]> git.itanic.dy.fi Git - linux-stable/blob - net/netfilter/nf_tables_api.c
netfilter: nf_tables: fix percpu memory leak at nf_tables_addchain()
[linux-stable] / net / 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
28 unsigned int nf_tables_net_id __read_mostly;
29
30 static LIST_HEAD(nf_tables_expressions);
31 static LIST_HEAD(nf_tables_objects);
32 static LIST_HEAD(nf_tables_flowtables);
33 static LIST_HEAD(nf_tables_destroy_list);
34 static DEFINE_SPINLOCK(nf_tables_destroy_list_lock);
35
36 enum {
37         NFT_VALIDATE_SKIP       = 0,
38         NFT_VALIDATE_NEED,
39         NFT_VALIDATE_DO,
40 };
41
42 static struct rhltable nft_objname_ht;
43
44 static u32 nft_chain_hash(const void *data, u32 len, u32 seed);
45 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed);
46 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *, const void *);
47
48 static u32 nft_objname_hash(const void *data, u32 len, u32 seed);
49 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed);
50 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *, const void *);
51
52 static const struct rhashtable_params nft_chain_ht_params = {
53         .head_offset            = offsetof(struct nft_chain, rhlhead),
54         .key_offset             = offsetof(struct nft_chain, name),
55         .hashfn                 = nft_chain_hash,
56         .obj_hashfn             = nft_chain_hash_obj,
57         .obj_cmpfn              = nft_chain_hash_cmp,
58         .automatic_shrinking    = true,
59 };
60
61 static const struct rhashtable_params nft_objname_ht_params = {
62         .head_offset            = offsetof(struct nft_object, rhlhead),
63         .key_offset             = offsetof(struct nft_object, key),
64         .hashfn                 = nft_objname_hash,
65         .obj_hashfn             = nft_objname_hash_obj,
66         .obj_cmpfn              = nft_objname_hash_cmp,
67         .automatic_shrinking    = true,
68 };
69
70 struct nft_audit_data {
71         struct nft_table *table;
72         int entries;
73         int op;
74         struct list_head list;
75 };
76
77 static const u8 nft2audit_op[NFT_MSG_MAX] = { // enum nf_tables_msg_types
78         [NFT_MSG_NEWTABLE]      = AUDIT_NFT_OP_TABLE_REGISTER,
79         [NFT_MSG_GETTABLE]      = AUDIT_NFT_OP_INVALID,
80         [NFT_MSG_DELTABLE]      = AUDIT_NFT_OP_TABLE_UNREGISTER,
81         [NFT_MSG_NEWCHAIN]      = AUDIT_NFT_OP_CHAIN_REGISTER,
82         [NFT_MSG_GETCHAIN]      = AUDIT_NFT_OP_INVALID,
83         [NFT_MSG_DELCHAIN]      = AUDIT_NFT_OP_CHAIN_UNREGISTER,
84         [NFT_MSG_NEWRULE]       = AUDIT_NFT_OP_RULE_REGISTER,
85         [NFT_MSG_GETRULE]       = AUDIT_NFT_OP_INVALID,
86         [NFT_MSG_DELRULE]       = AUDIT_NFT_OP_RULE_UNREGISTER,
87         [NFT_MSG_NEWSET]        = AUDIT_NFT_OP_SET_REGISTER,
88         [NFT_MSG_GETSET]        = AUDIT_NFT_OP_INVALID,
89         [NFT_MSG_DELSET]        = AUDIT_NFT_OP_SET_UNREGISTER,
90         [NFT_MSG_NEWSETELEM]    = AUDIT_NFT_OP_SETELEM_REGISTER,
91         [NFT_MSG_GETSETELEM]    = AUDIT_NFT_OP_INVALID,
92         [NFT_MSG_DELSETELEM]    = AUDIT_NFT_OP_SETELEM_UNREGISTER,
93         [NFT_MSG_NEWGEN]        = AUDIT_NFT_OP_GEN_REGISTER,
94         [NFT_MSG_GETGEN]        = AUDIT_NFT_OP_INVALID,
95         [NFT_MSG_TRACE]         = AUDIT_NFT_OP_INVALID,
96         [NFT_MSG_NEWOBJ]        = AUDIT_NFT_OP_OBJ_REGISTER,
97         [NFT_MSG_GETOBJ]        = AUDIT_NFT_OP_INVALID,
98         [NFT_MSG_DELOBJ]        = AUDIT_NFT_OP_OBJ_UNREGISTER,
99         [NFT_MSG_GETOBJ_RESET]  = AUDIT_NFT_OP_OBJ_RESET,
100         [NFT_MSG_NEWFLOWTABLE]  = AUDIT_NFT_OP_FLOWTABLE_REGISTER,
101         [NFT_MSG_GETFLOWTABLE]  = AUDIT_NFT_OP_INVALID,
102         [NFT_MSG_DELFLOWTABLE]  = AUDIT_NFT_OP_FLOWTABLE_UNREGISTER,
103 };
104
105 static void nft_validate_state_update(struct net *net, u8 new_validate_state)
106 {
107         struct nftables_pernet *nft_net = nft_pernet(net);
108
109         switch (nft_net->validate_state) {
110         case NFT_VALIDATE_SKIP:
111                 WARN_ON_ONCE(new_validate_state == NFT_VALIDATE_DO);
112                 break;
113         case NFT_VALIDATE_NEED:
114                 break;
115         case NFT_VALIDATE_DO:
116                 if (new_validate_state == NFT_VALIDATE_NEED)
117                         return;
118         }
119
120         nft_net->validate_state = new_validate_state;
121 }
122 static void nf_tables_trans_destroy_work(struct work_struct *w);
123 static DECLARE_WORK(trans_destroy_work, nf_tables_trans_destroy_work);
124
125 static void nft_ctx_init(struct nft_ctx *ctx,
126                          struct net *net,
127                          const struct sk_buff *skb,
128                          const struct nlmsghdr *nlh,
129                          u8 family,
130                          struct nft_table *table,
131                          struct nft_chain *chain,
132                          const struct nlattr * const *nla)
133 {
134         ctx->net        = net;
135         ctx->family     = family;
136         ctx->level      = 0;
137         ctx->table      = table;
138         ctx->chain      = chain;
139         ctx->nla        = nla;
140         ctx->portid     = NETLINK_CB(skb).portid;
141         ctx->report     = nlmsg_report(nlh);
142         ctx->flags      = nlh->nlmsg_flags;
143         ctx->seq        = nlh->nlmsg_seq;
144 }
145
146 static struct nft_trans *nft_trans_alloc_gfp(const struct nft_ctx *ctx,
147                                              int msg_type, u32 size, gfp_t gfp)
148 {
149         struct nft_trans *trans;
150
151         trans = kzalloc(sizeof(struct nft_trans) + size, gfp);
152         if (trans == NULL)
153                 return NULL;
154
155         INIT_LIST_HEAD(&trans->list);
156         trans->msg_type = msg_type;
157         trans->ctx      = *ctx;
158
159         return trans;
160 }
161
162 static struct nft_trans *nft_trans_alloc(const struct nft_ctx *ctx,
163                                          int msg_type, u32 size)
164 {
165         return nft_trans_alloc_gfp(ctx, msg_type, size, GFP_KERNEL);
166 }
167
168 static void nft_trans_destroy(struct nft_trans *trans)
169 {
170         list_del(&trans->list);
171         kfree(trans);
172 }
173
174 static void nft_set_trans_bind(const struct nft_ctx *ctx, struct nft_set *set)
175 {
176         struct nftables_pernet *nft_net;
177         struct net *net = ctx->net;
178         struct nft_trans *trans;
179
180         if (!nft_set_is_anonymous(set))
181                 return;
182
183         nft_net = nft_pernet(net);
184         list_for_each_entry_reverse(trans, &nft_net->commit_list, list) {
185                 switch (trans->msg_type) {
186                 case NFT_MSG_NEWSET:
187                         if (nft_trans_set(trans) == set)
188                                 nft_trans_set_bound(trans) = true;
189                         break;
190                 case NFT_MSG_NEWSETELEM:
191                         if (nft_trans_elem_set(trans) == set)
192                                 nft_trans_elem_set_bound(trans) = true;
193                         break;
194                 }
195         }
196 }
197
198 static int nft_netdev_register_hooks(struct net *net,
199                                      struct list_head *hook_list)
200 {
201         struct nft_hook *hook;
202         int err, j;
203
204         j = 0;
205         list_for_each_entry(hook, hook_list, list) {
206                 err = nf_register_net_hook(net, &hook->ops);
207                 if (err < 0)
208                         goto err_register;
209
210                 j++;
211         }
212         return 0;
213
214 err_register:
215         list_for_each_entry(hook, hook_list, list) {
216                 if (j-- <= 0)
217                         break;
218
219                 nf_unregister_net_hook(net, &hook->ops);
220         }
221         return err;
222 }
223
224 static void nft_netdev_unregister_hooks(struct net *net,
225                                         struct list_head *hook_list,
226                                         bool release_netdev)
227 {
228         struct nft_hook *hook, *next;
229
230         list_for_each_entry_safe(hook, next, hook_list, list) {
231                 nf_unregister_net_hook(net, &hook->ops);
232                 if (release_netdev) {
233                         list_del(&hook->list);
234                         kfree_rcu(hook, rcu);
235                 }
236         }
237 }
238
239 static int nf_tables_register_hook(struct net *net,
240                                    const struct nft_table *table,
241                                    struct nft_chain *chain)
242 {
243         struct nft_base_chain *basechain;
244         const struct nf_hook_ops *ops;
245
246         if (table->flags & NFT_TABLE_F_DORMANT ||
247             !nft_is_base_chain(chain))
248                 return 0;
249
250         basechain = nft_base_chain(chain);
251         ops = &basechain->ops;
252
253         if (basechain->type->ops_register)
254                 return basechain->type->ops_register(net, ops);
255
256         if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
257                 return nft_netdev_register_hooks(net, &basechain->hook_list);
258
259         return nf_register_net_hook(net, &basechain->ops);
260 }
261
262 static void __nf_tables_unregister_hook(struct net *net,
263                                         const struct nft_table *table,
264                                         struct nft_chain *chain,
265                                         bool release_netdev)
266 {
267         struct nft_base_chain *basechain;
268         const struct nf_hook_ops *ops;
269
270         if (table->flags & NFT_TABLE_F_DORMANT ||
271             !nft_is_base_chain(chain))
272                 return;
273         basechain = nft_base_chain(chain);
274         ops = &basechain->ops;
275
276         if (basechain->type->ops_unregister)
277                 return basechain->type->ops_unregister(net, ops);
278
279         if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
280                 nft_netdev_unregister_hooks(net, &basechain->hook_list,
281                                             release_netdev);
282         else
283                 nf_unregister_net_hook(net, &basechain->ops);
284 }
285
286 static void nf_tables_unregister_hook(struct net *net,
287                                       const struct nft_table *table,
288                                       struct nft_chain *chain)
289 {
290         return __nf_tables_unregister_hook(net, table, chain, false);
291 }
292
293 static void nft_trans_commit_list_add_tail(struct net *net, struct nft_trans *trans)
294 {
295         struct nftables_pernet *nft_net = nft_pernet(net);
296
297         list_add_tail(&trans->list, &nft_net->commit_list);
298 }
299
300 static int nft_trans_table_add(struct nft_ctx *ctx, int msg_type)
301 {
302         struct nft_trans *trans;
303
304         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_table));
305         if (trans == NULL)
306                 return -ENOMEM;
307
308         if (msg_type == NFT_MSG_NEWTABLE)
309                 nft_activate_next(ctx->net, ctx->table);
310
311         nft_trans_commit_list_add_tail(ctx->net, trans);
312         return 0;
313 }
314
315 static int nft_deltable(struct nft_ctx *ctx)
316 {
317         int err;
318
319         err = nft_trans_table_add(ctx, NFT_MSG_DELTABLE);
320         if (err < 0)
321                 return err;
322
323         nft_deactivate_next(ctx->net, ctx->table);
324         return err;
325 }
326
327 static struct nft_trans *nft_trans_chain_add(struct nft_ctx *ctx, int msg_type)
328 {
329         struct nft_trans *trans;
330
331         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_chain));
332         if (trans == NULL)
333                 return ERR_PTR(-ENOMEM);
334
335         if (msg_type == NFT_MSG_NEWCHAIN) {
336                 nft_activate_next(ctx->net, ctx->chain);
337
338                 if (ctx->nla[NFTA_CHAIN_ID]) {
339                         nft_trans_chain_id(trans) =
340                                 ntohl(nla_get_be32(ctx->nla[NFTA_CHAIN_ID]));
341                 }
342         }
343
344         nft_trans_commit_list_add_tail(ctx->net, trans);
345         return trans;
346 }
347
348 static int nft_delchain(struct nft_ctx *ctx)
349 {
350         struct nft_trans *trans;
351
352         trans = nft_trans_chain_add(ctx, NFT_MSG_DELCHAIN);
353         if (IS_ERR(trans))
354                 return PTR_ERR(trans);
355
356         ctx->table->use--;
357         nft_deactivate_next(ctx->net, ctx->chain);
358
359         return 0;
360 }
361
362 static void nft_rule_expr_activate(const struct nft_ctx *ctx,
363                                    struct nft_rule *rule)
364 {
365         struct nft_expr *expr;
366
367         expr = nft_expr_first(rule);
368         while (nft_expr_more(rule, expr)) {
369                 if (expr->ops->activate)
370                         expr->ops->activate(ctx, expr);
371
372                 expr = nft_expr_next(expr);
373         }
374 }
375
376 static void nft_rule_expr_deactivate(const struct nft_ctx *ctx,
377                                      struct nft_rule *rule,
378                                      enum nft_trans_phase phase)
379 {
380         struct nft_expr *expr;
381
382         expr = nft_expr_first(rule);
383         while (nft_expr_more(rule, expr)) {
384                 if (expr->ops->deactivate)
385                         expr->ops->deactivate(ctx, expr, phase);
386
387                 expr = nft_expr_next(expr);
388         }
389 }
390
391 static int
392 nf_tables_delrule_deactivate(struct nft_ctx *ctx, struct nft_rule *rule)
393 {
394         /* You cannot delete the same rule twice */
395         if (nft_is_active_next(ctx->net, rule)) {
396                 nft_deactivate_next(ctx->net, rule);
397                 ctx->chain->use--;
398                 return 0;
399         }
400         return -ENOENT;
401 }
402
403 static struct nft_trans *nft_trans_rule_add(struct nft_ctx *ctx, int msg_type,
404                                             struct nft_rule *rule)
405 {
406         struct nft_trans *trans;
407
408         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_rule));
409         if (trans == NULL)
410                 return NULL;
411
412         if (msg_type == NFT_MSG_NEWRULE && ctx->nla[NFTA_RULE_ID] != NULL) {
413                 nft_trans_rule_id(trans) =
414                         ntohl(nla_get_be32(ctx->nla[NFTA_RULE_ID]));
415         }
416         nft_trans_rule(trans) = rule;
417         nft_trans_commit_list_add_tail(ctx->net, trans);
418
419         return trans;
420 }
421
422 static int nft_delrule(struct nft_ctx *ctx, struct nft_rule *rule)
423 {
424         struct nft_flow_rule *flow;
425         struct nft_trans *trans;
426         int err;
427
428         trans = nft_trans_rule_add(ctx, NFT_MSG_DELRULE, rule);
429         if (trans == NULL)
430                 return -ENOMEM;
431
432         if (ctx->chain->flags & NFT_CHAIN_HW_OFFLOAD) {
433                 flow = nft_flow_rule_create(ctx->net, rule);
434                 if (IS_ERR(flow)) {
435                         nft_trans_destroy(trans);
436                         return PTR_ERR(flow);
437                 }
438
439                 nft_trans_flow_rule(trans) = flow;
440         }
441
442         err = nf_tables_delrule_deactivate(ctx, rule);
443         if (err < 0) {
444                 nft_trans_destroy(trans);
445                 return err;
446         }
447         nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_PREPARE);
448
449         return 0;
450 }
451
452 static int nft_delrule_by_chain(struct nft_ctx *ctx)
453 {
454         struct nft_rule *rule;
455         int err;
456
457         list_for_each_entry(rule, &ctx->chain->rules, list) {
458                 if (!nft_is_active_next(ctx->net, rule))
459                         continue;
460
461                 err = nft_delrule(ctx, rule);
462                 if (err < 0)
463                         return err;
464         }
465         return 0;
466 }
467
468 static int nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
469                              struct nft_set *set)
470 {
471         struct nft_trans *trans;
472
473         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_set));
474         if (trans == NULL)
475                 return -ENOMEM;
476
477         if (msg_type == NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID] != NULL) {
478                 nft_trans_set_id(trans) =
479                         ntohl(nla_get_be32(ctx->nla[NFTA_SET_ID]));
480                 nft_activate_next(ctx->net, set);
481         }
482         nft_trans_set(trans) = set;
483         nft_trans_commit_list_add_tail(ctx->net, trans);
484
485         return 0;
486 }
487
488 static int nft_delset(const struct nft_ctx *ctx, struct nft_set *set)
489 {
490         int err;
491
492         err = nft_trans_set_add(ctx, NFT_MSG_DELSET, set);
493         if (err < 0)
494                 return err;
495
496         nft_deactivate_next(ctx->net, set);
497         ctx->table->use--;
498
499         return err;
500 }
501
502 static int nft_trans_obj_add(struct nft_ctx *ctx, int msg_type,
503                              struct nft_object *obj)
504 {
505         struct nft_trans *trans;
506
507         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_obj));
508         if (trans == NULL)
509                 return -ENOMEM;
510
511         if (msg_type == NFT_MSG_NEWOBJ)
512                 nft_activate_next(ctx->net, obj);
513
514         nft_trans_obj(trans) = obj;
515         nft_trans_commit_list_add_tail(ctx->net, trans);
516
517         return 0;
518 }
519
520 static int nft_delobj(struct nft_ctx *ctx, struct nft_object *obj)
521 {
522         int err;
523
524         err = nft_trans_obj_add(ctx, NFT_MSG_DELOBJ, obj);
525         if (err < 0)
526                 return err;
527
528         nft_deactivate_next(ctx->net, obj);
529         ctx->table->use--;
530
531         return err;
532 }
533
534 static int nft_trans_flowtable_add(struct nft_ctx *ctx, int msg_type,
535                                    struct nft_flowtable *flowtable)
536 {
537         struct nft_trans *trans;
538
539         trans = nft_trans_alloc(ctx, msg_type,
540                                 sizeof(struct nft_trans_flowtable));
541         if (trans == NULL)
542                 return -ENOMEM;
543
544         if (msg_type == NFT_MSG_NEWFLOWTABLE)
545                 nft_activate_next(ctx->net, flowtable);
546
547         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
548         nft_trans_flowtable(trans) = flowtable;
549         nft_trans_commit_list_add_tail(ctx->net, trans);
550
551         return 0;
552 }
553
554 static int nft_delflowtable(struct nft_ctx *ctx,
555                             struct nft_flowtable *flowtable)
556 {
557         int err;
558
559         err = nft_trans_flowtable_add(ctx, NFT_MSG_DELFLOWTABLE, flowtable);
560         if (err < 0)
561                 return err;
562
563         nft_deactivate_next(ctx->net, flowtable);
564         ctx->table->use--;
565
566         return err;
567 }
568
569 static void __nft_reg_track_clobber(struct nft_regs_track *track, u8 dreg)
570 {
571         int i;
572
573         for (i = track->regs[dreg].num_reg; i > 0; i--)
574                 __nft_reg_track_cancel(track, dreg - i);
575 }
576
577 static void __nft_reg_track_update(struct nft_regs_track *track,
578                                    const struct nft_expr *expr,
579                                    u8 dreg, u8 num_reg)
580 {
581         track->regs[dreg].selector = expr;
582         track->regs[dreg].bitwise = NULL;
583         track->regs[dreg].num_reg = num_reg;
584 }
585
586 void nft_reg_track_update(struct nft_regs_track *track,
587                           const struct nft_expr *expr, u8 dreg, u8 len)
588 {
589         unsigned int regcount;
590         int i;
591
592         __nft_reg_track_clobber(track, dreg);
593
594         regcount = DIV_ROUND_UP(len, NFT_REG32_SIZE);
595         for (i = 0; i < regcount; i++, dreg++)
596                 __nft_reg_track_update(track, expr, dreg, i);
597 }
598 EXPORT_SYMBOL_GPL(nft_reg_track_update);
599
600 void nft_reg_track_cancel(struct nft_regs_track *track, u8 dreg, u8 len)
601 {
602         unsigned int regcount;
603         int i;
604
605         __nft_reg_track_clobber(track, dreg);
606
607         regcount = DIV_ROUND_UP(len, NFT_REG32_SIZE);
608         for (i = 0; i < regcount; i++, dreg++)
609                 __nft_reg_track_cancel(track, dreg);
610 }
611 EXPORT_SYMBOL_GPL(nft_reg_track_cancel);
612
613 void __nft_reg_track_cancel(struct nft_regs_track *track, u8 dreg)
614 {
615         track->regs[dreg].selector = NULL;
616         track->regs[dreg].bitwise = NULL;
617         track->regs[dreg].num_reg = 0;
618 }
619 EXPORT_SYMBOL_GPL(__nft_reg_track_cancel);
620
621 /*
622  * Tables
623  */
624
625 static struct nft_table *nft_table_lookup(const struct net *net,
626                                           const struct nlattr *nla,
627                                           u8 family, u8 genmask, u32 nlpid)
628 {
629         struct nftables_pernet *nft_net;
630         struct nft_table *table;
631
632         if (nla == NULL)
633                 return ERR_PTR(-EINVAL);
634
635         nft_net = nft_pernet(net);
636         list_for_each_entry_rcu(table, &nft_net->tables, list,
637                                 lockdep_is_held(&nft_net->commit_mutex)) {
638                 if (!nla_strcmp(nla, table->name) &&
639                     table->family == family &&
640                     nft_active_genmask(table, genmask)) {
641                         if (nft_table_has_owner(table) &&
642                             nlpid && table->nlpid != nlpid)
643                                 return ERR_PTR(-EPERM);
644
645                         return table;
646                 }
647         }
648
649         return ERR_PTR(-ENOENT);
650 }
651
652 static struct nft_table *nft_table_lookup_byhandle(const struct net *net,
653                                                    const struct nlattr *nla,
654                                                    u8 genmask, u32 nlpid)
655 {
656         struct nftables_pernet *nft_net;
657         struct nft_table *table;
658
659         nft_net = nft_pernet(net);
660         list_for_each_entry(table, &nft_net->tables, list) {
661                 if (be64_to_cpu(nla_get_be64(nla)) == table->handle &&
662                     nft_active_genmask(table, genmask)) {
663                         if (nft_table_has_owner(table) &&
664                             nlpid && table->nlpid != nlpid)
665                                 return ERR_PTR(-EPERM);
666
667                         return table;
668                 }
669         }
670
671         return ERR_PTR(-ENOENT);
672 }
673
674 static inline u64 nf_tables_alloc_handle(struct nft_table *table)
675 {
676         return ++table->hgenerator;
677 }
678
679 static const struct nft_chain_type *chain_type[NFPROTO_NUMPROTO][NFT_CHAIN_T_MAX];
680
681 static const struct nft_chain_type *
682 __nft_chain_type_get(u8 family, enum nft_chain_types type)
683 {
684         if (family >= NFPROTO_NUMPROTO ||
685             type >= NFT_CHAIN_T_MAX)
686                 return NULL;
687
688         return chain_type[family][type];
689 }
690
691 static const struct nft_chain_type *
692 __nf_tables_chain_type_lookup(const struct nlattr *nla, u8 family)
693 {
694         const struct nft_chain_type *type;
695         int i;
696
697         for (i = 0; i < NFT_CHAIN_T_MAX; i++) {
698                 type = __nft_chain_type_get(family, i);
699                 if (!type)
700                         continue;
701                 if (!nla_strcmp(nla, type->name))
702                         return type;
703         }
704         return NULL;
705 }
706
707 struct nft_module_request {
708         struct list_head        list;
709         char                    module[MODULE_NAME_LEN];
710         bool                    done;
711 };
712
713 #ifdef CONFIG_MODULES
714 __printf(2, 3) int nft_request_module(struct net *net, const char *fmt,
715                                       ...)
716 {
717         char module_name[MODULE_NAME_LEN];
718         struct nftables_pernet *nft_net;
719         struct nft_module_request *req;
720         va_list args;
721         int ret;
722
723         va_start(args, fmt);
724         ret = vsnprintf(module_name, MODULE_NAME_LEN, fmt, args);
725         va_end(args);
726         if (ret >= MODULE_NAME_LEN)
727                 return 0;
728
729         nft_net = nft_pernet(net);
730         list_for_each_entry(req, &nft_net->module_list, list) {
731                 if (!strcmp(req->module, module_name)) {
732                         if (req->done)
733                                 return 0;
734
735                         /* A request to load this module already exists. */
736                         return -EAGAIN;
737                 }
738         }
739
740         req = kmalloc(sizeof(*req), GFP_KERNEL);
741         if (!req)
742                 return -ENOMEM;
743
744         req->done = false;
745         strlcpy(req->module, module_name, MODULE_NAME_LEN);
746         list_add_tail(&req->list, &nft_net->module_list);
747
748         return -EAGAIN;
749 }
750 EXPORT_SYMBOL_GPL(nft_request_module);
751 #endif
752
753 static void lockdep_nfnl_nft_mutex_not_held(void)
754 {
755 #ifdef CONFIG_PROVE_LOCKING
756         if (debug_locks)
757                 WARN_ON_ONCE(lockdep_nfnl_is_held(NFNL_SUBSYS_NFTABLES));
758 #endif
759 }
760
761 static const struct nft_chain_type *
762 nf_tables_chain_type_lookup(struct net *net, const struct nlattr *nla,
763                             u8 family, bool autoload)
764 {
765         const struct nft_chain_type *type;
766
767         type = __nf_tables_chain_type_lookup(nla, family);
768         if (type != NULL)
769                 return type;
770
771         lockdep_nfnl_nft_mutex_not_held();
772 #ifdef CONFIG_MODULES
773         if (autoload) {
774                 if (nft_request_module(net, "nft-chain-%u-%.*s", family,
775                                        nla_len(nla),
776                                        (const char *)nla_data(nla)) == -EAGAIN)
777                         return ERR_PTR(-EAGAIN);
778         }
779 #endif
780         return ERR_PTR(-ENOENT);
781 }
782
783 static __be16 nft_base_seq(const struct net *net)
784 {
785         struct nftables_pernet *nft_net = nft_pernet(net);
786
787         return htons(nft_net->base_seq & 0xffff);
788 }
789
790 static const struct nla_policy nft_table_policy[NFTA_TABLE_MAX + 1] = {
791         [NFTA_TABLE_NAME]       = { .type = NLA_STRING,
792                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
793         [NFTA_TABLE_FLAGS]      = { .type = NLA_U32 },
794         [NFTA_TABLE_HANDLE]     = { .type = NLA_U64 },
795         [NFTA_TABLE_USERDATA]   = { .type = NLA_BINARY,
796                                     .len = NFT_USERDATA_MAXLEN }
797 };
798
799 static int nf_tables_fill_table_info(struct sk_buff *skb, struct net *net,
800                                      u32 portid, u32 seq, int event, u32 flags,
801                                      int family, const struct nft_table *table)
802 {
803         struct nlmsghdr *nlh;
804
805         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
806         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
807                            NFNETLINK_V0, nft_base_seq(net));
808         if (!nlh)
809                 goto nla_put_failure;
810
811         if (nla_put_string(skb, NFTA_TABLE_NAME, table->name) ||
812             nla_put_be32(skb, NFTA_TABLE_FLAGS,
813                          htonl(table->flags & NFT_TABLE_F_MASK)) ||
814             nla_put_be32(skb, NFTA_TABLE_USE, htonl(table->use)) ||
815             nla_put_be64(skb, NFTA_TABLE_HANDLE, cpu_to_be64(table->handle),
816                          NFTA_TABLE_PAD))
817                 goto nla_put_failure;
818         if (nft_table_has_owner(table) &&
819             nla_put_be32(skb, NFTA_TABLE_OWNER, htonl(table->nlpid)))
820                 goto nla_put_failure;
821
822         if (table->udata) {
823                 if (nla_put(skb, NFTA_TABLE_USERDATA, table->udlen, table->udata))
824                         goto nla_put_failure;
825         }
826
827         nlmsg_end(skb, nlh);
828         return 0;
829
830 nla_put_failure:
831         nlmsg_trim(skb, nlh);
832         return -1;
833 }
834
835 struct nftnl_skb_parms {
836         bool report;
837 };
838 #define NFT_CB(skb)     (*(struct nftnl_skb_parms*)&((skb)->cb))
839
840 static void nft_notify_enqueue(struct sk_buff *skb, bool report,
841                                struct list_head *notify_list)
842 {
843         NFT_CB(skb).report = report;
844         list_add_tail(&skb->list, notify_list);
845 }
846
847 static void nf_tables_table_notify(const struct nft_ctx *ctx, int event)
848 {
849         struct nftables_pernet *nft_net;
850         struct sk_buff *skb;
851         u16 flags = 0;
852         int err;
853
854         if (!ctx->report &&
855             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
856                 return;
857
858         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
859         if (skb == NULL)
860                 goto err;
861
862         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
863                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
864
865         err = nf_tables_fill_table_info(skb, ctx->net, ctx->portid, ctx->seq,
866                                         event, flags, ctx->family, ctx->table);
867         if (err < 0) {
868                 kfree_skb(skb);
869                 goto err;
870         }
871
872         nft_net = nft_pernet(ctx->net);
873         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
874         return;
875 err:
876         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
877 }
878
879 static int nf_tables_dump_tables(struct sk_buff *skb,
880                                  struct netlink_callback *cb)
881 {
882         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
883         struct nftables_pernet *nft_net;
884         const struct nft_table *table;
885         unsigned int idx = 0, s_idx = cb->args[0];
886         struct net *net = sock_net(skb->sk);
887         int family = nfmsg->nfgen_family;
888
889         rcu_read_lock();
890         nft_net = nft_pernet(net);
891         cb->seq = READ_ONCE(nft_net->base_seq);
892
893         list_for_each_entry_rcu(table, &nft_net->tables, list) {
894                 if (family != NFPROTO_UNSPEC && family != table->family)
895                         continue;
896
897                 if (idx < s_idx)
898                         goto cont;
899                 if (idx > s_idx)
900                         memset(&cb->args[1], 0,
901                                sizeof(cb->args) - sizeof(cb->args[0]));
902                 if (!nft_is_active(net, table))
903                         continue;
904                 if (nf_tables_fill_table_info(skb, net,
905                                               NETLINK_CB(cb->skb).portid,
906                                               cb->nlh->nlmsg_seq,
907                                               NFT_MSG_NEWTABLE, NLM_F_MULTI,
908                                               table->family, table) < 0)
909                         goto done;
910
911                 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
912 cont:
913                 idx++;
914         }
915 done:
916         rcu_read_unlock();
917         cb->args[0] = idx;
918         return skb->len;
919 }
920
921 static int nft_netlink_dump_start_rcu(struct sock *nlsk, struct sk_buff *skb,
922                                       const struct nlmsghdr *nlh,
923                                       struct netlink_dump_control *c)
924 {
925         int err;
926
927         if (!try_module_get(THIS_MODULE))
928                 return -EINVAL;
929
930         rcu_read_unlock();
931         err = netlink_dump_start(nlsk, skb, nlh, c);
932         rcu_read_lock();
933         module_put(THIS_MODULE);
934
935         return err;
936 }
937
938 /* called with rcu_read_lock held */
939 static int nf_tables_gettable(struct sk_buff *skb, const struct nfnl_info *info,
940                               const struct nlattr * const nla[])
941 {
942         struct netlink_ext_ack *extack = info->extack;
943         u8 genmask = nft_genmask_cur(info->net);
944         u8 family = info->nfmsg->nfgen_family;
945         const struct nft_table *table;
946         struct net *net = info->net;
947         struct sk_buff *skb2;
948         int err;
949
950         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
951                 struct netlink_dump_control c = {
952                         .dump = nf_tables_dump_tables,
953                         .module = THIS_MODULE,
954                 };
955
956                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
957         }
958
959         table = nft_table_lookup(net, nla[NFTA_TABLE_NAME], family, genmask, 0);
960         if (IS_ERR(table)) {
961                 NL_SET_BAD_ATTR(extack, nla[NFTA_TABLE_NAME]);
962                 return PTR_ERR(table);
963         }
964
965         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
966         if (!skb2)
967                 return -ENOMEM;
968
969         err = nf_tables_fill_table_info(skb2, net, NETLINK_CB(skb).portid,
970                                         info->nlh->nlmsg_seq, NFT_MSG_NEWTABLE,
971                                         0, family, table);
972         if (err < 0)
973                 goto err_fill_table_info;
974
975         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
976
977 err_fill_table_info:
978         kfree_skb(skb2);
979         return err;
980 }
981
982 static void nft_table_disable(struct net *net, struct nft_table *table, u32 cnt)
983 {
984         struct nft_chain *chain;
985         u32 i = 0;
986
987         list_for_each_entry(chain, &table->chains, list) {
988                 if (!nft_is_active_next(net, chain))
989                         continue;
990                 if (!nft_is_base_chain(chain))
991                         continue;
992
993                 if (cnt && i++ == cnt)
994                         break;
995
996                 nf_tables_unregister_hook(net, table, chain);
997         }
998 }
999
1000 static int nf_tables_table_enable(struct net *net, struct nft_table *table)
1001 {
1002         struct nft_chain *chain;
1003         int err, i = 0;
1004
1005         list_for_each_entry(chain, &table->chains, list) {
1006                 if (!nft_is_active_next(net, chain))
1007                         continue;
1008                 if (!nft_is_base_chain(chain))
1009                         continue;
1010
1011                 err = nf_tables_register_hook(net, table, chain);
1012                 if (err < 0)
1013                         goto err_register_hooks;
1014
1015                 i++;
1016         }
1017         return 0;
1018
1019 err_register_hooks:
1020         if (i)
1021                 nft_table_disable(net, table, i);
1022         return err;
1023 }
1024
1025 static void nf_tables_table_disable(struct net *net, struct nft_table *table)
1026 {
1027         table->flags &= ~NFT_TABLE_F_DORMANT;
1028         nft_table_disable(net, table, 0);
1029         table->flags |= NFT_TABLE_F_DORMANT;
1030 }
1031
1032 #define __NFT_TABLE_F_INTERNAL          (NFT_TABLE_F_MASK + 1)
1033 #define __NFT_TABLE_F_WAS_DORMANT       (__NFT_TABLE_F_INTERNAL << 0)
1034 #define __NFT_TABLE_F_WAS_AWAKEN        (__NFT_TABLE_F_INTERNAL << 1)
1035 #define __NFT_TABLE_F_UPDATE            (__NFT_TABLE_F_WAS_DORMANT | \
1036                                          __NFT_TABLE_F_WAS_AWAKEN)
1037
1038 static int nf_tables_updtable(struct nft_ctx *ctx)
1039 {
1040         struct nft_trans *trans;
1041         u32 flags;
1042         int ret;
1043
1044         if (!ctx->nla[NFTA_TABLE_FLAGS])
1045                 return 0;
1046
1047         flags = ntohl(nla_get_be32(ctx->nla[NFTA_TABLE_FLAGS]));
1048         if (flags & ~NFT_TABLE_F_MASK)
1049                 return -EOPNOTSUPP;
1050
1051         if (flags == ctx->table->flags)
1052                 return 0;
1053
1054         if ((nft_table_has_owner(ctx->table) &&
1055              !(flags & NFT_TABLE_F_OWNER)) ||
1056             (!nft_table_has_owner(ctx->table) &&
1057              flags & NFT_TABLE_F_OWNER))
1058                 return -EOPNOTSUPP;
1059
1060         trans = nft_trans_alloc(ctx, NFT_MSG_NEWTABLE,
1061                                 sizeof(struct nft_trans_table));
1062         if (trans == NULL)
1063                 return -ENOMEM;
1064
1065         if ((flags & NFT_TABLE_F_DORMANT) &&
1066             !(ctx->table->flags & NFT_TABLE_F_DORMANT)) {
1067                 ctx->table->flags |= NFT_TABLE_F_DORMANT;
1068                 if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE))
1069                         ctx->table->flags |= __NFT_TABLE_F_WAS_AWAKEN;
1070         } else if (!(flags & NFT_TABLE_F_DORMANT) &&
1071                    ctx->table->flags & NFT_TABLE_F_DORMANT) {
1072                 ctx->table->flags &= ~NFT_TABLE_F_DORMANT;
1073                 if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE)) {
1074                         ret = nf_tables_table_enable(ctx->net, ctx->table);
1075                         if (ret < 0)
1076                                 goto err_register_hooks;
1077
1078                         ctx->table->flags |= __NFT_TABLE_F_WAS_DORMANT;
1079                 }
1080         }
1081
1082         nft_trans_table_update(trans) = true;
1083         nft_trans_commit_list_add_tail(ctx->net, trans);
1084
1085         return 0;
1086
1087 err_register_hooks:
1088         nft_trans_destroy(trans);
1089         return ret;
1090 }
1091
1092 static u32 nft_chain_hash(const void *data, u32 len, u32 seed)
1093 {
1094         const char *name = data;
1095
1096         return jhash(name, strlen(name), seed);
1097 }
1098
1099 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed)
1100 {
1101         const struct nft_chain *chain = data;
1102
1103         return nft_chain_hash(chain->name, 0, seed);
1104 }
1105
1106 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *arg,
1107                               const void *ptr)
1108 {
1109         const struct nft_chain *chain = ptr;
1110         const char *name = arg->key;
1111
1112         return strcmp(chain->name, name);
1113 }
1114
1115 static u32 nft_objname_hash(const void *data, u32 len, u32 seed)
1116 {
1117         const struct nft_object_hash_key *k = data;
1118
1119         seed ^= hash_ptr(k->table, 32);
1120
1121         return jhash(k->name, strlen(k->name), seed);
1122 }
1123
1124 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed)
1125 {
1126         const struct nft_object *obj = data;
1127
1128         return nft_objname_hash(&obj->key, 0, seed);
1129 }
1130
1131 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *arg,
1132                                 const void *ptr)
1133 {
1134         const struct nft_object_hash_key *k = arg->key;
1135         const struct nft_object *obj = ptr;
1136
1137         if (obj->key.table != k->table)
1138                 return -1;
1139
1140         return strcmp(obj->key.name, k->name);
1141 }
1142
1143 static bool nft_supported_family(u8 family)
1144 {
1145         return false
1146 #ifdef CONFIG_NF_TABLES_INET
1147                 || family == NFPROTO_INET
1148 #endif
1149 #ifdef CONFIG_NF_TABLES_IPV4
1150                 || family == NFPROTO_IPV4
1151 #endif
1152 #ifdef CONFIG_NF_TABLES_ARP
1153                 || family == NFPROTO_ARP
1154 #endif
1155 #ifdef CONFIG_NF_TABLES_NETDEV
1156                 || family == NFPROTO_NETDEV
1157 #endif
1158 #if IS_ENABLED(CONFIG_NF_TABLES_BRIDGE)
1159                 || family == NFPROTO_BRIDGE
1160 #endif
1161 #ifdef CONFIG_NF_TABLES_IPV6
1162                 || family == NFPROTO_IPV6
1163 #endif
1164                 ;
1165 }
1166
1167 static int nf_tables_newtable(struct sk_buff *skb, const struct nfnl_info *info,
1168                               const struct nlattr * const nla[])
1169 {
1170         struct nftables_pernet *nft_net = nft_pernet(info->net);
1171         struct netlink_ext_ack *extack = info->extack;
1172         u8 genmask = nft_genmask_next(info->net);
1173         u8 family = info->nfmsg->nfgen_family;
1174         struct net *net = info->net;
1175         const struct nlattr *attr;
1176         struct nft_table *table;
1177         struct nft_ctx ctx;
1178         u32 flags = 0;
1179         int err;
1180
1181         if (!nft_supported_family(family))
1182                 return -EOPNOTSUPP;
1183
1184         lockdep_assert_held(&nft_net->commit_mutex);
1185         attr = nla[NFTA_TABLE_NAME];
1186         table = nft_table_lookup(net, attr, family, genmask,
1187                                  NETLINK_CB(skb).portid);
1188         if (IS_ERR(table)) {
1189                 if (PTR_ERR(table) != -ENOENT)
1190                         return PTR_ERR(table);
1191         } else {
1192                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
1193                         NL_SET_BAD_ATTR(extack, attr);
1194                         return -EEXIST;
1195                 }
1196                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
1197                         return -EOPNOTSUPP;
1198
1199                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1200
1201                 return nf_tables_updtable(&ctx);
1202         }
1203
1204         if (nla[NFTA_TABLE_FLAGS]) {
1205                 flags = ntohl(nla_get_be32(nla[NFTA_TABLE_FLAGS]));
1206                 if (flags & ~NFT_TABLE_F_MASK)
1207                         return -EOPNOTSUPP;
1208         }
1209
1210         err = -ENOMEM;
1211         table = kzalloc(sizeof(*table), GFP_KERNEL_ACCOUNT);
1212         if (table == NULL)
1213                 goto err_kzalloc;
1214
1215         table->name = nla_strdup(attr, GFP_KERNEL_ACCOUNT);
1216         if (table->name == NULL)
1217                 goto err_strdup;
1218
1219         if (nla[NFTA_TABLE_USERDATA]) {
1220                 table->udata = nla_memdup(nla[NFTA_TABLE_USERDATA], GFP_KERNEL_ACCOUNT);
1221                 if (table->udata == NULL)
1222                         goto err_table_udata;
1223
1224                 table->udlen = nla_len(nla[NFTA_TABLE_USERDATA]);
1225         }
1226
1227         err = rhltable_init(&table->chains_ht, &nft_chain_ht_params);
1228         if (err)
1229                 goto err_chain_ht;
1230
1231         INIT_LIST_HEAD(&table->chains);
1232         INIT_LIST_HEAD(&table->sets);
1233         INIT_LIST_HEAD(&table->objects);
1234         INIT_LIST_HEAD(&table->flowtables);
1235         table->family = family;
1236         table->flags = flags;
1237         table->handle = ++nft_net->table_handle;
1238         if (table->flags & NFT_TABLE_F_OWNER)
1239                 table->nlpid = NETLINK_CB(skb).portid;
1240
1241         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1242         err = nft_trans_table_add(&ctx, NFT_MSG_NEWTABLE);
1243         if (err < 0)
1244                 goto err_trans;
1245
1246         list_add_tail_rcu(&table->list, &nft_net->tables);
1247         return 0;
1248 err_trans:
1249         rhltable_destroy(&table->chains_ht);
1250 err_chain_ht:
1251         kfree(table->udata);
1252 err_table_udata:
1253         kfree(table->name);
1254 err_strdup:
1255         kfree(table);
1256 err_kzalloc:
1257         return err;
1258 }
1259
1260 static int nft_flush_table(struct nft_ctx *ctx)
1261 {
1262         struct nft_flowtable *flowtable, *nft;
1263         struct nft_chain *chain, *nc;
1264         struct nft_object *obj, *ne;
1265         struct nft_set *set, *ns;
1266         int err;
1267
1268         list_for_each_entry(chain, &ctx->table->chains, list) {
1269                 if (!nft_is_active_next(ctx->net, chain))
1270                         continue;
1271
1272                 if (nft_chain_is_bound(chain))
1273                         continue;
1274
1275                 ctx->chain = chain;
1276
1277                 err = nft_delrule_by_chain(ctx);
1278                 if (err < 0)
1279                         goto out;
1280         }
1281
1282         list_for_each_entry_safe(set, ns, &ctx->table->sets, list) {
1283                 if (!nft_is_active_next(ctx->net, set))
1284                         continue;
1285
1286                 if (nft_set_is_anonymous(set) &&
1287                     !list_empty(&set->bindings))
1288                         continue;
1289
1290                 err = nft_delset(ctx, set);
1291                 if (err < 0)
1292                         goto out;
1293         }
1294
1295         list_for_each_entry_safe(flowtable, nft, &ctx->table->flowtables, list) {
1296                 if (!nft_is_active_next(ctx->net, flowtable))
1297                         continue;
1298
1299                 err = nft_delflowtable(ctx, flowtable);
1300                 if (err < 0)
1301                         goto out;
1302         }
1303
1304         list_for_each_entry_safe(obj, ne, &ctx->table->objects, list) {
1305                 if (!nft_is_active_next(ctx->net, obj))
1306                         continue;
1307
1308                 err = nft_delobj(ctx, obj);
1309                 if (err < 0)
1310                         goto out;
1311         }
1312
1313         list_for_each_entry_safe(chain, nc, &ctx->table->chains, list) {
1314                 if (!nft_is_active_next(ctx->net, chain))
1315                         continue;
1316
1317                 if (nft_chain_is_bound(chain))
1318                         continue;
1319
1320                 ctx->chain = chain;
1321
1322                 err = nft_delchain(ctx);
1323                 if (err < 0)
1324                         goto out;
1325         }
1326
1327         err = nft_deltable(ctx);
1328 out:
1329         return err;
1330 }
1331
1332 static int nft_flush(struct nft_ctx *ctx, int family)
1333 {
1334         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
1335         const struct nlattr * const *nla = ctx->nla;
1336         struct nft_table *table, *nt;
1337         int err = 0;
1338
1339         list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
1340                 if (family != AF_UNSPEC && table->family != family)
1341                         continue;
1342
1343                 ctx->family = table->family;
1344
1345                 if (!nft_is_active_next(ctx->net, table))
1346                         continue;
1347
1348                 if (nft_table_has_owner(table) && table->nlpid != ctx->portid)
1349                         continue;
1350
1351                 if (nla[NFTA_TABLE_NAME] &&
1352                     nla_strcmp(nla[NFTA_TABLE_NAME], table->name) != 0)
1353                         continue;
1354
1355                 ctx->table = table;
1356
1357                 err = nft_flush_table(ctx);
1358                 if (err < 0)
1359                         goto out;
1360         }
1361 out:
1362         return err;
1363 }
1364
1365 static int nf_tables_deltable(struct sk_buff *skb, const struct nfnl_info *info,
1366                               const struct nlattr * const nla[])
1367 {
1368         struct netlink_ext_ack *extack = info->extack;
1369         u8 genmask = nft_genmask_next(info->net);
1370         u8 family = info->nfmsg->nfgen_family;
1371         struct net *net = info->net;
1372         const struct nlattr *attr;
1373         struct nft_table *table;
1374         struct nft_ctx ctx;
1375
1376         nft_ctx_init(&ctx, net, skb, info->nlh, 0, NULL, NULL, nla);
1377         if (family == AF_UNSPEC ||
1378             (!nla[NFTA_TABLE_NAME] && !nla[NFTA_TABLE_HANDLE]))
1379                 return nft_flush(&ctx, family);
1380
1381         if (nla[NFTA_TABLE_HANDLE]) {
1382                 attr = nla[NFTA_TABLE_HANDLE];
1383                 table = nft_table_lookup_byhandle(net, attr, genmask,
1384                                                   NETLINK_CB(skb).portid);
1385         } else {
1386                 attr = nla[NFTA_TABLE_NAME];
1387                 table = nft_table_lookup(net, attr, family, genmask,
1388                                          NETLINK_CB(skb).portid);
1389         }
1390
1391         if (IS_ERR(table)) {
1392                 NL_SET_BAD_ATTR(extack, attr);
1393                 return PTR_ERR(table);
1394         }
1395
