031  *      @obj_type: object type
7032  *
7033  *      Unregisters the object type for use with nf_tables.
7034  */
7035 void nft_unregister_obj(struct nft_object_type *obj_type)
7036 {
7037         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7038         list_del_rcu(&obj_type->list);
7039         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7040 }
7041 EXPORT_SYMBOL_GPL(nft_unregister_obj);
7042
7043 struct nft_object *nft_obj_lookup(const struct net *net,
7044                                   const struct nft_table *table,
7045                                   const struct nlattr *nla, u32 objtype,
7046                                   u8 genmask)
7047 {
7048         struct nft_object_hash_key k = { .table = table };
7049         char search[NFT_OBJ_MAXNAMELEN];
7050         struct rhlist_head *tmp, *list;
7051         struct nft_object *obj;
7052
7053         nla_strscpy(search, nla, sizeof(search));
7054         k.name = search;
7055
7056         WARN_ON_ONCE(!rcu_read_lock_held() &&
7057                      !lockdep_commit_lock_is_held(net));
7058
7059         rcu_read_lock();
7060         list = rhltable_lookup(&nft_objname_ht, &k, nft_objname_ht_params);
7061         if (!list)
7062                 goto out;
7063
7064         rhl_for_each_entry_rcu(obj, tmp, list, rhlhead) {
7065                 if (objtype == obj->ops->type->type &&
7066                     nft_active_genmask(obj, genmask)) {
7067                         rcu_read_unlock();
7068                         return obj;
7069                 }
7070         }
7071 out:
7072         rcu_read_unlock();
7073         return ERR_PTR(-ENOENT);
7074 }
7075 EXPORT_SYMBOL_GPL(nft_obj_lookup);
7076
7077 static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
7078                                                   const struct nlattr *nla,
7079                                                   u32 objtype, u8 genmask)
7080 {
7081         struct nft_object *obj;
7082
7083         list_for_each_entry(obj, &table->objects, list) {
7084                 if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
7085                     objtype == obj->ops->type->type &&
7086                     nft_active_genmask(obj, genmask))
7087                         return obj;
7088         }
7089         return ERR_PTR(-ENOENT);
7090 }
7091
7092 static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
7093         [NFTA_OBJ_TABLE]        = { .type = NLA_STRING,
7094                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
7095         [NFTA_OBJ_NAME]         = { .type = NLA_STRING,
7096                                     .len = NFT_OBJ_MAXNAMELEN - 1 },
7097         [NFTA_OBJ_TYPE]         = { .type = NLA_U32 },
7098         [NFTA_OBJ_DATA]         = { .type = NLA_NESTED },
7099         [NFTA_OBJ_HANDLE]       = { .type = NLA_U64},
7100         [NFTA_OBJ_USERDATA]     = { .type = NLA_BINARY,
7101                                     .len = NFT_USERDATA_MAXLEN },
7102 };
7103
7104 static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
7105                                        const struct nft_object_type *type,
7106                                        const struct nlattr *attr)
7107 {
7108         struct nlattr **tb;
7109         const struct nft_object_ops *ops;
7110         struct nft_object *obj;
7111         int err = -ENOMEM;
7112
7113         tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
7114         if (!tb)
7115                 goto err1;
7116
7117         if (attr) {
7118                 err = nla_parse_nested_deprecated(tb, type->maxattr, attr,
7119                                                   type->policy, NULL);
7120                 if (err < 0)
7121                         goto err2;
7122         } else {
7123                 memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
7124         }
7125
7126         if (type->select_ops) {
7127                 ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
7128                 if (IS_ERR(ops)) {
7129                         err = PTR_ERR(ops);
7130                         goto err2;
7131                 }
7132         } else {
7133                 ops = type->ops;
7134         }
7135
7136         err = -ENOMEM;
7137         obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL_ACCOUNT);
7138         if (!obj)
7139                 goto err2;
7140
7141         err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
7142         if (err < 0)
7143                 goto err3;
7144
7145         obj->ops = ops;
7146
7147         kfree(tb);
7148         return obj;
7149 err3:
7150         kfree(obj);
7151 err2:
7152         kfree(tb);
7153 err1:
7154         return ERR_PTR(err);
7155 }
7156
7157 static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
7158                            struct nft_object *obj, bool reset)
7159 {
7160         struct nlattr *nest;
7161
7162         nest = nla_nest_start_noflag(skb, attr);
7163         if (!nest)
7164                 goto nla_put_failure;
7165         if (obj->ops->dump(skb, obj, reset) < 0)
7166                 goto nla_put_failure;
7167         nla_nest_end(skb, nest);
7168         return 0;
7169
7170 nla_put_failure:
7171         return -1;
7172 }
7173
7174 static const struct nft_object_type *__nft_obj_type_get(u32 objtype, u8 family)
7175 {
7176         const struct nft_object_type *type;
7177
7178         list_for_each_entry_rcu(type, &nf_tables_objects, list) {
7179                 if (type->family != NFPROTO_UNSPEC &&
7180                     type->family != family)
7181                         continue;
7182
7183                 if (objtype == type->type)
7184                         return type;
7185         }
7186         return NULL;
7187 }
7188
7189 static const struct nft_object_type *
7190 nft_obj_type_get(struct net *net, u32 objtype, u8 family)
7191 {
7192         const struct nft_object_type *type;
7193
7194         rcu_read_lock();
7195         type = __nft_obj_type_get(objtype, family);
7196         if (type != NULL && try_module_get(type->owner)) {
7197                 rcu_read_unlock();
7198                 return type;
7199         }
7200         rcu_read_unlock();
7201
7202         lockdep_nfnl_nft_mutex_not_held();
7203 #ifdef CONFIG_MODULES
7204         if (type == NULL) {
7205                 if (nft_request_module(net, "nft-obj-%u", objtype) == -EAGAIN)
7206                         return ERR_PTR(-EAGAIN);
7207         }
7208 #endif
7209         return ERR_PTR(-ENOENT);
7210 }
7211
7212 static int nf_tables_updobj(const struct nft_ctx *ctx,
7213                             const struct nft_object_type *type,
7214                             const struct nlattr *attr,
7215                             struct nft_object *obj)
7216 {
7217         struct nft_object *newobj;
7218         struct nft_trans *trans;
7219         int err = -ENOMEM;
7220
7221         if (!try_module_get(type->owner))
7222                 return -ENOENT;
7223
7224         trans = nft_trans_alloc(ctx, NFT_MSG_NEWOBJ,
7225                                 sizeof(struct nft_trans_obj));
7226         if (!trans)
7227                 goto err_trans;
7228
7229         newobj = nft_obj_init(ctx, type, attr);
7230         if (IS_ERR(newobj)) {
7231                 err = PTR_ERR(newobj);
7232                 goto err_free_trans;
7233         }
7234
7235         nft_trans_obj(trans) = obj;
7236         nft_trans_obj_update(trans) = true;
7237         nft_trans_obj_newobj(trans) = newobj;
7238         nft_trans_commit_list_add_tail(ctx->net, trans);
7239
7240         return 0;
7241
7242 err_free_trans:
7243         kfree(trans);
7244 err_trans:
7245         module_put(type->owner);
7246         return err;
7247 }
7248
7249 static int nf_tables_newobj(struct sk_buff *skb, const struct nfnl_info *info,
7250                             const struct nlattr * const nla[])
7251 {
7252         struct netlink_ext_ack *extack = info->extack;
7253         u8 genmask = nft_genmask_next(info->net);
7254         u8 family = info->nfmsg->nfgen_family;
7255         const struct nft_object_type *type;
7256         struct net *net = info->net;
7257         struct nft_table *table;
7258         struct nft_object *obj;
7259         struct nft_ctx ctx;
7260         u32 objtype;
7261         int err;
7262
7263         if (!nla[NFTA_OBJ_TYPE] ||
7264             !nla[NFTA_OBJ_NAME] ||
7265             !nla[NFTA_OBJ_DATA])
7266                 return -EINVAL;
7267
7268         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
7269                                  NETLINK_CB(skb).portid);
7270         if (IS_ERR(table)) {
7271                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
7272                 return PTR_ERR(table);
7273         }
7274
7275         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7276         obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
7277         if (IS_ERR(obj)) {
7278                 err = PTR_ERR(obj);
7279                 if (err != -ENOENT) {
7280                         NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
7281                         return err;
7282                 }
7283         } else {
7284                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
7285                         NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
7286                         return -EEXIST;
7287                 }
7288                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
7289                         return -EOPNOTSUPP;
7290
7291                 type = __nft_obj_type_get(objtype, family);
7292                 if (WARN_ON_ONCE(!type))
7293                         return -ENOENT;
7294
7295                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7296
7297                 return nf_tables_updobj(&ctx, type, nla[NFTA_OBJ_DATA], obj);
7298         }
7299
7300         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7301
7302         if (!nft_use_inc(&table->use))
7303                 return -EMFILE;
7304
7305         type = nft_obj_type_get(net, objtype, family);
7306         if (IS_ERR(type)) {
7307                 err = PTR_ERR(type);
7308                 goto err_type;
7309         }
7310
7311         obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
7312         if (IS_ERR(obj)) {
7313                 err = PTR_ERR(obj);
7314                 goto err_init;
7315         }
7316         obj->key.table = table;
7317         obj->handle = nf_tables_alloc_handle(table);
7318
7319         obj->key.name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL_ACCOUNT);
7320         if (!obj->key.name) {
7321                 err = -ENOMEM;
7322                 goto err_strdup;
7323         }
7324
7325         if (nla[NFTA_OBJ_USERDATA]) {
7326                 obj->udata = nla_memdup(nla[NFTA_OBJ_USERDATA], GFP_KERNEL_ACCOUNT);
7327                 if (obj->udata == NULL)
7328                         goto err_userdata;
7329
7330                 obj->udlen = nla_len(nla[NFTA_OBJ_USERDATA]);
7331         }
7332
7333         err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
7334         if (err < 0)
7335                 goto err_trans;
7336
7337         err = rhltable_insert(&nft_objname_ht, &obj->rhlhead,
7338                               nft_objname_ht_params);
7339         if (err < 0)
7340                 goto err_obj_ht;
7341
7342         list_add_tail_rcu(&obj->list, &table->objects);
7343
7344         return 0;
7345 err_obj_ht:
7346         /* queued in transaction log */
7347         INIT_LIST_HEAD(&obj->list);
7348         return err;
7349 err_trans:
7350         kfree(obj->udata);
7351 err_userdata:
7352         kfree(obj->key.name);
7353 err_strdup:
7354         if (obj->ops->destroy)
7355                 obj->ops->destroy(&ctx, obj);
7356         kfree(obj);
7357 err_init:
7358         module_put(type->owner);
7359 err_type:
7360         nft_use_dec_restore(&table->use);
7361
7362         return err;
7363 }
7364
7365 static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
7366                                    u32 portid, u32 seq, int event, u32 flags,
7367                                    int family, const struct nft_table *table,
7368                                    struct nft_object *obj, bool reset)
7369 {
7370         struct nlmsghdr *nlh;
7371
7372         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
7373         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
7374                            NFNETLINK_V0, nft_base_seq(net));
7375         if (!nlh)
7376                 goto nla_put_failure;
7377
7378         if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
7379             nla_put_string(skb, NFTA_OBJ_NAME, obj->key.name) ||
7380             nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
7381             nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
7382             nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset) ||
7383             nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
7384                          NFTA_OBJ_PAD))
7385                 goto nla_put_failure;
7386
7387         if (obj->udata &&
7388             nla_put(skb, NFTA_OBJ_USERDATA, obj->udlen, obj->udata))
7389                 goto nla_put_failure;
7390
7391         nlmsg_end(skb, nlh);
7392         return 0;
7393
7394 nla_put_failure:
7395         nlmsg_trim(skb, nlh);
7396         return -1;
7397 }
7398
7399 static void audit_log_obj_reset(const struct nft_table *table,
7400                                 unsigned int base_seq, unsigned int nentries)
7401 {
7402         char *buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, base_seq);
7403
7404         audit_log_nfcfg(buf, table->family, nentries,
7405                         AUDIT_NFT_OP_OBJ_RESET, GFP_ATOMIC);
7406         kfree(buf);
7407 }
7408
7409 struct nft_obj_filter {
7410         char            *table;
7411         u32             type;
7412 };
7413
7414 static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
7415 {
7416         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
7417         const struct nft_table *table;
7418         unsigned int idx = 0, s_idx = cb->args[0];
7419         struct nft_obj_filter *filter = cb->data;
7420         struct net *net = sock_net(skb->sk);
7421         int family = nfmsg->nfgen_family;
7422         struct nftables_pernet *nft_net;
7423         unsigned int entries = 0;
7424         struct nft_object *obj;
7425         bool reset = false;
7426         int rc = 0;
7427
7428         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
7429                 reset = true;
7430
7431         rcu_read_lock();
7432         nft_net = nft_pernet(net);
7433         cb->seq = READ_ONCE(nft_net->base_seq);
7434
7435         list_for_each_entry_rcu(table, &nft_net->tables, list) {
7436                 if (family != NFPROTO_UNSPEC && family != table->family)
7437                         continue;
7438
7439                 entries = 0;
7440                 list_for_each_entry_rcu(obj, &table->objects, list) {
7441                         if (!nft_is_active(net, obj))
7442                                 goto cont;
7443                         if (idx < s_idx)
7444                                 goto cont;
7445                         if (idx > s_idx)
7446                                 memset(&cb->args[1], 0,
7447                                        sizeof(cb->args) - sizeof(cb->args[0]));
7448                         if (filter && filter->table &&
7449                             strcmp(filter->table, table->name))
7450                                 goto cont;
7451                         if (filter &&
7452                             filter->type != NFT_OBJECT_UNSPEC &&
7453                             obj->ops->type->type != filter->type)
7454                                 goto cont;
7455
7456                         rc = nf_tables_fill_obj_info(skb, net,
7457                                                      NETLINK_CB(cb->skb).portid,
7458                                                      cb->nlh->nlmsg_seq,
7459                                                      NFT_MSG_NEWOBJ,
7460                                                      NLM_F_MULTI | NLM_F_APPEND,
7461                                                      table->family, table,
7462                                                      obj, reset);
7463                         if (rc < 0)
7464                                 break;
7465
7466                         entries++;
7467                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
7468 cont:
7469                         idx++;
7470                 }
7471                 if (reset && entries)
7472                         audit_log_obj_reset(table, nft_net->base_seq, entries);
7473                 if (rc < 0)
7474                         break;
7475         }
7476         rcu_read_unlock();
7477
7478         cb->args[0] = idx;
7479         return skb->len;
7480 }
7481
7482 static int nf_tables_dump_obj_start(struct netlink_callback *cb)
7483 {
7484         const struct nlattr * const *nla = cb->data;
7485         struct nft_obj_filter *filter = NULL;
7486
7487         if (nla[NFTA_OBJ_TABLE] || nla[NFTA_OBJ_TYPE]) {
7488                 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
7489                 if (!filter)
7490                         return -ENOMEM;
7491
7492                 if (nla[NFTA_OBJ_TABLE]) {
7493                         filter->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
7494                         if (!filter->table) {
7495                                 kfree(filter);
7496                                 return -ENOMEM;
7497                         }
7498                 }
7499
7500                 if (nla[NFTA_OBJ_TYPE])
7501                         filter->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7502         }
7503
7504         cb->data = filter;
7505         return 0;
7506 }
7507
7508 static int nf_tables_dump_obj_done(struct netlink_callback *cb)
7509 {
7510         struct nft_obj_filter *filter = cb->data;
7511
7512         if (filter) {
7513                 kfree(filter->table);
7514                 kfree(filter);
7515         }
7516
7517         return 0;
7518 }
7519
7520 /* called with rcu_read_lock held */
7521 static int nf_tables_getobj(struct sk_buff *skb, const struct nfnl_info *info,
7522                             const struct nlattr * const nla[])
7523 {
7524         struct netlink_ext_ack *extack = info->extack;
7525         u8 genmask = nft_genmask_cur(info->net);
7526         u8 family = info->nfmsg->nfgen_family;
7527         const struct nft_table *table;
7528         struct net *net = info->net;
7529         struct nft_object *obj;
7530         struct sk_buff *skb2;
7531         bool reset = false;
7532         u32 objtype;
7533         int err;
7534
7535         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
7536                 struct netlink_dump_control c = {
7537                         .start = nf_tables_dump_obj_start,
7538                         .dump = nf_tables_dump_obj,
7539                         .done = nf_tables_dump_obj_done,
7540                         .module = THIS_MODULE,
7541                         .data = (void *)nla,
7542                 };
7543
7544                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
7545         }
7546
7547         if (!nla[NFTA_OBJ_NAME] ||
7548             !nla[NFTA_OBJ_TYPE])
7549                 return -EINVAL;
7550
7551         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask, 0);
7552         if (IS_ERR(table)) {
7553                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
7554                 return PTR_ERR(table);
7555         }
7556
7557         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7558         obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
7559         if (IS_ERR(obj)) {
7560                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
7561                 return PTR_ERR(obj);
7562         }
7563
7564         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7565         if (!skb2)
7566                 return -ENOMEM;
7567
7568         if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
7569                 reset = true;
7570
7571         if (reset) {
7572                 const struct nftables_pernet *nft_net;
7573                 char *buf;
7574
7575                 nft_net = nft_pernet(net);
7576                 buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, nft_net->base_seq);
7577
7578                 audit_log_nfcfg(buf,
7579                                 family,
7580                                 1,
7581                                 AUDIT_NFT_OP_OBJ_RESET,
7582                                 GFP_ATOMIC);
7583                 kfree(buf);
7584         }
7585
7586         err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
7587                                       info->nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
7588                                       family, table, obj, reset);
7589         if (err < 0)
7590                 goto err_fill_obj_info;
7591
7592         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7593
7594 err_fill_obj_info:
7595         kfree_skb(skb2);
7596         return err;
7597 }
7598
7599 static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
7600 {
7601         if (obj->ops->destroy)
7602                 obj->ops->destroy(ctx, obj);
7603
7604         module_put(obj->ops->type->owner);
7605         kfree(obj->key.name);
7606         kfree(obj->udata);
7607         kfree(obj);
7608 }
7609
7610 static int nf_tables_delobj(struct sk_buff *skb, const struct nfnl_info *info,
7611                             const struct nlattr * const nla[])
7612 {
7613         struct netlink_ext_ack *extack = info->extack;
7614         u8 genmask = nft_genmask_next(info->net);
7615         u8 family = info->nfmsg->nfgen_family;
7616         struct net *net = info->net;
7617         const struct nlattr *attr;
7618         struct nft_table *table;
7619         struct nft_object *obj;
7620         struct nft_ctx ctx;
7621         u32 objtype;
7622
7623         if (!nla[NFTA_OBJ_TYPE] ||
7624             (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
7625                 return -EINVAL;
7626
7627         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
7628                                  NETLINK_CB(skb).portid);
7629         if (IS_ERR(table)) {
7630                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
7631                 return PTR_ERR(table);
7632         }
7633
7634         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7635         if (nla[NFTA_OBJ_HANDLE]) {
7636                 attr = nla[NFTA_OBJ_HANDLE];
7637                 obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
7638         } else {
7639                 attr = nla[NFTA_OBJ_NAME];
7640                 obj = nft_obj_lookup(net, table, attr, objtype, genmask);
7641         }
7642
7643         if (IS_ERR(obj)) {
7644                 NL_SET_BAD_ATTR(extack, attr);
7645                 return PTR_ERR(obj);
7646         }
7647         if (obj->use > 0) {
7648                 NL_SET_BAD_ATTR(extack, attr);
7649                 return -EBUSY;
7650         }
7651
7652         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7653
7654         return nft_delobj(&ctx, obj);
7655 }
7656
7657 static void
7658 __nft_obj_notify(struct net *net, const struct nft_table *table,
7659                  struct nft_object *obj, u32 portid, u32 seq, int event,
7660                  u16 flags, int family, int report, gfp_t gfp)
7661 {
7662         struct nftables_pernet *nft_net = nft_pernet(net);
7663         struct sk_buff *skb;
7664         int err;
7665
7666         if (!report &&
7667             !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
7668                 return;
7669
7670         skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
7671         if (skb == NULL)
7672                 goto err;
7673
7674         err = nf_tables_fill_obj_info(skb, net, portid, seq, event,
7675                                       flags & (NLM_F_CREATE | NLM_F_EXCL),
7676                                       family, table, obj, false);
7677         if (err < 0) {
7678                 kfree_skb(skb);
7679                 goto err;
7680         }
7681
7682         nft_notify_enqueue(skb, report, &nft_net->notify_list);
7683         return;
7684 err:
7685         nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
7686 }
7687
7688 void nft_obj_notify(struct net *net, const struct nft_table *table,
7689                     struct nft_object *obj, u32 portid, u32 seq, int event,
7690                     u16 flags, int family, int report, gfp_t gfp)
7691 {
7692         struct nftables_pernet *nft_net = nft_pernet(net);
7693         char *buf = kasprintf(gfp, "%s:%u",
7694                               table->name, nft_net->base_seq);
7695
7696         audit_log_nfcfg(buf,
7697                         family,
7698                         obj->handle,
7699                         event == NFT_MSG_NEWOBJ ?
