]> Git Repo - linux.git/blob - net/netfilter/nf_tables_api.c
RISC-V: replace cbom instructions with an insn-def
[linux.git] / net / netfilter / nf_tables_api.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2007-2009 Patrick McHardy <[email protected]>
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         strscpy(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(&cpu_stats->syncp);
1538                         pkts = cpu_stats->pkts;
1539                         bytes = cpu_stats->bytes;
1540                 } while (u64_stats_fetch_retry(&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, bool reset)
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, reset) < 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, bool reset)
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, reset) < 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 int nft_expr_inner_parse(const struct nft_ctx *ctx, const struct nlattr *nla,
2861                          struct nft_expr_info *info)
2862 {
2863         struct nlattr *tb[NFTA_EXPR_MAX + 1];
2864         const struct nft_expr_type *type;
2865         int err;
2866
2867         err = nla_parse_nested_deprecated(tb, NFTA_EXPR_MAX, nla,
2868                                           nft_expr_policy, NULL);
2869         if (err < 0)
2870                 return err;
2871
2872         if (!tb[NFTA_EXPR_DATA])
2873                 return -EINVAL;
2874
2875         type = __nft_expr_type_get(ctx->family, tb[NFTA_EXPR_NAME]);
2876         if (!type)
2877                 return -ENOENT;
2878
2879         if (!type->inner_ops)
2880                 return -EOPNOTSUPP;
2881
2882         err = nla_parse_nested_deprecated(info->tb, type->maxattr,
2883                                           tb[NFTA_EXPR_DATA],
2884                                           type->policy, NULL);
2885         if (err < 0)
2886                 goto err_nla_parse;
2887
2888         info->attr = nla;
2889         info->ops = type->inner_ops;
2890
2891         return 0;
2892
2893 err_nla_parse:
2894         return err;
2895 }
2896
2897 static int nf_tables_newexpr(const struct nft_ctx *ctx,
2898                              const struct nft_expr_info *expr_info,
2899                              struct nft_expr *expr)
2900 {
2901         const struct nft_expr_ops *ops = expr_info->ops;
2902         int err;
2903
2904         expr->ops = ops;
2905         if (ops->init) {
2906                 err = ops->init(ctx, expr, (const struct nlattr **)expr_info->tb);
2907                 if (err < 0)
2908                         goto err1;
2909         }
2910
2911         return 0;
2912 err1:
2913         expr->ops = NULL;
2914         return err;
2915 }
2916
2917 static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
2918                                    struct nft_expr *expr)
2919 {
2920         const struct nft_expr_type *type = expr->ops->type;
2921
2922         if (expr->ops->destroy)
2923                 expr->ops->destroy(ctx, expr);
2924         module_put(type->owner);
2925 }
2926
2927 static struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
2928                                       const struct nlattr *nla)
2929 {
2930         struct nft_expr_info expr_info;
2931         struct nft_expr *expr;
2932         struct module *owner;
2933         int err;
2934
2935         err = nf_tables_expr_parse(ctx, nla, &expr_info);
2936         if (err < 0)
2937                 goto err_expr_parse;
2938
2939         err = -EOPNOTSUPP;
2940         if (!(expr_info.ops->type->flags & NFT_EXPR_STATEFUL))
2941                 goto err_expr_stateful;
2942
2943         err = -ENOMEM;
2944         expr = kzalloc(expr_info.ops->size, GFP_KERNEL_ACCOUNT);
2945         if (expr == NULL)
2946                 goto err_expr_stateful;
2947
2948         err = nf_tables_newexpr(ctx, &expr_info, expr);
2949         if (err < 0)
2950                 goto err_expr_new;
2951
2952         return expr;
2953 err_expr_new:
2954         kfree(expr);
2955 err_expr_stateful:
2956         owner = expr_info.ops->type->owner;
2957         if (expr_info.ops->type->release_ops)
2958                 expr_info.ops->type->release_ops(expr_info.ops);
2959
2960         module_put(owner);
2961 err_expr_parse:
2962         return ERR_PTR(err);
2963 }
2964
2965 int nft_expr_clone(struct nft_expr *dst, struct nft_expr *src)
2966 {
2967         int err;
2968
2969         if (src->ops->clone) {
2970                 dst->ops = src->ops;
2971                 err = src->ops->clone(dst, src);
2972                 if (err < 0)
2973                         return err;
2974         } else {
2975                 memcpy(dst, src, src->ops->size);
2976         }
2977
2978         __module_get(src->ops->type->owner);
2979
2980         return 0;
2981 }
2982
2983 void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
2984 {
2985         nf_tables_expr_destroy(ctx, expr);
2986         kfree(expr);
2987 }
2988
2989 /*
2990  * Rules
2991  */
2992
2993 static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
2994                                           u64 handle)
2995 {
2996         struct nft_rule *rule;
2997
2998         // FIXME: this sucks
2999         list_for_each_entry_rcu(rule, &chain->rules, list) {
3000                 if (handle == rule->handle)
3001                         return rule;
3002         }
3003
3004         return ERR_PTR(-ENOENT);
3005 }
3006
3007 static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
3008                                         const struct nlattr *nla)
3009 {
3010         if (nla == NULL)
3011                 return ERR_PTR(-EINVAL);
3012
3013         return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
3014 }
3015
3016 static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
3017         [NFTA_RULE_TABLE]       = { .type = NLA_STRING,
3018                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
3019         [NFTA_RULE_CHAIN]       = { .type = NLA_STRING,
3020                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
3021         [NFTA_RULE_HANDLE]      = { .type = NLA_U64 },
3022         [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
3023         [NFTA_RULE_COMPAT]      = { .type = NLA_NESTED },
3024         [NFTA_RULE_POSITION]    = { .type = NLA_U64 },
3025         [NFTA_RULE_USERDATA]    = { .type = NLA_BINARY,
3026                                     .len = NFT_USERDATA_MAXLEN },
3027         [NFTA_RULE_ID]          = { .type = NLA_U32 },
3028         [NFTA_RULE_POSITION_ID] = { .type = NLA_U32 },
3029         [NFTA_RULE_CHAIN_ID]    = { .type = NLA_U32 },
3030 };
3031
3032 static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
3033                                     u32 portid, u32 seq, int event,
3034                                     u32 flags, int family,
3035                                     const struct nft_table *table,
3036                                     const struct nft_chain *chain,
3037                                     const struct nft_rule *rule, u64 handle,
3038                                     bool reset)
3039 {
3040         struct nlmsghdr *nlh;
3041         const struct nft_expr *expr, *next;
3042         struct nlattr *list;
3043         u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3044
3045         nlh = nfnl_msg_put(skb, portid, seq, type, flags, family, NFNETLINK_V0,
3046                            nft_base_seq(net));
3047         if (!nlh)
3048                 goto nla_put_failure;
3049
3050         if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
3051                 goto nla_put_failure;
3052         if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
3053                 goto nla_put_failure;
3054         if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
3055                          NFTA_RULE_PAD))
3056                 goto nla_put_failure;
3057
3058         if (event != NFT_MSG_DELRULE && handle) {
3059                 if (nla_put_be64(skb, NFTA_RULE_POSITION, cpu_to_be64(handle),
3060                                  NFTA_RULE_PAD))
3061                         goto nla_put_failure;
3062         }
3063
3064         if (chain->flags & NFT_CHAIN_HW_OFFLOAD)
3065                 nft_flow_rule_stats(chain, rule);
3066
3067         list = nla_nest_start_noflag(skb, NFTA_RULE_EXPRESSIONS);
3068         if (list == NULL)
3069                 goto nla_put_failure;
3070         nft_rule_for_each_expr(expr, next, rule) {
3071                 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr, reset) < 0)
3072                         goto nla_put_failure;
3073         }
3074         nla_nest_end(skb, list);
3075
3076         if (rule->udata) {
3077                 struct nft_userdata *udata = nft_userdata(rule);
3078                 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
3079                             udata->data) < 0)
3080                         goto nla_put_failure;
3081         }
3082
3083         nlmsg_end(skb, nlh);
3084         return 0;
3085
3086 nla_put_failure:
3087         nlmsg_trim(skb, nlh);
3088         return -1;
3089 }
3090
3091 static void nf_tables_rule_notify(const struct nft_ctx *ctx,
3092                                   const struct nft_rule *rule, int event)
3093 {
3094         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3095         const struct nft_rule *prule;
3096         struct sk_buff *skb;
3097         u64 handle = 0;
3098         u16 flags = 0;
3099         int err;
3100
3101         if (!ctx->report &&
3102             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3103                 return;
3104
3105         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
3106         if (skb == NULL)
3107                 goto err;
3108
3109         if (event == NFT_MSG_NEWRULE &&
3110             !list_is_first(&rule->list, &ctx->chain->rules) &&
3111             !list_is_last(&rule->list, &ctx->chain->rules)) {
3112                 prule = list_prev_entry(rule, list);
3113                 handle = prule->handle;
3114         }
3115         if (ctx->flags & (NLM_F_APPEND | NLM_F_REPLACE))
3116                 flags |= NLM_F_APPEND;
3117         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
3118                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
3119
3120         err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
3121                                        event, flags, ctx->family, ctx->table,
3122                                        ctx->chain, rule, handle, false);
3123         if (err < 0) {
3124                 kfree_skb(skb);
3125                 goto err;
3126         }
3127
3128         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
3129         return;
3130 err:
3131         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
3132 }
3133
3134 struct nft_rule_dump_ctx {
3135         char *table;
3136         char *chain;
3137 };
3138
3139 static int __nf_tables_dump_rules(struct sk_buff *skb,
3140                                   unsigned int *idx,
3141                                   struct netlink_callback *cb,
3142                                   const struct nft_table *table,
3143                                   const struct nft_chain *chain,
3144                                   bool reset)
3145 {
3146         struct net *net = sock_net(skb->sk);
3147         const struct nft_rule *rule, *prule;
3148         unsigned int s_idx = cb->args[0];
3149         u64 handle;
3150
3151         prule = NULL;
3152         list_for_each_entry_rcu(rule, &chain->rules, list) {
3153                 if (!nft_is_active(net, rule))
3154                         goto cont_skip;
3155                 if (*idx < s_idx)
3156                         goto cont;
3157                 if (*idx > s_idx) {
3158                         memset(&cb->args[1], 0,
3159                                         sizeof(cb->args) - sizeof(cb->args[0]));
3160                 }
3161                 if (prule)
3162                         handle = prule->handle;
3163                 else
3164                         handle = 0;
3165
3166                 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
3167                                         cb->nlh->nlmsg_seq,
3168                                         NFT_MSG_NEWRULE,
3169                                         NLM_F_MULTI | NLM_F_APPEND,
3170                                         table->family,
3171                                         table, chain, rule, handle, reset) < 0)
3172                         return 1;
3173
3174                 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
3175 cont:
3176                 prule = rule;
3177 cont_skip:
3178                 (*idx)++;
3179         }
3180         return 0;
3181 }
3182
3183 static int nf_tables_dump_rules(struct sk_buff *skb,
3184                                 struct netlink_callback *cb)
3185 {
3186         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3187         const struct nft_rule_dump_ctx *ctx = cb->data;
3188         struct nft_table *table;
3189         const struct nft_chain *chain;
3190         unsigned int idx = 0;
3191         struct net *net = sock_net(skb->sk);
3192         int family = nfmsg->nfgen_family;
3193         struct nftables_pernet *nft_net;
3194         bool reset = false;
3195
3196         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETRULE_RESET)
3197                 reset = true;
3198
3199         rcu_read_lock();
3200         nft_net = nft_pernet(net);
3201         cb->seq = READ_ONCE(nft_net->base_seq);
3202
3203         list_for_each_entry_rcu(table, &nft_net->tables, list) {
3204                 if (family != NFPROTO_UNSPEC && family != table->family)
3205                         continue;
3206
3207                 if (ctx && ctx->table && strcmp(ctx->table, table->name) != 0)
3208                         continue;
3209
3210                 if (ctx && ctx->table && ctx->chain) {
3211                         struct rhlist_head *list, *tmp;
3212
3213                         list = rhltable_lookup(&table->chains_ht, ctx->chain,
3214                                                nft_chain_ht_params);
3215                         if (!list)
3216                                 goto done;
3217
3218                         rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
3219                                 if (!nft_is_active(net, chain))
3220                                         continue;
3221                                 __nf_tables_dump_rules(skb, &idx,
3222                                                        cb, table, chain, reset);
3223                                 break;
3224                         }
3225                         goto done;
3226                 }
3227
3228                 list_for_each_entry_rcu(chain, &table->chains, list) {
3229                         if (__nf_tables_dump_rules(skb, &idx,
3230                                                    cb, table, chain, reset))
3231                                 goto done;
3232                 }
3233
3234                 if (ctx && ctx->table)
3235                         break;
3236         }
3237 done:
3238         rcu_read_unlock();
3239
3240         cb->args[0] = idx;
3241         return skb->len;
3242 }
3243
3244 static int nf_tables_dump_rules_start(struct netlink_callback *cb)
3245 {
3246         const struct nlattr * const *nla = cb->data;
3247         struct nft_rule_dump_ctx *ctx = NULL;
3248
3249         if (nla[NFTA_RULE_TABLE] || nla[NFTA_RULE_CHAIN]) {
3250                 ctx = kzalloc(sizeof(*ctx), GFP_ATOMIC);
3251                 if (!ctx)
3252                         return -ENOMEM;
3253
3254                 if (nla[NFTA_RULE_TABLE]) {
3255                         ctx->table = nla_strdup(nla[NFTA_RULE_TABLE],
3256                                                         GFP_ATOMIC);
3257                         if (!ctx->table) {
3258                                 kfree(ctx);
3259                                 return -ENOMEM;
3260                         }
3261                 }
3262                 if (nla[NFTA_RULE_CHAIN]) {
3263                         ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN],
3264                                                 GFP_ATOMIC);
3265                         if (!ctx->chain) {
3266                                 kfree(ctx->table);
3267                                 kfree(ctx);
3268                                 return -ENOMEM;
3269                         }
3270                 }
3271         }
3272
3273         cb->data = ctx;
3274         return 0;
3275 }
3276
3277 static int nf_tables_dump_rules_done(struct netlink_callback *cb)
3278 {
3279         struct nft_rule_dump_ctx *ctx = cb->data;
3280
3281         if (ctx) {
3282                 kfree(ctx->table);
3283                 kfree(ctx->chain);
3284                 kfree(ctx);
3285         }
3286         return 0;
3287 }
3288
3289 /* called with rcu_read_lock held */
3290 static int nf_tables_getrule(struct sk_buff *skb, const struct nfnl_info *info,
3291                              const struct nlattr * const nla[])
3292 {
3293         struct netlink_ext_ack *extack = info->extack;
3294         u8 genmask = nft_genmask_cur(info->net);
3295         u8 family = info->nfmsg->nfgen_family;
3296         const struct nft_chain *chain;
3297         const struct nft_rule *rule;
3298         struct net *net = info->net;
3299         struct nft_table *table;
3300         struct sk_buff *skb2;
3301         bool reset = false;
3302         int err;
3303
3304         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3305                 struct netlink_dump_control c = {
3306                         .start= nf_tables_dump_rules_start,
3307                         .dump = nf_tables_dump_rules,
3308                         .done = nf_tables_dump_rules_done,
3309                         .module = THIS_MODULE,
3310                         .data = (void *)nla,
3311                 };
3312
3313                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
3314         }
3315
3316         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask, 0);
3317         if (IS_ERR(table)) {
3318                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3319                 return PTR_ERR(table);
3320         }
3321
3322         chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
3323         if (IS_ERR(chain)) {
3324                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3325                 return PTR_ERR(chain);
3326         }
3327
3328         rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3329         if (IS_ERR(rule)) {
3330                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3331                 return PTR_ERR(rule);
3332         }
3333
3334         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3335         if (!skb2)
3336                 return -ENOMEM;
3337
3338         if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_GETRULE_RESET)
3339                 reset = true;
3340
3341         err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
3342                                        info->nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
3343                                        family, table, chain, rule, 0, reset);
3344         if (err < 0)
3345                 goto err_fill_rule_info;
3346
3347         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
3348
3349 err_fill_rule_info:
3350         kfree_skb(skb2);
3351         return err;
3352 }
3353
3354 static void nf_tables_rule_destroy(const struct nft_ctx *ctx,
3355                                    struct nft_rule *rule)
3356 {
3357         struct nft_expr *expr, *next;
3358
3359         /*
3360          * Careful: some expressions might not be initialized in case this
3361          * is called on error from nf_tables_newrule().
3362          */
3363         expr = nft_expr_first(rule);
3364         while (nft_expr_more(rule, expr)) {
3365                 next = nft_expr_next(expr);
3366                 nf_tables_expr_destroy(ctx, expr);
3367                 expr = next;
3368         }
3369         kfree(rule);
3370 }
3371
3372 void nf_tables_rule_release(const struct nft_ctx *ctx, struct nft_rule *rule)
3373 {
3374         nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_RELEASE);
3375         nf_tables_rule_destroy(ctx, rule);
3376 }
3377
3378 int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
3379 {
3380         struct nft_expr *expr, *last;
3381         const struct nft_data *data;
3382         struct nft_rule *rule;
3383         int err;
3384
3385         if (ctx->level == NFT_JUMP_STACK_SIZE)
3386                 return -EMLINK;
3387
3388         list_for_each_entry(rule, &chain->rules, list) {
3389                 if (!nft_is_active_next(ctx->net, rule))
3390                         continue;
3391
3392                 nft_rule_for_each_expr(expr, last, rule) {
3393                         if (!expr->ops->validate)
3394                                 continue;
3395
3396                         err = expr->ops->validate(ctx, expr, &data);
3397                         if (err < 0)
3398                                 return err;
3399                 }
3400
3401                 cond_resched();
3402         }
3403
3404         return 0;
3405 }
3406 EXPORT_SYMBOL_GPL(nft_chain_validate);
3407
3408 static int nft_table_validate(struct net *net, const struct nft_table *table)
3409 {
3410         struct nft_chain *chain;
3411         struct nft_ctx ctx = {
3412                 .net    = net,
3413                 .family = table->family,
3414         };
3415         int err;
3416
3417         list_for_each_entry(chain, &table->chains, list) {
3418                 if (!nft_is_base_chain(chain))
3419                         continue;
3420
3421                 ctx.chain = chain;
3422                 err = nft_chain_validate(&ctx, chain);
3423                 if (err < 0)
3424                         return err;
3425         }
3426
3427         return 0;
3428 }
3429
3430 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3431                                              const struct nft_chain *chain,
3432                                              const struct nlattr *nla);
3433
3434 #define NFT_RULE_MAXEXPRS       128
3435
3436 static int nf_tables_newrule(struct sk_buff *skb, const struct nfnl_info *info,
3437                              const struct nlattr * const nla[])
3438 {
3439         struct nftables_pernet *nft_net = nft_pernet(info->net);
3440         struct netlink_ext_ack *extack = info->extack;
3441         unsigned int size, i, n, ulen = 0, usize = 0;
3442         u8 genmask = nft_genmask_next(info->net);
3443         struct nft_rule *rule, *old_rule = NULL;
3444         struct nft_expr_info *expr_info = NULL;
3445         u8 family = info->nfmsg->nfgen_family;
3446         struct nft_flow_rule *flow = NULL;
3447         struct net *net = info->net;
3448         struct nft_userdata *udata;
3449         struct nft_table *table;
3450         struct nft_chain *chain;
3451         struct nft_trans *trans;
3452         u64 handle, pos_handle;
3453         struct nft_expr *expr;
3454         struct nft_ctx ctx;
3455         struct nlattr *tmp;
3456         int err, rem;
3457
3458         lockdep_assert_held(&nft_net->commit_mutex);
3459
3460         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3461                                  NETLINK_CB(skb).portid);
3462         if (IS_ERR(table)) {
3463                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3464                 return PTR_ERR(table);
3465         }
3466
3467         if (nla[NFTA_RULE_CHAIN]) {
3468                 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3469                                          genmask);
3470                 if (IS_ERR(chain)) {
3471                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3472                         return PTR_ERR(chain);
3473                 }
3474                 if (nft_chain_is_bound(chain))
3475                         return -EOPNOTSUPP;
3476
3477         } else if (nla[NFTA_RULE_CHAIN_ID]) {
3478                 chain = nft_chain_lookup_byid(net, table, nla[NFTA_RULE_CHAIN_ID]);
3479                 if (IS_ERR(chain)) {
3480                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN_ID]);
3481                         return PTR_ERR(chain);
3482                 }
3483         } else {
3484                 return -EINVAL;
3485         }
3486
3487         if (nla[NFTA_RULE_HANDLE]) {
3488                 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
3489                 rule = __nft_rule_lookup(chain, handle);
3490                 if (IS_ERR(rule)) {
3491                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3492                         return PTR_ERR(rule);
3493                 }
3494
3495                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
3496                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3497                         return -EEXIST;
3498                 }
3499                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
3500                         old_rule = rule;
3501                 else
3502                         return -EOPNOTSUPP;
3503         } else {
3504                 if (!(info->nlh->nlmsg_flags & NLM_F_CREATE) ||
3505                     info->nlh->nlmsg_flags & NLM_F_REPLACE)
3506                         return -EINVAL;
3507                 handle = nf_tables_alloc_handle(table);
3508
3509                 if (chain->use == UINT_MAX)
3510                         return -EOVERFLOW;
3511
3512                 if (nla[NFTA_RULE_POSITION]) {
3513                         pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
3514                         old_rule = __nft_rule_lookup(chain, pos_handle);
3515                         if (IS_ERR(old_rule)) {
3516                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
3517                                 return PTR_ERR(old_rule);
3518                         }
3519                 } else if (nla[NFTA_RULE_POSITION_ID]) {
3520                         old_rule = nft_rule_lookup_byid(net, chain, nla[NFTA_RULE_POSITION_ID]);
3521                         if (IS_ERR(old_rule)) {
3522                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION_ID]);
3523                                 return PTR_ERR(old_rule);
3524                         }
3525                 }
3526         }
3527
3528         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3529
3530         n = 0;
3531         size = 0;
3532         if (nla[NFTA_RULE_EXPRESSIONS]) {
3533                 expr_info = kvmalloc_array(NFT_RULE_MAXEXPRS,
3534                                            sizeof(struct nft_expr_info),
3535                                            GFP_KERNEL);
3536                 if (!expr_info)
3537                         return -ENOMEM;
3538
3539                 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
3540                         err = -EINVAL;
3541                         if (nla_type(tmp) != NFTA_LIST_ELEM)
3542                                 goto err_release_expr;
3543                         if (n == NFT_RULE_MAXEXPRS)
3544                                 goto err_release_expr;
3545                         err = nf_tables_expr_parse(&ctx, tmp, &expr_info[n]);
3546                         if (err < 0) {
3547                                 NL_SET_BAD_ATTR(extack, tmp);
3548                                 goto err_release_expr;
3549                         }
3550                         size += expr_info[n].ops->size;
3551                         n++;
3552                 }
3553         }
3554         /* Check for overflow of dlen field */
3555         err = -EFBIG;
3556         if (size >= 1 << 12)
3557                 goto err_release_expr;
3558
3559         if (nla[NFTA_RULE_USERDATA]) {
3560                 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
3561                 if (ulen > 0)
3562                         usize = sizeof(struct nft_userdata) + ulen;
3563         }
3564
3565         err = -ENOMEM;
3566         rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL_ACCOUNT);
3567         if (rule == NULL)
3568                 goto err_release_expr;
3569
3570         nft_activate_next(net, rule);
3571
3572         rule->handle = handle;
3573         rule->dlen   = size;
3574         rule->udata  = ulen ? 1 : 0;
3575
3576         if (ulen) {
3577                 udata = nft_userdata(rule);
3578                 udata->len = ulen - 1;
3579                 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
3580         }
3581
3582         expr = nft_expr_first(rule);
3583         for (i = 0; i < n; i++) {
3584                 err = nf_tables_newexpr(&ctx, &expr_info[i], expr);
3585                 if (err < 0) {
3586                         NL_SET_BAD_ATTR(extack, expr_info[i].attr);
3587                         goto err_release_rule;
3588                 }
3589
3590                 if (expr_info[i].ops->validate)
3591                         nft_validate_state_update(net, NFT_VALIDATE_NEED);
3592
3593                 expr_info[i].ops = NULL;
3594                 expr = nft_expr_next(expr);
3595         }
3596
3597         if (chain->flags & NFT_CHAIN_HW_OFFLOAD) {
3598                 flow = nft_flow_rule_create(net, rule);
3599                 if (IS_ERR(flow)) {
3600                         err = PTR_ERR(flow);
3601                         goto err_release_rule;
3602                 }
3603         }
3604
3605         if (info->nlh->nlmsg_flags & NLM_F_REPLACE) {
3606                 err = nft_delrule(&ctx, old_rule);
3607                 if (err < 0)
3608                         goto err_destroy_flow_rule;
3609
3610                 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3611                 if (trans == NULL) {
3612                         err = -ENOMEM;
3613                         goto err_destroy_flow_rule;
3614                 }
3615                 list_add_tail_rcu(&rule->list, &old_rule->list);
3616         } else {
3617                 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3618                 if (!trans) {
3619                         err = -ENOMEM;
3620                         goto err_destroy_flow_rule;
3621                 }
3622
3623                 if (info->nlh->nlmsg_flags & NLM_F_APPEND) {
3624                         if (old_rule)
3625                                 list_add_rcu(&rule->list, &old_rule->list);
3626                         else
3627                                 list_add_tail_rcu(&rule->list, &chain->rules);
3628                  } else {
3629                         if (old_rule)
3630                                 list_add_tail_rcu(&rule->list, &old_rule->list);
3631                         else
3632                                 list_add_rcu(&rule->list, &chain->rules);
3633                 }
3634         }
3635         kvfree(expr_info);
3636         chain->use++;
3637
3638         if (flow)
3639                 nft_trans_flow_rule(trans) = flow;
3640
3641         if (nft_net->validate_state == NFT_VALIDATE_DO)
3642                 return nft_table_validate(net, table);
3643
3644         return 0;
3645
3646 err_destroy_flow_rule:
3647         if (flow)
3648                 nft_flow_rule_destroy(flow);
3649 err_release_rule:
3650         nf_tables_rule_release(&ctx, rule);
3651 err_release_expr:
3652         for (i = 0; i < n; i++) {
3653                 if (expr_info[i].ops) {
3654                         module_put(expr_info[i].ops->type->owner);
3655                         if (expr_info[i].ops->type->release_ops)
3656                                 expr_info[i].ops->type->release_ops(expr_info[i].ops);
3657                 }
3658         }
3659         kvfree(expr_info);
3660
3661         return err;
3662 }
3663
3664 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3665                                              const struct nft_chain *chain,
3666                                              const struct nlattr *nla)
3667 {
3668         struct nftables_pernet *nft_net = nft_pernet(net);
3669         u32 id = ntohl(nla_get_be32(nla));
3670         struct nft_trans *trans;
3671
3672         list_for_each_entry(trans, &nft_net->commit_list, list) {
3673                 struct nft_rule *rule = nft_trans_rule(trans);
3674
3675                 if (trans->msg_type == NFT_MSG_NEWRULE &&
3676                     trans->ctx.chain == chain &&
3677                     id == nft_trans_rule_id(trans))
3678                         return rule;
3679         }
3680         return ERR_PTR(-ENOENT);
3681 }
3682
3683 static int nf_tables_delrule(struct sk_buff *skb, const struct nfnl_info *info,
3684                              const struct nlattr * const nla[])
3685 {
3686         struct netlink_ext_ack *extack = info->extack;
3687         u8 genmask = nft_genmask_next(info->net);
3688         u8 family = info->nfmsg->nfgen_family;
3689         struct nft_chain *chain = NULL;
3690         struct net *net = info->net;
3691         struct nft_table *table;
3692         struct nft_rule *rule;
3693         struct nft_ctx ctx;
3694         int err = 0;
3695
3696         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3697                                  NETLINK_CB(skb).portid);
3698         if (IS_ERR(table)) {
3699                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3700                 return PTR_ERR(table);
3701         }
3702
3703         if (nla[NFTA_RULE_CHAIN]) {
3704                 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3705                                          genmask);
3706                 if (IS_ERR(chain)) {
3707                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3708                         return PTR_ERR(chain);
3709                 }
3710                 if (nft_chain_is_bound(chain))
3711                         return -EOPNOTSUPP;
3712         }
3713
3714         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3715
3716         if (chain) {
3717                 if (nla[NFTA_RULE_HANDLE]) {
3718                         rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3719                         if (IS_ERR(rule)) {
3720                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3721                                 return PTR_ERR(rule);
3722                         }
3723
3724                         err = nft_delrule(&ctx, rule);
3725                 } else if (nla[NFTA_RULE_ID]) {
3726                         rule = nft_rule_lookup_byid(net, chain, nla[NFTA_RULE_ID]);
3727                         if (IS_ERR(rule)) {
3728                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
3729                                 return PTR_ERR(rule);
3730                         }
3731
3732                         err = nft_delrule(&ctx, rule);
3733                 } else {
3734                         err = nft_delrule_by_chain(&ctx);
3735                 }
3736         } else {
3737                 list_for_each_entry(chain, &table->chains, list) {
3738                         if (!nft_is_active_next(net, chain))
3739                                 continue;
3740
3741                         ctx.chain = chain;
3742                         err = nft_delrule_by_chain(&ctx);
3743                         if (err < 0)
3744                                 break;
3745                 }
3746         }
3747
3748         return err;
3749 }
3750
3751 /*
3752  * Sets
3753  */
3754 static const struct nft_set_type *nft_set_types[] = {
3755         &nft_set_hash_fast_type,
3756         &nft_set_hash_type,
3757         &nft_set_rhash_type,
3758         &nft_set_bitmap_type,
3759         &nft_set_rbtree_type,
3760 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
3761         &nft_set_pipapo_avx2_type,
3762 #endif
3763         &nft_set_pipapo_type,
3764 };
3765
3766 #define NFT_SET_FEATURES        (NFT_SET_INTERVAL | NFT_SET_MAP | \
3767                                  NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
3768                                  NFT_SET_EVAL)
3769
3770 static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
3771 {
3772         return (flags & type->features) == (flags & NFT_SET_FEATURES);
3773 }
3774
3775 /*
3776  * Select a set implementation based on the data characteristics and the
3777  * given policy. The total memory use might not be known if no size is
3778  * given, in that case the amount of memory per element is used.