1396         if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
1397             table->use > 0)
1398                 return -EBUSY;
1399
1400         ctx.family = family;
1401         ctx.table = table;
1402
1403         return nft_flush_table(&ctx);
1404 }
1405
1406 static void nf_tables_table_destroy(struct nft_ctx *ctx)
1407 {
1408         if (WARN_ON(ctx->table->use > 0))
1409                 return;
1410
1411         rhltable_destroy(&ctx->table->chains_ht);
1412         kfree(ctx->table->name);
1413         kfree(ctx->table->udata);
1414         kfree(ctx->table);
1415 }
1416
1417 void nft_register_chain_type(const struct nft_chain_type *ctype)
1418 {
1419         nfnl_lock(NFNL_SUBSYS_NFTABLES);
1420         if (WARN_ON(__nft_chain_type_get(ctype->family, ctype->type))) {
1421                 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1422                 return;
1423         }
1424         chain_type[ctype->family][ctype->type] = ctype;
1425         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1426 }
1427 EXPORT_SYMBOL_GPL(nft_register_chain_type);
1428
1429 void nft_unregister_chain_type(const struct nft_chain_type *ctype)
1430 {
1431         nfnl_lock(NFNL_SUBSYS_NFTABLES);
1432         chain_type[ctype->family][ctype->type] = NULL;
1433         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1434 }
1435 EXPORT_SYMBOL_GPL(nft_unregister_chain_type);
1436
1437 /*
1438  * Chains
1439  */
1440
1441 static struct nft_chain *
1442 nft_chain_lookup_byhandle(const struct nft_table *table, u64 handle, u8 genmask)
1443 {
1444         struct nft_chain *chain;
1445
1446         list_for_each_entry(chain, &table->chains, list) {
1447                 if (chain->handle == handle &&
1448                     nft_active_genmask(chain, genmask))
1449                         return chain;
1450         }
1451
1452         return ERR_PTR(-ENOENT);
1453 }
1454
1455 static bool lockdep_commit_lock_is_held(const struct net *net)
1456 {
1457 #ifdef CONFIG_PROVE_LOCKING
1458         struct nftables_pernet *nft_net = nft_pernet(net);
1459
1460         return lockdep_is_held(&nft_net->commit_mutex);
1461 #else
1462         return true;
1463 #endif
1464 }
1465
1466 static struct nft_chain *nft_chain_lookup(struct net *net,
1467                                           struct nft_table *table,
1468                                           const struct nlattr *nla, u8 genmask)
1469 {
1470         char search[NFT_CHAIN_MAXNAMELEN + 1];
1471         struct rhlist_head *tmp, *list;
1472         struct nft_chain *chain;
1473
1474         if (nla == NULL)
1475                 return ERR_PTR(-EINVAL);
1476
1477         nla_strscpy(search, nla, sizeof(search));
1478
1479         WARN_ON(!rcu_read_lock_held() &&
1480                 !lockdep_commit_lock_is_held(net));
1481
1482         chain = ERR_PTR(-ENOENT);
1483         rcu_read_lock();
1484         list = rhltable_lookup(&table->chains_ht, search, nft_chain_ht_params);
1485         if (!list)
1486                 goto out_unlock;
1487
1488         rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
1489                 if (nft_active_genmask(chain, genmask))
1490                         goto out_unlock;
1491         }
1492         chain = ERR_PTR(-ENOENT);
1493 out_unlock:
1494         rcu_read_unlock();
1495         return chain;
1496 }
1497
1498 static const struct nla_policy nft_chain_policy[NFTA_CHAIN_MAX + 1] = {
1499         [NFTA_CHAIN_TABLE]      = { .type = NLA_STRING,
1500                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
1501         [NFTA_CHAIN_HANDLE]     = { .type = NLA_U64 },
1502         [NFTA_CHAIN_NAME]       = { .type = NLA_STRING,
1503                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
1504         [NFTA_CHAIN_HOOK]       = { .type = NLA_NESTED },
1505         [NFTA_CHAIN_POLICY]     = { .type = NLA_U32 },
1506         [NFTA_CHAIN_TYPE]       = { .type = NLA_STRING,
1507                                     .len = NFT_MODULE_AUTOLOAD_LIMIT },
1508         [NFTA_CHAIN_COUNTERS]   = { .type = NLA_NESTED },
1509         [NFTA_CHAIN_FLAGS]      = { .type = NLA_U32 },
1510         [NFTA_CHAIN_ID]         = { .type = NLA_U32 },
1511         [NFTA_CHAIN_USERDATA]   = { .type = NLA_BINARY,
1512                                     .len = NFT_USERDATA_MAXLEN },
1513 };
1514
1515 static const struct nla_policy nft_hook_policy[NFTA_HOOK_MAX + 1] = {
1516         [NFTA_HOOK_HOOKNUM]     = { .type = NLA_U32 },
1517         [NFTA_HOOK_PRIORITY]    = { .type = NLA_U32 },
1518         [NFTA_HOOK_DEV]         = { .type = NLA_STRING,
1519                                     .len = IFNAMSIZ - 1 },
1520 };
1521
1522 static int nft_dump_stats(struct sk_buff *skb, struct nft_stats __percpu *stats)
1523 {
1524         struct nft_stats *cpu_stats, total;
1525         struct nlattr *nest;
1526         unsigned int seq;
1527         u64 pkts, bytes;
1528         int cpu;
1529
1530         if (!stats)
1531                 return 0;
1532
1533         memset(&total, 0, sizeof(total));
1534         for_each_possible_cpu(cpu) {
1535                 cpu_stats = per_cpu_ptr(stats, cpu);
1536                 do {
1537                         seq = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1538                         pkts = cpu_stats->pkts;
1539                         bytes = cpu_stats->bytes;
1540                 } while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, seq));
1541                 total.pkts += pkts;
1542                 total.bytes += bytes;
1543         }
1544         nest = nla_nest_start_noflag(skb, NFTA_CHAIN_COUNTERS);
1545         if (nest == NULL)
1546                 goto nla_put_failure;
1547
1548         if (nla_put_be64(skb, NFTA_COUNTER_PACKETS, cpu_to_be64(total.pkts),
1549                          NFTA_COUNTER_PAD) ||
1550             nla_put_be64(skb, NFTA_COUNTER_BYTES, cpu_to_be64(total.bytes),
1551                          NFTA_COUNTER_PAD))
1552                 goto nla_put_failure;
1553
1554         nla_nest_end(skb, nest);
1555         return 0;
1556
1557 nla_put_failure:
1558         return -ENOSPC;
1559 }
1560
1561 static int nft_dump_basechain_hook(struct sk_buff *skb, int family,
1562                                    const struct nft_base_chain *basechain)
1563 {
1564         const struct nf_hook_ops *ops = &basechain->ops;
1565         struct nft_hook *hook, *first = NULL;
1566         struct nlattr *nest, *nest_devs;
1567         int n = 0;
1568
1569         nest = nla_nest_start_noflag(skb, NFTA_CHAIN_HOOK);
1570         if (nest == NULL)
1571                 goto nla_put_failure;
1572         if (nla_put_be32(skb, NFTA_HOOK_HOOKNUM, htonl(ops->hooknum)))
1573                 goto nla_put_failure;
1574         if (nla_put_be32(skb, NFTA_HOOK_PRIORITY, htonl(ops->priority)))
1575                 goto nla_put_failure;
1576
1577         if (nft_base_chain_netdev(family, ops->hooknum)) {
1578                 nest_devs = nla_nest_start_noflag(skb, NFTA_HOOK_DEVS);
1579                 list_for_each_entry(hook, &basechain->hook_list, list) {
1580                         if (!first)
1581                                 first = hook;
1582
1583                         if (nla_put_string(skb, NFTA_DEVICE_NAME,
1584                                            hook->ops.dev->name))
1585                                 goto nla_put_failure;
1586                         n++;
1587                 }
1588                 nla_nest_end(skb, nest_devs);
1589
1590                 if (n == 1 &&
1591                     nla_put_string(skb, NFTA_HOOK_DEV, first->ops.dev->name))
1592                         goto nla_put_failure;
1593         }
1594         nla_nest_end(skb, nest);
1595
1596         return 0;
1597 nla_put_failure:
1598         return -1;
1599 }
1600
1601 static int nf_tables_fill_chain_info(struct sk_buff *skb, struct net *net,
1602                                      u32 portid, u32 seq, int event, u32 flags,
1603                                      int family, const struct nft_table *table,
1604                                      const struct nft_chain *chain)
1605 {
1606         struct nlmsghdr *nlh;
1607
1608         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
1609         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
1610                            NFNETLINK_V0, nft_base_seq(net));
1611         if (!nlh)
1612                 goto nla_put_failure;
1613
1614         if (nla_put_string(skb, NFTA_CHAIN_TABLE, table->name))
1615                 goto nla_put_failure;
1616         if (nla_put_be64(skb, NFTA_CHAIN_HANDLE, cpu_to_be64(chain->handle),
1617                          NFTA_CHAIN_PAD))
1618                 goto nla_put_failure;
1619         if (nla_put_string(skb, NFTA_CHAIN_NAME, chain->name))
1620                 goto nla_put_failure;
1621
1622         if (nft_is_base_chain(chain)) {
1623                 const struct nft_base_chain *basechain = nft_base_chain(chain);
1624                 struct nft_stats __percpu *stats;
1625
1626                 if (nft_dump_basechain_hook(skb, family, basechain))
1627                         goto nla_put_failure;
1628
1629                 if (nla_put_be32(skb, NFTA_CHAIN_POLICY,
1630                                  htonl(basechain->policy)))
1631                         goto nla_put_failure;
1632
1633                 if (nla_put_string(skb, NFTA_CHAIN_TYPE, basechain->type->name))
1634                         goto nla_put_failure;
1635
1636                 stats = rcu_dereference_check(basechain->stats,
1637                                               lockdep_commit_lock_is_held(net));
1638                 if (nft_dump_stats(skb, stats))
1639                         goto nla_put_failure;
1640         }
1641
1642         if (chain->flags &&
1643             nla_put_be32(skb, NFTA_CHAIN_FLAGS, htonl(chain->flags)))
1644                 goto nla_put_failure;
1645
1646         if (nla_put_be32(skb, NFTA_CHAIN_USE, htonl(chain->use)))
1647                 goto nla_put_failure;
1648
1649         if (chain->udata &&
1650             nla_put(skb, NFTA_CHAIN_USERDATA, chain->udlen, chain->udata))
1651                 goto nla_put_failure;
1652
1653         nlmsg_end(skb, nlh);
1654         return 0;
1655
1656 nla_put_failure:
1657         nlmsg_trim(skb, nlh);
1658         return -1;
1659 }
1660
1661 static void nf_tables_chain_notify(const struct nft_ctx *ctx, int event)
1662 {
1663         struct nftables_pernet *nft_net;
1664         struct sk_buff *skb;
1665         u16 flags = 0;
1666         int err;
1667
1668         if (!ctx->report &&
1669             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1670                 return;
1671
1672         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1673         if (skb == NULL)
1674                 goto err;
1675
1676         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
1677                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
1678
1679         err = nf_tables_fill_chain_info(skb, ctx->net, ctx->portid, ctx->seq,
1680                                         event, flags, ctx->family, ctx->table,
1681                                         ctx->chain);
1682         if (err < 0) {
1683                 kfree_skb(skb);
1684                 goto err;
1685         }
1686
1687         nft_net = nft_pernet(ctx->net);
1688         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
1689         return;
1690 err:
1691         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1692 }
1693
1694 static int nf_tables_dump_chains(struct sk_buff *skb,
1695                                  struct netlink_callback *cb)
1696 {
1697         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1698         unsigned int idx = 0, s_idx = cb->args[0];
1699         struct net *net = sock_net(skb->sk);
1700         int family = nfmsg->nfgen_family;
1701         struct nftables_pernet *nft_net;
1702         const struct nft_table *table;
1703         const struct nft_chain *chain;
1704
1705         rcu_read_lock();
1706         nft_net = nft_pernet(net);
1707         cb->seq = READ_ONCE(nft_net->base_seq);
1708
1709         list_for_each_entry_rcu(table, &nft_net->tables, list) {
1710                 if (family != NFPROTO_UNSPEC && family != table->family)
1711                         continue;
1712
1713                 list_for_each_entry_rcu(chain, &table->chains, list) {
1714                         if (idx < s_idx)
1715                                 goto cont;
1716                         if (idx > s_idx)
1717                                 memset(&cb->args[1], 0,
1718                                        sizeof(cb->args) - sizeof(cb->args[0]));
1719                         if (!nft_is_active(net, chain))
1720                                 continue;
1721                         if (nf_tables_fill_chain_info(skb, net,
1722                                                       NETLINK_CB(cb->skb).portid,
1723                                                       cb->nlh->nlmsg_seq,
1724                                                       NFT_MSG_NEWCHAIN,
1725                                                       NLM_F_MULTI,
1726                                                       table->family, table,
1727                                                       chain) < 0)
1728                                 goto done;
1729
1730                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1731 cont:
1732                         idx++;
1733                 }
1734         }
1735 done:
1736         rcu_read_unlock();
1737         cb->args[0] = idx;
1738         return skb->len;
1739 }
1740
1741 /* called with rcu_read_lock held */
1742 static int nf_tables_getchain(struct sk_buff *skb, const struct nfnl_info *info,
1743                               const struct nlattr * const nla[])
1744 {
1745         struct netlink_ext_ack *extack = info->extack;
1746         u8 genmask = nft_genmask_cur(info->net);
1747         u8 family = info->nfmsg->nfgen_family;
1748         const struct nft_chain *chain;
1749         struct net *net = info->net;
1750         struct nft_table *table;
1751         struct sk_buff *skb2;
1752         int err;
1753
1754         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1755                 struct netlink_dump_control c = {
1756                         .dump = nf_tables_dump_chains,
1757                         .module = THIS_MODULE,
1758                 };
1759
1760                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
1761         }
1762
1763         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask, 0);
1764         if (IS_ERR(table)) {
1765                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
1766                 return PTR_ERR(table);
1767         }
1768
1769         chain = nft_chain_lookup(net, table, nla[NFTA_CHAIN_NAME], genmask);
1770         if (IS_ERR(chain)) {
1771                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
1772                 return PTR_ERR(chain);
1773         }
1774
1775         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1776         if (!skb2)
1777                 return -ENOMEM;
1778
1779         err = nf_tables_fill_chain_info(skb2, net, NETLINK_CB(skb).portid,
1780                                         info->nlh->nlmsg_seq, NFT_MSG_NEWCHAIN,
1781                                         0, family, table, chain);
1782         if (err < 0)
1783                 goto err_fill_chain_info;
1784
1785         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
1786
1787 err_fill_chain_info:
1788         kfree_skb(skb2);
1789         return err;
1790 }
1791
1792 static const struct nla_policy nft_counter_policy[NFTA_COUNTER_MAX + 1] = {
1793         [NFTA_COUNTER_PACKETS]  = { .type = NLA_U64 },
1794         [NFTA_COUNTER_BYTES]    = { .type = NLA_U64 },
1795 };
1796
1797 static struct nft_stats __percpu *nft_stats_alloc(const struct nlattr *attr)
1798 {
1799         struct nlattr *tb[NFTA_COUNTER_MAX+1];
1800         struct nft_stats __percpu *newstats;
1801         struct nft_stats *stats;
1802         int err;
1803
1804         err = nla_parse_nested_deprecated(tb, NFTA_COUNTER_MAX, attr,
1805                                           nft_counter_policy, NULL);
1806         if (err < 0)
1807                 return ERR_PTR(err);
1808
1809         if (!tb[NFTA_COUNTER_BYTES] || !tb[NFTA_COUNTER_PACKETS])
1810                 return ERR_PTR(-EINVAL);
1811
1812         newstats = netdev_alloc_pcpu_stats(struct nft_stats);
1813         if (newstats == NULL)
1814                 return ERR_PTR(-ENOMEM);
1815
1816         /* Restore old counters on this cpu, no problem. Per-cpu statistics
1817          * are not exposed to userspace.
1818          */
1819         preempt_disable();
1820         stats = this_cpu_ptr(newstats);
1821         stats->bytes = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_BYTES]));
1822         stats->pkts = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_PACKETS]));
1823         preempt_enable();
1824
1825         return newstats;
1826 }
1827
1828 static void nft_chain_stats_replace(struct nft_trans *trans)
1829 {
1830         struct nft_base_chain *chain = nft_base_chain(trans->ctx.chain);
1831
1832         if (!nft_trans_chain_stats(trans))
1833                 return;
1834
1835         nft_trans_chain_stats(trans) =
1836                 rcu_replace_pointer(chain->stats, nft_trans_chain_stats(trans),
1837                                     lockdep_commit_lock_is_held(trans->ctx.net));
1838
1839         if (!nft_trans_chain_stats(trans))
1840                 static_branch_inc(&nft_counters_enabled);
1841 }
1842
1843 static void nf_tables_chain_free_chain_rules(struct nft_chain *chain)
1844 {
1845         struct nft_rule_blob *g0 = rcu_dereference_raw(chain->blob_gen_0);
1846         struct nft_rule_blob *g1 = rcu_dereference_raw(chain->blob_gen_1);
1847
1848         if (g0 != g1)
1849                 kvfree(g1);
1850         kvfree(g0);
1851
1852         /* should be NULL either via abort or via successful commit */
1853         WARN_ON_ONCE(chain->blob_next);
1854         kvfree(chain->blob_next);
1855 }
1856
1857 void nf_tables_chain_destroy(struct nft_ctx *ctx)
1858 {
1859         struct nft_chain *chain = ctx->chain;
1860         struct nft_hook *hook, *next;
1861
1862         if (WARN_ON(chain->use > 0))
1863                 return;
1864
1865         /* no concurrent access possible anymore */
1866         nf_tables_chain_free_chain_rules(chain);
1867
1868         if (nft_is_base_chain(chain)) {
1869                 struct nft_base_chain *basechain = nft_base_chain(chain);
1870
1871                 if (nft_base_chain_netdev(ctx->family, basechain->ops.hooknum)) {
1872                         list_for_each_entry_safe(hook, next,
1873                                                  &basechain->hook_list, list) {
1874                                 list_del_rcu(&hook->list);
1875                                 kfree_rcu(hook, rcu);
1876                         }
1877                 }
1878                 module_put(basechain->type->owner);
1879                 if (rcu_access_pointer(basechain->stats)) {
1880                         static_branch_dec(&nft_counters_enabled);
1881                         free_percpu(rcu_dereference_raw(basechain->stats));
1882                 }
1883                 kfree(chain->name);
1884                 kfree(chain->udata);
1885                 kfree(basechain);
1886         } else {
1887                 kfree(chain->name);
1888                 kfree(chain->udata);
1889                 kfree(chain);
1890         }
1891 }
1892
1893 static struct nft_hook *nft_netdev_hook_alloc(struct net *net,
1894                                               const struct nlattr *attr)
1895 {
1896         struct net_device *dev;
1897         char ifname[IFNAMSIZ];
1898         struct nft_hook *hook;
1899         int err;
1900
1901         hook = kmalloc(sizeof(struct nft_hook), GFP_KERNEL_ACCOUNT);
1902         if (!hook) {
1903                 err = -ENOMEM;
1904                 goto err_hook_alloc;
1905         }
1906
1907         nla_strscpy(ifname, attr, IFNAMSIZ);
1908         /* nf_tables_netdev_event() is called under rtnl_mutex, this is
1909          * indirectly serializing all the other holders of the commit_mutex with
1910          * the rtnl_mutex.
1911          */
1912         dev = __dev_get_by_name(net, ifname);
1913         if (!dev) {
1914                 err = -ENOENT;
1915                 goto err_hook_dev;
1916         }
1917         hook->ops.dev = dev;
1918
1919         return hook;
1920
1921 err_hook_dev:
1922         kfree(hook);
1923 err_hook_alloc:
1924         return ERR_PTR(err);
1925 }
1926
1927 static struct nft_hook *nft_hook_list_find(struct list_head *hook_list,
1928                                            const struct nft_hook *this)
1929 {
1930         struct nft_hook *hook;
1931
1932         list_for_each_entry(hook, hook_list, list) {
1933                 if (this->ops.dev == hook->ops.dev)
1934                         return hook;
1935         }
1936
1937         return NULL;
1938 }
1939
1940 static int nf_tables_parse_netdev_hooks(struct net *net,
1941                                         const struct nlattr *attr,
1942                                         struct list_head *hook_list)
1943 {
1944         struct nft_hook *hook, *next;
1945         const struct nlattr *tmp;
1946         int rem, n = 0, err;
1947
1948         nla_for_each_nested(tmp, attr, rem) {
1949                 if (nla_type(tmp) != NFTA_DEVICE_NAME) {
1950                         err = -EINVAL;
1951                         goto err_hook;
1952                 }
1953
1954                 hook = nft_netdev_hook_alloc(net, tmp);
1955                 if (IS_ERR(hook)) {
1956                         err = PTR_ERR(hook);
1957                         goto err_hook;
1958                 }
1959                 if (nft_hook_list_find(hook_list, hook)) {
1960                         kfree(hook);
1961                         err = -EEXIST;
1962                         goto err_hook;
1963                 }
1964                 list_add_tail(&hook->list, hook_list);
1965                 n++;
1966
1967                 if (n == NFT_NETDEVICE_MAX) {
1968                         err = -EFBIG;
1969                         goto err_hook;
1970                 }
1971         }
1972
1973         return 0;
1974
1975 err_hook:
1976         list_for_each_entry_safe(hook, next, hook_list, list) {
1977                 list_del(&hook->list);
1978                 kfree(hook);
1979         }
1980         return err;
1981 }
1982
1983 struct nft_chain_hook {
1984         u32                             num;
1985         s32                             priority;
1986         const struct nft_chain_type     *type;
1987         struct list_head                list;
1988 };
1989
1990 static int nft_chain_parse_netdev(struct net *net,
1991                                   struct nlattr *tb[],
1992                                   struct list_head *hook_list)
1993 {
1994         struct nft_hook *hook;
1995         int err;
1996
1997         if (tb[NFTA_HOOK_DEV]) {
1998                 hook = nft_netdev_hook_alloc(net, tb[NFTA_HOOK_DEV]);
1999                 if (IS_ERR(hook))
2000                         return PTR_ERR(hook);
2001
2002                 list_add_tail(&hook->list, hook_list);
2003         } else if (tb[NFTA_HOOK_DEVS]) {
2004                 err = nf_tables_parse_netdev_hooks(net, tb[NFTA_HOOK_DEVS],
2005                                                    hook_list);
2006                 if (err < 0)
2007                         return err;
2008
2009                 if (list_empty(hook_list))
2010                         return -EINVAL;
2011         } else {
2012                 return -EINVAL;
2013         }
2014
2015         return 0;
2016 }
2017
2018 static int nft_chain_parse_hook(struct net *net,
2019                                 const struct nlattr * const nla[],
2020                                 struct nft_chain_hook *hook, u8 family,
2021                                 struct netlink_ext_ack *extack, bool autoload)
2022 {
2023         struct nftables_pernet *nft_net = nft_pernet(net);
2024         struct nlattr *ha[NFTA_HOOK_MAX + 1];
2025         const struct nft_chain_type *type;
2026         int err;
2027
2028         lockdep_assert_held(&nft_net->commit_mutex);
2029         lockdep_nfnl_nft_mutex_not_held();
2030
2031         err = nla_parse_nested_deprecated(ha, NFTA_HOOK_MAX,
2032                                           nla[NFTA_CHAIN_HOOK],
2033                                           nft_hook_policy, NULL);
2034         if (err < 0)
2035                 return err;
2036
2037         if (ha[NFTA_HOOK_HOOKNUM] == NULL ||
2038             ha[NFTA_HOOK_PRIORITY] == NULL)
2039                 return -EINVAL;
2040
2041         hook->num = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
2042         hook->priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
2043
2044         type = __nft_chain_type_get(family, NFT_CHAIN_T_DEFAULT);
2045         if (!type)
2046                 return -EOPNOTSUPP;
2047
2048         if (nla[NFTA_CHAIN_TYPE]) {
2049                 type = nf_tables_chain_type_lookup(net, nla[NFTA_CHAIN_TYPE],
2050                                                    family, autoload);
2051                 if (IS_ERR(type)) {
2052                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
2053                         return PTR_ERR(type);
2054                 }
2055         }
2056         if (hook->num >= NFT_MAX_HOOKS || !(type->hook_mask & (1 << hook->num)))
2057                 return -EOPNOTSUPP;
2058
2059         if (type->type == NFT_CHAIN_T_NAT &&
2060             hook->priority <= NF_IP_PRI_CONNTRACK)
2061                 return -EOPNOTSUPP;
2062
2063         if (!try_module_get(type->owner)) {
2064                 if (nla[NFTA_CHAIN_TYPE])
2065                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
2066                 return -ENOENT;
2067         }
2068
2069         hook->type = type;
2070
2071         INIT_LIST_HEAD(&hook->list);
2072         if (nft_base_chain_netdev(family, hook->num)) {
2073                 err = nft_chain_parse_netdev(net, ha, &hook->list);
2074                 if (err < 0) {
2075                         module_put(type->owner);
2076                         return err;
2077                 }
2078         } else if (ha[NFTA_HOOK_DEV] || ha[NFTA_HOOK_DEVS]) {
2079                 module_put(type->owner);
2080                 return -EOPNOTSUPP;
2081         }
2082
2083         return 0;
2084 }
2085
2086 static void nft_chain_release_hook(struct nft_chain_hook *hook)
2087 {
2088         struct nft_hook *h, *next;
2089
2090         list_for_each_entry_safe(h, next, &hook->list, list) {
2091                 list_del(&h->list);
2092                 kfree(h);
2093         }
2094         module_put(hook->type->owner);
2095 }
2096
2097 struct nft_rules_old {
2098         struct rcu_head h;
2099         struct nft_rule_blob *blob;
2100 };
2101
2102 static void nft_last_rule(struct nft_rule_blob *blob, const void *ptr)
2103 {
2104         struct nft_rule_dp *prule;
2105
2106         prule = (struct nft_rule_dp *)ptr;
2107         prule->is_last = 1;
2108         /* blob size does not include the trailer rule */
2109 }
2110
2111 static struct nft_rule_blob *nf_tables_chain_alloc_rules(unsigned int size)
2112 {
2113         struct nft_rule_blob *blob;
2114
2115         /* size must include room for the last rule */
2116         if (size < offsetof(struct nft_rule_dp, data))
2117                 return NULL;
2118
2119         size += sizeof(struct nft_rule_blob) + sizeof(struct nft_rules_old);
2120         if (size > INT_MAX)
2121                 return NULL;
2122
2123         blob = kvmalloc(size, GFP_KERNEL_ACCOUNT);
2124         if (!blob)
2125                 return NULL;
2126
2127         blob->size = 0;
2128         nft_last_rule(blob, blob->data);
2129
2130         return blob;
2131 }
2132
2133 static void nft_basechain_hook_init(struct nf_hook_ops *ops, u8 family,
2134                                     const struct nft_chain_hook *hook,
2135                                     struct nft_chain *chain)
2136 {
2137         ops->pf                 = family;
2138         ops->hooknum            = hook->num;
2139         ops->priority           = hook->priority;
2140         ops->priv               = chain;
2141         ops->hook               = hook->type->hooks[ops->hooknum];
2142         ops->hook_ops_type      = NF_HOOK_OP_NF_TABLES;
2143 }
2144
2145 static int nft_basechain_init(struct nft_base_chain *basechain, u8 family,
2146                               struct nft_chain_hook *hook, u32 flags)
2147 {
2148         struct nft_chain *chain;
2149         struct nft_hook *h;
2150
2151         basechain->type = hook->type;
2152         INIT_LIST_HEAD(&basechain->hook_list);
2153         chain = &basechain->chain;
2154
2155         if (nft_base_chain_netdev(family, hook->num)) {
2156                 list_splice_init(&hook->list, &basechain->hook_list);
2157                 list_for_each_entry(h, &basechain->hook_list, list)
2158                         nft_basechain_hook_init(&h->ops, family, hook, chain);
2159
2160                 basechain->ops.hooknum  = hook->num;
2161                 basechain->ops.priority = hook->priority;
2162         } else {
2163                 nft_basechain_hook_init(&basechain->ops, family, hook, chain);
2164         }
2165
2166         chain->flags |= NFT_CHAIN_BASE | flags;
2167         basechain->policy = NF_ACCEPT;
2168         if (chain->flags & NFT_CHAIN_HW_OFFLOAD &&
2169             !nft_chain_offload_support(basechain)) {
2170                 list_splice_init(&basechain->hook_list, &hook->list);
2171                 return -EOPNOTSUPP;
2172         }
2173
2174         flow_block_init(&basechain->flow_block);
2175
2176         return 0;
2177 }
2178
2179 static int nft_chain_add(struct nft_table *table, struct nft_chain *chain)
2180 {
2181         int err;
2182
2183         err = rhltable_insert_key(&table->chains_ht, chain->name,
2184                                   &chain->rhlhead, nft_chain_ht_params);
2185         if (err)
2186                 return err;
2187
2188         list_add_tail_rcu(&chain->list, &table->chains);
2189
2190         return 0;
2191 }
2192
2193 static u64 chain_id;
2194
2195 static int nf_tables_addchain(struct nft_ctx *ctx, u8 family, u8 genmask,
2196                               u8 policy, u32 flags,
2197                               struct netlink_ext_ack *extack)
2198 {
2199         const struct nlattr * const *nla = ctx->nla;
2200         struct nft_table *table = ctx->table;
2201         struct nft_base_chain *basechain;
2202         struct net *net = ctx->net;
2203         char name[NFT_NAME_MAXLEN];
2204         struct nft_rule_blob *blob;
2205         struct nft_trans *trans;
2206         struct nft_chain *chain;
2207         unsigned int data_size;
2208         int err;
2209
2210         if (table->use == UINT_MAX)
2211                 return -EOVERFLOW;
2212
2213         if (nla[NFTA_CHAIN_HOOK]) {
2214                 struct nft_stats __percpu *stats = NULL;
2215                 struct nft_chain_hook hook;
2216
2217                 if (flags & NFT_CHAIN_BINDING)
2218                         return -EOPNOTSUPP;
2219
2220                 err = nft_chain_parse_hook(net, nla, &hook, family, extack,
2221                                            true);
2222                 if (err < 0)
2223                         return err;
2224
2225                 basechain = kzalloc(sizeof(*basechain), GFP_KERNEL_ACCOUNT);
2226                 if (basechain == NULL) {
2227                         nft_chain_release_hook(&hook);
2228                         return -ENOMEM;
2229                 }
2230                 chain = &basechain->chain;
2231
2232                 if (nla[NFTA_CHAIN_COUNTERS]) {
2233                         stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2234                         if (IS_ERR(stats)) {
2235                                 nft_chain_release_hook(&hook);
2236                                 kfree(basechain);
2237                                 return PTR_ERR(stats);
2238                         }
2239                         rcu_assign_pointer(basechain->stats, stats);
2240                 }
2241
2242                 err = nft_basechain_init(basechain, family, &hook, flags);
2243                 if (err < 0) {
2244                         nft_chain_release_hook(&hook);
2245                         kfree(basechain);
2246                         free_percpu(stats);
2247                         return err;
2248                 }
2249                 if (stats)
2250                         static_branch_inc(&nft_counters_enabled);
2251         } else {
2252                 if (flags & NFT_CHAIN_BASE)
2253                         return -EINVAL;
2254                 if (flags & NFT_CHAIN_HW_OFFLOAD)
2255                         return -EOPNOTSUPP;
2256
2257                 chain = kzalloc(sizeof(*chain), GFP_KERNEL_ACCOUNT);
2258                 if (chain == NULL)
2259                         return -ENOMEM;
2260
2261                 chain->flags = flags;
2262         }
2263         ctx->chain = chain;
2264
2265         INIT_LIST_HEAD(&chain->rules);
2266         chain->handle = nf_tables_alloc_handle(table);
2267         chain->table = table;
2268
2269         if (nla[NFTA_CHAIN_NAME]) {
2270                 chain->name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL_ACCOUNT);
2271         } else {
2272                 if (!(flags & NFT_CHAIN_BINDING)) {
2273                         err = -EINVAL;
2274                         goto err_destroy_chain;
2275                 }
2276
2277                 snprintf(name, sizeof(name), "__chain%llu", ++chain_id);
2278                 chain->name = kstrdup(name, GFP_KERNEL_ACCOUNT);
2279         }
2280
2281         if (!chain->name) {
2282                 err = -ENOMEM;
2283                 goto err_destroy_chain;
2284         }
2285
2286         if (nla[NFTA_CHAIN_USERDATA]) {
2287                 chain->udata = nla_memdup(nla[NFTA_CHAIN_USERDATA], GFP_KERNEL_ACCOUNT);
2288                 if (chain->udata == NULL) {
2289                         err = -ENOMEM;
2290                         goto err_destroy_chain;
2291                 }
2292                 chain->udlen = nla_len(nla[NFTA_CHAIN_USERDATA]);
2293         }
2294
2295         data_size = offsetof(struct nft_rule_dp, data); /* last rule */
2296         blob = nf_tables_chain_alloc_rules(data_size);
2297         if (!blob) {
2298                 err = -ENOMEM;
2299                 goto err_destroy_chain;
2300         }
2301
2302         RCU_INIT_POINTER(chain->blob_gen_0, blob);
2303         RCU_INIT_POINTER(chain->blob_gen_1, blob);
2304
2305         err = nf_tables_register_hook(net, table, chain);
2306         if (err < 0)
2307                 goto err_destroy_chain;
2308
2309         trans = nft_trans_chain_add(ctx, NFT_MSG_NEWCHAIN);
2310         if (IS_ERR(trans)) {
2311                 err = PTR_ERR(trans);
2312                 goto err_unregister_hook;
2313         }
2314
2315         nft_trans_chain_policy(trans) = NFT_CHAIN_POLICY_UNSET;
2316         if (nft_is_base_chain(chain))
2317                 nft_trans_chain_policy(trans) = policy;
2318
2319         err = nft_chain_add(table, chain);
2320         if (err < 0) {
2321                 nft_trans_destroy(trans);
2322                 goto err_unregister_hook;
2323         }
2324
2325         table->use++;
2326
2327         return 0;
2328 err_unregister_hook:
2329         nf_tables_unregister_hook(net, table, chain);
2330 err_destroy_chain:
2331         nf_tables_chain_destroy(ctx);
2332
2333         return err;
2334 }
2335
2336 static bool nft_hook_list_equal(struct list_head *hook_list1,
2337                                 struct list_head *hook_list2)
2338 {
2339         struct nft_hook *hook;
2340         int n = 0, m = 0;
2341
2342         n = 0;
2343         list_for_each_entry(hook, hook_list2, list) {
2344                 if (!nft_hook_list_find(hook_list1, hook))
2345                         return false;
2346
2347                 n++;
2348         }
2349         list_for_each_entry(hook, hook_list1, list)
2350                 m++;
2351
2352         return n == m;
2353 }
2354
2355 static int nf_tables_updchain(struct nft_ctx *ctx, u8 genmask, u8 policy,
2356                               u32 flags, const struct nlattr *attr,
2357                               struct netlink_ext_ack *extack)
2358 {
2359         const struct nlattr * const *nla = ctx->nla;
2360         struct nft_table *table = ctx->table;
2361         struct nft_chain *chain = ctx->chain;
2362         struct nft_base_chain *basechain;
2363         struct nft_stats *stats = NULL;
2364         struct nft_chain_hook hook;
2365         struct nf_hook_ops *ops;
2366         struct nft_trans *trans;
2367         int err;
2368
2369         if (chain->flags ^ flags)
2370                 return -EOPNOTSUPP;
2371
2372         if (nla[NFTA_CHAIN_HOOK]) {
2373                 if (!nft_is_base_chain(chain)) {
2374                         NL_SET_BAD_ATTR(extack, attr);
2375                         return -EEXIST;
2376                 }
2377                 err = nft_chain_parse_hook(ctx->net, nla, &hook, ctx->family,
2378                                            extack, false);
2379                 if (err < 0)
2380                         return err;
2381
2382                 basechain = nft_base_chain(chain);
2383                 if (basechain->type != hook.type) {
2384                         nft_chain_release_hook(&hook);
2385                         NL_SET_BAD_ATTR(extack, attr);
2386                         return -EEXIST;
2387                 }
2388
2389                 if (nft_base_chain_netdev(ctx->family, hook.num)) {
2390                         if (!nft_hook_list_equal(&basechain->hook_list,
2391                                                  &hook.list)) {
2392                                 nft_chain_release_hook(&hook);
2393                                 NL_SET_BAD_ATTR(extack, attr);
2394                                 return -EEXIST;
2395                         }
2396                 } else {
2397                         ops = &basechain->ops;
2398                         if (ops->hooknum != hook.num ||
2399                             ops->priority != hook.priority) {
2400                                 nft_chain_release_hook(&hook);
2401                                 NL_SET_BAD_ATTR(extack, attr);
2402                                 return -EEXIST;
2403                         }
2404                 }
2405                 nft_chain_release_hook(&hook);
2406         }
2407
2408         if (nla[NFTA_CHAIN_HANDLE] &&
2409             nla[NFTA_CHAIN_NAME]) {
2410                 struct nft_chain *chain2;
2411
2412                 chain2 = nft_chain_lookup(ctx->net, table,
2413                                           nla[NFTA_CHAIN_NAME], genmask);
2414                 if (!IS_ERR(chain2)) {
2415                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2416                         return -EEXIST;
2417                 }
2418         }
2419
2420         if (nla[NFTA_CHAIN_COUNTERS]) {
2421                 if (!nft_is_base_chain(chain))
2422                         return -EOPNOTSUPP;
2423
2424                 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2425                 if (IS_ERR(stats))
2426                         return PTR_ERR(stats);
2427         }
2428
2429         err = -ENOMEM;
2430         trans = nft_trans_alloc(ctx, NFT_MSG_NEWCHAIN,
2431                                 sizeof(struct nft_trans_chain));
2432         if (trans == NULL)
2433                 goto err;
2434
2435         nft_trans_chain_stats(trans) = stats;
2436         nft_trans_chain_update(trans) = true;
2437
2438         if (nla[NFTA_CHAIN_POLICY])
2439                 nft_trans_chain_policy(trans) = policy;
2440         else
2441                 nft_trans_chain_policy(trans) = -1;
2442
2443         if (nla[NFTA_CHAIN_HANDLE] &&
2444             nla[NFTA_CHAIN_NAME]) {
2445                 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
2446                 struct nft_trans *tmp;
2447                 char *name;
2448
2449                 err = -ENOMEM;
2450                 name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL_ACCOUNT);
2451                 if (!name)
2452                         goto err;
2453
2454                 err = -EEXIST;
2455                 list_for_each_entry(tmp, &nft_net->commit_list, list) {
2456                         if (tmp->msg_type == NFT_MSG_NEWCHAIN &&
2457                             tmp->ctx.table == table &&
2458                             nft_trans_chain_update(tmp) &&
2459                             nft_trans_chain_name(tmp) &&
2460                             strcmp(name, nft_trans_chain_name(tmp)) == 0) {
2461                                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2462                                 kfree(name);
2463                                 goto err;
2464                         }
2465                 }
2466
2467                 nft_trans_chain_name(trans) = name;
2468         }
2469         nft_trans_commit_list_add_tail(ctx->net, trans);
2470
2471         return 0;
2472 err:
2473         free_percpu(stats);
2474         kfree(trans);
2475         return err;
2476 }
2477
2478 static struct nft_chain *nft_chain_lookup_byid(const struct net *net,
2479                                                const struct nft_table *table,
2480                                                const struct nlattr *nla)
2481 {
2482         struct nftables_pernet *nft_net = nft_pernet(net);
2483         u32 id = ntohl(nla_get_be32(nla));
2484         struct nft_trans *trans;
2485
2486         list_for_each_entry(trans, &nft_net->commit_list, list) {
2487                 struct nft_chain *chain = trans->ctx.chain;
2488
2489                 if (trans->msg_type == NFT_MSG_NEWCHAIN &&
2490                     chain->table == table &&
2491                     id == nft_trans_chain_id(trans))
2492                         return chain;
2493         }
2494         return ERR_PTR(-ENOENT);
2495 }
2496
2497 static int nf_tables_newchain(struct sk_buff *skb, const struct nfnl_info *info,
2498                               const struct nlattr * const nla[])
2499 {
2500         struct nftables_pernet *nft_net = nft_pernet(info->net);
2501         struct netlink_ext_ack *extack = info->extack;
2502         u8 genmask = nft_genmask_next(info->net);
2503         u8 family = info->nfmsg->nfgen_family;
2504         struct nft_chain *chain = NULL;
2505         struct net *net = info->net;
2506         const struct nlattr *attr;
2507         struct nft_table *table;
2508         u8 policy = NF_ACCEPT;
2509         struct nft_ctx ctx;
2510         u64 handle = 0;
2511         u32 flags = 0;
2512
2513         lockdep_assert_held(&nft_net->commit_mutex);
2514
2515         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2516                                  NETLINK_CB(skb).portid);
2517         if (IS_ERR(table)) {
2518                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2519                 return PTR_ERR(table);
2520         }
2521
2522         chain = NULL;
2523         attr = nla[NFTA_CHAIN_NAME];
2524
2525         if (nla[NFTA_CHAIN_HANDLE]) {
2526                 handle = be64_to_cpu(nla_get_be64(nla[NFTA_CHAIN_HANDLE]));
2527                 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2528                 if (IS_ERR(chain)) {
2529                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_HANDLE]);
2530                         return PTR_ERR(chain);
2531                 }
2532                 attr = nla[NFTA_CHAIN_HANDLE];
2533         } else if (nla[NFTA_CHAIN_NAME]) {
2534                 chain = nft_chain_lookup(net, table, attr, genmask);
2535                 if (IS_ERR(chain)) {
2536                         if (PTR_ERR(chain) != -ENOENT) {
2537                                 NL_SET_BAD_ATTR(extack, attr);
2538                                 return PTR_ERR(chain);
2539                         }
2540                         chain = NULL;
2541                 }
2542         } else if (!nla[NFTA_CHAIN_ID]) {
2543                 return -EINVAL;
2544         }
2545
2546         if (nla[NFTA_CHAIN_POLICY]) {
2547                 if (chain != NULL &&
2548                     !nft_is_base_chain(chain)) {
2549                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2550                         return -EOPNOTSUPP;
2551                 }
2552
2553                 if (chain == NULL &&
2554                     nla[NFTA_CHAIN_HOOK] == NULL) {
2555                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2556                         return -EOPNOTSUPP;
2557                 }
2558
2559                 policy = ntohl(nla_get_be32(nla[NFTA_CHAIN_POLICY]));
2560                 switch (policy) {
2561                 case NF_DROP:
2562                 case NF_ACCEPT:
2563                         break;
2564                 default:
2565                         return -EINVAL;
2566                 }
2567         }
2568
2569         if (nla[NFTA_CHAIN_FLAGS])
2570                 flags = ntohl(nla_get_be32(nla[NFTA_CHAIN_FLAGS]));
2571         else if (chain)
2572                 flags = chain->flags;
2573
2574         if (flags & ~NFT_CHAIN_FLAGS)
2575                 return -EOPNOTSUPP;
2576
2577         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2578
2579         if (chain != NULL) {
2580                 if (chain->flags & NFT_CHAIN_BINDING)
2581                         return -EINVAL;
2582
2583                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
2584                         NL_SET_BAD_ATTR(extack, attr);
2585                         return -EEXIST;
2586                 }
2587                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
2588                         return -EOPNOTSUPP;
2589
2590                 flags |= chain->flags & NFT_CHAIN_BASE;
2591                 return nf_tables_updchain(&ctx, genmask, policy, flags, attr,
2592                                           extack);
2593         }
2594
2595         return nf_tables_addchain(&ctx, family, genmask, policy, flags, extack);
2596 }
2597
2598 static int nf_tables_delchain(struct sk_buff *skb, const struct nfnl_info *info,
2599                               const struct nlattr * const nla[])
2600 {
2601         struct netlink_ext_ack *extack = info->extack;
2602         u8 genmask = nft_genmask_next(info->net);
2603         u8 family = info->nfmsg->nfgen_family;
2604         struct net *net = info->net;
2605         const struct nlattr *attr;
2606         struct nft_table *table;
2607         struct nft_chain *chain;
2608         struct nft_rule *rule;
2609         struct nft_ctx ctx;
2610         u64 handle;
2611         u32 use;
2612         int err;
2613
2614         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2615                                  NETLINK_CB(skb).portid);
2616         if (IS_ERR(table)) {
2617                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2618                 return PTR_ERR(table);
2619         }
2620
2621         if (nla[NFTA_CHAIN_HANDLE]) {
2622                 attr = nla[NFTA_CHAIN_HANDLE];
2623                 handle = be64_to_cpu(nla_get_be64(attr));
2624                 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2625         } else {
2626                 attr = nla[NFTA_CHAIN_NAME];
2627                 chain = nft_chain_lookup(net, table, attr, genmask);
2628         }
2629         if (IS_ERR(chain)) {
2630                 NL_SET_BAD_ATTR(extack, attr);
2631                 return PTR_ERR(chain);
2632         }
2633
2634         if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
2635             chain->use > 0)
2636                 return -EBUSY;
2637
2638         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2639
2640         use = chain->use;
2641         list_for_each_entry(rule, &chain->rules, list) {
2642                 if (!nft_is_active_next(net, rule))
2643                         continue;
2644                 use--;
2645
2646                 err = nft_delrule(&ctx, rule);
2647                 if (err < 0)
2648                         return err;
2649         }
2650
2651         /* There are rules and elements that are still holding references to us,
2652          * we cannot do a recursive removal in this case.