7700                                  AUDIT_NFT_OP_OBJ_REGISTER :
7701                                  AUDIT_NFT_OP_OBJ_UNREGISTER,
7702                         gfp);
7703         kfree(buf);
7704
7705         __nft_obj_notify(net, table, obj, portid, seq, event,
7706                          flags, family, report, gfp);
7707 }
7708 EXPORT_SYMBOL_GPL(nft_obj_notify);
7709
7710 static void nf_tables_obj_notify(const struct nft_ctx *ctx,
7711                                  struct nft_object *obj, int event)
7712 {
7713         __nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid,
7714                          ctx->seq, event, ctx->flags, ctx->family,
7715                          ctx->report, GFP_KERNEL);
7716 }
7717
7718 /*
7719  * Flow tables
7720  */
7721 void nft_register_flowtable_type(struct nf_flowtable_type *type)
7722 {
7723         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7724         list_add_tail_rcu(&type->list, &nf_tables_flowtables);
7725         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7726 }
7727 EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
7728
7729 void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
7730 {
7731         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7732         list_del_rcu(&type->list);
7733         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7734 }
7735 EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
7736
7737 static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
7738         [NFTA_FLOWTABLE_TABLE]          = { .type = NLA_STRING,
7739                                             .len = NFT_NAME_MAXLEN - 1 },
7740         [NFTA_FLOWTABLE_NAME]           = { .type = NLA_STRING,
7741                                             .len = NFT_NAME_MAXLEN - 1 },
7742         [NFTA_FLOWTABLE_HOOK]           = { .type = NLA_NESTED },
7743         [NFTA_FLOWTABLE_HANDLE]         = { .type = NLA_U64 },
7744         [NFTA_FLOWTABLE_FLAGS]          = { .type = NLA_U32 },
7745 };
7746
7747 struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
7748                                            const struct nlattr *nla, u8 genmask)
7749 {
7750         struct nft_flowtable *flowtable;
7751
7752         list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7753                 if (!nla_strcmp(nla, flowtable->name) &&
7754                     nft_active_genmask(flowtable, genmask))
7755                         return flowtable;
7756         }
7757         return ERR_PTR(-ENOENT);
7758 }
7759 EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
7760
7761 void nf_tables_deactivate_flowtable(const struct nft_ctx *ctx,
7762                                     struct nft_flowtable *flowtable,
7763                                     enum nft_trans_phase phase)
7764 {
7765         switch (phase) {
7766         case NFT_TRANS_PREPARE_ERROR:
7767         case NFT_TRANS_PREPARE:
7768         case NFT_TRANS_ABORT:
7769         case NFT_TRANS_RELEASE:
7770                 nft_use_dec(&flowtable->use);
7771                 fallthrough;
7772         default:
7773                 return;
7774         }
7775 }
7776 EXPORT_SYMBOL_GPL(nf_tables_deactivate_flowtable);
7777
7778 static struct nft_flowtable *
7779 nft_flowtable_lookup_byhandle(const struct nft_table *table,
7780                               const struct nlattr *nla, u8 genmask)
7781 {
7782        struct nft_flowtable *flowtable;
7783
7784        list_for_each_entry(flowtable, &table->flowtables, list) {
7785                if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
7786                    nft_active_genmask(flowtable, genmask))
7787                        return flowtable;
7788        }
7789        return ERR_PTR(-ENOENT);
7790 }
7791
7792 struct nft_flowtable_hook {
7793         u32                     num;
7794         int                     priority;
7795         struct list_head        list;
7796 };
7797
7798 static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
7799         [NFTA_FLOWTABLE_HOOK_NUM]       = { .type = NLA_U32 },
7800         [NFTA_FLOWTABLE_HOOK_PRIORITY]  = { .type = NLA_U32 },
7801         [NFTA_FLOWTABLE_HOOK_DEVS]      = { .type = NLA_NESTED },
7802 };
7803
7804 static int nft_flowtable_parse_hook(const struct nft_ctx *ctx,
7805                                     const struct nlattr *attr,
7806                                     struct nft_flowtable_hook *flowtable_hook,
7807                                     struct nft_flowtable *flowtable, bool add)
7808 {
7809         struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
7810         struct nft_hook *hook;
7811         int hooknum, priority;
7812         int err;
7813
7814         INIT_LIST_HEAD(&flowtable_hook->list);
7815
7816         err = nla_parse_nested_deprecated(tb, NFTA_FLOWTABLE_HOOK_MAX, attr,
7817                                           nft_flowtable_hook_policy, NULL);
7818         if (err < 0)
7819                 return err;
7820
7821         if (add) {
7822                 if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
7823                     !tb[NFTA_FLOWTABLE_HOOK_PRIORITY])
7824                         return -EINVAL;
7825
7826                 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7827                 if (hooknum != NF_NETDEV_INGRESS)
7828                         return -EOPNOTSUPP;
7829
7830                 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7831
7832                 flowtable_hook->priority        = priority;
7833                 flowtable_hook->num             = hooknum;
7834         } else {
7835                 if (tb[NFTA_FLOWTABLE_HOOK_NUM]) {
7836                         hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7837                         if (hooknum != flowtable->hooknum)
7838                                 return -EOPNOTSUPP;
7839                 }
7840
7841                 if (tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
7842                         priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7843                         if (priority != flowtable->data.priority)
7844                                 return -EOPNOTSUPP;
7845                 }
7846
7847                 flowtable_hook->priority        = flowtable->data.priority;
7848                 flowtable_hook->num             = flowtable->hooknum;
7849         }
7850
7851         if (tb[NFTA_FLOWTABLE_HOOK_DEVS]) {
7852                 err = nf_tables_parse_netdev_hooks(ctx->net,
7853                                                    tb[NFTA_FLOWTABLE_HOOK_DEVS],
7854                                                    &flowtable_hook->list);
7855                 if (err < 0)
7856                         return err;
7857         }
7858
7859         list_for_each_entry(hook, &flowtable_hook->list, list) {
7860                 hook->ops.pf            = NFPROTO_NETDEV;
7861                 hook->ops.hooknum       = flowtable_hook->num;
7862                 hook->ops.priority      = flowtable_hook->priority;
7863                 hook->ops.priv          = &flowtable->data;
7864                 hook->ops.hook          = flowtable->data.type->hook;
7865         }
7866
7867         return err;
7868 }
7869
7870 /* call under rcu_read_lock */
7871 static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
7872 {
7873         const struct nf_flowtable_type *type;
7874
7875         list_for_each_entry_rcu(type, &nf_tables_flowtables, list) {
7876                 if (family == type->family)
7877                         return type;
7878         }
7879         return NULL;
7880 }
7881
7882 static const struct nf_flowtable_type *
7883 nft_flowtable_type_get(struct net *net, u8 family)
7884 {
7885         const struct nf_flowtable_type *type;
7886
7887         rcu_read_lock();
7888         type = __nft_flowtable_type_get(family);
7889         if (type != NULL && try_module_get(type->owner)) {
7890                 rcu_read_unlock();
7891                 return type;
7892         }
7893         rcu_read_unlock();
7894
7895         lockdep_nfnl_nft_mutex_not_held();
7896 #ifdef CONFIG_MODULES
7897         if (type == NULL) {
7898                 if (nft_request_module(net, "nf-flowtable-%u", family) == -EAGAIN)
7899                         return ERR_PTR(-EAGAIN);
7900         }
7901 #endif
7902         return ERR_PTR(-ENOENT);
7903 }
7904
7905 /* Only called from error and netdev event paths. */
7906 static void nft_unregister_flowtable_hook(struct net *net,
7907                                           struct nft_flowtable *flowtable,
7908                                           struct nft_hook *hook)
7909 {
7910         nf_unregister_net_hook(net, &hook->ops);
7911         flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7912                                     FLOW_BLOCK_UNBIND);
7913 }
7914
7915 static void __nft_unregister_flowtable_net_hooks(struct net *net,
7916                                                  struct list_head *hook_list,
7917                                                  bool release_netdev)
7918 {
7919         struct nft_hook *hook, *next;
7920
7921         list_for_each_entry_safe(hook, next, hook_list, list) {
7922                 nf_unregister_net_hook(net, &hook->ops);
7923                 if (release_netdev) {
7924                         list_del(&hook->list);
7925                         kfree_rcu(hook, rcu);
7926                 }
7927         }
7928 }
7929
7930 static void nft_unregister_flowtable_net_hooks(struct net *net,
7931                                                struct list_head *hook_list)
7932 {
7933         __nft_unregister_flowtable_net_hooks(net, hook_list, false);
7934 }
7935
7936 static int nft_register_flowtable_net_hooks(struct net *net,
7937                                             struct nft_table *table,
7938                                             struct list_head *hook_list,
7939                                             struct nft_flowtable *flowtable)
7940 {
7941         struct nft_hook *hook, *hook2, *next;
7942         struct nft_flowtable *ft;
7943         int err, i = 0;
7944
7945         list_for_each_entry(hook, hook_list, list) {
7946                 list_for_each_entry(ft, &table->flowtables, list) {
7947                         if (!nft_is_active_next(net, ft))
7948                                 continue;
7949
7950                         list_for_each_entry(hook2, &ft->hook_list, list) {
7951                                 if (hook->ops.dev == hook2->ops.dev &&
7952                                     hook->ops.pf == hook2->ops.pf) {
7953                                         err = -EEXIST;
7954                                         goto err_unregister_net_hooks;
7955                                 }
7956                         }
7957                 }
7958
7959                 err = flowtable->data.type->setup(&flowtable->data,
7960                                                   hook->ops.dev,
7961                                                   FLOW_BLOCK_BIND);
7962                 if (err < 0)
7963                         goto err_unregister_net_hooks;
7964
7965                 err = nf_register_net_hook(net, &hook->ops);
7966                 if (err < 0) {
7967                         flowtable->data.type->setup(&flowtable->data,
7968                                                     hook->ops.dev,
7969                                                     FLOW_BLOCK_UNBIND);
7970                         goto err_unregister_net_hooks;
7971                 }
7972
7973                 i++;
7974         }
7975
7976         return 0;
7977
7978 err_unregister_net_hooks:
7979         list_for_each_entry_safe(hook, next, hook_list, list) {
7980                 if (i-- <= 0)
7981                         break;
7982
7983                 nft_unregister_flowtable_hook(net, flowtable, hook);
7984                 list_del_rcu(&hook->list);
7985                 kfree_rcu(hook, rcu);
7986         }
7987
7988         return err;
7989 }
7990
7991 static void nft_hooks_destroy(struct list_head *hook_list)
7992 {
7993         struct nft_hook *hook, *next;
7994
7995         list_for_each_entry_safe(hook, next, hook_list, list) {
7996                 list_del_rcu(&hook->list);
7997                 kfree_rcu(hook, rcu);
7998         }
7999 }
8000
8001 static int nft_flowtable_update(struct nft_ctx *ctx, const struct nlmsghdr *nlh,
8002                                 struct nft_flowtable *flowtable)
8003 {
8004         const struct nlattr * const *nla = ctx->nla;
8005         struct nft_flowtable_hook flowtable_hook;
8006         struct nft_hook *hook, *next;
8007         struct nft_trans *trans;
8008         bool unregister = false;
8009         u32 flags;
8010         int err;
8011
8012         err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
8013                                        &flowtable_hook, flowtable, false);
8014         if (err < 0)
8015                 return err;
8016
8017         list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
8018                 if (nft_hook_list_find(&flowtable->hook_list, hook)) {
8019                         list_del(&hook->list);
8020                         kfree(hook);
8021                 }
8022         }
8023
8024         if (nla[NFTA_FLOWTABLE_FLAGS]) {
8025                 flags = ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
8026                 if (flags & ~NFT_FLOWTABLE_MASK) {
8027                         err = -EOPNOTSUPP;
8028                         goto err_flowtable_update_hook;
8029                 }
8030                 if ((flowtable->data.flags & NFT_FLOWTABLE_HW_OFFLOAD) ^
8031                     (flags & NFT_FLOWTABLE_HW_OFFLOAD)) {
8032                         err = -EOPNOTSUPP;
8033                         goto err_flowtable_update_hook;
8034                 }
8035         } else {
8036                 flags = flowtable->data.flags;
8037         }
8038
8039         err = nft_register_flowtable_net_hooks(ctx->net, ctx->table,
8040                                                &flowtable_hook.list, flowtable);
8041         if (err < 0)
8042                 goto err_flowtable_update_hook;
8043
8044         trans = nft_trans_alloc(ctx, NFT_MSG_NEWFLOWTABLE,
8045                                 sizeof(struct nft_trans_flowtable));
8046         if (!trans) {
8047                 unregister = true;
8048                 err = -ENOMEM;
8049                 goto err_flowtable_update_hook;
8050         }
8051
8052         nft_trans_flowtable_flags(trans) = flags;
8053         nft_trans_flowtable(trans) = flowtable;
8054         nft_trans_flowtable_update(trans) = true;
8055         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
8056         list_splice(&flowtable_hook.list, &nft_trans_flowtable_hooks(trans));
8057
8058         nft_trans_commit_list_add_tail(ctx->net, trans);
8059
8060         return 0;
8061
8062 err_flowtable_update_hook:
8063         list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
8064                 if (unregister)
8065                         nft_unregister_flowtable_hook(ctx->net, flowtable, hook);
8066                 list_del_rcu(&hook->list);
8067                 kfree_rcu(hook, rcu);
8068         }
8069
8070         return err;
8071
8072 }
8073
8074 static int nf_tables_newflowtable(struct sk_buff *skb,
8075                                   const struct nfnl_info *info,
8076                                   const struct nlattr * const nla[])
8077 {
8078         struct netlink_ext_ack *extack = info->extack;
8079         struct nft_flowtable_hook flowtable_hook;
8080         u8 genmask = nft_genmask_next(info->net);
8081         u8 family = info->nfmsg->nfgen_family;
8082         const struct nf_flowtable_type *type;
8083         struct nft_flowtable *flowtable;
8084         struct net *net = info->net;
8085         struct nft_table *table;
8086         struct nft_trans *trans;
8087         struct nft_ctx ctx;
8088         int err;
8089
8090         if (!nla[NFTA_FLOWTABLE_TABLE] ||
8091             !nla[NFTA_FLOWTABLE_NAME] ||
8092             !nla[NFTA_FLOWTABLE_HOOK])
8093                 return -EINVAL;
8094
8095         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
8096                                  genmask, NETLINK_CB(skb).portid);
8097         if (IS_ERR(table)) {
8098                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
8099                 return PTR_ERR(table);
8100         }
8101
8102         flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
8103                                          genmask);
8104         if (IS_ERR(flowtable)) {
8105                 err = PTR_ERR(flowtable);
8106                 if (err != -ENOENT) {
8107                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
8108                         return err;
8109                 }
8110         } else {
8111                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
8112                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
8113                         return -EEXIST;
8114                 }
8115
8116                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
8117
8118                 return nft_flowtable_update(&ctx, info->nlh, flowtable);
8119         }
8120
8121         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
8122
8123         if (!nft_use_inc(&table->use))
8124                 return -EMFILE;
8125
8126         flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL_ACCOUNT);
8127         if (!flowtable) {
8128                 err = -ENOMEM;
8129                 goto flowtable_alloc;
8130         }
8131
8132         flowtable->table = table;
8133         flowtable->handle = nf_tables_alloc_handle(table);
8134         INIT_LIST_HEAD(&flowtable->hook_list);
8135
8136         flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL_ACCOUNT);
8137         if (!flowtable->name) {
8138                 err = -ENOMEM;
8139                 goto err1;
8140         }
8141
8142         type = nft_flowtable_type_get(net, family);
8143         if (IS_ERR(type)) {
8144                 err = PTR_ERR(type);
8145                 goto err2;
8146         }
8147
8148         if (nla[NFTA_FLOWTABLE_FLAGS]) {
8149                 flowtable->data.flags =
8150                         ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
8151                 if (flowtable->data.flags & ~NFT_FLOWTABLE_MASK) {
8152                         err = -EOPNOTSUPP;
8153                         goto err3;
8154                 }
8155         }
8156
8157         write_pnet(&flowtable->data.net, net);
8158         flowtable->data.type = type;
8159         err = type->init(&flowtable->data);
8160         if (err < 0)
8161                 goto err3;
8162
8163         err = nft_flowtable_parse_hook(&ctx, nla[NFTA_FLOWTABLE_HOOK],