3779  */
3780 static const struct nft_set_ops *
3781 nft_select_set_ops(const struct nft_ctx *ctx,
3782                    const struct nlattr * const nla[],
3783                    const struct nft_set_desc *desc,
3784                    enum nft_set_policies policy)
3785 {
3786         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3787         const struct nft_set_ops *ops, *bops;
3788         struct nft_set_estimate est, best;
3789         const struct nft_set_type *type;
3790         u32 flags = 0;
3791         int i;
3792
3793         lockdep_assert_held(&nft_net->commit_mutex);
3794         lockdep_nfnl_nft_mutex_not_held();
3795
3796         if (nla[NFTA_SET_FLAGS] != NULL)
3797                 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
3798
3799         bops        = NULL;
3800         best.size   = ~0;
3801         best.lookup = ~0;
3802         best.space  = ~0;
3803
3804         for (i = 0; i < ARRAY_SIZE(nft_set_types); i++) {
3805                 type = nft_set_types[i];
3806                 ops = &type->ops;
3807
3808                 if (!nft_set_ops_candidate(type, flags))
3809                         continue;
3810                 if (!ops->estimate(desc, flags, &est))
3811                         continue;
3812
3813                 switch (policy) {
3814                 case NFT_SET_POL_PERFORMANCE:
3815                         if (est.lookup < best.lookup)
3816                                 break;
3817                         if (est.lookup == best.lookup &&
3818                             est.space < best.space)
3819                                 break;
3820                         continue;
3821                 case NFT_SET_POL_MEMORY:
3822                         if (!desc->size) {
3823                                 if (est.space < best.space)
3824                                         break;
3825                                 if (est.space == best.space &&
3826                                     est.lookup < best.lookup)
3827                                         break;
3828                         } else if (est.size < best.size || !bops) {
3829                                 break;
3830                         }
3831                         continue;
3832                 default:
3833                         break;
3834                 }
3835
3836                 bops = ops;
3837                 best = est;
3838         }
3839
3840         if (bops != NULL)
3841                 return bops;
3842
3843         return ERR_PTR(-EOPNOTSUPP);
3844 }
3845
3846 static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
3847         [NFTA_SET_TABLE]                = { .type = NLA_STRING,
3848                                             .len = NFT_TABLE_MAXNAMELEN - 1 },
3849         [NFTA_SET_NAME]                 = { .type = NLA_STRING,
3850                                             .len = NFT_SET_MAXNAMELEN - 1 },
3851         [NFTA_SET_FLAGS]                = { .type = NLA_U32 },
3852         [NFTA_SET_KEY_TYPE]             = { .type = NLA_U32 },
3853         [NFTA_SET_KEY_LEN]              = { .type = NLA_U32 },
3854         [NFTA_SET_DATA_TYPE]            = { .type = NLA_U32 },
3855         [NFTA_SET_DATA_LEN]             = { .type = NLA_U32 },
3856         [NFTA_SET_POLICY]               = { .type = NLA_U32 },
3857         [NFTA_SET_DESC]                 = { .type = NLA_NESTED },
3858         [NFTA_SET_ID]                   = { .type = NLA_U32 },
3859         [NFTA_SET_TIMEOUT]              = { .type = NLA_U64 },
3860         [NFTA_SET_GC_INTERVAL]          = { .type = NLA_U32 },
3861         [NFTA_SET_USERDATA]             = { .type = NLA_BINARY,
3862                                             .len  = NFT_USERDATA_MAXLEN },
3863         [NFTA_SET_OBJ_TYPE]             = { .type = NLA_U32 },
3864         [NFTA_SET_HANDLE]               = { .type = NLA_U64 },
3865         [NFTA_SET_EXPR]                 = { .type = NLA_NESTED },
3866         [NFTA_SET_EXPRESSIONS]          = { .type = NLA_NESTED },
3867 };
3868
3869 static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
3870         [NFTA_SET_DESC_SIZE]            = { .type = NLA_U32 },
3871         [NFTA_SET_DESC_CONCAT]          = { .type = NLA_NESTED },
3872 };
3873
3874 static struct nft_set *nft_set_lookup(const struct nft_table *table,
3875                                       const struct nlattr *nla, u8 genmask)
3876 {
3877         struct nft_set *set;
3878
3879         if (nla == NULL)
3880                 return ERR_PTR(-EINVAL);
3881
3882         list_for_each_entry_rcu(set, &table->sets, list) {
3883                 if (!nla_strcmp(nla, set->name) &&
3884                     nft_active_genmask(set, genmask))
3885                         return set;
3886         }
3887         return ERR_PTR(-ENOENT);
3888 }
3889
3890 static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
3891                                                const struct nlattr *nla,
3892                                                u8 genmask)
3893 {
3894         struct nft_set *set;
3895
3896         list_for_each_entry(set, &table->sets, list) {
3897                 if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
3898                     nft_active_genmask(set, genmask))
3899                         return set;
3900         }
3901         return ERR_PTR(-ENOENT);
3902 }
3903
3904 static struct nft_set *nft_set_lookup_byid(const struct net *net,
3905                                            const struct nft_table *table,
3906                                            const struct nlattr *nla, u8 genmask)
3907 {
3908         struct nftables_pernet *nft_net = nft_pernet(net);
3909         u32 id = ntohl(nla_get_be32(nla));
3910         struct nft_trans *trans;
3911
3912         list_for_each_entry(trans, &nft_net->commit_list, list) {
3913                 if (trans->msg_type == NFT_MSG_NEWSET) {
3914                         struct nft_set *set = nft_trans_set(trans);
3915
3916                         if (id == nft_trans_set_id(trans) &&
3917                             set->table == table &&
3918                             nft_active_genmask(set, genmask))
3919                                 return set;
3920                 }
3921         }
3922         return ERR_PTR(-ENOENT);
3923 }
3924
3925 struct nft_set *nft_set_lookup_global(const struct net *net,
3926                                       const struct nft_table *table,
3927                                       const struct nlattr *nla_set_name,
3928                                       const struct nlattr *nla_set_id,
3929                                       u8 genmask)
3930 {
3931         struct nft_set *set;
3932
3933         set = nft_set_lookup(table, nla_set_name, genmask);
3934         if (IS_ERR(set)) {
3935                 if (!nla_set_id)
3936                         return set;
3937
3938                 set = nft_set_lookup_byid(net, table, nla_set_id, genmask);
3939         }
3940         return set;
3941 }
3942 EXPORT_SYMBOL_GPL(nft_set_lookup_global);
3943
3944 static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
3945                                     const char *name)
3946 {
3947         const struct nft_set *i;
3948         const char *p;
3949         unsigned long *inuse;
3950         unsigned int n = 0, min = 0;
3951
3952         p = strchr(name, '%');
3953         if (p != NULL) {
3954                 if (p[1] != 'd' || strchr(p + 2, '%'))
3955                         return -EINVAL;
3956
3957                 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
3958                 if (inuse == NULL)
3959                         return -ENOMEM;
3960 cont:
3961                 list_for_each_entry(i, &ctx->table->sets, list) {
3962                         int tmp;
3963
3964                         if (!nft_is_active_next(ctx->net, i))
3965                                 continue;
3966                         if (!sscanf(i->name, name, &tmp))
3967                                 continue;
3968                         if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
3969                                 continue;
3970
3971                         set_bit(tmp - min, inuse);
3972                 }
3973
3974                 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
3975                 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
3976                         min += BITS_PER_BYTE * PAGE_SIZE;
3977                         memset(inuse, 0, PAGE_SIZE);
3978                         goto cont;
3979                 }
3980                 free_page((unsigned long)inuse);
3981         }
3982
3983         set->name = kasprintf(GFP_KERNEL_ACCOUNT, name, min + n);
3984         if (!set->name)
3985                 return -ENOMEM;
3986
3987         list_for_each_entry(i, &ctx->table->sets, list) {
3988                 if (!nft_is_active_next(ctx->net, i))
3989                         continue;
3990                 if (!strcmp(set->name, i->name)) {
3991                         kfree(set->name);
3992                         set->name = NULL;
3993                         return -ENFILE;
3994                 }
3995         }
3996         return 0;
3997 }
3998
3999 int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
4000 {
4001         u64 ms = be64_to_cpu(nla_get_be64(nla));
4002         u64 max = (u64)(~((u64)0));
4003
4004         max = div_u64(max, NSEC_PER_MSEC);
4005         if (ms >= max)
4006                 return -ERANGE;
4007
4008         ms *= NSEC_PER_MSEC;
4009         *result = nsecs_to_jiffies64(ms);
4010         return 0;
4011 }
4012
4013 __be64 nf_jiffies64_to_msecs(u64 input)
4014 {
4015         return cpu_to_be64(jiffies64_to_msecs(input));
4016 }
4017
4018 static int nf_tables_fill_set_concat(struct sk_buff *skb,
4019                                      const struct nft_set *set)
4020 {
4021         struct nlattr *concat, *field;
4022         int i;
4023
4024         concat = nla_nest_start_noflag(skb, NFTA_SET_DESC_CONCAT);
4025         if (!concat)
4026                 return -ENOMEM;
4027
4028         for (i = 0; i < set->field_count; i++) {
4029                 field = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4030                 if (!field)
4031                         return -ENOMEM;
4032
4033                 if (nla_put_be32(skb, NFTA_SET_FIELD_LEN,
4034                                  htonl(set->field_len[i])))
4035                         return -ENOMEM;
4036
4037                 nla_nest_end(skb, field);
4038         }
4039
4040         nla_nest_end(skb, concat);
4041
4042         return 0;
4043 }
4044
4045 static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
4046                               const struct nft_set *set, u16 event, u16 flags)
4047 {
4048         struct nlmsghdr *nlh;
4049         u32 portid = ctx->portid;
4050         struct nlattr *nest;
4051         u32 seq = ctx->seq;
4052         int i;
4053
4054         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
4055         nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
4056                            NFNETLINK_V0, nft_base_seq(ctx->net));
4057         if (!nlh)
4058                 goto nla_put_failure;
4059
4060         if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
4061                 goto nla_put_failure;
4062         if (nla_put_string(skb, NFTA_SET_NAME, set->name))
4063                 goto nla_put_failure;
4064         if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
4065                          NFTA_SET_PAD))
4066                 goto nla_put_failure;
4067         if (set->flags != 0)
4068                 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
4069                         goto nla_put_failure;
4070
4071         if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
4072                 goto nla_put_failure;
4073         if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
4074                 goto nla_put_failure;
4075         if (set->flags & NFT_SET_MAP) {
4076                 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
4077                         goto nla_put_failure;
4078                 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
4079                         goto nla_put_failure;
4080         }
4081         if (set->flags & NFT_SET_OBJECT &&
4082             nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
4083                 goto nla_put_failure;
4084
4085         if (set->timeout &&
4086             nla_put_be64(skb, NFTA_SET_TIMEOUT,
4087                          nf_jiffies64_to_msecs(set->timeout),
4088                          NFTA_SET_PAD))
4089                 goto nla_put_failure;
4090         if (set->gc_int &&
4091             nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(set->gc_int)))
4092                 goto nla_put_failure;
4093
4094         if (set->policy != NFT_SET_POL_PERFORMANCE) {
4095                 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
4096                         goto nla_put_failure;
4097         }
4098
4099         if (set->udata &&
4100             nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
4101                 goto nla_put_failure;
4102
4103         nest = nla_nest_start_noflag(skb, NFTA_SET_DESC);
4104         if (!nest)
4105                 goto nla_put_failure;
4106         if (set->size &&
4107             nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
4108                 goto nla_put_failure;
4109
4110         if (set->field_count > 1 &&
4111             nf_tables_fill_set_concat(skb, set))
4112                 goto nla_put_failure;
4113
4114         nla_nest_end(skb, nest);
4115
4116         if (set->num_exprs == 1) {
4117                 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPR);
4118                 if (nf_tables_fill_expr_info(skb, set->exprs[0], false) < 0)
4119                         goto nla_put_failure;
4120
4121                 nla_nest_end(skb, nest);
4122         } else if (set->num_exprs > 1) {
4123                 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPRESSIONS);
4124                 if (nest == NULL)
4125                         goto nla_put_failure;
4126
4127                 for (i = 0; i < set->num_exprs; i++) {
4128                         if (nft_expr_dump(skb, NFTA_LIST_ELEM,
4129                                           set->exprs[i], false) < 0)
4130                                 goto nla_put_failure;
4131                 }
4132                 nla_nest_end(skb, nest);
4133         }
4134
4135         nlmsg_end(skb, nlh);
4136         return 0;
4137
4138 nla_put_failure:
4139         nlmsg_trim(skb, nlh);
4140         return -1;
4141 }
4142
4143 static void nf_tables_set_notify(const struct nft_ctx *ctx,
4144                                  const struct nft_set *set, int event,
4145                                  gfp_t gfp_flags)
4146 {
4147         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
4148         u32 portid = ctx->portid;
4149         struct sk_buff *skb;
4150         u16 flags = 0;
4151         int err;
4152
4153         if (!ctx->report &&
4154             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
4155                 return;
4156
4157         skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
4158         if (skb == NULL)
4159                 goto err;
4160
4161         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
4162                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
4163
4164         err = nf_tables_fill_set(skb, ctx, set, event, flags);
4165         if (err < 0) {
4166                 kfree_skb(skb);
4167                 goto err;
4168         }
4169
4170         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
4171         return;
4172 err:
4173         nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
4174 }
4175
4176 static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
4177 {
4178         const struct nft_set *set;
4179         unsigned int idx, s_idx = cb->args[0];
4180         struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
4181         struct net *net = sock_net(skb->sk);
4182         struct nft_ctx *ctx = cb->data, ctx_set;
4183         struct nftables_pernet *nft_net;
4184
4185         if (cb->args[1])
4186                 return skb->len;
4187
4188         rcu_read_lock();
4189         nft_net = nft_pernet(net);
4190         cb->seq = READ_ONCE(nft_net->base_seq);
4191
4192         list_for_each_entry_rcu(table, &nft_net->tables, list) {
4193                 if (ctx->family != NFPROTO_UNSPEC &&
4194                     ctx->family != table->family)
4195                         continue;
4196
4197                 if (ctx->table && ctx->table != table)
4198                         continue;
4199
4200                 if (cur_table) {
4201                         if (cur_table != table)
4202                                 continue;
4203
4204                         cur_table = NULL;
4205                 }
4206                 idx = 0;
4207                 list_for_each_entry_rcu(set, &table->sets, list) {
4208                         if (idx < s_idx)
4209                                 goto cont;
4210                         if (!nft_is_active(net, set))
4211                                 goto cont;
4212
4213                         ctx_set = *ctx;
4214                         ctx_set.table = table;
4215                         ctx_set.family = table->family;
4216
4217                         if (nf_tables_fill_set(skb, &ctx_set, set,
4218                                                NFT_MSG_NEWSET,
4219                                                NLM_F_MULTI) < 0) {
4220                                 cb->args[0] = idx;
4221                                 cb->args[2] = (unsigned long) table;
4222                                 goto done;
4223                         }
4224                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
4225 cont:
4226                         idx++;
4227                 }
4228                 if (s_idx)
4229                         s_idx = 0;
4230         }
4231         cb->args[1] = 1;
4232 done:
4233         rcu_read_unlock();
4234         return skb->len;
4235 }
4236
4237 static int nf_tables_dump_sets_start(struct netlink_callback *cb)
4238 {
4239         struct nft_ctx *ctx_dump = NULL;
4240
4241         ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
4242         if (ctx_dump == NULL)
4243                 return -ENOMEM;
4244
4245         cb->data = ctx_dump;
4246         return 0;
4247 }
4248
4249 static int nf_tables_dump_sets_done(struct netlink_callback *cb)
4250 {
4251         kfree(cb->data);
4252         return 0;
4253 }
4254
4255 /* called with rcu_read_lock held */
4256 static int nf_tables_getset(struct sk_buff *skb, const struct nfnl_info *info,
4257                             const struct nlattr * const nla[])
4258 {
4259         struct netlink_ext_ack *extack = info->extack;
4260         u8 genmask = nft_genmask_cur(info->net);
4261         u8 family = info->nfmsg->nfgen_family;
4262         struct nft_table *table = NULL;
4263         struct net *net = info->net;
4264         const struct nft_set *set;
4265         struct sk_buff *skb2;
4266         struct nft_ctx ctx;
4267         int err;
4268
4269         if (nla[NFTA_SET_TABLE]) {
4270                 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4271                                          genmask, 0);
4272                 if (IS_ERR(table)) {
4273                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4274                         return PTR_ERR(table);
4275                 }
4276         }
4277
4278         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4279
4280         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
4281                 struct netlink_dump_control c = {
4282                         .start = nf_tables_dump_sets_start,
4283                         .dump = nf_tables_dump_sets,
4284                         .done = nf_tables_dump_sets_done,
4285                         .data = &ctx,
4286                         .module = THIS_MODULE,
4287                 };
4288
4289                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
4290         }
4291
4292         /* Only accept unspec with dump */
4293         if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4294                 return -EAFNOSUPPORT;
4295         if (!nla[NFTA_SET_TABLE])
4296                 return -EINVAL;
4297
4298         set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4299         if (IS_ERR(set))
4300                 return PTR_ERR(set);
4301
4302         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
4303         if (skb2 == NULL)
4304                 return -ENOMEM;
4305
4306         err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
4307         if (err < 0)
4308                 goto err_fill_set_info;
4309
4310         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
4311
4312 err_fill_set_info:
4313         kfree_skb(skb2);
4314         return err;
4315 }
4316
4317 static const struct nla_policy nft_concat_policy[NFTA_SET_FIELD_MAX + 1] = {
4318         [NFTA_SET_FIELD_LEN]    = { .type = NLA_U32 },
4319 };
4320
4321 static int nft_set_desc_concat_parse(const struct nlattr *attr,
4322                                      struct nft_set_desc *desc)
4323 {
4324         struct nlattr *tb[NFTA_SET_FIELD_MAX + 1];
4325         u32 len;
4326         int err;
4327
4328         if (desc->field_count >= ARRAY_SIZE(desc->field_len))
4329                 return -E2BIG;
4330
4331         err = nla_parse_nested_deprecated(tb, NFTA_SET_FIELD_MAX, attr,
4332                                           nft_concat_policy, NULL);
4333         if (err < 0)
4334                 return err;
4335
4336         if (!tb[NFTA_SET_FIELD_LEN])
4337                 return -EINVAL;
4338
4339         len = ntohl(nla_get_be32(tb[NFTA_SET_FIELD_LEN]));
4340         if (!len || len > U8_MAX)
4341                 return -EINVAL;
4342
4343         desc->field_len[desc->field_count++] = len;
4344
4345         return 0;
4346 }
4347
4348 static int nft_set_desc_concat(struct nft_set_desc *desc,
4349                                const struct nlattr *nla)
4350 {
4351         struct nlattr *attr;
4352         u32 num_regs = 0;
4353         int rem, err, i;
4354
4355         nla_for_each_nested(attr, nla, rem) {
4356                 if (nla_type(attr) != NFTA_LIST_ELEM)
4357                         return -EINVAL;
4358
4359                 err = nft_set_desc_concat_parse(attr, desc);
4360                 if (err < 0)
4361                         return err;
4362         }
4363
4364         for (i = 0; i < desc->field_count; i++)
4365                 num_regs += DIV_ROUND_UP(desc->field_len[i], sizeof(u32));
4366
4367         if (num_regs > NFT_REG32_COUNT)
4368                 return -E2BIG;
4369
4370         return 0;
4371 }
4372
4373 static int nf_tables_set_desc_parse(struct nft_set_desc *desc,
4374                                     const struct nlattr *nla)
4375 {
4376         struct nlattr *da[NFTA_SET_DESC_MAX + 1];
4377         int err;
4378
4379         err = nla_parse_nested_deprecated(da, NFTA_SET_DESC_MAX, nla,
4380                                           nft_set_desc_policy, NULL);
4381         if (err < 0)
4382                 return err;
4383
4384         if (da[NFTA_SET_DESC_SIZE] != NULL)
4385                 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
4386         if (da[NFTA_SET_DESC_CONCAT])
4387                 err = nft_set_desc_concat(desc, da[NFTA_SET_DESC_CONCAT]);
4388
4389         return err;
4390 }
4391
4392 static int nf_tables_newset(struct sk_buff *skb, const struct nfnl_info *info,
4393                             const struct nlattr * const nla[])
4394 {
4395         u32 ktype, dtype, flags, policy, gc_int, objtype;
4396         struct netlink_ext_ack *extack = info->extack;
4397         u8 genmask = nft_genmask_next(info->net);
4398         u8 family = info->nfmsg->nfgen_family;
4399         const struct nft_set_ops *ops;
4400         struct nft_expr *expr = NULL;
4401         struct net *net = info->net;
4402         struct nft_set_desc desc;
4403         struct nft_table *table;
4404         unsigned char *udata;
4405         struct nft_set *set;
4406         struct nft_ctx ctx;
4407         size_t alloc_size;
4408         u64 timeout;
4409         char *name;
4410         int err, i;
4411         u16 udlen;
4412         u64 size;
4413
4414         if (nla[NFTA_SET_TABLE] == NULL ||
4415             nla[NFTA_SET_NAME] == NULL ||
4416             nla[NFTA_SET_KEY_LEN] == NULL ||
4417             nla[NFTA_SET_ID] == NULL)
4418                 return -EINVAL;
4419
4420         memset(&desc, 0, sizeof(desc));
4421
4422         ktype = NFT_DATA_VALUE;
4423         if (nla[NFTA_SET_KEY_TYPE] != NULL) {
4424                 ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
4425                 if ((ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
4426                         return -EINVAL;
4427         }
4428
4429         desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
4430         if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
4431                 return -EINVAL;
4432
4433         flags = 0;
4434         if (nla[NFTA_SET_FLAGS] != NULL) {
4435                 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
4436                 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
4437                               NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
4438                               NFT_SET_MAP | NFT_SET_EVAL |
4439                               NFT_SET_OBJECT | NFT_SET_CONCAT | NFT_SET_EXPR))
4440                         return -EOPNOTSUPP;
4441                 /* Only one of these operations is supported */
4442                 if ((flags & (NFT_SET_MAP | NFT_SET_OBJECT)) ==
4443                              (NFT_SET_MAP | NFT_SET_OBJECT))
4444                         return -EOPNOTSUPP;
4445                 if ((flags & (NFT_SET_EVAL | NFT_SET_OBJECT)) ==
4446                              (NFT_SET_EVAL | NFT_SET_OBJECT))
4447                         return -EOPNOTSUPP;
4448         }
4449
4450         dtype = 0;
4451         if (nla[NFTA_SET_DATA_TYPE] != NULL) {
4452                 if (!(flags & NFT_SET_MAP))
4453                         return -EINVAL;
4454
4455                 dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
4456                 if ((dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
4457                     dtype != NFT_DATA_VERDICT)
4458                         return -EINVAL;
4459
4460                 if (dtype != NFT_DATA_VERDICT) {
4461                         if (nla[NFTA_SET_DATA_LEN] == NULL)
4462                                 return -EINVAL;
4463                         desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
4464                         if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
4465                                 return -EINVAL;
4466                 } else
4467                         desc.dlen = sizeof(struct nft_verdict);
4468         } else if (flags & NFT_SET_MAP)
4469                 return -EINVAL;
4470
4471         if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
4472                 if (!(flags & NFT_SET_OBJECT))
4473                         return -EINVAL;
4474
4475                 objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
4476                 if (objtype == NFT_OBJECT_UNSPEC ||
4477                     objtype > NFT_OBJECT_MAX)
4478                         return -EOPNOTSUPP;
4479         } else if (flags & NFT_SET_OBJECT)
4480                 return -EINVAL;
4481         else
4482                 objtype = NFT_OBJECT_UNSPEC;
4483
4484         timeout = 0;
4485         if (nla[NFTA_SET_TIMEOUT] != NULL) {
4486                 if (!(flags & NFT_SET_TIMEOUT))
4487                         return -EINVAL;
4488
4489                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &timeout);
4490                 if (err)
4491                         return err;
4492         }
4493         gc_int = 0;
4494         if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
4495                 if (!(flags & NFT_SET_TIMEOUT))
4496                         return -EINVAL;
4497                 gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
4498         }
4499
4500         policy = NFT_SET_POL_PERFORMANCE;
4501         if (nla[NFTA_SET_POLICY] != NULL)
4502                 policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
4503
4504         if (nla[NFTA_SET_DESC] != NULL) {
4505                 err = nf_tables_set_desc_parse(&desc, nla[NFTA_SET_DESC]);
4506                 if (err < 0)
4507                         return err;
4508
4509                 if (desc.field_count > 1 && !(flags & NFT_SET_CONCAT))
4510                         return -EINVAL;
4511         } else if (flags & NFT_SET_CONCAT) {
4512                 return -EINVAL;
4513         }
4514
4515         if (nla[NFTA_SET_EXPR] || nla[NFTA_SET_EXPRESSIONS])
4516                 desc.expr = true;
4517
4518         table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask,
4519                                  NETLINK_CB(skb).portid);
4520         if (IS_ERR(table)) {
4521                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4522                 return PTR_ERR(table);
4523         }
4524
4525         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4526
4527         set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4528         if (IS_ERR(set)) {
4529                 if (PTR_ERR(set) != -ENOENT) {
4530                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4531                         return PTR_ERR(set);
4532                 }
4533         } else {
4534                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
4535                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4536                         return -EEXIST;
4537                 }
4538                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
4539                         return -EOPNOTSUPP;
4540
4541                 return 0;
4542         }
4543
4544         if (!(info->nlh->nlmsg_flags & NLM_F_CREATE))
4545                 return -ENOENT;
4546
4547         ops = nft_select_set_ops(&ctx, nla, &desc, policy);
4548         if (IS_ERR(ops))
4549                 return PTR_ERR(ops);
4550
4551         udlen = 0;
4552         if (nla[NFTA_SET_USERDATA])
4553                 udlen = nla_len(nla[NFTA_SET_USERDATA]);
4554
4555         size = 0;
4556         if (ops->privsize != NULL)
4557                 size = ops->privsize(nla, &desc);
4558         alloc_size = sizeof(*set) + size + udlen;
4559         if (alloc_size < size || alloc_size > INT_MAX)
4560                 return -ENOMEM;
4561         set = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT);
4562         if (!set)
4563                 return -ENOMEM;
4564
4565         name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL_ACCOUNT);
4566         if (!name) {
4567                 err = -ENOMEM;
4568                 goto err_set_name;
4569         }
4570