2653          */
2654         if (use > 0) {
2655                 NL_SET_BAD_ATTR(extack, attr);
2656                 return -EBUSY;
2657         }
2658
2659         return nft_delchain(&ctx);
2660 }
2661
2662 /*
2663  * Expressions
2664  */
2665
2666 /**
2667  *      nft_register_expr - register nf_tables expr type
2668  *      @type: expr type
2669  *
2670  *      Registers the expr type for use with nf_tables. Returns zero on
2671  *      success or a negative errno code otherwise.
2672  */
2673 int nft_register_expr(struct nft_expr_type *type)
2674 {
2675         nfnl_lock(NFNL_SUBSYS_NFTABLES);
2676         if (type->family == NFPROTO_UNSPEC)
2677                 list_add_tail_rcu(&type->list, &nf_tables_expressions);
2678         else
2679                 list_add_rcu(&type->list, &nf_tables_expressions);
2680         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2681         return 0;
2682 }
2683 EXPORT_SYMBOL_GPL(nft_register_expr);
2684
2685 /**
2686  *      nft_unregister_expr - unregister nf_tables expr type
2687  *      @type: expr type
2688  *
2689  *      Unregisters the expr typefor use with nf_tables.
2690  */
2691 void nft_unregister_expr(struct nft_expr_type *type)
2692 {
2693         nfnl_lock(NFNL_SUBSYS_NFTABLES);
2694         list_del_rcu(&type->list);
2695         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2696 }
2697 EXPORT_SYMBOL_GPL(nft_unregister_expr);
2698
2699 static const struct nft_expr_type *__nft_expr_type_get(u8 family,
2700                                                        struct nlattr *nla)
2701 {
2702         const struct nft_expr_type *type, *candidate = NULL;
2703
2704         list_for_each_entry(type, &nf_tables_expressions, list) {
2705                 if (!nla_strcmp(nla, type->name)) {
2706                         if (!type->family && !candidate)
2707                                 candidate = type;
2708                         else if (type->family == family)
2709                                 candidate = type;
2710                 }
2711         }
2712         return candidate;
2713 }
2714
2715 #ifdef CONFIG_MODULES
2716 static int nft_expr_type_request_module(struct net *net, u8 family,
2717                                         struct nlattr *nla)
2718 {
2719         if (nft_request_module(net, "nft-expr-%u-%.*s", family,
2720                                nla_len(nla), (char *)nla_data(nla)) == -EAGAIN)
2721                 return -EAGAIN;
2722
2723         return 0;
2724 }
2725 #endif
2726
2727 static const struct nft_expr_type *nft_expr_type_get(struct net *net,
2728                                                      u8 family,
2729                                                      struct nlattr *nla)
2730 {
2731         const struct nft_expr_type *type;
2732
2733         if (nla == NULL)
2734                 return ERR_PTR(-EINVAL);
2735
2736         type = __nft_expr_type_get(family, nla);
2737         if (type != NULL && try_module_get(type->owner))
2738                 return type;
2739
2740         lockdep_nfnl_nft_mutex_not_held();
2741 #ifdef CONFIG_MODULES
2742         if (type == NULL) {
2743                 if (nft_expr_type_request_module(net, family, nla) == -EAGAIN)
2744                         return ERR_PTR(-EAGAIN);
2745
2746                 if (nft_request_module(net, "nft-expr-%.*s",
2747                                        nla_len(nla),
2748                                        (char *)nla_data(nla)) == -EAGAIN)
2749                         return ERR_PTR(-EAGAIN);
2750         }
2751 #endif
2752         return ERR_PTR(-ENOENT);
2753 }
2754
2755 static const struct nla_policy nft_expr_policy[NFTA_EXPR_MAX + 1] = {
2756         [NFTA_EXPR_NAME]        = { .type = NLA_STRING,
2757                                     .len = NFT_MODULE_AUTOLOAD_LIMIT },
2758         [NFTA_EXPR_DATA]        = { .type = NLA_NESTED },
2759 };
2760
2761 static int nf_tables_fill_expr_info(struct sk_buff *skb,
2762                                     const struct nft_expr *expr)
2763 {
2764         if (nla_put_string(skb, NFTA_EXPR_NAME, expr->ops->type->name))
2765                 goto nla_put_failure;
2766
2767         if (expr->ops->dump) {
2768                 struct nlattr *data = nla_nest_start_noflag(skb,
2769                                                             NFTA_EXPR_DATA);
2770                 if (data == NULL)
2771                         goto nla_put_failure;
2772                 if (expr->ops->dump(skb, expr) < 0)
2773                         goto nla_put_failure;
2774                 nla_nest_end(skb, data);
2775         }
2776
2777         return skb->len;
2778
2779 nla_put_failure:
2780         return -1;
2781 };
2782
2783 int nft_expr_dump(struct sk_buff *skb, unsigned int attr,
2784                   const struct nft_expr *expr)
2785 {
2786         struct nlattr *nest;
2787
2788         nest = nla_nest_start_noflag(skb, attr);
2789         if (!nest)
2790                 goto nla_put_failure;
2791         if (nf_tables_fill_expr_info(skb, expr) < 0)
2792                 goto nla_put_failure;
2793         nla_nest_end(skb, nest);
2794         return 0;
2795
2796 nla_put_failure:
2797         return -1;
2798 }
2799
2800 struct nft_expr_info {
2801         const struct nft_expr_ops       *ops;
2802         const struct nlattr             *attr;
2803         struct nlattr                   *tb[NFT_EXPR_MAXATTR + 1];
2804 };
2805
2806 static int nf_tables_expr_parse(const struct nft_ctx *ctx,
2807                                 const struct nlattr *nla,
2808                                 struct nft_expr_info *info)
2809 {
2810         const struct nft_expr_type *type;
2811         const struct nft_expr_ops *ops;
2812         struct nlattr *tb[NFTA_EXPR_MAX + 1];
2813         int err;
2814
2815         err = nla_parse_nested_deprecated(tb, NFTA_EXPR_MAX, nla,
2816                                           nft_expr_policy, NULL);
2817         if (err < 0)
2818                 return err;
2819
2820         type = nft_expr_type_get(ctx->net, ctx->family, tb[NFTA_EXPR_NAME]);
2821         if (IS_ERR(type))
2822                 return PTR_ERR(type);
2823
2824         if (tb[NFTA_EXPR_DATA]) {
2825                 err = nla_parse_nested_deprecated(info->tb, type->maxattr,
2826                                                   tb[NFTA_EXPR_DATA],
2827                                                   type->policy, NULL);
2828                 if (err < 0)
2829                         goto err1;
2830         } else
2831                 memset(info->tb, 0, sizeof(info->tb[0]) * (type->maxattr + 1));
2832
2833         if (type->select_ops != NULL) {
2834                 ops = type->select_ops(ctx,
2835                                        (const struct nlattr * const *)info->tb);
2836                 if (IS_ERR(ops)) {
2837                         err = PTR_ERR(ops);
2838 #ifdef CONFIG_MODULES
2839                         if (err == -EAGAIN)
2840                                 if (nft_expr_type_request_module(ctx->net,
2841                                                                  ctx->family,
2842                                                                  tb[NFTA_EXPR_NAME]) != -EAGAIN)
2843                                         err = -ENOENT;
2844 #endif
2845                         goto err1;
2846                 }
2847         } else
2848                 ops = type->ops;
2849
2850         info->attr = nla;
2851         info->ops = ops;
2852
2853         return 0;
2854
2855 err1:
2856         module_put(type->owner);
2857         return err;
2858 }
2859
2860 static int nf_tables_newexpr(const struct nft_ctx *ctx,
2861                              const struct nft_expr_info *expr_info,
2862                              struct nft_expr *expr)
2863 {
2864         const struct nft_expr_ops *ops = expr_info->ops;
2865         int err;
2866
2867         expr->ops = ops;
2868         if (ops->init) {
2869                 err = ops->init(ctx, expr, (const struct nlattr **)expr_info->tb);
2870                 if (err < 0)
2871                         goto err1;
2872         }
2873
2874         return 0;
2875 err1:
2876         expr->ops = NULL;
2877         return err;
2878 }
2879
2880 static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
2881                                    struct nft_expr *expr)
2882 {
2883         const struct nft_expr_type *type = expr->ops->type;
2884
2885         if (expr->ops->destroy)
2886                 expr->ops->destroy(ctx, expr);
2887         module_put(type->owner);
2888 }
2889
2890 static struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
2891                                       const struct nlattr *nla)
2892 {
2893         struct nft_expr_info expr_info;
2894         struct nft_expr *expr;
2895         struct module *owner;
2896         int err;
2897
2898         err = nf_tables_expr_parse(ctx, nla, &expr_info);
2899         if (err < 0)
2900                 goto err_expr_parse;
2901
2902         err = -EOPNOTSUPP;
2903         if (!(expr_info.ops->type->flags & NFT_EXPR_STATEFUL))
2904                 goto err_expr_stateful;
2905
2906         err = -ENOMEM;
2907         expr = kzalloc(expr_info.ops->size, GFP_KERNEL_ACCOUNT);
2908         if (expr == NULL)
2909                 goto err_expr_stateful;
2910
2911         err = nf_tables_newexpr(ctx, &expr_info, expr);
2912         if (err < 0)
2913                 goto err_expr_new;
2914
2915         return expr;
2916 err_expr_new:
2917         kfree(expr);
2918 err_expr_stateful:
2919         owner = expr_info.ops->type->owner;
2920         if (expr_info.ops->type->release_ops)
2921                 expr_info.ops->type->release_ops(expr_info.ops);
2922
2923         module_put(owner);
2924 err_expr_parse:
2925         return ERR_PTR(err);
2926 }
2927
2928 int nft_expr_clone(struct nft_expr *dst, struct nft_expr *src)
2929 {
2930         int err;
2931
2932         if (src->ops->clone) {
2933                 dst->ops = src->ops;
2934                 err = src->ops->clone(dst, src);
2935                 if (err < 0)
2936                         return err;
2937         } else {
2938                 memcpy(dst, src, src->ops->size);
2939         }
2940
2941         __module_get(src->ops->type->owner);
2942
2943         return 0;
2944 }
2945
2946 void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
2947 {
2948         nf_tables_expr_destroy(ctx, expr);
2949         kfree(expr);
2950 }
2951
2952 /*
2953  * Rules
2954  */
2955
2956 static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
2957                                           u64 handle)
2958 {
2959         struct nft_rule *rule;
2960
2961         // FIXME: this sucks
2962         list_for_each_entry_rcu(rule, &chain->rules, list) {
2963                 if (handle == rule->handle)
2964                         return rule;
2965         }
2966
2967         return ERR_PTR(-ENOENT);
2968 }
2969
2970 static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
2971                                         const struct nlattr *nla)
2972 {
2973         if (nla == NULL)
2974                 return ERR_PTR(-EINVAL);
2975
2976         return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
2977 }
2978
2979 static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
2980         [NFTA_RULE_TABLE]       = { .type = NLA_STRING,
2981                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
2982         [NFTA_RULE_CHAIN]       = { .type = NLA_STRING,
2983                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
2984         [NFTA_RULE_HANDLE]      = { .type = NLA_U64 },
2985         [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
2986         [NFTA_RULE_COMPAT]      = { .type = NLA_NESTED },
2987         [NFTA_RULE_POSITION]    = { .type = NLA_U64 },
2988         [NFTA_RULE_USERDATA]    = { .type = NLA_BINARY,
2989                                     .len = NFT_USERDATA_MAXLEN },
2990         [NFTA_RULE_ID]          = { .type = NLA_U32 },
2991         [NFTA_RULE_POSITION_ID] = { .type = NLA_U32 },
2992         [NFTA_RULE_CHAIN_ID]    = { .type = NLA_U32 },
2993 };
2994
2995 static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
2996                                     u32 portid, u32 seq, int event,
2997                                     u32 flags, int family,
2998                                     const struct nft_table *table,
2999                                     const struct nft_chain *chain,
3000                                     const struct nft_rule *rule, u64 handle)
3001 {
3002         struct nlmsghdr *nlh;
3003         const struct nft_expr *expr, *next;
3004         struct nlattr *list;
3005         u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3006
3007         nlh = nfnl_msg_put(skb, portid, seq, type, flags, family, NFNETLINK_V0,
3008                            nft_base_seq(net));
3009         if (!nlh)
3010                 goto nla_put_failure;
3011
3012         if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
3013                 goto nla_put_failure;
3014         if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
3015                 goto nla_put_failure;
3016         if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
3017                          NFTA_RULE_PAD))
3018                 goto nla_put_failure;
3019
3020         if (event != NFT_MSG_DELRULE && handle) {
3021                 if (nla_put_be64(skb, NFTA_RULE_POSITION, cpu_to_be64(handle),
3022                                  NFTA_RULE_PAD))
3023                         goto nla_put_failure;
3024         }
3025
3026         if (chain->flags & NFT_CHAIN_HW_OFFLOAD)
3027                 nft_flow_rule_stats(chain, rule);
3028
3029         list = nla_nest_start_noflag(skb, NFTA_RULE_EXPRESSIONS);
3030         if (list == NULL)
3031                 goto nla_put_failure;
3032         nft_rule_for_each_expr(expr, next, rule) {
3033                 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
3034                         goto nla_put_failure;
3035         }
3036         nla_nest_end(skb, list);
3037
3038         if (rule->udata) {
3039                 struct nft_userdata *udata = nft_userdata(rule);
3040                 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
3041                             udata->data) < 0)
3042                         goto nla_put_failure;
3043         }
3044
3045         nlmsg_end(skb, nlh);
3046         return 0;
3047
3048 nla_put_failure:
3049         nlmsg_trim(skb, nlh);
3050         return -1;
3051 }
3052
3053 static void nf_tables_rule_notify(const struct nft_ctx *ctx,
3054                                   const struct nft_rule *rule, int event)
3055 {
3056         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3057         const struct nft_rule *prule;
3058         struct sk_buff *skb;
3059         u64 handle = 0;
3060         u16 flags = 0;
3061         int err;
3062
3063         if (!ctx->report &&
3064             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3065                 return;
3066
3067         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
3068         if (skb == NULL)
3069                 goto err;
3070
3071         if (event == NFT_MSG_NEWRULE &&
3072             !list_is_first(&rule->list, &ctx->chain->rules) &&
3073             !list_is_last(&rule->list, &ctx->chain->rules)) {
3074                 prule = list_prev_entry(rule, list);
3075                 handle = prule->handle;
3076         }
3077         if (ctx->flags & (NLM_F_APPEND | NLM_F_REPLACE))
3078                 flags |= NLM_F_APPEND;
3079         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
3080                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
3081
3082         err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
3083                                        event, flags, ctx->family, ctx->table,
3084                                        ctx->chain, rule, handle);
3085         if (err < 0) {
3086                 kfree_skb(skb);
3087                 goto err;
3088         }
3089
3090         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
3091         return;
3092 err:
3093         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
3094 }
3095
3096 struct nft_rule_dump_ctx {
3097         char *table;
3098         char *chain;
3099 };
3100
3101 static int __nf_tables_dump_rules(struct sk_buff *skb,
3102                                   unsigned int *idx,
3103                                   struct netlink_callback *cb,
3104                                   const struct nft_table *table,
3105                                   const struct nft_chain *chain)
3106 {
3107         struct net *net = sock_net(skb->sk);
3108         const struct nft_rule *rule, *prule;
3109         unsigned int s_idx = cb->args[0];
3110         u64 handle;
3111
3112         prule = NULL;
3113         list_for_each_entry_rcu(rule, &chain->rules, list) {
3114                 if (!nft_is_active(net, rule))
3115                         goto cont_skip;
3116                 if (*idx < s_idx)
3117                         goto cont;
3118                 if (*idx > s_idx) {
3119                         memset(&cb->args[1], 0,
3120                                         sizeof(cb->args) - sizeof(cb->args[0]));
3121                 }
3122                 if (prule)
3123                         handle = prule->handle;
3124                 else
3125                         handle = 0;
3126
3127                 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
3128                                         cb->nlh->nlmsg_seq,
3129                                         NFT_MSG_NEWRULE,
3130                                         NLM_F_MULTI | NLM_F_APPEND,
3131                                         table->family,
3132                                         table, chain, rule, handle) < 0)
3133                         return 1;
3134
3135                 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
3136 cont:
3137                 prule = rule;
3138 cont_skip:
3139                 (*idx)++;
3140         }
3141         return 0;
3142 }
3143
3144 static int nf_tables_dump_rules(struct sk_buff *skb,
3145                                 struct netlink_callback *cb)
3146 {
3147         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3148         const struct nft_rule_dump_ctx *ctx = cb->data;
3149         struct nft_table *table;
3150         const struct nft_chain *chain;
3151         unsigned int idx = 0;
3152         struct net *net = sock_net(skb->sk);
3153         int family = nfmsg->nfgen_family;
3154         struct nftables_pernet *nft_net;
3155
3156         rcu_read_lock();
3157         nft_net = nft_pernet(net);
3158         cb->seq = READ_ONCE(nft_net->base_seq);
3159
3160         list_for_each_entry_rcu(table, &nft_net->tables, list) {
3161                 if (family != NFPROTO_UNSPEC && family != table->family)
3162                         continue;
3163
3164                 if (ctx && ctx->table && strcmp(ctx->table, table->name) != 0)
3165                         continue;
3166
3167                 if (ctx && ctx->table && ctx->chain) {
3168                         struct rhlist_head *list, *tmp;
3169
3170                         list = rhltable_lookup(&table->chains_ht, ctx->chain,
3171                                                nft_chain_ht_params);
3172                         if (!list)
3173                                 goto done;
3174
3175                         rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
3176                                 if (!nft_is_active(net, chain))
3177                                         continue;
3178                                 __nf_tables_dump_rules(skb, &idx,
3179                                                        cb, table, chain);
3180                                 break;
3181                         }
3182                         goto done;
3183                 }
3184
3185                 list_for_each_entry_rcu(chain, &table->chains, list) {
3186                         if (__nf_tables_dump_rules(skb, &idx, cb, table, chain))
3187                                 goto done;
3188                 }
3189
3190                 if (ctx && ctx->table)
3191                         break;
3192         }
3193 done:
3194         rcu_read_unlock();
3195
3196         cb->args[0] = idx;
3197         return skb->len;
3198 }
3199
3200 static int nf_tables_dump_rules_start(struct netlink_callback *cb)
3201 {
3202         const struct nlattr * const *nla = cb->data;
3203         struct nft_rule_dump_ctx *ctx = NULL;
3204
3205         if (nla[NFTA_RULE_TABLE] || nla[NFTA_RULE_CHAIN]) {
3206                 ctx = kzalloc(sizeof(*ctx), GFP_ATOMIC);
3207                 if (!ctx)
3208                         return -ENOMEM;
3209
3210                 if (nla[NFTA_RULE_TABLE]) {
3211                         ctx->table = nla_strdup(nla[NFTA_RULE_TABLE],
3212                                                         GFP_ATOMIC);
3213                         if (!ctx->table) {
3214                                 kfree(ctx);
3215                                 return -ENOMEM;
3216                         }
3217                 }
3218                 if (nla[NFTA_RULE_CHAIN]) {
3219                         ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN],
3220                                                 GFP_ATOMIC);
3221                         if (!ctx->chain) {
3222                                 kfree(ctx->table);
3223                                 kfree(ctx);
3224                                 return -ENOMEM;
3225                         }
3226                 }
3227         }
3228
3229         cb->data = ctx;
3230         return 0;
3231 }
3232
3233 static int nf_tables_dump_rules_done(struct netlink_callback *cb)
3234 {
3235         struct nft_rule_dump_ctx *ctx = cb->data;
3236
3237         if (ctx) {
3238                 kfree(ctx->table);
3239                 kfree(ctx->chain);
3240                 kfree(ctx);
3241         }
3242         return 0;
3243 }
3244
3245 /* called with rcu_read_lock held */
3246 static int nf_tables_getrule(struct sk_buff *skb, const struct nfnl_info *info,
3247                              const struct nlattr * const nla[])
3248 {
3249         struct netlink_ext_ack *extack = info->extack;
3250         u8 genmask = nft_genmask_cur(info->net);
3251         u8 family = info->nfmsg->nfgen_family;
3252         const struct nft_chain *chain;
3253         const struct nft_rule *rule;
3254         struct net *net = info->net;
3255         struct nft_table *table;
3256         struct sk_buff *skb2;
3257         int err;
3258
3259         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3260                 struct netlink_dump_control c = {
3261                         .start= nf_tables_dump_rules_start,
3262                         .dump = nf_tables_dump_rules,
3263                         .done = nf_tables_dump_rules_done,
3264                         .module = THIS_MODULE,
3265                         .data = (void *)nla,
3266                 };
3267
3268                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
3269         }
3270
3271         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask, 0);
3272         if (IS_ERR(table)) {
3273                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3274                 return PTR_ERR(table);
3275         }
3276
3277         chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
3278         if (IS_ERR(chain)) {
3279                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3280                 return PTR_ERR(chain);
3281         }
3282
3283         rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3284         if (IS_ERR(rule)) {
3285                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3286                 return PTR_ERR(rule);
3287         }
3288
3289         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3290         if (!skb2)
3291                 return -ENOMEM;
3292
3293         err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
3294                                        info->nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
3295                                        family, table, chain, rule, 0);
3296         if (err < 0)
3297                 goto err_fill_rule_info;
3298
3299         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
3300
3301 err_fill_rule_info:
3302         kfree_skb(skb2);
3303         return err;
3304 }
3305
3306 static void nf_tables_rule_destroy(const struct nft_ctx *ctx,
3307                                    struct nft_rule *rule)
3308 {
3309         struct nft_expr *expr, *next;
3310
3311         /*
3312          * Careful: some expressions might not be initialized in case this
3313          * is called on error from nf_tables_newrule().
3314          */
3315         expr = nft_expr_first(rule);
3316         while (nft_expr_more(rule, expr)) {
3317                 next = nft_expr_next(expr);
3318                 nf_tables_expr_destroy(ctx, expr);
3319                 expr = next;
3320         }
3321         kfree(rule);
3322 }
3323
3324 void nf_tables_rule_release(const struct nft_ctx *ctx, struct nft_rule *rule)
3325 {
3326         nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_RELEASE);
3327         nf_tables_rule_destroy(ctx, rule);
3328 }
3329
3330 int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
3331 {
3332         struct nft_expr *expr, *last;
3333         const struct nft_data *data;
3334         struct nft_rule *rule;
3335         int err;
3336
3337         if (ctx->level == NFT_JUMP_STACK_SIZE)
3338                 return -EMLINK;
3339
3340         list_for_each_entry(rule, &chain->rules, list) {
3341                 if (!nft_is_active_next(ctx->net, rule))
3342                         continue;
3343
3344                 nft_rule_for_each_expr(expr, last, rule) {
3345                         if (!expr->ops->validate)
3346                                 continue;
3347
3348                         err = expr->ops->validate(ctx, expr, &data);
3349                         if (err < 0)
3350                                 return err;
3351                 }
3352
3353                 cond_resched();
3354         }
3355
3356         return 0;
3357 }
3358 EXPORT_SYMBOL_GPL(nft_chain_validate);
3359
3360 static int nft_table_validate(struct net *net, const struct nft_table *table)
3361 {
3362         struct nft_chain *chain;
3363         struct nft_ctx ctx = {
3364                 .net    = net,
3365                 .family = table->family,
3366         };
3367         int err;
3368
3369         list_for_each_entry(chain, &table->chains, list) {
3370                 if (!nft_is_base_chain(chain))
3371                         continue;
3372
3373                 ctx.chain = chain;
3374                 err = nft_chain_validate(&ctx, chain);
3375                 if (err < 0)
3376                         return err;
3377         }
3378
3379         return 0;
3380 }
3381
3382 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3383                                              const struct nft_chain *chain,
3384                                              const struct nlattr *nla);
3385
3386 #define NFT_RULE_MAXEXPRS       128
3387
3388 static int nf_tables_newrule(struct sk_buff *skb, const struct nfnl_info *info,
3389                              const struct nlattr * const nla[])
3390 {
3391         struct nftables_pernet *nft_net = nft_pernet(info->net);
3392         struct netlink_ext_ack *extack = info->extack;
3393         unsigned int size, i, n, ulen = 0, usize = 0;
3394         u8 genmask = nft_genmask_next(info->net);
3395         struct nft_rule *rule, *old_rule = NULL;
3396         struct nft_expr_info *expr_info = NULL;
3397         u8 family = info->nfmsg->nfgen_family;
3398         struct nft_flow_rule *flow = NULL;
3399         struct net *net = info->net;
3400         struct nft_userdata *udata;
3401         struct nft_table *table;
3402         struct nft_chain *chain;
3403         struct nft_trans *trans;
3404         u64 handle, pos_handle;
3405         struct nft_expr *expr;
3406         struct nft_ctx ctx;
3407         struct nlattr *tmp;
3408         int err, rem;
3409
3410         lockdep_assert_held(&nft_net->commit_mutex);
3411
3412         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3413                                  NETLINK_CB(skb).portid);
3414         if (IS_ERR(table)) {
3415                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3416                 return PTR_ERR(table);
3417         }
3418
3419         if (nla[NFTA_RULE_CHAIN]) {
3420                 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3421                                          genmask);
3422                 if (IS_ERR(chain)) {
3423                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3424                         return PTR_ERR(chain);
3425                 }
3426                 if (nft_chain_is_bound(chain))
3427                         return -EOPNOTSUPP;
3428
3429         } else if (nla[NFTA_RULE_CHAIN_ID]) {
3430                 chain = nft_chain_lookup_byid(net, table, nla[NFTA_RULE_CHAIN_ID]);
3431                 if (IS_ERR(chain)) {
3432                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN_ID]);
3433                         return PTR_ERR(chain);
3434                 }
3435         } else {
3436                 return -EINVAL;
3437         }
3438
3439         if (nla[NFTA_RULE_HANDLE]) {
3440                 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
3441                 rule = __nft_rule_lookup(chain, handle);
3442                 if (IS_ERR(rule)) {
3443                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3444                         return PTR_ERR(rule);
3445                 }
3446
3447                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
3448                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3449                         return -EEXIST;
3450                 }
3451                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
3452                         old_rule = rule;
3453                 else
3454                         return -EOPNOTSUPP;
3455         } else {
3456                 if (!(info->nlh->nlmsg_flags & NLM_F_CREATE) ||
3457                     info->nlh->nlmsg_flags & NLM_F_REPLACE)
3458                         return -EINVAL;
3459                 handle = nf_tables_alloc_handle(table);
3460
3461                 if (chain->use == UINT_MAX)
3462                         return -EOVERFLOW;
3463
3464                 if (nla[NFTA_RULE_POSITION]) {
3465                         pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
3466                         old_rule = __nft_rule_lookup(chain, pos_handle);
3467                         if (IS_ERR(old_rule)) {
3468                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
3469                                 return PTR_ERR(old_rule);
3470                         }
3471                 } else if (nla[NFTA_RULE_POSITION_ID]) {
3472                         old_rule = nft_rule_lookup_byid(net, chain, nla[NFTA_RULE_POSITION_ID]);
3473                         if (IS_ERR(old_rule)) {
3474                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION_ID]);
3475                                 return PTR_ERR(old_rule);
3476                         }
3477                 }
3478         }
3479
3480         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3481
3482         n = 0;
3483         size = 0;
3484         if (nla[NFTA_RULE_EXPRESSIONS]) {
3485                 expr_info = kvmalloc_array(NFT_RULE_MAXEXPRS,
3486                                            sizeof(struct nft_expr_info),
3487                                            GFP_KERNEL);
3488                 if (!expr_info)
3489                         return -ENOMEM;
3490
3491                 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
3492                         err = -EINVAL;
3493                         if (nla_type(tmp) != NFTA_LIST_ELEM)
3494                                 goto err_release_expr;
3495                         if (n == NFT_RULE_MAXEXPRS)
3496                                 goto err_release_expr;
3497                         err = nf_tables_expr_parse(&ctx, tmp, &expr_info[n]);
3498                         if (err < 0) {
3499                                 NL_SET_BAD_ATTR(extack, tmp);
3500                                 goto err_release_expr;
3501                         }
3502                         size += expr_info[n].ops->size;
3503                         n++;
3504                 }
3505         }
3506         /* Check for overflow of dlen field */
3507         err = -EFBIG;
3508         if (size >= 1 << 12)
3509                 goto err_release_expr;
3510
3511         if (nla[NFTA_RULE_USERDATA]) {
3512                 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
3513                 if (ulen > 0)
3514                         usize = sizeof(struct nft_userdata) + ulen;
3515         }
3516
3517         err = -ENOMEM;
3518         rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL_ACCOUNT);
3519         if (rule == NULL)
3520                 goto err_release_expr;
3521
3522         nft_activate_next(net, rule);
3523
3524         rule->handle = handle;
3525         rule->dlen   = size;
3526         rule->udata  = ulen ? 1 : 0;
3527
3528         if (ulen) {
3529                 udata = nft_userdata(rule);
3530                 udata->len = ulen - 1;
3531                 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
3532         }
3533
3534         expr = nft_expr_first(rule);
3535         for (i = 0; i < n; i++) {
3536                 err = nf_tables_newexpr(&ctx, &expr_info[i], expr);
3537                 if (err < 0) {
3538                         NL_SET_BAD_ATTR(extack, expr_info[i].attr);
3539                         goto err_release_rule;
3540                 }
3541
3542                 if (expr_info[i].ops->validate)
3543                         nft_validate_state_update(net, NFT_VALIDATE_NEED);
3544
3545                 expr_info[i].ops = NULL;
3546                 expr = nft_expr_next(expr);
3547         }
3548
3549         if (chain->flags & NFT_CHAIN_HW_OFFLOAD) {
3550                 flow = nft_flow_rule_create(net, rule);
3551                 if (IS_ERR(flow)) {
3552                         err = PTR_ERR(flow);
3553                         goto err_release_rule;
3554                 }
3555         }
3556
3557         if (info->nlh->nlmsg_flags & NLM_F_REPLACE) {
3558                 err = nft_delrule(&ctx, old_rule);
3559                 if (err < 0)
3560                         goto err_destroy_flow_rule;
3561
3562                 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3563                 if (trans == NULL) {
3564                         err = -ENOMEM;
3565                         goto err_destroy_flow_rule;
3566                 }
3567                 list_add_tail_rcu(&rule->list, &old_rule->list);
3568         } else {
3569                 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3570                 if (!trans) {
3571                         err = -ENOMEM;
3572                         goto err_destroy_flow_rule;
3573                 }
3574
3575                 if (info->nlh->nlmsg_flags & NLM_F_APPEND) {
3576                         if (old_rule)
3577                                 list_add_rcu(&rule->list, &old_rule->list);
3578                         else
3579                                 list_add_tail_rcu(&rule->list, &chain->rules);
3580                  } else {
3581                         if (old_rule)
3582                                 list_add_tail_rcu(&rule->list, &old_rule->list);
3583                         else
3584                                 list_add_rcu(&rule->list, &chain->rules);
3585                 }
3586         }
3587         kvfree(expr_info);
3588         chain->use++;
3589
3590         if (flow)
3591                 nft_trans_flow_rule(trans) = flow;
3592
3593         if (nft_net->validate_state == NFT_VALIDATE_DO)
3594                 return nft_table_validate(net, table);
3595
3596         return 0;
3597
3598 err_destroy_flow_rule:
3599         if (flow)
3600                 nft_flow_rule_destroy(flow);
3601 err_release_rule:
3602         nf_tables_rule_release(&ctx, rule);
3603 err_release_expr:
3604         for (i = 0; i < n; i++) {
3605                 if (expr_info[i].ops) {
3606                         module_put(expr_info[i].ops->type->owner);
3607                         if (expr_info[i].ops->type->release_ops)
3608                                 expr_info[i].ops->type->release_ops(expr_info[i].ops);
3609                 }
3610         }
3611         kvfree(expr_info);
3612
3613         return err;
3614 }
3615
3616 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3617                                              const struct nft_chain *chain,
3618                                              const struct nlattr *nla)
3619 {
3620         struct nftables_pernet *nft_net = nft_pernet(net);
3621         u32 id = ntohl(nla_get_be32(nla));
3622         struct nft_trans *trans;
3623
3624         list_for_each_entry(trans, &nft_net->commit_list, list) {
3625                 struct nft_rule *rule = nft_trans_rule(trans);
3626
3627                 if (trans->msg_type == NFT_MSG_NEWRULE &&
3628                     trans->ctx.chain == chain &&
3629                     id == nft_trans_rule_id(trans))
3630                         return rule;
3631         }
3632         return ERR_PTR(-ENOENT);
3633 }
3634
3635 static int nf_tables_delrule(struct sk_buff *skb, const struct nfnl_info *info,
3636                              const struct nlattr * const nla[])
3637 {
3638         struct netlink_ext_ack *extack = info->extack;
3639         u8 genmask = nft_genmask_next(info->net);
3640         u8 family = info->nfmsg->nfgen_family;
3641         struct nft_chain *chain = NULL;
3642         struct net *net = info->net;
3643         struct nft_table *table;
3644         struct nft_rule *rule;
3645         struct nft_ctx ctx;
3646         int err = 0;
3647
3648         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3649                                  NETLINK_CB(skb).portid);
3650         if (IS_ERR(table)) {
3651                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3652                 return PTR_ERR(table);
3653         }
3654
3655         if (nla[NFTA_RULE_CHAIN]) {
3656                 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3657                                          genmask);
3658                 if (IS_ERR(chain)) {
3659                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3660                         return PTR_ERR(chain);
3661                 }
3662                 if (nft_chain_is_bound(chain))
3663                         return -EOPNOTSUPP;
3664         }
3665
3666         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3667
3668         if (chain) {
3669                 if (nla[NFTA_RULE_HANDLE]) {
3670                         rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3671                         if (IS_ERR(rule)) {
3672                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3673                                 return PTR_ERR(rule);
3674                         }
3675
3676                         err = nft_delrule(&ctx, rule);
3677                 } else if (nla[NFTA_RULE_ID]) {
3678                         rule = nft_rule_lookup_byid(net, chain, nla[NFTA_RULE_ID]);
3679                         if (IS_ERR(rule)) {
3680                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
3681                                 return PTR_ERR(rule);
3682                         }
3683
3684                         err = nft_delrule(&ctx, rule);
3685                 } else {
3686                         err = nft_delrule_by_chain(&ctx);
3687                 }
3688         } else {
3689                 list_for_each_entry(chain, &table->chains, list) {
3690                         if (!nft_is_active_next(net, chain))
3691                                 continue;
3692
3693                         ctx.chain = chain;
3694                         err = nft_delrule_by_chain(&ctx);
3695                         if (err < 0)
3696                                 break;
3697                 }
3698         }
3699
3700         return err;
3701 }
3702
3703 /*
3704  * Sets
3705  */
3706 static const struct nft_set_type *nft_set_types[] = {
3707         &nft_set_hash_fast_type,
3708         &nft_set_hash_type,
3709         &nft_set_rhash_type,
3710         &nft_set_bitmap_type,
3711         &nft_set_rbtree_type,
3712 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
3713         &nft_set_pipapo_avx2_type,
3714 #endif
3715         &nft_set_pipapo_type,
3716 };
3717
3718 #define NFT_SET_FEATURES        (NFT_SET_INTERVAL | NFT_SET_MAP | \
3719                                  NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
3720                                  NFT_SET_EVAL)
3721
3722 static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
3723 {
3724         return (flags & type->features) == (flags & NFT_SET_FEATURES);
3725 }
3726
3727 /*
3728  * Select a set implementation based on the data characteristics and the
3729  * given policy. The total memory use might not be known if no size is
3730  * given, in that case the amount of memory per element is used.