8164                                        &flowtable_hook, flowtable, true);
8165         if (err < 0)
8166                 goto err_flowtable_parse_hooks;
8167
8168         list_splice(&flowtable_hook.list, &flowtable->hook_list);
8169         flowtable->data.priority = flowtable_hook.priority;
8170         flowtable->hooknum = flowtable_hook.num;
8171
8172         trans = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
8173         if (IS_ERR(trans)) {
8174                 err = PTR_ERR(trans);
8175                 goto err_flowtable_trans;
8176         }
8177
8178         /* This must be LAST to ensure no packets are walking over this flowtable. */
8179         err = nft_register_flowtable_net_hooks(ctx.net, table,
8180                                                &flowtable->hook_list,
8181                                                flowtable);
8182         if (err < 0)
8183                 goto err_flowtable_hooks;
8184
8185         list_add_tail_rcu(&flowtable->list, &table->flowtables);
8186
8187         return 0;
8188
8189 err_flowtable_hooks:
8190         nft_trans_destroy(trans);
8191 err_flowtable_trans:
8192         nft_hooks_destroy(&flowtable->hook_list);
8193 err_flowtable_parse_hooks:
8194         flowtable->data.type->free(&flowtable->data);
8195 err3:
8196         module_put(type->owner);
8197 err2:
8198         kfree(flowtable->name);
8199 err1:
8200         kfree(flowtable);
8201 flowtable_alloc:
8202         nft_use_dec_restore(&table->use);
8203
8204         return err;
8205 }
8206
8207 static void nft_flowtable_hook_release(struct nft_flowtable_hook *flowtable_hook)
8208 {
8209         struct nft_hook *this, *next;
8210
8211         list_for_each_entry_safe(this, next, &flowtable_hook->list, list) {
8212                 list_del(&this->list);
8213                 kfree(this);
8214         }
8215 }
8216
8217 static int nft_delflowtable_hook(struct nft_ctx *ctx,
8218                                  struct nft_flowtable *flowtable)
8219 {
8220         const struct nlattr * const *nla = ctx->nla;
8221         struct nft_flowtable_hook flowtable_hook;
8222         LIST_HEAD(flowtable_del_list);
8223         struct nft_hook *this, *hook;
8224         struct nft_trans *trans;
8225         int err;
8226
8227         err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
8228                                        &flowtable_hook, flowtable, false);
8229         if (err < 0)
8230                 return err;
8231
8232         list_for_each_entry(this, &flowtable_hook.list, list) {
8233                 hook = nft_hook_list_find(&flowtable->hook_list, this);
8234                 if (!hook) {
8235                         err = -ENOENT;
8236                         goto err_flowtable_del_hook;
8237                 }
8238                 list_move(&hook->list, &flowtable_del_list);
8239         }
8240
8241         trans = nft_trans_alloc(ctx, NFT_MSG_DELFLOWTABLE,
8242                                 sizeof(struct nft_trans_flowtable));
8243         if (!trans) {
8244                 err = -ENOMEM;
8245                 goto err_flowtable_del_hook;
8246         }
8247
8248         nft_trans_flowtable(trans) = flowtable;
8249         nft_trans_flowtable_update(trans) = true;
8250         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
8251         list_splice(&flowtable_del_list, &nft_trans_flowtable_hooks(trans));
8252         nft_flowtable_hook_release(&flowtable_hook);
8253
8254         nft_trans_commit_list_add_tail(ctx->net, trans);
8255
8256         return 0;
8257
8258 err_flowtable_del_hook:
8259         list_splice(&flowtable_del_list, &flowtable->hook_list);
8260         nft_flowtable_hook_release(&flowtable_hook);
8261
8262         return err;
8263 }
8264
8265 static int nf_tables_delflowtable(struct sk_buff *skb,
8266                                   const struct nfnl_info *info,
8267                                   const struct nlattr * const nla[])
8268 {
8269         struct netlink_ext_ack *extack = info->extack;
8270         u8 genmask = nft_genmask_next(info->net);
8271         u8 family = info->nfmsg->nfgen_family;
8272         struct nft_flowtable *flowtable;
8273         struct net *net = info->net;
8274         const struct nlattr *attr;
8275         struct nft_table *table;
8276         struct nft_ctx ctx;
8277
8278         if (!nla[NFTA_FLOWTABLE_TABLE] ||
8279             (!nla[NFTA_FLOWTABLE_NAME] &&
8280              !nla[NFTA_FLOWTABLE_HANDLE]))
8281                 return -EINVAL;
8282
8283         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
8284                                  genmask, NETLINK_CB(skb).portid);
8285         if (IS_ERR(table)) {
8286                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
8287                 return PTR_ERR(table);
8288         }
8289
8290         if (nla[NFTA_FLOWTABLE_HANDLE]) {
8291                 attr = nla[NFTA_FLOWTABLE_HANDLE];
8292                 flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
8293         } else {
8294                 attr = nla[NFTA_FLOWTABLE_NAME];
8295                 flowtable = nft_flowtable_lookup(table, attr, genmask);
8296         }
8297
8298         if (IS_ERR(flowtable)) {
8299                 NL_SET_BAD_ATTR(extack, attr);
8300                 return PTR_ERR(flowtable);
8301         }
8302
8303         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
8304
8305         if (nla[NFTA_FLOWTABLE_HOOK])
8306                 return nft_delflowtable_hook(&ctx, flowtable);
8307
8308         if (flowtable->use > 0) {
8309                 NL_SET_BAD_ATTR(extack, attr);
8310                 return -EBUSY;
8311         }
8312
8313         return nft_delflowtable(&ctx, flowtable);
8314 }
8315
8316 static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
8317                                          u32 portid, u32 seq, int event,
8318                                          u32 flags, int family,
8319                                          struct nft_flowtable *flowtable,
8320                                          struct list_head *hook_list)
8321 {
8322         struct nlattr *nest, *nest_devs;
8323         struct nft_hook *hook;
8324         struct nlmsghdr *nlh;
8325
8326         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
8327         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
8328                            NFNETLINK_V0, nft_base_seq(net));
8329         if (!nlh)
8330                 goto nla_put_failure;
8331
8332         if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
8333             nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
8334             nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
8335             nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
8336                          NFTA_FLOWTABLE_PAD) ||
8337             nla_put_be32(skb, NFTA_FLOWTABLE_FLAGS, htonl(flowtable->data.flags)))
8338                 goto nla_put_failure;
8339
8340         nest = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK);
8341         if (!nest)
8342                 goto nla_put_failure;
8343         if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
8344             nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->data.priority)))
8345                 goto nla_put_failure;
8346
8347         nest_devs = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK_DEVS);
8348         if (!nest_devs)
8349                 goto nla_put_failure;
8350
8351         list_for_each_entry_rcu(hook, hook_list, list) {
8352                 if (nla_put_string(skb, NFTA_DEVICE_NAME, hook->ops.dev->name))
8353                         goto nla_put_failure;
8354         }
8355         nla_nest_end(skb, nest_devs);
8356         nla_nest_end(skb, nest);
8357
8358         nlmsg_end(skb, nlh);
8359         return 0;
8360
8361 nla_put_failure:
8362         nlmsg_trim(skb, nlh);
8363         return -1;
8364 }
8365
8366 struct nft_flowtable_filter {
8367         char            *table;
8368 };
8369
8370 static int nf_tables_dump_flowtable(struct sk_buff *skb,
8371                                     struct netlink_callback *cb)
8372 {
8373         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
8374         struct nft_flowtable_filter *filter = cb->data;
8375         unsigned int idx = 0, s_idx = cb->args[0];
8376         struct net *net = sock_net(skb->sk);
8377         int family = nfmsg->nfgen_family;
8378         struct nft_flowtable *flowtable;
8379         struct nftables_pernet *nft_net;
8380         const struct nft_table *table;
8381
8382         rcu_read_lock();
8383         nft_net = nft_pernet(net);
8384         cb->seq = READ_ONCE(nft_net->base_seq);
8385
8386         list_for_each_entry_rcu(table, &nft_net->tables, list) {
8387                 if (family != NFPROTO_UNSPEC && family != table->family)
8388                         continue;
8389
8390                 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
8391                         if (!nft_is_active(net, flowtable))
8392                                 goto cont;
8393                         if (idx < s_idx)
8394                                 goto cont;
8395                         if (idx > s_idx)
8396                                 memset(&cb->args[1], 0,
8397                                        sizeof(cb->args) - sizeof(cb->args[0]));
8398                         if (filter && filter->table &&
8399                             strcmp(filter->table, table->name))
8400                                 goto cont;
8401
8402                         if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
8403                                                           cb->nlh->nlmsg_seq,
8404                                                           NFT_MSG_NEWFLOWTABLE,
8405                                                           NLM_F_MULTI | NLM_F_APPEND,
8406                                                           table->family,
8407                                                           flowtable,
8408                                                           &flowtable->hook_list) < 0)
8409                                 goto done;
8410
8411                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
8412 cont:
8413                         idx++;
8414                 }
8415         }
8416 done:
8417         rcu_read_unlock();
8418
8419         cb->args[0] = idx;
8420         return skb->len;
8421 }
8422
8423 static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
8424 {
8425         const struct nlattr * const *nla = cb->data;
8426         struct nft_flowtable_filter *filter = NULL;
8427
8428         if (nla[NFTA_FLOWTABLE_TABLE]) {
8429                 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
8430                 if (!filter)
8431                         return -ENOMEM;
8432
8433                 filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
8434                                            GFP_ATOMIC);
8435                 if (!filter->table) {
8436                         kfree(filter);
8437                         return -ENOMEM;
8438                 }
8439         }
8440
8441         cb->data = filter;
8442         return 0;
8443 }
8444
8445 static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
8446 {
8447         struct nft_flowtable_filter *filter = cb->data;
8448
8449         if (!filter)
8450                 return 0;
8451
8452         kfree(filter->table);
8453         kfree(filter);
8454
8455         return 0;
8456 }
8457
8458 /* called with rcu_read_lock held */
8459 static int nf_tables_getflowtable(struct sk_buff *skb,
8460                                   const struct nfnl_info *info,
8461                                   const struct nlattr * const nla[])
8462 {
8463         u8 genmask = nft_genmask_cur(info->net);
8464         u8 family = info->nfmsg->nfgen_family;
8465         struct nft_flowtable *flowtable;
8466         const struct nft_table *table;
8467         struct net *net = info->net;
8468         struct sk_buff *skb2;
8469         int err;
8470
8471         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
8472                 struct netlink_dump_control c = {
8473                         .start = nf_tables_dump_flowtable_start,
8474                         .dump = nf_tables_dump_flowtable,
8475                         .done = nf_tables_dump_flowtable_done,
8476                         .module = THIS_MODULE,
8477                         .data = (void *)nla,
8478                 };
8479
8480                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
8481         }
8482
8483         if (!nla[NFTA_FLOWTABLE_NAME])
8484                 return -EINVAL;
8485
8486         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
8487                                  genmask, 0);
8488         if (IS_ERR(table))
8489                 return PTR_ERR(table);
8490
8491         flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
8492                                          genmask);
8493         if (IS_ERR(flowtable))
8494                 return PTR_ERR(flowtable);
8495
8496         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
8497         if (!skb2)
8498                 return -ENOMEM;
8499
8500         err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
8501                                             info->nlh->nlmsg_seq,
8502                                             NFT_MSG_NEWFLOWTABLE, 0, family,
8503                                             flowtable, &flowtable->hook_list);
8504         if (err < 0)
8505                 goto err_fill_flowtable_info;
8506
8507         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
8508
8509 err_fill_flowtable_info:
8510         kfree_skb(skb2);
8511         return err;
8512 }
8513
8514 static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
8515                                        struct nft_flowtable *flowtable,
8516                                        struct list_head *hook_list,
8517                                        int event)
8518 {
8519         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
8520         struct sk_buff *skb;
8521         u16 flags = 0;
8522         int err;
8523
8524         if (!ctx->report &&
8525             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
8526                 return;
8527
8528         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
8529         if (skb == NULL)
8530                 goto err;
8531
8532         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
8533                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
8534
8535         err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
8536                                             ctx->seq, event, flags,
8537                                             ctx->family, flowtable, hook_list);
8538         if (err < 0) {
8539                 kfree_skb(skb);
8540                 goto err;
8541         }
8542
8543         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
8544         return;
8545 err:
8546         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
8547 }
8548
8549 static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
8550 {
8551         struct nft_hook *hook, *next;
8552
8553         flowtable->data.type->free(&flowtable->data);
8554         list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
8555                 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
8556                                             FLOW_BLOCK_UNBIND);
8557                 list_del_rcu(&hook->list);
8558                 kfree(hook);
8559         }
8560         kfree(flowtable->name);
8561         module_put(flowtable->data.type->owner);
8562         kfree(flowtable);
8563 }
8564
8565 static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
8566                                    u32 portid, u32 seq)
8567 {
8568         struct nftables_pernet *nft_net = nft_pernet(net);
8569         struct nlmsghdr *nlh;
8570         char buf[TASK_COMM_LEN];
8571         int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
8572
8573         nlh = nfnl_msg_put(skb, portid, seq, event, 0, AF_UNSPEC,
8574                            NFNETLINK_V0, nft_base_seq(net));
8575         if (!nlh)
8576                 goto nla_put_failure;
8577
8578         if (nla_put_be32(skb, NFTA_GEN_ID, htonl(nft_net->base_seq)) ||
8579             nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
8580             nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
8581                 goto nla_put_failure;
8582
8583         nlmsg_end(skb, nlh);
8584         return 0;
8585
8586 nla_put_failure:
8587         nlmsg_trim(skb, nlh);
8588         return -EMSGSIZE;
8589 }
8590
8591 static void nft_flowtable_event(unsigned long event, struct net_device *dev,
8592                                 struct nft_flowtable *flowtable)
8593 {
8594         struct nft_hook *hook;
8595
8596         list_for_each_entry(hook, &flowtable->hook_list, list) {
8597                 if (hook->ops.dev != dev)
8598                         continue;
8599
8600                 /* flow_offload_netdev_event() cleans up entries for us. */
8601                 nft_unregister_flowtable_hook(dev_net(dev), flowtable, hook);
8602                 list_del_rcu(&hook->list);
8603                 kfree_rcu(hook, rcu);
8604                 break;
8605         }
8606 }
8607
8608 static int nf_tables_flowtable_event(struct notifier_block *this,
8609                                      unsigned long event, void *ptr)
8610 {
8611         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
8612         struct nft_flowtable *flowtable;
8613         struct nftables_pernet *nft_net;
8614         struct nft_table *table;
8615         struct net *net;
8616
8617         if (event != NETDEV_UNREGISTER)
8618                 return 0;
8619
8620         net = dev_net(dev);
8621         nft_net = nft_pernet(net);
8622         mutex_lock(&nft_net->commit_mutex);
8623         list_for_each_entry(table, &nft_net->tables, list) {
8624                 list_for_each_entry(flowtable, &table->flowtables, list) {
8625                         nft_flowtable_event(event, dev, flowtable);
8626                 }
8627         }
8628         mutex_unlock(&nft_net->commit_mutex);
8629
8630         return NOTIFY_DONE;
8631 }
8632
8633 static struct notifier_block nf_tables_flowtable_notifier = {
8634         .notifier_call  = nf_tables_flowtable_event,
8635 };
8636
8637 static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
8638                                  int event)
8639 {
8640         struct nlmsghdr *nlh = nlmsg_hdr(skb);
8641         struct sk_buff *skb2;
8642         int err;
8643
8644         if (!nlmsg_report(nlh) &&