4571         err = nf_tables_set_alloc_name(&ctx, set, name);
4572         kfree(name);
4573         if (err < 0)
4574                 goto err_set_name;
4575
4576         udata = NULL;
4577         if (udlen) {
4578                 udata = set->data + size;
4579                 nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
4580         }
4581
4582         INIT_LIST_HEAD(&set->bindings);
4583         INIT_LIST_HEAD(&set->catchall_list);
4584         set->table = table;
4585         write_pnet(&set->net, net);
4586         set->ops = ops;
4587         set->ktype = ktype;
4588         set->klen = desc.klen;
4589         set->dtype = dtype;
4590         set->objtype = objtype;
4591         set->dlen = desc.dlen;
4592         set->flags = flags;
4593         set->size = desc.size;
4594         set->policy = policy;
4595         set->udlen = udlen;
4596         set->udata = udata;
4597         set->timeout = timeout;
4598         set->gc_int = gc_int;
4599
4600         set->field_count = desc.field_count;
4601         for (i = 0; i < desc.field_count; i++)
4602                 set->field_len[i] = desc.field_len[i];
4603
4604         err = ops->init(set, &desc, nla);
4605         if (err < 0)
4606                 goto err_set_init;
4607
4608         if (nla[NFTA_SET_EXPR]) {
4609                 expr = nft_set_elem_expr_alloc(&ctx, set, nla[NFTA_SET_EXPR]);
4610                 if (IS_ERR(expr)) {
4611                         err = PTR_ERR(expr);
4612                         goto err_set_expr_alloc;
4613                 }
4614                 set->exprs[0] = expr;
4615                 set->num_exprs++;
4616         } else if (nla[NFTA_SET_EXPRESSIONS]) {
4617                 struct nft_expr *expr;
4618                 struct nlattr *tmp;
4619                 int left;
4620
4621                 if (!(flags & NFT_SET_EXPR)) {
4622                         err = -EINVAL;
4623                         goto err_set_expr_alloc;
4624                 }
4625                 i = 0;
4626                 nla_for_each_nested(tmp, nla[NFTA_SET_EXPRESSIONS], left) {
4627                         if (i == NFT_SET_EXPR_MAX) {
4628                                 err = -E2BIG;
4629                                 goto err_set_expr_alloc;
4630                         }
4631                         if (nla_type(tmp) != NFTA_LIST_ELEM) {
4632                                 err = -EINVAL;
4633                                 goto err_set_expr_alloc;
4634                         }
4635                         expr = nft_set_elem_expr_alloc(&ctx, set, tmp);
4636                         if (IS_ERR(expr)) {
4637                                 err = PTR_ERR(expr);
4638                                 goto err_set_expr_alloc;
4639                         }
4640                         set->exprs[i++] = expr;
4641                         set->num_exprs++;
4642                 }
4643         }
4644
4645         set->handle = nf_tables_alloc_handle(table);
4646
4647         err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
4648         if (err < 0)
4649                 goto err_set_expr_alloc;
4650
4651         list_add_tail_rcu(&set->list, &table->sets);
4652         table->use++;
4653         return 0;
4654
4655 err_set_expr_alloc:
4656         for (i = 0; i < set->num_exprs; i++)
4657                 nft_expr_destroy(&ctx, set->exprs[i]);
4658
4659         ops->destroy(set);
4660 err_set_init:
4661         kfree(set->name);
4662 err_set_name:
4663         kvfree(set);
4664         return err;
4665 }
4666
4667 struct nft_set_elem_catchall {
4668         struct list_head        list;
4669         struct rcu_head         rcu;
4670         void                    *elem;
4671 };
4672
4673 static void nft_set_catchall_destroy(const struct nft_ctx *ctx,
4674                                      struct nft_set *set)
4675 {
4676         struct nft_set_elem_catchall *next, *catchall;
4677
4678         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
4679                 list_del_rcu(&catchall->list);
4680                 nft_set_elem_destroy(set, catchall->elem, true);
4681                 kfree_rcu(catchall, rcu);
4682         }
4683 }
4684
4685 static void nft_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
4686 {
4687         int i;
4688
4689         if (WARN_ON(set->use > 0))
4690                 return;
4691
4692         for (i = 0; i < set->num_exprs; i++)
4693                 nft_expr_destroy(ctx, set->exprs[i]);
4694
4695         set->ops->destroy(set);
4696         nft_set_catchall_destroy(ctx, set);
4697         kfree(set->name);
4698         kvfree(set);
4699 }
4700
4701 static int nf_tables_delset(struct sk_buff *skb, const struct nfnl_info *info,
4702                             const struct nlattr * const nla[])
4703 {
4704         struct netlink_ext_ack *extack = info->extack;
4705         u8 genmask = nft_genmask_next(info->net);
4706         u8 family = info->nfmsg->nfgen_family;
4707         struct net *net = info->net;
4708         const struct nlattr *attr;
4709         struct nft_table *table;
4710         struct nft_set *set;
4711         struct nft_ctx ctx;
4712
4713         if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4714                 return -EAFNOSUPPORT;
4715
4716         table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4717                                  genmask, NETLINK_CB(skb).portid);
4718         if (IS_ERR(table)) {
4719                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4720                 return PTR_ERR(table);
4721         }
4722
4723         if (nla[NFTA_SET_HANDLE]) {
4724                 attr = nla[NFTA_SET_HANDLE];
4725                 set = nft_set_lookup_byhandle(table, attr, genmask);
4726         } else {
4727                 attr = nla[NFTA_SET_NAME];
4728                 set = nft_set_lookup(table, attr, genmask);
4729         }
4730
4731         if (IS_ERR(set)) {
4732                 NL_SET_BAD_ATTR(extack, attr);
4733                 return PTR_ERR(set);
4734         }
4735         if (set->use ||
4736             (info->nlh->nlmsg_flags & NLM_F_NONREC &&
4737              atomic_read(&set->nelems) > 0)) {
4738                 NL_SET_BAD_ATTR(extack, attr);
4739                 return -EBUSY;
4740         }
4741
4742         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4743
4744         return nft_delset(&ctx, set);
4745 }
4746
4747 static int nft_validate_register_store(const struct nft_ctx *ctx,
4748                                        enum nft_registers reg,
4749                                        const struct nft_data *data,
4750                                        enum nft_data_types type,
4751                                        unsigned int len);
4752
4753 static int nft_setelem_data_validate(const struct nft_ctx *ctx,
4754                                      struct nft_set *set,
4755                                      struct nft_set_elem *elem)
4756 {
4757         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4758         enum nft_registers dreg;
4759
4760         dreg = nft_type_to_reg(set->dtype);
4761         return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
4762                                            set->dtype == NFT_DATA_VERDICT ?
4763                                            NFT_DATA_VERDICT : NFT_DATA_VALUE,
4764                                            set->dlen);
4765 }
4766
4767 static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
4768                                         struct nft_set *set,
4769                                         const struct nft_set_iter *iter,
4770                                         struct nft_set_elem *elem)
4771 {
4772         return nft_setelem_data_validate(ctx, set, elem);
4773 }
4774
4775 static int nft_set_catchall_bind_check(const struct nft_ctx *ctx,
4776                                        struct nft_set *set)
4777 {
4778         u8 genmask = nft_genmask_next(ctx->net);
4779         struct nft_set_elem_catchall *catchall;
4780         struct nft_set_elem elem;
4781         struct nft_set_ext *ext;
4782         int ret = 0;
4783
4784         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4785                 ext = nft_set_elem_ext(set, catchall->elem);
4786                 if (!nft_set_elem_active(ext, genmask))
4787                         continue;
4788
4789                 elem.priv = catchall->elem;
4790                 ret = nft_setelem_data_validate(ctx, set, &elem);
4791                 if (ret < 0)
4792                         break;
4793         }
4794
4795         return ret;
4796 }
4797
4798 int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
4799                        struct nft_set_binding *binding)
4800 {
4801         struct nft_set_binding *i;
4802         struct nft_set_iter iter;
4803
4804         if (set->use == UINT_MAX)
4805                 return -EOVERFLOW;
4806
4807         if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
4808                 return -EBUSY;
4809
4810         if (binding->flags & NFT_SET_MAP) {
4811                 /* If the set is already bound to the same chain all
4812                  * jumps are already validated for that chain.
4813                  */
4814                 list_for_each_entry(i, &set->bindings, list) {
4815                         if (i->flags & NFT_SET_MAP &&
4816                             i->chain == binding->chain)
4817                                 goto bind;
4818                 }
4819
4820                 iter.genmask    = nft_genmask_next(ctx->net);
4821                 iter.skip       = 0;
4822                 iter.count      = 0;
4823                 iter.err        = 0;
4824                 iter.fn         = nf_tables_bind_check_setelem;
4825
4826                 set->ops->walk(ctx, set, &iter);
4827                 if (!iter.err)
4828                         iter.err = nft_set_catchall_bind_check(ctx, set);
4829
4830                 if (iter.err < 0)
4831                         return iter.err;
4832         }
4833 bind:
4834         binding->chain = ctx->chain;
4835         list_add_tail_rcu(&binding->list, &set->bindings);
4836         nft_set_trans_bind(ctx, set);
4837         set->use++;
4838
4839         return 0;
4840 }
4841 EXPORT_SYMBOL_GPL(nf_tables_bind_set);
4842
4843 static void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
4844                                  struct nft_set_binding *binding, bool event)
4845 {
4846         list_del_rcu(&binding->list);
4847
4848         if (list_empty(&set->bindings) && nft_set_is_anonymous(set)) {
4849                 list_del_rcu(&set->list);
4850                 if (event)
4851                         nf_tables_set_notify(ctx, set, NFT_MSG_DELSET,
4852                                              GFP_KERNEL);
4853         }
4854 }
4855
4856 void nf_tables_deactivate_set(const struct nft_ctx *ctx, struct nft_set *set,
4857                               struct nft_set_binding *binding,
4858                               enum nft_trans_phase phase)
4859 {
4860         switch (phase) {
4861         case NFT_TRANS_PREPARE:
4862                 set->use--;
4863                 return;
4864         case NFT_TRANS_ABORT:
4865         case NFT_TRANS_RELEASE:
4866                 set->use--;
4867                 fallthrough;
4868         default:
4869                 nf_tables_unbind_set(ctx, set, binding,
4870                                      phase == NFT_TRANS_COMMIT);
4871         }
4872 }
4873 EXPORT_SYMBOL_GPL(nf_tables_deactivate_set);
4874
4875 void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
4876 {
4877         if (list_empty(&set->bindings) && nft_set_is_anonymous(set))
4878                 nft_set_destroy(ctx, set);
4879 }
4880 EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
4881
4882 const struct nft_set_ext_type nft_set_ext_types[] = {
4883         [NFT_SET_EXT_KEY]               = {
4884                 .align  = __alignof__(u32),
4885         },
4886         [NFT_SET_EXT_DATA]              = {
4887                 .align  = __alignof__(u32),
4888         },
4889         [NFT_SET_EXT_EXPRESSIONS]       = {
4890                 .align  = __alignof__(struct nft_set_elem_expr),
4891         },
4892         [NFT_SET_EXT_OBJREF]            = {
4893                 .len    = sizeof(struct nft_object *),
4894                 .align  = __alignof__(struct nft_object *),
4895         },
4896         [NFT_SET_EXT_FLAGS]             = {
4897                 .len    = sizeof(u8),
4898                 .align  = __alignof__(u8),
4899         },
4900         [NFT_SET_EXT_TIMEOUT]           = {
4901                 .len    = sizeof(u64),
4902                 .align  = __alignof__(u64),
4903         },
4904         [NFT_SET_EXT_EXPIRATION]        = {
4905                 .len    = sizeof(u64),
4906                 .align  = __alignof__(u64),
4907         },
4908         [NFT_SET_EXT_USERDATA]          = {
4909                 .len    = sizeof(struct nft_userdata),
4910                 .align  = __alignof__(struct nft_userdata),
4911         },
4912         [NFT_SET_EXT_KEY_END]           = {
4913                 .align  = __alignof__(u32),
4914         },
4915 };
4916
4917 /*
4918  * Set elements
4919  */
4920
4921 static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
4922         [NFTA_SET_ELEM_KEY]             = { .type = NLA_NESTED },
4923         [NFTA_SET_ELEM_DATA]            = { .type = NLA_NESTED },
4924         [NFTA_SET_ELEM_FLAGS]           = { .type = NLA_U32 },
4925         [NFTA_SET_ELEM_TIMEOUT]         = { .type = NLA_U64 },
4926         [NFTA_SET_ELEM_EXPIRATION]      = { .type = NLA_U64 },
4927         [NFTA_SET_ELEM_USERDATA]        = { .type = NLA_BINARY,
4928                                             .len = NFT_USERDATA_MAXLEN },
4929         [NFTA_SET_ELEM_EXPR]            = { .type = NLA_NESTED },
4930         [NFTA_SET_ELEM_OBJREF]          = { .type = NLA_STRING,
4931                                             .len = NFT_OBJ_MAXNAMELEN - 1 },
4932         [NFTA_SET_ELEM_KEY_END]         = { .type = NLA_NESTED },
4933         [NFTA_SET_ELEM_EXPRESSIONS]     = { .type = NLA_NESTED },
4934 };
4935
4936 static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
4937         [NFTA_SET_ELEM_LIST_TABLE]      = { .type = NLA_STRING,
4938                                             .len = NFT_TABLE_MAXNAMELEN - 1 },
4939         [NFTA_SET_ELEM_LIST_SET]        = { .type = NLA_STRING,
4940                                             .len = NFT_SET_MAXNAMELEN - 1 },
4941         [NFTA_SET_ELEM_LIST_ELEMENTS]   = { .type = NLA_NESTED },
4942         [NFTA_SET_ELEM_LIST_SET_ID]     = { .type = NLA_U32 },
4943 };
4944
4945 static int nft_set_elem_expr_dump(struct sk_buff *skb,
4946                                   const struct nft_set *set,
4947                                   const struct nft_set_ext *ext)
4948 {
4949         struct nft_set_elem_expr *elem_expr;
4950         u32 size, num_exprs = 0;
4951         struct nft_expr *expr;
4952         struct nlattr *nest;
4953
4954         elem_expr = nft_set_ext_expr(ext);
4955         nft_setelem_expr_foreach(expr, elem_expr, size)
4956                 num_exprs++;
4957
4958         if (num_exprs == 1) {
4959                 expr = nft_setelem_expr_at(elem_expr, 0);
4960                 if (nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, expr, false) < 0)
4961                         return -1;
4962
4963                 return 0;
4964         } else if (num_exprs > 1) {
4965                 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_EXPRESSIONS);
4966                 if (nest == NULL)
4967                         goto nla_put_failure;
4968
4969                 nft_setelem_expr_foreach(expr, elem_expr, size) {
4970                         expr = nft_setelem_expr_at(elem_expr, size);
4971                         if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr, false) < 0)
4972                                 goto nla_put_failure;
4973                 }
4974                 nla_nest_end(skb, nest);
4975         }
4976         return 0;
4977
4978 nla_put_failure:
4979         return -1;
4980 }
4981
4982 static int nf_tables_fill_setelem(struct sk_buff *skb,
4983                                   const struct nft_set *set,
4984                                   const struct nft_set_elem *elem)
4985 {
4986         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4987         unsigned char *b = skb_tail_pointer(skb);
4988         struct nlattr *nest;
4989
4990         nest = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4991         if (nest == NULL)
4992                 goto nla_put_failure;
4993
4994         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
4995             nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
4996                           NFT_DATA_VALUE, set->klen) < 0)
4997                 goto nla_put_failure;
4998
4999         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
5000             nft_data_dump(skb, NFTA_SET_ELEM_KEY_END, nft_set_ext_key_end(ext),
5001                           NFT_DATA_VALUE, set->klen) < 0)
5002                 goto nla_put_failure;
5003
5004         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5005             nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
5006                           set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
5007                           set->dlen) < 0)
5008                 goto nla_put_failure;
5009
5010         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS) &&
5011             nft_set_elem_expr_dump(skb, set, ext))
5012                 goto nla_put_failure;
5013
5014         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
5015             nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
5016                            (*nft_set_ext_obj(ext))->key.name) < 0)
5017                 goto nla_put_failure;
5018
5019         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
5020             nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
5021                          htonl(*nft_set_ext_flags(ext))))
5022                 goto nla_put_failure;
5023
5024         if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
5025             nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
5026                          nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
5027                          NFTA_SET_ELEM_PAD))
5028                 goto nla_put_failure;
5029
5030         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5031                 u64 expires, now = get_jiffies_64();
5032
5033                 expires = *nft_set_ext_expiration(ext);
5034                 if (time_before64(now, expires))
5035                         expires -= now;
5036                 else
5037                         expires = 0;
5038
5039                 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
5040                                  nf_jiffies64_to_msecs(expires),
5041                                  NFTA_SET_ELEM_PAD))
5042                         goto nla_put_failure;
5043         }
5044
5045         if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
5046                 struct nft_userdata *udata;
5047
5048                 udata = nft_set_ext_userdata(ext);
5049                 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
5050                             udata->len + 1, udata->data))
5051                         goto nla_put_failure;
5052         }
5053
5054         nla_nest_end(skb, nest);
5055         return 0;
5056
5057 nla_put_failure:
5058         nlmsg_trim(skb, b);
5059         return -EMSGSIZE;
5060 }
5061
5062 struct nft_set_dump_args {
5063         const struct netlink_callback   *cb;
5064         struct nft_set_iter             iter;
5065         struct sk_buff                  *skb;
5066 };
5067
5068 static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
5069                                   struct nft_set *set,
5070                                   const struct nft_set_iter *iter,
5071                                   struct nft_set_elem *elem)
5072 {
5073         struct nft_set_dump_args *args;
5074
5075         args = container_of(iter, struct nft_set_dump_args, iter);
5076         return nf_tables_fill_setelem(args->skb, set, elem);
5077 }
5078
5079 struct nft_set_dump_ctx {
5080         const struct nft_set    *set;
5081         struct nft_ctx          ctx;
5082 };
5083
5084 static int nft_set_catchall_dump(struct net *net, struct sk_buff *skb,
5085                                  const struct nft_set *set)
5086 {
5087         struct nft_set_elem_catchall *catchall;
5088         u8 genmask = nft_genmask_cur(net);
5089         struct nft_set_elem elem;
5090         struct nft_set_ext *ext;
5091         int ret = 0;
5092
5093         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5094                 ext = nft_set_elem_ext(set, catchall->elem);
5095                 if (!nft_set_elem_active(ext, genmask) ||
5096                     nft_set_elem_expired(ext))
5097                         continue;
5098
5099                 elem.priv = catchall->elem;
5100                 ret = nf_tables_fill_setelem(skb, set, &elem);
5101                 break;
5102         }
5103
5104         return ret;
5105 }
5106
5107 static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
5108 {
5109         struct nft_set_dump_ctx *dump_ctx = cb->data;
5110         struct net *net = sock_net(skb->sk);
5111         struct nftables_pernet *nft_net;
5112         struct nft_table *table;
5113         struct nft_set *set;
5114         struct nft_set_dump_args args;
5115         bool set_found = false;
5116         struct nlmsghdr *nlh;
5117         struct nlattr *nest;
5118         u32 portid, seq;
5119         int event;
5120
5121         rcu_read_lock();
5122         nft_net = nft_pernet(net);
5123         cb->seq = READ_ONCE(nft_net->base_seq);
5124
5125         list_for_each_entry_rcu(table, &nft_net->tables, list) {
5126                 if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
5127                     dump_ctx->ctx.family != table->family)
5128                         continue;
5129
5130                 if (table != dump_ctx->ctx.table)
5131                         continue;
5132
5133                 list_for_each_entry_rcu(set, &table->sets, list) {
5134                         if (set == dump_ctx->set) {
5135                                 set_found = true;
5136                                 break;
5137                         }
5138                 }
5139                 break;
5140         }
5141
5142         if (!set_found) {
5143                 rcu_read_unlock();
5144                 return -ENOENT;
5145         }
5146
5147         event  = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
5148         portid = NETLINK_CB(cb->skb).portid;
5149         seq    = cb->nlh->nlmsg_seq;
5150
5151         nlh = nfnl_msg_put(skb, portid, seq, event, NLM_F_MULTI,
5152                            table->family, NFNETLINK_V0, nft_base_seq(net));
5153         if (!nlh)
5154                 goto nla_put_failure;
5155
5156         if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
5157                 goto nla_put_failure;
5158         if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
5159                 goto nla_put_failure;
5160
5161         nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5162         if (nest == NULL)
5163                 goto nla_put_failure;
5164
5165         args.cb                 = cb;
5166         args.skb                = skb;
5167         args.iter.genmask       = nft_genmask_cur(net);
5168         args.iter.skip          = cb->args[0];
5169         args.iter.count         = 0;
5170         args.iter.err           = 0;
5171         args.iter.fn            = nf_tables_dump_setelem;
5172         set->ops->walk(&dump_ctx->ctx, set, &args.iter);
5173
5174         if (!args.iter.err && args.iter.count == cb->args[0])
5175                 args.iter.err = nft_set_catchall_dump(net, skb, set);
5176         rcu_read_unlock();
5177
5178         nla_nest_end(skb, nest);
5179         nlmsg_end(skb, nlh);
5180
5181         if (args.iter.err && args.iter.err != -EMSGSIZE)
5182                 return args.iter.err;
5183         if (args.iter.count == cb->args[0])
5184                 return 0;
5185
5186         cb->args[0] = args.iter.count;
5187         return skb->len;
5188
5189 nla_put_failure:
5190         rcu_read_unlock();
5191         return -ENOSPC;
5192 }
5193
5194 static int nf_tables_dump_set_start(struct netlink_callback *cb)
5195 {
5196         struct nft_set_dump_ctx *dump_ctx = cb->data;
5197
5198         cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
5199
5200         return cb->data ? 0 : -ENOMEM;
5201 }
5202
5203 static int nf_tables_dump_set_done(struct netlink_callback *cb)
5204 {
5205         kfree(cb->data);
5206         return 0;
5207 }
5208
5209 static int nf_tables_fill_setelem_info(struct sk_buff *skb,
5210                                        const struct nft_ctx *ctx, u32 seq,
5211                                        u32 portid, int event, u16 flags,
5212                                        const struct nft_set *set,
5213                                        const struct nft_set_elem *elem)
5214 {
5215         struct nlmsghdr *nlh;
5216         struct nlattr *nest;
5217         int err;
5218
5219         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
5220         nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
5221                            NFNETLINK_V0, nft_base_seq(ctx->net));
5222         if (!nlh)
5223                 goto nla_put_failure;
5224
5225         if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
5226                 goto nla_put_failure;
5227         if (nla_put_string(skb, NFTA_SET_NAME, set->name))
5228                 goto nla_put_failure;
5229
5230         nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5231         if (nest == NULL)
5232                 goto nla_put_failure;
5233
5234         err = nf_tables_fill_setelem(skb, set, elem);
5235         if (err < 0)
5236                 goto nla_put_failure;
5237
5238         nla_nest_end(skb, nest);
5239
5240         nlmsg_end(skb, nlh);
5241         return 0;
5242
5243 nla_put_failure:
5244         nlmsg_trim(skb, nlh);
5245         return -1;
5246 }
5247
5248 static int nft_setelem_parse_flags(const struct nft_set *set,
5249                                    const struct nlattr *attr, u32 *flags)
5250 {
5251         if (attr == NULL)
5252                 return 0;
5253
5254         *flags = ntohl(nla_get_be32(attr));
5255         if (*flags & ~(NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5256                 return -EOPNOTSUPP;
5257         if (!(set->flags & NFT_SET_INTERVAL) &&
5258             *flags & NFT_SET_ELEM_INTERVAL_END)
5259                 return -EINVAL;
5260         if ((*flags & (NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL)) ==
5261             (NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5262                 return -EINVAL;
5263
5264         return 0;
5265 }
5266
5267 static int nft_setelem_parse_key(struct nft_ctx *ctx, struct nft_set *set,
5268                                  struct nft_data *key, struct nlattr *attr)
5269 {
5270         struct nft_data_desc desc = {
5271                 .type   = NFT_DATA_VALUE,
5272                 .size   = NFT_DATA_VALUE_MAXLEN,
5273                 .len    = set->klen,
5274         };
5275
5276         return nft_data_init(ctx, key, &desc, attr);
5277 }
5278
5279 static int nft_setelem_parse_data(struct nft_ctx *ctx, struct nft_set *set,
5280                                   struct nft_data_desc *desc,
5281                                   struct nft_data *data,
5282                                   struct nlattr *attr)
5283 {
5284         u32 dtype;
5285
5286         if (set->dtype == NFT_DATA_VERDICT)
5287                 dtype = NFT_DATA_VERDICT;
5288         else
5289                 dtype = NFT_DATA_VALUE;
5290
5291         desc->type = dtype;
5292         desc->size = NFT_DATA_VALUE_MAXLEN;
5293         desc->len = set->dlen;
5294         desc->flags = NFT_DATA_DESC_SETELEM;
5295
5296         return nft_data_init(ctx, data, desc, attr);
5297 }
5298
5299 static void *nft_setelem_catchall_get(const struct net *net,
5300                                       const struct nft_set *set)
5301 {
5302         struct nft_set_elem_catchall *catchall;
5303         u8 genmask = nft_genmask_cur(net);
5304         struct nft_set_ext *ext;
5305         void *priv = NULL;
5306
5307         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5308                 ext = nft_set_elem_ext(set, catchall->elem);
5309                 if (!nft_set_elem_active(ext, genmask) ||
5310                     nft_set_elem_expired(ext))
5311                         continue;
5312
5313                 priv = catchall->elem;
5314                 break;
5315         }
5316
5317         return priv;
5318 }
5319
5320 static int nft_setelem_get(struct nft_ctx *ctx, struct nft_set *set,
5321                            struct nft_set_elem *elem, u32 flags)
5322 {
5323         void *priv;
5324
5325         if (!(flags & NFT_SET_ELEM_CATCHALL)) {
5326                 priv = set->ops->get(ctx->net, set, elem, flags);
5327                 if (IS_ERR(priv))
5328                         return PTR_ERR(priv);
5329         } else {
5330                 priv = nft_setelem_catchall_get(ctx->net, set);
5331                 if (!priv)
5332                         return -ENOENT;
5333         }
5334         elem->priv = priv;
5335
5336         return 0;
5337 }
5338
5339 static int nft_get_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5340                             const struct nlattr *attr)
5341 {
5342         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5343         struct nft_set_elem elem;
5344         struct sk_buff *skb;
5345         uint32_t flags = 0;
5346         int err;
5347
5348         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5349                                           nft_set_elem_policy, NULL);
5350         if (err < 0)
5351                 return err;
5352
5353         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5354         if (err < 0)
5355                 return err;
5356
5357         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5358                 return -EINVAL;
5359
5360         if (nla[NFTA_SET_ELEM_KEY]) {
5361                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5362                                             nla[NFTA_SET_ELEM_KEY]);
5363                 if (err < 0)
5364                         return err;
5365         }
5366
5367         if (nla[NFTA_SET_ELEM_KEY_END]) {
5368                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5369                                             nla[NFTA_SET_ELEM_KEY_END]);
5370                 if (err < 0)
5371                         return err;
5372         }
5373
5374         err = nft_setelem_get(ctx, set, &elem, flags);
5375         if (err < 0)
5376                 return err;
5377
5378         err = -ENOMEM;
5379         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