3731  */
3732 static const struct nft_set_ops *
3733 nft_select_set_ops(const struct nft_ctx *ctx,
3734                    const struct nlattr * const nla[],
3735                    const struct nft_set_desc *desc,
3736                    enum nft_set_policies policy)
3737 {
3738         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3739         const struct nft_set_ops *ops, *bops;
3740         struct nft_set_estimate est, best;
3741         const struct nft_set_type *type;
3742         u32 flags = 0;
3743         int i;
3744
3745         lockdep_assert_held(&nft_net->commit_mutex);
3746         lockdep_nfnl_nft_mutex_not_held();
3747
3748         if (nla[NFTA_SET_FLAGS] != NULL)
3749                 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
3750
3751         bops        = NULL;
3752         best.size   = ~0;
3753         best.lookup = ~0;
3754         best.space  = ~0;
3755
3756         for (i = 0; i < ARRAY_SIZE(nft_set_types); i++) {
3757                 type = nft_set_types[i];
3758                 ops = &type->ops;
3759
3760                 if (!nft_set_ops_candidate(type, flags))
3761                         continue;
3762                 if (!ops->estimate(desc, flags, &est))
3763                         continue;
3764
3765                 switch (policy) {
3766                 case NFT_SET_POL_PERFORMANCE:
3767                         if (est.lookup < best.lookup)
3768                                 break;
3769                         if (est.lookup == best.lookup &&
3770                             est.space < best.space)
3771                                 break;
3772                         continue;
3773                 case NFT_SET_POL_MEMORY:
3774                         if (!desc->size) {
3775                                 if (est.space < best.space)
3776                                         break;
3777                                 if (est.space == best.space &&
3778                                     est.lookup < best.lookup)
3779                                         break;
3780                         } else if (est.size < best.size || !bops) {
3781                                 break;
3782                         }
3783                         continue;
3784                 default:
3785                         break;
3786                 }
3787
3788                 bops = ops;
3789                 best = est;
3790         }
3791
3792         if (bops != NULL)
3793                 return bops;
3794
3795         return ERR_PTR(-EOPNOTSUPP);
3796 }
3797
3798 static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
3799         [NFTA_SET_TABLE]                = { .type = NLA_STRING,
3800                                             .len = NFT_TABLE_MAXNAMELEN - 1 },
3801         [NFTA_SET_NAME]                 = { .type = NLA_STRING,
3802                                             .len = NFT_SET_MAXNAMELEN - 1 },
3803         [NFTA_SET_FLAGS]                = { .type = NLA_U32 },
3804         [NFTA_SET_KEY_TYPE]             = { .type = NLA_U32 },
3805         [NFTA_SET_KEY_LEN]              = { .type = NLA_U32 },
3806         [NFTA_SET_DATA_TYPE]            = { .type = NLA_U32 },
3807         [NFTA_SET_DATA_LEN]             = { .type = NLA_U32 },
3808         [NFTA_SET_POLICY]               = { .type = NLA_U32 },
3809         [NFTA_SET_DESC]                 = { .type = NLA_NESTED },
3810         [NFTA_SET_ID]                   = { .type = NLA_U32 },
3811         [NFTA_SET_TIMEOUT]              = { .type = NLA_U64 },
3812         [NFTA_SET_GC_INTERVAL]          = { .type = NLA_U32 },
3813         [NFTA_SET_USERDATA]             = { .type = NLA_BINARY,
3814                                             .len  = NFT_USERDATA_MAXLEN },
3815         [NFTA_SET_OBJ_TYPE]             = { .type = NLA_U32 },
3816         [NFTA_SET_HANDLE]               = { .type = NLA_U64 },
3817         [NFTA_SET_EXPR]                 = { .type = NLA_NESTED },
3818         [NFTA_SET_EXPRESSIONS]          = { .type = NLA_NESTED },
3819 };
3820
3821 static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
3822         [NFTA_SET_DESC_SIZE]            = { .type = NLA_U32 },
3823         [NFTA_SET_DESC_CONCAT]          = { .type = NLA_NESTED },
3824 };
3825
3826 static struct nft_set *nft_set_lookup(const struct nft_table *table,
3827                                       const struct nlattr *nla, u8 genmask)
3828 {
3829         struct nft_set *set;
3830
3831         if (nla == NULL)
3832                 return ERR_PTR(-EINVAL);
3833
3834         list_for_each_entry_rcu(set, &table->sets, list) {
3835                 if (!nla_strcmp(nla, set->name) &&
3836                     nft_active_genmask(set, genmask))
3837                         return set;
3838         }
3839         return ERR_PTR(-ENOENT);
3840 }
3841
3842 static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
3843                                                const struct nlattr *nla,
3844                                                u8 genmask)
3845 {
3846         struct nft_set *set;
3847
3848         list_for_each_entry(set, &table->sets, list) {
3849                 if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
3850                     nft_active_genmask(set, genmask))
3851                         return set;
3852         }
3853         return ERR_PTR(-ENOENT);
3854 }
3855
3856 static struct nft_set *nft_set_lookup_byid(const struct net *net,
3857                                            const struct nft_table *table,
3858                                            const struct nlattr *nla, u8 genmask)
3859 {
3860         struct nftables_pernet *nft_net = nft_pernet(net);
3861         u32 id = ntohl(nla_get_be32(nla));
3862         struct nft_trans *trans;
3863
3864         list_for_each_entry(trans, &nft_net->commit_list, list) {
3865                 if (trans->msg_type == NFT_MSG_NEWSET) {
3866                         struct nft_set *set = nft_trans_set(trans);
3867
3868                         if (id == nft_trans_set_id(trans) &&
3869                             set->table == table &&
3870                             nft_active_genmask(set, genmask))
3871                                 return set;
3872                 }
3873         }
3874         return ERR_PTR(-ENOENT);
3875 }
3876
3877 struct nft_set *nft_set_lookup_global(const struct net *net,
3878                                       const struct nft_table *table,
3879                                       const struct nlattr *nla_set_name,
3880                                       const struct nlattr *nla_set_id,
3881                                       u8 genmask)
3882 {
3883         struct nft_set *set;
3884
3885         set = nft_set_lookup(table, nla_set_name, genmask);
3886         if (IS_ERR(set)) {
3887                 if (!nla_set_id)
3888                         return set;
3889
3890                 set = nft_set_lookup_byid(net, table, nla_set_id, genmask);
3891         }
3892         return set;
3893 }
3894 EXPORT_SYMBOL_GPL(nft_set_lookup_global);
3895
3896 static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
3897                                     const char *name)
3898 {
3899         const struct nft_set *i;
3900         const char *p;
3901         unsigned long *inuse;
3902         unsigned int n = 0, min = 0;
3903
3904         p = strchr(name, '%');
3905         if (p != NULL) {
3906                 if (p[1] != 'd' || strchr(p + 2, '%'))
3907                         return -EINVAL;
3908
3909                 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
3910                 if (inuse == NULL)
3911                         return -ENOMEM;
3912 cont:
3913                 list_for_each_entry(i, &ctx->table->sets, list) {
3914                         int tmp;
3915
3916                         if (!nft_is_active_next(ctx->net, i))
3917                                 continue;
3918                         if (!sscanf(i->name, name, &tmp))
3919                                 continue;
3920                         if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
3921                                 continue;
3922
3923                         set_bit(tmp - min, inuse);
3924                 }
3925
3926                 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
3927                 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
3928                         min += BITS_PER_BYTE * PAGE_SIZE;
3929                         memset(inuse, 0, PAGE_SIZE);
3930                         goto cont;
3931                 }
3932                 free_page((unsigned long)inuse);
3933         }
3934
3935         set->name = kasprintf(GFP_KERNEL_ACCOUNT, name, min + n);
3936         if (!set->name)
3937                 return -ENOMEM;
3938
3939         list_for_each_entry(i, &ctx->table->sets, list) {
3940                 if (!nft_is_active_next(ctx->net, i))
3941                         continue;
3942                 if (!strcmp(set->name, i->name)) {
3943                         kfree(set->name);
3944                         set->name = NULL;
3945                         return -ENFILE;
3946                 }
3947         }
3948         return 0;
3949 }
3950
3951 int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
3952 {
3953         u64 ms = be64_to_cpu(nla_get_be64(nla));
3954         u64 max = (u64)(~((u64)0));
3955
3956         max = div_u64(max, NSEC_PER_MSEC);
3957         if (ms >= max)
3958                 return -ERANGE;
3959
3960         ms *= NSEC_PER_MSEC;
3961         *result = nsecs_to_jiffies64(ms);
3962         return 0;
3963 }
3964
3965 __be64 nf_jiffies64_to_msecs(u64 input)
3966 {
3967         return cpu_to_be64(jiffies64_to_msecs(input));
3968 }
3969
3970 static int nf_tables_fill_set_concat(struct sk_buff *skb,
3971                                      const struct nft_set *set)
3972 {
3973         struct nlattr *concat, *field;
3974         int i;
3975
3976         concat = nla_nest_start_noflag(skb, NFTA_SET_DESC_CONCAT);
3977         if (!concat)
3978                 return -ENOMEM;
3979
3980         for (i = 0; i < set->field_count; i++) {
3981                 field = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
3982                 if (!field)
3983                         return -ENOMEM;
3984
3985                 if (nla_put_be32(skb, NFTA_SET_FIELD_LEN,
3986                                  htonl(set->field_len[i])))
3987                         return -ENOMEM;
3988
3989                 nla_nest_end(skb, field);
3990         }
3991
3992         nla_nest_end(skb, concat);
3993
3994         return 0;
3995 }
3996
3997 static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
3998                               const struct nft_set *set, u16 event, u16 flags)
3999 {
4000         struct nlmsghdr *nlh;
4001         u32 portid = ctx->portid;
4002         struct nlattr *nest;
4003         u32 seq = ctx->seq;
4004         int i;
4005
4006         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
4007         nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
4008                            NFNETLINK_V0, nft_base_seq(ctx->net));
4009         if (!nlh)
4010                 goto nla_put_failure;
4011
4012         if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
4013                 goto nla_put_failure;
4014         if (nla_put_string(skb, NFTA_SET_NAME, set->name))
4015                 goto nla_put_failure;
4016         if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
4017                          NFTA_SET_PAD))
4018                 goto nla_put_failure;
4019         if (set->flags != 0)
4020                 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
4021                         goto nla_put_failure;
4022
4023         if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
4024                 goto nla_put_failure;
4025         if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
4026                 goto nla_put_failure;
4027         if (set->flags & NFT_SET_MAP) {
4028                 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
4029                         goto nla_put_failure;
4030                 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
4031                         goto nla_put_failure;
4032         }
4033         if (set->flags & NFT_SET_OBJECT &&
4034             nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
4035                 goto nla_put_failure;
4036
4037         if (set->timeout &&
4038             nla_put_be64(skb, NFTA_SET_TIMEOUT,
4039                          nf_jiffies64_to_msecs(set->timeout),
4040                          NFTA_SET_PAD))
4041                 goto nla_put_failure;
4042         if (set->gc_int &&
4043             nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(set->gc_int)))
4044                 goto nla_put_failure;
4045
4046         if (set->policy != NFT_SET_POL_PERFORMANCE) {
4047                 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
4048                         goto nla_put_failure;
4049         }
4050
4051         if (set->udata &&
4052             nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
4053                 goto nla_put_failure;
4054
4055         nest = nla_nest_start_noflag(skb, NFTA_SET_DESC);
4056         if (!nest)
4057                 goto nla_put_failure;
4058         if (set->size &&
4059             nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
4060                 goto nla_put_failure;
4061
4062         if (set->field_count > 1 &&
4063             nf_tables_fill_set_concat(skb, set))
4064                 goto nla_put_failure;
4065
4066         nla_nest_end(skb, nest);
4067
4068         if (set->num_exprs == 1) {
4069                 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPR);
4070                 if (nf_tables_fill_expr_info(skb, set->exprs[0]) < 0)
4071                         goto nla_put_failure;
4072
4073                 nla_nest_end(skb, nest);
4074         } else if (set->num_exprs > 1) {
4075                 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPRESSIONS);
4076                 if (nest == NULL)
4077                         goto nla_put_failure;
4078
4079                 for (i = 0; i < set->num_exprs; i++) {
4080                         if (nft_expr_dump(skb, NFTA_LIST_ELEM,
4081                                           set->exprs[i]) < 0)
4082                                 goto nla_put_failure;
4083                 }
4084                 nla_nest_end(skb, nest);
4085         }
4086
4087         nlmsg_end(skb, nlh);
4088         return 0;
4089
4090 nla_put_failure:
4091         nlmsg_trim(skb, nlh);
4092         return -1;
4093 }
4094
4095 static void nf_tables_set_notify(const struct nft_ctx *ctx,
4096                                  const struct nft_set *set, int event,
4097                                  gfp_t gfp_flags)
4098 {
4099         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
4100         u32 portid = ctx->portid;
4101         struct sk_buff *skb;
4102         u16 flags = 0;
4103         int err;
4104
4105         if (!ctx->report &&
4106             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
4107                 return;
4108
4109         skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
4110         if (skb == NULL)
4111                 goto err;
4112
4113         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
4114                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
4115
4116         err = nf_tables_fill_set(skb, ctx, set, event, flags);
4117         if (err < 0) {
4118                 kfree_skb(skb);
4119                 goto err;
4120         }
4121
4122         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
4123         return;
4124 err:
4125         nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
4126 }
4127
4128 static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
4129 {
4130         const struct nft_set *set;
4131         unsigned int idx, s_idx = cb->args[0];
4132         struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
4133         struct net *net = sock_net(skb->sk);
4134         struct nft_ctx *ctx = cb->data, ctx_set;
4135         struct nftables_pernet *nft_net;
4136
4137         if (cb->args[1])
4138                 return skb->len;
4139
4140         rcu_read_lock();
4141         nft_net = nft_pernet(net);
4142         cb->seq = READ_ONCE(nft_net->base_seq);
4143
4144         list_for_each_entry_rcu(table, &nft_net->tables, list) {
4145                 if (ctx->family != NFPROTO_UNSPEC &&
4146                     ctx->family != table->family)
4147                         continue;
4148
4149                 if (ctx->table && ctx->table != table)
4150                         continue;
4151
4152                 if (cur_table) {
4153                         if (cur_table != table)
4154                                 continue;
4155
4156                         cur_table = NULL;
4157                 }
4158                 idx = 0;
4159                 list_for_each_entry_rcu(set, &table->sets, list) {
4160                         if (idx < s_idx)
4161                                 goto cont;
4162                         if (!nft_is_active(net, set))
4163                                 goto cont;
4164
4165                         ctx_set = *ctx;
4166                         ctx_set.table = table;
4167                         ctx_set.family = table->family;
4168
4169                         if (nf_tables_fill_set(skb, &ctx_set, set,
4170                                                NFT_MSG_NEWSET,
4171                                                NLM_F_MULTI) < 0) {
4172                                 cb->args[0] = idx;
4173                                 cb->args[2] = (unsigned long) table;
4174                                 goto done;
4175                         }
4176                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
4177 cont:
4178                         idx++;
4179                 }
4180                 if (s_idx)
4181                         s_idx = 0;
4182         }
4183         cb->args[1] = 1;
4184 done:
4185         rcu_read_unlock();
4186         return skb->len;
4187 }
4188
4189 static int nf_tables_dump_sets_start(struct netlink_callback *cb)
4190 {
4191         struct nft_ctx *ctx_dump = NULL;
4192
4193         ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
4194         if (ctx_dump == NULL)
4195                 return -ENOMEM;
4196
4197         cb->data = ctx_dump;
4198         return 0;
4199 }
4200
4201 static int nf_tables_dump_sets_done(struct netlink_callback *cb)
4202 {
4203         kfree(cb->data);
4204         return 0;
4205 }
4206
4207 /* called with rcu_read_lock held */
4208 static int nf_tables_getset(struct sk_buff *skb, const struct nfnl_info *info,
4209                             const struct nlattr * const nla[])
4210 {
4211         struct netlink_ext_ack *extack = info->extack;
4212         u8 genmask = nft_genmask_cur(info->net);
4213         u8 family = info->nfmsg->nfgen_family;
4214         struct nft_table *table = NULL;
4215         struct net *net = info->net;
4216         const struct nft_set *set;
4217         struct sk_buff *skb2;
4218         struct nft_ctx ctx;
4219         int err;
4220
4221         if (nla[NFTA_SET_TABLE]) {
4222                 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4223                                          genmask, 0);
4224                 if (IS_ERR(table)) {
4225                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4226                         return PTR_ERR(table);
4227                 }
4228         }
4229
4230         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4231
4232         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
4233                 struct netlink_dump_control c = {
4234                         .start = nf_tables_dump_sets_start,
4235                         .dump = nf_tables_dump_sets,
4236                         .done = nf_tables_dump_sets_done,
4237                         .data = &ctx,
4238                         .module = THIS_MODULE,
4239                 };
4240
4241                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
4242         }
4243
4244         /* Only accept unspec with dump */
4245         if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4246                 return -EAFNOSUPPORT;
4247         if (!nla[NFTA_SET_TABLE])
4248                 return -EINVAL;
4249
4250         set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4251         if (IS_ERR(set))
4252                 return PTR_ERR(set);
4253
4254         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
4255         if (skb2 == NULL)
4256                 return -ENOMEM;
4257
4258         err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
4259         if (err < 0)
4260                 goto err_fill_set_info;
4261
4262         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
4263
4264 err_fill_set_info:
4265         kfree_skb(skb2);
4266         return err;
4267 }
4268
4269 static const struct nla_policy nft_concat_policy[NFTA_SET_FIELD_MAX + 1] = {
4270         [NFTA_SET_FIELD_LEN]    = { .type = NLA_U32 },
4271 };
4272
4273 static int nft_set_desc_concat_parse(const struct nlattr *attr,
4274                                      struct nft_set_desc *desc)
4275 {
4276         struct nlattr *tb[NFTA_SET_FIELD_MAX + 1];
4277         u32 len;
4278         int err;
4279
4280         if (desc->field_count >= ARRAY_SIZE(desc->field_len))
4281                 return -E2BIG;
4282
4283         err = nla_parse_nested_deprecated(tb, NFTA_SET_FIELD_MAX, attr,
4284                                           nft_concat_policy, NULL);
4285         if (err < 0)
4286                 return err;
4287
4288         if (!tb[NFTA_SET_FIELD_LEN])
4289                 return -EINVAL;
4290
4291         len = ntohl(nla_get_be32(tb[NFTA_SET_FIELD_LEN]));
4292         if (!len || len > U8_MAX)
4293                 return -EINVAL;
4294
4295         desc->field_len[desc->field_count++] = len;
4296
4297         return 0;
4298 }
4299
4300 static int nft_set_desc_concat(struct nft_set_desc *desc,
4301                                const struct nlattr *nla)
4302 {
4303         struct nlattr *attr;
4304         u32 num_regs = 0;
4305         int rem, err, i;
4306
4307         nla_for_each_nested(attr, nla, rem) {
4308                 if (nla_type(attr) != NFTA_LIST_ELEM)
4309                         return -EINVAL;
4310
4311                 err = nft_set_desc_concat_parse(attr, desc);
4312                 if (err < 0)
4313                         return err;
4314         }
4315
4316         for (i = 0; i < desc->field_count; i++)
4317                 num_regs += DIV_ROUND_UP(desc->field_len[i], sizeof(u32));
4318
4319         if (num_regs > NFT_REG32_COUNT)
4320                 return -E2BIG;
4321
4322         return 0;
4323 }
4324
4325 static int nf_tables_set_desc_parse(struct nft_set_desc *desc,
4326                                     const struct nlattr *nla)
4327 {
4328         struct nlattr *da[NFTA_SET_DESC_MAX + 1];
4329         int err;
4330
4331         err = nla_parse_nested_deprecated(da, NFTA_SET_DESC_MAX, nla,
4332                                           nft_set_desc_policy, NULL);
4333         if (err < 0)
4334                 return err;
4335
4336         if (da[NFTA_SET_DESC_SIZE] != NULL)
4337                 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
4338         if (da[NFTA_SET_DESC_CONCAT])
4339                 err = nft_set_desc_concat(desc, da[NFTA_SET_DESC_CONCAT]);
4340
4341         return err;
4342 }
4343
4344 static int nf_tables_newset(struct sk_buff *skb, const struct nfnl_info *info,
4345                             const struct nlattr * const nla[])
4346 {
4347         u32 ktype, dtype, flags, policy, gc_int, objtype;
4348         struct netlink_ext_ack *extack = info->extack;
4349         u8 genmask = nft_genmask_next(info->net);
4350         u8 family = info->nfmsg->nfgen_family;
4351         const struct nft_set_ops *ops;
4352         struct nft_expr *expr = NULL;
4353         struct net *net = info->net;
4354         struct nft_set_desc desc;
4355         struct nft_table *table;
4356         unsigned char *udata;
4357         struct nft_set *set;
4358         struct nft_ctx ctx;
4359         size_t alloc_size;
4360         u64 timeout;
4361         char *name;
4362         int err, i;
4363         u16 udlen;
4364         u64 size;
4365
4366         if (nla[NFTA_SET_TABLE] == NULL ||
4367             nla[NFTA_SET_NAME] == NULL ||
4368             nla[NFTA_SET_KEY_LEN] == NULL ||
4369             nla[NFTA_SET_ID] == NULL)
4370                 return -EINVAL;
4371
4372         memset(&desc, 0, sizeof(desc));
4373
4374         ktype = NFT_DATA_VALUE;
4375         if (nla[NFTA_SET_KEY_TYPE] != NULL) {
4376                 ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
4377                 if ((ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
4378                         return -EINVAL;
4379         }
4380
4381         desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
4382         if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
4383                 return -EINVAL;
4384
4385         flags = 0;
4386         if (nla[NFTA_SET_FLAGS] != NULL) {
4387                 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
4388                 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
4389                               NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
4390                               NFT_SET_MAP | NFT_SET_EVAL |
4391                               NFT_SET_OBJECT | NFT_SET_CONCAT | NFT_SET_EXPR))
4392                         return -EOPNOTSUPP;
4393                 /* Only one of these operations is supported */
4394                 if ((flags & (NFT_SET_MAP | NFT_SET_OBJECT)) ==
4395                              (NFT_SET_MAP | NFT_SET_OBJECT))
4396                         return -EOPNOTSUPP;
4397                 if ((flags & (NFT_SET_EVAL | NFT_SET_OBJECT)) ==
4398                              (NFT_SET_EVAL | NFT_SET_OBJECT))
4399                         return -EOPNOTSUPP;
4400         }
4401
4402         dtype = 0;
4403         if (nla[NFTA_SET_DATA_TYPE] != NULL) {
4404                 if (!(flags & NFT_SET_MAP))
4405                         return -EINVAL;
4406
4407                 dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
4408                 if ((dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
4409                     dtype != NFT_DATA_VERDICT)
4410                         return -EINVAL;
4411
4412                 if (dtype != NFT_DATA_VERDICT) {
4413                         if (nla[NFTA_SET_DATA_LEN] == NULL)
4414                                 return -EINVAL;
4415                         desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
4416                         if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
4417                                 return -EINVAL;
4418                 } else
4419                         desc.dlen = sizeof(struct nft_verdict);
4420         } else if (flags & NFT_SET_MAP)
4421                 return -EINVAL;
4422
4423         if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
4424                 if (!(flags & NFT_SET_OBJECT))
4425                         return -EINVAL;
4426
4427                 objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
4428                 if (objtype == NFT_OBJECT_UNSPEC ||
4429                     objtype > NFT_OBJECT_MAX)
4430                         return -EOPNOTSUPP;
4431         } else if (flags & NFT_SET_OBJECT)
4432                 return -EINVAL;
4433         else
4434                 objtype = NFT_OBJECT_UNSPEC;
4435
4436         timeout = 0;
4437         if (nla[NFTA_SET_TIMEOUT] != NULL) {
4438                 if (!(flags & NFT_SET_TIMEOUT))
4439                         return -EINVAL;
4440
4441                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &timeout);
4442                 if (err)
4443                         return err;
4444         }
4445         gc_int = 0;
4446         if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
4447                 if (!(flags & NFT_SET_TIMEOUT))
4448                         return -EINVAL;
4449                 gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
4450         }
4451
4452         policy = NFT_SET_POL_PERFORMANCE;
4453         if (nla[NFTA_SET_POLICY] != NULL)
4454                 policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
4455
4456         if (nla[NFTA_SET_DESC] != NULL) {
4457                 err = nf_tables_set_desc_parse(&desc, nla[NFTA_SET_DESC]);
4458                 if (err < 0)
4459                         return err;
4460
4461                 if (desc.field_count > 1 && !(flags & NFT_SET_CONCAT))
4462                         return -EINVAL;
4463         } else if (flags & NFT_SET_CONCAT) {
4464                 return -EINVAL;
4465         }
4466
4467         if (nla[NFTA_SET_EXPR] || nla[NFTA_SET_EXPRESSIONS])
4468                 desc.expr = true;
4469
4470         table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask,
4471                                  NETLINK_CB(skb).portid);
4472         if (IS_ERR(table)) {
4473                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4474                 return PTR_ERR(table);
4475         }
4476
4477         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4478
4479         set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4480         if (IS_ERR(set)) {
4481                 if (PTR_ERR(set) != -ENOENT) {
4482                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4483                         return PTR_ERR(set);
4484                 }
4485         } else {
4486                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
4487                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4488                         return -EEXIST;
4489                 }
4490                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
4491                         return -EOPNOTSUPP;
4492
4493                 return 0;
4494         }
4495
4496         if (!(info->nlh->nlmsg_flags & NLM_F_CREATE))
4497                 return -ENOENT;
4498
4499         ops = nft_select_set_ops(&ctx, nla, &desc, policy);
4500         if (IS_ERR(ops))
4501                 return PTR_ERR(ops);
4502
4503         udlen = 0;
4504         if (nla[NFTA_SET_USERDATA])
4505                 udlen = nla_len(nla[NFTA_SET_USERDATA]);
4506
4507         size = 0;
4508         if (ops->privsize != NULL)
4509                 size = ops->privsize(nla, &desc);
4510         alloc_size = sizeof(*set) + size + udlen;
4511         if (alloc_size < size || alloc_size > INT_MAX)
4512                 return -ENOMEM;
4513         set = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT);
4514         if (!set)
4515                 return -ENOMEM;
4516
4517         name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL_ACCOUNT);
4518         if (!name) {
4519                 err = -ENOMEM;
4520                 goto err_set_name;
4521         }
4522
4523         err = nf_tables_set_alloc_name(&ctx, set, name);
4524         kfree(name);
4525         if (err < 0)
4526                 goto err_set_name;
4527
4528         udata = NULL;
4529         if (udlen) {
4530                 udata = set->data + size;
4531                 nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
4532         }
4533
4534         INIT_LIST_HEAD(&set->bindings);
4535         INIT_LIST_HEAD(&set->catchall_list);
4536         set->table = table;
4537         write_pnet(&set->net, net);
4538         set->ops = ops;
4539         set->ktype = ktype;
4540         set->klen = desc.klen;
4541         set->dtype = dtype;
4542         set->objtype = objtype;
4543         set->dlen = desc.dlen;
4544         set->flags = flags;
4545         set->size = desc.size;
4546         set->policy = policy;
4547         set->udlen = udlen;
4548         set->udata = udata;
4549         set->timeout = timeout;
4550         set->gc_int = gc_int;
4551
4552         set->field_count = desc.field_count;
4553         for (i = 0; i < desc.field_count; i++)
4554                 set->field_len[i] = desc.field_len[i];
4555
4556         err = ops->init(set, &desc, nla);
4557         if (err < 0)
4558                 goto err_set_init;
4559
4560         if (nla[NFTA_SET_EXPR]) {
4561                 expr = nft_set_elem_expr_alloc(&ctx, set, nla[NFTA_SET_EXPR]);
4562                 if (IS_ERR(expr)) {
4563                         err = PTR_ERR(expr);
4564                         goto err_set_expr_alloc;
4565                 }
4566                 set->exprs[0] = expr;
4567                 set->num_exprs++;
4568         } else if (nla[NFTA_SET_EXPRESSIONS]) {
4569                 struct nft_expr *expr;
4570                 struct nlattr *tmp;
4571                 int left;
4572
4573                 if (!(flags & NFT_SET_EXPR)) {
4574                         err = -EINVAL;
4575                         goto err_set_expr_alloc;
4576                 }
4577                 i = 0;
4578                 nla_for_each_nested(tmp, nla[NFTA_SET_EXPRESSIONS], left) {
4579                         if (i == NFT_SET_EXPR_MAX) {
4580                                 err = -E2BIG;
4581                                 goto err_set_expr_alloc;
4582                         }
4583                         if (nla_type(tmp) != NFTA_LIST_ELEM) {
4584                                 err = -EINVAL;
4585                                 goto err_set_expr_alloc;
4586                         }
4587                         expr = nft_set_elem_expr_alloc(&ctx, set, tmp);
4588                         if (IS_ERR(expr)) {
4589                                 err = PTR_ERR(expr);
4590                                 goto err_set_expr_alloc;
4591                         }
4592                         set->exprs[i++] = expr;
4593                         set->num_exprs++;
4594                 }
4595         }
4596
4597         set->handle = nf_tables_alloc_handle(table);
4598
4599         err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
4600         if (err < 0)
4601                 goto err_set_expr_alloc;
4602
4603         list_add_tail_rcu(&set->list, &table->sets);
4604         table->use++;
4605         return 0;
4606
4607 err_set_expr_alloc:
4608         for (i = 0; i < set->num_exprs; i++)
4609                 nft_expr_destroy(&ctx, set->exprs[i]);
4610
4611         ops->destroy(set);
4612 err_set_init:
4613         kfree(set->name);
4614 err_set_name:
4615         kvfree(set);
4616         return err;
4617 }
4618
4619 struct nft_set_elem_catchall {
4620         struct list_head        list;
4621         struct rcu_head         rcu;
4622         void                    *elem;
4623 };
4624
4625 static void nft_set_catchall_destroy(const struct nft_ctx *ctx,
4626                                      struct nft_set *set)
4627 {
4628         struct nft_set_elem_catchall *next, *catchall;
4629
4630         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
4631                 list_del_rcu(&catchall->list);
4632                 nft_set_elem_destroy(set, catchall->elem, true);
4633                 kfree_rcu(catchall, rcu);
4634         }
4635 }
4636
4637 static void nft_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
4638 {
4639         int i;
4640
4641         if (WARN_ON(set->use > 0))
4642                 return;
4643
4644         for (i = 0; i < set->num_exprs; i++)
4645                 nft_expr_destroy(ctx, set->exprs[i]);
4646
4647         set->ops->destroy(set);
4648         nft_set_catchall_destroy(ctx, set);
4649         kfree(set->name);
4650         kvfree(set);
4651 }
4652
4653 static int nf_tables_delset(struct sk_buff *skb, const struct nfnl_info *info,
4654                             const struct nlattr * const nla[])
4655 {
4656         struct netlink_ext_ack *extack = info->extack;
4657         u8 genmask = nft_genmask_next(info->net);
4658         u8 family = info->nfmsg->nfgen_family;
4659         struct net *net = info->net;
4660         const struct nlattr *attr;
4661         struct nft_table *table;
4662         struct nft_set *set;
4663         struct nft_ctx ctx;
4664
4665         if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4666                 return -EAFNOSUPPORT;
4667
4668         table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4669                                  genmask, NETLINK_CB(skb).portid);
4670         if (IS_ERR(table)) {
4671                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4672                 return PTR_ERR(table);
4673         }
4674
4675         if (nla[NFTA_SET_HANDLE]) {
4676                 attr = nla[NFTA_SET_HANDLE];
4677                 set = nft_set_lookup_byhandle(table, attr, genmask);
4678         } else {
4679                 attr = nla[NFTA_SET_NAME];
4680                 set = nft_set_lookup(table, attr, genmask);
4681         }
4682
4683         if (IS_ERR(set)) {
4684                 NL_SET_BAD_ATTR(extack, attr);
4685                 return PTR_ERR(set);
4686         }
4687         if (set->use ||
4688             (info->nlh->nlmsg_flags & NLM_F_NONREC &&
4689              atomic_read(&set->nelems) > 0)) {
4690                 NL_SET_BAD_ATTR(extack, attr);
4691                 return -EBUSY;
4692         }
4693
4694         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4695
4696         return nft_delset(&ctx, set);
4697 }
4698
4699 static int nft_validate_register_store(const struct nft_ctx *ctx,
4700                                        enum nft_registers reg,
4701                                        const struct nft_data *data,
4702                                        enum nft_data_types type,
4703                                        unsigned int len);
4704
4705 static int nft_setelem_data_validate(const struct nft_ctx *ctx,
4706                                      struct nft_set *set,
4707                                      struct nft_set_elem *elem)
4708 {
4709         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4710         enum nft_registers dreg;
4711
4712         dreg = nft_type_to_reg(set->dtype);
4713         return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
4714                                            set->dtype == NFT_DATA_VERDICT ?
4715                                            NFT_DATA_VERDICT : NFT_DATA_VALUE,
4716                                            set->dlen);
4717 }
4718
4719 static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
4720                                         struct nft_set *set,
4721                                         const struct nft_set_iter *iter,
4722                                         struct nft_set_elem *elem)
4723 {
4724         return nft_setelem_data_validate(ctx, set, elem);
4725 }
4726
4727 static int nft_set_catchall_bind_check(const struct nft_ctx *ctx,
4728                                        struct nft_set *set)
4729 {
4730         u8 genmask = nft_genmask_next(ctx->net);
4731         struct nft_set_elem_catchall *catchall;
4732         struct nft_set_elem elem;
4733         struct nft_set_ext *ext;
4734         int ret = 0;
4735
4736         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4737                 ext = nft_set_elem_ext(set, catchall->elem);
4738                 if (!nft_set_elem_active(ext, genmask))
4739                         continue;
4740
4741                 elem.priv = catchall->elem;
4742                 ret = nft_setelem_data_validate(ctx, set, &elem);
4743                 if (ret < 0)
4744                         break;
4745         }
4746
4747         return ret;
4748 }
4749
4750 int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
4751                        struct nft_set_binding *binding)
4752 {
4753         struct nft_set_binding *i;
4754         struct nft_set_iter iter;
4755
4756         if (set->use == UINT_MAX)
4757                 return -EOVERFLOW;
4758
4759         if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
4760                 return -EBUSY;
4761
4762         if (binding->flags & NFT_SET_MAP) {
4763                 /* If the set is already bound to the same chain all
4764                  * jumps are already validated for that chain.