8645             !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
8646                 return;
8647
8648         skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
8649         if (skb2 == NULL)
8650                 goto err;
8651
8652         err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
8653                                       nlh->nlmsg_seq);
8654         if (err < 0) {
8655                 kfree_skb(skb2);
8656                 goto err;
8657         }
8658
8659         nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
8660                        nlmsg_report(nlh), GFP_KERNEL);
8661         return;
8662 err:
8663         nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
8664                           -ENOBUFS);
8665 }
8666
8667 static int nf_tables_getgen(struct sk_buff *skb, const struct nfnl_info *info,
8668                             const struct nlattr * const nla[])
8669 {
8670         struct sk_buff *skb2;
8671         int err;
8672
8673         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
8674         if (skb2 == NULL)
8675                 return -ENOMEM;
8676
8677         err = nf_tables_fill_gen_info(skb2, info->net, NETLINK_CB(skb).portid,
8678                                       info->nlh->nlmsg_seq);
8679         if (err < 0)
8680                 goto err_fill_gen_info;
8681
8682         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
8683
8684 err_fill_gen_info:
8685         kfree_skb(skb2);
8686         return err;
8687 }
8688
8689 static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
8690         [NFT_MSG_NEWTABLE] = {
8691                 .call           = nf_tables_newtable,
8692                 .type           = NFNL_CB_BATCH,
8693                 .attr_count     = NFTA_TABLE_MAX,
8694                 .policy         = nft_table_policy,
8695         },
8696         [NFT_MSG_GETTABLE] = {
8697                 .call           = nf_tables_gettable,
8698                 .type           = NFNL_CB_RCU,
8699                 .attr_count     = NFTA_TABLE_MAX,
8700                 .policy         = nft_table_policy,
8701         },
8702         [NFT_MSG_DELTABLE] = {
8703                 .call           = nf_tables_deltable,
8704                 .type           = NFNL_CB_BATCH,
8705                 .attr_count     = NFTA_TABLE_MAX,
8706                 .policy         = nft_table_policy,
8707         },
8708         [NFT_MSG_NEWCHAIN] = {
8709                 .call           = nf_tables_newchain,
8710                 .type           = NFNL_CB_BATCH,
8711                 .attr_count     = NFTA_CHAIN_MAX,
8712                 .policy         = nft_chain_policy,
8713         },
8714         [NFT_MSG_GETCHAIN] = {
8715                 .call           = nf_tables_getchain,
8716                 .type           = NFNL_CB_RCU,
8717                 .attr_count     = NFTA_CHAIN_MAX,
8718                 .policy         = nft_chain_policy,
8719         },
8720         [NFT_MSG_DELCHAIN] = {
8721                 .call           = nf_tables_delchain,
8722                 .type           = NFNL_CB_BATCH,
8723                 .attr_count     = NFTA_CHAIN_MAX,
8724                 .policy         = nft_chain_policy,
8725         },
8726         [NFT_MSG_NEWRULE] = {
8727                 .call           = nf_tables_newrule,
8728                 .type           = NFNL_CB_BATCH,
8729                 .attr_count     = NFTA_RULE_MAX,
8730                 .policy         = nft_rule_policy,
8731         },
8732         [NFT_MSG_GETRULE] = {
8733                 .call           = nf_tables_getrule,
8734                 .type           = NFNL_CB_RCU,
8735                 .attr_count     = NFTA_RULE_MAX,
8736                 .policy         = nft_rule_policy,
8737         },
8738         [NFT_MSG_DELRULE] = {
8739                 .call           = nf_tables_delrule,
8740                 .type           = NFNL_CB_BATCH,
8741                 .attr_count     = NFTA_RULE_MAX,
8742                 .policy         = nft_rule_policy,
8743         },
8744         [NFT_MSG_NEWSET] = {
8745                 .call           = nf_tables_newset,
8746                 .type           = NFNL_CB_BATCH,
8747                 .attr_count     = NFTA_SET_MAX,
8748                 .policy         = nft_set_policy,
8749         },
8750         [NFT_MSG_GETSET] = {
8751                 .call           = nf_tables_getset,
8752                 .type           = NFNL_CB_RCU,
8753                 .attr_count     = NFTA_SET_MAX,
8754                 .policy         = nft_set_policy,
8755         },
8756         [NFT_MSG_DELSET] = {
8757                 .call           = nf_tables_delset,
8758                 .type           = NFNL_CB_BATCH,
8759                 .attr_count     = NFTA_SET_MAX,
8760                 .policy         = nft_set_policy,
8761         },
8762         [NFT_MSG_NEWSETELEM] = {
8763                 .call           = nf_tables_newsetelem,
8764                 .type           = NFNL_CB_BATCH,
8765                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8766                 .policy         = nft_set_elem_list_policy,
8767         },
8768         [NFT_MSG_GETSETELEM] = {
8769                 .call           = nf_tables_getsetelem,
8770                 .type           = NFNL_CB_RCU,
8771                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8772                 .policy         = nft_set_elem_list_policy,
8773         },
8774         [NFT_MSG_DELSETELEM] = {
8775                 .call           = nf_tables_delsetelem,
8776                 .type           = NFNL_CB_BATCH,
8777                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8778                 .policy         = nft_set_elem_list_policy,
8779         },
8780         [NFT_MSG_GETGEN] = {
8781                 .call           = nf_tables_getgen,
8782                 .type           = NFNL_CB_RCU,
8783         },
8784         [NFT_MSG_NEWOBJ] = {
8785                 .call           = nf_tables_newobj,
8786                 .type           = NFNL_CB_BATCH,
8787                 .attr_count     = NFTA_OBJ_MAX,
8788                 .policy         = nft_obj_policy,
8789         },
8790         [NFT_MSG_GETOBJ] = {
8791                 .call           = nf_tables_getobj,
8792                 .type           = NFNL_CB_RCU,
8793                 .attr_count     = NFTA_OBJ_MAX,
8794                 .policy         = nft_obj_policy,
8795         },
8796         [NFT_MSG_DELOBJ] = {
8797                 .call           = nf_tables_delobj,
8798                 .type           = NFNL_CB_BATCH,
8799                 .attr_count     = NFTA_OBJ_MAX,
8800                 .policy         = nft_obj_policy,
8801         },
8802         [NFT_MSG_GETOBJ_RESET] = {
8803                 .call           = nf_tables_getobj,
8804                 .type           = NFNL_CB_RCU,
8805                 .attr_count     = NFTA_OBJ_MAX,
8806                 .policy         = nft_obj_policy,
8807         },
8808         [NFT_MSG_NEWFLOWTABLE] = {
8809                 .call           = nf_tables_newflowtable,
8810                 .type           = NFNL_CB_BATCH,
8811                 .attr_count     = NFTA_FLOWTABLE_MAX,
8812                 .policy         = nft_flowtable_policy,
8813         },
8814         [NFT_MSG_GETFLOWTABLE] = {
8815                 .call           = nf_tables_getflowtable,
8816                 .type           = NFNL_CB_RCU,
8817                 .attr_count     = NFTA_FLOWTABLE_MAX,
8818                 .policy         = nft_flowtable_policy,
8819         },
8820         [NFT_MSG_DELFLOWTABLE] = {
8821                 .call           = nf_tables_delflowtable,
8822                 .type           = NFNL_CB_BATCH,
8823                 .attr_count     = NFTA_FLOWTABLE_MAX,
8824                 .policy         = nft_flowtable_policy,
8825         },
8826 };
8827
8828 static int nf_tables_validate(struct net *net)
8829 {
8830         struct nftables_pernet *nft_net = nft_pernet(net);
8831         struct nft_table *table;
8832
8833         switch (nft_net->validate_state) {
8834         case NFT_VALIDATE_SKIP:
8835                 break;
8836         case NFT_VALIDATE_NEED:
8837                 nft_validate_state_update(net, NFT_VALIDATE_DO);
8838                 fallthrough;
8839         case NFT_VALIDATE_DO:
8840                 list_for_each_entry(table, &nft_net->tables, list) {
8841                         if (nft_table_validate(net, table) < 0)
8842                                 return -EAGAIN;
8843                 }
8844
8845                 nft_validate_state_update(net, NFT_VALIDATE_SKIP);
8846                 break;
8847         }
8848
8849         return 0;
8850 }
8851
8852 /* a drop policy has to be deferred until all rules have been activated,
8853  * otherwise a large ruleset that contains a drop-policy base chain will
8854  * cause all packets to get dropped until the full transaction has been
8855  * processed.
8856  *
8857  * We defer the drop policy until the transaction has been finalized.
8858  */
8859 static void nft_chain_commit_drop_policy(struct nft_trans *trans)
8860 {
8861         struct nft_base_chain *basechain;
8862
8863         if (nft_trans_chain_policy(trans) != NF_DROP)
8864                 return;
8865
8866         if (!nft_is_base_chain(trans->ctx.chain))
8867                 return;
8868
8869         basechain = nft_base_chain(trans->ctx.chain);
8870         basechain->policy = NF_DROP;
8871 }
8872
8873 static void nft_chain_commit_update(struct nft_trans *trans)
8874 {
8875         struct nft_base_chain *basechain;
8876
8877         if (nft_trans_chain_name(trans)) {
8878                 rhltable_remove(&trans->ctx.table->chains_ht,
8879                                 &trans->ctx.chain->rhlhead,
8880                                 nft_chain_ht_params);
8881                 swap(trans->ctx.chain->name, nft_trans_chain_name(trans));
8882                 rhltable_insert_key(&trans->ctx.table->chains_ht,
8883                                     trans->ctx.chain->name,
8884                                     &trans->ctx.chain->rhlhead,
8885                                     nft_chain_ht_params);
8886         }
8887
8888         if (!nft_is_base_chain(trans->ctx.chain))
8889                 return;
8890
8891         nft_chain_stats_replace(trans);
8892
8893         basechain = nft_base_chain(trans->ctx.chain);
8894
8895         switch (nft_trans_chain_policy(trans)) {
8896         case NF_DROP:
8897         case NF_ACCEPT:
8898                 basechain->policy = nft_trans_chain_policy(trans);
8899                 break;
8900         }
8901 }
8902
8903 static void nft_obj_commit_update(struct nft_trans *trans)
8904 {
8905         struct nft_object *newobj;
8906         struct nft_object *obj;
8907
8908         obj = nft_trans_obj(trans);
8909         newobj = nft_trans_obj_newobj(trans);
8910
8911         if (obj->ops->update)
8912                 obj->ops->update(obj, newobj);
8913
8914         nft_obj_destroy(&trans->ctx, newobj);
8915 }
8916
8917 static void nft_commit_release(struct nft_trans *trans)
8918 {
8919         switch (trans->msg_type) {
8920         case NFT_MSG_DELTABLE:
8921                 nf_tables_table_destroy(&trans->ctx);
8922                 break;
8923         case NFT_MSG_NEWCHAIN:
8924                 free_percpu(nft_trans_chain_stats(trans));
8925                 kfree(nft_trans_chain_name(trans));
8926                 break;
8927         case NFT_MSG_DELCHAIN:
8928                 nf_tables_chain_destroy(&trans->ctx);
8929                 break;
8930         case NFT_MSG_DELRULE:
8931                 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8932                 break;
8933         case NFT_MSG_DELSET:
8934                 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8935                 break;
8936         case NFT_MSG_DELSETELEM:
8937                 nf_tables_set_elem_destroy(&trans->ctx,
8938                                            nft_trans_elem_set(trans),
8939                                            nft_trans_elem(trans).priv);
8940                 break;
8941         case NFT_MSG_DELOBJ:
8942                 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8943                 break;
8944         case NFT_MSG_DELFLOWTABLE:
8945                 if (nft_trans_flowtable_update(trans))
8946                         nft_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8947                 else
8948                         nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8949                 break;
8950         }
8951
8952         if (trans->put_net)
8953                 put_net(trans->ctx.net);
8954
8955         kfree(trans);
8956 }
8957
8958 static void nf_tables_trans_destroy_work(struct work_struct *w)
8959 {
8960         struct nft_trans *trans, *next;
8961         LIST_HEAD(head);
8962
8963         spin_lock(&nf_tables_destroy_list_lock);
8964         list_splice_init(&nf_tables_destroy_list, &head);
8965         spin_unlock(&nf_tables_destroy_list_lock);
8966
8967         if (list_empty(&head))
8968                 return;
8969
8970         synchronize_rcu();
8971
8972         list_for_each_entry_safe(trans, next, &head, list) {
8973                 nft_trans_list_del(trans);
8974                 nft_commit_release(trans);
8975         }
8976 }
8977
8978 void nf_tables_trans_destroy_flush_work(void)
8979 {
8980         flush_work(&trans_destroy_work);
8981 }
8982 EXPORT_SYMBOL_GPL(nf_tables_trans_destroy_flush_work);
8983
8984 static bool nft_expr_reduce(struct nft_regs_track *track,
8985                             const struct nft_expr *expr)
8986 {
8987         return false;
8988 }
8989
8990 static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
8991 {
8992         const struct nft_expr *expr, *last;
8993         struct nft_regs_track track = {};
8994         unsigned int size, data_size;
8995         void *data, *data_boundary;
8996         struct nft_rule_dp *prule;
8997         struct nft_rule *rule;
8998
8999         /* already handled or inactive chain? */
9000         if (chain->blob_next || !nft_is_active_next(net, chain))
9001                 return 0;
9002
9003         data_size = 0;
9004         list_for_each_entry(rule, &chain->rules, list) {
9005                 if (nft_is_active_next(net, rule)) {
9006                         data_size += sizeof(*prule) + rule->dlen;
9007                         if (data_size > INT_MAX)
9008                                 return -ENOMEM;
9009                 }
9010         }
9011         data_size += offsetof(struct nft_rule_dp, data);        /* last rule */
9012
9013         chain->blob_next = nf_tables_chain_alloc_rules(data_size);
9014         if (!chain->blob_next)
9015                 return -ENOMEM;
9016
9017         data = (void *)chain->blob_next->data;
9018         data_boundary = data + data_size;
9019         size = 0;
9020
9021         list_for_each_entry(rule, &chain->rules, list) {
9022                 if (!nft_is_active_next(net, rule))
9023                         continue;
9024
9025                 prule = (struct nft_rule_dp *)data;
9026                 data += offsetof(struct nft_rule_dp, data);
9027                 if (WARN_ON_ONCE(data > data_boundary))
9028                         return -ENOMEM;
9029
9030                 size = 0;
9031                 track.last = nft_expr_last(rule);
9032                 nft_rule_for_each_expr(expr, last, rule) {
9033                         track.cur = expr;
9034
9035                         if (nft_expr_reduce(&track, expr)) {
9036                                 expr = track.cur;
9037                                 continue;
9038                         }
9039
9040                         if (WARN_ON_ONCE(data + size + expr->ops->size > data_boundary))
9041                                 return -ENOMEM;
9042
9043                         memcpy(data + size, expr, expr->ops->size);
9044                         size += expr->ops->size;
9045                 }
9046                 if (WARN_ON_ONCE(size >= 1 << 12))
9047                         return -ENOMEM;
9048
9049                 prule->handle = rule->handle;
9050                 prule->dlen = size;
9051                 prule->is_last = 0;
9052
9053                 data += size;
9054                 size = 0;
9055                 chain->blob_next->size += (unsigned long)(data - (void *)prule);
9056         }
9057
9058         prule = (struct nft_rule_dp *)data;
9059         data += offsetof(struct nft_rule_dp, data);
9060         if (WARN_ON_ONCE(data > data_boundary))
9061                 return -ENOMEM;
9062
9063         nft_last_rule(chain->blob_next, prule);
9064
9065         return 0;
9066 }
9067
9068 static void nf_tables_commit_chain_prepare_cancel(struct net *net)
9069 {
9070         struct nftables_pernet *nft_net = nft_pernet(net);
9071         struct nft_trans *trans, *next;
9072
9073         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
9074                 struct nft_chain *chain = trans->ctx.chain;
9075
9076                 if (trans->msg_type == NFT_MSG_NEWRULE ||
9077                     trans->msg_type == NFT_MSG_DELRULE) {
9078                         kvfree(chain->blob_next);
9079                         chain->blob_next = NULL;
9080                 }
9081         }
9082 }
9083
9084 static void __nf_tables_commit_chain_free_rules_old(struct rcu_head *h)
9085 {
9086         struct nft_rules_old *o = container_of(h, struct nft_rules_old, h);
9087
9088         kvfree(o->blob);
9089 }
9090
9091 static void nf_tables_commit_chain_free_rules_old(struct nft_rule_blob *blob)
9092 {
9093         struct nft_rules_old *old;
9094
9095         /* rcu_head is after end marker */
9096         old = (void *)blob + sizeof(*blob) + blob->size;
9097         old->blob = blob;
9098
9099         call_rcu(&old->h, __nf_tables_commit_chain_free_rules_old);
9100 }
9101
9102 static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
9103 {
9104         struct nft_rule_blob *g0, *g1;
9105         bool next_genbit;
9106
9107         next_genbit = nft_gencursor_next(net);
9108
9109         g0 = rcu_dereference_protected(chain->blob_gen_0,
9110                                        lockdep_commit_lock_is_held(net));
9111         g1 = rcu_dereference_protected(chain->blob_gen_1,
9112                                        lockdep_commit_lock_is_held(net));
9113
9114         /* No changes to this chain? */
9115         if (chain->blob_next == NULL) {
9116                 /* chain had no change in last or next generation */
9117                 if (g0 == g1)
9118                         return;
9119                 /*
9120                  * chain had no change in this generation; make sure next
9121                  * one uses same rules as current generation.