5380         if (skb == NULL)
5381                 return err;
5382
5383         err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
5384                                           NFT_MSG_NEWSETELEM, 0, set, &elem);
5385         if (err < 0)
5386                 goto err_fill_setelem;
5387
5388         return nfnetlink_unicast(skb, ctx->net, ctx->portid);
5389
5390 err_fill_setelem:
5391         kfree_skb(skb);
5392         return err;
5393 }
5394
5395 /* called with rcu_read_lock held */
5396 static int nf_tables_getsetelem(struct sk_buff *skb,
5397                                 const struct nfnl_info *info,
5398                                 const struct nlattr * const nla[])
5399 {
5400         struct netlink_ext_ack *extack = info->extack;
5401         u8 genmask = nft_genmask_cur(info->net);
5402         u8 family = info->nfmsg->nfgen_family;
5403         struct net *net = info->net;
5404         struct nft_table *table;
5405         struct nft_set *set;
5406         struct nlattr *attr;
5407         struct nft_ctx ctx;
5408         int rem, err = 0;
5409
5410         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
5411                                  genmask, NETLINK_CB(skb).portid);
5412         if (IS_ERR(table)) {
5413                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
5414                 return PTR_ERR(table);
5415         }
5416
5417         set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
5418         if (IS_ERR(set))
5419                 return PTR_ERR(set);
5420
5421         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5422
5423         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
5424                 struct netlink_dump_control c = {
5425                         .start = nf_tables_dump_set_start,
5426                         .dump = nf_tables_dump_set,
5427                         .done = nf_tables_dump_set_done,
5428                         .module = THIS_MODULE,
5429                 };
5430                 struct nft_set_dump_ctx dump_ctx = {
5431                         .set = set,
5432                         .ctx = ctx,
5433                 };
5434
5435                 c.data = &dump_ctx;
5436                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
5437         }
5438
5439         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
5440                 return -EINVAL;
5441
5442         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
5443                 err = nft_get_set_elem(&ctx, set, attr);
5444                 if (err < 0) {
5445                         NL_SET_BAD_ATTR(extack, attr);
5446                         break;
5447                 }
5448         }
5449
5450         return err;
5451 }
5452
5453 static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
5454                                      const struct nft_set *set,
5455                                      const struct nft_set_elem *elem,
5456                                      int event)
5457 {
5458         struct nftables_pernet *nft_net;
5459         struct net *net = ctx->net;
5460         u32 portid = ctx->portid;
5461         struct sk_buff *skb;
5462         u16 flags = 0;
5463         int err;
5464
5465         if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5466                 return;
5467
5468         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5469         if (skb == NULL)
5470                 goto err;
5471
5472         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
5473                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
5474
5475         err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
5476                                           set, elem);
5477         if (err < 0) {
5478                 kfree_skb(skb);
5479                 goto err;
5480         }
5481
5482         nft_net = nft_pernet(net);
5483         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
5484         return;
5485 err:
5486         nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
5487 }
5488
5489 static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
5490                                               int msg_type,
5491                                               struct nft_set *set)
5492 {
5493         struct nft_trans *trans;
5494
5495         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
5496         if (trans == NULL)
5497                 return NULL;
5498
5499         nft_trans_elem_set(trans) = set;
5500         return trans;
5501 }
5502
5503 struct nft_expr *nft_set_elem_expr_alloc(const struct nft_ctx *ctx,
5504                                          const struct nft_set *set,
5505                                          const struct nlattr *attr)
5506 {
5507         struct nft_expr *expr;
5508         int err;
5509
5510         expr = nft_expr_init(ctx, attr);
5511         if (IS_ERR(expr))
5512                 return expr;
5513
5514         err = -EOPNOTSUPP;
5515         if (expr->ops->type->flags & NFT_EXPR_GC) {
5516                 if (set->flags & NFT_SET_TIMEOUT)
5517                         goto err_set_elem_expr;
5518                 if (!set->ops->gc_init)
5519                         goto err_set_elem_expr;
5520                 set->ops->gc_init(set);
5521         }
5522
5523         return expr;
5524
5525 err_set_elem_expr:
5526         nft_expr_destroy(ctx, expr);
5527         return ERR_PTR(err);
5528 }
5529
5530 static int nft_set_ext_check(const struct nft_set_ext_tmpl *tmpl, u8 id, u32 len)
5531 {
5532         len += nft_set_ext_types[id].len;
5533         if (len > tmpl->ext_len[id] ||
5534             len > U8_MAX)
5535                 return -1;
5536
5537         return 0;
5538 }
5539
5540 static int nft_set_ext_memcpy(const struct nft_set_ext_tmpl *tmpl, u8 id,
5541                               void *to, const void *from, u32 len)
5542 {
5543         if (nft_set_ext_check(tmpl, id, len) < 0)
5544                 return -1;
5545
5546         memcpy(to, from, len);
5547
5548         return 0;
5549 }
5550
5551 void *nft_set_elem_init(const struct nft_set *set,
5552                         const struct nft_set_ext_tmpl *tmpl,
5553                         const u32 *key, const u32 *key_end,
5554                         const u32 *data, u64 timeout, u64 expiration, gfp_t gfp)
5555 {
5556         struct nft_set_ext *ext;
5557         void *elem;
5558
5559         elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
5560         if (elem == NULL)
5561                 return ERR_PTR(-ENOMEM);
5562
5563         ext = nft_set_elem_ext(set, elem);
5564         nft_set_ext_init(ext, tmpl);
5565
5566         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
5567             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_KEY,
5568                                nft_set_ext_key(ext), key, set->klen) < 0)
5569                 goto err_ext_check;
5570
5571         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
5572             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_KEY_END,
5573                                nft_set_ext_key_end(ext), key_end, set->klen) < 0)
5574                 goto err_ext_check;
5575
5576         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5577             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_DATA,
5578                                nft_set_ext_data(ext), data, set->dlen) < 0)
5579                 goto err_ext_check;
5580
5581         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5582                 *nft_set_ext_expiration(ext) = get_jiffies_64() + expiration;
5583                 if (expiration == 0)
5584                         *nft_set_ext_expiration(ext) += timeout;
5585         }
5586         if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
5587                 *nft_set_ext_timeout(ext) = timeout;
5588
5589         return elem;
5590
5591 err_ext_check:
5592         kfree(elem);
5593
5594         return ERR_PTR(-EINVAL);
5595 }
5596
5597 static void __nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5598                                         struct nft_expr *expr)
5599 {
5600         if (expr->ops->destroy_clone) {
5601                 expr->ops->destroy_clone(ctx, expr);
5602                 module_put(expr->ops->type->owner);
5603         } else {
5604                 nf_tables_expr_destroy(ctx, expr);
5605         }
5606 }
5607
5608 static void nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5609                                       struct nft_set_elem_expr *elem_expr)
5610 {
5611         struct nft_expr *expr;
5612         u32 size;
5613
5614         nft_setelem_expr_foreach(expr, elem_expr, size)
5615                 __nft_set_elem_expr_destroy(ctx, expr);
5616 }
5617
5618 void nft_set_elem_destroy(const struct nft_set *set, void *elem,
5619                           bool destroy_expr)
5620 {
5621         struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5622         struct nft_ctx ctx = {
5623                 .net    = read_pnet(&set->net),
5624                 .family = set->table->family,
5625         };
5626
5627         nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
5628         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5629                 nft_data_release(nft_set_ext_data(ext), set->dtype);
5630         if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5631                 nft_set_elem_expr_destroy(&ctx, nft_set_ext_expr(ext));
5632
5633         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
5634                 (*nft_set_ext_obj(ext))->use--;
5635         kfree(elem);
5636 }
5637 EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
5638
5639 /* Only called from commit path, nft_setelem_data_deactivate() already deals
5640  * with the refcounting from the preparation phase.
5641  */
5642 static void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
5643                                        const struct nft_set *set, void *elem)
5644 {
5645         struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5646
5647         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5648                 nft_set_elem_expr_destroy(ctx, nft_set_ext_expr(ext));
5649
5650         kfree(elem);
5651 }
5652
5653 int nft_set_elem_expr_clone(const struct nft_ctx *ctx, struct nft_set *set,
5654                             struct nft_expr *expr_array[])
5655 {
5656         struct nft_expr *expr;
5657         int err, i, k;
5658
5659         for (i = 0; i < set->num_exprs; i++) {
5660                 expr = kzalloc(set->exprs[i]->ops->size, GFP_KERNEL_ACCOUNT);
5661                 if (!expr)
5662                         goto err_expr;
5663
5664                 err = nft_expr_clone(expr, set->exprs[i]);
5665                 if (err < 0) {
5666                         kfree(expr);
5667                         goto err_expr;
5668                 }
5669                 expr_array[i] = expr;
5670         }
5671
5672         return 0;
5673
5674 err_expr:
5675         for (k = i - 1; k >= 0; k--)
5676                 nft_expr_destroy(ctx, expr_array[k]);
5677
5678         return -ENOMEM;
5679 }
5680
5681 static int nft_set_elem_expr_setup(struct nft_ctx *ctx,
5682                                    const struct nft_set_ext_tmpl *tmpl,
5683                                    const struct nft_set_ext *ext,
5684                                    struct nft_expr *expr_array[],
5685                                    u32 num_exprs)
5686 {
5687         struct nft_set_elem_expr *elem_expr = nft_set_ext_expr(ext);
5688         u32 len = sizeof(struct nft_set_elem_expr);
5689         struct nft_expr *expr;
5690         int i, err;
5691
5692         if (num_exprs == 0)
5693                 return 0;
5694
5695         for (i = 0; i < num_exprs; i++)
5696                 len += expr_array[i]->ops->size;
5697
5698         if (nft_set_ext_check(tmpl, NFT_SET_EXT_EXPRESSIONS, len) < 0)
5699                 return -EINVAL;
5700
5701         for (i = 0; i < num_exprs; i++) {
5702                 expr = nft_setelem_expr_at(elem_expr, elem_expr->size);
5703                 err = nft_expr_clone(expr, expr_array[i]);
5704                 if (err < 0)
5705                         goto err_elem_expr_setup;
5706
5707                 elem_expr->size += expr_array[i]->ops->size;
5708                 nft_expr_destroy(ctx, expr_array[i]);
5709                 expr_array[i] = NULL;
5710         }
5711
5712         return 0;
5713
5714 err_elem_expr_setup:
5715         for (; i < num_exprs; i++) {
5716                 nft_expr_destroy(ctx, expr_array[i]);
5717                 expr_array[i] = NULL;
5718         }
5719
5720         return -ENOMEM;
5721 }
5722
5723 struct nft_set_ext *nft_set_catchall_lookup(const struct net *net,
5724                                             const struct nft_set *set)
5725 {
5726         struct nft_set_elem_catchall *catchall;
5727         u8 genmask = nft_genmask_cur(net);
5728         struct nft_set_ext *ext;
5729
5730         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5731                 ext = nft_set_elem_ext(set, catchall->elem);
5732                 if (nft_set_elem_active(ext, genmask) &&
5733                     !nft_set_elem_expired(ext))
5734                         return ext;
5735         }
5736
5737         return NULL;
5738 }
5739 EXPORT_SYMBOL_GPL(nft_set_catchall_lookup);
5740
5741 void *nft_set_catchall_gc(const struct nft_set *set)
5742 {
5743         struct nft_set_elem_catchall *catchall, *next;
5744         struct nft_set_ext *ext;
5745         void *elem = NULL;
5746
5747         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5748                 ext = nft_set_elem_ext(set, catchall->elem);
5749
5750                 if (!nft_set_elem_expired(ext) ||
5751                     nft_set_elem_mark_busy(ext))
5752                         continue;
5753
5754                 elem = catchall->elem;
5755                 list_del_rcu(&catchall->list);
5756                 kfree_rcu(catchall, rcu);
5757                 break;
5758         }
5759
5760         return elem;
5761 }
5762 EXPORT_SYMBOL_GPL(nft_set_catchall_gc);
5763
5764 static int nft_setelem_catchall_insert(const struct net *net,
5765                                        struct nft_set *set,
5766                                        const struct nft_set_elem *elem,
5767                                        struct nft_set_ext **pext)
5768 {
5769         struct nft_set_elem_catchall *catchall;
5770         u8 genmask = nft_genmask_next(net);
5771         struct nft_set_ext *ext;
5772
5773         list_for_each_entry(catchall, &set->catchall_list, list) {
5774                 ext = nft_set_elem_ext(set, catchall->elem);
5775                 if (nft_set_elem_active(ext, genmask)) {
5776                         *pext = ext;
5777                         return -EEXIST;
5778                 }
5779         }
5780
5781         catchall = kmalloc(sizeof(*catchall), GFP_KERNEL);
5782         if (!catchall)
5783                 return -ENOMEM;
5784
5785         catchall->elem = elem->priv;
5786         list_add_tail_rcu(&catchall->list, &set->catchall_list);
5787
5788         return 0;
5789 }
5790
5791 static int nft_setelem_insert(const struct net *net,
5792                               struct nft_set *set,
5793                               const struct nft_set_elem *elem,
5794                               struct nft_set_ext **ext, unsigned int flags)
5795 {
5796         int ret;
5797
5798         if (flags & NFT_SET_ELEM_CATCHALL)
5799                 ret = nft_setelem_catchall_insert(net, set, elem, ext);
5800         else
5801                 ret = set->ops->insert(net, set, elem, ext);
5802
5803         return ret;
5804 }
5805
5806 static bool nft_setelem_is_catchall(const struct nft_set *set,
5807                                     const struct nft_set_elem *elem)
5808 {
5809         struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5810
5811         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
5812             *nft_set_ext_flags(ext) & NFT_SET_ELEM_CATCHALL)
5813                 return true;
5814
5815         return false;
5816 }
5817
5818 static void nft_setelem_activate(struct net *net, struct nft_set *set,
5819                                  struct nft_set_elem *elem)
5820 {
5821         struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5822
5823         if (nft_setelem_is_catchall(set, elem)) {
5824                 nft_set_elem_change_active(net, set, ext);
5825                 nft_set_elem_clear_busy(ext);
5826         } else {
5827                 set->ops->activate(net, set, elem);
5828         }
5829 }
5830
5831 static int nft_setelem_catchall_deactivate(const struct net *net,
5832                                            struct nft_set *set,
5833                                            struct nft_set_elem *elem)
5834 {
5835         struct nft_set_elem_catchall *catchall;
5836         struct nft_set_ext *ext;
5837
5838         list_for_each_entry(catchall, &set->catchall_list, list) {
5839                 ext = nft_set_elem_ext(set, catchall->elem);
5840                 if (!nft_is_active(net, ext) ||
5841                     nft_set_elem_mark_busy(ext))
5842                         continue;
5843
5844                 kfree(elem->priv);
5845                 elem->priv = catchall->elem;
5846                 nft_set_elem_change_active(net, set, ext);
5847                 return 0;
5848         }
5849
5850         return -ENOENT;
5851 }
5852
5853 static int __nft_setelem_deactivate(const struct net *net,
5854                                     struct nft_set *set,
5855                                     struct nft_set_elem *elem)
5856 {
5857         void *priv;
5858
5859         priv = set->ops->deactivate(net, set, elem);
5860         if (!priv)
5861                 return -ENOENT;
5862
5863         kfree(elem->priv);
5864         elem->priv = priv;
5865         set->ndeact++;
5866
5867         return 0;
5868 }
5869
5870 static int nft_setelem_deactivate(const struct net *net,
5871                                   struct nft_set *set,
5872                                   struct nft_set_elem *elem, u32 flags)
5873 {
5874         int ret;
5875
5876         if (flags & NFT_SET_ELEM_CATCHALL)
5877                 ret = nft_setelem_catchall_deactivate(net, set, elem);
5878         else
5879                 ret = __nft_setelem_deactivate(net, set, elem);
5880
5881         return ret;
5882 }
5883
5884 static void nft_setelem_catchall_remove(const struct net *net,
5885                                         const struct nft_set *set,
5886                                         const struct nft_set_elem *elem)
5887 {
5888         struct nft_set_elem_catchall *catchall, *next;
5889
5890         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5891                 if (catchall->elem == elem->priv) {
5892                         list_del_rcu(&catchall->list);
5893                         kfree_rcu(catchall, rcu);
5894                         break;
5895                 }
5896         }
5897 }
5898
5899 static void nft_setelem_remove(const struct net *net,
5900                                const struct nft_set *set,
5901                                const struct nft_set_elem *elem)
5902 {
5903         if (nft_setelem_is_catchall(set, elem))
5904                 nft_setelem_catchall_remove(net, set, elem);
5905         else
5906                 set->ops->remove(net, set, elem);
5907 }
5908
5909 static bool nft_setelem_valid_key_end(const struct nft_set *set,
5910                                       struct nlattr **nla, u32 flags)
5911 {
5912         if ((set->flags & (NFT_SET_CONCAT | NFT_SET_INTERVAL)) ==
5913                           (NFT_SET_CONCAT | NFT_SET_INTERVAL)) {
5914                 if (flags & NFT_SET_ELEM_INTERVAL_END)
5915                         return false;
5916
5917                 if (nla[NFTA_SET_ELEM_KEY_END] &&
5918                     flags & NFT_SET_ELEM_CATCHALL)
5919                         return false;
5920         } else {
5921                 if (nla[NFTA_SET_ELEM_KEY_END])
5922                         return false;
5923         }
5924
5925         return true;
5926 }
5927
5928 static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5929                             const struct nlattr *attr, u32 nlmsg_flags)
5930 {
5931         struct nft_expr *expr_array[NFT_SET_EXPR_MAX] = {};
5932         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5933         u8 genmask = nft_genmask_next(ctx->net);
5934         u32 flags = 0, size = 0, num_exprs = 0;
5935         struct nft_set_ext_tmpl tmpl;
5936         struct nft_set_ext *ext, *ext2;
5937         struct nft_set_elem elem;
5938         struct nft_set_binding *binding;
5939         struct nft_object *obj = NULL;
5940         struct nft_userdata *udata;
5941         struct nft_data_desc desc;
5942         enum nft_registers dreg;
5943         struct nft_trans *trans;
5944         u64 timeout;
5945         u64 expiration;
5946         int err, i;
5947         u8 ulen;
5948
5949         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5950                                           nft_set_elem_policy, NULL);
5951         if (err < 0)
5952                 return err;
5953
5954         nft_set_ext_prepare(&tmpl);
5955
5956         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5957         if (err < 0)
5958                 return err;
5959
5960         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5961                 return -EINVAL;
5962
5963         if (flags != 0) {
5964                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
5965                 if (err < 0)
5966                         return err;
5967         }
5968
5969         if (set->flags & NFT_SET_MAP) {
5970                 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
5971                     !(flags & NFT_SET_ELEM_INTERVAL_END))
5972                         return -EINVAL;
5973         } else {
5974                 if (nla[NFTA_SET_ELEM_DATA] != NULL)
5975                         return -EINVAL;
5976         }
5977
5978         if (set->flags & NFT_SET_OBJECT) {
5979                 if (!nla[NFTA_SET_ELEM_OBJREF] &&
5980                     !(flags & NFT_SET_ELEM_INTERVAL_END))
5981                         return -EINVAL;
5982         } else {
5983                 if (nla[NFTA_SET_ELEM_OBJREF])
5984                         return -EINVAL;
5985         }
5986
5987         if (!nft_setelem_valid_key_end(set, nla, flags))
5988                 return -EINVAL;
5989
5990         if ((flags & NFT_SET_ELEM_INTERVAL_END) &&
5991              (nla[NFTA_SET_ELEM_DATA] ||
5992               nla[NFTA_SET_ELEM_OBJREF] ||
5993               nla[NFTA_SET_ELEM_TIMEOUT] ||
5994               nla[NFTA_SET_ELEM_EXPIRATION] ||
5995               nla[NFTA_SET_ELEM_USERDATA] ||
5996               nla[NFTA_SET_ELEM_EXPR] ||
5997               nla[NFTA_SET_ELEM_KEY_END] ||
5998               nla[NFTA_SET_ELEM_EXPRESSIONS]))
5999                 return -EINVAL;
6000
6001         timeout = 0;
6002         if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
6003                 if (!(set->flags & NFT_SET_TIMEOUT))
6004                         return -EINVAL;
6005                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
6006                                             &timeout);
6007                 if (err)
6008                         return err;
6009         } else if (set->flags & NFT_SET_TIMEOUT &&
6010                    !(flags & NFT_SET_ELEM_INTERVAL_END)) {
6011                 timeout = set->timeout;
6012         }
6013
6014         expiration = 0;
6015         if (nla[NFTA_SET_ELEM_EXPIRATION] != NULL) {
6016                 if (!(set->flags & NFT_SET_TIMEOUT))
6017                         return -EINVAL;
6018                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_EXPIRATION],
6019                                             &expiration);
6020                 if (err)
6021                         return err;
6022         }
6023
6024         if (nla[NFTA_SET_ELEM_EXPR]) {
6025                 struct nft_expr *expr;
6026
6027                 if (set->num_exprs && set->num_exprs != 1)
6028                         return -EOPNOTSUPP;
6029
6030                 expr = nft_set_elem_expr_alloc(ctx, set,
6031                                                nla[NFTA_SET_ELEM_EXPR]);
6032                 if (IS_ERR(expr))
6033                         return PTR_ERR(expr);
6034
6035                 expr_array[0] = expr;
6036                 num_exprs = 1;
6037
6038                 if (set->num_exprs && set->exprs[0]->ops != expr->ops) {
6039                         err = -EOPNOTSUPP;
6040                         goto err_set_elem_expr;
6041                 }
6042         } else if (nla[NFTA_SET_ELEM_EXPRESSIONS]) {
6043                 struct nft_expr *expr;
6044                 struct nlattr *tmp;
6045                 int left;
6046
6047                 i = 0;
6048                 nla_for_each_nested(tmp, nla[NFTA_SET_ELEM_EXPRESSIONS], left) {
6049                         if (i == NFT_SET_EXPR_MAX ||
6050                             (set->num_exprs && set->num_exprs == i)) {
6051                                 err = -E2BIG;
6052                                 goto err_set_elem_expr;
6053                         }
6054                         if (nla_type(tmp) != NFTA_LIST_ELEM) {
6055                                 err = -EINVAL;
6056                                 goto err_set_elem_expr;
6057                         }
6058                         expr = nft_set_elem_expr_alloc(ctx, set, tmp);
6059                         if (IS_ERR(expr)) {
6060                                 err = PTR_ERR(expr);
6061                                 goto err_set_elem_expr;
6062                         }
6063                         expr_array[i] = expr;
6064                         num_exprs++;
6065
6066                         if (set->num_exprs && expr->ops != set->exprs[i]->ops) {
6067                                 err = -EOPNOTSUPP;
6068                                 goto err_set_elem_expr;
6069                         }
6070                         i++;
6071                 }
6072                 if (set->num_exprs && set->num_exprs != i) {
6073                         err = -EOPNOTSUPP;
6074                         goto err_set_elem_expr;
6075                 }
6076         } else if (set->num_exprs > 0 &&
6077                    !(flags & NFT_SET_ELEM_INTERVAL_END)) {
6078                 err = nft_set_elem_expr_clone(ctx, set, expr_array);
6079                 if (err < 0)
6080                         goto err_set_elem_expr_clone;
6081
6082                 num_exprs = set->num_exprs;
6083         }
6084
6085         if (nla[NFTA_SET_ELEM_KEY]) {
6086                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6087                                             nla[NFTA_SET_ELEM_KEY]);
6088                 if (err < 0)
6089                         goto err_set_elem_expr;
6090
6091                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6092                 if (err < 0)
6093                         goto err_parse_key;
6094         }
6095
6096         if (nla[NFTA_SET_ELEM_KEY_END]) {
6097                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6098                                             nla[NFTA_SET_ELEM_KEY_END]);
6099                 if (err < 0)
6100                         goto err_parse_key;
6101
6102                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6103                 if (err < 0)
6104                         goto err_parse_key_end;
6105         }
6106
6107         if (timeout > 0) {
6108                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
6109                 if (err < 0)
6110                         goto err_parse_key_end;
6111
6112                 if (timeout != set->timeout) {
6113                         err = nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
6114                         if (err < 0)
6115                                 goto err_parse_key_end;
6116                 }
6117         }
6118
6119         if (num_exprs) {
6120                 for (i = 0; i < num_exprs; i++)
6121                         size += expr_array[i]->ops->size;
6122
6123                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_EXPRESSIONS,
6124                                              sizeof(struct nft_set_elem_expr) + size);
6125                 if (err < 0)
6126                         goto err_parse_key_end;
6127         }
6128
6129         if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
6130                 obj = nft_obj_lookup(ctx->net, ctx->table,
6131                                      nla[NFTA_SET_ELEM_OBJREF],
6132                                      set->objtype, genmask);
6133                 if (IS_ERR(obj)) {
6134                         err = PTR_ERR(obj);
6135                         goto err_parse_key_end;
6136                 }
6137                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
6138                 if (err < 0)
6139                         goto err_parse_key_end;
6140         }
6141
6142         if (nla[NFTA_SET_ELEM_DATA] != NULL) {
6143                 err = nft_setelem_parse_data(ctx, set, &desc, &elem.data.val,
6144                                              nla[NFTA_SET_ELEM_DATA]);
6145                 if (err < 0)
6146                         goto err_parse_key_end;
6147
6148                 dreg = nft_type_to_reg(set->dtype);
6149                 list_for_each_entry(binding, &set->bindings, list) {
6150                         struct nft_ctx bind_ctx = {
6151                                 .net    = ctx->net,
6152                                 .family = ctx->family,
6153                                 .table  = ctx->table,
6154                                 .chain  = (struct nft_chain *)binding->chain,
6155                         };
6156
6157                         if (!(binding->flags & NFT_SET_MAP))
6158                                 continue;
6159
6160                         err = nft_validate_register_store(&bind_ctx, dreg,
6161                                                           &elem.data.val,
6162                                                           desc.type, desc.len);
6163                         if (err < 0)
6164                                 goto err_parse_data;
6165
6166                         if (desc.type == NFT_DATA_VERDICT &&
6167                             (elem.data.val.verdict.code == NFT_GOTO ||
6168                              elem.data.val.verdict.code == NFT_JUMP))
6169                                 nft_validate_state_update(ctx->net,
6170                                                           NFT_VALIDATE_NEED);
6171                 }
6172
6173                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, desc.len);
6174                 if (err < 0)
6175                         goto err_parse_data;
6176         }
6177
6178         /* The full maximum length of userdata can exceed the maximum
6179          * offset value (U8_MAX) for following extensions, therefor it
6180          * must be the last extension added.