4765                  */
4766                 list_for_each_entry(i, &set->bindings, list) {
4767                         if (i->flags & NFT_SET_MAP &&
4768                             i->chain == binding->chain)
4769                                 goto bind;
4770                 }
4771
4772                 iter.genmask    = nft_genmask_next(ctx->net);
4773                 iter.skip       = 0;
4774                 iter.count      = 0;
4775                 iter.err        = 0;
4776                 iter.fn         = nf_tables_bind_check_setelem;
4777
4778                 set->ops->walk(ctx, set, &iter);
4779                 if (!iter.err)
4780                         iter.err = nft_set_catchall_bind_check(ctx, set);
4781
4782                 if (iter.err < 0)
4783                         return iter.err;
4784         }
4785 bind:
4786         binding->chain = ctx->chain;
4787         list_add_tail_rcu(&binding->list, &set->bindings);
4788         nft_set_trans_bind(ctx, set);
4789         set->use++;
4790
4791         return 0;
4792 }
4793 EXPORT_SYMBOL_GPL(nf_tables_bind_set);
4794
4795 static void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
4796                                  struct nft_set_binding *binding, bool event)
4797 {
4798         list_del_rcu(&binding->list);
4799
4800         if (list_empty(&set->bindings) && nft_set_is_anonymous(set)) {
4801                 list_del_rcu(&set->list);
4802                 if (event)
4803                         nf_tables_set_notify(ctx, set, NFT_MSG_DELSET,
4804                                              GFP_KERNEL);
4805         }
4806 }
4807
4808 void nf_tables_deactivate_set(const struct nft_ctx *ctx, struct nft_set *set,
4809                               struct nft_set_binding *binding,
4810                               enum nft_trans_phase phase)
4811 {
4812         switch (phase) {
4813         case NFT_TRANS_PREPARE:
4814                 set->use--;
4815                 return;
4816         case NFT_TRANS_ABORT:
4817         case NFT_TRANS_RELEASE:
4818                 set->use--;
4819                 fallthrough;
4820         default:
4821                 nf_tables_unbind_set(ctx, set, binding,
4822                                      phase == NFT_TRANS_COMMIT);
4823         }
4824 }
4825 EXPORT_SYMBOL_GPL(nf_tables_deactivate_set);
4826
4827 void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
4828 {
4829         if (list_empty(&set->bindings) && nft_set_is_anonymous(set))
4830                 nft_set_destroy(ctx, set);
4831 }
4832 EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
4833
4834 const struct nft_set_ext_type nft_set_ext_types[] = {
4835         [NFT_SET_EXT_KEY]               = {
4836                 .align  = __alignof__(u32),
4837         },
4838         [NFT_SET_EXT_DATA]              = {
4839                 .align  = __alignof__(u32),
4840         },
4841         [NFT_SET_EXT_EXPRESSIONS]       = {
4842                 .align  = __alignof__(struct nft_set_elem_expr),
4843         },
4844         [NFT_SET_EXT_OBJREF]            = {
4845                 .len    = sizeof(struct nft_object *),
4846                 .align  = __alignof__(struct nft_object *),
4847         },
4848         [NFT_SET_EXT_FLAGS]             = {
4849                 .len    = sizeof(u8),
4850                 .align  = __alignof__(u8),
4851         },
4852         [NFT_SET_EXT_TIMEOUT]           = {
4853                 .len    = sizeof(u64),
4854                 .align  = __alignof__(u64),
4855         },
4856         [NFT_SET_EXT_EXPIRATION]        = {
4857                 .len    = sizeof(u64),
4858                 .align  = __alignof__(u64),
4859         },
4860         [NFT_SET_EXT_USERDATA]          = {
4861                 .len    = sizeof(struct nft_userdata),
4862                 .align  = __alignof__(struct nft_userdata),
4863         },
4864         [NFT_SET_EXT_KEY_END]           = {
4865                 .align  = __alignof__(u32),
4866         },
4867 };
4868
4869 /*
4870  * Set elements
4871  */
4872
4873 static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
4874         [NFTA_SET_ELEM_KEY]             = { .type = NLA_NESTED },
4875         [NFTA_SET_ELEM_DATA]            = { .type = NLA_NESTED },
4876         [NFTA_SET_ELEM_FLAGS]           = { .type = NLA_U32 },
4877         [NFTA_SET_ELEM_TIMEOUT]         = { .type = NLA_U64 },
4878         [NFTA_SET_ELEM_EXPIRATION]      = { .type = NLA_U64 },
4879         [NFTA_SET_ELEM_USERDATA]        = { .type = NLA_BINARY,
4880                                             .len = NFT_USERDATA_MAXLEN },
4881         [NFTA_SET_ELEM_EXPR]            = { .type = NLA_NESTED },
4882         [NFTA_SET_ELEM_OBJREF]          = { .type = NLA_STRING,
4883                                             .len = NFT_OBJ_MAXNAMELEN - 1 },
4884         [NFTA_SET_ELEM_KEY_END]         = { .type = NLA_NESTED },
4885         [NFTA_SET_ELEM_EXPRESSIONS]     = { .type = NLA_NESTED },
4886 };
4887
4888 static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
4889         [NFTA_SET_ELEM_LIST_TABLE]      = { .type = NLA_STRING,
4890                                             .len = NFT_TABLE_MAXNAMELEN - 1 },
4891         [NFTA_SET_ELEM_LIST_SET]        = { .type = NLA_STRING,
4892                                             .len = NFT_SET_MAXNAMELEN - 1 },
4893         [NFTA_SET_ELEM_LIST_ELEMENTS]   = { .type = NLA_NESTED },
4894         [NFTA_SET_ELEM_LIST_SET_ID]     = { .type = NLA_U32 },
4895 };
4896
4897 static int nft_set_elem_expr_dump(struct sk_buff *skb,
4898                                   const struct nft_set *set,
4899                                   const struct nft_set_ext *ext)
4900 {
4901         struct nft_set_elem_expr *elem_expr;
4902         u32 size, num_exprs = 0;
4903         struct nft_expr *expr;
4904         struct nlattr *nest;
4905
4906         elem_expr = nft_set_ext_expr(ext);
4907         nft_setelem_expr_foreach(expr, elem_expr, size)
4908                 num_exprs++;
4909
4910         if (num_exprs == 1) {
4911                 expr = nft_setelem_expr_at(elem_expr, 0);
4912                 if (nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, expr) < 0)
4913                         return -1;
4914
4915                 return 0;
4916         } else if (num_exprs > 1) {
4917                 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_EXPRESSIONS);
4918                 if (nest == NULL)
4919                         goto nla_put_failure;
4920
4921                 nft_setelem_expr_foreach(expr, elem_expr, size) {
4922                         expr = nft_setelem_expr_at(elem_expr, size);
4923                         if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
4924                                 goto nla_put_failure;
4925                 }
4926                 nla_nest_end(skb, nest);
4927         }
4928         return 0;
4929
4930 nla_put_failure:
4931         return -1;
4932 }
4933
4934 static int nf_tables_fill_setelem(struct sk_buff *skb,
4935                                   const struct nft_set *set,
4936                                   const struct nft_set_elem *elem)
4937 {
4938         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4939         unsigned char *b = skb_tail_pointer(skb);
4940         struct nlattr *nest;
4941
4942         nest = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4943         if (nest == NULL)
4944                 goto nla_put_failure;
4945
4946         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
4947             nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
4948                           NFT_DATA_VALUE, set->klen) < 0)
4949                 goto nla_put_failure;
4950
4951         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
4952             nft_data_dump(skb, NFTA_SET_ELEM_KEY_END, nft_set_ext_key_end(ext),
4953                           NFT_DATA_VALUE, set->klen) < 0)
4954                 goto nla_put_failure;
4955
4956         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
4957             nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
4958                           set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
4959                           set->dlen) < 0)
4960                 goto nla_put_failure;
4961
4962         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS) &&
4963             nft_set_elem_expr_dump(skb, set, ext))
4964                 goto nla_put_failure;
4965
4966         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
4967             nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
4968                            (*nft_set_ext_obj(ext))->key.name) < 0)
4969                 goto nla_put_failure;
4970
4971         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
4972             nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
4973                          htonl(*nft_set_ext_flags(ext))))
4974                 goto nla_put_failure;
4975
4976         if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
4977             nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
4978                          nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
4979                          NFTA_SET_ELEM_PAD))
4980                 goto nla_put_failure;
4981
4982         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
4983                 u64 expires, now = get_jiffies_64();
4984
4985                 expires = *nft_set_ext_expiration(ext);
4986                 if (time_before64(now, expires))
4987                         expires -= now;
4988                 else
4989                         expires = 0;
4990
4991                 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
4992                                  nf_jiffies64_to_msecs(expires),
4993                                  NFTA_SET_ELEM_PAD))
4994                         goto nla_put_failure;
4995         }
4996
4997         if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
4998                 struct nft_userdata *udata;
4999
5000                 udata = nft_set_ext_userdata(ext);
5001                 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
5002                             udata->len + 1, udata->data))
5003                         goto nla_put_failure;
5004         }
5005
5006         nla_nest_end(skb, nest);
5007         return 0;
5008
5009 nla_put_failure:
5010         nlmsg_trim(skb, b);
5011         return -EMSGSIZE;
5012 }
5013
5014 struct nft_set_dump_args {
5015         const struct netlink_callback   *cb;
5016         struct nft_set_iter             iter;
5017         struct sk_buff                  *skb;
5018 };
5019
5020 static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
5021                                   struct nft_set *set,
5022                                   const struct nft_set_iter *iter,
5023                                   struct nft_set_elem *elem)
5024 {
5025         struct nft_set_dump_args *args;
5026
5027         args = container_of(iter, struct nft_set_dump_args, iter);
5028         return nf_tables_fill_setelem(args->skb, set, elem);
5029 }
5030
5031 struct nft_set_dump_ctx {
5032         const struct nft_set    *set;
5033         struct nft_ctx          ctx;
5034 };
5035
5036 static int nft_set_catchall_dump(struct net *net, struct sk_buff *skb,
5037                                  const struct nft_set *set)
5038 {
5039         struct nft_set_elem_catchall *catchall;
5040         u8 genmask = nft_genmask_cur(net);
5041         struct nft_set_elem elem;
5042         struct nft_set_ext *ext;
5043         int ret = 0;
5044
5045         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5046                 ext = nft_set_elem_ext(set, catchall->elem);
5047                 if (!nft_set_elem_active(ext, genmask) ||
5048                     nft_set_elem_expired(ext))
5049                         continue;
5050
5051                 elem.priv = catchall->elem;
5052                 ret = nf_tables_fill_setelem(skb, set, &elem);
5053                 break;
5054         }
5055
5056         return ret;
5057 }
5058
5059 static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
5060 {
5061         struct nft_set_dump_ctx *dump_ctx = cb->data;
5062         struct net *net = sock_net(skb->sk);
5063         struct nftables_pernet *nft_net;
5064         struct nft_table *table;
5065         struct nft_set *set;
5066         struct nft_set_dump_args args;
5067         bool set_found = false;
5068         struct nlmsghdr *nlh;
5069         struct nlattr *nest;
5070         u32 portid, seq;
5071         int event;
5072
5073         rcu_read_lock();
5074         nft_net = nft_pernet(net);
5075         cb->seq = READ_ONCE(nft_net->base_seq);
5076
5077         list_for_each_entry_rcu(table, &nft_net->tables, list) {
5078                 if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
5079                     dump_ctx->ctx.family != table->family)
5080                         continue;
5081
5082                 if (table != dump_ctx->ctx.table)
5083                         continue;
5084
5085                 list_for_each_entry_rcu(set, &table->sets, list) {
5086                         if (set == dump_ctx->set) {
5087                                 set_found = true;
5088                                 break;
5089                         }
5090                 }
5091                 break;
5092         }
5093
5094         if (!set_found) {
5095                 rcu_read_unlock();
5096                 return -ENOENT;
5097         }
5098
5099         event  = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
5100         portid = NETLINK_CB(cb->skb).portid;
5101         seq    = cb->nlh->nlmsg_seq;
5102
5103         nlh = nfnl_msg_put(skb, portid, seq, event, NLM_F_MULTI,
5104                            table->family, NFNETLINK_V0, nft_base_seq(net));
5105         if (!nlh)
5106                 goto nla_put_failure;
5107
5108         if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
5109                 goto nla_put_failure;
5110         if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
5111                 goto nla_put_failure;
5112
5113         nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5114         if (nest == NULL)
5115                 goto nla_put_failure;
5116
5117         args.cb                 = cb;
5118         args.skb                = skb;
5119         args.iter.genmask       = nft_genmask_cur(net);
5120         args.iter.skip          = cb->args[0];
5121         args.iter.count         = 0;
5122         args.iter.err           = 0;
5123         args.iter.fn            = nf_tables_dump_setelem;
5124         set->ops->walk(&dump_ctx->ctx, set, &args.iter);
5125
5126         if (!args.iter.err && args.iter.count == cb->args[0])
5127                 args.iter.err = nft_set_catchall_dump(net, skb, set);
5128         rcu_read_unlock();
5129
5130         nla_nest_end(skb, nest);
5131         nlmsg_end(skb, nlh);
5132
5133         if (args.iter.err && args.iter.err != -EMSGSIZE)
5134                 return args.iter.err;
5135         if (args.iter.count == cb->args[0])
5136                 return 0;
5137
5138         cb->args[0] = args.iter.count;
5139         return skb->len;
5140
5141 nla_put_failure:
5142         rcu_read_unlock();
5143         return -ENOSPC;
5144 }
5145
5146 static int nf_tables_dump_set_start(struct netlink_callback *cb)
5147 {
5148         struct nft_set_dump_ctx *dump_ctx = cb->data;
5149
5150         cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
5151
5152         return cb->data ? 0 : -ENOMEM;
5153 }
5154
5155 static int nf_tables_dump_set_done(struct netlink_callback *cb)
5156 {
5157         kfree(cb->data);
5158         return 0;
5159 }
5160
5161 static int nf_tables_fill_setelem_info(struct sk_buff *skb,
5162                                        const struct nft_ctx *ctx, u32 seq,
5163                                        u32 portid, int event, u16 flags,
5164                                        const struct nft_set *set,
5165                                        const struct nft_set_elem *elem)
5166 {
5167         struct nlmsghdr *nlh;
5168         struct nlattr *nest;
5169         int err;
5170
5171         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
5172         nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
5173                            NFNETLINK_V0, nft_base_seq(ctx->net));
5174         if (!nlh)
5175                 goto nla_put_failure;
5176
5177         if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
5178                 goto nla_put_failure;
5179         if (nla_put_string(skb, NFTA_SET_NAME, set->name))
5180                 goto nla_put_failure;
5181
5182         nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5183         if (nest == NULL)
5184                 goto nla_put_failure;
5185
5186         err = nf_tables_fill_setelem(skb, set, elem);
5187         if (err < 0)
5188                 goto nla_put_failure;
5189
5190         nla_nest_end(skb, nest);
5191
5192         nlmsg_end(skb, nlh);
5193         return 0;
5194
5195 nla_put_failure:
5196         nlmsg_trim(skb, nlh);
5197         return -1;
5198 }
5199
5200 static int nft_setelem_parse_flags(const struct nft_set *set,
5201                                    const struct nlattr *attr, u32 *flags)
5202 {
5203         if (attr == NULL)
5204                 return 0;
5205
5206         *flags = ntohl(nla_get_be32(attr));
5207         if (*flags & ~(NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5208                 return -EOPNOTSUPP;
5209         if (!(set->flags & NFT_SET_INTERVAL) &&
5210             *flags & NFT_SET_ELEM_INTERVAL_END)
5211                 return -EINVAL;
5212         if ((*flags & (NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL)) ==
5213             (NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5214                 return -EINVAL;
5215
5216         return 0;
5217 }
5218
5219 static int nft_setelem_parse_key(struct nft_ctx *ctx, struct nft_set *set,
5220                                  struct nft_data *key, struct nlattr *attr)
5221 {
5222         struct nft_data_desc desc = {
5223                 .type   = NFT_DATA_VALUE,
5224                 .size   = NFT_DATA_VALUE_MAXLEN,
5225                 .len    = set->klen,
5226         };
5227
5228         return nft_data_init(ctx, key, &desc, attr);
5229 }
5230
5231 static int nft_setelem_parse_data(struct nft_ctx *ctx, struct nft_set *set,
5232                                   struct nft_data_desc *desc,
5233                                   struct nft_data *data,
5234                                   struct nlattr *attr)
5235 {
5236         u32 dtype;
5237
5238         if (set->dtype == NFT_DATA_VERDICT)
5239                 dtype = NFT_DATA_VERDICT;
5240         else
5241                 dtype = NFT_DATA_VALUE;
5242
5243         desc->type = dtype;
5244         desc->size = NFT_DATA_VALUE_MAXLEN;
5245         desc->len = set->dlen;
5246         desc->flags = NFT_DATA_DESC_SETELEM;
5247
5248         return nft_data_init(ctx, data, desc, attr);
5249 }
5250
5251 static void *nft_setelem_catchall_get(const struct net *net,
5252                                       const struct nft_set *set)
5253 {
5254         struct nft_set_elem_catchall *catchall;
5255         u8 genmask = nft_genmask_cur(net);
5256         struct nft_set_ext *ext;
5257         void *priv = NULL;
5258
5259         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5260                 ext = nft_set_elem_ext(set, catchall->elem);
5261                 if (!nft_set_elem_active(ext, genmask) ||
5262                     nft_set_elem_expired(ext))
5263                         continue;
5264
5265                 priv = catchall->elem;
5266                 break;
5267         }
5268
5269         return priv;
5270 }
5271
5272 static int nft_setelem_get(struct nft_ctx *ctx, struct nft_set *set,
5273                            struct nft_set_elem *elem, u32 flags)
5274 {
5275         void *priv;
5276
5277         if (!(flags & NFT_SET_ELEM_CATCHALL)) {
5278                 priv = set->ops->get(ctx->net, set, elem, flags);
5279                 if (IS_ERR(priv))
5280                         return PTR_ERR(priv);
5281         } else {
5282                 priv = nft_setelem_catchall_get(ctx->net, set);
5283                 if (!priv)
5284                         return -ENOENT;
5285         }
5286         elem->priv = priv;
5287
5288         return 0;
5289 }
5290
5291 static int nft_get_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5292                             const struct nlattr *attr)
5293 {
5294         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5295         struct nft_set_elem elem;
5296         struct sk_buff *skb;
5297         uint32_t flags = 0;
5298         int err;
5299
5300         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5301                                           nft_set_elem_policy, NULL);
5302         if (err < 0)
5303                 return err;
5304
5305         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5306         if (err < 0)
5307                 return err;
5308
5309         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5310                 return -EINVAL;
5311
5312         if (nla[NFTA_SET_ELEM_KEY]) {
5313                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5314                                             nla[NFTA_SET_ELEM_KEY]);
5315                 if (err < 0)
5316                         return err;
5317         }
5318
5319         if (nla[NFTA_SET_ELEM_KEY_END]) {
5320                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5321                                             nla[NFTA_SET_ELEM_KEY_END]);
5322                 if (err < 0)
5323                         return err;
5324         }
5325
5326         err = nft_setelem_get(ctx, set, &elem, flags);
5327         if (err < 0)
5328                 return err;
5329
5330         err = -ENOMEM;
5331         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
5332         if (skb == NULL)
5333                 return err;
5334
5335         err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
5336                                           NFT_MSG_NEWSETELEM, 0, set, &elem);
5337         if (err < 0)
5338                 goto err_fill_setelem;
5339
5340         return nfnetlink_unicast(skb, ctx->net, ctx->portid);
5341
5342 err_fill_setelem:
5343         kfree_skb(skb);
5344         return err;
5345 }
5346
5347 /* called with rcu_read_lock held */
5348 static int nf_tables_getsetelem(struct sk_buff *skb,
5349                                 const struct nfnl_info *info,
5350                                 const struct nlattr * const nla[])
5351 {
5352         struct netlink_ext_ack *extack = info->extack;
5353         u8 genmask = nft_genmask_cur(info->net);
5354         u8 family = info->nfmsg->nfgen_family;
5355         struct net *net = info->net;
5356         struct nft_table *table;
5357         struct nft_set *set;
5358         struct nlattr *attr;
5359         struct nft_ctx ctx;
5360         int rem, err = 0;
5361
5362         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
5363                                  genmask, NETLINK_CB(skb).portid);
5364         if (IS_ERR(table)) {
5365                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
5366                 return PTR_ERR(table);
5367         }
5368
5369         set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
5370         if (IS_ERR(set))
5371                 return PTR_ERR(set);
5372
5373         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5374
5375         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
5376                 struct netlink_dump_control c = {
5377                         .start = nf_tables_dump_set_start,
5378                         .dump = nf_tables_dump_set,
5379                         .done = nf_tables_dump_set_done,
5380                         .module = THIS_MODULE,
5381                 };
5382                 struct nft_set_dump_ctx dump_ctx = {
5383                         .set = set,
5384                         .ctx = ctx,
5385                 };
5386
5387                 c.data = &dump_ctx;
5388                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
5389         }
5390
5391         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
5392                 return -EINVAL;
5393
5394         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
5395                 err = nft_get_set_elem(&ctx, set, attr);
5396                 if (err < 0) {
5397                         NL_SET_BAD_ATTR(extack, attr);
5398                         break;
5399                 }
5400         }
5401
5402         return err;
5403 }
5404
5405 static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
5406                                      const struct nft_set *set,
5407                                      const struct nft_set_elem *elem,
5408                                      int event)
5409 {
5410         struct nftables_pernet *nft_net;
5411         struct net *net = ctx->net;
5412         u32 portid = ctx->portid;
5413         struct sk_buff *skb;
5414         u16 flags = 0;
5415         int err;
5416
5417         if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5418                 return;
5419
5420         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5421         if (skb == NULL)
5422                 goto err;
5423
5424         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
5425                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
5426
5427         err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
5428                                           set, elem);
5429         if (err < 0) {
5430                 kfree_skb(skb);
5431                 goto err;
5432         }
5433
5434         nft_net = nft_pernet(net);
5435         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
5436         return;
5437 err:
5438         nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
5439 }
5440
5441 static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
5442                                               int msg_type,
5443                                               struct nft_set *set)
5444 {
5445         struct nft_trans *trans;
5446
5447         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
5448         if (trans == NULL)
5449                 return NULL;
5450
5451         nft_trans_elem_set(trans) = set;
5452         return trans;
5453 }
5454
5455 struct nft_expr *nft_set_elem_expr_alloc(const struct nft_ctx *ctx,
5456                                          const struct nft_set *set,
5457                                          const struct nlattr *attr)
5458 {
5459         struct nft_expr *expr;
5460         int err;
5461
5462         expr = nft_expr_init(ctx, attr);
5463         if (IS_ERR(expr))
5464                 return expr;
5465
5466         err = -EOPNOTSUPP;
5467         if (expr->ops->type->flags & NFT_EXPR_GC) {
5468                 if (set->flags & NFT_SET_TIMEOUT)
5469                         goto err_set_elem_expr;
5470                 if (!set->ops->gc_init)
5471                         goto err_set_elem_expr;
5472                 set->ops->gc_init(set);
5473         }
5474
5475         return expr;
5476
5477 err_set_elem_expr:
5478         nft_expr_destroy(ctx, expr);
5479         return ERR_PTR(err);
5480 }
5481
5482 void *nft_set_elem_init(const struct nft_set *set,
5483                         const struct nft_set_ext_tmpl *tmpl,
5484                         const u32 *key, const u32 *key_end,
5485                         const u32 *data, u64 timeout, u64 expiration, gfp_t gfp)
5486 {
5487         struct nft_set_ext *ext;
5488         void *elem;
5489
5490         elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
5491         if (elem == NULL)
5492                 return NULL;
5493
5494         ext = nft_set_elem_ext(set, elem);
5495         nft_set_ext_init(ext, tmpl);
5496
5497         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY))
5498                 memcpy(nft_set_ext_key(ext), key, set->klen);
5499         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END))
5500                 memcpy(nft_set_ext_key_end(ext), key_end, set->klen);
5501         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5502                 memcpy(nft_set_ext_data(ext), data, set->dlen);
5503         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5504                 *nft_set_ext_expiration(ext) = get_jiffies_64() + expiration;
5505                 if (expiration == 0)
5506                         *nft_set_ext_expiration(ext) += timeout;
5507         }
5508         if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
5509                 *nft_set_ext_timeout(ext) = timeout;
5510
5511         return elem;
5512 }
5513
5514 static void __nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5515                                         struct nft_expr *expr)
5516 {
5517         if (expr->ops->destroy_clone) {
5518                 expr->ops->destroy_clone(ctx, expr);
5519                 module_put(expr->ops->type->owner);
5520         } else {
5521                 nf_tables_expr_destroy(ctx, expr);
5522         }
5523 }
5524
5525 static void nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5526                                       struct nft_set_elem_expr *elem_expr)
5527 {
5528         struct nft_expr *expr;
5529         u32 size;
5530
5531         nft_setelem_expr_foreach(expr, elem_expr, size)
5532                 __nft_set_elem_expr_destroy(ctx, expr);
5533 }
5534
5535 void nft_set_elem_destroy(const struct nft_set *set, void *elem,
5536                           bool destroy_expr)
5537 {
5538         struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5539         struct nft_ctx ctx = {
5540                 .net    = read_pnet(&set->net),
5541                 .family = set->table->family,
5542         };
5543
5544         nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
5545         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5546                 nft_data_release(nft_set_ext_data(ext), set->dtype);
5547         if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5548                 nft_set_elem_expr_destroy(&ctx, nft_set_ext_expr(ext));
5549
5550         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
5551                 (*nft_set_ext_obj(ext))->use--;
5552         kfree(elem);
5553 }
5554 EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
5555
5556 /* Only called from commit path, nft_setelem_data_deactivate() already deals
5557  * with the refcounting from the preparation phase.
5558  */
5559 static void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
5560                                        const struct nft_set *set, void *elem)
5561 {
5562         struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5563
5564         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5565                 nft_set_elem_expr_destroy(ctx, nft_set_ext_expr(ext));
5566
5567         kfree(elem);
5568 }
5569
5570 int nft_set_elem_expr_clone(const struct nft_ctx *ctx, struct nft_set *set,
5571                             struct nft_expr *expr_array[])
5572 {
5573         struct nft_expr *expr;
5574         int err, i, k;
5575
5576         for (i = 0; i < set->num_exprs; i++) {
5577                 expr = kzalloc(set->exprs[i]->ops->size, GFP_KERNEL_ACCOUNT);
5578                 if (!expr)
5579                         goto err_expr;
5580
5581                 err = nft_expr_clone(expr, set->exprs[i]);
5582                 if (err < 0) {
5583                         kfree(expr);
5584                         goto err_expr;
5585                 }
5586                 expr_array[i] = expr;
5587         }
5588
5589         return 0;
5590
5591 err_expr:
5592         for (k = i - 1; k >= 0; k--)
5593                 nft_expr_destroy(ctx, expr_array[k]);
5594
5595         return -ENOMEM;
5596 }
5597
5598 static int nft_set_elem_expr_setup(struct nft_ctx *ctx,
5599                                    const struct nft_set_ext *ext,
5600                                    struct nft_expr *expr_array[],
5601                                    u32 num_exprs)
5602 {
5603         struct nft_set_elem_expr *elem_expr = nft_set_ext_expr(ext);
5604         struct nft_expr *expr;
5605         int i, err;
5606
5607         for (i = 0; i < num_exprs; i++) {
5608                 expr = nft_setelem_expr_at(elem_expr, elem_expr->size);
5609                 err = nft_expr_clone(expr, expr_array[i]);
5610                 if (err < 0)
5611                         goto err_elem_expr_setup;
5612
5613                 elem_expr->size += expr_array[i]->ops->size;
5614                 nft_expr_destroy(ctx, expr_array[i]);
5615                 expr_array[i] = NULL;
5616         }
5617
5618         return 0;
5619
5620 err_elem_expr_setup:
5621         for (; i < num_exprs; i++) {
5622                 nft_expr_destroy(ctx, expr_array[i]);
5623                 expr_array[i] = NULL;
5624         }
5625
5626         return -ENOMEM;
5627 }
5628
5629 struct nft_set_ext *nft_set_catchall_lookup(const struct net *net,
5630                                             const struct nft_set *set)
5631 {
5632         struct nft_set_elem_catchall *catchall;
5633         u8 genmask = nft_genmask_cur(net);
5634         struct nft_set_ext *ext;
5635
5636         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5637                 ext = nft_set_elem_ext(set, catchall->elem);
5638                 if (nft_set_elem_active(ext, genmask) &&
5639                     !nft_set_elem_expired(ext))
5640                         return ext;
5641         }
5642
5643         return NULL;
5644 }
5645 EXPORT_SYMBOL_GPL(nft_set_catchall_lookup);
5646
5647 void *nft_set_catchall_gc(const struct nft_set *set)
5648 {
5649         struct nft_set_elem_catchall *catchall, *next;
5650         struct nft_set_ext *ext;
5651         void *elem = NULL;
5652
5653         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5654                 ext = nft_set_elem_ext(set, catchall->elem);
5655
5656                 if (!nft_set_elem_expired(ext) ||
5657                     nft_set_elem_mark_busy(ext))
5658                         continue;
5659
5660                 elem = catchall->elem;
5661                 list_del_rcu(&catchall->list);
5662                 kfree_rcu(catchall, rcu);
5663                 break;
5664         }
5665
5666         return elem;
5667 }
5668 EXPORT_SYMBOL_GPL(nft_set_catchall_gc);
5669
5670 static int nft_setelem_catchall_insert(const struct net *net,
5671                                        struct nft_set *set,
5672                                        const struct nft_set_elem *elem,
5673                                        struct nft_set_ext **pext)
5674 {
5675         struct nft_set_elem_catchall *catchall;
5676         u8 genmask = nft_genmask_next(net);
5677         struct nft_set_ext *ext;
5678
5679         list_for_each_entry(catchall, &set->catchall_list, list) {
5680                 ext = nft_set_elem_ext(set, catchall->elem);
5681                 if (nft_set_elem_active(ext, genmask)) {
5682                         *pext = ext;
5683                         return -EEXIST;
5684                 }
5685         }
5686
5687         catchall = kmalloc(sizeof(*catchall), GFP_KERNEL);
5688         if (!catchall)
5689                 return -ENOMEM;
5690
5691         catchall->elem = elem->priv;
5692         list_add_tail_rcu(&catchall->list, &set->catchall_list);
5693
5694         return 0;
5695 }
5696
5697 static int nft_setelem_insert(const struct net *net,
5698                               struct nft_set *set,
5699                               const struct nft_set_elem *elem,
5700                               struct nft_set_ext **ext, unsigned int flags)
5701 {
5702         int ret;
5703
5704         if (flags & NFT_SET_ELEM_CATCHALL)
5705                 ret = nft_setelem_catchall_insert(net, set, elem, ext);
5706         else
5707                 ret = set->ops->insert(net, set, elem, ext);
5708
5709         return ret;
5710 }
5711
5712 static bool nft_setelem_is_catchall(const struct nft_set *set,
5713                                     const struct nft_set_elem *elem)
5714 {
5715         struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5716
5717         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
5718             *nft_set_ext_flags(ext) & NFT_SET_ELEM_CATCHALL)
5719                 return true;
5720
5721         return false;
5722 }
5723
5724 static void nft_setelem_activate(struct net *net, struct nft_set *set,
5725                                  struct nft_set_elem *elem)
5726 {
5727         struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5728
5729         if (nft_setelem_is_catchall(set, elem)) {
5730                 nft_set_elem_change_active(net, set, ext);
5731                 nft_set_elem_clear_busy(ext);
5732         } else {
5733                 set->ops->activate(net, set, elem);
5734         }
5735 }
5736
5737 static int nft_setelem_catchall_deactivate(const struct net *net,
5738                                            struct nft_set *set,
5739                                            struct nft_set_elem *elem)
5740 {
5741         struct nft_set_elem_catchall *catchall;
5742         struct nft_set_ext *ext;
5743
5744         list_for_each_entry(catchall, &set->catchall_list, list) {
5745                 ext = nft_set_elem_ext(set, catchall->elem);
5746                 if (!nft_is_active(net, ext) ||
5747                     nft_set_elem_mark_busy(ext))
5748                         continue;
5749
5750                 kfree(elem->priv);
5751                 elem->priv = catchall->elem;
5752                 nft_set_elem_change_active(net, set, ext);
5753                 return 0;
5754         }
5755
5756         return -ENOENT;
5757 }
5758
5759 static int __nft_setelem_deactivate(const struct net *net,
5760                                     struct nft_set *set,
5761                                     struct nft_set_elem *elem)
5762 {
5763         void *priv;
5764
5765         priv = set->ops->deactivate(net, set, elem);
5766         if (!priv)
5767                 return -ENOENT;
5768
5769         kfree(elem->priv);
5770         elem->priv = priv;
5771         set->ndeact++;
5772
5773         return 0;
5774 }
5775
5776 static int nft_setelem_deactivate(const struct net *net,
5777                                   struct nft_set *set,
5778                                   struct nft_set_elem *elem, u32 flags)
5779 {
5780         int ret;
5781
5782         if (flags & NFT_SET_ELEM_CATCHALL)
5783                 ret = nft_setelem_catchall_deactivate(net, set, elem);
5784         else
5785                 ret = __nft_setelem_deactivate(net, set, elem);
5786
5787         return ret;
5788 }
5789
5790 static void nft_setelem_catchall_remove(const struct net *net,
5791                                         const struct nft_set *set,
5792                                         const struct nft_set_elem *elem)
5793 {
5794         struct nft_set_elem_catchall *catchall, *next;
5795
5796         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5797                 if (catchall->elem == elem->priv) {
5798                         list_del_rcu(&catchall->list);
5799                         kfree_rcu(catchall, rcu);
5800                         break;
5801                 }
5802         }
5803 }
5804
5805 static void nft_setelem_remove(const struct net *net,
5806                                const struct nft_set *set,
5807                                const struct nft_set_elem *elem)
5808 {
5809         if (nft_setelem_is_catchall(set, elem))
5810                 nft_setelem_catchall_remove(net, set, elem);
5811         else
5812                 set->ops->remove(net, set, elem);
5813 }
5814
5815 static bool nft_setelem_valid_key_end(const struct nft_set *set,
5816                                       struct nlattr **nla, u32 flags)
5817 {
5818         if ((set->flags & (NFT_SET_CONCAT | NFT_SET_INTERVAL)) ==
5819                           (NFT_SET_CONCAT | NFT_SET_INTERVAL)) {
5820                 if (flags & NFT_SET_ELEM_INTERVAL_END)
5821                         return false;
5822                 if (!nla[NFTA_SET_ELEM_KEY_END] &&
5823                     !(flags & NFT_SET_ELEM_CATCHALL))
5824                         return false;
5825         } else {
5826                 if (nla[NFTA_SET_ELEM_KEY_END])
5827                         return false;
5828         }
5829
5830         return true;
5831 }
5832
5833 static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5834                             const struct nlattr *attr, u32 nlmsg_flags)
5835 {
5836         struct nft_expr *expr_array[NFT_SET_EXPR_MAX] = {};
5837         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5838         u8 genmask = nft_genmask_next(ctx->net);
5839         u32 flags = 0, size = 0, num_exprs = 0;
5840         struct nft_set_ext_tmpl tmpl;
5841         struct nft_set_ext *ext, *ext2;
5842         struct nft_set_elem elem;
5843         struct nft_set_binding *binding;
5844         struct nft_object *obj = NULL;
5845         struct nft_userdata *udata;
5846         struct nft_data_desc desc;
5847         enum nft_registers dreg;
5848         struct nft_trans *trans;
5849         u64 timeout;
5850         u64 expiration;
5851         int err, i;
5852         u8 ulen;
5853
5854         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5855                                           nft_set_elem_policy, NULL);
5856         if (err < 0)
5857                 return err;
5858
5859         nft_set_ext_prepare(&tmpl);
5860
5861         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5862         if (err < 0)
5863                 return err;
5864
5865         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5866                 return -EINVAL;
5867
5868         if (flags != 0) {
5869                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
5870                 if (err < 0)
5871                         return err;
5872         }
5873
5874         if (set->flags & NFT_SET_MAP) {
5875                 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
5876                     !(flags & NFT_SET_ELEM_INTERVAL_END))
5877                         return -EINVAL;
5878         } else {
5879                 if (nla[NFTA_SET_ELEM_DATA] != NULL)
5880                         return -EINVAL;
5881         }
5882
5883         if (set->flags & NFT_SET_OBJECT) {
5884                 if (!nla[NFTA_SET_ELEM_OBJREF] &&
5885                     !(flags & NFT_SET_ELEM_INTERVAL_END))
5886                         return -EINVAL;
5887         } else {
5888                 if (nla[NFTA_SET_ELEM_OBJREF])
5889                         return -EINVAL;
5890         }
5891
5892         if (!nft_setelem_valid_key_end(set, nla, flags))
5893                 return -EINVAL;
5894
5895         if ((flags & NFT_SET_ELEM_INTERVAL_END) &&
5896              (nla[NFTA_SET_ELEM_DATA] ||
5897               nla[NFTA_SET_ELEM_OBJREF] ||
5898               nla[NFTA_SET_ELEM_TIMEOUT] ||
5899               nla[NFTA_SET_ELEM_EXPIRATION] ||
5900               nla[NFTA_SET_ELEM_USERDATA] ||
5901               nla[NFTA_SET_ELEM_EXPR] ||
5902               nla[NFTA_SET_ELEM_KEY_END] ||
5903               nla[NFTA_SET_ELEM_EXPRESSIONS]))
5904                 return -EINVAL;
5905
5906         timeout = 0;
5907         if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
5908                 if (!(set->flags & NFT_SET_TIMEOUT))
5909                         return -EINVAL;
5910                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
5911                                             &timeout);
5912                 if (err)
5913                         return err;
5914         } else if (set->flags & NFT_SET_TIMEOUT) {
5915                 timeout = set->timeout;
5916         }
5917
5918         expiration = 0;
5919         if (nla[NFTA_SET_ELEM_EXPIRATION] != NULL) {
5920                 if (!(set->flags & NFT_SET_TIMEOUT))
5921                         return -EINVAL;
5922                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_EXPIRATION],
5923                                             &expiration);
5924                 if (err)
5925                         return err;
5926         }
5927
5928         if (nla[NFTA_SET_ELEM_EXPR]) {
5929                 struct nft_expr *expr;
5930
5931                 if (set->num_exprs && set->num_exprs != 1)
5932                         return -EOPNOTSUPP;
5933
5934                 expr = nft_set_elem_expr_alloc(ctx, set,
5935                                                nla[NFTA_SET_ELEM_EXPR]);
5936                 if (IS_ERR(expr))
5937                         return PTR_ERR(expr);
5938
5939                 expr_array[0] = expr;
5940                 num_exprs = 1;
5941
5942                 if (set->num_exprs && set->exprs[0]->ops != expr->ops) {
5943                         err = -EOPNOTSUPP;
5944                         goto err_set_elem_expr;
5945                 }
5946         } else if (nla[NFTA_SET_ELEM_EXPRESSIONS]) {
5947                 struct nft_expr *expr;
5948                 struct nlattr *tmp;
5949                 int left;
5950
5951                 i = 0;
5952                 nla_for_each_nested(tmp, nla[NFTA_SET_ELEM_EXPRESSIONS], left) {
5953                         if (i == NFT_SET_EXPR_MAX ||
5954                             (set->num_exprs && set->num_exprs == i)) {
5955                                 err = -E2BIG;
5956                                 goto err_set_elem_expr;
5957                         }
5958                         if (nla_type(tmp) != NFTA_LIST_ELEM) {
5959                                 err = -EINVAL;
5960                                 goto err_set_elem_expr;
5961                         }
5962                         expr = nft_set_elem_expr_alloc(ctx, set, tmp);
5963                         if (IS_ERR(expr)) {
5964                                 err = PTR_ERR(expr);
5965                                 goto err_set_elem_expr;
5966                         }
5967                         expr_array[i] = expr;
5968                         num_exprs++;
5969
5970                         if (set->num_exprs && expr->ops != set->exprs[i]->ops) {
5971                                 err = -EOPNOTSUPP;
5972                                 goto err_set_elem_expr;
5973                         }
5974                         i++;
5975                 }
5976                 if (set->num_exprs && set->num_exprs != i) {
5977                         err = -EOPNOTSUPP;
5978                         goto err_set_elem_expr;
5979                 }
5980         } else if (set->num_exprs > 0) {
5981                 err = nft_set_elem_expr_clone(ctx, set, expr_array);
5982                 if (err < 0)
5983                         goto err_set_elem_expr_clone;
5984
5985                 num_exprs = set->num_exprs;
5986         }
5987
5988         if (nla[NFTA_SET_ELEM_KEY]) {
5989                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5990                                             nla[NFTA_SET_ELEM_KEY]);
5991                 if (err < 0)
5992                         goto err_set_elem_expr;
5993
5994                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
5995                 if (err < 0)
5996                         goto err_parse_key;
5997         }
5998
5999         if (nla[NFTA_SET_ELEM_KEY_END]) {
6000                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6001                                             nla[NFTA_SET_ELEM_KEY_END]);
6002                 if (err < 0)
6003                         goto err_parse_key;
6004
6005                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6006                 if (err < 0)
6007                         goto err_parse_key_end;
6008         }
6009
6010         if (timeout > 0) {
6011                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
6012                 if (err < 0)
6013                         goto err_parse_key_end;
6014
6015                 if (timeout != set->timeout) {
6016                         err = nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
6017                         if (err < 0)
6018                                 goto err_parse_key_end;
6019                 }
6020         }
6021
6022         if (num_exprs) {
6023                 for (i = 0; i < num_exprs; i++)
6024                         size += expr_array[i]->ops->size;
6025
6026                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_EXPRESSIONS,
6027                                              sizeof(struct nft_set_elem_expr) + size);
6028                 if (err < 0)
6029                         goto err_parse_key_end;
6030         }
6031
6032         if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
6033                 obj = nft_obj_lookup(ctx->net, ctx->table,
6034                                      nla[NFTA_SET_ELEM_OBJREF],
6035                                      set->objtype, genmask);
6036                 if (IS_ERR(obj)) {
6037                         err = PTR_ERR(obj);
6038                         goto err_parse_key_end;
6039                 }
6040                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
6041                 if (err < 0)
6042                         goto err_parse_key_end;
6043         }
6044
6045         if (nla[NFTA_SET_ELEM_DATA] != NULL) {
6046                 err = nft_setelem_parse_data(ctx, set, &desc, &elem.data.val,
6047                                              nla[NFTA_SET_ELEM_DATA]);
6048                 if (err < 0)
6049                         goto err_parse_key_end;
6050
6051                 dreg = nft_type_to_reg(set->dtype);
6052                 list_for_each_entry(binding, &set->bindings, list) {
6053                         struct nft_ctx bind_ctx = {
6054                                 .net    = ctx->net,
6055                                 .family = ctx->family,
6056                                 .table  = ctx->table,
6057                                 .chain  = (struct nft_chain *)binding->chain,
6058                         };
6059
6060                         if (!(binding->flags & NFT_SET_MAP))
6061                                 continue;
6062
6063                         err = nft_validate_register_store(&bind_ctx, dreg,
6064                                                           &elem.data.val,
6065                                                           desc.type, desc.len);
6066                         if (err < 0)
6067                                 goto err_parse_data;
6068
6069                         if (desc.type == NFT_DATA_VERDICT &&
6070                             (elem.data.val.verdict.code == NFT_GOTO ||
6071                              elem.data.val.verdict.code == NFT_JUMP))
6072                                 nft_validate_state_update(ctx->net,
6073                                                           NFT_VALIDATE_NEED);
6074                 }
6075
6076                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, desc.len);
6077                 if (err < 0)
6078                         goto err_parse_data;
6079         }
6080
6081         /* The full maximum length of userdata can exceed the maximum
6082          * offset value (U8_MAX) for following extensions, therefor it
6083          * must be the last extension added.
6084          */
6085         ulen = 0;
6086         if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
6087                 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
6088                 if (ulen > 0) {
6089                         err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
6090                                                      ulen);
6091                         if (err < 0)
6092                                 goto err_parse_data;
6093                 }
6094         }
6095
6096         err = -ENOMEM;
6097         elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6098                                       elem.key_end.val.data, elem.data.val.data,
6099                                       timeout, expiration, GFP_KERNEL_ACCOUNT);
6100         if (elem.priv == NULL)
6101                 goto err_parse_data;
6102
6103         ext = nft_set_elem_ext(set, elem.priv);
6104         if (flags)
6105                 *nft_set_ext_flags(ext) = flags;
6106         if (ulen > 0) {
6107                 udata = nft_set_ext_userdata(ext);
6108                 udata->len = ulen - 1;
6109                 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
6110         }
6111         if (obj) {
6112                 *nft_set_ext_obj(ext) = obj;
6113                 obj->use++;
6114         }
6115         err = nft_set_elem_expr_setup(ctx, ext, expr_array, num_exprs);
6116         if (err < 0)
6117                 goto err_elem_expr;
6118
6119         trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
6120         if (trans == NULL) {
6121                 err = -ENOMEM;
6122                 goto err_elem_expr;
6123         }
6124
6125         ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
6126
6127         err = nft_setelem_insert(ctx->net, set, &elem, &ext2, flags);
6128         if (err) {
6129                 if (err == -EEXIST) {
6130                         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
6131                             nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
6132                             nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
6133                             nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF))
6134                                 goto err_element_clash;
6135                         if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
6136                              nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
6137                              memcmp(nft_set_ext_data(ext),
6138                                     nft_set_ext_data(ext2), set->dlen) != 0) ||
6139                             (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
6140                              nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
6141                              *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
6142                                 goto err_element_clash;
6143                         else if (!(nlmsg_flags & NLM_F_EXCL))
6144                                 err = 0;
6145                 } else if (err == -ENOTEMPTY) {
6146                         /* ENOTEMPTY reports overlapping between this element
6147                          * and an existing one.
6148                          */
6149                         err = -EEXIST;
6150                 }
6151                 goto err_element_clash;
6152         }
6153
6154         if (!(flags & NFT_SET_ELEM_CATCHALL) && set->size &&
6155             !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact)) {
6156                 err = -ENFILE;
6157                 goto err_set_full;
6158         }
6159
6160         nft_trans_elem(trans) = elem;
6161         nft_trans_commit_list_add_tail(ctx->net, trans);
6162         return 0;
6163
6164 err_set_full:
6165         nft_setelem_remove(ctx->net, set, &elem);
6166 err_element_clash:
6167         kfree(trans);
6168 err_elem_expr:
6169         if (obj)
6170                 obj->use--;
6171
6172         nf_tables_set_elem_destroy(ctx, set, elem.priv);
6173 err_parse_data:
6174         if (nla[NFTA_SET_ELEM_DATA] != NULL)
6175                 nft_data_release(&elem.data.val, desc.type);
6176 err_parse_key_end:
6177         nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
6178 err_parse_key:
6179         nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6180 err_set_elem_expr:
6181         for (i = 0; i < num_exprs && expr_array[i]; i++)
6182                 nft_expr_destroy(ctx, expr_array[i]);
6183 err_set_elem_expr_clone:
6184         return err;
6185 }
6186
6187 static int nf_tables_newsetelem(struct sk_buff *skb,
6188                                 const struct nfnl_info *info,
6189                                 const struct nlattr * const nla[])
6190 {
6191         struct nftables_pernet *nft_net = nft_pernet(info->net);
6192         struct netlink_ext_ack *extack = info->extack;
6193         u8 genmask = nft_genmask_next(info->net);
6194         u8 family = info->nfmsg->nfgen_family;
6195         struct net *net = info->net;
6196         const struct nlattr *attr;
6197         struct nft_table *table;
6198         struct nft_set *set;
6199         struct nft_ctx ctx;
6200         int rem, err;
6201
6202         if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
6203                 return -EINVAL;
6204
6205         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6206                                  genmask, NETLINK_CB(skb).portid);
6207         if (IS_ERR(table)) {
6208                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6209                 return PTR_ERR(table);
6210         }
6211
6212         set = nft_set_lookup_global(net, table, nla[NFTA_SET_ELEM_LIST_SET],
6213                                     nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
6214         if (IS_ERR(set))
6215                 return PTR_ERR(set);
6216
6217         if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6218                 return -EBUSY;
6219
6220         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6221
6222         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6223                 err = nft_add_set_elem(&ctx, set, attr, info->nlh->nlmsg_flags);
6224                 if (err < 0) {
6225                         NL_SET_BAD_ATTR(extack, attr);
6226                         return err;
6227                 }
6228         }
6229
6230         if (nft_net->validate_state == NFT_VALIDATE_DO)
6231                 return nft_table_validate(net, table);
6232
6233         return 0;
6234 }
6235
6236 /**
6237  *      nft_data_hold - hold a nft_data item
6238  *
6239  *      @data: struct nft_data to release
6240  *      @type: type of data
6241  *
6242  *      Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
6243  *      NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
6244  *      NFT_GOTO verdicts. This function must be called on active data objects
6245  *      from the second phase of the commit protocol.