9122                  */
9123                 if (next_genbit) {
9124                         rcu_assign_pointer(chain->blob_gen_1, g0);
9125                         nf_tables_commit_chain_free_rules_old(g1);
9126                 } else {
9127                         rcu_assign_pointer(chain->blob_gen_0, g1);
9128                         nf_tables_commit_chain_free_rules_old(g0);
9129                 }
9130
9131                 return;
9132         }
9133
9134         if (next_genbit)
9135                 rcu_assign_pointer(chain->blob_gen_1, chain->blob_next);
9136         else
9137                 rcu_assign_pointer(chain->blob_gen_0, chain->blob_next);
9138
9139         chain->blob_next = NULL;
9140
9141         if (g0 == g1)
9142                 return;
9143
9144         if (next_genbit)
9145                 nf_tables_commit_chain_free_rules_old(g1);
9146         else
9147                 nf_tables_commit_chain_free_rules_old(g0);
9148 }
9149
9150 static void nft_obj_del(struct nft_object *obj)
9151 {
9152         rhltable_remove(&nft_objname_ht, &obj->rhlhead, nft_objname_ht_params);
9153         list_del_rcu(&obj->list);
9154 }
9155
9156 void nft_chain_del(struct nft_chain *chain)
9157 {
9158         struct nft_table *table = chain->table;
9159
9160         WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
9161                                      nft_chain_ht_params));
9162         list_del_rcu(&chain->list);
9163 }
9164
9165 static void nft_trans_gc_setelem_remove(struct nft_ctx *ctx,
9166                                         struct nft_trans_gc *trans)
9167 {
9168         void **priv = trans->priv;
9169         unsigned int i;
9170
9171         for (i = 0; i < trans->count; i++) {
9172                 struct nft_set_elem elem = {
9173                         .priv = priv[i],
9174                 };
9175
9176                 nft_setelem_data_deactivate(ctx->net, trans->set, &elem);
9177                 nft_setelem_remove(ctx->net, trans->set, &elem);
9178         }
9179 }
9180
9181 void nft_trans_gc_destroy(struct nft_trans_gc *trans)
9182 {
9183         nft_set_put(trans->set);
9184         put_net(trans->net);
9185         kfree(trans);
9186 }
9187
9188 static void nft_trans_gc_trans_free(struct rcu_head *rcu)
9189 {
9190         struct nft_set_elem elem = {};
9191         struct nft_trans_gc *trans;
9192         struct nft_ctx ctx = {};
9193         unsigned int i;
9194
9195         trans = container_of(rcu, struct nft_trans_gc, rcu);
9196         ctx.net = read_pnet(&trans->set->net);
9197
9198         for (i = 0; i < trans->count; i++) {
9199                 elem.priv = trans->priv[i];
9200                 if (!nft_setelem_is_catchall(trans->set, &elem))
9201                         atomic_dec(&trans->set->nelems);
9202
9203                 nf_tables_set_elem_destroy(&ctx, trans->set, elem.priv);
9204         }
9205
9206         nft_trans_gc_destroy(trans);
9207 }
9208
9209 static bool nft_trans_gc_work_done(struct nft_trans_gc *trans)
9210 {
9211         struct nftables_pernet *nft_net;
9212         struct nft_ctx ctx = {};
9213
9214         nft_net = nft_pernet(trans->net);
9215
9216         mutex_lock(&nft_net->commit_mutex);
9217
9218         /* Check for race with transaction, otherwise this batch refers to
9219          * stale objects that might not be there anymore. Skip transaction if
9220          * set has been destroyed from control plane transaction in case gc
9221          * worker loses race.
9222          */
9223         if (READ_ONCE(nft_net->gc_seq) != trans->seq || trans->set->dead) {
9224                 mutex_unlock(&nft_net->commit_mutex);
9225                 return false;
9226         }
9227
9228         ctx.net = trans->net;
9229         ctx.table = trans->set->table;
9230
9231         nft_trans_gc_setelem_remove(&ctx, trans);
9232         mutex_unlock(&nft_net->commit_mutex);
9233
9234         return true;
9235 }
9236
9237 static void nft_trans_gc_work(struct work_struct *work)
9238 {
9239         struct nft_trans_gc *trans, *next;
9240         LIST_HEAD(trans_gc_list);
9241
9242         spin_lock(&nf_tables_gc_list_lock);
9243         list_splice_init(&nf_tables_gc_list, &trans_gc_list);
9244         spin_unlock(&nf_tables_gc_list_lock);
9245
9246         list_for_each_entry_safe(trans, next, &trans_gc_list, list) {
9247                 list_del(&trans->list);
9248                 if (!nft_trans_gc_work_done(trans)) {
9249                         nft_trans_gc_destroy(trans);
9250                         continue;
9251                 }
9252                 call_rcu(&trans->rcu, nft_trans_gc_trans_free);
9253         }
9254 }
9255
9256 struct nft_trans_gc *nft_trans_gc_alloc(struct nft_set *set,
9257                                         unsigned int gc_seq, gfp_t gfp)
9258 {
9259         struct net *net = read_pnet(&set->net);
9260         struct nft_trans_gc *trans;
9261
9262         trans = kzalloc(sizeof(*trans), gfp);
9263         if (!trans)
9264                 return NULL;
9265
9266         trans->net = maybe_get_net(net);
9267         if (!trans->net) {
9268                 kfree(trans);
9269                 return NULL;
9270         }
9271
9272         refcount_inc(&set->refs);
9273         trans->set = set;
9274         trans->seq = gc_seq;
9275
9276         return trans;
9277 }
9278
9279 void nft_trans_gc_elem_add(struct nft_trans_gc *trans, void *priv)
9280 {
9281         trans->priv[trans->count++] = priv;
9282 }
9283
9284 static void nft_trans_gc_queue_work(struct nft_trans_gc *trans)
9285 {
9286         spin_lock(&nf_tables_gc_list_lock);
9287         list_add_tail(&trans->list, &nf_tables_gc_list);
9288         spin_unlock(&nf_tables_gc_list_lock);
9289
9290         schedule_work(&trans_gc_work);
9291 }
9292
9293 static int nft_trans_gc_space(struct nft_trans_gc *trans)
9294 {
9295         return NFT_TRANS_GC_BATCHCOUNT - trans->count;
9296 }
9297
9298 struct nft_trans_gc *nft_trans_gc_queue_async(struct nft_trans_gc *gc,
9299                                               unsigned int gc_seq, gfp_t gfp)
9300 {
9301         struct nft_set *set;
9302
9303         if (nft_trans_gc_space(gc))
9304                 return gc;
9305
9306         set = gc->set;
9307         nft_trans_gc_queue_work(gc);
9308
9309         return nft_trans_gc_alloc(set, gc_seq, gfp);
9310 }
9311
9312 void nft_trans_gc_queue_async_done(struct nft_trans_gc *trans)
9313 {
9314         if (trans->count == 0) {
9315                 nft_trans_gc_destroy(trans);
9316                 return;
9317         }
9318
9319         nft_trans_gc_queue_work(trans);
9320 }
9321
9322 struct nft_trans_gc *nft_trans_gc_queue_sync(struct nft_trans_gc *gc, gfp_t gfp)
9323 {
9324         struct nft_set *set;
9325
9326         if (WARN_ON_ONCE(!lockdep_commit_lock_is_held(gc->net)))
9327                 return NULL;
9328
9329         if (nft_trans_gc_space(gc))
9330                 return gc;
9331
9332         set = gc->set;
9333         call_rcu(&gc->rcu, nft_trans_gc_trans_free);
9334
9335         return nft_trans_gc_alloc(set, 0, gfp);
9336 }
9337
9338 void nft_trans_gc_queue_sync_done(struct nft_trans_gc *trans)
9339 {
9340         WARN_ON_ONCE(!lockdep_commit_lock_is_held(trans->net));
9341
9342         if (trans->count == 0) {
9343                 nft_trans_gc_destroy(trans);
9344                 return;
9345         }
9346
9347         call_rcu(&trans->rcu, nft_trans_gc_trans_free);
9348 }
9349
9350 struct nft_trans_gc *nft_trans_gc_catchall_async(struct nft_trans_gc *gc,
9351                                                  unsigned int gc_seq)
9352 {
9353         struct nft_set_elem_catchall *catchall;
9354         const struct nft_set *set = gc->set;
9355         struct nft_set_ext *ext;
9356
9357         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
9358                 ext = nft_set_elem_ext(set, catchall->elem);
9359
9360                 if (!nft_set_elem_expired(ext))
9361                         continue;
9362                 if (nft_set_elem_is_dead(ext))
9363                         goto dead_elem;
9364
9365                 nft_set_elem_dead(ext);
9366 dead_elem:
9367                 gc = nft_trans_gc_queue_async(gc, gc_seq, GFP_ATOMIC);
9368                 if (!gc)
9369                         return NULL;
9370
9371                 nft_trans_gc_elem_add(gc, catchall->elem);
9372         }
9373
9374         return gc;
9375 }
9376
9377 struct nft_trans_gc *nft_trans_gc_catchall_sync(struct nft_trans_gc *gc)
9378 {
9379         struct nft_set_elem_catchall *catchall, *next;
9380         const struct nft_set *set = gc->set;
9381         struct nft_set_elem elem;
9382         struct nft_set_ext *ext;
9383
9384         WARN_ON_ONCE(!lockdep_commit_lock_is_held(gc->net));
9385
9386         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
9387                 ext = nft_set_elem_ext(set, catchall->elem);
9388
9389                 if (!nft_set_elem_expired(ext))
9390                         continue;
9391
9392                 gc = nft_trans_gc_queue_sync(gc, GFP_KERNEL);
9393                 if (!gc)
9394                         return NULL;
9395
9396                 memset(&elem, 0, sizeof(elem));
9397                 elem.priv = catchall->elem;
9398
9399                 nft_setelem_data_deactivate(gc->net, gc->set, &elem);
9400                 nft_setelem_catchall_destroy(catchall);
9401                 nft_trans_gc_elem_add(gc, elem.priv);
9402         }
9403
9404         return gc;
9405 }
9406
9407 static void nf_tables_module_autoload_cleanup(struct net *net)
9408 {
9409         struct nftables_pernet *nft_net = nft_pernet(net);
9410         struct nft_module_request *req, *next;
9411
9412         WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
9413         list_for_each_entry_safe(req, next, &nft_net->module_list, list) {
9414                 WARN_ON_ONCE(!req->done);
9415                 list_del(&req->list);
9416                 kfree(req);
9417         }
9418 }
9419
9420 static void nf_tables_commit_release(struct net *net)
9421 {
9422         struct nftables_pernet *nft_net = nft_pernet(net);
9423         struct nft_trans *trans;
9424
9425         /* all side effects have to be made visible.
9426          * For example, if a chain named 'foo' has been deleted, a
9427          * new transaction must not find it anymore.
9428          *
9429          * Memory reclaim happens asynchronously from work queue
9430          * to prevent expensive synchronize_rcu() in commit phase.
9431          */
9432         if (list_empty(&nft_net->commit_list)) {
9433                 nf_tables_module_autoload_cleanup(net);
9434                 mutex_unlock(&nft_net->commit_mutex);
9435                 return;
9436         }
9437
9438         trans = list_last_entry(&nft_net->commit_list,
9439                                 struct nft_trans, list);
9440         get_net(trans->ctx.net);
9441         WARN_ON_ONCE(trans->put_net);
9442
9443         trans->put_net = true;
9444         spin_lock(&nf_tables_destroy_list_lock);
9445         list_splice_tail_init(&nft_net->commit_list, &nf_tables_destroy_list);
9446         spin_unlock(&nf_tables_destroy_list_lock);
9447
9448         nf_tables_module_autoload_cleanup(net);
9449         schedule_work(&trans_destroy_work);
9450
9451         mutex_unlock(&nft_net->commit_mutex);
9452 }
9453
9454 static void nft_commit_notify(struct net *net, u32 portid)
9455 {
9456         struct nftables_pernet *nft_net = nft_pernet(net);
9457         struct sk_buff *batch_skb = NULL, *nskb, *skb;
9458         unsigned char *data;
9459         int len;
9460
9461         list_for_each_entry_safe(skb, nskb, &nft_net->notify_list, list) {
9462                 if (!batch_skb) {
9463 new_batch:
9464                         batch_skb = skb;
9465                         len = NLMSG_GOODSIZE - skb->len;
9466                         list_del(&skb->list);
9467                         continue;
9468                 }
9469                 len -= skb->len;
9470                 if (len > 0 && NFT_CB(skb).report == NFT_CB(batch_skb).report) {
9471                         data = skb_put(batch_skb, skb->len);
9472                         memcpy(data, skb->data, skb->len);
9473                         list_del(&skb->list);
9474                         kfree_skb(skb);
9475                         continue;
9476                 }
9477                 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
9478                                NFT_CB(batch_skb).report, GFP_KERNEL);
9479                 goto new_batch;
9480         }
9481
9482         if (batch_skb) {
9483                 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
9484                                NFT_CB(batch_skb).report, GFP_KERNEL);
9485         }
9486
9487         WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
9488 }
9489
9490 static int nf_tables_commit_audit_alloc(struct list_head *adl,
9491                                         struct nft_table *table)
9492 {
9493         struct nft_audit_data *adp;
9494
9495         list_for_each_entry(adp, adl, list) {
9496                 if (adp->table == table)
9497                         return 0;
9498         }
9499         adp = kzalloc(sizeof(*adp), GFP_KERNEL);
9500         if (!adp)
9501                 return -ENOMEM;
9502         adp->table = table;
9503         list_add(&adp->list, adl);
9504         return 0;
9505 }
9506
9507 static void nf_tables_commit_audit_free(struct list_head *adl)
9508 {
9509         struct nft_audit_data *adp, *adn;
9510
9511         list_for_each_entry_safe(adp, adn, adl, list) {
9512                 list_del(&adp->list);
9513                 kfree(adp);
9514         }
9515 }
9516
9517 static void nf_tables_commit_audit_collect(struct list_head *adl,
9518                                            struct nft_table *table, u32 op)
9519 {
9520         struct nft_audit_data *adp;
9521
9522         list_for_each_entry(adp, adl, list) {
9523                 if (adp->table == table)
9524                         goto found;
9525         }
9526         WARN_ONCE(1, "table=%s not expected in commit list", table->name);
9527         return;
9528 found:
9529         adp->entries++;
9530         if (!adp->op || adp->op > op)
9531                 adp->op = op;
9532 }
9533
9534 #define AUNFTABLENAMELEN (NFT_TABLE_MAXNAMELEN + 22)
9535
9536 static void nf_tables_commit_audit_log(struct list_head *adl, u32 generation)
9537 {
9538         struct nft_audit_data *adp, *adn;
9539         char aubuf[AUNFTABLENAMELEN];
9540
9541         list_for_each_entry_safe(adp, adn, adl, list) {
9542                 snprintf(aubuf, AUNFTABLENAMELEN, "%s:%u", adp->table->name,
9543                          generation);
9544                 audit_log_nfcfg(aubuf, adp->table->family, adp->entries,
9545                                 nft2audit_op[adp->op], GFP_KERNEL);
9546                 list_del(&adp->list);
9547                 kfree(adp);
9548         }
9549 }
9550
9551 static void nft_set_commit_update(struct list_head *set_update_list)
9552 {
9553         struct nft_set *set, *next;
9554
9555         list_for_each_entry_safe(set, next, set_update_list, pending_update) {
9556                 list_del_init(&set->pending_update);
9557
9558                 if (!set->ops->commit || set->dead)
9559                         continue;
9560
9561                 set->ops->commit(set);
9562         }
9563 }
9564
9565 static unsigned int nft_gc_seq_begin(struct nftables_pernet *nft_net)
9566 {
9567         unsigned int gc_seq;
9568
9569         /* Bump gc counter, it becomes odd, this is the busy mark. */
9570         gc_seq = READ_ONCE(nft_net->gc_seq);
9571         WRITE_ONCE(nft_net->gc_seq, ++gc_seq);
9572
9573         return gc_seq;
9574 }
9575
9576 static void nft_gc_seq_end(struct nftables_pernet *nft_net, unsigned int gc_seq)
9577 {
9578         WRITE_ONCE(nft_net->gc_seq, ++gc_seq);
9579 }
9580
9581 static int nf_tables_commit(struct net *net, struct sk_buff *skb)
9582 {
9583         struct nftables_pernet *nft_net = nft_pernet(net);
9584         struct nft_trans *trans, *next;
9585         unsigned int base_seq, gc_seq;
9586         LIST_HEAD(set_update_list);
9587         struct nft_trans_elem *te;
9588         struct nft_chain *chain;
9589         struct nft_table *table;
9590         LIST_HEAD(adl);
9591         int err;
9592
9593         if (list_empty(&nft_net->commit_list)) {
9594                 mutex_unlock(&nft_net->commit_mutex);
9595                 return 0;
9596         }
9597
9598         list_for_each_entry(trans, &nft_net->binding_list, binding_list) {
9599                 switch (trans->msg_type) {
9600                 case NFT_MSG_NEWSET:
9601                         if (!nft_trans_set_update(trans) &&
9602                             nft_set_is_anonymous(nft_trans_set(trans)) &&
9603                             !nft_trans_set_bound(trans)) {
9604                                 pr_warn_once("nftables ruleset with unbound set\n");
9605                                 return -EINVAL;
9606                         }
9607                         break;
9608                 case NFT_MSG_NEWCHAIN:
9609                         if (!nft_trans_chain_update(trans) &&
9610                             nft_chain_binding(nft_trans_chain(trans)) &&
9611                             !nft_trans_chain_bound(trans)) {
9612                                 pr_warn_once("nftables ruleset with unbound chain\n");
9613                                 return -EINVAL;
9614                         }
9615                         break;
9616                 }
9617         }
9618
9619         /* 0. Validate ruleset, otherwise roll back for error reporting. */
9620         if (nf_tables_validate(net) < 0)
9621                 return -EAGAIN;
9622
9623         err = nft_flow_rule_offload_commit(net);
9624         if (err < 0)
9625                 return err;
9626
9627         /* 1.  Allocate space for next generation rules_gen_X[] */
9628         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
9629                 int ret;
9630
9631                 ret = nf_tables_commit_audit_alloc(&adl, trans->ctx.table);
9632                 if (ret) {
9633                         nf_tables_commit_chain_prepare_cancel(net);
9634                         nf_tables_commit_audit_free(&adl);
9635                         return ret;
9636                 }
9637                 if (trans->msg_type == NFT_MSG_NEWRULE ||
9638                     trans->msg_type == NFT_MSG_DELRULE) {
9639                         chain = trans->ctx.chain;
9640
9641                         ret = nf_tables_commit_chain_prepare(net, chain);
9642                         if (ret < 0) {
9643                                 nf_tables_commit_chain_prepare_cancel(net);
9644                                 nf_tables_commit_audit_free(&adl);
9645                                 return ret;
9646                         }
9647                 }
9648         }
9649
9650         /* step 2.  Make rules_gen_X visible to packet path */
9651         list_for_each_entry(table, &nft_net->tables, list) {
9652                 list_for_each_entry(chain, &table->chains, list)
9653                         nf_tables_commit_chain(net, chain);
9654         }
9655
9656         /*
9657          * Bump generation counter, invalidate any dump in progress.