6181          */
6182         ulen = 0;
6183         if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
6184                 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
6185                 if (ulen > 0) {
6186                         err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
6187                                                      ulen);
6188                         if (err < 0)
6189                                 goto err_parse_data;
6190                 }
6191         }
6192
6193         elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6194                                       elem.key_end.val.data, elem.data.val.data,
6195                                       timeout, expiration, GFP_KERNEL_ACCOUNT);
6196         if (IS_ERR(elem.priv)) {
6197                 err = PTR_ERR(elem.priv);
6198                 goto err_parse_data;
6199         }
6200
6201         ext = nft_set_elem_ext(set, elem.priv);
6202         if (flags)
6203                 *nft_set_ext_flags(ext) = flags;
6204
6205         if (ulen > 0) {
6206                 if (nft_set_ext_check(&tmpl, NFT_SET_EXT_USERDATA, ulen) < 0) {
6207                         err = -EINVAL;
6208                         goto err_elem_userdata;
6209                 }
6210                 udata = nft_set_ext_userdata(ext);
6211                 udata->len = ulen - 1;
6212                 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
6213         }
6214         if (obj) {
6215                 *nft_set_ext_obj(ext) = obj;
6216                 obj->use++;
6217         }
6218         err = nft_set_elem_expr_setup(ctx, &tmpl, ext, expr_array, num_exprs);
6219         if (err < 0)
6220                 goto err_elem_free;
6221
6222         trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
6223         if (trans == NULL) {
6224                 err = -ENOMEM;
6225                 goto err_elem_free;
6226         }
6227
6228         ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
6229
6230         err = nft_setelem_insert(ctx->net, set, &elem, &ext2, flags);
6231         if (err) {
6232                 if (err == -EEXIST) {
6233                         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
6234                             nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
6235                             nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
6236                             nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF))
6237                                 goto err_element_clash;
6238                         if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
6239                              nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
6240                              memcmp(nft_set_ext_data(ext),
6241                                     nft_set_ext_data(ext2), set->dlen) != 0) ||
6242                             (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
6243                              nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
6244                              *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
6245                                 goto err_element_clash;
6246                         else if (!(nlmsg_flags & NLM_F_EXCL))
6247                                 err = 0;
6248                 } else if (err == -ENOTEMPTY) {
6249                         /* ENOTEMPTY reports overlapping between this element
6250                          * and an existing one.
6251                          */
6252                         err = -EEXIST;
6253                 }
6254                 goto err_element_clash;
6255         }
6256
6257         if (!(flags & NFT_SET_ELEM_CATCHALL) && set->size &&
6258             !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact)) {
6259                 err = -ENFILE;
6260                 goto err_set_full;
6261         }
6262
6263         nft_trans_elem(trans) = elem;
6264         nft_trans_commit_list_add_tail(ctx->net, trans);
6265         return 0;
6266
6267 err_set_full:
6268         nft_setelem_remove(ctx->net, set, &elem);
6269 err_element_clash:
6270         kfree(trans);
6271 err_elem_free:
6272         if (obj)
6273                 obj->use--;
6274 err_elem_userdata:
6275         nf_tables_set_elem_destroy(ctx, set, elem.priv);
6276 err_parse_data:
6277         if (nla[NFTA_SET_ELEM_DATA] != NULL)
6278                 nft_data_release(&elem.data.val, desc.type);
6279 err_parse_key_end:
6280         nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
6281 err_parse_key:
6282         nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6283 err_set_elem_expr:
6284         for (i = 0; i < num_exprs && expr_array[i]; i++)
6285                 nft_expr_destroy(ctx, expr_array[i]);
6286 err_set_elem_expr_clone:
6287         return err;
6288 }
6289
6290 static int nf_tables_newsetelem(struct sk_buff *skb,
6291                                 const struct nfnl_info *info,
6292                                 const struct nlattr * const nla[])
6293 {
6294         struct nftables_pernet *nft_net = nft_pernet(info->net);
6295         struct netlink_ext_ack *extack = info->extack;
6296         u8 genmask = nft_genmask_next(info->net);
6297         u8 family = info->nfmsg->nfgen_family;
6298         struct net *net = info->net;
6299         const struct nlattr *attr;
6300         struct nft_table *table;
6301         struct nft_set *set;
6302         struct nft_ctx ctx;
6303         int rem, err;
6304
6305         if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
6306                 return -EINVAL;
6307
6308         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6309                                  genmask, NETLINK_CB(skb).portid);
6310         if (IS_ERR(table)) {
6311                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6312                 return PTR_ERR(table);
6313         }
6314
6315         set = nft_set_lookup_global(net, table, nla[NFTA_SET_ELEM_LIST_SET],
6316                                     nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
6317         if (IS_ERR(set))
6318                 return PTR_ERR(set);
6319
6320         if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6321                 return -EBUSY;
6322
6323         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6324
6325         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6326                 err = nft_add_set_elem(&ctx, set, attr, info->nlh->nlmsg_flags);
6327                 if (err < 0) {
6328                         NL_SET_BAD_ATTR(extack, attr);
6329                         return err;
6330                 }
6331         }
6332
6333         if (nft_net->validate_state == NFT_VALIDATE_DO)
6334                 return nft_table_validate(net, table);
6335
6336         return 0;
6337 }
6338
6339 /**
6340  *      nft_data_hold - hold a nft_data item
6341  *
6342  *      @data: struct nft_data to release
6343  *      @type: type of data
6344  *
6345  *      Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
6346  *      NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
6347  *      NFT_GOTO verdicts. This function must be called on active data objects
6348  *      from the second phase of the commit protocol.
6349  */
6350 void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
6351 {
6352         struct nft_chain *chain;
6353         struct nft_rule *rule;
6354
6355         if (type == NFT_DATA_VERDICT) {
6356                 switch (data->verdict.code) {
6357                 case NFT_JUMP:
6358                 case NFT_GOTO:
6359                         chain = data->verdict.chain;
6360                         chain->use++;
6361
6362                         if (!nft_chain_is_bound(chain))
6363                                 break;
6364
6365                         chain->table->use++;
6366                         list_for_each_entry(rule, &chain->rules, list)
6367                                 chain->use++;
6368
6369                         nft_chain_add(chain->table, chain);
6370                         break;
6371                 }
6372         }
6373 }
6374
6375 static void nft_setelem_data_activate(const struct net *net,
6376                                       const struct nft_set *set,
6377                                       struct nft_set_elem *elem)
6378 {
6379         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6380
6381         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6382                 nft_data_hold(nft_set_ext_data(ext), set->dtype);
6383         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6384                 (*nft_set_ext_obj(ext))->use++;
6385 }
6386
6387 static void nft_setelem_data_deactivate(const struct net *net,
6388                                         const struct nft_set *set,
6389                                         struct nft_set_elem *elem)
6390 {
6391         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6392
6393         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6394                 nft_data_release(nft_set_ext_data(ext), set->dtype);
6395         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6396                 (*nft_set_ext_obj(ext))->use--;
6397 }
6398
6399 static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
6400                            const struct nlattr *attr)
6401 {
6402         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6403         struct nft_set_ext_tmpl tmpl;
6404         struct nft_set_elem elem;
6405         struct nft_set_ext *ext;
6406         struct nft_trans *trans;
6407         u32 flags = 0;
6408         int err;
6409
6410         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6411                                           nft_set_elem_policy, NULL);
6412         if (err < 0)
6413                 return err;
6414
6415         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6416         if (err < 0)
6417                 return err;
6418
6419         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
6420                 return -EINVAL;
6421
6422         if (!nft_setelem_valid_key_end(set, nla, flags))
6423                 return -EINVAL;
6424
6425         nft_set_ext_prepare(&tmpl);
6426
6427         if (flags != 0) {
6428                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
6429                 if (err < 0)
6430                         return err;
6431         }
6432
6433         if (nla[NFTA_SET_ELEM_KEY]) {
6434                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6435                                             nla[NFTA_SET_ELEM_KEY]);
6436                 if (err < 0)
6437                         return err;
6438
6439                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6440                 if (err < 0)
6441                         goto fail_elem;
6442         }
6443
6444         if (nla[NFTA_SET_ELEM_KEY_END]) {
6445                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6446                                             nla[NFTA_SET_ELEM_KEY_END]);
6447                 if (err < 0)
6448                         goto fail_elem;
6449
6450                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6451                 if (err < 0)
6452                         goto fail_elem_key_end;
6453         }
6454
6455         err = -ENOMEM;
6456         elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6457                                       elem.key_end.val.data, NULL, 0, 0,
6458                                       GFP_KERNEL_ACCOUNT);
6459         if (IS_ERR(elem.priv)) {
6460                 err = PTR_ERR(elem.priv);
6461                 goto fail_elem_key_end;
6462         }
6463
6464         ext = nft_set_elem_ext(set, elem.priv);
6465         if (flags)
6466                 *nft_set_ext_flags(ext) = flags;
6467
6468         trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
6469         if (trans == NULL)
6470                 goto fail_trans;
6471
6472         err = nft_setelem_deactivate(ctx->net, set, &elem, flags);
6473         if (err < 0)
6474                 goto fail_ops;
6475
6476         nft_setelem_data_deactivate(ctx->net, set, &elem);
6477
6478         nft_trans_elem(trans) = elem;
6479         nft_trans_commit_list_add_tail(ctx->net, trans);
6480         return 0;
6481
6482 fail_ops:
6483         kfree(trans);
6484 fail_trans:
6485         kfree(elem.priv);
6486 fail_elem_key_end:
6487         nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
6488 fail_elem:
6489         nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6490         return err;
6491 }
6492
6493 static int nft_setelem_flush(const struct nft_ctx *ctx,
6494                              struct nft_set *set,
6495                              const struct nft_set_iter *iter,
6496                              struct nft_set_elem *elem)
6497 {
6498         struct nft_trans *trans;
6499         int err;
6500
6501         trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6502                                     sizeof(struct nft_trans_elem), GFP_ATOMIC);
6503         if (!trans)
6504                 return -ENOMEM;
6505
6506         if (!set->ops->flush(ctx->net, set, elem->priv)) {
6507                 err = -ENOENT;
6508                 goto err1;
6509         }
6510         set->ndeact++;
6511
6512         nft_setelem_data_deactivate(ctx->net, set, elem);
6513         nft_trans_elem_set(trans) = set;
6514         nft_trans_elem(trans) = *elem;
6515         nft_trans_commit_list_add_tail(ctx->net, trans);
6516
6517         return 0;
6518 err1:
6519         kfree(trans);
6520         return err;
6521 }
6522
6523 static int __nft_set_catchall_flush(const struct nft_ctx *ctx,
6524                                     struct nft_set *set,
6525                                     struct nft_set_elem *elem)
6526 {
6527         struct nft_trans *trans;
6528
6529         trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6530                                     sizeof(struct nft_trans_elem), GFP_KERNEL);
6531         if (!trans)
6532                 return -ENOMEM;
6533
6534         nft_setelem_data_deactivate(ctx->net, set, elem);
6535         nft_trans_elem_set(trans) = set;
6536         nft_trans_elem(trans) = *elem;
6537         nft_trans_commit_list_add_tail(ctx->net, trans);
6538
6539         return 0;
6540 }
6541
6542 static int nft_set_catchall_flush(const struct nft_ctx *ctx,
6543                                   struct nft_set *set)
6544 {
6545         u8 genmask = nft_genmask_next(ctx->net);
6546         struct nft_set_elem_catchall *catchall;
6547         struct nft_set_elem elem;
6548         struct nft_set_ext *ext;
6549         int ret = 0;
6550
6551         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6552                 ext = nft_set_elem_ext(set, catchall->elem);
6553                 if (!nft_set_elem_active(ext, genmask) ||
6554                     nft_set_elem_mark_busy(ext))
6555                         continue;
6556
6557                 elem.priv = catchall->elem;
6558                 ret = __nft_set_catchall_flush(ctx, set, &elem);
6559                 if (ret < 0)
6560                         break;
6561         }
6562
6563         return ret;
6564 }
6565
6566 static int nft_set_flush(struct nft_ctx *ctx, struct nft_set *set, u8 genmask)
6567 {
6568         struct nft_set_iter iter = {
6569                 .genmask        = genmask,
6570                 .fn             = nft_setelem_flush,
6571         };
6572
6573         set->ops->walk(ctx, set, &iter);
6574         if (!iter.err)
6575                 iter.err = nft_set_catchall_flush(ctx, set);
6576
6577         return iter.err;
6578 }
6579
6580 static int nf_tables_delsetelem(struct sk_buff *skb,
6581                                 const struct nfnl_info *info,
6582                                 const struct nlattr * const nla[])
6583 {
6584         struct netlink_ext_ack *extack = info->extack;
6585         u8 genmask = nft_genmask_next(info->net);
6586         u8 family = info->nfmsg->nfgen_family;
6587         struct net *net = info->net;
6588         const struct nlattr *attr;
6589         struct nft_table *table;
6590         struct nft_set *set;
6591         struct nft_ctx ctx;
6592         int rem, err = 0;
6593
6594         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6595                                  genmask, NETLINK_CB(skb).portid);
6596         if (IS_ERR(table)) {
6597                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6598                 return PTR_ERR(table);
6599         }
6600
6601         set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
6602         if (IS_ERR(set))
6603                 return PTR_ERR(set);
6604         if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6605                 return -EBUSY;
6606
6607         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6608
6609         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
6610                 return nft_set_flush(&ctx, set, genmask);
6611
6612         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6613                 err = nft_del_setelem(&ctx, set, attr);
6614                 if (err < 0) {
6615                         NL_SET_BAD_ATTR(extack, attr);
6616                         break;
6617                 }
6618         }
6619         return err;
6620 }
6621
6622 void nft_set_gc_batch_release(struct rcu_head *rcu)
6623 {
6624         struct nft_set_gc_batch *gcb;
6625         unsigned int i;
6626
6627         gcb = container_of(rcu, struct nft_set_gc_batch, head.rcu);
6628         for (i = 0; i < gcb->head.cnt; i++)
6629                 nft_set_elem_destroy(gcb->head.set, gcb->elems[i], true);
6630         kfree(gcb);
6631 }
6632
6633 struct nft_set_gc_batch *nft_set_gc_batch_alloc(const struct nft_set *set,
6634                                                 gfp_t gfp)
6635 {
6636         struct nft_set_gc_batch *gcb;
6637
6638         gcb = kzalloc(sizeof(*gcb), gfp);
6639         if (gcb == NULL)
6640                 return gcb;
6641         gcb->head.set = set;
6642         return gcb;
6643 }
6644
6645 /*
6646  * Stateful objects
6647  */
6648
6649 /**
6650  *      nft_register_obj- register nf_tables stateful object type
6651  *      @obj_type: object type
6652  *
6653  *      Registers the object type for use with nf_tables. Returns zero on
6654  *      success or a negative errno code otherwise.
6655  */
6656 int nft_register_obj(struct nft_object_type *obj_type)
6657 {
6658         if (obj_type->type == NFT_OBJECT_UNSPEC)
6659                 return -EINVAL;
6660
6661         nfnl_lock(NFNL_SUBSYS_NFTABLES);
6662         list_add_rcu(&obj_type->list, &nf_tables_objects);
6663         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6664         return 0;
6665 }
6666 EXPORT_SYMBOL_GPL(nft_register_obj);
6667
6668 /**
6669  *      nft_unregister_obj - unregister nf_tables object type
6670  *      @obj_type: object type
6671  *
6672  *      Unregisters the object type for use with nf_tables.