6246  */
6247 void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
6248 {
6249         struct nft_chain *chain;
6250         struct nft_rule *rule;
6251
6252         if (type == NFT_DATA_VERDICT) {
6253                 switch (data->verdict.code) {
6254                 case NFT_JUMP:
6255                 case NFT_GOTO:
6256                         chain = data->verdict.chain;
6257                         chain->use++;
6258
6259                         if (!nft_chain_is_bound(chain))
6260                                 break;
6261
6262                         chain->table->use++;
6263                         list_for_each_entry(rule, &chain->rules, list)
6264                                 chain->use++;
6265
6266                         nft_chain_add(chain->table, chain);
6267                         break;
6268                 }
6269         }
6270 }
6271
6272 static void nft_setelem_data_activate(const struct net *net,
6273                                       const struct nft_set *set,
6274                                       struct nft_set_elem *elem)
6275 {
6276         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6277
6278         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6279                 nft_data_hold(nft_set_ext_data(ext), set->dtype);
6280         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6281                 (*nft_set_ext_obj(ext))->use++;
6282 }
6283
6284 static void nft_setelem_data_deactivate(const struct net *net,
6285                                         const struct nft_set *set,
6286                                         struct nft_set_elem *elem)
6287 {
6288         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6289
6290         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6291                 nft_data_release(nft_set_ext_data(ext), set->dtype);
6292         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6293                 (*nft_set_ext_obj(ext))->use--;
6294 }
6295
6296 static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
6297                            const struct nlattr *attr)
6298 {
6299         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6300         struct nft_set_ext_tmpl tmpl;
6301         struct nft_set_elem elem;
6302         struct nft_set_ext *ext;
6303         struct nft_trans *trans;
6304         u32 flags = 0;
6305         int err;
6306
6307         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6308                                           nft_set_elem_policy, NULL);
6309         if (err < 0)
6310                 return err;
6311
6312         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6313         if (err < 0)
6314                 return err;
6315
6316         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
6317                 return -EINVAL;
6318
6319         if (!nft_setelem_valid_key_end(set, nla, flags))
6320                 return -EINVAL;
6321
6322         nft_set_ext_prepare(&tmpl);
6323
6324         if (flags != 0) {
6325                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
6326                 if (err < 0)
6327                         return err;
6328         }
6329
6330         if (nla[NFTA_SET_ELEM_KEY]) {
6331                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6332                                             nla[NFTA_SET_ELEM_KEY]);
6333                 if (err < 0)
6334                         return err;
6335
6336                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6337                 if (err < 0)
6338                         goto fail_elem;
6339         }
6340
6341         if (nla[NFTA_SET_ELEM_KEY_END]) {
6342                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6343                                             nla[NFTA_SET_ELEM_KEY_END]);
6344                 if (err < 0)
6345                         goto fail_elem;
6346
6347                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6348                 if (err < 0)
6349                         goto fail_elem_key_end;
6350         }
6351
6352         err = -ENOMEM;
6353         elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6354                                       elem.key_end.val.data, NULL, 0, 0,
6355                                       GFP_KERNEL_ACCOUNT);
6356         if (elem.priv == NULL)
6357                 goto fail_elem_key_end;
6358
6359         ext = nft_set_elem_ext(set, elem.priv);
6360         if (flags)
6361                 *nft_set_ext_flags(ext) = flags;
6362
6363         trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
6364         if (trans == NULL)
6365                 goto fail_trans;
6366
6367         err = nft_setelem_deactivate(ctx->net, set, &elem, flags);
6368         if (err < 0)
6369                 goto fail_ops;
6370
6371         nft_setelem_data_deactivate(ctx->net, set, &elem);
6372
6373         nft_trans_elem(trans) = elem;
6374         nft_trans_commit_list_add_tail(ctx->net, trans);
6375         return 0;
6376
6377 fail_ops:
6378         kfree(trans);
6379 fail_trans:
6380         kfree(elem.priv);
6381 fail_elem_key_end:
6382         nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
6383 fail_elem:
6384         nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6385         return err;
6386 }
6387
6388 static int nft_setelem_flush(const struct nft_ctx *ctx,
6389                              struct nft_set *set,
6390                              const struct nft_set_iter *iter,
6391                              struct nft_set_elem *elem)
6392 {
6393         struct nft_trans *trans;
6394         int err;
6395
6396         trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6397                                     sizeof(struct nft_trans_elem), GFP_ATOMIC);
6398         if (!trans)
6399                 return -ENOMEM;
6400
6401         if (!set->ops->flush(ctx->net, set, elem->priv)) {
6402                 err = -ENOENT;
6403                 goto err1;
6404         }
6405         set->ndeact++;
6406
6407         nft_setelem_data_deactivate(ctx->net, set, elem);
6408         nft_trans_elem_set(trans) = set;
6409         nft_trans_elem(trans) = *elem;
6410         nft_trans_commit_list_add_tail(ctx->net, trans);
6411
6412         return 0;
6413 err1:
6414         kfree(trans);
6415         return err;
6416 }
6417
6418 static int __nft_set_catchall_flush(const struct nft_ctx *ctx,
6419                                     struct nft_set *set,
6420                                     struct nft_set_elem *elem)
6421 {
6422         struct nft_trans *trans;
6423
6424         trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6425                                     sizeof(struct nft_trans_elem), GFP_KERNEL);
6426         if (!trans)
6427                 return -ENOMEM;
6428
6429         nft_setelem_data_deactivate(ctx->net, set, elem);
6430         nft_trans_elem_set(trans) = set;
6431         nft_trans_elem(trans) = *elem;
6432         nft_trans_commit_list_add_tail(ctx->net, trans);
6433
6434         return 0;
6435 }
6436
6437 static int nft_set_catchall_flush(const struct nft_ctx *ctx,
6438                                   struct nft_set *set)
6439 {
6440         u8 genmask = nft_genmask_next(ctx->net);
6441         struct nft_set_elem_catchall *catchall;
6442         struct nft_set_elem elem;
6443         struct nft_set_ext *ext;
6444         int ret = 0;
6445
6446         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6447                 ext = nft_set_elem_ext(set, catchall->elem);
6448                 if (!nft_set_elem_active(ext, genmask) ||
6449                     nft_set_elem_mark_busy(ext))
6450                         continue;
6451
6452                 elem.priv = catchall->elem;
6453                 ret = __nft_set_catchall_flush(ctx, set, &elem);
6454                 if (ret < 0)
6455                         break;
6456         }
6457
6458         return ret;
6459 }
6460
6461 static int nft_set_flush(struct nft_ctx *ctx, struct nft_set *set, u8 genmask)
6462 {
6463         struct nft_set_iter iter = {
6464                 .genmask        = genmask,
6465                 .fn             = nft_setelem_flush,
6466         };
6467
6468         set->ops->walk(ctx, set, &iter);
6469         if (!iter.err)
6470                 iter.err = nft_set_catchall_flush(ctx, set);
6471
6472         return iter.err;
6473 }
6474
6475 static int nf_tables_delsetelem(struct sk_buff *skb,
6476                                 const struct nfnl_info *info,
6477                                 const struct nlattr * const nla[])
6478 {
6479         struct netlink_ext_ack *extack = info->extack;
6480         u8 genmask = nft_genmask_next(info->net);
6481         u8 family = info->nfmsg->nfgen_family;
6482         struct net *net = info->net;
6483         const struct nlattr *attr;
6484         struct nft_table *table;
6485         struct nft_set *set;
6486         struct nft_ctx ctx;
6487         int rem, err = 0;
6488
6489         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6490                                  genmask, NETLINK_CB(skb).portid);
6491         if (IS_ERR(table)) {
6492                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6493                 return PTR_ERR(table);
6494         }
6495
6496         set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
6497         if (IS_ERR(set))
6498                 return PTR_ERR(set);
6499         if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6500                 return -EBUSY;
6501
6502         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6503
6504         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
6505                 return nft_set_flush(&ctx, set, genmask);
6506
6507         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6508                 err = nft_del_setelem(&ctx, set, attr);
6509                 if (err < 0) {
6510                         NL_SET_BAD_ATTR(extack, attr);
6511                         break;
6512                 }
6513         }
6514         return err;
6515 }
6516
6517 void nft_set_gc_batch_release(struct rcu_head *rcu)
6518 {
6519         struct nft_set_gc_batch *gcb;
6520         unsigned int i;
6521
6522         gcb = container_of(rcu, struct nft_set_gc_batch, head.rcu);
6523         for (i = 0; i < gcb->head.cnt; i++)
6524                 nft_set_elem_destroy(gcb->head.set, gcb->elems[i], true);
6525         kfree(gcb);
6526 }
6527
6528 struct nft_set_gc_batch *nft_set_gc_batch_alloc(const struct nft_set *set,
6529                                                 gfp_t gfp)
6530 {
6531         struct nft_set_gc_batch *gcb;
6532
6533         gcb = kzalloc(sizeof(*gcb), gfp);
6534         if (gcb == NULL)
6535                 return gcb;
6536         gcb->head.set = set;
6537         return gcb;
6538 }
6539
6540 /*
6541  * Stateful objects
6542  */
6543
6544 /**
6545  *      nft_register_obj- register nf_tables stateful object type
6546  *      @obj_type: object type
6547  *
6548  *      Registers the object type for use with nf_tables. Returns zero on
6549  *      success or a negative errno code otherwise.
6550  */
6551 int nft_register_obj(struct nft_object_type *obj_type)
6552 {
6553         if (obj_type->type == NFT_OBJECT_UNSPEC)
6554                 return -EINVAL;
6555
6556         nfnl_lock(NFNL_SUBSYS_NFTABLES);
6557         list_add_rcu(&obj_type->list, &nf_tables_objects);
6558         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6559         return 0;
6560 }
6561 EXPORT_SYMBOL_GPL(nft_register_obj);
6562
6563 /**
6564  *      nft_unregister_obj - unregister nf_tables object type
6565  *      @obj_type: object type
6566  *
6567  *      Unregisters the object type for use with nf_tables.
6568  */
6569 void nft_unregister_obj(struct nft_object_type *obj_type)
6570 {
6571         nfnl_lock(NFNL_SUBSYS_NFTABLES);
6572         list_del_rcu(&obj_type->list);
6573         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6574 }
6575 EXPORT_SYMBOL_GPL(nft_unregister_obj);
6576
6577 struct nft_object *nft_obj_lookup(const struct net *net,
6578                                   const struct nft_table *table,
6579                                   const struct nlattr *nla, u32 objtype,
6580                                   u8 genmask)
6581 {
6582         struct nft_object_hash_key k = { .table = table };
6583         char search[NFT_OBJ_MAXNAMELEN];
6584         struct rhlist_head *tmp, *list;
6585         struct nft_object *obj;
6586
6587         nla_strscpy(search, nla, sizeof(search));
6588         k.name = search;
6589
6590         WARN_ON_ONCE(!rcu_read_lock_held() &&
6591                      !lockdep_commit_lock_is_held(net));
6592
6593         rcu_read_lock();
6594         list = rhltable_lookup(&nft_objname_ht, &k, nft_objname_ht_params);
6595         if (!list)
6596                 goto out;
6597
6598         rhl_for_each_entry_rcu(obj, tmp, list, rhlhead) {
6599                 if (objtype == obj->ops->type->type &&
6600                     nft_active_genmask(obj, genmask)) {
6601                         rcu_read_unlock();
6602                         return obj;
6603                 }
6604         }
6605 out:
6606         rcu_read_unlock();
6607         return ERR_PTR(-ENOENT);
6608 }
6609 EXPORT_SYMBOL_GPL(nft_obj_lookup);
6610
6611 static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
6612                                                   const struct nlattr *nla,
6613                                                   u32 objtype, u8 genmask)
6614 {
6615         struct nft_object *obj;
6616
6617         list_for_each_entry(obj, &table->objects, list) {
6618                 if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
6619                     objtype == obj->ops->type->type &&
6620                     nft_active_genmask(obj, genmask))
6621                         return obj;
6622         }
6623         return ERR_PTR(-ENOENT);
6624 }
6625
6626 static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
6627         [NFTA_OBJ_TABLE]        = { .type = NLA_STRING,
6628                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
6629         [NFTA_OBJ_NAME]         = { .type = NLA_STRING,
6630                                     .len = NFT_OBJ_MAXNAMELEN - 1 },
6631         [NFTA_OBJ_TYPE]         = { .type = NLA_U32 },
6632         [NFTA_OBJ_DATA]         = { .type = NLA_NESTED },
6633         [NFTA_OBJ_HANDLE]       = { .type = NLA_U64},
6634         [NFTA_OBJ_USERDATA]     = { .type = NLA_BINARY,
6635                                     .len = NFT_USERDATA_MAXLEN },
6636 };
6637
6638 static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
6639                                        const struct nft_object_type *type,
6640                                        const struct nlattr *attr)
6641 {
6642         struct nlattr **tb;
6643         const struct nft_object_ops *ops;
6644         struct nft_object *obj;
6645         int err = -ENOMEM;
6646
6647         tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
6648         if (!tb)
6649                 goto err1;
6650
6651         if (attr) {
6652                 err = nla_parse_nested_deprecated(tb, type->maxattr, attr,
6653                                                   type->policy, NULL);
6654                 if (err < 0)
6655                         goto err2;
6656         } else {
6657                 memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
6658         }
6659
6660         if (type->select_ops) {
6661                 ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
6662                 if (IS_ERR(ops)) {
6663                         err = PTR_ERR(ops);
6664                         goto err2;
6665                 }
6666         } else {
6667                 ops = type->ops;
6668         }
6669
6670         err = -ENOMEM;
6671         obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL_ACCOUNT);
6672         if (!obj)
6673                 goto err2;
6674
6675         err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
6676         if (err < 0)
6677                 goto err3;
6678
6679         obj->ops = ops;
6680
6681         kfree(tb);
6682         return obj;
6683 err3:
6684         kfree(obj);
6685 err2:
6686         kfree(tb);
6687 err1:
6688         return ERR_PTR(err);
6689 }
6690
6691 static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
6692                            struct nft_object *obj, bool reset)
6693 {
6694         struct nlattr *nest;
6695
6696         nest = nla_nest_start_noflag(skb, attr);
6697         if (!nest)
6698                 goto nla_put_failure;
6699         if (obj->ops->dump(skb, obj, reset) < 0)
6700                 goto nla_put_failure;
6701         nla_nest_end(skb, nest);
6702         return 0;
6703
6704 nla_put_failure:
6705         return -1;
6706 }
6707
6708 static const struct nft_object_type *__nft_obj_type_get(u32 objtype)
6709 {
6710         const struct nft_object_type *type;
6711
6712         list_for_each_entry(type, &nf_tables_objects, list) {
6713                 if (objtype == type->type)
6714                         return type;
6715         }
6716         return NULL;
6717 }
6718
6719 static const struct nft_object_type *
6720 nft_obj_type_get(struct net *net, u32 objtype)
6721 {
6722         const struct nft_object_type *type;
6723
6724         type = __nft_obj_type_get(objtype);
6725         if (type != NULL && try_module_get(type->owner))
6726                 return type;
6727
6728         lockdep_nfnl_nft_mutex_not_held();
6729 #ifdef CONFIG_MODULES
6730         if (type == NULL) {
6731                 if (nft_request_module(net, "nft-obj-%u", objtype) == -EAGAIN)
6732                         return ERR_PTR(-EAGAIN);
6733         }
6734 #endif
6735         return ERR_PTR(-ENOENT);
6736 }
6737
6738 static int nf_tables_updobj(const struct nft_ctx *ctx,
6739                             const struct nft_object_type *type,
6740                             const struct nlattr *attr,
6741                             struct nft_object *obj)
6742 {
6743         struct nft_object *newobj;
6744         struct nft_trans *trans;
6745         int err = -ENOMEM;
6746
6747         if (!try_module_get(type->owner))
6748                 return -ENOENT;
6749
6750         trans = nft_trans_alloc(ctx, NFT_MSG_NEWOBJ,
6751                                 sizeof(struct nft_trans_obj));
6752         if (!trans)
6753                 goto err_trans;
6754
6755         newobj = nft_obj_init(ctx, type, attr);
6756         if (IS_ERR(newobj)) {
6757                 err = PTR_ERR(newobj);
6758                 goto err_free_trans;
6759         }
6760
6761         nft_trans_obj(trans) = obj;
6762         nft_trans_obj_update(trans) = true;
6763         nft_trans_obj_newobj(trans) = newobj;
6764         nft_trans_commit_list_add_tail(ctx->net, trans);
6765
6766         return 0;
6767
6768 err_free_trans:
6769         kfree(trans);
6770 err_trans:
6771         module_put(type->owner);
6772         return err;
6773 }
6774
6775 static int nf_tables_newobj(struct sk_buff *skb, const struct nfnl_info *info,
6776                             const struct nlattr * const nla[])
6777 {
6778         struct netlink_ext_ack *extack = info->extack;
6779         u8 genmask = nft_genmask_next(info->net);
6780         u8 family = info->nfmsg->nfgen_family;
6781         const struct nft_object_type *type;
6782         struct net *net = info->net;
6783         struct nft_table *table;
6784         struct nft_object *obj;
6785         struct nft_ctx ctx;
6786         u32 objtype;
6787         int err;
6788
6789         if (!nla[NFTA_OBJ_TYPE] ||
6790             !nla[NFTA_OBJ_NAME] ||
6791             !nla[NFTA_OBJ_DATA])
6792                 return -EINVAL;
6793
6794         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
6795                                  NETLINK_CB(skb).portid);
6796         if (IS_ERR(table)) {
6797                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6798                 return PTR_ERR(table);
6799         }
6800
6801         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6802         obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6803         if (IS_ERR(obj)) {
6804                 err = PTR_ERR(obj);
6805                 if (err != -ENOENT) {
6806                         NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6807                         return err;
6808                 }
6809         } else {
6810                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
6811                         NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6812                         return -EEXIST;
6813                 }
6814                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
6815                         return -EOPNOTSUPP;
6816
6817                 type = __nft_obj_type_get(objtype);
6818                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6819
6820                 return nf_tables_updobj(&ctx, type, nla[NFTA_OBJ_DATA], obj);
6821         }
6822
6823         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6824
6825         type = nft_obj_type_get(net, objtype);
6826         if (IS_ERR(type))
6827                 return PTR_ERR(type);
6828
6829         obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
6830         if (IS_ERR(obj)) {
6831                 err = PTR_ERR(obj);
6832                 goto err_init;
6833         }
6834         obj->key.table = table;
6835         obj->handle = nf_tables_alloc_handle(table);
6836
6837         obj->key.name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL_ACCOUNT);
6838         if (!obj->key.name) {
6839                 err = -ENOMEM;
6840                 goto err_strdup;
6841         }
6842
6843         if (nla[NFTA_OBJ_USERDATA]) {
6844                 obj->udata = nla_memdup(nla[NFTA_OBJ_USERDATA], GFP_KERNEL);
6845                 if (obj->udata == NULL)
6846                         goto err_userdata;
6847
6848                 obj->udlen = nla_len(nla[NFTA_OBJ_USERDATA]);
6849         }
6850
6851         err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
6852         if (err < 0)
6853                 goto err_trans;
6854
6855         err = rhltable_insert(&nft_objname_ht, &obj->rhlhead,
6856                               nft_objname_ht_params);
6857         if (err < 0)
6858                 goto err_obj_ht;
6859
6860         list_add_tail_rcu(&obj->list, &table->objects);
6861         table->use++;
6862         return 0;
6863 err_obj_ht:
6864         /* queued in transaction log */
6865         INIT_LIST_HEAD(&obj->list);
6866         return err;
6867 err_trans:
6868         kfree(obj->udata);
6869 err_userdata:
6870         kfree(obj->key.name);
6871 err_strdup:
6872         if (obj->ops->destroy)
6873                 obj->ops->destroy(&ctx, obj);
6874         kfree(obj);
6875 err_init:
6876         module_put(type->owner);
6877         return err;
6878 }
6879
6880 static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
6881                                    u32 portid, u32 seq, int event, u32 flags,
6882                                    int family, const struct nft_table *table,
6883                                    struct nft_object *obj, bool reset)
6884 {
6885         struct nlmsghdr *nlh;
6886
6887         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
6888         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
6889                            NFNETLINK_V0, nft_base_seq(net));
6890         if (!nlh)
6891                 goto nla_put_failure;
6892
6893         if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
6894             nla_put_string(skb, NFTA_OBJ_NAME, obj->key.name) ||
6895             nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
6896             nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
6897             nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset) ||
6898             nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
6899                          NFTA_OBJ_PAD))
6900                 goto nla_put_failure;
6901
6902         if (obj->udata &&
6903             nla_put(skb, NFTA_OBJ_USERDATA, obj->udlen, obj->udata))
6904                 goto nla_put_failure;
6905
6906         nlmsg_end(skb, nlh);
6907         return 0;
6908
6909 nla_put_failure:
6910         nlmsg_trim(skb, nlh);
6911         return -1;
6912 }
6913
6914 struct nft_obj_filter {
6915         char            *table;
6916         u32             type;
6917 };
6918
6919 static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
6920 {
6921         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
6922         const struct nft_table *table;
6923         unsigned int idx = 0, s_idx = cb->args[0];
6924         struct nft_obj_filter *filter = cb->data;
6925         struct net *net = sock_net(skb->sk);
6926         int family = nfmsg->nfgen_family;
6927         struct nftables_pernet *nft_net;
6928         struct nft_object *obj;
6929         bool reset = false;
6930
6931         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6932                 reset = true;
6933
6934         rcu_read_lock();
6935         nft_net = nft_pernet(net);
6936         cb->seq = READ_ONCE(nft_net->base_seq);
6937
6938         list_for_each_entry_rcu(table, &nft_net->tables, list) {
6939                 if (family != NFPROTO_UNSPEC && family != table->family)
6940                         continue;
6941
6942                 list_for_each_entry_rcu(obj, &table->objects, list) {
6943                         if (!nft_is_active(net, obj))
6944                                 goto cont;
6945                         if (idx < s_idx)
6946                                 goto cont;
6947                         if (idx > s_idx)
6948                                 memset(&cb->args[1], 0,
6949                                        sizeof(cb->args) - sizeof(cb->args[0]));
6950                         if (filter && filter->table &&
6951                             strcmp(filter->table, table->name))
6952                                 goto cont;
6953                         if (filter &&
6954                             filter->type != NFT_OBJECT_UNSPEC &&
6955                             obj->ops->type->type != filter->type)
6956                                 goto cont;
6957                         if (reset) {
6958                                 char *buf = kasprintf(GFP_ATOMIC,
6959                                                       "%s:%u",
6960                                                       table->name,
6961                                                       nft_net->base_seq);
6962
6963                                 audit_log_nfcfg(buf,
6964                                                 family,
6965                                                 obj->handle,
6966                                                 AUDIT_NFT_OP_OBJ_RESET,
6967                                                 GFP_ATOMIC);
6968                                 kfree(buf);
6969                         }
6970
6971                         if (nf_tables_fill_obj_info(skb, net, NETLINK_CB(cb->skb).portid,
6972                                                     cb->nlh->nlmsg_seq,
6973                                                     NFT_MSG_NEWOBJ,
6974                                                     NLM_F_MULTI | NLM_F_APPEND,
6975                                                     table->family, table,
6976                                                     obj, reset) < 0)
6977                                 goto done;
6978
6979                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
6980 cont:
6981                         idx++;
6982                 }
6983         }
6984 done:
6985         rcu_read_unlock();
6986
6987         cb->args[0] = idx;
6988         return skb->len;
6989 }
6990
6991 static int nf_tables_dump_obj_start(struct netlink_callback *cb)
6992 {
6993         const struct nlattr * const *nla = cb->data;
6994         struct nft_obj_filter *filter = NULL;
6995
6996         if (nla[NFTA_OBJ_TABLE] || nla[NFTA_OBJ_TYPE]) {
6997                 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
6998                 if (!filter)
6999                         return -ENOMEM;
7000
7001                 if (nla[NFTA_OBJ_TABLE]) {
7002                         filter->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
7003                         if (!filter->table) {
7004                                 kfree(filter);
7005                                 return -ENOMEM;
7006                         }
7007                 }
7008
7009                 if (nla[NFTA_OBJ_TYPE])
7010                         filter->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7011         }
7012
7013         cb->data = filter;
7014         return 0;
7015 }
7016
7017 static int nf_tables_dump_obj_done(struct netlink_callback *cb)
7018 {
7019         struct nft_obj_filter *filter = cb->data;
7020
7021         if (filter) {
7022                 kfree(filter->table);
7023                 kfree(filter);
7024         }
7025
7026         return 0;
7027 }
7028
7029 /* called with rcu_read_lock held */
7030 static int nf_tables_getobj(struct sk_buff *skb, const struct nfnl_info *info,
7031                             const struct nlattr * const nla[])
7032 {
7033         struct netlink_ext_ack *extack = info->extack;
7034         u8 genmask = nft_genmask_cur(info->net);
7035         u8 family = info->nfmsg->nfgen_family;
7036         const struct nft_table *table;
7037         struct net *net = info->net;
7038         struct nft_object *obj;
7039         struct sk_buff *skb2;
7040         bool reset = false;
7041         u32 objtype;
7042         int err;
7043
7044         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
7045                 struct netlink_dump_control c = {
7046                         .start = nf_tables_dump_obj_start,
7047                         .dump = nf_tables_dump_obj,
7048                         .done = nf_tables_dump_obj_done,
7049                         .module = THIS_MODULE,
7050                         .data = (void *)nla,
7051                 };
7052
7053                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
7054         }
7055
7056         if (!nla[NFTA_OBJ_NAME] ||
7057             !nla[NFTA_OBJ_TYPE])
7058                 return -EINVAL;
7059
7060         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask, 0);
7061         if (IS_ERR(table)) {
7062                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
7063                 return PTR_ERR(table);
7064         }
7065
7066         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7067         obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
7068         if (IS_ERR(obj)) {
7069                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
7070                 return PTR_ERR(obj);
7071         }
7072
7073         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7074         if (!skb2)
7075                 return -ENOMEM;
7076
7077         if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
7078                 reset = true;
7079
7080         if (reset) {
7081                 const struct nftables_pernet *nft_net;
7082                 char *buf;
7083
7084                 nft_net = nft_pernet(net);
7085                 buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, nft_net->base_seq);
7086
7087                 audit_log_nfcfg(buf,
7088                                 family,
7089                                 obj->handle,
7090                                 AUDIT_NFT_OP_OBJ_RESET,
7091                                 GFP_ATOMIC);
7092                 kfree(buf);
7093         }
7094
7095         err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
7096                                       info->nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
7097                                       family, table, obj, reset);
7098         if (err < 0)
7099                 goto err_fill_obj_info;
7100
7101         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7102
7103 err_fill_obj_info:
7104         kfree_skb(skb2);
7105         return err;
7106 }
7107
7108 static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
7109 {
7110         if (obj->ops->destroy)
7111                 obj->ops->destroy(ctx, obj);
7112
7113         module_put(obj->ops->type->owner);
7114         kfree(obj->key.name);
7115         kfree(obj->udata);
7116         kfree(obj);
7117 }
7118
7119 static int nf_tables_delobj(struct sk_buff *skb, const struct nfnl_info *info,
7120                             const struct nlattr * const nla[])
7121 {
7122         struct netlink_ext_ack *extack = info->extack;
7123         u8 genmask = nft_genmask_next(info->net);
7124         u8 family = info->nfmsg->nfgen_family;
7125         struct net *net = info->net;
7126         const struct nlattr *attr;
7127         struct nft_table *table;
7128         struct nft_object *obj;
7129         struct nft_ctx ctx;
7130         u32 objtype;
7131
7132         if (!nla[NFTA_OBJ_TYPE] ||
7133             (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
7134                 return -EINVAL;
7135
7136         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
7137                                  NETLINK_CB(skb).portid);
7138         if (IS_ERR(table)) {
7139                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
7140                 return PTR_ERR(table);
7141         }
7142
7143         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7144         if (nla[NFTA_OBJ_HANDLE]) {
7145                 attr = nla[NFTA_OBJ_HANDLE];
7146                 obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
7147         } else {
7148                 attr = nla[NFTA_OBJ_NAME];
7149                 obj = nft_obj_lookup(net, table, attr, objtype, genmask);
7150         }
7151
7152         if (IS_ERR(obj)) {
7153                 NL_SET_BAD_ATTR(extack, attr);
7154                 return PTR_ERR(obj);
7155         }
7156         if (obj->use > 0) {
7157                 NL_SET_BAD_ATTR(extack, attr);
7158                 return -EBUSY;
7159         }
7160
7161         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7162
7163         return nft_delobj(&ctx, obj);
7164 }
7165
7166 void nft_obj_notify(struct net *net, const struct nft_table *table,
7167                     struct nft_object *obj, u32 portid, u32 seq, int event,
7168                     u16 flags, int family, int report, gfp_t gfp)
7169 {
7170         struct nftables_pernet *nft_net = nft_pernet(net);
7171         struct sk_buff *skb;
7172         int err;
7173         char *buf = kasprintf(gfp, "%s:%u",
7174                               table->name, nft_net->base_seq);
7175
7176         audit_log_nfcfg(buf,
7177                         family,
7178                         obj->handle,
7179                         event == NFT_MSG_NEWOBJ ?
7180                                  AUDIT_NFT_OP_OBJ_REGISTER :
7181                                  AUDIT_NFT_OP_OBJ_UNREGISTER,
7182                         gfp);
7183         kfree(buf);
7184
7185         if (!report &&
7186             !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
7187                 return;
7188
7189         skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
7190         if (skb == NULL)
7191                 goto err;
7192
7193         err = nf_tables_fill_obj_info(skb, net, portid, seq, event,
7194                                       flags & (NLM_F_CREATE | NLM_F_EXCL),
7195                                       family, table, obj, false);
7196         if (err < 0) {
7197                 kfree_skb(skb);
7198                 goto err;
7199         }
7200
7201         nft_notify_enqueue(skb, report, &nft_net->notify_list);
7202         return;
7203 err:
7204         nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
7205 }
7206 EXPORT_SYMBOL_GPL(nft_obj_notify);
7207
7208 static void nf_tables_obj_notify(const struct nft_ctx *ctx,
7209                                  struct nft_object *obj, int event)
7210 {
7211         nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid, ctx->seq, event,
7212                        ctx->flags, ctx->family, ctx->report, GFP_KERNEL);
7213 }
7214
7215 /*
7216  * Flow tables
7217  */
7218 void nft_register_flowtable_type(struct nf_flowtable_type *type)
7219 {
7220         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7221         list_add_tail_rcu(&type->list, &nf_tables_flowtables);
7222         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7223 }
7224 EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
7225
7226 void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
7227 {
7228         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7229         list_del_rcu(&type->list);
7230         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7231 }
7232 EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
7233
7234 static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
7235         [NFTA_FLOWTABLE_TABLE]          = { .type = NLA_STRING,
7236                                             .len = NFT_NAME_MAXLEN - 1 },
7237         [NFTA_FLOWTABLE_NAME]           = { .type = NLA_STRING,
7238                                             .len = NFT_NAME_MAXLEN - 1 },
7239         [NFTA_FLOWTABLE_HOOK]           = { .type = NLA_NESTED },
7240         [NFTA_FLOWTABLE_HANDLE]         = { .type = NLA_U64 },
7241         [NFTA_FLOWTABLE_FLAGS]          = { .type = NLA_U32 },
7242 };
7243
7244 struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
7245                                            const struct nlattr *nla, u8 genmask)
7246 {
7247         struct nft_flowtable *flowtable;
7248
7249         list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7250                 if (!nla_strcmp(nla, flowtable->name) &&
7251                     nft_active_genmask(flowtable, genmask))
7252                         return flowtable;
7253         }
7254         return ERR_PTR(-ENOENT);
7255 }
7256 EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
7257
7258 void nf_tables_deactivate_flowtable(const struct nft_ctx *ctx,
7259                                     struct nft_flowtable *flowtable,
7260                                     enum nft_trans_phase phase)
7261 {
7262         switch (phase) {
7263         case NFT_TRANS_PREPARE:
7264         case NFT_TRANS_ABORT:
7265         case NFT_TRANS_RELEASE:
7266                 flowtable->use--;
7267                 fallthrough;
7268         default:
7269                 return;
7270         }
7271 }
7272 EXPORT_SYMBOL_GPL(nf_tables_deactivate_flowtable);
7273
7274 static struct nft_flowtable *
7275 nft_flowtable_lookup_byhandle(const struct nft_table *table,
7276                               const struct nlattr *nla, u8 genmask)
7277 {
7278        struct nft_flowtable *flowtable;
7279
7280        list_for_each_entry(flowtable, &table->flowtables, list) {
7281                if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
7282                    nft_active_genmask(flowtable, genmask))
7283                        return flowtable;
7284        }
7285        return ERR_PTR(-ENOENT);
7286 }
7287
7288 struct nft_flowtable_hook {
7289         u32                     num;
7290         int                     priority;
7291         struct list_head        list;
7292 };
7293
7294 static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
7295         [NFTA_FLOWTABLE_HOOK_NUM]       = { .type = NLA_U32 },
7296         [NFTA_FLOWTABLE_HOOK_PRIORITY]  = { .type = NLA_U32 },
7297         [NFTA_FLOWTABLE_HOOK_DEVS]      = { .type = NLA_NESTED },
7298 };
7299
7300 static int nft_flowtable_parse_hook(const struct nft_ctx *ctx,
7301                                     const struct nlattr *attr,
7302                                     struct nft_flowtable_hook *flowtable_hook,
7303                                     struct nft_flowtable *flowtable, bool add)
7304 {
7305         struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
7306         struct nft_hook *hook;
7307         int hooknum, priority;
7308         int err;
7309
7310         INIT_LIST_HEAD(&flowtable_hook->list);
7311
7312         err = nla_parse_nested_deprecated(tb, NFTA_FLOWTABLE_HOOK_MAX, attr,
7313                                           nft_flowtable_hook_policy, NULL);
7314         if (err < 0)
7315                 return err;
7316
7317         if (add) {
7318                 if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
7319                     !tb[NFTA_FLOWTABLE_HOOK_PRIORITY])
7320                         return -EINVAL;
7321
7322                 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7323                 if (hooknum != NF_NETDEV_INGRESS)
7324                         return -EOPNOTSUPP;
7325
7326                 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7327
7328                 flowtable_hook->priority        = priority;
7329                 flowtable_hook->num             = hooknum;
7330         } else {
7331                 if (tb[NFTA_FLOWTABLE_HOOK_NUM]) {
7332                         hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7333                         if (hooknum != flowtable->hooknum)
7334                                 return -EOPNOTSUPP;
7335                 }
7336
7337                 if (tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
7338                         priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7339                         if (priority != flowtable->data.priority)
7340                                 return -EOPNOTSUPP;
7341                 }
7342
7343                 flowtable_hook->priority        = flowtable->data.priority;
7344                 flowtable_hook->num             = flowtable->hooknum;
7345         }
7346
7347         if (tb[NFTA_FLOWTABLE_HOOK_DEVS]) {
7348                 err = nf_tables_parse_netdev_hooks(ctx->net,
7349                                                    tb[NFTA_FLOWTABLE_HOOK_DEVS],
7350                                                    &flowtable_hook->list);
7351                 if (err < 0)
7352                         return err;
7353         }
7354
7355         list_for_each_entry(hook, &flowtable_hook->list, list) {
7356                 hook->ops.pf            = NFPROTO_NETDEV;
7357                 hook->ops.hooknum       = flowtable_hook->num;
7358                 hook->ops.priority      = flowtable_hook->priority;
7359                 hook->ops.priv          = &flowtable->data;
7360                 hook->ops.hook          = flowtable->data.type->hook;
7361         }
7362
7363         return err;
7364 }
7365
7366 static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
7367 {
7368         const struct nf_flowtable_type *type;
7369
7370         list_for_each_entry(type, &nf_tables_flowtables, list) {
7371                 if (family == type->family)
7372                         return type;
7373         }
7374         return NULL;
7375 }
7376
7377 static const struct nf_flowtable_type *
7378 nft_flowtable_type_get(struct net *net, u8 family)
7379 {
7380         const struct nf_flowtable_type *type;
7381
7382         type = __nft_flowtable_type_get(family);
7383         if (type != NULL && try_module_get(type->owner))
7384                 return type;
7385
7386         lockdep_nfnl_nft_mutex_not_held();
7387 #ifdef CONFIG_MODULES
7388         if (type == NULL) {
7389                 if (nft_request_module(net, "nf-flowtable-%u", family) == -EAGAIN)
7390                         return ERR_PTR(-EAGAIN);
7391         }
7392 #endif
7393         return ERR_PTR(-ENOENT);
7394 }
7395
7396 /* Only called from error and netdev event paths. */
7397 static void nft_unregister_flowtable_hook(struct net *net,
7398                                           struct nft_flowtable *flowtable,
7399                                           struct nft_hook *hook)
7400 {
7401         nf_unregister_net_hook(net, &hook->ops);
7402         flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7403                                     FLOW_BLOCK_UNBIND);
7404 }
7405
7406 static void __nft_unregister_flowtable_net_hooks(struct net *net,
7407                                                  struct list_head *hook_list,
7408                                                  bool release_netdev)
7409 {
7410         struct nft_hook *hook, *next;
7411
7412         list_for_each_entry_safe(hook, next, hook_list, list) {
7413                 nf_unregister_net_hook(net, &hook->ops);
7414                 if (release_netdev) {
7415                         list_del(&hook->list);
7416                         kfree_rcu(hook, rcu);
7417                 }
7418         }
7419 }
7420
7421 static void nft_unregister_flowtable_net_hooks(struct net *net,
7422                                                struct list_head *hook_list)
7423 {
7424         __nft_unregister_flowtable_net_hooks(net, hook_list, false);
7425 }
7426
7427 static int nft_register_flowtable_net_hooks(struct net *net,
7428                                             struct nft_table *table,
7429                                             struct list_head *hook_list,
7430                                             struct nft_flowtable *flowtable)
7431 {
7432         struct nft_hook *hook, *hook2, *next;
7433         struct nft_flowtable *ft;
7434         int err, i = 0;
7435
7436         list_for_each_entry(hook, hook_list, list) {
7437                 list_for_each_entry(ft, &table->flowtables, list) {
7438                         if (!nft_is_active_next(net, ft))
7439                                 continue;
7440
7441                         list_for_each_entry(hook2, &ft->hook_list, list) {
7442                                 if (hook->ops.dev == hook2->ops.dev &&
7443                                     hook->ops.pf == hook2->ops.pf) {
7444                                         err = -EEXIST;
7445                                         goto err_unregister_net_hooks;
7446                                 }
7447                         }
7448                 }
7449
7450                 err = flowtable->data.type->setup(&flowtable->data,
7451                                                   hook->ops.dev,
7452                                                   FLOW_BLOCK_BIND);
7453                 if (err < 0)
7454                         goto err_unregister_net_hooks;
7455
7456                 err = nf_register_net_hook(net, &hook->ops);
7457                 if (err < 0) {
7458                         flowtable->data.type->setup(&flowtable->data,
7459                                                     hook->ops.dev,
7460                                                     FLOW_BLOCK_UNBIND);
7461                         goto err_unregister_net_hooks;
7462                 }
7463
7464                 i++;
7465         }
7466
7467         return 0;
7468
7469 err_unregister_net_hooks:
7470         list_for_each_entry_safe(hook, next, hook_list, list) {
7471                 if (i-- <= 0)
7472                         break;
7473
7474                 nft_unregister_flowtable_hook(net, flowtable, hook);
7475                 list_del_rcu(&hook->list);
7476                 kfree_rcu(hook, rcu);
7477         }
7478
7479         return err;
7480 }
7481
7482 static void nft_flowtable_hooks_destroy(struct list_head *hook_list)
7483 {
7484         struct nft_hook *hook, *next;
7485
7486         list_for_each_entry_safe(hook, next, hook_list, list) {
7487                 list_del_rcu(&hook->list);
7488                 kfree_rcu(hook, rcu);
7489         }
7490 }
7491
7492 static int nft_flowtable_update(struct nft_ctx *ctx, const struct nlmsghdr *nlh,
7493                                 struct nft_flowtable *flowtable)
7494 {
7495         const struct nlattr * const *nla = ctx->nla;
7496         struct nft_flowtable_hook flowtable_hook;
7497         struct nft_hook *hook, *next;
7498         struct nft_trans *trans;
7499         bool unregister = false;
7500         u32 flags;
7501         int err;
7502
7503         err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7504                                        &flowtable_hook, flowtable, false);
7505         if (err < 0)
7506                 return err;
7507
7508         list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7509                 if (nft_hook_list_find(&flowtable->hook_list, hook)) {
7510                         list_del(&hook->list);
7511                         kfree(hook);
7512                 }
7513         }
7514
7515         if (nla[NFTA_FLOWTABLE_FLAGS]) {
7516                 flags = ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7517                 if (flags & ~NFT_FLOWTABLE_MASK) {
7518                         err = -EOPNOTSUPP;
7519                         goto err_flowtable_update_hook;
7520                 }
7521                 if ((flowtable->data.flags & NFT_FLOWTABLE_HW_OFFLOAD) ^
7522                     (flags & NFT_FLOWTABLE_HW_OFFLOAD)) {
7523                         err = -EOPNOTSUPP;
7524                         goto err_flowtable_update_hook;
7525                 }
7526         } else {
7527                 flags = flowtable->data.flags;
7528         }
7529
7530         err = nft_register_flowtable_net_hooks(ctx->net, ctx->table,
7531                                                &flowtable_hook.list, flowtable);
7532         if (err < 0)
7533                 goto err_flowtable_update_hook;
7534
7535         trans = nft_trans_alloc(ctx, NFT_MSG_NEWFLOWTABLE,
7536                                 sizeof(struct nft_trans_flowtable));
7537         if (!trans) {
7538                 unregister = true;
7539                 err = -ENOMEM;
7540                 goto err_flowtable_update_hook;
7541         }
7542
7543         nft_trans_flowtable_flags(trans) = flags;
7544         nft_trans_flowtable(trans) = flowtable;
7545         nft_trans_flowtable_update(trans) = true;
7546         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7547         list_splice(&flowtable_hook.list, &nft_trans_flowtable_hooks(trans));
7548
7549         nft_trans_commit_list_add_tail(ctx->net, trans);
7550
7551         return 0;
7552
7553 err_flowtable_update_hook:
7554         list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7555                 if (unregister)
7556                         nft_unregister_flowtable_hook(ctx->net, flowtable, hook);
7557                 list_del_rcu(&hook->list);
7558                 kfree_rcu(hook, rcu);
7559         }
7560
7561         return err;
7562
7563 }
7564
7565 static int nf_tables_newflowtable(struct sk_buff *skb,
7566                                   const struct nfnl_info *info,
7567                                   const struct nlattr * const nla[])
7568 {
7569         struct netlink_ext_ack *extack = info->extack;
7570         struct nft_flowtable_hook flowtable_hook;
7571         u8 genmask = nft_genmask_next(info->net);
7572         u8 family = info->nfmsg->nfgen_family;
7573         const struct nf_flowtable_type *type;
7574         struct nft_flowtable *flowtable;
7575         struct nft_hook *hook, *next;
7576         struct net *net = info->net;
7577         struct nft_table *table;
7578         struct nft_ctx ctx;
7579         int err;
7580
7581         if (!nla[NFTA_FLOWTABLE_TABLE] ||
7582             !nla[NFTA_FLOWTABLE_NAME] ||
7583             !nla[NFTA_FLOWTABLE_HOOK])
7584                 return -EINVAL;
7585
7586         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7587                                  genmask, NETLINK_CB(skb).portid);
7588         if (IS_ERR(table)) {
7589                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7590                 return PTR_ERR(table);
7591         }
7592
7593         flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7594                                          genmask);
7595         if (IS_ERR(flowtable)) {
7596                 err = PTR_ERR(flowtable);
7597                 if (err != -ENOENT) {
7598                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7599                         return err;
7600                 }
7601         } else {
7602                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
7603                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7604                         return -EEXIST;
7605                 }
7606
7607                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7608
7609                 return nft_flowtable_update(&ctx, info->nlh, flowtable);