9658          * Cannot fail after this point.
9659          */
9660         base_seq = READ_ONCE(nft_net->base_seq);
9661         while (++base_seq == 0)
9662                 ;
9663
9664         WRITE_ONCE(nft_net->base_seq, base_seq);
9665
9666         gc_seq = nft_gc_seq_begin(nft_net);
9667
9668         /* step 3. Start new generation, rules_gen_X now in use. */
9669         net->nft.gencursor = nft_gencursor_next(net);
9670
9671         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
9672                 nf_tables_commit_audit_collect(&adl, trans->ctx.table,
9673                                                trans->msg_type);
9674                 switch (trans->msg_type) {
9675                 case NFT_MSG_NEWTABLE:
9676                         if (nft_trans_table_update(trans)) {
9677                                 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
9678                                         nft_trans_destroy(trans);
9679                                         break;
9680                                 }
9681                                 if (trans->ctx.table->flags & NFT_TABLE_F_DORMANT)
9682                                         nf_tables_table_disable(net, trans->ctx.table);
9683
9684                                 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
9685                         } else {
9686                                 nft_clear(net, trans->ctx.table);
9687                         }
9688                         nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
9689                         nft_trans_destroy(trans);
9690                         break;
9691                 case NFT_MSG_DELTABLE:
9692                         list_del_rcu(&trans->ctx.table->list);
9693                         nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
9694                         break;
9695                 case NFT_MSG_NEWCHAIN:
9696                         if (nft_trans_chain_update(trans)) {
9697                                 nft_chain_commit_update(trans);
9698                                 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
9699                                 /* trans destroyed after rcu grace period */
9700                         } else {
9701                                 nft_chain_commit_drop_policy(trans);
9702                                 nft_clear(net, trans->ctx.chain);
9703                                 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
9704                                 nft_trans_destroy(trans);
9705                         }
9706                         break;
9707                 case NFT_MSG_DELCHAIN:
9708                         nft_chain_del(trans->ctx.chain);
9709                         nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
9710                         nf_tables_unregister_hook(trans->ctx.net,
9711                                                   trans->ctx.table,
9712                                                   trans->ctx.chain);
9713                         break;
9714                 case NFT_MSG_NEWRULE:
9715                         nft_clear(trans->ctx.net, nft_trans_rule(trans));
9716                         nf_tables_rule_notify(&trans->ctx,
9717                                               nft_trans_rule(trans),
9718                                               NFT_MSG_NEWRULE);
9719                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9720                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9721
9722                         nft_trans_destroy(trans);
9723                         break;
9724                 case NFT_MSG_DELRULE:
9725                         list_del_rcu(&nft_trans_rule(trans)->list);
9726                         nf_tables_rule_notify(&trans->ctx,
9727                                               nft_trans_rule(trans),
9728                                               NFT_MSG_DELRULE);
9729                         nft_rule_expr_deactivate(&trans->ctx,
9730                                                  nft_trans_rule(trans),
9731                                                  NFT_TRANS_COMMIT);
9732
9733                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9734                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9735                         break;
9736                 case NFT_MSG_NEWSET:
9737                         if (nft_trans_set_update(trans)) {
9738                                 struct nft_set *set = nft_trans_set(trans);
9739
9740                                 WRITE_ONCE(set->timeout, nft_trans_set_timeout(trans));
9741                                 WRITE_ONCE(set->gc_int, nft_trans_set_gc_int(trans));
9742                         } else {
9743                                 nft_clear(net, nft_trans_set(trans));
9744                                 /* This avoids hitting -EBUSY when deleting the table
9745                                  * from the transaction.
9746                                  */
9747                                 if (nft_set_is_anonymous(nft_trans_set(trans)) &&
9748                                     !list_empty(&nft_trans_set(trans)->bindings))
9749                                         nft_use_dec(&trans->ctx.table->use);
9750                         }
9751                         nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
9752                                              NFT_MSG_NEWSET, GFP_KERNEL);
9753                         nft_trans_destroy(trans);
9754                         break;
9755                 case NFT_MSG_DELSET:
9756                         nft_trans_set(trans)->dead = 1;
9757                         list_del_rcu(&nft_trans_set(trans)->list);
9758                         nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
9759                                              NFT_MSG_DELSET, GFP_KERNEL);
9760                         break;
9761                 case NFT_MSG_NEWSETELEM:
9762                         te = (struct nft_trans_elem *)trans->data;
9763
9764                         nft_setelem_activate(net, te->set, &te->elem);
9765                         nf_tables_setelem_notify(&trans->ctx, te->set,
9766                                                  &te->elem,
9767                                                  NFT_MSG_NEWSETELEM);
9768                         if (te->set->ops->commit &&
9769                             list_empty(&te->set->pending_update)) {
9770                                 list_add_tail(&te->set->pending_update,
9771                                               &set_update_list);
9772                         }
9773                         nft_trans_destroy(trans);
9774                         break;
9775                 case NFT_MSG_DELSETELEM:
9776                         te = (struct nft_trans_elem *)trans->data;
9777
9778                         nf_tables_setelem_notify(&trans->ctx, te->set,
9779                                                  &te->elem,
9780                                                  NFT_MSG_DELSETELEM);
9781                         nft_setelem_remove(net, te->set, &te->elem);
9782                         if (!nft_setelem_is_catchall(te->set, &te->elem)) {
9783                                 atomic_dec(&te->set->nelems);
9784                                 te->set->ndeact--;
9785                         }
9786                         if (te->set->ops->commit &&
9787                             list_empty(&te->set->pending_update)) {
9788                                 list_add_tail(&te->set->pending_update,
9789                                               &set_update_list);
9790                         }
9791                         break;
9792                 case NFT_MSG_NEWOBJ:
9793                         if (nft_trans_obj_update(trans)) {
9794                                 nft_obj_commit_update(trans);
9795                                 nf_tables_obj_notify(&trans->ctx,
9796                                                      nft_trans_obj(trans),
9797                                                      NFT_MSG_NEWOBJ);
9798                         } else {
9799                                 nft_clear(net, nft_trans_obj(trans));
9800                                 nf_tables_obj_notify(&trans->ctx,
9801                                                      nft_trans_obj(trans),
9802                                                      NFT_MSG_NEWOBJ);
9803                                 nft_trans_destroy(trans);
9804                         }
9805                         break;
9806                 case NFT_MSG_DELOBJ:
9807                         nft_obj_del(nft_trans_obj(trans));
9808                         nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
9809                                              NFT_MSG_DELOBJ);
9810                         break;
9811                 case NFT_MSG_NEWFLOWTABLE:
9812                         if (nft_trans_flowtable_update(trans)) {
9813                                 nft_trans_flowtable(trans)->data.flags =
9814                                         nft_trans_flowtable_flags(trans);
9815                                 nf_tables_flowtable_notify(&trans->ctx,
9816                                                            nft_trans_flowtable(trans),
9817                                                            &nft_trans_flowtable_hooks(trans),
9818                                                            NFT_MSG_NEWFLOWTABLE);
9819                                 list_splice(&nft_trans_flowtable_hooks(trans),
9820                                             &nft_trans_flowtable(trans)->hook_list);
9821                         } else {
9822                                 nft_clear(net, nft_trans_flowtable(trans));
9823                                 nf_tables_flowtable_notify(&trans->ctx,
9824                                                            nft_trans_flowtable(trans),
9825                                                            &nft_trans_flowtable(trans)->hook_list,
9826                                                            NFT_MSG_NEWFLOWTABLE);
9827                         }
9828                         nft_trans_destroy(trans);
9829                         break;
9830                 case NFT_MSG_DELFLOWTABLE:
9831                         if (nft_trans_flowtable_update(trans)) {
9832                                 nf_tables_flowtable_notify(&trans->ctx,
9833                                                            nft_trans_flowtable(trans),
9834                                                            &nft_trans_flowtable_hooks(trans),
9835                                                            NFT_MSG_DELFLOWTABLE);
9836                                 nft_unregister_flowtable_net_hooks(net,
9837                                                                    &nft_trans_flowtable_hooks(trans));
9838                         } else {
9839                                 list_del_rcu(&nft_trans_flowtable(trans)->list);
9840                                 nf_tables_flowtable_notify(&trans->ctx,
9841                                                            nft_trans_flowtable(trans),
9842                                                            &nft_trans_flowtable(trans)->hook_list,
9843                                                            NFT_MSG_DELFLOWTABLE);
9844                                 nft_unregister_flowtable_net_hooks(net,
9845                                                 &nft_trans_flowtable(trans)->hook_list);
9846                         }
9847                         break;
9848                 }
9849         }
9850
9851         nft_set_commit_update(&set_update_list);
9852
9853         nft_commit_notify(net, NETLINK_CB(skb).portid);
9854         nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
9855         nf_tables_commit_audit_log(&adl, nft_net->base_seq);
9856
9857         nft_gc_seq_end(nft_net, gc_seq);
9858         nf_tables_commit_release(net);
9859
9860         return 0;
9861 }
9862
9863 static void nf_tables_module_autoload(struct net *net)
9864 {
9865         struct nftables_pernet *nft_net = nft_pernet(net);
9866         struct nft_module_request *req, *next;
9867         LIST_HEAD(module_list);
9868
9869         list_splice_init(&nft_net->module_list, &module_list);
9870         mutex_unlock(&nft_net->commit_mutex);
9871         list_for_each_entry_safe(req, next, &module_list, list) {
9872                 request_module("%s", req->module);
9873                 req->done = true;
9874         }
9875         mutex_lock(&nft_net->commit_mutex);
9876         list_splice(&module_list, &nft_net->module_list);
9877 }
9878
9879 static void nf_tables_abort_release(struct nft_trans *trans)
9880 {
9881         switch (trans->msg_type) {
9882         case NFT_MSG_NEWTABLE:
9883                 nf_tables_table_destroy(&trans->ctx);
9884                 break;
9885         case NFT_MSG_NEWCHAIN:
9886                 nf_tables_chain_destroy(&trans->ctx);
9887                 break;
9888         case NFT_MSG_NEWRULE:
9889                 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
9890                 break;
9891         case NFT_MSG_NEWSET:
9892                 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
9893                 break;
9894         case NFT_MSG_NEWSETELEM:
9895                 nft_set_elem_destroy(nft_trans_elem_set(trans),
9896                                      nft_trans_elem(trans).priv, true);
9897                 break;
9898         case NFT_MSG_NEWOBJ:
9899                 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
9900                 break;
9901         case NFT_MSG_NEWFLOWTABLE:
9902                 if (nft_trans_flowtable_update(trans))
9903                         nft_hooks_destroy(&nft_trans_flowtable_hooks(trans));
9904                 else
9905                         nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
9906                 break;
9907         }
9908         kfree(trans);
9909 }
9910
9911 static void nft_set_abort_update(struct list_head *set_update_list)
9912 {
9913         struct nft_set *set, *next;
9914
9915         list_for_each_entry_safe(set, next, set_update_list, pending_update) {
9916                 list_del_init(&set->pending_update);
9917
9918                 if (!set->ops->abort)
9919                         continue;
9920
9921                 set->ops->abort(set);
9922         }
9923 }
9924
9925 static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)
9926 {
9927         struct nftables_pernet *nft_net = nft_pernet(net);
9928         struct nft_trans *trans, *next;
9929         LIST_HEAD(set_update_list);
9930         struct nft_trans_elem *te;
9931         int err = 0;
9932
9933         if (action == NFNL_ABORT_VALIDATE &&
9934             nf_tables_validate(net) < 0)
9935                 err = -EAGAIN;
9936
9937         list_for_each_entry_safe_reverse(trans, next, &nft_net->commit_list,
9938                                          list) {
9939                 switch (trans->msg_type) {
9940                 case NFT_MSG_NEWTABLE:
9941                         if (nft_trans_table_update(trans)) {
9942                                 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
9943                                         nft_trans_destroy(trans);
9944                                         break;
9945                                 }
9946                                 if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_DORMANT) {
9947                                         nf_tables_table_disable(net, trans->ctx.table);
9948                                         trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
9949                                 } else if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_AWAKEN) {
9950                                         trans->ctx.table->flags &= ~NFT_TABLE_F_DORMANT;
9951                                 }
9952                                 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
9953                                 nft_trans_destroy(trans);
9954                         } else {
9955                                 list_del_rcu(&trans->ctx.table->list);
9956                         }
9957                         break;
9958                 case NFT_MSG_DELTABLE:
9959                         nft_clear(trans->ctx.net, trans->ctx.table);
9960                         nft_trans_destroy(trans);
9961                         break;
9962                 case NFT_MSG_NEWCHAIN:
9963                         if (nft_trans_chain_update(trans)) {
9964                                 free_percpu(nft_trans_chain_stats(trans));
9965                                 kfree(nft_trans_chain_name(trans));
9966                                 nft_trans_destroy(trans);
9967                         } else {
9968                                 if (nft_trans_chain_bound(trans)) {
9969                                         nft_trans_destroy(trans);
9970                                         break;
9971                                 }
9972                                 nft_use_dec_restore(&trans->ctx.table->use);
9973                                 nft_chain_del(trans->ctx.chain);
9974                                 nf_tables_unregister_hook(trans->ctx.net,
9975                                                           trans->ctx.table,
9976                                                           trans->ctx.chain);
9977                         }
9978                         break;
9979                 case NFT_MSG_DELCHAIN:
9980                         nft_use_inc_restore(&trans->ctx.table->use);
9981                         nft_clear(trans->ctx.net, trans->ctx.chain);
9982                         nft_trans_destroy(trans);
9983                         break;
9984                 case NFT_MSG_NEWRULE:
9985                         if (nft_trans_rule_bound(trans)) {
9986                                 nft_trans_destroy(trans);
9987                                 break;
9988                         }
9989                         nft_use_dec_restore(&trans->ctx.chain->use);
9990                         list_del_rcu(&nft_trans_rule(trans)->list);
9991                         nft_rule_expr_deactivate(&trans->ctx,
9992                                                  nft_trans_rule(trans),
9993                                                  NFT_TRANS_ABORT);
9994                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9995                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9996                         break;
9997                 case NFT_MSG_DELRULE:
9998                         nft_use_inc_restore(&trans->ctx.chain->use);
9999                         nft_clear(trans->ctx.net, nft_trans_rule(trans));
10000                         nft_rule_expr_activate(&trans->ctx, nft_trans_rule(trans));
10001                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
10002                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
10003
10004                         nft_trans_destroy(trans);
10005                         break;
10006                 case NFT_MSG_NEWSET:
10007                         if (nft_trans_set_update(trans)) {
10008                                 nft_trans_destroy(trans);
10009                                 break;
10010                         }
10011                         nft_use_dec_restore(&trans->ctx.table->use);
10012                         if (nft_trans_set_bound(trans)) {
10013                                 nft_trans_destroy(trans);
10014                                 break;
10015                         }
10016                         nft_trans_set(trans)->dead = 1;
10017                         list_del_rcu(&nft_trans_set(trans)->list);
10018                         break;
10019                 case NFT_MSG_DELSET:
10020                         nft_use_inc_restore(&trans->ctx.table->use);
10021                         nft_clear(trans->ctx.net, nft_trans_set(trans));
10022                         if (nft_trans_set(trans)->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
10023                                 nft_map_activate(&trans->ctx, nft_trans_set(trans));
10024
10025                         nft_trans_destroy(trans);
10026                         break;
10027                 case NFT_MSG_NEWSETELEM:
10028                         if (nft_trans_elem_set_bound(trans)) {
10029                                 nft_trans_destroy(trans);
10030                                 break;
10031                         }
10032                         te = (struct nft_trans_elem *)trans->data;
10033                         nft_setelem_remove(net, te->set, &te->elem);
10034                         if (!nft_setelem_is_catchall(te->set, &te->elem))
10035                                 atomic_dec(&te->set->nelems);
10036
10037                         if (te->set->ops->abort &&
10038                             list_empty(&te->set->pending_update)) {
10039                                 list_add_tail(&te->set->pending_update,
10040                                               &set_update_list);
10041                         }
10042                         break;
10043                 case NFT_MSG_DELSETELEM:
10044                         te = (struct nft_trans_elem *)trans->data;
10045
10046                         nft_setelem_data_activate(net, te->set, &te->elem);
10047                         nft_setelem_activate(net, te->set, &te->elem);
10048                         if (!nft_setelem_is_catchall(te->set, &te->elem))
10049                                 te->set->ndeact--;
10050
10051                         if (te->set->ops->abort &&
10052                             list_empty(&te->set->pending_update)) {
10053                                 list_add_tail(&te->set->pending_update,
10054                                               &set_update_list);
10055                         }
10056                         nft_trans_destroy(trans);
10057                         break;
10058                 case NFT_MSG_NEWOBJ:
10059                         if (nft_trans_obj_update(trans)) {
10060                                 nft_obj_destroy(&trans->ctx, nft_trans_obj_newobj(trans));
10061                                 nft_trans_destroy(trans);
10062                         } else {
10063                                 nft_use_dec_restore(&trans->ctx.table->use);
10064                                 nft_obj_del(nft_trans_obj(trans));
10065                         }
10066                         break;
10067                 case NFT_MSG_DELOBJ:
10068                         nft_use_inc_restore(&trans->ctx.table->use);
10069                         nft_clear(trans->ctx.net, nft_trans_obj(trans));
10070                         nft_trans_destroy(trans);
10071                         break;
10072                 case NFT_MSG_NEWFLOWTABLE:
10073                         if (nft_trans_flowtable_update(trans)) {
10074                                 nft_unregister_flowtable_net_hooks(net,
10075                                                 &nft_trans_flowtable_hooks(trans));
10076                         } else {
10077                                 nft_use_dec_restore(&trans->ctx.table->use);
10078                                 list_del_rcu(&nft_trans_flowtable(trans)->list);
10079                                 nft_unregister_flowtable_net_hooks(net,
10080                                                 &nft_trans_flowtable(trans)->hook_list);
10081                         }
10082                         break;
10083                 case NFT_MSG_DELFLOWTABLE:
10084                         if (nft_trans_flowtable_update(trans)) {
10085                                 list_splice(&nft_trans_flowtable_hooks(trans),
10086                                             &nft_trans_flowtable(trans)->hook_list);
10087                         } else {
10088                                 nft_use_inc_restore(&trans->ctx.table->use);
10089                                 nft_clear(trans->ctx.net, nft_trans_flowtable(trans));
10090                         }
10091                         nft_trans_destroy(trans);
10092                         break;
10093                 }
10094         }
10095
10096         nft_set_abort_update(&set_update_list);
10097
10098         synchronize_rcu();
10099
10100         list_for_each_entry_safe_reverse(trans, next,
10101                                          &nft_net->commit_list, list) {
10102                 nft_trans_list_del(trans);
10103                 nf_tables_abort_release(trans);
10104         }
10105
10106         return err;
10107 }
10108
10109 static int nf_tables_abort(struct net *net, struct sk_buff *skb,
10110                            enum nfnl_abort_action action)
10111 {
10112         struct nftables_pernet *nft_net = nft_pernet(net);
10113         unsigned int gc_seq;
10114         int ret;
10115
10116         gc_seq = nft_gc_seq_begin(nft_net);
10117         ret = __nf_tables_abort(net, action);
10118         nft_gc_seq_end(nft_net, gc_seq);
10119
10120         WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
10121
10122         /* module autoload needs to happen after GC sequence update because it
10123          * temporarily releases and grabs mutex again.