6673  */
6674 void nft_unregister_obj(struct nft_object_type *obj_type)
6675 {
6676         nfnl_lock(NFNL_SUBSYS_NFTABLES);
6677         list_del_rcu(&obj_type->list);
6678         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6679 }
6680 EXPORT_SYMBOL_GPL(nft_unregister_obj);
6681
6682 struct nft_object *nft_obj_lookup(const struct net *net,
6683                                   const struct nft_table *table,
6684                                   const struct nlattr *nla, u32 objtype,
6685                                   u8 genmask)
6686 {
6687         struct nft_object_hash_key k = { .table = table };
6688         char search[NFT_OBJ_MAXNAMELEN];
6689         struct rhlist_head *tmp, *list;
6690         struct nft_object *obj;
6691
6692         nla_strscpy(search, nla, sizeof(search));
6693         k.name = search;
6694
6695         WARN_ON_ONCE(!rcu_read_lock_held() &&
6696                      !lockdep_commit_lock_is_held(net));
6697
6698         rcu_read_lock();
6699         list = rhltable_lookup(&nft_objname_ht, &k, nft_objname_ht_params);
6700         if (!list)
6701                 goto out;
6702
6703         rhl_for_each_entry_rcu(obj, tmp, list, rhlhead) {
6704                 if (objtype == obj->ops->type->type &&
6705                     nft_active_genmask(obj, genmask)) {
6706                         rcu_read_unlock();
6707                         return obj;
6708                 }
6709         }
6710 out:
6711         rcu_read_unlock();
6712         return ERR_PTR(-ENOENT);
6713 }
6714 EXPORT_SYMBOL_GPL(nft_obj_lookup);
6715
6716 static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
6717                                                   const struct nlattr *nla,
6718                                                   u32 objtype, u8 genmask)
6719 {
6720         struct nft_object *obj;
6721
6722         list_for_each_entry(obj, &table->objects, list) {
6723                 if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
6724                     objtype == obj->ops->type->type &&
6725                     nft_active_genmask(obj, genmask))
6726                         return obj;
6727         }
6728         return ERR_PTR(-ENOENT);
6729 }
6730
6731 static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
6732         [NFTA_OBJ_TABLE]        = { .type = NLA_STRING,
6733                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
6734         [NFTA_OBJ_NAME]         = { .type = NLA_STRING,
6735                                     .len = NFT_OBJ_MAXNAMELEN - 1 },
6736         [NFTA_OBJ_TYPE]         = { .type = NLA_U32 },
6737         [NFTA_OBJ_DATA]         = { .type = NLA_NESTED },
6738         [NFTA_OBJ_HANDLE]       = { .type = NLA_U64},
6739         [NFTA_OBJ_USERDATA]     = { .type = NLA_BINARY,
6740                                     .len = NFT_USERDATA_MAXLEN },
6741 };
6742
6743 static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
6744                                        const struct nft_object_type *type,
6745                                        const struct nlattr *attr)
6746 {
6747         struct nlattr **tb;
6748         const struct nft_object_ops *ops;
6749         struct nft_object *obj;
6750         int err = -ENOMEM;
6751
6752         tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
6753         if (!tb)
6754                 goto err1;
6755
6756         if (attr) {
6757                 err = nla_parse_nested_deprecated(tb, type->maxattr, attr,
6758                                                   type->policy, NULL);
6759                 if (err < 0)
6760                         goto err2;
6761         } else {
6762                 memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
6763         }
6764
6765         if (type->select_ops) {
6766                 ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
6767                 if (IS_ERR(ops)) {
6768                         err = PTR_ERR(ops);
6769                         goto err2;
6770                 }
6771         } else {
6772                 ops = type->ops;
6773         }
6774
6775         err = -ENOMEM;
6776         obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL_ACCOUNT);
6777         if (!obj)
6778                 goto err2;
6779
6780         err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
6781         if (err < 0)
6782                 goto err3;
6783
6784         obj->ops = ops;
6785
6786         kfree(tb);
6787         return obj;
6788 err3:
6789         kfree(obj);
6790 err2:
6791         kfree(tb);
6792 err1:
6793         return ERR_PTR(err);
6794 }
6795
6796 static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
6797                            struct nft_object *obj, bool reset)
6798 {
6799         struct nlattr *nest;
6800
6801         nest = nla_nest_start_noflag(skb, attr);
6802         if (!nest)
6803                 goto nla_put_failure;
6804         if (obj->ops->dump(skb, obj, reset) < 0)
6805                 goto nla_put_failure;
6806         nla_nest_end(skb, nest);
6807         return 0;
6808
6809 nla_put_failure:
6810         return -1;
6811 }
6812
6813 static const struct nft_object_type *__nft_obj_type_get(u32 objtype)
6814 {
6815         const struct nft_object_type *type;
6816
6817         list_for_each_entry(type, &nf_tables_objects, list) {
6818                 if (objtype == type->type)
6819                         return type;
6820         }
6821         return NULL;
6822 }
6823
6824 static const struct nft_object_type *
6825 nft_obj_type_get(struct net *net, u32 objtype)
6826 {
6827         const struct nft_object_type *type;
6828
6829         type = __nft_obj_type_get(objtype);
6830         if (type != NULL && try_module_get(type->owner))
6831                 return type;
6832
6833         lockdep_nfnl_nft_mutex_not_held();
6834 #ifdef CONFIG_MODULES
6835         if (type == NULL) {
6836                 if (nft_request_module(net, "nft-obj-%u", objtype) == -EAGAIN)
6837                         return ERR_PTR(-EAGAIN);
6838         }
6839 #endif
6840         return ERR_PTR(-ENOENT);
6841 }
6842
6843 static int nf_tables_updobj(const struct nft_ctx *ctx,
6844                             const struct nft_object_type *type,
6845                             const struct nlattr *attr,
6846                             struct nft_object *obj)
6847 {
6848         struct nft_object *newobj;
6849         struct nft_trans *trans;
6850         int err = -ENOMEM;
6851
6852         if (!try_module_get(type->owner))
6853                 return -ENOENT;
6854
6855         trans = nft_trans_alloc(ctx, NFT_MSG_NEWOBJ,
6856                                 sizeof(struct nft_trans_obj));
6857         if (!trans)
6858                 goto err_trans;
6859
6860         newobj = nft_obj_init(ctx, type, attr);
6861         if (IS_ERR(newobj)) {
6862                 err = PTR_ERR(newobj);
6863                 goto err_free_trans;
6864         }
6865
6866         nft_trans_obj(trans) = obj;
6867         nft_trans_obj_update(trans) = true;
6868         nft_trans_obj_newobj(trans) = newobj;
6869         nft_trans_commit_list_add_tail(ctx->net, trans);
6870
6871         return 0;
6872
6873 err_free_trans:
6874         kfree(trans);
6875 err_trans:
6876         module_put(type->owner);
6877         return err;
6878 }
6879
6880 static int nf_tables_newobj(struct sk_buff *skb, const struct nfnl_info *info,
6881                             const struct nlattr * const nla[])
6882 {
6883         struct netlink_ext_ack *extack = info->extack;
6884         u8 genmask = nft_genmask_next(info->net);
6885         u8 family = info->nfmsg->nfgen_family;
6886         const struct nft_object_type *type;
6887         struct net *net = info->net;
6888         struct nft_table *table;
6889         struct nft_object *obj;
6890         struct nft_ctx ctx;
6891         u32 objtype;
6892         int err;
6893
6894         if (!nla[NFTA_OBJ_TYPE] ||
6895             !nla[NFTA_OBJ_NAME] ||
6896             !nla[NFTA_OBJ_DATA])
6897                 return -EINVAL;
6898
6899         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
6900                                  NETLINK_CB(skb).portid);
6901         if (IS_ERR(table)) {
6902                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6903                 return PTR_ERR(table);
6904         }
6905
6906         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6907         obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6908         if (IS_ERR(obj)) {
6909                 err = PTR_ERR(obj);
6910                 if (err != -ENOENT) {
6911                         NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6912                         return err;
6913                 }
6914         } else {
6915                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
6916                         NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6917                         return -EEXIST;
6918                 }
6919                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
6920                         return -EOPNOTSUPP;
6921
6922                 type = __nft_obj_type_get(objtype);
6923                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6924
6925                 return nf_tables_updobj(&ctx, type, nla[NFTA_OBJ_DATA], obj);
6926         }
6927
6928         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6929
6930         type = nft_obj_type_get(net, objtype);
6931         if (IS_ERR(type))
6932                 return PTR_ERR(type);
6933
6934         obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
6935         if (IS_ERR(obj)) {
6936                 err = PTR_ERR(obj);
6937                 goto err_init;
6938         }
6939         obj->key.table = table;
6940         obj->handle = nf_tables_alloc_handle(table);
6941
6942         obj->key.name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL_ACCOUNT);
6943         if (!obj->key.name) {
6944                 err = -ENOMEM;
6945                 goto err_strdup;
6946         }
6947
6948         if (nla[NFTA_OBJ_USERDATA]) {
6949                 obj->udata = nla_memdup(nla[NFTA_OBJ_USERDATA], GFP_KERNEL);
6950                 if (obj->udata == NULL)
6951                         goto err_userdata;
6952
6953                 obj->udlen = nla_len(nla[NFTA_OBJ_USERDATA]);
6954         }
6955
6956         err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
6957         if (err < 0)
6958                 goto err_trans;
6959
6960         err = rhltable_insert(&nft_objname_ht, &obj->rhlhead,
6961                               nft_objname_ht_params);
6962         if (err < 0)
6963                 goto err_obj_ht;
6964
6965         list_add_tail_rcu(&obj->list, &table->objects);
6966         table->use++;
6967         return 0;
6968 err_obj_ht:
6969         /* queued in transaction log */
6970         INIT_LIST_HEAD(&obj->list);
6971         return err;
6972 err_trans:
6973         kfree(obj->udata);
6974 err_userdata:
6975         kfree(obj->key.name);
6976 err_strdup:
6977         if (obj->ops->destroy)
6978                 obj->ops->destroy(&ctx, obj);
6979         kfree(obj);
6980 err_init:
6981         module_put(type->owner);
6982         return err;
6983 }
6984
6985 static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
6986                                    u32 portid, u32 seq, int event, u32 flags,
6987                                    int family, const struct nft_table *table,
6988                                    struct nft_object *obj, bool reset)
6989 {
6990         struct nlmsghdr *nlh;
6991
6992         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
6993         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
6994                            NFNETLINK_V0, nft_base_seq(net));
6995         if (!nlh)
6996                 goto nla_put_failure;
6997
6998         if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
6999             nla_put_string(skb, NFTA_OBJ_NAME, obj->key.name) ||
7000             nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
7001             nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
7002             nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset) ||
7003             nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
7004                          NFTA_OBJ_PAD))
7005                 goto nla_put_failure;
7006
7007         if (obj->udata &&
7008             nla_put(skb, NFTA_OBJ_USERDATA, obj->udlen, obj->udata))
7009                 goto nla_put_failure;
7010
7011         nlmsg_end(skb, nlh);
7012         return 0;
7013
7014 nla_put_failure:
7015         nlmsg_trim(skb, nlh);
7016         return -1;
7017 }
7018
7019 struct nft_obj_filter {
7020         char            *table;
7021         u32             type;
7022 };
7023
7024 static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
7025 {
7026         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
7027         const struct nft_table *table;
7028         unsigned int idx = 0, s_idx = cb->args[0];
7029         struct nft_obj_filter *filter = cb->data;
7030         struct net *net = sock_net(skb->sk);
7031         int family = nfmsg->nfgen_family;
7032         struct nftables_pernet *nft_net;
7033         struct nft_object *obj;
7034         bool reset = false;
7035
7036         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
7037                 reset = true;
7038
7039         rcu_read_lock();
7040         nft_net = nft_pernet(net);
7041         cb->seq = READ_ONCE(nft_net->base_seq);
7042
7043         list_for_each_entry_rcu(table, &nft_net->tables, list) {
7044                 if (family != NFPROTO_UNSPEC && family != table->family)
7045                         continue;
7046
7047                 list_for_each_entry_rcu(obj, &table->objects, list) {
7048                         if (!nft_is_active(net, obj))
7049                                 goto cont;
7050                         if (idx < s_idx)
7051                                 goto cont;
7052                         if (idx > s_idx)
7053                                 memset(&cb->args[1], 0,
7054                                        sizeof(cb->args) - sizeof(cb->args[0]));
7055                         if (filter && filter->table &&
7056                             strcmp(filter->table, table->name))
7057                                 goto cont;
7058                         if (filter &&
7059                             filter->type != NFT_OBJECT_UNSPEC &&
7060                             obj->ops->type->type != filter->type)
7061                                 goto cont;
7062                         if (reset) {
7063                                 char *buf = kasprintf(GFP_ATOMIC,
7064                                                       "%s:%u",
7065                                                       table->name,
7066                                                       nft_net->base_seq);
7067
7068                                 audit_log_nfcfg(buf,
7069                                                 family,
7070                                                 obj->handle,
7071                                                 AUDIT_NFT_OP_OBJ_RESET,
7072                                                 GFP_ATOMIC);
7073                                 kfree(buf);
7074                         }
7075
7076                         if (nf_tables_fill_obj_info(skb, net, NETLINK_CB(cb->skb).portid,
7077                                                     cb->nlh->nlmsg_seq,
7078                                                     NFT_MSG_NEWOBJ,
7079                                                     NLM_F_MULTI | NLM_F_APPEND,
7080                                                     table->family, table,
7081                                                     obj, reset) < 0)
7082                                 goto done;
7083
7084                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
7085 cont:
7086                         idx++;
7087                 }
7088         }
7089 done:
7090         rcu_read_unlock();
7091
7092         cb->args[0] = idx;
7093         return skb->len;
7094 }
7095
7096 static int nf_tables_dump_obj_start(struct netlink_callback *cb)
7097 {
7098         const struct nlattr * const *nla = cb->data;
7099         struct nft_obj_filter *filter = NULL;
7100
7101         if (nla[NFTA_OBJ_TABLE] || nla[NFTA_OBJ_TYPE]) {
7102                 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
7103                 if (!filter)
7104                         return -ENOMEM;
7105
7106                 if (nla[NFTA_OBJ_TABLE]) {
7107                         filter->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
7108                         if (!filter->table) {
7109                                 kfree(filter);
7110                                 return -ENOMEM;
7111                         }
7112                 }
7113
7114                 if (nla[NFTA_OBJ_TYPE])
7115                         filter->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7116         }
7117
7118         cb->data = filter;
7119         return 0;
7120 }
7121
7122 static int nf_tables_dump_obj_done(struct netlink_callback *cb)
7123 {
7124         struct nft_obj_filter *filter = cb->data;
7125
7126         if (filter) {
7127                 kfree(filter->table);
7128                 kfree(filter);
7129         }
7130
7131         return 0;
7132 }
7133
7134 /* called with rcu_read_lock held */
7135 static int nf_tables_getobj(struct sk_buff *skb, const struct nfnl_info *info,
7136                             const struct nlattr * const nla[])
7137 {
7138         struct netlink_ext_ack *extack = info->extack;
7139         u8 genmask = nft_genmask_cur(info->net);
7140         u8 family = info->nfmsg->nfgen_family;
7141         const struct nft_table *table;
7142         struct net *net = info->net;
7143         struct nft_object *obj;
7144         struct sk_buff *skb2;
7145         bool reset = false;
7146         u32 objtype;
7147         int err;
7148
7149         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
7150                 struct netlink_dump_control c = {
7151                         .start = nf_tables_dump_obj_start,
7152                         .dump = nf_tables_dump_obj,
7153                         .done = nf_tables_dump_obj_done,
7154                         .module = THIS_MODULE,
7155                         .data = (void *)nla,
7156                 };
7157
7158                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
7159         }
7160
7161         if (!nla[NFTA_OBJ_NAME] ||
7162             !nla[NFTA_OBJ_TYPE])
7163                 return -EINVAL;
7164
7165         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask, 0);
7166         if (IS_ERR(table)) {
7167                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
7168                 return PTR_ERR(table);
7169         }
7170
7171         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7172         obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
7173         if (IS_ERR(obj)) {
7174                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
7175                 return PTR_ERR(obj);
7176         }
7177
7178         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7179         if (!skb2)
7180                 return -ENOMEM;
7181
7182         if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
7183                 reset = true;
7184
7185         if (reset) {
7186                 const struct nftables_pernet *nft_net;
7187                 char *buf;
7188
7189                 nft_net = nft_pernet(net);
7190                 buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, nft_net->base_seq);
7191
7192                 audit_log_nfcfg(buf,
7193                                 family,
7194                                 obj->handle,
7195                                 AUDIT_NFT_OP_OBJ_RESET,
7196                                 GFP_ATOMIC);
7197                 kfree(buf);
7198         }
7199
7200         err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
7201                                       info->nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
7202                                       family, table, obj, reset);
7203         if (err < 0)
7204                 goto err_fill_obj_info;
7205
7206         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7207
7208 err_fill_obj_info:
7209         kfree_skb(skb2);
7210         return err;
7211 }
7212
7213 static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
7214 {
7215         if (obj->ops->destroy)
7216                 obj->ops->destroy(ctx, obj);
7217
7218         module_put(obj->ops->type->owner);
7219         kfree(obj->key.name);
7220         kfree(obj->udata);
7221         kfree(obj);
7222 }
7223
7224 static int nf_tables_delobj(struct sk_buff *skb, const struct nfnl_info *info,
7225                             const struct nlattr * const nla[])
7226 {
7227         struct netlink_ext_ack *extack = info->extack;
7228         u8 genmask = nft_genmask_next(info->net);
7229         u8 family = info->nfmsg->nfgen_family;
7230         struct net *net = info->net;
7231         const struct nlattr *attr;
7232         struct nft_table *table;
7233         struct nft_object *obj;
7234         struct nft_ctx ctx;
7235         u32 objtype;
7236
7237         if (!nla[NFTA_OBJ_TYPE] ||
7238             (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
7239                 return -EINVAL;
7240
7241         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
7242                                  NETLINK_CB(skb).portid);
7243         if (IS_ERR(table)) {
7244                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
7245                 return PTR_ERR(table);
7246         }
7247
7248         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7249         if (nla[NFTA_OBJ_HANDLE]) {
7250                 attr = nla[NFTA_OBJ_HANDLE];
7251                 obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
7252         } else {
7253                 attr = nla[NFTA_OBJ_NAME];
7254                 obj = nft_obj_lookup(net, table, attr, objtype, genmask);
7255         }
7256
7257         if (IS_ERR(obj)) {
7258                 NL_SET_BAD_ATTR(extack, attr);
7259                 return PTR_ERR(obj);
7260         }
7261         if (obj->use > 0) {
7262                 NL_SET_BAD_ATTR(extack, attr);
7263                 return -EBUSY;
7264         }
7265
7266         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7267
7268         return nft_delobj(&ctx, obj);
7269 }
7270
7271 void nft_obj_notify(struct net *net, const struct nft_table *table,
7272                     struct nft_object *obj, u32 portid, u32 seq, int event,
7273                     u16 flags, int family, int report, gfp_t gfp)
7274 {
7275         struct nftables_pernet *nft_net = nft_pernet(net);
7276         struct sk_buff *skb;
7277         int err;
7278         char *buf = kasprintf(gfp, "%s:%u",
7279                               table->name, nft_net->base_seq);
7280
7281         audit_log_nfcfg(buf,
7282                         family,
7283                         obj->handle,
7284                         event == NFT_MSG_NEWOBJ ?
7285                                  AUDIT_NFT_OP_OBJ_REGISTER :
7286                                  AUDIT_NFT_OP_OBJ_UNREGISTER,
7287                         gfp);
7288         kfree(buf);
7289
7290         if (!report &&
7291             !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
7292                 return;
7293
7294         skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
7295         if (skb == NULL)
7296                 goto err;
7297
7298         err = nf_tables_fill_obj_info(skb, net, portid, seq, event,
7299                                       flags & (NLM_F_CREATE | NLM_F_EXCL),
7300                                       family, table, obj, false);
7301         if (err < 0) {
7302                 kfree_skb(skb);
7303                 goto err;
7304         }
7305
7306         nft_notify_enqueue(skb, report, &nft_net->notify_list);
7307         return;
7308 err:
7309         nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
7310 }
7311 EXPORT_SYMBOL_GPL(nft_obj_notify);
7312
7313 static void nf_tables_obj_notify(const struct nft_ctx *ctx,
7314                                  struct nft_object *obj, int event)
7315 {
7316         nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid, ctx->seq, event,
7317                        ctx->flags, ctx->family, ctx->report, GFP_KERNEL);
7318 }
7319
7320 /*
7321  * Flow tables
7322  */
7323 void nft_register_flowtable_type(struct nf_flowtable_type *type)
7324 {
7325         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7326         list_add_tail_rcu(&type->list, &nf_tables_flowtables);
7327         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7328 }
7329 EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
7330
7331 void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
7332 {
7333         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7334         list_del_rcu(&type->list);
7335         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7336 }
7337 EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
7338
7339 static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
7340         [NFTA_FLOWTABLE_TABLE]          = { .type = NLA_STRING,
7341                                             .len = NFT_NAME_MAXLEN - 1 },
7342         [NFTA_FLOWTABLE_NAME]           = { .type = NLA_STRING,
7343                                             .len = NFT_NAME_MAXLEN - 1 },
7344         [NFTA_FLOWTABLE_HOOK]           = { .type = NLA_NESTED },
7345         [NFTA_FLOWTABLE_HANDLE]         = { .type = NLA_U64 },
7346         [NFTA_FLOWTABLE_FLAGS]          = { .type = NLA_U32 },
7347 };
7348
7349 struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
7350                                            const struct nlattr *nla, u8 genmask)
7351 {
7352         struct nft_flowtable *flowtable;
7353
7354         list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7355                 if (!nla_strcmp(nla, flowtable->name) &&
7356                     nft_active_genmask(flowtable, genmask))
7357                         return flowtable;
7358         }
7359         return ERR_PTR(-ENOENT);
7360 }
7361 EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
7362
7363 void nf_tables_deactivate_flowtable(const struct nft_ctx *ctx,
7364                                     struct nft_flowtable *flowtable,
7365                                     enum nft_trans_phase phase)
7366 {
7367         switch (phase) {
7368         case NFT_TRANS_PREPARE:
7369         case NFT_TRANS_ABORT:
7370         case NFT_TRANS_RELEASE:
7371                 flowtable->use--;
7372                 fallthrough;
7373         default:
7374                 return;
7375         }
7376 }
7377 EXPORT_SYMBOL_GPL(nf_tables_deactivate_flowtable);
7378
7379 static struct nft_flowtable *
7380 nft_flowtable_lookup_byhandle(const struct nft_table *table,
7381                               const struct nlattr *nla, u8 genmask)
7382 {
7383        struct nft_flowtable *flowtable;
7384
7385        list_for_each_entry(flowtable, &table->flowtables, list) {
7386                if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
7387                    nft_active_genmask(flowtable, genmask))
7388                        return flowtable;
7389        }
7390        return ERR_PTR(-ENOENT);
7391 }
7392
7393 struct nft_flowtable_hook {
7394         u32                     num;
7395         int                     priority;
7396         struct list_head        list;
7397 };
7398
7399 static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
7400         [NFTA_FLOWTABLE_HOOK_NUM]       = { .type = NLA_U32 },
7401         [NFTA_FLOWTABLE_HOOK_PRIORITY]  = { .type = NLA_U32 },
7402         [NFTA_FLOWTABLE_HOOK_DEVS]      = { .type = NLA_NESTED },
7403 };
7404
7405 static int nft_flowtable_parse_hook(const struct nft_ctx *ctx,
7406                                     const struct nlattr *attr,
7407                                     struct nft_flowtable_hook *flowtable_hook,
7408                                     struct nft_flowtable *flowtable, bool add)
7409 {
7410         struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
7411         struct nft_hook *hook;
7412         int hooknum, priority;
7413         int err;
7414
7415         INIT_LIST_HEAD(&flowtable_hook->list);
7416
7417         err = nla_parse_nested_deprecated(tb, NFTA_FLOWTABLE_HOOK_MAX, attr,
7418                                           nft_flowtable_hook_policy, NULL);
7419         if (err < 0)
7420                 return err;
7421
7422         if (add) {
7423                 if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
7424                     !tb[NFTA_FLOWTABLE_HOOK_PRIORITY])
7425                         return -EINVAL;
7426
7427                 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7428                 if (hooknum != NF_NETDEV_INGRESS)
7429                         return -EOPNOTSUPP;
7430
7431                 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7432
7433                 flowtable_hook->priority        = priority;
7434                 flowtable_hook->num             = hooknum;
7435         } else {
7436                 if (tb[NFTA_FLOWTABLE_HOOK_NUM]) {
7437                         hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7438                         if (hooknum != flowtable->hooknum)
7439                                 return -EOPNOTSUPP;
7440                 }
7441
7442                 if (tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
7443                         priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7444                         if (priority != flowtable->data.priority)
7445                                 return -EOPNOTSUPP;
7446                 }
7447
7448                 flowtable_hook->priority        = flowtable->data.priority;
7449                 flowtable_hook->num             = flowtable->hooknum;
7450         }
7451
7452         if (tb[NFTA_FLOWTABLE_HOOK_DEVS]) {
7453                 err = nf_tables_parse_netdev_hooks(ctx->net,
7454                                                    tb[NFTA_FLOWTABLE_HOOK_DEVS],
7455                                                    &flowtable_hook->list);
7456                 if (err < 0)
7457                         return err;
7458         }
7459
7460         list_for_each_entry(hook, &flowtable_hook->list, list) {
7461                 hook->ops.pf            = NFPROTO_NETDEV;
7462                 hook->ops.hooknum       = flowtable_hook->num;
7463                 hook->ops.priority      = flowtable_hook->priority;
7464                 hook->ops.priv          = &flowtable->data;
7465                 hook->ops.hook          = flowtable->data.type->hook;
7466         }
7467
7468         return err;
7469 }
7470
7471 static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
7472 {
7473         const struct nf_flowtable_type *type;
7474
7475         list_for_each_entry(type, &nf_tables_flowtables, list) {
7476                 if (family == type->family)
7477                         return type;
7478         }
7479         return NULL;
7480 }
7481
7482 static const struct nf_flowtable_type *
7483 nft_flowtable_type_get(struct net *net, u8 family)
7484 {
7485         const struct nf_flowtable_type *type;
7486
7487         type = __nft_flowtable_type_get(family);
7488         if (type != NULL && try_module_get(type->owner))
7489                 return type;
7490
7491         lockdep_nfnl_nft_mutex_not_held();
7492 #ifdef CONFIG_MODULES
7493         if (type == NULL) {
7494                 if (nft_request_module(net, "nf-flowtable-%u", family) == -EAGAIN)
7495                         return ERR_PTR(-EAGAIN);
7496         }
7497 #endif
7498         return ERR_PTR(-ENOENT);
7499 }
7500
7501 /* Only called from error and netdev event paths. */
7502 static void nft_unregister_flowtable_hook(struct net *net,
7503                                           struct nft_flowtable *flowtable,
7504                                           struct nft_hook *hook)
7505 {
7506         nf_unregister_net_hook(net, &hook->ops);
7507         flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7508                                     FLOW_BLOCK_UNBIND);
7509 }
7510
7511 static void __nft_unregister_flowtable_net_hooks(struct net *net,
7512                                                  struct list_head *hook_list,
7513                                                  bool release_netdev)
7514 {
7515         struct nft_hook *hook, *next;
7516
7517         list_for_each_entry_safe(hook, next, hook_list, list) {
7518                 nf_unregister_net_hook(net, &hook->ops);
7519                 if (release_netdev) {
7520                         list_del(&hook->list);
7521                         kfree_rcu(hook, rcu);
7522                 }
7523         }
7524 }
7525
7526 static void nft_unregister_flowtable_net_hooks(struct net *net,
7527                                                struct list_head *hook_list)
7528 {
7529         __nft_unregister_flowtable_net_hooks(net, hook_list, false);
7530 }
7531
7532 static int nft_register_flowtable_net_hooks(struct net *net,
7533                                             struct nft_table *table,
7534                                             struct list_head *hook_list,
7535                                             struct nft_flowtable *flowtable)
7536 {
7537         struct nft_hook *hook, *hook2, *next;
7538         struct nft_flowtable *ft;
7539         int err, i = 0;
7540
7541         list_for_each_entry(hook, hook_list, list) {
7542                 list_for_each_entry(ft, &table->flowtables, list) {
7543                         if (!nft_is_active_next(net, ft))
7544                                 continue;
7545
7546                         list_for_each_entry(hook2, &ft->hook_list, list) {
7547                                 if (hook->ops.dev == hook2->ops.dev &&
7548                                     hook->ops.pf == hook2->ops.pf) {
7549                                         err = -EEXIST;
7550                                         goto err_unregister_net_hooks;
7551                                 }
7552                         }
7553                 }
7554
7555                 err = flowtable->data.type->setup(&flowtable->data,
7556                                                   hook->ops.dev,
7557                                                   FLOW_BLOCK_BIND);
7558                 if (err < 0)
7559                         goto err_unregister_net_hooks;
7560
7561                 err = nf_register_net_hook(net, &hook->ops);
7562                 if (err < 0) {
7563                         flowtable->data.type->setup(&flowtable->data,
7564                                                     hook->ops.dev,
7565                                                     FLOW_BLOCK_UNBIND);