7610         }
7611
7612         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7613
7614         flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL_ACCOUNT);
7615         if (!flowtable)
7616                 return -ENOMEM;
7617
7618         flowtable->table = table;
7619         flowtable->handle = nf_tables_alloc_handle(table);
7620         INIT_LIST_HEAD(&flowtable->hook_list);
7621
7622         flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL_ACCOUNT);
7623         if (!flowtable->name) {
7624                 err = -ENOMEM;
7625                 goto err1;
7626         }
7627
7628         type = nft_flowtable_type_get(net, family);
7629         if (IS_ERR(type)) {
7630                 err = PTR_ERR(type);
7631                 goto err2;
7632         }
7633
7634         if (nla[NFTA_FLOWTABLE_FLAGS]) {
7635                 flowtable->data.flags =
7636                         ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7637                 if (flowtable->data.flags & ~NFT_FLOWTABLE_MASK) {
7638                         err = -EOPNOTSUPP;
7639                         goto err3;
7640                 }
7641         }
7642
7643         write_pnet(&flowtable->data.net, net);
7644         flowtable->data.type = type;
7645         err = type->init(&flowtable->data);
7646         if (err < 0)
7647                 goto err3;
7648
7649         err = nft_flowtable_parse_hook(&ctx, nla[NFTA_FLOWTABLE_HOOK],
7650                                        &flowtable_hook, flowtable, true);
7651         if (err < 0)
7652                 goto err4;
7653
7654         list_splice(&flowtable_hook.list, &flowtable->hook_list);
7655         flowtable->data.priority = flowtable_hook.priority;
7656         flowtable->hooknum = flowtable_hook.num;
7657
7658         err = nft_register_flowtable_net_hooks(ctx.net, table,
7659                                                &flowtable->hook_list,
7660                                                flowtable);
7661         if (err < 0) {
7662                 nft_flowtable_hooks_destroy(&flowtable->hook_list);
7663                 goto err4;
7664         }
7665
7666         err = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
7667         if (err < 0)
7668                 goto err5;
7669
7670         list_add_tail_rcu(&flowtable->list, &table->flowtables);
7671         table->use++;
7672
7673         return 0;
7674 err5:
7675         list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7676                 nft_unregister_flowtable_hook(net, flowtable, hook);
7677                 list_del_rcu(&hook->list);
7678                 kfree_rcu(hook, rcu);
7679         }
7680 err4:
7681         flowtable->data.type->free(&flowtable->data);
7682 err3:
7683         module_put(type->owner);
7684 err2:
7685         kfree(flowtable->name);
7686 err1:
7687         kfree(flowtable);
7688         return err;
7689 }
7690
7691 static void nft_flowtable_hook_release(struct nft_flowtable_hook *flowtable_hook)
7692 {
7693         struct nft_hook *this, *next;
7694
7695         list_for_each_entry_safe(this, next, &flowtable_hook->list, list) {
7696                 list_del(&this->list);
7697                 kfree(this);
7698         }
7699 }
7700
7701 static int nft_delflowtable_hook(struct nft_ctx *ctx,
7702                                  struct nft_flowtable *flowtable)
7703 {
7704         const struct nlattr * const *nla = ctx->nla;
7705         struct nft_flowtable_hook flowtable_hook;
7706         LIST_HEAD(flowtable_del_list);
7707         struct nft_hook *this, *hook;
7708         struct nft_trans *trans;
7709         int err;
7710
7711         err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7712                                        &flowtable_hook, flowtable, false);
7713         if (err < 0)
7714                 return err;
7715
7716         list_for_each_entry(this, &flowtable_hook.list, list) {
7717                 hook = nft_hook_list_find(&flowtable->hook_list, this);
7718                 if (!hook) {
7719                         err = -ENOENT;
7720                         goto err_flowtable_del_hook;
7721                 }
7722                 list_move(&hook->list, &flowtable_del_list);
7723         }
7724
7725         trans = nft_trans_alloc(ctx, NFT_MSG_DELFLOWTABLE,
7726                                 sizeof(struct nft_trans_flowtable));
7727         if (!trans) {
7728                 err = -ENOMEM;
7729                 goto err_flowtable_del_hook;
7730         }
7731
7732         nft_trans_flowtable(trans) = flowtable;
7733         nft_trans_flowtable_update(trans) = true;
7734         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7735         list_splice(&flowtable_del_list, &nft_trans_flowtable_hooks(trans));
7736         nft_flowtable_hook_release(&flowtable_hook);
7737
7738         nft_trans_commit_list_add_tail(ctx->net, trans);
7739
7740         return 0;
7741
7742 err_flowtable_del_hook:
7743         list_splice(&flowtable_del_list, &flowtable->hook_list);
7744         nft_flowtable_hook_release(&flowtable_hook);
7745
7746         return err;
7747 }
7748
7749 static int nf_tables_delflowtable(struct sk_buff *skb,
7750                                   const struct nfnl_info *info,
7751                                   const struct nlattr * const nla[])
7752 {
7753         struct netlink_ext_ack *extack = info->extack;
7754         u8 genmask = nft_genmask_next(info->net);
7755         u8 family = info->nfmsg->nfgen_family;
7756         struct nft_flowtable *flowtable;
7757         struct net *net = info->net;
7758         const struct nlattr *attr;
7759         struct nft_table *table;
7760         struct nft_ctx ctx;
7761
7762         if (!nla[NFTA_FLOWTABLE_TABLE] ||
7763             (!nla[NFTA_FLOWTABLE_NAME] &&
7764              !nla[NFTA_FLOWTABLE_HANDLE]))
7765                 return -EINVAL;
7766
7767         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7768                                  genmask, NETLINK_CB(skb).portid);
7769         if (IS_ERR(table)) {
7770                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7771                 return PTR_ERR(table);
7772         }
7773
7774         if (nla[NFTA_FLOWTABLE_HANDLE]) {
7775                 attr = nla[NFTA_FLOWTABLE_HANDLE];
7776                 flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
7777         } else {
7778                 attr = nla[NFTA_FLOWTABLE_NAME];
7779                 flowtable = nft_flowtable_lookup(table, attr, genmask);
7780         }
7781
7782         if (IS_ERR(flowtable)) {
7783                 NL_SET_BAD_ATTR(extack, attr);
7784                 return PTR_ERR(flowtable);
7785         }
7786
7787         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7788
7789         if (nla[NFTA_FLOWTABLE_HOOK])
7790                 return nft_delflowtable_hook(&ctx, flowtable);
7791
7792         if (flowtable->use > 0) {
7793                 NL_SET_BAD_ATTR(extack, attr);
7794                 return -EBUSY;
7795         }
7796
7797         return nft_delflowtable(&ctx, flowtable);
7798 }
7799
7800 static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
7801                                          u32 portid, u32 seq, int event,
7802                                          u32 flags, int family,
7803                                          struct nft_flowtable *flowtable,
7804                                          struct list_head *hook_list)
7805 {
7806         struct nlattr *nest, *nest_devs;
7807         struct nft_hook *hook;
7808         struct nlmsghdr *nlh;
7809
7810         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
7811         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
7812                            NFNETLINK_V0, nft_base_seq(net));
7813         if (!nlh)
7814                 goto nla_put_failure;
7815
7816         if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
7817             nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
7818             nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
7819             nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
7820                          NFTA_FLOWTABLE_PAD) ||
7821             nla_put_be32(skb, NFTA_FLOWTABLE_FLAGS, htonl(flowtable->data.flags)))
7822                 goto nla_put_failure;
7823
7824         nest = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK);
7825         if (!nest)
7826                 goto nla_put_failure;
7827         if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
7828             nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->data.priority)))
7829                 goto nla_put_failure;
7830
7831         nest_devs = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK_DEVS);
7832         if (!nest_devs)
7833                 goto nla_put_failure;
7834
7835         list_for_each_entry_rcu(hook, hook_list, list) {
7836                 if (nla_put_string(skb, NFTA_DEVICE_NAME, hook->ops.dev->name))
7837                         goto nla_put_failure;
7838         }
7839         nla_nest_end(skb, nest_devs);
7840         nla_nest_end(skb, nest);
7841
7842         nlmsg_end(skb, nlh);
7843         return 0;
7844
7845 nla_put_failure:
7846         nlmsg_trim(skb, nlh);
7847         return -1;
7848 }
7849
7850 struct nft_flowtable_filter {
7851         char            *table;
7852 };
7853
7854 static int nf_tables_dump_flowtable(struct sk_buff *skb,
7855                                     struct netlink_callback *cb)
7856 {
7857         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
7858         struct nft_flowtable_filter *filter = cb->data;
7859         unsigned int idx = 0, s_idx = cb->args[0];
7860         struct net *net = sock_net(skb->sk);
7861         int family = nfmsg->nfgen_family;
7862         struct nft_flowtable *flowtable;
7863         struct nftables_pernet *nft_net;
7864         const struct nft_table *table;
7865
7866         rcu_read_lock();
7867         nft_net = nft_pernet(net);
7868         cb->seq = READ_ONCE(nft_net->base_seq);
7869
7870         list_for_each_entry_rcu(table, &nft_net->tables, list) {
7871                 if (family != NFPROTO_UNSPEC && family != table->family)
7872                         continue;
7873
7874                 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7875                         if (!nft_is_active(net, flowtable))
7876                                 goto cont;
7877                         if (idx < s_idx)
7878                                 goto cont;
7879                         if (idx > s_idx)
7880                                 memset(&cb->args[1], 0,
7881                                        sizeof(cb->args) - sizeof(cb->args[0]));
7882                         if (filter && filter->table &&
7883                             strcmp(filter->table, table->name))
7884                                 goto cont;
7885
7886                         if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
7887                                                           cb->nlh->nlmsg_seq,
7888                                                           NFT_MSG_NEWFLOWTABLE,
7889                                                           NLM_F_MULTI | NLM_F_APPEND,
7890                                                           table->family,
7891                                                           flowtable,
7892                                                           &flowtable->hook_list) < 0)
7893                                 goto done;
7894
7895                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
7896 cont:
7897                         idx++;
7898                 }
7899         }
7900 done:
7901         rcu_read_unlock();
7902
7903         cb->args[0] = idx;
7904         return skb->len;
7905 }
7906
7907 static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
7908 {
7909         const struct nlattr * const *nla = cb->data;
7910         struct nft_flowtable_filter *filter = NULL;
7911
7912         if (nla[NFTA_FLOWTABLE_TABLE]) {
7913                 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
7914                 if (!filter)
7915                         return -ENOMEM;
7916
7917                 filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
7918                                            GFP_ATOMIC);
7919                 if (!filter->table) {
7920                         kfree(filter);
7921                         return -ENOMEM;
7922                 }
7923         }
7924
7925         cb->data = filter;
7926         return 0;
7927 }
7928
7929 static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
7930 {
7931         struct nft_flowtable_filter *filter = cb->data;
7932
7933         if (!filter)
7934                 return 0;
7935
7936         kfree(filter->table);
7937         kfree(filter);
7938
7939         return 0;
7940 }
7941
7942 /* called with rcu_read_lock held */
7943 static int nf_tables_getflowtable(struct sk_buff *skb,
7944                                   const struct nfnl_info *info,
7945                                   const struct nlattr * const nla[])
7946 {
7947         u8 genmask = nft_genmask_cur(info->net);
7948         u8 family = info->nfmsg->nfgen_family;
7949         struct nft_flowtable *flowtable;
7950         const struct nft_table *table;
7951         struct net *net = info->net;
7952         struct sk_buff *skb2;
7953         int err;
7954
7955         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
7956                 struct netlink_dump_control c = {
7957                         .start = nf_tables_dump_flowtable_start,
7958                         .dump = nf_tables_dump_flowtable,
7959                         .done = nf_tables_dump_flowtable_done,
7960                         .module = THIS_MODULE,
7961                         .data = (void *)nla,
7962                 };
7963
7964                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
7965         }
7966
7967         if (!nla[NFTA_FLOWTABLE_NAME])
7968                 return -EINVAL;
7969
7970         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7971                                  genmask, 0);
7972         if (IS_ERR(table))
7973                 return PTR_ERR(table);
7974
7975         flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7976                                          genmask);
7977         if (IS_ERR(flowtable))
7978                 return PTR_ERR(flowtable);
7979
7980         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7981         if (!skb2)
7982                 return -ENOMEM;
7983
7984         err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
7985                                             info->nlh->nlmsg_seq,
7986                                             NFT_MSG_NEWFLOWTABLE, 0, family,
7987                                             flowtable, &flowtable->hook_list);
7988         if (err < 0)
7989                 goto err_fill_flowtable_info;
7990
7991         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7992
7993 err_fill_flowtable_info:
7994         kfree_skb(skb2);
7995         return err;
7996 }
7997
7998 static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
7999                                        struct nft_flowtable *flowtable,
8000                                        struct list_head *hook_list,
8001                                        int event)
8002 {
8003         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
8004         struct sk_buff *skb;
8005         u16 flags = 0;
8006         int err;
8007
8008         if (!ctx->report &&
8009             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
8010                 return;
8011
8012         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
8013         if (skb == NULL)
8014                 goto err;
8015
8016         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
8017                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
8018
8019         err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
8020                                             ctx->seq, event, flags,
8021                                             ctx->family, flowtable, hook_list);
8022         if (err < 0) {
8023                 kfree_skb(skb);
8024                 goto err;
8025         }
8026
8027         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
8028         return;
8029 err:
8030         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
8031 }
8032
8033 static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
8034 {
8035         struct nft_hook *hook, *next;
8036
8037         flowtable->data.type->free(&flowtable->data);
8038         list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
8039                 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
8040                                             FLOW_BLOCK_UNBIND);
8041                 list_del_rcu(&hook->list);
8042                 kfree(hook);
8043         }
8044         kfree(flowtable->name);
8045         module_put(flowtable->data.type->owner);
8046         kfree(flowtable);
8047 }
8048
8049 static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
8050                                    u32 portid, u32 seq)
8051 {
8052         struct nftables_pernet *nft_net = nft_pernet(net);
8053         struct nlmsghdr *nlh;
8054         char buf[TASK_COMM_LEN];
8055         int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
8056
8057         nlh = nfnl_msg_put(skb, portid, seq, event, 0, AF_UNSPEC,
8058                            NFNETLINK_V0, nft_base_seq(net));
8059         if (!nlh)
8060                 goto nla_put_failure;
8061
8062         if (nla_put_be32(skb, NFTA_GEN_ID, htonl(nft_net->base_seq)) ||
8063             nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
8064             nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
8065                 goto nla_put_failure;
8066
8067         nlmsg_end(skb, nlh);
8068         return 0;
8069
8070 nla_put_failure:
8071         nlmsg_trim(skb, nlh);
8072         return -EMSGSIZE;
8073 }
8074
8075 static void nft_flowtable_event(unsigned long event, struct net_device *dev,
8076                                 struct nft_flowtable *flowtable)
8077 {
8078         struct nft_hook *hook;
8079
8080         list_for_each_entry(hook, &flowtable->hook_list, list) {
8081                 if (hook->ops.dev != dev)
8082                         continue;
8083
8084                 /* flow_offload_netdev_event() cleans up entries for us. */
8085                 nft_unregister_flowtable_hook(dev_net(dev), flowtable, hook);
8086                 list_del_rcu(&hook->list);
8087                 kfree_rcu(hook, rcu);
8088                 break;
8089         }
8090 }
8091
8092 static int nf_tables_flowtable_event(struct notifier_block *this,
8093                                      unsigned long event, void *ptr)
8094 {
8095         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
8096         struct nft_flowtable *flowtable;
8097         struct nftables_pernet *nft_net;
8098         struct nft_table *table;
8099         struct net *net;
8100
8101         if (event != NETDEV_UNREGISTER)
8102                 return 0;
8103
8104         net = dev_net(dev);
8105         nft_net = nft_pernet(net);
8106         mutex_lock(&nft_net->commit_mutex);
8107         list_for_each_entry(table, &nft_net->tables, list) {
8108                 list_for_each_entry(flowtable, &table->flowtables, list) {
8109                         nft_flowtable_event(event, dev, flowtable);
8110                 }
8111         }
8112         mutex_unlock(&nft_net->commit_mutex);
8113
8114         return NOTIFY_DONE;
8115 }
8116
8117 static struct notifier_block nf_tables_flowtable_notifier = {
8118         .notifier_call  = nf_tables_flowtable_event,
8119 };
8120
8121 static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
8122                                  int event)
8123 {
8124         struct nlmsghdr *nlh = nlmsg_hdr(skb);
8125         struct sk_buff *skb2;
8126         int err;
8127
8128         if (!nlmsg_report(nlh) &&
8129             !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
8130                 return;
8131
8132         skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
8133         if (skb2 == NULL)
8134                 goto err;
8135
8136         err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
8137                                       nlh->nlmsg_seq);
8138         if (err < 0) {
8139                 kfree_skb(skb2);
8140                 goto err;
8141         }
8142
8143         nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
8144                        nlmsg_report(nlh), GFP_KERNEL);
8145         return;
8146 err:
8147         nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
8148                           -ENOBUFS);
8149 }
8150
8151 static int nf_tables_getgen(struct sk_buff *skb, const struct nfnl_info *info,
8152                             const struct nlattr * const nla[])
8153 {
8154         struct sk_buff *skb2;
8155         int err;
8156
8157         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
8158         if (skb2 == NULL)
8159                 return -ENOMEM;
8160
8161         err = nf_tables_fill_gen_info(skb2, info->net, NETLINK_CB(skb).portid,
8162                                       info->nlh->nlmsg_seq);
8163         if (err < 0)
8164                 goto err_fill_gen_info;
8165
8166         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
8167
8168 err_fill_gen_info:
8169         kfree_skb(skb2);
8170         return err;
8171 }
8172
8173 static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
8174         [NFT_MSG_NEWTABLE] = {
8175                 .call           = nf_tables_newtable,
8176                 .type           = NFNL_CB_BATCH,
8177                 .attr_count     = NFTA_TABLE_MAX,
8178                 .policy         = nft_table_policy,
8179         },
8180         [NFT_MSG_GETTABLE] = {
8181                 .call           = nf_tables_gettable,
8182                 .type           = NFNL_CB_RCU,
8183                 .attr_count     = NFTA_TABLE_MAX,
8184                 .policy         = nft_table_policy,
8185         },
8186         [NFT_MSG_DELTABLE] = {
8187                 .call           = nf_tables_deltable,
8188                 .type           = NFNL_CB_BATCH,
8189                 .attr_count     = NFTA_TABLE_MAX,
8190                 .policy         = nft_table_policy,
8191         },
8192         [NFT_MSG_NEWCHAIN] = {
8193                 .call           = nf_tables_newchain,
8194                 .type           = NFNL_CB_BATCH,
8195                 .attr_count     = NFTA_CHAIN_MAX,
8196                 .policy         = nft_chain_policy,
8197         },
8198         [NFT_MSG_GETCHAIN] = {
8199                 .call           = nf_tables_getchain,
8200                 .type           = NFNL_CB_RCU,
8201                 .attr_count     = NFTA_CHAIN_MAX,
8202                 .policy         = nft_chain_policy,
8203         },
8204         [NFT_MSG_DELCHAIN] = {
8205                 .call           = nf_tables_delchain,
8206                 .type           = NFNL_CB_BATCH,
8207                 .attr_count     = NFTA_CHAIN_MAX,
8208                 .policy         = nft_chain_policy,
8209         },
8210         [NFT_MSG_NEWRULE] = {
8211                 .call           = nf_tables_newrule,
8212                 .type           = NFNL_CB_BATCH,
8213                 .attr_count     = NFTA_RULE_MAX,
8214                 .policy         = nft_rule_policy,
8215         },
8216         [NFT_MSG_GETRULE] = {
8217                 .call           = nf_tables_getrule,
8218                 .type           = NFNL_CB_RCU,
8219                 .attr_count     = NFTA_RULE_MAX,
8220                 .policy         = nft_rule_policy,
8221         },
8222         [NFT_MSG_DELRULE] = {
8223                 .call           = nf_tables_delrule,
8224                 .type           = NFNL_CB_BATCH,
8225                 .attr_count     = NFTA_RULE_MAX,
8226                 .policy         = nft_rule_policy,
8227         },
8228         [NFT_MSG_NEWSET] = {
8229                 .call           = nf_tables_newset,
8230                 .type           = NFNL_CB_BATCH,
8231                 .attr_count     = NFTA_SET_MAX,
8232                 .policy         = nft_set_policy,
8233         },
8234         [NFT_MSG_GETSET] = {
8235                 .call           = nf_tables_getset,
8236                 .type           = NFNL_CB_RCU,
8237                 .attr_count     = NFTA_SET_MAX,
8238                 .policy         = nft_set_policy,
8239         },
8240         [NFT_MSG_DELSET] = {
8241                 .call           = nf_tables_delset,
8242                 .type           = NFNL_CB_BATCH,
8243                 .attr_count     = NFTA_SET_MAX,
8244                 .policy         = nft_set_policy,
8245         },
8246         [NFT_MSG_NEWSETELEM] = {
8247                 .call           = nf_tables_newsetelem,
8248                 .type           = NFNL_CB_BATCH,
8249                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8250                 .policy         = nft_set_elem_list_policy,
8251         },
8252         [NFT_MSG_GETSETELEM] = {
8253                 .call           = nf_tables_getsetelem,
8254                 .type           = NFNL_CB_RCU,
8255                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8256                 .policy         = nft_set_elem_list_policy,
8257         },
8258         [NFT_MSG_DELSETELEM] = {
8259                 .call           = nf_tables_delsetelem,
8260                 .type           = NFNL_CB_BATCH,
8261                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8262                 .policy         = nft_set_elem_list_policy,
8263         },
8264         [NFT_MSG_GETGEN] = {
8265                 .call           = nf_tables_getgen,
8266                 .type           = NFNL_CB_RCU,
8267         },
8268         [NFT_MSG_NEWOBJ] = {
8269                 .call           = nf_tables_newobj,
8270                 .type           = NFNL_CB_BATCH,
8271                 .attr_count     = NFTA_OBJ_MAX,
8272                 .policy         = nft_obj_policy,
8273         },
8274         [NFT_MSG_GETOBJ] = {
8275                 .call           = nf_tables_getobj,
8276                 .type           = NFNL_CB_RCU,
8277                 .attr_count     = NFTA_OBJ_MAX,
8278                 .policy         = nft_obj_policy,
8279         },
8280         [NFT_MSG_DELOBJ] = {
8281                 .call           = nf_tables_delobj,
8282                 .type           = NFNL_CB_BATCH,
8283                 .attr_count     = NFTA_OBJ_MAX,
8284                 .policy         = nft_obj_policy,
8285         },
8286         [NFT_MSG_GETOBJ_RESET] = {
8287                 .call           = nf_tables_getobj,
8288                 .type           = NFNL_CB_RCU,
8289                 .attr_count     = NFTA_OBJ_MAX,
8290                 .policy         = nft_obj_policy,
8291         },
8292         [NFT_MSG_NEWFLOWTABLE] = {
8293                 .call           = nf_tables_newflowtable,
8294                 .type           = NFNL_CB_BATCH,
8295                 .attr_count     = NFTA_FLOWTABLE_MAX,
8296                 .policy         = nft_flowtable_policy,
8297         },
8298         [NFT_MSG_GETFLOWTABLE] = {
8299                 .call           = nf_tables_getflowtable,
8300                 .type           = NFNL_CB_RCU,
8301                 .attr_count     = NFTA_FLOWTABLE_MAX,
8302                 .policy         = nft_flowtable_policy,
8303         },
8304         [NFT_MSG_DELFLOWTABLE] = {
8305                 .call           = nf_tables_delflowtable,
8306                 .type           = NFNL_CB_BATCH,
8307                 .attr_count     = NFTA_FLOWTABLE_MAX,
8308                 .policy         = nft_flowtable_policy,
8309         },
8310 };
8311
8312 static int nf_tables_validate(struct net *net)
8313 {
8314         struct nftables_pernet *nft_net = nft_pernet(net);
8315         struct nft_table *table;
8316
8317         switch (nft_net->validate_state) {
8318         case NFT_VALIDATE_SKIP:
8319                 break;
8320         case NFT_VALIDATE_NEED:
8321                 nft_validate_state_update(net, NFT_VALIDATE_DO);
8322                 fallthrough;
8323         case NFT_VALIDATE_DO:
8324                 list_for_each_entry(table, &nft_net->tables, list) {
8325                         if (nft_table_validate(net, table) < 0)
8326                                 return -EAGAIN;
8327                 }
8328                 break;
8329         }
8330
8331         return 0;
8332 }
8333
8334 /* a drop policy has to be deferred until all rules have been activated,
8335  * otherwise a large ruleset that contains a drop-policy base chain will
8336  * cause all packets to get dropped until the full transaction has been
8337  * processed.
8338  *
8339  * We defer the drop policy until the transaction has been finalized.
8340  */
8341 static void nft_chain_commit_drop_policy(struct nft_trans *trans)
8342 {
8343         struct nft_base_chain *basechain;
8344
8345         if (nft_trans_chain_policy(trans) != NF_DROP)
8346                 return;
8347
8348         if (!nft_is_base_chain(trans->ctx.chain))
8349                 return;
8350
8351         basechain = nft_base_chain(trans->ctx.chain);
8352         basechain->policy = NF_DROP;
8353 }
8354
8355 static void nft_chain_commit_update(struct nft_trans *trans)
8356 {
8357         struct nft_base_chain *basechain;
8358
8359         if (nft_trans_chain_name(trans)) {
8360                 rhltable_remove(&trans->ctx.table->chains_ht,
8361                                 &trans->ctx.chain->rhlhead,
8362                                 nft_chain_ht_params);
8363                 swap(trans->ctx.chain->name, nft_trans_chain_name(trans));
8364                 rhltable_insert_key(&trans->ctx.table->chains_ht,
8365                                     trans->ctx.chain->name,
8366                                     &trans->ctx.chain->rhlhead,
8367                                     nft_chain_ht_params);
8368         }
8369
8370         if (!nft_is_base_chain(trans->ctx.chain))
8371                 return;
8372
8373         nft_chain_stats_replace(trans);
8374
8375         basechain = nft_base_chain(trans->ctx.chain);
8376
8377         switch (nft_trans_chain_policy(trans)) {
8378         case NF_DROP:
8379         case NF_ACCEPT:
8380                 basechain->policy = nft_trans_chain_policy(trans);
8381                 break;
8382         }
8383 }
8384
8385 static void nft_obj_commit_update(struct nft_trans *trans)
8386 {
8387         struct nft_object *newobj;
8388         struct nft_object *obj;
8389
8390         obj = nft_trans_obj(trans);
8391         newobj = nft_trans_obj_newobj(trans);
8392
8393         if (obj->ops->update)
8394                 obj->ops->update(obj, newobj);
8395
8396         nft_obj_destroy(&trans->ctx, newobj);
8397 }
8398
8399 static void nft_commit_release(struct nft_trans *trans)
8400 {
8401         switch (trans->msg_type) {
8402         case NFT_MSG_DELTABLE:
8403                 nf_tables_table_destroy(&trans->ctx);
8404                 break;
8405         case NFT_MSG_NEWCHAIN:
8406                 free_percpu(nft_trans_chain_stats(trans));
8407                 kfree(nft_trans_chain_name(trans));
8408                 break;
8409         case NFT_MSG_DELCHAIN:
8410                 nf_tables_chain_destroy(&trans->ctx);
8411                 break;
8412         case NFT_MSG_DELRULE:
8413                 if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
8414                         nft_flow_rule_destroy(nft_trans_flow_rule(trans));
8415
8416                 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8417                 break;
8418         case NFT_MSG_DELSET:
8419                 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8420                 break;
8421         case NFT_MSG_DELSETELEM:
8422                 nf_tables_set_elem_destroy(&trans->ctx,
8423                                            nft_trans_elem_set(trans),
8424                                            nft_trans_elem(trans).priv);
8425                 break;
8426         case NFT_MSG_DELOBJ:
8427                 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8428                 break;
8429         case NFT_MSG_DELFLOWTABLE:
8430                 if (nft_trans_flowtable_update(trans))
8431                         nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8432                 else
8433                         nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8434                 break;
8435         }
8436
8437         if (trans->put_net)
8438                 put_net(trans->ctx.net);
8439
8440         kfree(trans);
8441 }
8442
8443 static void nf_tables_trans_destroy_work(struct work_struct *w)
8444 {
8445         struct nft_trans *trans, *next;
8446         LIST_HEAD(head);
8447
8448         spin_lock(&nf_tables_destroy_list_lock);
8449         list_splice_init(&nf_tables_destroy_list, &head);
8450         spin_unlock(&nf_tables_destroy_list_lock);
8451
8452         if (list_empty(&head))
8453                 return;
8454
8455         synchronize_rcu();
8456
8457         list_for_each_entry_safe(trans, next, &head, list) {
8458                 list_del(&trans->list);
8459                 nft_commit_release(trans);
8460         }
8461 }
8462
8463 void nf_tables_trans_destroy_flush_work(void)
8464 {
8465         flush_work(&trans_destroy_work);
8466 }
8467 EXPORT_SYMBOL_GPL(nf_tables_trans_destroy_flush_work);
8468
8469 static bool nft_expr_reduce(struct nft_regs_track *track,
8470                             const struct nft_expr *expr)
8471 {
8472         return false;
8473 }
8474
8475 static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
8476 {
8477         const struct nft_expr *expr, *last;
8478         struct nft_regs_track track = {};
8479         unsigned int size, data_size;
8480         void *data, *data_boundary;
8481         struct nft_rule_dp *prule;
8482         struct nft_rule *rule;
8483
8484         /* already handled or inactive chain? */
8485         if (chain->blob_next || !nft_is_active_next(net, chain))
8486                 return 0;
8487
8488         data_size = 0;
8489         list_for_each_entry(rule, &chain->rules, list) {
8490                 if (nft_is_active_next(net, rule)) {
8491                         data_size += sizeof(*prule) + rule->dlen;
8492                         if (data_size > INT_MAX)
8493                                 return -ENOMEM;
8494                 }
8495         }
8496         data_size += offsetof(struct nft_rule_dp, data);        /* last rule */
8497
8498         chain->blob_next = nf_tables_chain_alloc_rules(data_size);
8499         if (!chain->blob_next)
8500                 return -ENOMEM;
8501
8502         data = (void *)chain->blob_next->data;
8503         data_boundary = data + data_size;
8504         size = 0;
8505
8506         list_for_each_entry(rule, &chain->rules, list) {
8507                 if (!nft_is_active_next(net, rule))
8508                         continue;
8509
8510                 prule = (struct nft_rule_dp *)data;
8511                 data += offsetof(struct nft_rule_dp, data);
8512                 if (WARN_ON_ONCE(data > data_boundary))
8513                         return -ENOMEM;
8514
8515                 size = 0;
8516                 track.last = nft_expr_last(rule);
8517                 nft_rule_for_each_expr(expr, last, rule) {
8518                         track.cur = expr;
8519
8520                         if (nft_expr_reduce(&track, expr)) {
8521                                 expr = track.cur;
8522                                 continue;
8523                         }
8524
8525                         if (WARN_ON_ONCE(data + expr->ops->size > data_boundary))
8526                                 return -ENOMEM;
8527
8528                         memcpy(data + size, expr, expr->ops->size);
8529                         size += expr->ops->size;
8530                 }
8531                 if (WARN_ON_ONCE(size >= 1 << 12))
8532                         return -ENOMEM;
8533
8534                 prule->handle = rule->handle;
8535                 prule->dlen = size;
8536                 prule->is_last = 0;
8537
8538                 data += size;
8539                 size = 0;
8540                 chain->blob_next->size += (unsigned long)(data - (void *)prule);
8541         }
8542
8543         prule = (struct nft_rule_dp *)data;
8544         data += offsetof(struct nft_rule_dp, data);
8545         if (WARN_ON_ONCE(data > data_boundary))
8546                 return -ENOMEM;
8547
8548         nft_last_rule(chain->blob_next, prule);
8549
8550         return 0;
8551 }
8552
8553 static void nf_tables_commit_chain_prepare_cancel(struct net *net)
8554 {
8555         struct nftables_pernet *nft_net = nft_pernet(net);
8556         struct nft_trans *trans, *next;
8557
8558         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8559                 struct nft_chain *chain = trans->ctx.chain;
8560
8561                 if (trans->msg_type == NFT_MSG_NEWRULE ||
8562                     trans->msg_type == NFT_MSG_DELRULE) {
8563                         kvfree(chain->blob_next);
8564                         chain->blob_next = NULL;
8565                 }
8566         }
8567 }
8568
8569 static void __nf_tables_commit_chain_free_rules_old(struct rcu_head *h)
8570 {
8571         struct nft_rules_old *o = container_of(h, struct nft_rules_old, h);
8572
8573         kvfree(o->blob);
8574 }
8575
8576 static void nf_tables_commit_chain_free_rules_old(struct nft_rule_blob *blob)
8577 {
8578         struct nft_rules_old *old;
8579
8580         /* rcu_head is after end marker */
8581         old = (void *)blob + sizeof(*blob) + blob->size;
8582         old->blob = blob;
8583
8584         call_rcu(&old->h, __nf_tables_commit_chain_free_rules_old);
8585 }
8586
8587 static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
8588 {
8589         struct nft_rule_blob *g0, *g1;
8590         bool next_genbit;
8591
8592         next_genbit = nft_gencursor_next(net);
8593
8594         g0 = rcu_dereference_protected(chain->blob_gen_0,
8595                                        lockdep_commit_lock_is_held(net));
8596         g1 = rcu_dereference_protected(chain->blob_gen_1,
8597                                        lockdep_commit_lock_is_held(net));
8598
8599         /* No changes to this chain? */
8600         if (chain->blob_next == NULL) {
8601                 /* chain had no change in last or next generation */
8602                 if (g0 == g1)
8603                         return;
8604                 /*
8605                  * chain had no change in this generation; make sure next
8606                  * one uses same rules as current generation.
8607                  */
8608                 if (next_genbit) {
8609                         rcu_assign_pointer(chain->blob_gen_1, g0);
8610                         nf_tables_commit_chain_free_rules_old(g1);
8611                 } else {
8612                         rcu_assign_pointer(chain->blob_gen_0, g1);
8613                         nf_tables_commit_chain_free_rules_old(g0);
8614                 }
8615
8616                 return;
8617         }
8618
8619         if (next_genbit)
8620                 rcu_assign_pointer(chain->blob_gen_1, chain->blob_next);
8621         else
8622                 rcu_assign_pointer(chain->blob_gen_0, chain->blob_next);
8623
8624         chain->blob_next = NULL;
8625
8626         if (g0 == g1)
8627                 return;
8628
8629         if (next_genbit)
8630                 nf_tables_commit_chain_free_rules_old(g1);
8631         else
8632                 nf_tables_commit_chain_free_rules_old(g0);
8633 }
8634
8635 static void nft_obj_del(struct nft_object *obj)
8636 {
8637         rhltable_remove(&nft_objname_ht, &obj->rhlhead, nft_objname_ht_params);
8638         list_del_rcu(&obj->list);
8639 }
8640
8641 void nft_chain_del(struct nft_chain *chain)
8642 {
8643         struct nft_table *table = chain->table;
8644
8645         WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
8646                                      nft_chain_ht_params));
8647         list_del_rcu(&chain->list);
8648 }
8649
8650 static void nf_tables_module_autoload_cleanup(struct net *net)
8651 {
8652         struct nftables_pernet *nft_net = nft_pernet(net);
8653         struct nft_module_request *req, *next;
8654
8655         WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
8656         list_for_each_entry_safe(req, next, &nft_net->module_list, list) {
8657                 WARN_ON_ONCE(!req->done);
8658                 list_del(&req->list);
8659                 kfree(req);
8660         }
8661 }
8662
8663 static void nf_tables_commit_release(struct net *net)
8664 {
8665         struct nftables_pernet *nft_net = nft_pernet(net);
8666         struct nft_trans *trans;
8667
8668         /* all side effects have to be made visible.
8669          * For example, if a chain named 'foo' has been deleted, a
8670          * new transaction must not find it anymore.
8671          *
8672          * Memory reclaim happens asynchronously from work queue
8673          * to prevent expensive synchronize_rcu() in commit phase.
8674          */
8675         if (list_empty(&nft_net->commit_list)) {
8676                 nf_tables_module_autoload_cleanup(net);
8677                 mutex_unlock(&nft_net->commit_mutex);
8678                 return;
8679         }
8680
8681         trans = list_last_entry(&nft_net->commit_list,
8682                                 struct nft_trans, list);
8683         get_net(trans->ctx.net);
8684         WARN_ON_ONCE(trans->put_net);
8685
8686         trans->put_net = true;
8687         spin_lock(&nf_tables_destroy_list_lock);
8688         list_splice_tail_init(&nft_net->commit_list, &nf_tables_destroy_list);
8689         spin_unlock(&nf_tables_destroy_list_lock);
8690
8691         nf_tables_module_autoload_cleanup(net);
8692         schedule_work(&trans_destroy_work);
8693
8694         mutex_unlock(&nft_net->commit_mutex);
8695 }
8696
8697 static void nft_commit_notify(struct net *net, u32 portid)
8698 {
8699         struct nftables_pernet *nft_net = nft_pernet(net);
8700         struct sk_buff *batch_skb = NULL, *nskb, *skb;
8701         unsigned char *data;
8702         int len;
8703
8704         list_for_each_entry_safe(skb, nskb, &nft_net->notify_list, list) {
8705                 if (!batch_skb) {
8706 new_batch:
8707                         batch_skb = skb;
8708                         len = NLMSG_GOODSIZE - skb->len;
8709                         list_del(&skb->list);
8710                         continue;
8711                 }
8712                 len -= skb->len;
8713                 if (len > 0 && NFT_CB(skb).report == NFT_CB(batch_skb).report) {
8714                         data = skb_put(batch_skb, skb->len);
8715                         memcpy(data, skb->data, skb->len);
8716                         list_del(&skb->list);
8717                         kfree_skb(skb);
8718                         continue;
8719                 }
8720                 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8721                                NFT_CB(batch_skb).report, GFP_KERNEL);
8722                 goto new_batch;
8723         }
8724
8725         if (batch_skb) {
8726                 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8727                                NFT_CB(batch_skb).report, GFP_KERNEL);
8728         }
8729
8730         WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
8731 }
8732
8733 static int nf_tables_commit_audit_alloc(struct list_head *adl,
8734                                         struct nft_table *table)
8735 {
8736         struct nft_audit_data *adp;
8737
8738         list_for_each_entry(adp, adl, list) {
8739                 if (adp->table == table)
8740                         return 0;
8741         }
8742         adp = kzalloc(sizeof(*adp), GFP_KERNEL);
8743         if (!adp)
8744                 return -ENOMEM;
8745         adp->table = table;
8746         list_add(&adp->list, adl);
8747         return 0;
8748 }
8749
8750 static void nf_tables_commit_audit_free(struct list_head *adl)
8751 {
8752         struct nft_audit_data *adp, *adn;
8753
8754         list_for_each_entry_safe(adp, adn, adl, list) {
8755                 list_del(&adp->list);
8756                 kfree(adp);
8757         }
8758 }
8759
8760 static void nf_tables_commit_audit_collect(struct list_head *adl,
8761                                            struct nft_table *table, u32 op)
8762 {
8763         struct nft_audit_data *adp;
8764
8765         list_for_each_entry(adp, adl, list) {
8766                 if (adp->table == table)
8767                         goto found;
8768         }
8769         WARN_ONCE(1, "table=%s not expected in commit list", table->name);
8770         return;
8771 found:
8772         adp->entries++;
8773         if (!adp->op || adp->op > op)
8774                 adp->op = op;
8775 }
8776
8777 #define AUNFTABLENAMELEN (NFT_TABLE_MAXNAMELEN + 22)
8778
8779 static void nf_tables_commit_audit_log(struct list_head *adl, u32 generation)
8780 {
8781         struct nft_audit_data *adp, *adn;
8782         char aubuf[AUNFTABLENAMELEN];
8783
8784         list_for_each_entry_safe(adp, adn, adl, list) {
8785                 snprintf(aubuf, AUNFTABLENAMELEN, "%s:%u", adp->table->name,
8786                          generation);
8787                 audit_log_nfcfg(aubuf, adp->table->family, adp->entries,
8788                                 nft2audit_op[adp->op], GFP_KERNEL);
8789                 list_del(&adp->list);
8790                 kfree(adp);
8791         }
8792 }
8793
8794 static int nf_tables_commit(struct net *net, struct sk_buff *skb)
8795 {
8796         struct nftables_pernet *nft_net = nft_pernet(net);
8797         struct nft_trans *trans, *next;
8798         struct nft_trans_elem *te;
8799         struct nft_chain *chain;
8800         struct nft_table *table;
8801         unsigned int base_seq;
8802         LIST_HEAD(adl);
8803         int err;
8804
8805         if (list_empty(&nft_net->commit_list)) {
8806                 mutex_unlock(&nft_net->commit_mutex);
8807                 return 0;
8808         }
8809
8810         /* 0. Validate ruleset, otherwise roll back for error reporting. */
8811         if (nf_tables_validate(net) < 0)
8812                 return -EAGAIN;
8813
8814         err = nft_flow_rule_offload_commit(net);
8815         if (err < 0)
8816                 return err;
8817
8818         /* 1.  Allocate space for next generation rules_gen_X[] */
8819         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8820                 int ret;
8821
8822                 ret = nf_tables_commit_audit_alloc(&adl, trans->ctx.table);
8823                 if (ret) {
8824                         nf_tables_commit_chain_prepare_cancel(net);
8825                         nf_tables_commit_audit_free(&adl);
8826                         return ret;
8827                 }
8828                 if (trans->msg_type == NFT_MSG_NEWRULE ||
8829                     trans->msg_type == NFT_MSG_DELRULE) {
8830                         chain = trans->ctx.chain;
8831
8832                         ret = nf_tables_commit_chain_prepare(net, chain);
8833                         if (ret < 0) {
8834                                 nf_tables_commit_chain_prepare_cancel(net);
8835                                 nf_tables_commit_audit_free(&adl);
8836                                 return ret;
8837                         }
8838                 }
8839         }
8840
8841         /* step 2.  Make rules_gen_X visible to packet path */
8842         list_for_each_entry(table, &nft_net->tables, list) {
8843                 list_for_each_entry(chain, &table->chains, list)
8844                         nf_tables_commit_chain(net, chain);
8845         }
8846
8847         /*
8848          * Bump generation counter, invalidate any dump in progress.
8849          * Cannot fail after this point.
8850          */
8851         base_seq = READ_ONCE(nft_net->base_seq);
8852         while (++base_seq == 0)
8853                 ;
8854
8855         WRITE_ONCE(nft_net->base_seq, base_seq);
8856
8857         /* step 3. Start new generation, rules_gen_X now in use. */
8858         net->nft.gencursor = nft_gencursor_next(net);
8859
8860         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8861                 nf_tables_commit_audit_collect(&adl, trans->ctx.table,
8862                                                trans->msg_type);
8863                 switch (trans->msg_type) {
8864                 case NFT_MSG_NEWTABLE:
8865                         if (nft_trans_table_update(trans)) {
8866                                 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
8867                                         nft_trans_destroy(trans);
8868                                         break;
8869                                 }
8870                                 if (trans->ctx.table->flags & NFT_TABLE_F_DORMANT)
8871                                         nf_tables_table_disable(net, trans->ctx.table);
8872
8873                                 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
8874                         } else {
8875                                 nft_clear(net, trans->ctx.table);
8876                         }
8877                         nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
8878                         nft_trans_destroy(trans);
8879                         break;
8880                 case NFT_MSG_DELTABLE:
8881                         list_del_rcu(&trans->ctx.table->list);
8882                         nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
8883                         break;
8884                 case NFT_MSG_NEWCHAIN:
8885                         if (nft_trans_chain_update(trans)) {
8886                                 nft_chain_commit_update(trans);
8887                                 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8888                                 /* trans destroyed after rcu grace period */
8889                         } else {
8890                                 nft_chain_commit_drop_policy(trans);
8891                                 nft_clear(net, trans->ctx.chain);
8892                                 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8893                                 nft_trans_destroy(trans);
8894                         }
8895                         break;
8896                 case NFT_MSG_DELCHAIN:
8897                         nft_chain_del(trans->ctx.chain);
8898                         nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
8899                         nf_tables_unregister_hook(trans->ctx.net,
8900                                                   trans->ctx.table,
8901                                                   trans->ctx.chain);
8902                         break;
8903                 case NFT_MSG_NEWRULE:
8904                         nft_clear(trans->ctx.net, nft_trans_rule(trans));
8905                         nf_tables_rule_notify(&trans->ctx,
8906                                               nft_trans_rule(trans),
8907                                               NFT_MSG_NEWRULE);
8908                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
8909                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
8910
8911                         nft_trans_destroy(trans);
8912                         break;
8913                 case NFT_MSG_DELRULE:
8914                         list_del_rcu(&nft_trans_rule(trans)->list);
8915                         nf_tables_rule_notify(&trans->ctx,
8916                                               nft_trans_rule(trans),
8917                                               NFT_MSG_DELRULE);
8918                         nft_rule_expr_deactivate(&trans->ctx,
8919                                                  nft_trans_rule(trans),
8920                                                  NFT_TRANS_COMMIT);
8921                         break;
8922                 case NFT_MSG_NEWSET:
8923                         nft_clear(net, nft_trans_set(trans));
8924                         /* This avoids hitting -EBUSY when deleting the table
8925                          * from the transaction.