10124          */
10125         if (action == NFNL_ABORT_AUTOLOAD)
10126                 nf_tables_module_autoload(net);
10127         else
10128                 nf_tables_module_autoload_cleanup(net);
10129
10130         mutex_unlock(&nft_net->commit_mutex);
10131
10132         return ret;
10133 }
10134
10135 static bool nf_tables_valid_genid(struct net *net, u32 genid)
10136 {
10137         struct nftables_pernet *nft_net = nft_pernet(net);
10138         bool genid_ok;
10139
10140         mutex_lock(&nft_net->commit_mutex);
10141
10142         genid_ok = genid == 0 || nft_net->base_seq == genid;
10143         if (!genid_ok)
10144                 mutex_unlock(&nft_net->commit_mutex);
10145
10146         /* else, commit mutex has to be released by commit or abort function */
10147         return genid_ok;
10148 }
10149
10150 static const struct nfnetlink_subsystem nf_tables_subsys = {
10151         .name           = "nf_tables",
10152         .subsys_id      = NFNL_SUBSYS_NFTABLES,
10153         .cb_count       = NFT_MSG_MAX,
10154         .cb             = nf_tables_cb,
10155         .commit         = nf_tables_commit,
10156         .abort          = nf_tables_abort,
10157         .valid_genid    = nf_tables_valid_genid,
10158         .owner          = THIS_MODULE,
10159 };
10160
10161 int nft_chain_validate_dependency(const struct nft_chain *chain,
10162                                   enum nft_chain_types type)
10163 {
10164         const struct nft_base_chain *basechain;
10165
10166         if (nft_is_base_chain(chain)) {
10167                 basechain = nft_base_chain(chain);
10168                 if (basechain->type->type != type)
10169                         return -EOPNOTSUPP;
10170         }
10171         return 0;
10172 }
10173 EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
10174
10175 int nft_chain_validate_hooks(const struct nft_chain *chain,
10176                              unsigned int hook_flags)
10177 {
10178         struct nft_base_chain *basechain;
10179
10180         if (nft_is_base_chain(chain)) {
10181                 basechain = nft_base_chain(chain);
10182
10183                 if ((1 << basechain->ops.hooknum) & hook_flags)
10184                         return 0;
10185
10186                 return -EOPNOTSUPP;
10187         }
10188
10189         return 0;
10190 }
10191 EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
10192
10193 /*
10194  * Loop detection - walk through the ruleset beginning at the destination chain
10195  * of a new jump until either the source chain is reached (loop) or all
10196  * reachable chains have been traversed.
10197  *
10198  * The loop check is performed whenever a new jump verdict is added to an
10199  * expression or verdict map or a verdict map is bound to a new chain.
10200  */
10201
10202 static int nf_tables_check_loops(const struct nft_ctx *ctx,
10203                                  const struct nft_chain *chain);
10204
10205 static int nft_check_loops(const struct nft_ctx *ctx,
10206                            const struct nft_set_ext *ext)
10207 {
10208         const struct nft_data *data;
10209         int ret;
10210
10211         data = nft_set_ext_data(ext);
10212         switch (data->verdict.code) {
10213         case NFT_JUMP:
10214         case NFT_GOTO:
10215                 ret = nf_tables_check_loops(ctx, data->verdict.chain);
10216                 break;
10217         default:
10218                 ret = 0;
10219                 break;
10220         }
10221
10222         return ret;
10223 }
10224
10225 static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
10226                                         struct nft_set *set,
10227                                         const struct nft_set_iter *iter,
10228                                         struct nft_set_elem *elem)
10229 {
10230         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
10231
10232         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
10233             *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
10234                 return 0;
10235
10236         return nft_check_loops(ctx, ext);
10237 }
10238
10239 static int nft_set_catchall_loops(const struct nft_ctx *ctx,
10240                                   struct nft_set *set)
10241 {
10242         u8 genmask = nft_genmask_next(ctx->net);
10243         struct nft_set_elem_catchall *catchall;
10244         struct nft_set_ext *ext;
10245         int ret = 0;
10246
10247         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
10248                 ext = nft_set_elem_ext(set, catchall->elem);
10249                 if (!nft_set_elem_active(ext, genmask))
10250                         continue;
10251
10252                 ret = nft_check_loops(ctx, ext);
10253                 if (ret < 0)
10254                         return ret;
10255         }
10256
10257         return ret;
10258 }
10259
10260 static int nf_tables_check_loops(const struct nft_ctx *ctx,
10261                                  const struct nft_chain *chain)
10262 {
10263         const struct nft_rule *rule;
10264         const struct nft_expr *expr, *last;
10265         struct nft_set *set;
10266         struct nft_set_binding *binding;
10267         struct nft_set_iter iter;
10268
10269         if (ctx->chain == chain)
10270                 return -ELOOP;
10271
10272         list_for_each_entry(rule, &chain->rules, list) {
10273                 nft_rule_for_each_expr(expr, last, rule) {
10274                         struct nft_immediate_expr *priv;
10275                         const struct nft_data *data;
10276                         int err;
10277
10278                         if (strcmp(expr->ops->type->name, "immediate"))
10279                                 continue;
10280
10281                         priv = nft_expr_priv(expr);
10282                         if (priv->dreg != NFT_REG_VERDICT)
10283                                 continue;
10284
10285                         data = &priv->data;
10286                         switch (data->verdict.code) {
10287                         case NFT_JUMP:
10288                         case NFT_GOTO:
10289                                 err = nf_tables_check_loops(ctx,
10290                                                         data->verdict.chain);
10291                                 if (err < 0)
10292                                         return err;
10293                                 break;
10294                         default:
10295                                 break;
10296                         }
10297                 }
10298         }
10299
10300         list_for_each_entry(set, &ctx->table->sets, list) {
10301                 if (!nft_is_active_next(ctx->net, set))
10302                         continue;
10303                 if (!(set->flags & NFT_SET_MAP) ||
10304                     set->dtype != NFT_DATA_VERDICT)
10305                         continue;
10306
10307                 list_for_each_entry(binding, &set->bindings, list) {
10308                         if (!(binding->flags & NFT_SET_MAP) ||
10309                             binding->chain != chain)
10310                                 continue;
10311
10312                         iter.genmask    = nft_genmask_next(ctx->net);
10313                         iter.skip       = 0;
10314                         iter.count      = 0;
10315                         iter.err        = 0;
10316                         iter.fn         = nf_tables_loop_check_setelem;
10317
10318                         set->ops->walk(ctx, set, &iter);
10319                         if (!iter.err)
10320                                 iter.err = nft_set_catchall_loops(ctx, set);
10321
10322                         if (iter.err < 0)
10323                                 return iter.err;
10324                 }
10325         }
10326
10327         return 0;
10328 }
10329
10330 /**
10331  *      nft_parse_u32_check - fetch u32 attribute and check for maximum value
10332  *
10333  *      @attr: netlink attribute to fetch value from
10334  *      @max: maximum value to be stored in dest
10335  *      @dest: pointer to the variable
10336  *
10337  *      Parse, check and store a given u32 netlink attribute into variable.
10338  *      This function returns -ERANGE if the value goes over maximum value.
10339  *      Otherwise a 0 is returned and the attribute value is stored in the
10340  *      destination variable.
10341  */
10342 int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
10343 {
10344         u32 val;
10345
10346         val = ntohl(nla_get_be32(attr));
10347         if (val > max)
10348                 return -ERANGE;
10349
10350         *dest = val;
10351         return 0;
10352 }
10353 EXPORT_SYMBOL_GPL(nft_parse_u32_check);
10354
10355 static int nft_parse_register(const struct nlattr *attr, u32 *preg)
10356 {
10357         unsigned int reg;
10358
10359         reg = ntohl(nla_get_be32(attr));
10360         switch (reg) {
10361         case NFT_REG_VERDICT...NFT_REG_4:
10362                 *preg = reg * NFT_REG_SIZE / NFT_REG32_SIZE;
10363                 break;
10364         case NFT_REG32_00...NFT_REG32_15:
10365                 *preg = reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
10366                 break;
10367         default:
10368                 return -ERANGE;
10369         }
10370
10371         return 0;
10372 }
10373
10374 /**
10375  *      nft_dump_register - dump a register value to a netlink attribute
10376  *
10377  *      @skb: socket buffer
10378  *      @attr: attribute number
10379  *      @reg: register number
10380  *
10381  *      Construct a netlink attribute containing the register number. For
10382  *      compatibility reasons, register numbers being a multiple of 4 are
10383  *      translated to the corresponding 128 bit register numbers.
10384  */
10385 int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
10386 {
10387         if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
10388                 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
10389         else
10390                 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
10391
10392         return nla_put_be32(skb, attr, htonl(reg));
10393 }
10394 EXPORT_SYMBOL_GPL(nft_dump_register);
10395
10396 static int nft_validate_register_load(enum nft_registers reg, unsigned int len)
10397 {
10398         if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
10399                 return -EINVAL;
10400         if (len == 0)
10401                 return -EINVAL;
10402         if (reg * NFT_REG32_SIZE + len > sizeof_field(struct nft_regs, data))
10403                 return -ERANGE;
10404
10405         return 0;
10406 }
10407
10408 int nft_parse_register_load(const struct nlattr *attr, u8 *sreg, u32 len)
10409 {
10410         u32 reg;
10411         int err;
10412
10413         err = nft_parse_register(attr, &reg);
10414         if (err < 0)
10415                 return err;
10416
10417         err = nft_validate_register_load(reg, len);
10418         if (err < 0)
10419                 return err;
10420
10421         *sreg = reg;
10422         return 0;
10423 }
10424 EXPORT_SYMBOL_GPL(nft_parse_register_load);
10425
10426 static int nft_validate_register_store(const struct nft_ctx *ctx,
10427                                        enum nft_registers reg,
10428                                        const struct nft_data *data,
10429                                        enum nft_data_types type,
10430                                        unsigned int len)
10431 {
10432         int err;
10433
10434         switch (reg) {
10435         case NFT_REG_VERDICT:
10436                 if (type != NFT_DATA_VERDICT)
10437                         return -EINVAL;
10438
10439                 if (data != NULL &&
10440                     (data->verdict.code == NFT_GOTO ||
10441                      data->verdict.code == NFT_JUMP)) {
10442                         err = nf_tables_check_loops(ctx, data->verdict.chain);
10443                         if (err < 0)
10444                                 return err;
10445                 }
10446
10447                 return 0;
10448         default:
10449                 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
10450                         return -EINVAL;
10451                 if (len == 0)
10452                         return -EINVAL;
10453                 if (reg * NFT_REG32_SIZE + len >
10454                     sizeof_field(struct nft_regs, data))
10455                         return -ERANGE;
10456
10457                 if (data != NULL && type != NFT_DATA_VALUE)
10458                         return -EINVAL;
10459                 return 0;
10460         }
10461 }
10462
10463 int nft_parse_register_store(const struct nft_ctx *ctx,
10464                              const struct nlattr *attr, u8 *dreg,
10465                              const struct nft_data *data,
10466                              enum nft_data_types type, unsigned int len)
10467 {
10468         int err;
10469         u32 reg;
10470
10471         err = nft_parse_register(attr, &reg);
10472         if (err < 0)
10473                 return err;
10474
10475         err = nft_validate_register_store(ctx, reg, data, type, len);
10476         if (err < 0)
10477                 return err;
10478
10479         *dreg = reg;
10480         return 0;
10481 }
10482 EXPORT_SYMBOL_GPL(nft_parse_register_store);
10483
10484 static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
10485         [NFTA_VERDICT_CODE]     = { .type = NLA_U32 },
10486         [NFTA_VERDICT_CHAIN]    = { .type = NLA_STRING,
10487                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
10488         [NFTA_VERDICT_CHAIN_ID] = { .type = NLA_U32 },
10489 };
10490
10491 static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
10492                             struct nft_data_desc *desc, const struct nlattr *nla)
10493 {
10494         u8 genmask = nft_genmask_next(ctx->net);
10495         struct nlattr *tb[NFTA_VERDICT_MAX + 1];
10496         struct nft_chain *chain;
10497         int err;
10498
10499         err = nla_parse_nested_deprecated(tb, NFTA_VERDICT_MAX, nla,
10500                                           nft_verdict_policy, NULL);
10501         if (err < 0)
10502                 return err;
10503
10504         if (!tb[NFTA_VERDICT_CODE])
10505                 return -EINVAL;
10506
10507         /* zero padding hole for memcmp */
10508         memset(data, 0, sizeof(*data));
10509         data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
10510
10511         switch (data->verdict.code) {
10512         case NF_ACCEPT:
10513         case NF_DROP:
10514         case NF_QUEUE:
10515                 break;
10516         case NFT_CONTINUE:
10517         case NFT_BREAK:
10518         case NFT_RETURN:
10519                 break;
10520         case NFT_JUMP:
10521         case NFT_GOTO:
10522                 if (tb[NFTA_VERDICT_CHAIN]) {
10523                         chain = nft_chain_lookup(ctx->net, ctx->table,
10524                                                  tb[NFTA_VERDICT_CHAIN],
10525                                                  genmask);
10526                 } else if (tb[NFTA_VERDICT_CHAIN_ID]) {
10527                         chain = nft_chain_lookup_byid(ctx->net, ctx->table,
10528                                                       tb[NFTA_VERDICT_CHAIN_ID],
10529                                                       genmask);
10530                         if (IS_ERR(chain))
10531                                 return PTR_ERR(chain);
10532                 } else {
10533                         return -EINVAL;
10534                 }
10535
10536                 if (IS_ERR(chain))
10537                         return PTR_ERR(chain);
10538                 if (nft_is_base_chain(chain))
10539                         return -EOPNOTSUPP;
10540                 if (nft_chain_is_bound(chain))
10541                         return -EINVAL;
10542                 if (desc->flags & NFT_DATA_DESC_SETELEM &&
10543                     chain->flags & NFT_CHAIN_BINDING)
10544                         return -EINVAL;
10545                 if (!nft_use_inc(&chain->use))
10546                         return -EMFILE;
10547
10548                 data->verdict.chain = chain;
10549                 break;
10550         default:
10551                 return -EINVAL;
10552         }
10553
10554         desc->len = sizeof(data->verdict);
10555
10556         return 0;
10557 }
10558
10559 static void nft_verdict_uninit(const struct nft_data *data)
10560 {
10561         struct nft_chain *chain;
10562
10563         switch (data->verdict.code) {
10564         case NFT_JUMP:
10565         case NFT_GOTO:
10566                 chain = data->verdict.chain;
10567                 nft_use_dec(&chain->use);
10568                 break;
10569         }
10570 }
10571
10572 int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
10573 {
10574         struct nlattr *nest;
10575
10576         nest = nla_nest_start_noflag(skb, type);
10577         if (!nest)
10578                 goto nla_put_failure;
10579
10580         if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
10581                 goto nla_put_failure;
10582
10583         switch (v->code) {
10584         case NFT_JUMP:
10585         case NFT_GOTO:
10586                 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
10587                                    v->chain->name))
10588                         goto nla_put_failure;
10589         }
10590         nla_nest_end(skb, nest);
10591         return 0;
10592
10593 nla_put_failure:
10594         return -1;
10595 }
10596
10597 static int nft_value_init(const struct nft_ctx *ctx,
10598                           struct nft_data *data, struct nft_data_desc *desc,
10599                           const struct nlattr *nla)
10600 {
10601         unsigned int len;
10602
10603         len = nla_len(nla);
10604         if (len == 0)
10605                 return -EINVAL;
10606         if (len > desc->size)
10607                 return -EOVERFLOW;
10608         if (desc->len) {
10609                 if (len != desc->len)
10610                         return -EINVAL;
10611         } else {
10612                 desc->len = len;
10613         }
10614
10615         nla_memcpy(data->data, nla, len);
10616
10617         return 0;
10618 }
10619
10620 static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
10621                           unsigned int len)
10622 {
10623         return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
10624 }
10625
10626 static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
10627         [NFTA_DATA_VALUE]       = { .type = NLA_BINARY },
10628         [NFTA_DATA_VERDICT]     = { .type = NLA_NESTED },
10629 };
10630
10631 /**
10632  *      nft_data_init - parse nf_tables data netlink attributes
10633  *
10634  *      @ctx: context of the expression using the data
10635  *      @data: destination struct nft_data
10636  *      @desc: data description
10637  *      @nla: netlink attribute containing data
10638  *
10639  *      Parse the netlink data attributes and initialize a struct nft_data.