7566                         goto err_unregister_net_hooks;
7567                 }
7568
7569                 i++;
7570         }
7571
7572         return 0;
7573
7574 err_unregister_net_hooks:
7575         list_for_each_entry_safe(hook, next, hook_list, list) {
7576                 if (i-- <= 0)
7577                         break;
7578
7579                 nft_unregister_flowtable_hook(net, flowtable, hook);
7580                 list_del_rcu(&hook->list);
7581                 kfree_rcu(hook, rcu);
7582         }
7583
7584         return err;
7585 }
7586
7587 static void nft_flowtable_hooks_destroy(struct list_head *hook_list)
7588 {
7589         struct nft_hook *hook, *next;
7590
7591         list_for_each_entry_safe(hook, next, hook_list, list) {
7592                 list_del_rcu(&hook->list);
7593                 kfree_rcu(hook, rcu);
7594         }
7595 }
7596
7597 static int nft_flowtable_update(struct nft_ctx *ctx, const struct nlmsghdr *nlh,
7598                                 struct nft_flowtable *flowtable)
7599 {
7600         const struct nlattr * const *nla = ctx->nla;
7601         struct nft_flowtable_hook flowtable_hook;
7602         struct nft_hook *hook, *next;
7603         struct nft_trans *trans;
7604         bool unregister = false;
7605         u32 flags;
7606         int err;
7607
7608         err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7609                                        &flowtable_hook, flowtable, false);
7610         if (err < 0)
7611                 return err;
7612
7613         list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7614                 if (nft_hook_list_find(&flowtable->hook_list, hook)) {
7615                         list_del(&hook->list);
7616                         kfree(hook);
7617                 }
7618         }
7619
7620         if (nla[NFTA_FLOWTABLE_FLAGS]) {
7621                 flags = ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7622                 if (flags & ~NFT_FLOWTABLE_MASK) {
7623                         err = -EOPNOTSUPP;
7624                         goto err_flowtable_update_hook;
7625                 }
7626                 if ((flowtable->data.flags & NFT_FLOWTABLE_HW_OFFLOAD) ^
7627                     (flags & NFT_FLOWTABLE_HW_OFFLOAD)) {
7628                         err = -EOPNOTSUPP;
7629                         goto err_flowtable_update_hook;
7630                 }
7631         } else {
7632                 flags = flowtable->data.flags;
7633         }
7634
7635         err = nft_register_flowtable_net_hooks(ctx->net, ctx->table,
7636                                                &flowtable_hook.list, flowtable);
7637         if (err < 0)
7638                 goto err_flowtable_update_hook;
7639
7640         trans = nft_trans_alloc(ctx, NFT_MSG_NEWFLOWTABLE,
7641                                 sizeof(struct nft_trans_flowtable));
7642         if (!trans) {
7643                 unregister = true;
7644                 err = -ENOMEM;
7645                 goto err_flowtable_update_hook;
7646         }
7647
7648         nft_trans_flowtable_flags(trans) = flags;
7649         nft_trans_flowtable(trans) = flowtable;
7650         nft_trans_flowtable_update(trans) = true;
7651         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7652         list_splice(&flowtable_hook.list, &nft_trans_flowtable_hooks(trans));
7653
7654         nft_trans_commit_list_add_tail(ctx->net, trans);
7655
7656         return 0;
7657
7658 err_flowtable_update_hook:
7659         list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7660                 if (unregister)
7661                         nft_unregister_flowtable_hook(ctx->net, flowtable, hook);
7662                 list_del_rcu(&hook->list);
7663                 kfree_rcu(hook, rcu);
7664         }
7665
7666         return err;
7667
7668 }
7669
7670 static int nf_tables_newflowtable(struct sk_buff *skb,
7671                                   const struct nfnl_info *info,
7672                                   const struct nlattr * const nla[])
7673 {
7674         struct netlink_ext_ack *extack = info->extack;
7675         struct nft_flowtable_hook flowtable_hook;
7676         u8 genmask = nft_genmask_next(info->net);
7677         u8 family = info->nfmsg->nfgen_family;
7678         const struct nf_flowtable_type *type;
7679         struct nft_flowtable *flowtable;
7680         struct nft_hook *hook, *next;
7681         struct net *net = info->net;
7682         struct nft_table *table;
7683         struct nft_ctx ctx;
7684         int err;
7685
7686         if (!nla[NFTA_FLOWTABLE_TABLE] ||
7687             !nla[NFTA_FLOWTABLE_NAME] ||
7688             !nla[NFTA_FLOWTABLE_HOOK])
7689                 return -EINVAL;
7690
7691         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7692                                  genmask, NETLINK_CB(skb).portid);
7693         if (IS_ERR(table)) {
7694                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7695                 return PTR_ERR(table);
7696         }
7697
7698         flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7699                                          genmask);
7700         if (IS_ERR(flowtable)) {
7701                 err = PTR_ERR(flowtable);
7702                 if (err != -ENOENT) {
7703                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7704                         return err;
7705                 }
7706         } else {
7707                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
7708                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7709                         return -EEXIST;
7710                 }
7711
7712                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7713
7714                 return nft_flowtable_update(&ctx, info->nlh, flowtable);
7715         }
7716
7717         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7718
7719         flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL_ACCOUNT);
7720         if (!flowtable)
7721                 return -ENOMEM;
7722
7723         flowtable->table = table;
7724         flowtable->handle = nf_tables_alloc_handle(table);
7725         INIT_LIST_HEAD(&flowtable->hook_list);
7726
7727         flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL_ACCOUNT);
7728         if (!flowtable->name) {
7729                 err = -ENOMEM;
7730                 goto err1;
7731         }
7732
7733         type = nft_flowtable_type_get(net, family);
7734         if (IS_ERR(type)) {
7735                 err = PTR_ERR(type);
7736                 goto err2;
7737         }
7738
7739         if (nla[NFTA_FLOWTABLE_FLAGS]) {
7740                 flowtable->data.flags =
7741                         ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7742                 if (flowtable->data.flags & ~NFT_FLOWTABLE_MASK) {
7743                         err = -EOPNOTSUPP;
7744                         goto err3;
7745                 }
7746         }
7747
7748         write_pnet(&flowtable->data.net, net);
7749         flowtable->data.type = type;
7750         err = type->init(&flowtable->data);
7751         if (err < 0)
7752                 goto err3;
7753
7754         err = nft_flowtable_parse_hook(&ctx, nla[NFTA_FLOWTABLE_HOOK],
7755                                        &flowtable_hook, flowtable, true);
7756         if (err < 0)
7757                 goto err4;
7758
7759         list_splice(&flowtable_hook.list, &flowtable->hook_list);
7760         flowtable->data.priority = flowtable_hook.priority;
7761         flowtable->hooknum = flowtable_hook.num;
7762
7763         err = nft_register_flowtable_net_hooks(ctx.net, table,
7764                                                &flowtable->hook_list,
7765                                                flowtable);
7766         if (err < 0) {
7767                 nft_flowtable_hooks_destroy(&flowtable->hook_list);
7768                 goto err4;
7769         }
7770
7771         err = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
7772         if (err < 0)
7773                 goto err5;
7774
7775         list_add_tail_rcu(&flowtable->list, &table->flowtables);
7776         table->use++;
7777
7778         return 0;
7779 err5:
7780         list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7781                 nft_unregister_flowtable_hook(net, flowtable, hook);
7782                 list_del_rcu(&hook->list);
7783                 kfree_rcu(hook, rcu);
7784         }
7785 err4:
7786         flowtable->data.type->free(&flowtable->data);
7787 err3:
7788         module_put(type->owner);
7789 err2:
7790         kfree(flowtable->name);
7791 err1:
7792         kfree(flowtable);
7793         return err;
7794 }
7795
7796 static void nft_flowtable_hook_release(struct nft_flowtable_hook *flowtable_hook)
7797 {
7798         struct nft_hook *this, *next;
7799
7800         list_for_each_entry_safe(this, next, &flowtable_hook->list, list) {
7801                 list_del(&this->list);
7802                 kfree(this);
7803         }
7804 }
7805
7806 static int nft_delflowtable_hook(struct nft_ctx *ctx,
7807                                  struct nft_flowtable *flowtable)
7808 {
7809         const struct nlattr * const *nla = ctx->nla;
7810         struct nft_flowtable_hook flowtable_hook;
7811         LIST_HEAD(flowtable_del_list);
7812         struct nft_hook *this, *hook;
7813         struct nft_trans *trans;
7814         int err;
7815
7816         err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7817                                        &flowtable_hook, flowtable, false);
7818         if (err < 0)
7819                 return err;
7820
7821         list_for_each_entry(this, &flowtable_hook.list, list) {
7822                 hook = nft_hook_list_find(&flowtable->hook_list, this);
7823                 if (!hook) {
7824                         err = -ENOENT;
7825                         goto err_flowtable_del_hook;
7826                 }
7827                 list_move(&hook->list, &flowtable_del_list);
7828         }
7829
7830         trans = nft_trans_alloc(ctx, NFT_MSG_DELFLOWTABLE,
7831                                 sizeof(struct nft_trans_flowtable));
7832         if (!trans) {
7833                 err = -ENOMEM;
7834                 goto err_flowtable_del_hook;
7835         }
7836
7837         nft_trans_flowtable(trans) = flowtable;
7838         nft_trans_flowtable_update(trans) = true;
7839         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7840         list_splice(&flowtable_del_list, &nft_trans_flowtable_hooks(trans));
7841         nft_flowtable_hook_release(&flowtable_hook);
7842
7843         nft_trans_commit_list_add_tail(ctx->net, trans);
7844
7845         return 0;
7846
7847 err_flowtable_del_hook:
7848         list_splice(&flowtable_del_list, &flowtable->hook_list);
7849         nft_flowtable_hook_release(&flowtable_hook);
7850
7851         return err;
7852 }
7853
7854 static int nf_tables_delflowtable(struct sk_buff *skb,
7855                                   const struct nfnl_info *info,
7856                                   const struct nlattr * const nla[])
7857 {
7858         struct netlink_ext_ack *extack = info->extack;
7859         u8 genmask = nft_genmask_next(info->net);
7860         u8 family = info->nfmsg->nfgen_family;
7861         struct nft_flowtable *flowtable;
7862         struct net *net = info->net;
7863         const struct nlattr *attr;
7864         struct nft_table *table;
7865         struct nft_ctx ctx;
7866
7867         if (!nla[NFTA_FLOWTABLE_TABLE] ||
7868             (!nla[NFTA_FLOWTABLE_NAME] &&
7869              !nla[NFTA_FLOWTABLE_HANDLE]))
7870                 return -EINVAL;
7871
7872         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7873                                  genmask, NETLINK_CB(skb).portid);
7874         if (IS_ERR(table)) {
7875                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7876                 return PTR_ERR(table);
7877         }
7878
7879         if (nla[NFTA_FLOWTABLE_HANDLE]) {
7880                 attr = nla[NFTA_FLOWTABLE_HANDLE];
7881                 flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
7882         } else {
7883                 attr = nla[NFTA_FLOWTABLE_NAME];
7884                 flowtable = nft_flowtable_lookup(table, attr, genmask);
7885         }
7886
7887         if (IS_ERR(flowtable)) {
7888                 NL_SET_BAD_ATTR(extack, attr);
7889                 return PTR_ERR(flowtable);
7890         }
7891
7892         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7893
7894         if (nla[NFTA_FLOWTABLE_HOOK])
7895                 return nft_delflowtable_hook(&ctx, flowtable);
7896
7897         if (flowtable->use > 0) {
7898                 NL_SET_BAD_ATTR(extack, attr);
7899                 return -EBUSY;
7900         }
7901
7902         return nft_delflowtable(&ctx, flowtable);
7903 }
7904
7905 static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
7906                                          u32 portid, u32 seq, int event,
7907                                          u32 flags, int family,
7908                                          struct nft_flowtable *flowtable,
7909                                          struct list_head *hook_list)
7910 {
7911         struct nlattr *nest, *nest_devs;
7912         struct nft_hook *hook;
7913         struct nlmsghdr *nlh;
7914
7915         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
7916         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
7917                            NFNETLINK_V0, nft_base_seq(net));
7918         if (!nlh)
7919                 goto nla_put_failure;
7920
7921         if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
7922             nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
7923             nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
7924             nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
7925                          NFTA_FLOWTABLE_PAD) ||
7926             nla_put_be32(skb, NFTA_FLOWTABLE_FLAGS, htonl(flowtable->data.flags)))
7927                 goto nla_put_failure;
7928
7929         nest = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK);
7930         if (!nest)
7931                 goto nla_put_failure;
7932         if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
7933             nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->data.priority)))
7934                 goto nla_put_failure;
7935
7936         nest_devs = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK_DEVS);
7937         if (!nest_devs)
7938                 goto nla_put_failure;
7939
7940         list_for_each_entry_rcu(hook, hook_list, list) {
7941                 if (nla_put_string(skb, NFTA_DEVICE_NAME, hook->ops.dev->name))
7942                         goto nla_put_failure;
7943         }
7944         nla_nest_end(skb, nest_devs);
7945         nla_nest_end(skb, nest);
7946
7947         nlmsg_end(skb, nlh);
7948         return 0;
7949
7950 nla_put_failure:
7951         nlmsg_trim(skb, nlh);
7952         return -1;
7953 }
7954
7955 struct nft_flowtable_filter {
7956         char            *table;
7957 };
7958
7959 static int nf_tables_dump_flowtable(struct sk_buff *skb,
7960                                     struct netlink_callback *cb)
7961 {
7962         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
7963         struct nft_flowtable_filter *filter = cb->data;
7964         unsigned int idx = 0, s_idx = cb->args[0];
7965         struct net *net = sock_net(skb->sk);
7966         int family = nfmsg->nfgen_family;
7967         struct nft_flowtable *flowtable;
7968         struct nftables_pernet *nft_net;
7969         const struct nft_table *table;
7970
7971         rcu_read_lock();
7972         nft_net = nft_pernet(net);
7973         cb->seq = READ_ONCE(nft_net->base_seq);
7974
7975         list_for_each_entry_rcu(table, &nft_net->tables, list) {
7976                 if (family != NFPROTO_UNSPEC && family != table->family)
7977                         continue;
7978
7979                 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7980                         if (!nft_is_active(net, flowtable))
7981                                 goto cont;
7982                         if (idx < s_idx)
7983                                 goto cont;
7984                         if (idx > s_idx)
7985                                 memset(&cb->args[1], 0,
7986                                        sizeof(cb->args) - sizeof(cb->args[0]));
7987                         if (filter && filter->table &&
7988                             strcmp(filter->table, table->name))
7989                                 goto cont;
7990
7991                         if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
7992                                                           cb->nlh->nlmsg_seq,
7993                                                           NFT_MSG_NEWFLOWTABLE,
7994                                                           NLM_F_MULTI | NLM_F_APPEND,
7995                                                           table->family,
7996                                                           flowtable,
7997                                                           &flowtable->hook_list) < 0)
7998                                 goto done;
7999
8000                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
8001 cont:
8002                         idx++;
8003                 }
8004         }
8005 done:
8006         rcu_read_unlock();
8007
8008         cb->args[0] = idx;
8009         return skb->len;
8010 }
8011
8012 static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
8013 {
8014         const struct nlattr * const *nla = cb->data;
8015         struct nft_flowtable_filter *filter = NULL;
8016
8017         if (nla[NFTA_FLOWTABLE_TABLE]) {
8018                 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
8019                 if (!filter)
8020                         return -ENOMEM;
8021
8022                 filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
8023                                            GFP_ATOMIC);
8024                 if (!filter->table) {
8025                         kfree(filter);
8026                         return -ENOMEM;
8027                 }
8028         }
8029
8030         cb->data = filter;
8031         return 0;
8032 }
8033
8034 static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
8035 {
8036         struct nft_flowtable_filter *filter = cb->data;
8037
8038         if (!filter)
8039                 return 0;
8040
8041         kfree(filter->table);
8042         kfree(filter);
8043
8044         return 0;
8045 }
8046
8047 /* called with rcu_read_lock held */
8048 static int nf_tables_getflowtable(struct sk_buff *skb,
8049                                   const struct nfnl_info *info,
8050                                   const struct nlattr * const nla[])
8051 {
8052         u8 genmask = nft_genmask_cur(info->net);
8053         u8 family = info->nfmsg->nfgen_family;
8054         struct nft_flowtable *flowtable;
8055         const struct nft_table *table;
8056         struct net *net = info->net;
8057         struct sk_buff *skb2;
8058         int err;
8059
8060         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
8061                 struct netlink_dump_control c = {
8062                         .start = nf_tables_dump_flowtable_start,
8063                         .dump = nf_tables_dump_flowtable,
8064                         .done = nf_tables_dump_flowtable_done,
8065                         .module = THIS_MODULE,
8066                         .data = (void *)nla,
8067                 };
8068
8069                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
8070         }
8071
8072         if (!nla[NFTA_FLOWTABLE_NAME])
8073                 return -EINVAL;
8074
8075         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
8076                                  genmask, 0);
8077         if (IS_ERR(table))
8078                 return PTR_ERR(table);
8079
8080         flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
8081                                          genmask);
8082         if (IS_ERR(flowtable))
8083                 return PTR_ERR(flowtable);
8084
8085         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
8086         if (!skb2)
8087                 return -ENOMEM;
8088
8089         err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
8090                                             info->nlh->nlmsg_seq,
8091                                             NFT_MSG_NEWFLOWTABLE, 0, family,
8092                                             flowtable, &flowtable->hook_list);
8093         if (err < 0)
8094                 goto err_fill_flowtable_info;
8095
8096         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
8097
8098 err_fill_flowtable_info:
8099         kfree_skb(skb2);
8100         return err;
8101 }
8102
8103 static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
8104                                        struct nft_flowtable *flowtable,
8105                                        struct list_head *hook_list,
8106                                        int event)
8107 {
8108         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
8109         struct sk_buff *skb;
8110         u16 flags = 0;
8111         int err;
8112
8113         if (!ctx->report &&
8114             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
8115                 return;
8116
8117         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
8118         if (skb == NULL)
8119                 goto err;
8120
8121         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
8122                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
8123
8124         err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
8125                                             ctx->seq, event, flags,
8126                                             ctx->family, flowtable, hook_list);
8127         if (err < 0) {
8128                 kfree_skb(skb);
8129                 goto err;
8130         }
8131
8132         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
8133         return;
8134 err:
8135         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
8136 }
8137
8138 static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
8139 {
8140         struct nft_hook *hook, *next;
8141
8142         flowtable->data.type->free(&flowtable->data);
8143         list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
8144                 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
8145                                             FLOW_BLOCK_UNBIND);
8146                 list_del_rcu(&hook->list);
8147                 kfree(hook);
8148         }
8149         kfree(flowtable->name);
8150         module_put(flowtable->data.type->owner);
8151         kfree(flowtable);
8152 }
8153
8154 static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
8155                                    u32 portid, u32 seq)
8156 {
8157         struct nftables_pernet *nft_net = nft_pernet(net);
8158         struct nlmsghdr *nlh;
8159         char buf[TASK_COMM_LEN];
8160         int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
8161
8162         nlh = nfnl_msg_put(skb, portid, seq, event, 0, AF_UNSPEC,
8163                            NFNETLINK_V0, nft_base_seq(net));
8164         if (!nlh)
8165                 goto nla_put_failure;
8166
8167         if (nla_put_be32(skb, NFTA_GEN_ID, htonl(nft_net->base_seq)) ||
8168             nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
8169             nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
8170                 goto nla_put_failure;
8171
8172         nlmsg_end(skb, nlh);
8173         return 0;
8174
8175 nla_put_failure:
8176         nlmsg_trim(skb, nlh);
8177         return -EMSGSIZE;
8178 }
8179
8180 static void nft_flowtable_event(unsigned long event, struct net_device *dev,
8181                                 struct nft_flowtable *flowtable)
8182 {
8183         struct nft_hook *hook;
8184
8185         list_for_each_entry(hook, &flowtable->hook_list, list) {
8186                 if (hook->ops.dev != dev)
8187                         continue;
8188
8189                 /* flow_offload_netdev_event() cleans up entries for us. */
8190                 nft_unregister_flowtable_hook(dev_net(dev), flowtable, hook);
8191                 list_del_rcu(&hook->list);
8192                 kfree_rcu(hook, rcu);
8193                 break;
8194         }
8195 }
8196
8197 static int nf_tables_flowtable_event(struct notifier_block *this,
8198                                      unsigned long event, void *ptr)
8199 {
8200         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
8201         struct nft_flowtable *flowtable;
8202         struct nftables_pernet *nft_net;
8203         struct nft_table *table;
8204         struct net *net;
8205
8206         if (event != NETDEV_UNREGISTER)
8207                 return 0;
8208
8209         net = dev_net(dev);
8210         nft_net = nft_pernet(net);
8211         mutex_lock(&nft_net->commit_mutex);
8212         list_for_each_entry(table, &nft_net->tables, list) {
8213                 list_for_each_entry(flowtable, &table->flowtables, list) {
8214                         nft_flowtable_event(event, dev, flowtable);
8215                 }
8216         }
8217         mutex_unlock(&nft_net->commit_mutex);
8218
8219         return NOTIFY_DONE;
8220 }
8221
8222 static struct notifier_block nf_tables_flowtable_notifier = {
8223         .notifier_call  = nf_tables_flowtable_event,
8224 };
8225
8226 static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
8227                                  int event)
8228 {
8229         struct nlmsghdr *nlh = nlmsg_hdr(skb);
8230         struct sk_buff *skb2;
8231         int err;
8232
8233         if (!nlmsg_report(nlh) &&
8234             !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
8235                 return;
8236
8237         skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
8238         if (skb2 == NULL)
8239                 goto err;
8240
8241         err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
8242                                       nlh->nlmsg_seq);
8243         if (err < 0) {
8244                 kfree_skb(skb2);
8245                 goto err;
8246         }
8247
8248         nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
8249                        nlmsg_report(nlh), GFP_KERNEL);
8250         return;
8251 err:
8252         nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
8253                           -ENOBUFS);
8254 }
8255
8256 static int nf_tables_getgen(struct sk_buff *skb, const struct nfnl_info *info,
8257                             const struct nlattr * const nla[])
8258 {
8259         struct sk_buff *skb2;
8260         int err;
8261
8262         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
8263         if (skb2 == NULL)
8264                 return -ENOMEM;
8265
8266         err = nf_tables_fill_gen_info(skb2, info->net, NETLINK_CB(skb).portid,
8267                                       info->nlh->nlmsg_seq);
8268         if (err < 0)
8269                 goto err_fill_gen_info;
8270
8271         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
8272
8273 err_fill_gen_info:
8274         kfree_skb(skb2);
8275         return err;
8276 }
8277
8278 static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
8279         [NFT_MSG_NEWTABLE] = {
8280                 .call           = nf_tables_newtable,
8281                 .type           = NFNL_CB_BATCH,
8282                 .attr_count     = NFTA_TABLE_MAX,
8283                 .policy         = nft_table_policy,
8284         },
8285         [NFT_MSG_GETTABLE] = {
8286                 .call           = nf_tables_gettable,
8287                 .type           = NFNL_CB_RCU,
8288                 .attr_count     = NFTA_TABLE_MAX,
8289                 .policy         = nft_table_policy,
8290         },
8291         [NFT_MSG_DELTABLE] = {
8292                 .call           = nf_tables_deltable,
8293                 .type           = NFNL_CB_BATCH,
8294                 .attr_count     = NFTA_TABLE_MAX,
8295                 .policy         = nft_table_policy,
8296         },
8297         [NFT_MSG_NEWCHAIN] = {
8298                 .call           = nf_tables_newchain,
8299                 .type           = NFNL_CB_BATCH,
8300                 .attr_count     = NFTA_CHAIN_MAX,
8301                 .policy         = nft_chain_policy,
8302         },
8303         [NFT_MSG_GETCHAIN] = {
8304                 .call           = nf_tables_getchain,
8305                 .type           = NFNL_CB_RCU,
8306                 .attr_count     = NFTA_CHAIN_MAX,
8307                 .policy         = nft_chain_policy,
8308         },
8309         [NFT_MSG_DELCHAIN] = {
8310                 .call           = nf_tables_delchain,
8311                 .type           = NFNL_CB_BATCH,
8312                 .attr_count     = NFTA_CHAIN_MAX,
8313                 .policy         = nft_chain_policy,
8314         },
8315         [NFT_MSG_NEWRULE] = {
8316                 .call           = nf_tables_newrule,
8317                 .type           = NFNL_CB_BATCH,
8318                 .attr_count     = NFTA_RULE_MAX,
8319                 .policy         = nft_rule_policy,
8320         },
8321         [NFT_MSG_GETRULE] = {
8322                 .call           = nf_tables_getrule,
8323                 .type           = NFNL_CB_RCU,
8324                 .attr_count     = NFTA_RULE_MAX,
8325                 .policy         = nft_rule_policy,
8326         },
8327         [NFT_MSG_GETRULE_RESET] = {
8328                 .call           = nf_tables_getrule,
8329                 .type           = NFNL_CB_RCU,
8330                 .attr_count     = NFTA_RULE_MAX,
8331                 .policy         = nft_rule_policy,
8332         },
8333         [NFT_MSG_DELRULE] = {
8334                 .call           = nf_tables_delrule,
8335                 .type           = NFNL_CB_BATCH,
8336                 .attr_count     = NFTA_RULE_MAX,
8337                 .policy         = nft_rule_policy,
8338         },
8339         [NFT_MSG_NEWSET] = {
8340                 .call           = nf_tables_newset,
8341                 .type           = NFNL_CB_BATCH,
8342                 .attr_count     = NFTA_SET_MAX,
8343                 .policy         = nft_set_policy,
8344         },
8345         [NFT_MSG_GETSET] = {
8346                 .call           = nf_tables_getset,
8347                 .type           = NFNL_CB_RCU,
8348                 .attr_count     = NFTA_SET_MAX,
8349                 .policy         = nft_set_policy,
8350         },
8351         [NFT_MSG_DELSET] = {
8352                 .call           = nf_tables_delset,
8353                 .type           = NFNL_CB_BATCH,
8354                 .attr_count     = NFTA_SET_MAX,
8355                 .policy         = nft_set_policy,
8356         },
8357         [NFT_MSG_NEWSETELEM] = {
8358                 .call           = nf_tables_newsetelem,
8359                 .type           = NFNL_CB_BATCH,
8360                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8361                 .policy         = nft_set_elem_list_policy,
8362         },
8363         [NFT_MSG_GETSETELEM] = {
8364                 .call           = nf_tables_getsetelem,
8365                 .type           = NFNL_CB_RCU,
8366                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8367                 .policy         = nft_set_elem_list_policy,
8368         },
8369         [NFT_MSG_DELSETELEM] = {
8370                 .call           = nf_tables_delsetelem,
8371                 .type           = NFNL_CB_BATCH,
8372                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
8373                 .policy         = nft_set_elem_list_policy,
8374         },
8375         [NFT_MSG_GETGEN] = {
8376                 .call           = nf_tables_getgen,
8377                 .type           = NFNL_CB_RCU,
8378         },
8379         [NFT_MSG_NEWOBJ] = {
8380                 .call           = nf_tables_newobj,
8381                 .type           = NFNL_CB_BATCH,
8382                 .attr_count     = NFTA_OBJ_MAX,
8383                 .policy         = nft_obj_policy,
8384         },
8385         [NFT_MSG_GETOBJ] = {
8386                 .call           = nf_tables_getobj,
8387                 .type           = NFNL_CB_RCU,
8388                 .attr_count     = NFTA_OBJ_MAX,
8389                 .policy         = nft_obj_policy,
8390         },
8391         [NFT_MSG_DELOBJ] = {
8392                 .call           = nf_tables_delobj,
8393                 .type           = NFNL_CB_BATCH,
8394                 .attr_count     = NFTA_OBJ_MAX,
8395                 .policy         = nft_obj_policy,
8396         },
8397         [NFT_MSG_GETOBJ_RESET] = {
8398                 .call           = nf_tables_getobj,
8399                 .type           = NFNL_CB_RCU,
8400                 .attr_count     = NFTA_OBJ_MAX,
8401                 .policy         = nft_obj_policy,
8402         },
8403         [NFT_MSG_NEWFLOWTABLE] = {
8404                 .call           = nf_tables_newflowtable,
8405                 .type           = NFNL_CB_BATCH,
8406                 .attr_count     = NFTA_FLOWTABLE_MAX,
8407                 .policy         = nft_flowtable_policy,
8408         },
8409         [NFT_MSG_GETFLOWTABLE] = {
8410                 .call           = nf_tables_getflowtable,
8411                 .type           = NFNL_CB_RCU,
8412                 .attr_count     = NFTA_FLOWTABLE_MAX,
8413                 .policy         = nft_flowtable_policy,
8414         },
8415         [NFT_MSG_DELFLOWTABLE] = {
8416                 .call           = nf_tables_delflowtable,
8417                 .type           = NFNL_CB_BATCH,
8418                 .attr_count     = NFTA_FLOWTABLE_MAX,
8419                 .policy         = nft_flowtable_policy,
8420         },
8421 };
8422
8423 static int nf_tables_validate(struct net *net)
8424 {
8425         struct nftables_pernet *nft_net = nft_pernet(net);
8426         struct nft_table *table;
8427
8428         switch (nft_net->validate_state) {
8429         case NFT_VALIDATE_SKIP:
8430                 break;
8431         case NFT_VALIDATE_NEED:
8432                 nft_validate_state_update(net, NFT_VALIDATE_DO);
8433                 fallthrough;
8434         case NFT_VALIDATE_DO:
8435                 list_for_each_entry(table, &nft_net->tables, list) {
8436                         if (nft_table_validate(net, table) < 0)
8437                                 return -EAGAIN;
8438                 }
8439                 break;
8440         }
8441
8442         return 0;
8443 }
8444
8445 /* a drop policy has to be deferred until all rules have been activated,
8446  * otherwise a large ruleset that contains a drop-policy base chain will
8447  * cause all packets to get dropped until the full transaction has been
8448  * processed.
8449  *
8450  * We defer the drop policy until the transaction has been finalized.