8926                          */
8927                         if (nft_set_is_anonymous(nft_trans_set(trans)) &&
8928                             !list_empty(&nft_trans_set(trans)->bindings))
8929                                 trans->ctx.table->use--;
8930
8931                         nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
8932                                              NFT_MSG_NEWSET, GFP_KERNEL);
8933                         nft_trans_destroy(trans);
8934                         break;
8935                 case NFT_MSG_DELSET:
8936                         list_del_rcu(&nft_trans_set(trans)->list);
8937                         nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
8938                                              NFT_MSG_DELSET, GFP_KERNEL);
8939                         break;
8940                 case NFT_MSG_NEWSETELEM:
8941                         te = (struct nft_trans_elem *)trans->data;
8942
8943                         nft_setelem_activate(net, te->set, &te->elem);
8944                         nf_tables_setelem_notify(&trans->ctx, te->set,
8945                                                  &te->elem,
8946                                                  NFT_MSG_NEWSETELEM);
8947                         nft_trans_destroy(trans);
8948                         break;
8949                 case NFT_MSG_DELSETELEM:
8950                         te = (struct nft_trans_elem *)trans->data;
8951
8952                         nf_tables_setelem_notify(&trans->ctx, te->set,
8953                                                  &te->elem,
8954                                                  NFT_MSG_DELSETELEM);
8955                         nft_setelem_remove(net, te->set, &te->elem);
8956                         if (!nft_setelem_is_catchall(te->set, &te->elem)) {
8957                                 atomic_dec(&te->set->nelems);
8958                                 te->set->ndeact--;
8959                         }
8960                         break;
8961                 case NFT_MSG_NEWOBJ:
8962                         if (nft_trans_obj_update(trans)) {
8963                                 nft_obj_commit_update(trans);
8964                                 nf_tables_obj_notify(&trans->ctx,
8965                                                      nft_trans_obj(trans),
8966                                                      NFT_MSG_NEWOBJ);
8967                         } else {
8968                                 nft_clear(net, nft_trans_obj(trans));
8969                                 nf_tables_obj_notify(&trans->ctx,
8970                                                      nft_trans_obj(trans),
8971                                                      NFT_MSG_NEWOBJ);
8972                                 nft_trans_destroy(trans);
8973                         }
8974                         break;
8975                 case NFT_MSG_DELOBJ:
8976                         nft_obj_del(nft_trans_obj(trans));
8977                         nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
8978                                              NFT_MSG_DELOBJ);
8979                         break;
8980                 case NFT_MSG_NEWFLOWTABLE:
8981                         if (nft_trans_flowtable_update(trans)) {
8982                                 nft_trans_flowtable(trans)->data.flags =
8983                                         nft_trans_flowtable_flags(trans);
8984                                 nf_tables_flowtable_notify(&trans->ctx,
8985                                                            nft_trans_flowtable(trans),
8986                                                            &nft_trans_flowtable_hooks(trans),
8987                                                            NFT_MSG_NEWFLOWTABLE);
8988                                 list_splice(&nft_trans_flowtable_hooks(trans),
8989                                             &nft_trans_flowtable(trans)->hook_list);
8990                         } else {
8991                                 nft_clear(net, nft_trans_flowtable(trans));
8992                                 nf_tables_flowtable_notify(&trans->ctx,
8993                                                            nft_trans_flowtable(trans),
8994                                                            &nft_trans_flowtable(trans)->hook_list,
8995                                                            NFT_MSG_NEWFLOWTABLE);
8996                         }
8997                         nft_trans_destroy(trans);
8998                         break;
8999                 case NFT_MSG_DELFLOWTABLE:
9000                         if (nft_trans_flowtable_update(trans)) {
9001                                 nf_tables_flowtable_notify(&trans->ctx,
9002                                                            nft_trans_flowtable(trans),
9003                                                            &nft_trans_flowtable_hooks(trans),
9004                                                            NFT_MSG_DELFLOWTABLE);
9005                                 nft_unregister_flowtable_net_hooks(net,
9006                                                                    &nft_trans_flowtable_hooks(trans));
9007                         } else {
9008                                 list_del_rcu(&nft_trans_flowtable(trans)->list);
9009                                 nf_tables_flowtable_notify(&trans->ctx,
9010                                                            nft_trans_flowtable(trans),
9011                                                            &nft_trans_flowtable(trans)->hook_list,
9012                                                            NFT_MSG_DELFLOWTABLE);
9013                                 nft_unregister_flowtable_net_hooks(net,
9014                                                 &nft_trans_flowtable(trans)->hook_list);
9015                         }
9016                         break;
9017                 }
9018         }
9019
9020         nft_commit_notify(net, NETLINK_CB(skb).portid);
9021         nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
9022         nf_tables_commit_audit_log(&adl, nft_net->base_seq);
9023         nf_tables_commit_release(net);
9024
9025         return 0;
9026 }
9027
9028 static void nf_tables_module_autoload(struct net *net)
9029 {
9030         struct nftables_pernet *nft_net = nft_pernet(net);
9031         struct nft_module_request *req, *next;
9032         LIST_HEAD(module_list);
9033
9034         list_splice_init(&nft_net->module_list, &module_list);
9035         mutex_unlock(&nft_net->commit_mutex);
9036         list_for_each_entry_safe(req, next, &module_list, list) {
9037                 request_module("%s", req->module);
9038                 req->done = true;
9039         }
9040         mutex_lock(&nft_net->commit_mutex);
9041         list_splice(&module_list, &nft_net->module_list);
9042 }
9043
9044 static void nf_tables_abort_release(struct nft_trans *trans)
9045 {
9046         switch (trans->msg_type) {
9047         case NFT_MSG_NEWTABLE:
9048                 nf_tables_table_destroy(&trans->ctx);
9049                 break;
9050         case NFT_MSG_NEWCHAIN:
9051                 nf_tables_chain_destroy(&trans->ctx);
9052                 break;
9053         case NFT_MSG_NEWRULE:
9054                 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
9055                 break;
9056         case NFT_MSG_NEWSET:
9057                 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
9058                 break;
9059         case NFT_MSG_NEWSETELEM:
9060                 nft_set_elem_destroy(nft_trans_elem_set(trans),
9061                                      nft_trans_elem(trans).priv, true);
9062                 break;
9063         case NFT_MSG_NEWOBJ:
9064                 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
9065                 break;
9066         case NFT_MSG_NEWFLOWTABLE:
9067                 if (nft_trans_flowtable_update(trans))
9068                         nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
9069                 else
9070                         nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
9071                 break;
9072         }
9073         kfree(trans);
9074 }
9075
9076 static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)
9077 {
9078         struct nftables_pernet *nft_net = nft_pernet(net);
9079         struct nft_trans *trans, *next;
9080         struct nft_trans_elem *te;
9081
9082         if (action == NFNL_ABORT_VALIDATE &&
9083             nf_tables_validate(net) < 0)
9084                 return -EAGAIN;
9085
9086         list_for_each_entry_safe_reverse(trans, next, &nft_net->commit_list,
9087                                          list) {
9088                 switch (trans->msg_type) {
9089                 case NFT_MSG_NEWTABLE:
9090                         if (nft_trans_table_update(trans)) {
9091                                 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
9092                                         nft_trans_destroy(trans);
9093                                         break;
9094                                 }
9095                                 if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_DORMANT) {
9096                                         nf_tables_table_disable(net, trans->ctx.table);
9097                                         trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
9098                                 } else if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_AWAKEN) {
9099                                         trans->ctx.table->flags &= ~NFT_TABLE_F_DORMANT;
9100                                 }
9101                                 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
9102                                 nft_trans_destroy(trans);
9103                         } else {
9104                                 list_del_rcu(&trans->ctx.table->list);
9105                         }
9106                         break;
9107                 case NFT_MSG_DELTABLE:
9108                         nft_clear(trans->ctx.net, trans->ctx.table);
9109                         nft_trans_destroy(trans);
9110                         break;
9111                 case NFT_MSG_NEWCHAIN:
9112                         if (nft_trans_chain_update(trans)) {
9113                                 free_percpu(nft_trans_chain_stats(trans));
9114                                 kfree(nft_trans_chain_name(trans));
9115                                 nft_trans_destroy(trans);
9116                         } else {
9117                                 if (nft_chain_is_bound(trans->ctx.chain)) {
9118                                         nft_trans_destroy(trans);
9119                                         break;
9120                                 }
9121                                 trans->ctx.table->use--;
9122                                 nft_chain_del(trans->ctx.chain);
9123                                 nf_tables_unregister_hook(trans->ctx.net,
9124                                                           trans->ctx.table,
9125                                                           trans->ctx.chain);
9126                         }
9127                         break;
9128                 case NFT_MSG_DELCHAIN:
9129                         trans->ctx.table->use++;
9130                         nft_clear(trans->ctx.net, trans->ctx.chain);
9131                         nft_trans_destroy(trans);
9132                         break;
9133                 case NFT_MSG_NEWRULE:
9134                         trans->ctx.chain->use--;
9135                         list_del_rcu(&nft_trans_rule(trans)->list);
9136                         nft_rule_expr_deactivate(&trans->ctx,
9137                                                  nft_trans_rule(trans),
9138                                                  NFT_TRANS_ABORT);
9139                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9140                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9141                         break;
9142                 case NFT_MSG_DELRULE:
9143                         trans->ctx.chain->use++;
9144                         nft_clear(trans->ctx.net, nft_trans_rule(trans));
9145                         nft_rule_expr_activate(&trans->ctx, nft_trans_rule(trans));
9146                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9147                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9148
9149                         nft_trans_destroy(trans);
9150                         break;
9151                 case NFT_MSG_NEWSET:
9152                         trans->ctx.table->use--;
9153                         if (nft_trans_set_bound(trans)) {
9154                                 nft_trans_destroy(trans);
9155                                 break;
9156                         }
9157                         list_del_rcu(&nft_trans_set(trans)->list);
9158                         break;
9159                 case NFT_MSG_DELSET:
9160                         trans->ctx.table->use++;
9161                         nft_clear(trans->ctx.net, nft_trans_set(trans));
9162                         nft_trans_destroy(trans);
9163                         break;
9164                 case NFT_MSG_NEWSETELEM:
9165                         if (nft_trans_elem_set_bound(trans)) {
9166                                 nft_trans_destroy(trans);
9167                                 break;
9168                         }
9169                         te = (struct nft_trans_elem *)trans->data;
9170                         nft_setelem_remove(net, te->set, &te->elem);
9171                         if (!nft_setelem_is_catchall(te->set, &te->elem))
9172                                 atomic_dec(&te->set->nelems);
9173                         break;
9174                 case NFT_MSG_DELSETELEM:
9175                         te = (struct nft_trans_elem *)trans->data;
9176
9177                         nft_setelem_data_activate(net, te->set, &te->elem);
9178                         nft_setelem_activate(net, te->set, &te->elem);
9179                         if (!nft_setelem_is_catchall(te->set, &te->elem))
9180                                 te->set->ndeact--;
9181
9182                         nft_trans_destroy(trans);
9183                         break;
9184                 case NFT_MSG_NEWOBJ:
9185                         if (nft_trans_obj_update(trans)) {
9186                                 nft_obj_destroy(&trans->ctx, nft_trans_obj_newobj(trans));
9187                                 nft_trans_destroy(trans);
9188                         } else {
9189                                 trans->ctx.table->use--;
9190                                 nft_obj_del(nft_trans_obj(trans));
9191                         }
9192                         break;
9193                 case NFT_MSG_DELOBJ:
9194                         trans->ctx.table->use++;
9195                         nft_clear(trans->ctx.net, nft_trans_obj(trans));
9196                         nft_trans_destroy(trans);
9197                         break;
9198                 case NFT_MSG_NEWFLOWTABLE:
9199                         if (nft_trans_flowtable_update(trans)) {
9200                                 nft_unregister_flowtable_net_hooks(net,
9201                                                 &nft_trans_flowtable_hooks(trans));
9202                         } else {
9203                                 trans->ctx.table->use--;
9204                                 list_del_rcu(&nft_trans_flowtable(trans)->list);
9205                                 nft_unregister_flowtable_net_hooks(net,
9206                                                 &nft_trans_flowtable(trans)->hook_list);
9207                         }
9208                         break;
9209                 case NFT_MSG_DELFLOWTABLE:
9210                         if (nft_trans_flowtable_update(trans)) {
9211                                 list_splice(&nft_trans_flowtable_hooks(trans),
9212                                             &nft_trans_flowtable(trans)->hook_list);
9213                         } else {
9214                                 trans->ctx.table->use++;
9215                                 nft_clear(trans->ctx.net, nft_trans_flowtable(trans));
9216                         }
9217                         nft_trans_destroy(trans);
9218                         break;
9219                 }
9220         }
9221
9222         synchronize_rcu();
9223
9224         list_for_each_entry_safe_reverse(trans, next,
9225                                          &nft_net->commit_list, list) {
9226                 list_del(&trans->list);
9227                 nf_tables_abort_release(trans);
9228         }
9229
9230         if (action == NFNL_ABORT_AUTOLOAD)
9231                 nf_tables_module_autoload(net);
9232         else
9233                 nf_tables_module_autoload_cleanup(net);
9234
9235         return 0;
9236 }
9237
9238 static void nf_tables_cleanup(struct net *net)
9239 {
9240         nft_validate_state_update(net, NFT_VALIDATE_SKIP);
9241 }
9242
9243 static int nf_tables_abort(struct net *net, struct sk_buff *skb,
9244                            enum nfnl_abort_action action)
9245 {
9246         struct nftables_pernet *nft_net = nft_pernet(net);
9247         int ret = __nf_tables_abort(net, action);
9248
9249         mutex_unlock(&nft_net->commit_mutex);
9250
9251         return ret;
9252 }
9253
9254 static bool nf_tables_valid_genid(struct net *net, u32 genid)
9255 {
9256         struct nftables_pernet *nft_net = nft_pernet(net);
9257         bool genid_ok;
9258
9259         mutex_lock(&nft_net->commit_mutex);
9260
9261         genid_ok = genid == 0 || nft_net->base_seq == genid;
9262         if (!genid_ok)
9263                 mutex_unlock(&nft_net->commit_mutex);
9264
9265         /* else, commit mutex has to be released by commit or abort function */
9266         return genid_ok;
9267 }
9268
9269 static const struct nfnetlink_subsystem nf_tables_subsys = {
9270         .name           = "nf_tables",
9271         .subsys_id      = NFNL_SUBSYS_NFTABLES,
9272         .cb_count       = NFT_MSG_MAX,
9273         .cb             = nf_tables_cb,
9274         .commit         = nf_tables_commit,
9275         .abort          = nf_tables_abort,
9276         .cleanup        = nf_tables_cleanup,
9277         .valid_genid    = nf_tables_valid_genid,
9278         .owner          = THIS_MODULE,
9279 };
9280
9281 int nft_chain_validate_dependency(const struct nft_chain *chain,
9282                                   enum nft_chain_types type)
9283 {
9284         const struct nft_base_chain *basechain;
9285
9286         if (nft_is_base_chain(chain)) {
9287                 basechain = nft_base_chain(chain);
9288                 if (basechain->type->type != type)
9289                         return -EOPNOTSUPP;
9290         }
9291         return 0;
9292 }
9293 EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
9294
9295 int nft_chain_validate_hooks(const struct nft_chain *chain,
9296                              unsigned int hook_flags)
9297 {
9298         struct nft_base_chain *basechain;
9299
9300         if (nft_is_base_chain(chain)) {
9301                 basechain = nft_base_chain(chain);
9302
9303                 if ((1 << basechain->ops.hooknum) & hook_flags)
9304                         return 0;
9305
9306                 return -EOPNOTSUPP;
9307         }
9308
9309         return 0;
9310 }
9311 EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
9312
9313 /*
9314  * Loop detection - walk through the ruleset beginning at the destination chain
9315  * of a new jump until either the source chain is reached (loop) or all
9316  * reachable chains have been traversed.
9317  *
9318  * The loop check is performed whenever a new jump verdict is added to an
9319  * expression or verdict map or a verdict map is bound to a new chain.
9320  */
9321
9322 static int nf_tables_check_loops(const struct nft_ctx *ctx,
9323                                  const struct nft_chain *chain);
9324
9325 static int nft_check_loops(const struct nft_ctx *ctx,
9326                            const struct nft_set_ext *ext)
9327 {
9328         const struct nft_data *data;
9329         int ret;
9330
9331         data = nft_set_ext_data(ext);
9332         switch (data->verdict.code) {
9333         case NFT_JUMP:
9334         case NFT_GOTO:
9335                 ret = nf_tables_check_loops(ctx, data->verdict.chain);
9336                 break;
9337         default:
9338                 ret = 0;
9339                 break;
9340         }
9341
9342         return ret;
9343 }
9344
9345 static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
9346                                         struct nft_set *set,
9347                                         const struct nft_set_iter *iter,
9348                                         struct nft_set_elem *elem)
9349 {
9350         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
9351
9352         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
9353             *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
9354                 return 0;
9355
9356         return nft_check_loops(ctx, ext);
9357 }
9358
9359 static int nft_set_catchall_loops(const struct nft_ctx *ctx,
9360                                   struct nft_set *set)
9361 {
9362         u8 genmask = nft_genmask_next(ctx->net);
9363         struct nft_set_elem_catchall *catchall;
9364         struct nft_set_ext *ext;
9365         int ret = 0;
9366
9367         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
9368                 ext = nft_set_elem_ext(set, catchall->elem);
9369                 if (!nft_set_elem_active(ext, genmask))
9370                         continue;
9371
9372                 ret = nft_check_loops(ctx, ext);
9373                 if (ret < 0)
9374                         return ret;
9375         }
9376
9377         return ret;
9378 }
9379
9380 static int nf_tables_check_loops(const struct nft_ctx *ctx,
9381                                  const struct nft_chain *chain)
9382 {
9383         const struct nft_rule *rule;
9384         const struct nft_expr *expr, *last;
9385         struct nft_set *set;
9386         struct nft_set_binding *binding;
9387         struct nft_set_iter iter;
9388
9389         if (ctx->chain == chain)
9390                 return -ELOOP;
9391
9392         list_for_each_entry(rule, &chain->rules, list) {
9393                 nft_rule_for_each_expr(expr, last, rule) {
9394                         struct nft_immediate_expr *priv;
9395                         const struct nft_data *data;
9396                         int err;
9397
9398                         if (strcmp(expr->ops->type->name, "immediate"))
9399                                 continue;
9400
9401                         priv = nft_expr_priv(expr);
9402                         if (priv->dreg != NFT_REG_VERDICT)
9403                                 continue;
9404
9405                         data = &priv->data;
9406                         switch (data->verdict.code) {
9407                         case NFT_JUMP:
9408                         case NFT_GOTO:
9409                                 err = nf_tables_check_loops(ctx,
9410                                                         data->verdict.chain);
9411                                 if (err < 0)
9412                                         return err;
9413                                 break;
9414                         default:
9415                                 break;
9416                         }
9417                 }
9418         }
9419
9420         list_for_each_entry(set, &ctx->table->sets, list) {
9421                 if (!nft_is_active_next(ctx->net, set))
9422                         continue;
9423                 if (!(set->flags & NFT_SET_MAP) ||
9424                     set->dtype != NFT_DATA_VERDICT)
9425                         continue;
9426
9427                 list_for_each_entry(binding, &set->bindings, list) {
9428                         if (!(binding->flags & NFT_SET_MAP) ||
9429                             binding->chain != chain)
9430                                 continue;
9431
9432                         iter.genmask    = nft_genmask_next(ctx->net);
9433                         iter.skip       = 0;
9434                         iter.count      = 0;
9435                         iter.err        = 0;
9436                         iter.fn         = nf_tables_loop_check_setelem;
9437
9438                         set->ops->walk(ctx, set, &iter);
9439                         if (!iter.err)
9440                                 iter.err = nft_set_catchall_loops(ctx, set);
9441
9442                         if (iter.err < 0)
9443                                 return iter.err;
9444                 }
9445         }
9446
9447         return 0;
9448 }
9449
9450 /**
9451  *      nft_parse_u32_check - fetch u32 attribute and check for maximum value
9452  *
9453  *      @attr: netlink attribute to fetch value from
9454  *      @max: maximum value to be stored in dest
9455  *      @dest: pointer to the variable
9456  *
9457  *      Parse, check and store a given u32 netlink attribute into variable.
9458  *      This function returns -ERANGE if the value goes over maximum value.
9459  *      Otherwise a 0 is returned and the attribute value is stored in the
9460  *      destination variable.
9461  */
9462 int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
9463 {
9464         u32 val;
9465
9466         val = ntohl(nla_get_be32(attr));
9467         if (val > max)
9468                 return -ERANGE;
9469
9470         *dest = val;
9471         return 0;
9472 }
9473 EXPORT_SYMBOL_GPL(nft_parse_u32_check);
9474
9475 static int nft_parse_register(const struct nlattr *attr, u32 *preg)
9476 {
9477         unsigned int reg;
9478
9479         reg = ntohl(nla_get_be32(attr));
9480         switch (reg) {
9481         case NFT_REG_VERDICT...NFT_REG_4:
9482                 *preg = reg * NFT_REG_SIZE / NFT_REG32_SIZE;
9483                 break;
9484         case NFT_REG32_00...NFT_REG32_15:
9485                 *preg = reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
9486                 break;
9487         default:
9488                 return -ERANGE;
9489         }
9490
9491         return 0;
9492 }
9493
9494 /**
9495  *      nft_dump_register - dump a register value to a netlink attribute
9496  *
9497  *      @skb: socket buffer
9498  *      @attr: attribute number
9499  *      @reg: register number
9500  *
9501  *      Construct a netlink attribute containing the register number. For
9502  *      compatibility reasons, register numbers being a multiple of 4 are
9503  *      translated to the corresponding 128 bit register numbers.
9504  */
9505 int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
9506 {
9507         if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
9508                 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
9509         else
9510                 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
9511
9512         return nla_put_be32(skb, attr, htonl(reg));
9513 }
9514 EXPORT_SYMBOL_GPL(nft_dump_register);
9515
9516 static int nft_validate_register_load(enum nft_registers reg, unsigned int len)
9517 {
9518         if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9519                 return -EINVAL;
9520         if (len == 0)
9521                 return -EINVAL;
9522         if (reg * NFT_REG32_SIZE + len > sizeof_field(struct nft_regs, data))
9523                 return -ERANGE;
9524
9525         return 0;
9526 }
9527
9528 int nft_parse_register_load(const struct nlattr *attr, u8 *sreg, u32 len)
9529 {
9530         u32 reg;
9531         int err;
9532
9533         err = nft_parse_register(attr, &reg);
9534         if (err < 0)
9535                 return err;
9536
9537         err = nft_validate_register_load(reg, len);
9538         if (err < 0)
9539                 return err;
9540
9541         *sreg = reg;
9542         return 0;
9543 }
9544 EXPORT_SYMBOL_GPL(nft_parse_register_load);
9545
9546 static int nft_validate_register_store(const struct nft_ctx *ctx,
9547                                        enum nft_registers reg,
9548                                        const struct nft_data *data,
9549                                        enum nft_data_types type,
9550                                        unsigned int len)
9551 {
9552         int err;
9553
9554         switch (reg) {
9555         case NFT_REG_VERDICT:
9556                 if (type != NFT_DATA_VERDICT)
9557                         return -EINVAL;
9558
9559                 if (data != NULL &&
9560                     (data->verdict.code == NFT_GOTO ||
9561                      data->verdict.code == NFT_JUMP)) {
9562                         err = nf_tables_check_loops(ctx, data->verdict.chain);
9563                         if (err < 0)
9564                                 return err;
9565                 }
9566
9567                 return 0;
9568         default:
9569                 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9570                         return -EINVAL;
9571                 if (len == 0)
9572                         return -EINVAL;
9573                 if (reg * NFT_REG32_SIZE + len >
9574                     sizeof_field(struct nft_regs, data))
9575                         return -ERANGE;
9576
9577                 if (data != NULL && type != NFT_DATA_VALUE)
9578                         return -EINVAL;
9579                 return 0;
9580         }
9581 }
9582
9583 int nft_parse_register_store(const struct nft_ctx *ctx,
9584                              const struct nlattr *attr, u8 *dreg,
9585                              const struct nft_data *data,
9586                              enum nft_data_types type, unsigned int len)
9587 {
9588         int err;
9589         u32 reg;
9590
9591         err = nft_parse_register(attr, &reg);
9592         if (err < 0)
9593                 return err;
9594
9595         err = nft_validate_register_store(ctx, reg, data, type, len);
9596         if (err < 0)
9597                 return err;
9598
9599         *dreg = reg;
9600         return 0;
9601 }
9602 EXPORT_SYMBOL_GPL(nft_parse_register_store);
9603
9604 static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
9605         [NFTA_VERDICT_CODE]     = { .type = NLA_U32 },
9606         [NFTA_VERDICT_CHAIN]    = { .type = NLA_STRING,
9607                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
9608         [NFTA_VERDICT_CHAIN_ID] = { .type = NLA_U32 },
9609 };
9610
9611 static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
9612                             struct nft_data_desc *desc, const struct nlattr *nla)
9613 {
9614         u8 genmask = nft_genmask_next(ctx->net);
9615         struct nlattr *tb[NFTA_VERDICT_MAX + 1];
9616         struct nft_chain *chain;
9617         int err;
9618
9619         err = nla_parse_nested_deprecated(tb, NFTA_VERDICT_MAX, nla,
9620                                           nft_verdict_policy, NULL);
9621         if (err < 0)
9622                 return err;
9623
9624         if (!tb[NFTA_VERDICT_CODE])
9625                 return -EINVAL;
9626         data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
9627
9628         switch (data->verdict.code) {
9629         default:
9630                 switch (data->verdict.code & NF_VERDICT_MASK) {
9631                 case NF_ACCEPT:
9632                 case NF_DROP:
9633                 case NF_QUEUE:
9634                         break;
9635                 default:
9636                         return -EINVAL;
9637                 }
9638                 fallthrough;
9639         case NFT_CONTINUE:
9640         case NFT_BREAK:
9641         case NFT_RETURN:
9642                 break;
9643         case NFT_JUMP:
9644         case NFT_GOTO:
9645                 if (tb[NFTA_VERDICT_CHAIN]) {
9646                         chain = nft_chain_lookup(ctx->net, ctx->table,
9647                                                  tb[NFTA_VERDICT_CHAIN],
9648                                                  genmask);
9649                 } else if (tb[NFTA_VERDICT_CHAIN_ID]) {
9650                         chain = nft_chain_lookup_byid(ctx->net, ctx->table,
9651                                                       tb[NFTA_VERDICT_CHAIN_ID]);
9652                         if (IS_ERR(chain))
9653                                 return PTR_ERR(chain);
9654                 } else {
9655                         return -EINVAL;
9656                 }
9657
9658                 if (IS_ERR(chain))
9659                         return PTR_ERR(chain);
9660                 if (nft_is_base_chain(chain))
9661                         return -EOPNOTSUPP;
9662                 if (nft_chain_is_bound(chain))
9663                         return -EINVAL;
9664                 if (desc->flags & NFT_DATA_DESC_SETELEM &&
9665                     chain->flags & NFT_CHAIN_BINDING)
9666                         return -EINVAL;
9667
9668                 chain->use++;
9669                 data->verdict.chain = chain;
9670                 break;
9671         }
9672
9673         desc->len = sizeof(data->verdict);
9674
9675         return 0;
9676 }
9677
9678 static void nft_verdict_uninit(const struct nft_data *data)
9679 {
9680         struct nft_chain *chain;
9681         struct nft_rule *rule;
9682
9683         switch (data->verdict.code) {
9684         case NFT_JUMP:
9685         case NFT_GOTO:
9686                 chain = data->verdict.chain;
9687                 chain->use--;
9688
9689                 if (!nft_chain_is_bound(chain))
9690                         break;
9691
9692                 chain->table->use--;
9693                 list_for_each_entry(rule, &chain->rules, list)
9694                         chain->use--;
9695
9696                 nft_chain_del(chain);
9697                 break;
9698         }
9699 }
9700
9701 int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
9702 {
9703         struct nlattr *nest;
9704
9705         nest = nla_nest_start_noflag(skb, type);
9706         if (!nest)
9707                 goto nla_put_failure;
9708
9709         if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
9710                 goto nla_put_failure;
9711
9712         switch (v->code) {
9713         case NFT_JUMP:
9714         case NFT_GOTO:
9715                 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
9716                                    v->chain->name))
9717                         goto nla_put_failure;
9718         }
9719         nla_nest_end(skb, nest);
9720         return 0;
9721
9722 nla_put_failure:
9723         return -1;
9724 }
9725
9726 static int nft_value_init(const struct nft_ctx *ctx,
9727                           struct nft_data *data, struct nft_data_desc *desc,
9728                           const struct nlattr *nla)
9729 {
9730         unsigned int len;
9731
9732         len = nla_len(nla);
9733         if (len == 0)
9734                 return -EINVAL;
9735         if (len > desc->size)
9736                 return -EOVERFLOW;
9737         if (desc->len) {
9738                 if (len != desc->len)
9739                         return -EINVAL;
9740         } else {
9741                 desc->len = len;
9742         }
9743
9744         nla_memcpy(data->data, nla, len);
9745
9746         return 0;
9747 }
9748
9749 static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
9750                           unsigned int len)
9751 {
9752         return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
9753 }
9754
9755 static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
9756         [NFTA_DATA_VALUE]       = { .type = NLA_BINARY },
9757         [NFTA_DATA_VERDICT]     = { .type = NLA_NESTED },
9758 };
9759
9760 /**
9761  *      nft_data_init - parse nf_tables data netlink attributes
9762  *
9763  *      @ctx: context of the expression using the data
9764  *      @data: destination struct nft_data
9765  *      @desc: data description
9766  *      @nla: netlink attribute containing data
9767  *
9768  *      Parse the netlink data attributes and initialize a struct nft_data.
9769  *      The type and length of data are returned in the data description.
9770  *
9771  *      The caller can indicate that it only wants to accept data of type
9772  *      NFT_DATA_VALUE by passing NULL for the ctx argument.
9773  */
9774 int nft_data_init(const struct nft_ctx *ctx, struct nft_data *data,
9775                   struct nft_data_desc *desc, const struct nlattr *nla)
9776 {
9777         struct nlattr *tb[NFTA_DATA_MAX + 1];
9778         int err;
9779
9780         if (WARN_ON_ONCE(!desc->size))
9781                 return -EINVAL;
9782
9783         err = nla_parse_nested_deprecated(tb, NFTA_DATA_MAX, nla,
9784                                           nft_data_policy, NULL);
9785         if (err < 0)
9786                 return err;
9787
9788         if (tb[NFTA_DATA_VALUE]) {
9789                 if (desc->type != NFT_DATA_VALUE)
9790                         return -EINVAL;
9791
9792                 err = nft_value_init(ctx, data, desc, tb[NFTA_DATA_VALUE]);
9793         } else if (tb[NFTA_DATA_VERDICT] && ctx != NULL) {
9794                 if (desc->type != NFT_DATA_VERDICT)
9795                         return -EINVAL;
9796
9797                 err = nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
9798         } else {
9799                 err = -EINVAL;
9800         }
9801
9802         return err;
9803 }
9804 EXPORT_SYMBOL_GPL(nft_data_init);
9805
9806 /**
9807  *      nft_data_release - release a nft_data item
9808  *
9809  *      @data: struct nft_data to release
9810  *      @type: type of data
9811  *
9812  *      Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
9813  *      all others need to be released by calling this function.
9814  */
9815 void nft_data_release(const struct nft_data *data, enum nft_data_types type)
9816 {
9817         if (type < NFT_DATA_VERDICT)
9818                 return;
9819         switch (type) {
9820         case NFT_DATA_VERDICT:
9821                 return nft_verdict_uninit(data);
9822         default:
9823                 WARN_ON(1);
9824         }
9825 }
9826 EXPORT_SYMBOL_GPL(nft_data_release);
9827
9828 int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
9829                   enum nft_data_types type, unsigned int len)
9830 {
9831         struct nlattr *nest;
9832         int err;
9833
9834         nest = nla_nest_start_noflag(skb, attr);
9835         if (nest == NULL)
9836                 return -1;
9837
9838         switch (type) {
9839         case NFT_DATA_VALUE:
9840                 err = nft_value_dump(skb, data, len);
9841                 break;
9842         case NFT_DATA_VERDICT:
9843                 err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
9844                 break;
9845         default:
9846                 err = -EINVAL;
9847                 WARN_ON(1);
9848         }
9849
9850         nla_nest_end(skb, nest);
9851         return err;
9852 }
9853 EXPORT_SYMBOL_GPL(nft_data_dump);
9854
9855 int __nft_release_basechain(struct nft_ctx *ctx)
9856 {
9857         struct nft_rule *rule, *nr;
9858
9859         if (WARN_ON(!nft_is_base_chain(ctx->chain)))
9860                 return 0;
9861
9862         nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
9863         list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
9864                 list_del(&rule->list);
9865                 ctx->chain->use--;
9866                 nf_tables_rule_release(ctx, rule);
9867         }
9868         nft_chain_del(ctx->chain);
9869         ctx->table->use--;
9870         nf_tables_chain_destroy(ctx);
9871
9872         return 0;
9873 }
9874 EXPORT_SYMBOL_GPL(__nft_release_basechain);
9875
9876 static void __nft_release_hook(struct net *net, struct nft_table *table)
9877 {
9878         struct nft_flowtable *flowtable;
9879         struct nft_chain *chain;
9880
9881         list_for_each_entry(chain, &table->chains, list)
9882                 __nf_tables_unregister_hook(net, table, chain, true);
9883         list_for_each_entry(flowtable, &table->flowtables, list)
9884                 __nft_unregister_flowtable_net_hooks(net, &flowtable->hook_list,
9885                                                      true);
9886 }
9887
9888 static void __nft_release_hooks(struct net *net)
9889 {
9890         struct nftables_pernet *nft_net = nft_pernet(net);
9891         struct nft_table *table;
9892
9893         list_for_each_entry(table, &nft_net->tables, list) {
9894                 if (nft_table_has_owner(table))
9895                         continue;
9896
9897                 __nft_release_hook(net, table);
9898         }
9899 }
9900
9901 static void __nft_release_table(struct net *net, struct nft_table *table)
9902 {
9903         struct nft_flowtable *flowtable, *nf;
9904         struct nft_chain *chain, *nc;
9905         struct nft_object *obj, *ne;
9906         struct nft_rule *rule, *nr;
9907         struct nft_set *set, *ns;
9908         struct nft_ctx ctx = {
9909                 .net    = net,
9910                 .family = NFPROTO_NETDEV,
9911         };
9912
9913         ctx.family = table->family;
9914         ctx.table = table;
9915         list_for_each_entry(chain, &table->chains, list) {
9916                 ctx.chain = chain;
9917                 list_for_each_entry_safe(rule, nr, &chain->rules, list) {
9918                         list_del(&rule->list);
9919                         chain->use--;
9920                         nf_tables_rule_release(&ctx, rule);
9921                 }
9922         }
9923         list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
9924                 list_del(&flowtable->list);
9925                 table->use--;
9926                 nf_tables_flowtable_destroy(flowtable);
9927         }
9928         list_for_each_entry_safe(set, ns, &table->sets, list) {
9929                 list_del(&set->list);
9930                 table->use--;
9931                 nft_set_destroy(&ctx, set);
9932         }
9933         list_for_each_entry_safe(obj, ne, &table->objects, list) {
9934                 nft_obj_del(obj);
9935                 table->use--;
9936                 nft_obj_destroy(&ctx, obj);
9937         }
9938         list_for_each_entry_safe(chain, nc, &table->chains, list) {
9939                 ctx.chain = chain;
9940                 nft_chain_del(chain);
9941                 table->use--;
9942                 nf_tables_chain_destroy(&ctx);
9943         }
9944         nf_tables_table_destroy(&ctx);
9945 }
9946
9947 static void __nft_release_tables(struct net *net)
9948 {
9949         struct nftables_pernet *nft_net = nft_pernet(net);
9950         struct nft_table *table, *nt;
9951
9952         list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
9953                 if (nft_table_has_owner(table))
9954                         continue;
9955
9956                 list_del(&table->list);
9957
9958                 __nft_release_table(net, table);
9959         }
9960 }
9961
9962 static int nft_rcv_nl_event(struct notifier_block *this, unsigned long event,
9963                             void *ptr)
9964 {
9965         struct nft_table *table, *to_delete[8];
9966         struct nftables_pernet *nft_net;
9967         struct netlink_notify *n = ptr;
9968         struct net *net = n->net;
9969         unsigned int deleted;
9970         bool restart = false;
9971
9972         if (event != NETLINK_URELEASE || n->protocol != NETLINK_NETFILTER)
9973                 return NOTIFY_DONE;
9974
9975         nft_net = nft_pernet(net);
9976         deleted = 0;
9977         mutex_lock(&nft_net->commit_mutex);
9978 again:
9979         list_for_each_entry(table, &nft_net->tables, list) {
9980                 if (nft_table_has_owner(table) &&
9981                     n->portid == table->nlpid) {
9982                         __nft_release_hook(net, table);
9983                         list_del_rcu(&table->list);
9984                         to_delete[deleted++] = table;
9985                         if (deleted >= ARRAY_SIZE(to_delete))
9986                                 break;
9987                 }
9988         }
9989         if (deleted) {
9990                 restart = deleted >= ARRAY_SIZE(to_delete);
9991                 synchronize_rcu();
9992                 while (deleted)
9993                         __nft_release_table(net, to_delete[--deleted]);
9994
9995                 if (restart)
9996                         goto again;
9997         }
9998         mutex_unlock(&nft_net->commit_mutex);
9999
10000         return NOTIFY_DONE;
10001 }
10002
10003 static struct notifier_block nft_nl_notifier = {
10004         .notifier_call  = nft_rcv_nl_event,
10005 };
10006
10007 static int __net_init nf_tables_init_net(struct net *net)
10008 {
10009         struct nftables_pernet *nft_net = nft_pernet(net);
10010
10011         INIT_LIST_HEAD(&nft_net->tables);
10012         INIT_LIST_HEAD(&nft_net->commit_list);
10013         INIT_LIST_HEAD(&nft_net->module_list);
10014         INIT_LIST_HEAD(&nft_net->notify_list);
10015         mutex_init(&nft_net->commit_mutex);
10016         nft_net->base_seq = 1;
10017         nft_net->validate_state = NFT_VALIDATE_SKIP;
10018
10019         return 0;
10020 }
10021
10022 static void __net_exit nf_tables_pre_exit_net(struct net *net)
10023 {
10024         struct nftables_pernet *nft_net = nft_pernet(net);
10025
10026         mutex_lock(&nft_net->commit_mutex);
10027         __nft_release_hooks(net);
10028         mutex_unlock(&nft_net->commit_mutex);
10029 }
10030
10031 static void __net_exit nf_tables_exit_net(struct net *net)
10032 {
10033         struct nftables_pernet *nft_net = nft_pernet(net);
10034
10035         mutex_lock(&nft_net->commit_mutex);
10036         if (!list_empty(&nft_net->commit_list))
10037                 __nf_tables_abort(net, NFNL_ABORT_NONE);
10038         __nft_release_tables(net);
10039         mutex_unlock(&nft_net->commit_mutex);
10040         WARN_ON_ONCE(!list_empty(&nft_net->tables));
10041         WARN_ON_ONCE(!list_empty(&nft_net->module_list));
10042         WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
10043 }
10044
10045 static struct pernet_operations nf_tables_net_ops = {
10046         .init           = nf_tables_init_net,
10047         .pre_exit       = nf_tables_pre_exit_net,
10048         .exit           = nf_tables_exit_net,
10049         .id             = &nf_tables_net_id,
10050         .size           = sizeof(struct nftables_pernet),
10051 };
10052
10053 static int __init nf_tables_module_init(void)
10054 {
10055         int err;
10056
10057         err = register_pernet_subsys(&nf_tables_net_ops);
10058         if (err < 0)
10059                 return err;
10060
10061         err = nft_chain_filter_init();
10062         if (err < 0)
10063                 goto err_chain_filter;
10064
10065         err = nf_tables_core_module_init();
10066         if (err < 0)
10067                 goto err_core_module;
10068
10069         err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
10070         if (err < 0)
10071                 goto err_netdev_notifier;
10072
10073         err = rhltable_init(&nft_objname_ht, &nft_objname_ht_params);
10074         if (err < 0)
10075                 goto err_rht_objname;
10076
10077         err = nft_offload_init();
10078         if (err < 0)
10079                 goto err_offload;
10080
10081         err = netlink_register_notifier(&nft_nl_notifier);
10082         if (err < 0)
10083                 goto err_netlink_notifier;
10084
10085         /* must be last */
10086         err = nfnetlink_subsys_register(&nf_tables_subsys);
10087         if (err < 0)
10088                 goto err_nfnl_subsys;
10089
10090         nft_chain_route_init();
10091
10092         return err;
10093
10094 err_nfnl_subsys:
10095         netlink_unregister_notifier(&nft_nl_notifier);
10096 err_netlink_notifier:
10097         nft_offload_exit();
10098 err_offload:
10099         rhltable_destroy(&nft_objname_ht);
10100 err_rht_objname:
10101         unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
10102 err_netdev_notifier:
10103         nf_tables_core_module_exit();
10104 err_core_module:
10105         nft_chain_filter_fini();
10106 err_chain_filter:
10107         unregister_pernet_subsys(&nf_tables_net_ops);
10108         return err;
10109 }
10110
10111 static void __exit nf_tables_module_exit(void)
10112 {
10113         nfnetlink_subsys_unregister(&nf_tables_subsys);
10114         netlink_unregister_notifier(&nft_nl_notifier);
10115         nft_offload_exit();
10116         unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
10117         nft_chain_filter_fini();
10118         nft_chain_route_fini();
10119         unregister_pernet_subsys(&nf_tables_net_ops);
10120         cancel_work_sync(&trans_destroy_work);
10121         rcu_barrier();
10122         rhltable_destroy(&nft_objname_ht);
10123         nf_tables_core_module_exit();
10124 }
10125
10126 module_init(nf_tables_module_init);
10127 module_exit(nf_tables_module_exit);
10128
10129 MODULE_LICENSE("GPL");
10130 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
10131 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);