10640  *      The type and length of data are returned in the data description.
10641  *
10642  *      The caller can indicate that it only wants to accept data of type
10643  *      NFT_DATA_VALUE by passing NULL for the ctx argument.
10644  */
10645 int nft_data_init(const struct nft_ctx *ctx, struct nft_data *data,
10646                   struct nft_data_desc *desc, const struct nlattr *nla)
10647 {
10648         struct nlattr *tb[NFTA_DATA_MAX + 1];
10649         int err;
10650
10651         if (WARN_ON_ONCE(!desc->size))
10652                 return -EINVAL;
10653
10654         err = nla_parse_nested_deprecated(tb, NFTA_DATA_MAX, nla,
10655                                           nft_data_policy, NULL);
10656         if (err < 0)
10657                 return err;
10658
10659         if (tb[NFTA_DATA_VALUE]) {
10660                 if (desc->type != NFT_DATA_VALUE)
10661                         return -EINVAL;
10662
10663                 err = nft_value_init(ctx, data, desc, tb[NFTA_DATA_VALUE]);
10664         } else if (tb[NFTA_DATA_VERDICT] && ctx != NULL) {
10665                 if (desc->type != NFT_DATA_VERDICT)
10666                         return -EINVAL;
10667
10668                 err = nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
10669         } else {
10670                 err = -EINVAL;
10671         }
10672
10673         return err;
10674 }
10675 EXPORT_SYMBOL_GPL(nft_data_init);
10676
10677 /**
10678  *      nft_data_release - release a nft_data item
10679  *
10680  *      @data: struct nft_data to release
10681  *      @type: type of data
10682  *
10683  *      Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
10684  *      all others need to be released by calling this function.
10685  */
10686 void nft_data_release(const struct nft_data *data, enum nft_data_types type)
10687 {
10688         if (type < NFT_DATA_VERDICT)
10689                 return;
10690         switch (type) {
10691         case NFT_DATA_VERDICT:
10692                 return nft_verdict_uninit(data);
10693         default:
10694                 WARN_ON(1);
10695         }
10696 }
10697 EXPORT_SYMBOL_GPL(nft_data_release);
10698
10699 int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
10700                   enum nft_data_types type, unsigned int len)
10701 {
10702         struct nlattr *nest;
10703         int err;
10704
10705         nest = nla_nest_start_noflag(skb, attr);
10706         if (nest == NULL)
10707                 return -1;
10708
10709         switch (type) {
10710         case NFT_DATA_VALUE:
10711                 err = nft_value_dump(skb, data, len);
10712                 break;
10713         case NFT_DATA_VERDICT:
10714                 err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
10715                 break;
10716         default:
10717                 err = -EINVAL;
10718                 WARN_ON(1);
10719         }
10720
10721         nla_nest_end(skb, nest);
10722         return err;
10723 }
10724 EXPORT_SYMBOL_GPL(nft_data_dump);
10725
10726 int __nft_release_basechain(struct nft_ctx *ctx)
10727 {
10728         struct nft_rule *rule, *nr;
10729
10730         if (WARN_ON(!nft_is_base_chain(ctx->chain)))
10731                 return 0;
10732
10733         nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
10734         list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
10735                 list_del(&rule->list);
10736                 nft_use_dec(&ctx->chain->use);
10737                 nf_tables_rule_release(ctx, rule);
10738         }
10739         nft_chain_del(ctx->chain);
10740         nft_use_dec(&ctx->table->use);
10741         nf_tables_chain_destroy(ctx);
10742
10743         return 0;
10744 }
10745 EXPORT_SYMBOL_GPL(__nft_release_basechain);
10746
10747 static void __nft_release_hook(struct net *net, struct nft_table *table)
10748 {
10749         struct nft_flowtable *flowtable;
10750         struct nft_chain *chain;
10751
10752         list_for_each_entry(chain, &table->chains, list)
10753                 __nf_tables_unregister_hook(net, table, chain, true);
10754         list_for_each_entry(flowtable, &table->flowtables, list)
10755                 __nft_unregister_flowtable_net_hooks(net, &flowtable->hook_list,
10756                                                      true);
10757 }
10758
10759 static void __nft_release_hooks(struct net *net)
10760 {
10761         struct nftables_pernet *nft_net = nft_pernet(net);
10762         struct nft_table *table;
10763
10764         list_for_each_entry(table, &nft_net->tables, list) {
10765                 if (nft_table_has_owner(table))
10766                         continue;
10767
10768                 __nft_release_hook(net, table);
10769         }
10770 }
10771
10772 static void __nft_release_table(struct net *net, struct nft_table *table)
10773 {
10774         struct nft_flowtable *flowtable, *nf;
10775         struct nft_chain *chain, *nc;
10776         struct nft_object *obj, *ne;
10777         struct nft_rule *rule, *nr;
10778         struct nft_set *set, *ns;
10779         struct nft_ctx ctx = {
10780                 .net    = net,
10781                 .family = NFPROTO_NETDEV,
10782         };
10783
10784         ctx.family = table->family;
10785         ctx.table = table;
10786         list_for_each_entry(chain, &table->chains, list) {
10787                 if (nft_chain_binding(chain))
10788                         continue;
10789
10790                 ctx.chain = chain;
10791                 list_for_each_entry_safe(rule, nr, &chain->rules, list) {
10792                         list_del(&rule->list);
10793                         nft_use_dec(&chain->use);
10794                         nf_tables_rule_release(&ctx, rule);
10795                 }
10796         }
10797         list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
10798                 list_del(&flowtable->list);
10799                 nft_use_dec(&table->use);
10800                 nf_tables_flowtable_destroy(flowtable);
10801         }
10802         list_for_each_entry_safe(set, ns, &table->sets, list) {
10803                 list_del(&set->list);
10804                 nft_use_dec(&table->use);
10805                 if (set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
10806                         nft_map_deactivate(&ctx, set);
10807
10808                 nft_set_destroy(&ctx, set);
10809         }
10810         list_for_each_entry_safe(obj, ne, &table->objects, list) {
10811                 nft_obj_del(obj);
10812                 nft_use_dec(&table->use);
10813                 nft_obj_destroy(&ctx, obj);
10814         }
10815         list_for_each_entry_safe(chain, nc, &table->chains, list) {
10816                 ctx.chain = chain;
10817                 nft_chain_del(chain);
10818                 nft_use_dec(&table->use);
10819                 nf_tables_chain_destroy(&ctx);
10820         }
10821         nf_tables_table_destroy(&ctx);
10822 }
10823
10824 static void __nft_release_tables(struct net *net)
10825 {
10826         struct nftables_pernet *nft_net = nft_pernet(net);
10827         struct nft_table *table, *nt;
10828
10829         list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
10830                 if (nft_table_has_owner(table))
10831                         continue;
10832
10833                 list_del(&table->list);
10834
10835                 __nft_release_table(net, table);
10836         }
10837 }
10838
10839 static int nft_rcv_nl_event(struct notifier_block *this, unsigned long event,
10840                             void *ptr)
10841 {
10842         struct nft_table *table, *to_delete[8];
10843         struct nftables_pernet *nft_net;
10844         struct netlink_notify *n = ptr;
10845         struct net *net = n->net;
10846         unsigned int deleted;
10847         bool restart = false;
10848         unsigned int gc_seq;
10849
10850         if (event != NETLINK_URELEASE || n->protocol != NETLINK_NETFILTER)
10851                 return NOTIFY_DONE;
10852
10853         nft_net = nft_pernet(net);
10854         deleted = 0;
10855         mutex_lock(&nft_net->commit_mutex);
10856
10857         gc_seq = nft_gc_seq_begin(nft_net);
10858
10859         if (!list_empty(&nf_tables_destroy_list))
10860                 nf_tables_trans_destroy_flush_work();
10861 again:
10862         list_for_each_entry(table, &nft_net->tables, list) {
10863                 if (nft_table_has_owner(table) &&
10864                     n->portid == table->nlpid) {
10865                         __nft_release_hook(net, table);
10866                         list_del_rcu(&table->list);
10867                         to_delete[deleted++] = table;
10868                         if (deleted >= ARRAY_SIZE(to_delete))
10869                                 break;
10870                 }
10871         }
10872         if (deleted) {
10873                 restart = deleted >= ARRAY_SIZE(to_delete);
10874                 synchronize_rcu();
10875                 while (deleted)
10876                         __nft_release_table(net, to_delete[--deleted]);
10877
10878                 if (restart)
10879                         goto again;
10880         }
10881         nft_gc_seq_end(nft_net, gc_seq);
10882
10883         mutex_unlock(&nft_net->commit_mutex);
10884
10885         return NOTIFY_DONE;
10886 }
10887
10888 static struct notifier_block nft_nl_notifier = {
10889         .notifier_call  = nft_rcv_nl_event,
10890 };
10891
10892 static int __net_init nf_tables_init_net(struct net *net)
10893 {
10894         struct nftables_pernet *nft_net = nft_pernet(net);
10895
10896         INIT_LIST_HEAD(&nft_net->tables);
10897         INIT_LIST_HEAD(&nft_net->commit_list);
10898         INIT_LIST_HEAD(&nft_net->binding_list);
10899         INIT_LIST_HEAD(&nft_net->module_list);
10900         INIT_LIST_HEAD(&nft_net->notify_list);
10901         mutex_init(&nft_net->commit_mutex);
10902         nft_net->base_seq = 1;
10903         nft_net->validate_state = NFT_VALIDATE_SKIP;
10904         nft_net->gc_seq = 0;
10905
10906         return 0;
10907 }
10908
10909 static void __net_exit nf_tables_pre_exit_net(struct net *net)
10910 {
10911         struct nftables_pernet *nft_net = nft_pernet(net);
10912
10913         mutex_lock(&nft_net->commit_mutex);
10914         __nft_release_hooks(net);
10915         mutex_unlock(&nft_net->commit_mutex);
10916 }
10917
10918 static void __net_exit nf_tables_exit_net(struct net *net)
10919 {
10920         struct nftables_pernet *nft_net = nft_pernet(net);
10921         unsigned int gc_seq;
10922
10923         mutex_lock(&nft_net->commit_mutex);
10924
10925         gc_seq = nft_gc_seq_begin(nft_net);
10926
10927         WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
10928
10929         if (!list_empty(&nft_net->module_list))
10930                 nf_tables_module_autoload_cleanup(net);
10931
10932         __nft_release_tables(net);
10933
10934         nft_gc_seq_end(nft_net, gc_seq);
10935
10936         mutex_unlock(&nft_net->commit_mutex);
10937         WARN_ON_ONCE(!list_empty(&nft_net->tables));
10938         WARN_ON_ONCE(!list_empty(&nft_net->module_list));
10939         WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
10940 }
10941
10942 static void nf_tables_exit_batch(struct list_head *net_exit_list)
10943 {
10944         flush_work(&trans_gc_work);
10945 }
10946
10947 static struct pernet_operations nf_tables_net_ops = {
10948         .init           = nf_tables_init_net,
10949         .pre_exit       = nf_tables_pre_exit_net,
10950         .exit           = nf_tables_exit_net,
10951         .exit_batch     = nf_tables_exit_batch,
10952         .id             = &nf_tables_net_id,
10953         .size           = sizeof(struct nftables_pernet),
10954 };
10955
10956 static int __init nf_tables_module_init(void)
10957 {
10958         int err;
10959
10960         err = register_pernet_subsys(&nf_tables_net_ops);
10961         if (err < 0)
10962                 return err;
10963
10964         err = nft_chain_filter_init();
10965         if (err < 0)
10966                 goto err_chain_filter;
10967
10968         err = nf_tables_core_module_init();
10969         if (err < 0)
10970                 goto err_core_module;
10971
10972         err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
10973         if (err < 0)
10974                 goto err_netdev_notifier;
10975
10976         err = rhltable_init(&nft_objname_ht, &nft_objname_ht_params);
10977         if (err < 0)
10978                 goto err_rht_objname;
10979
10980         err = nft_offload_init();
10981         if (err < 0)
10982                 goto err_offload;
10983
10984         err = netlink_register_notifier(&nft_nl_notifier);
10985         if (err < 0)
10986                 goto err_netlink_notifier;
10987
10988         /* must be last */
10989         err = nfnetlink_subsys_register(&nf_tables_subsys);
10990         if (err < 0)
10991                 goto err_nfnl_subsys;
10992
10993         nft_chain_route_init();
10994
10995         return err;
10996
10997 err_nfnl_subsys:
10998         netlink_unregister_notifier(&nft_nl_notifier);
10999 err_netlink_notifier:
11000         nft_offload_exit();
11001 err_offload:
11002         rhltable_destroy(&nft_objname_ht);
11003 err_rht_objname:
11004         unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
11005 err_netdev_notifier:
11006         nf_tables_core_module_exit();
11007 err_core_module:
11008         nft_chain_filter_fini();
11009 err_chain_filter:
11010         unregister_pernet_subsys(&nf_tables_net_ops);
11011         return err;
11012 }
11013
11014 static void __exit nf_tables_module_exit(void)
11015 {
11016         nfnetlink_subsys_unregister(&nf_tables_subsys);
11017         netlink_unregister_notifier(&nft_nl_notifier);
11018         nft_offload_exit();
11019         unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
11020         nft_chain_filter_fini();
11021         nft_chain_route_fini();
11022         nf_tables_trans_destroy_flush_work();
11023         unregister_pernet_subsys(&nf_tables_net_ops);
11024         cancel_work_sync(&trans_gc_work);
11025         cancel_work_sync(&trans_destroy_work);
11026         rcu_barrier();
11027         rhltable_destroy(&nft_objname_ht);
11028         nf_tables_core_module_exit();
11029 }
11030
11031 module_init(nf_tables_module_init);
11032 module_exit(nf_tables_module_exit);
11033
11034 MODULE_LICENSE("GPL");
11035 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
11036 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);