8451  */
8452 static void nft_chain_commit_drop_policy(struct nft_trans *trans)
8453 {
8454         struct nft_base_chain *basechain;
8455
8456         if (nft_trans_chain_policy(trans) != NF_DROP)
8457                 return;
8458
8459         if (!nft_is_base_chain(trans->ctx.chain))
8460                 return;
8461
8462         basechain = nft_base_chain(trans->ctx.chain);
8463         basechain->policy = NF_DROP;
8464 }
8465
8466 static void nft_chain_commit_update(struct nft_trans *trans)
8467 {
8468         struct nft_base_chain *basechain;
8469
8470         if (nft_trans_chain_name(trans)) {
8471                 rhltable_remove(&trans->ctx.table->chains_ht,
8472                                 &trans->ctx.chain->rhlhead,
8473                                 nft_chain_ht_params);
8474                 swap(trans->ctx.chain->name, nft_trans_chain_name(trans));
8475                 rhltable_insert_key(&trans->ctx.table->chains_ht,
8476                                     trans->ctx.chain->name,
8477                                     &trans->ctx.chain->rhlhead,
8478                                     nft_chain_ht_params);
8479         }
8480
8481         if (!nft_is_base_chain(trans->ctx.chain))
8482                 return;
8483
8484         nft_chain_stats_replace(trans);
8485
8486         basechain = nft_base_chain(trans->ctx.chain);
8487
8488         switch (nft_trans_chain_policy(trans)) {
8489         case NF_DROP:
8490         case NF_ACCEPT:
8491                 basechain->policy = nft_trans_chain_policy(trans);
8492                 break;
8493         }
8494 }
8495
8496 static void nft_obj_commit_update(struct nft_trans *trans)
8497 {
8498         struct nft_object *newobj;
8499         struct nft_object *obj;
8500
8501         obj = nft_trans_obj(trans);
8502         newobj = nft_trans_obj_newobj(trans);
8503
8504         if (obj->ops->update)
8505                 obj->ops->update(obj, newobj);
8506
8507         nft_obj_destroy(&trans->ctx, newobj);
8508 }
8509
8510 static void nft_commit_release(struct nft_trans *trans)
8511 {
8512         switch (trans->msg_type) {
8513         case NFT_MSG_DELTABLE:
8514                 nf_tables_table_destroy(&trans->ctx);
8515                 break;
8516         case NFT_MSG_NEWCHAIN:
8517                 free_percpu(nft_trans_chain_stats(trans));
8518                 kfree(nft_trans_chain_name(trans));
8519                 break;
8520         case NFT_MSG_DELCHAIN:
8521                 nf_tables_chain_destroy(&trans->ctx);
8522                 break;
8523         case NFT_MSG_DELRULE:
8524                 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8525                 break;
8526         case NFT_MSG_DELSET:
8527                 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8528                 break;
8529         case NFT_MSG_DELSETELEM:
8530                 nf_tables_set_elem_destroy(&trans->ctx,
8531                                            nft_trans_elem_set(trans),
8532                                            nft_trans_elem(trans).priv);
8533                 break;
8534         case NFT_MSG_DELOBJ:
8535                 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8536                 break;
8537         case NFT_MSG_DELFLOWTABLE:
8538                 if (nft_trans_flowtable_update(trans))
8539                         nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8540                 else
8541                         nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8542                 break;
8543         }
8544
8545         if (trans->put_net)
8546                 put_net(trans->ctx.net);
8547
8548         kfree(trans);
8549 }
8550
8551 static void nf_tables_trans_destroy_work(struct work_struct *w)
8552 {
8553         struct nft_trans *trans, *next;
8554         LIST_HEAD(head);
8555
8556         spin_lock(&nf_tables_destroy_list_lock);
8557         list_splice_init(&nf_tables_destroy_list, &head);
8558         spin_unlock(&nf_tables_destroy_list_lock);
8559
8560         if (list_empty(&head))
8561                 return;
8562
8563         synchronize_rcu();
8564
8565         list_for_each_entry_safe(trans, next, &head, list) {
8566                 list_del(&trans->list);
8567                 nft_commit_release(trans);
8568         }
8569 }
8570
8571 void nf_tables_trans_destroy_flush_work(void)
8572 {
8573         flush_work(&trans_destroy_work);
8574 }
8575 EXPORT_SYMBOL_GPL(nf_tables_trans_destroy_flush_work);
8576
8577 static bool nft_expr_reduce(struct nft_regs_track *track,
8578                             const struct nft_expr *expr)
8579 {
8580         return false;
8581 }
8582
8583 static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
8584 {
8585         const struct nft_expr *expr, *last;
8586         struct nft_regs_track track = {};
8587         unsigned int size, data_size;
8588         void *data, *data_boundary;
8589         struct nft_rule_dp *prule;
8590         struct nft_rule *rule;
8591
8592         /* already handled or inactive chain? */
8593         if (chain->blob_next || !nft_is_active_next(net, chain))
8594                 return 0;
8595
8596         data_size = 0;
8597         list_for_each_entry(rule, &chain->rules, list) {
8598                 if (nft_is_active_next(net, rule)) {
8599                         data_size += sizeof(*prule) + rule->dlen;
8600                         if (data_size > INT_MAX)
8601                                 return -ENOMEM;
8602                 }
8603         }
8604         data_size += offsetof(struct nft_rule_dp, data);        /* last rule */
8605
8606         chain->blob_next = nf_tables_chain_alloc_rules(data_size);
8607         if (!chain->blob_next)
8608                 return -ENOMEM;
8609
8610         data = (void *)chain->blob_next->data;
8611         data_boundary = data + data_size;
8612         size = 0;
8613
8614         list_for_each_entry(rule, &chain->rules, list) {
8615                 if (!nft_is_active_next(net, rule))
8616                         continue;
8617
8618                 prule = (struct nft_rule_dp *)data;
8619                 data += offsetof(struct nft_rule_dp, data);
8620                 if (WARN_ON_ONCE(data > data_boundary))
8621                         return -ENOMEM;
8622
8623                 size = 0;
8624                 track.last = nft_expr_last(rule);
8625                 nft_rule_for_each_expr(expr, last, rule) {
8626                         track.cur = expr;
8627
8628                         if (nft_expr_reduce(&track, expr)) {
8629                                 expr = track.cur;
8630                                 continue;
8631                         }
8632
8633                         if (WARN_ON_ONCE(data + expr->ops->size > data_boundary))
8634                                 return -ENOMEM;
8635
8636                         memcpy(data + size, expr, expr->ops->size);
8637                         size += expr->ops->size;
8638                 }
8639                 if (WARN_ON_ONCE(size >= 1 << 12))
8640                         return -ENOMEM;
8641
8642                 prule->handle = rule->handle;
8643                 prule->dlen = size;
8644                 prule->is_last = 0;
8645
8646                 data += size;
8647                 size = 0;
8648                 chain->blob_next->size += (unsigned long)(data - (void *)prule);
8649         }
8650
8651         prule = (struct nft_rule_dp *)data;
8652         data += offsetof(struct nft_rule_dp, data);
8653         if (WARN_ON_ONCE(data > data_boundary))
8654                 return -ENOMEM;
8655
8656         nft_last_rule(chain->blob_next, prule);
8657
8658         return 0;
8659 }
8660
8661 static void nf_tables_commit_chain_prepare_cancel(struct net *net)
8662 {
8663         struct nftables_pernet *nft_net = nft_pernet(net);
8664         struct nft_trans *trans, *next;
8665
8666         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8667                 struct nft_chain *chain = trans->ctx.chain;
8668
8669                 if (trans->msg_type == NFT_MSG_NEWRULE ||
8670                     trans->msg_type == NFT_MSG_DELRULE) {
8671                         kvfree(chain->blob_next);
8672                         chain->blob_next = NULL;
8673                 }
8674         }
8675 }
8676
8677 static void __nf_tables_commit_chain_free_rules_old(struct rcu_head *h)
8678 {
8679         struct nft_rules_old *o = container_of(h, struct nft_rules_old, h);
8680
8681         kvfree(o->blob);
8682 }
8683
8684 static void nf_tables_commit_chain_free_rules_old(struct nft_rule_blob *blob)
8685 {
8686         struct nft_rules_old *old;
8687
8688         /* rcu_head is after end marker */
8689         old = (void *)blob + sizeof(*blob) + blob->size;
8690         old->blob = blob;
8691
8692         call_rcu(&old->h, __nf_tables_commit_chain_free_rules_old);
8693 }
8694
8695 static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
8696 {
8697         struct nft_rule_blob *g0, *g1;
8698         bool next_genbit;
8699
8700         next_genbit = nft_gencursor_next(net);
8701
8702         g0 = rcu_dereference_protected(chain->blob_gen_0,
8703                                        lockdep_commit_lock_is_held(net));
8704         g1 = rcu_dereference_protected(chain->blob_gen_1,
8705                                        lockdep_commit_lock_is_held(net));
8706
8707         /* No changes to this chain? */
8708         if (chain->blob_next == NULL) {
8709                 /* chain had no change in last or next generation */
8710                 if (g0 == g1)
8711                         return;
8712                 /*
8713                  * chain had no change in this generation; make sure next
8714                  * one uses same rules as current generation.
8715                  */
8716                 if (next_genbit) {
8717                         rcu_assign_pointer(chain->blob_gen_1, g0);
8718                         nf_tables_commit_chain_free_rules_old(g1);
8719                 } else {
8720                         rcu_assign_pointer(chain->blob_gen_0, g1);
8721                         nf_tables_commit_chain_free_rules_old(g0);
8722                 }
8723
8724                 return;
8725         }
8726
8727         if (next_genbit)
8728                 rcu_assign_pointer(chain->blob_gen_1, chain->blob_next);
8729         else
8730                 rcu_assign_pointer(chain->blob_gen_0, chain->blob_next);
8731
8732         chain->blob_next = NULL;
8733
8734         if (g0 == g1)
8735                 return;
8736
8737         if (next_genbit)
8738                 nf_tables_commit_chain_free_rules_old(g1);
8739         else
8740                 nf_tables_commit_chain_free_rules_old(g0);
8741 }
8742
8743 static void nft_obj_del(struct nft_object *obj)
8744 {
8745         rhltable_remove(&nft_objname_ht, &obj->rhlhead, nft_objname_ht_params);
8746         list_del_rcu(&obj->list);
8747 }
8748
8749 void nft_chain_del(struct nft_chain *chain)
8750 {
8751         struct nft_table *table = chain->table;
8752
8753         WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
8754                                      nft_chain_ht_params));
8755         list_del_rcu(&chain->list);
8756 }
8757
8758 static void nf_tables_module_autoload_cleanup(struct net *net)
8759 {
8760         struct nftables_pernet *nft_net = nft_pernet(net);
8761         struct nft_module_request *req, *next;
8762
8763         WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
8764         list_for_each_entry_safe(req, next, &nft_net->module_list, list) {
8765                 WARN_ON_ONCE(!req->done);
8766                 list_del(&req->list);
8767                 kfree(req);
8768         }
8769 }
8770
8771 static void nf_tables_commit_release(struct net *net)
8772 {
8773         struct nftables_pernet *nft_net = nft_pernet(net);
8774         struct nft_trans *trans;
8775
8776         /* all side effects have to be made visible.
8777          * For example, if a chain named 'foo' has been deleted, a
8778          * new transaction must not find it anymore.
8779          *
8780          * Memory reclaim happens asynchronously from work queue
8781          * to prevent expensive synchronize_rcu() in commit phase.
8782          */
8783         if (list_empty(&nft_net->commit_list)) {
8784                 nf_tables_module_autoload_cleanup(net);
8785                 mutex_unlock(&nft_net->commit_mutex);
8786                 return;
8787         }
8788
8789         trans = list_last_entry(&nft_net->commit_list,
8790                                 struct nft_trans, list);
8791         get_net(trans->ctx.net);
8792         WARN_ON_ONCE(trans->put_net);
8793
8794         trans->put_net = true;
8795         spin_lock(&nf_tables_destroy_list_lock);
8796         list_splice_tail_init(&nft_net->commit_list, &nf_tables_destroy_list);
8797         spin_unlock(&nf_tables_destroy_list_lock);
8798
8799         nf_tables_module_autoload_cleanup(net);
8800         schedule_work(&trans_destroy_work);
8801
8802         mutex_unlock(&nft_net->commit_mutex);
8803 }
8804
8805 static void nft_commit_notify(struct net *net, u32 portid)
8806 {
8807         struct nftables_pernet *nft_net = nft_pernet(net);
8808         struct sk_buff *batch_skb = NULL, *nskb, *skb;
8809         unsigned char *data;
8810         int len;
8811
8812         list_for_each_entry_safe(skb, nskb, &nft_net->notify_list, list) {
8813                 if (!batch_skb) {
8814 new_batch:
8815                         batch_skb = skb;
8816                         len = NLMSG_GOODSIZE - skb->len;
8817                         list_del(&skb->list);
8818                         continue;
8819                 }
8820                 len -= skb->len;
8821                 if (len > 0 && NFT_CB(skb).report == NFT_CB(batch_skb).report) {
8822                         data = skb_put(batch_skb, skb->len);
8823                         memcpy(data, skb->data, skb->len);
8824                         list_del(&skb->list);
8825                         kfree_skb(skb);
8826                         continue;
8827                 }
8828                 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8829                                NFT_CB(batch_skb).report, GFP_KERNEL);
8830                 goto new_batch;
8831         }
8832
8833         if (batch_skb) {
8834                 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8835                                NFT_CB(batch_skb).report, GFP_KERNEL);
8836         }
8837
8838         WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
8839 }
8840
8841 static int nf_tables_commit_audit_alloc(struct list_head *adl,
8842                                         struct nft_table *table)
8843 {
8844         struct nft_audit_data *adp;
8845
8846         list_for_each_entry(adp, adl, list) {
8847                 if (adp->table == table)
8848                         return 0;
8849         }
8850         adp = kzalloc(sizeof(*adp), GFP_KERNEL);
8851         if (!adp)
8852                 return -ENOMEM;
8853         adp->table = table;
8854         list_add(&adp->list, adl);
8855         return 0;
8856 }
8857
8858 static void nf_tables_commit_audit_free(struct list_head *adl)
8859 {
8860         struct nft_audit_data *adp, *adn;
8861
8862         list_for_each_entry_safe(adp, adn, adl, list) {
8863                 list_del(&adp->list);
8864                 kfree(adp);
8865         }
8866 }
8867
8868 static void nf_tables_commit_audit_collect(struct list_head *adl,
8869                                            struct nft_table *table, u32 op)
8870 {
8871         struct nft_audit_data *adp;
8872
8873         list_for_each_entry(adp, adl, list) {
8874                 if (adp->table == table)
8875                         goto found;
8876         }
8877         WARN_ONCE(1, "table=%s not expected in commit list", table->name);
8878         return;
8879 found:
8880         adp->entries++;
8881         if (!adp->op || adp->op > op)
8882                 adp->op = op;
8883 }
8884
8885 #define AUNFTABLENAMELEN (NFT_TABLE_MAXNAMELEN + 22)
8886
8887 static void nf_tables_commit_audit_log(struct list_head *adl, u32 generation)
8888 {
8889         struct nft_audit_data *adp, *adn;
8890         char aubuf[AUNFTABLENAMELEN];
8891
8892         list_for_each_entry_safe(adp, adn, adl, list) {
8893                 snprintf(aubuf, AUNFTABLENAMELEN, "%s:%u", adp->table->name,
8894                          generation);
8895                 audit_log_nfcfg(aubuf, adp->table->family, adp->entries,
8896                                 nft2audit_op[adp->op], GFP_KERNEL);
8897                 list_del(&adp->list);
8898                 kfree(adp);
8899         }
8900 }
8901
8902 static int nf_tables_commit(struct net *net, struct sk_buff *skb)
8903 {
8904         struct nftables_pernet *nft_net = nft_pernet(net);
8905         struct nft_trans *trans, *next;
8906         struct nft_trans_elem *te;
8907         struct nft_chain *chain;
8908         struct nft_table *table;
8909         unsigned int base_seq;
8910         LIST_HEAD(adl);
8911         int err;
8912
8913         if (list_empty(&nft_net->commit_list)) {
8914                 mutex_unlock(&nft_net->commit_mutex);
8915                 return 0;
8916         }
8917
8918         /* 0. Validate ruleset, otherwise roll back for error reporting. */
8919         if (nf_tables_validate(net) < 0)
8920                 return -EAGAIN;
8921
8922         err = nft_flow_rule_offload_commit(net);
8923         if (err < 0)
8924                 return err;
8925
8926         /* 1.  Allocate space for next generation rules_gen_X[] */
8927         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8928                 int ret;
8929
8930                 ret = nf_tables_commit_audit_alloc(&adl, trans->ctx.table);
8931                 if (ret) {
8932                         nf_tables_commit_chain_prepare_cancel(net);
8933                         nf_tables_commit_audit_free(&adl);
8934                         return ret;
8935                 }
8936                 if (trans->msg_type == NFT_MSG_NEWRULE ||
8937                     trans->msg_type == NFT_MSG_DELRULE) {
8938                         chain = trans->ctx.chain;
8939
8940                         ret = nf_tables_commit_chain_prepare(net, chain);
8941                         if (ret < 0) {
8942                                 nf_tables_commit_chain_prepare_cancel(net);
8943                                 nf_tables_commit_audit_free(&adl);
8944                                 return ret;
8945                         }
8946                 }
8947         }
8948
8949         /* step 2.  Make rules_gen_X visible to packet path */
8950         list_for_each_entry(table, &nft_net->tables, list) {
8951                 list_for_each_entry(chain, &table->chains, list)
8952                         nf_tables_commit_chain(net, chain);
8953         }
8954
8955         /*
8956          * Bump generation counter, invalidate any dump in progress.
8957          * Cannot fail after this point.
8958          */
8959         base_seq = READ_ONCE(nft_net->base_seq);
8960         while (++base_seq == 0)
8961                 ;
8962
8963         WRITE_ONCE(nft_net->base_seq, base_seq);
8964
8965         /* step 3. Start new generation, rules_gen_X now in use. */
8966         net->nft.gencursor = nft_gencursor_next(net);
8967
8968         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8969                 nf_tables_commit_audit_collect(&adl, trans->ctx.table,
8970                                                trans->msg_type);
8971                 switch (trans->msg_type) {
8972                 case NFT_MSG_NEWTABLE:
8973                         if (nft_trans_table_update(trans)) {
8974                                 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
8975                                         nft_trans_destroy(trans);
8976                                         break;
8977                                 }
8978                                 if (trans->ctx.table->flags & NFT_TABLE_F_DORMANT)
8979                                         nf_tables_table_disable(net, trans->ctx.table);
8980
8981                                 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
8982                         } else {
8983                                 nft_clear(net, trans->ctx.table);
8984                         }
8985                         nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
8986                         nft_trans_destroy(trans);
8987                         break;
8988                 case NFT_MSG_DELTABLE:
8989                         list_del_rcu(&trans->ctx.table->list);
8990                         nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
8991                         break;
8992                 case NFT_MSG_NEWCHAIN:
8993                         if (nft_trans_chain_update(trans)) {
8994                                 nft_chain_commit_update(trans);
8995                                 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8996                                 /* trans destroyed after rcu grace period */
8997                         } else {
8998                                 nft_chain_commit_drop_policy(trans);
8999                                 nft_clear(net, trans->ctx.chain);
9000                                 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
9001                                 nft_trans_destroy(trans);
9002                         }
9003                         break;
9004                 case NFT_MSG_DELCHAIN:
9005                         nft_chain_del(trans->ctx.chain);
9006                         nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
9007                         nf_tables_unregister_hook(trans->ctx.net,
9008                                                   trans->ctx.table,
9009                                                   trans->ctx.chain);
9010                         break;
9011                 case NFT_MSG_NEWRULE:
9012                         nft_clear(trans->ctx.net, nft_trans_rule(trans));
9013                         nf_tables_rule_notify(&trans->ctx,
9014                                               nft_trans_rule(trans),
9015                                               NFT_MSG_NEWRULE);
9016                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9017                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9018
9019                         nft_trans_destroy(trans);
9020                         break;
9021                 case NFT_MSG_DELRULE:
9022                         list_del_rcu(&nft_trans_rule(trans)->list);
9023                         nf_tables_rule_notify(&trans->ctx,
9024                                               nft_trans_rule(trans),
9025                                               NFT_MSG_DELRULE);
9026                         nft_rule_expr_deactivate(&trans->ctx,
9027                                                  nft_trans_rule(trans),
9028                                                  NFT_TRANS_COMMIT);
9029
9030                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9031                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9032                         break;
9033                 case NFT_MSG_NEWSET:
9034                         nft_clear(net, nft_trans_set(trans));
9035                         /* This avoids hitting -EBUSY when deleting the table
9036                          * from the transaction.
9037                          */
9038                         if (nft_set_is_anonymous(nft_trans_set(trans)) &&
9039                             !list_empty(&nft_trans_set(trans)->bindings))
9040                                 trans->ctx.table->use--;
9041
9042                         nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
9043                                              NFT_MSG_NEWSET, GFP_KERNEL);
9044                         nft_trans_destroy(trans);
9045                         break;
9046                 case NFT_MSG_DELSET:
9047                         list_del_rcu(&nft_trans_set(trans)->list);
9048                         nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
9049                                              NFT_MSG_DELSET, GFP_KERNEL);
9050                         break;
9051                 case NFT_MSG_NEWSETELEM:
9052                         te = (struct nft_trans_elem *)trans->data;
9053
9054                         nft_setelem_activate(net, te->set, &te->elem);
9055                         nf_tables_setelem_notify(&trans->ctx, te->set,
9056                                                  &te->elem,
9057                                                  NFT_MSG_NEWSETELEM);
9058                         nft_trans_destroy(trans);
9059                         break;
9060                 case NFT_MSG_DELSETELEM:
9061                         te = (struct nft_trans_elem *)trans->data;
9062
9063                         nf_tables_setelem_notify(&trans->ctx, te->set,
9064                                                  &te->elem,
9065                                                  NFT_MSG_DELSETELEM);
9066                         nft_setelem_remove(net, te->set, &te->elem);
9067                         if (!nft_setelem_is_catchall(te->set, &te->elem)) {
9068                                 atomic_dec(&te->set->nelems);
9069                                 te->set->ndeact--;
9070                         }
9071                         break;
9072                 case NFT_MSG_NEWOBJ:
9073                         if (nft_trans_obj_update(trans)) {
9074                                 nft_obj_commit_update(trans);
9075                                 nf_tables_obj_notify(&trans->ctx,
9076                                                      nft_trans_obj(trans),
9077                                                      NFT_MSG_NEWOBJ);
9078                         } else {
9079                                 nft_clear(net, nft_trans_obj(trans));
9080                                 nf_tables_obj_notify(&trans->ctx,
9081                                                      nft_trans_obj(trans),
9082                                                      NFT_MSG_NEWOBJ);
9083                                 nft_trans_destroy(trans);
9084                         }
9085                         break;
9086                 case NFT_MSG_DELOBJ:
9087                         nft_obj_del(nft_trans_obj(trans));
9088                         nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
9089                                              NFT_MSG_DELOBJ);
9090                         break;
9091                 case NFT_MSG_NEWFLOWTABLE:
9092                         if (nft_trans_flowtable_update(trans)) {
9093                                 nft_trans_flowtable(trans)->data.flags =
9094                                         nft_trans_flowtable_flags(trans);
9095                                 nf_tables_flowtable_notify(&trans->ctx,
9096                                                            nft_trans_flowtable(trans),
9097                                                            &nft_trans_flowtable_hooks(trans),
9098                                                            NFT_MSG_NEWFLOWTABLE);
9099                                 list_splice(&nft_trans_flowtable_hooks(trans),
9100                                             &nft_trans_flowtable(trans)->hook_list);
9101                         } else {
9102                                 nft_clear(net, nft_trans_flowtable(trans));
9103                                 nf_tables_flowtable_notify(&trans->ctx,
9104                                                            nft_trans_flowtable(trans),
9105                                                            &nft_trans_flowtable(trans)->hook_list,
9106                                                            NFT_MSG_NEWFLOWTABLE);
9107                         }
9108                         nft_trans_destroy(trans);
9109                         break;
9110                 case NFT_MSG_DELFLOWTABLE:
9111                         if (nft_trans_flowtable_update(trans)) {
9112                                 nf_tables_flowtable_notify(&trans->ctx,
9113                                                            nft_trans_flowtable(trans),
9114                                                            &nft_trans_flowtable_hooks(trans),
9115                                                            NFT_MSG_DELFLOWTABLE);
9116                                 nft_unregister_flowtable_net_hooks(net,
9117                                                                    &nft_trans_flowtable_hooks(trans));
9118                         } else {
9119                                 list_del_rcu(&nft_trans_flowtable(trans)->list);
9120                                 nf_tables_flowtable_notify(&trans->ctx,
9121                                                            nft_trans_flowtable(trans),
9122                                                            &nft_trans_flowtable(trans)->hook_list,
9123                                                            NFT_MSG_DELFLOWTABLE);
9124                                 nft_unregister_flowtable_net_hooks(net,
9125                                                 &nft_trans_flowtable(trans)->hook_list);
9126                         }
9127                         break;
9128                 }
9129         }
9130
9131         nft_commit_notify(net, NETLINK_CB(skb).portid);
9132         nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
9133         nf_tables_commit_audit_log(&adl, nft_net->base_seq);
9134         nf_tables_commit_release(net);
9135
9136         return 0;
9137 }
9138
9139 static void nf_tables_module_autoload(struct net *net)
9140 {
9141         struct nftables_pernet *nft_net = nft_pernet(net);
9142         struct nft_module_request *req, *next;
9143         LIST_HEAD(module_list);
9144
9145         list_splice_init(&nft_net->module_list, &module_list);
9146         mutex_unlock(&nft_net->commit_mutex);
9147         list_for_each_entry_safe(req, next, &module_list, list) {
9148                 request_module("%s", req->module);
9149                 req->done = true;
9150         }
9151         mutex_lock(&nft_net->commit_mutex);
9152         list_splice(&module_list, &nft_net->module_list);
9153 }
9154
9155 static void nf_tables_abort_release(struct nft_trans *trans)
9156 {
9157         switch (trans->msg_type) {
9158         case NFT_MSG_NEWTABLE:
9159                 nf_tables_table_destroy(&trans->ctx);
9160                 break;
9161         case NFT_MSG_NEWCHAIN:
9162                 nf_tables_chain_destroy(&trans->ctx);
9163                 break;
9164         case NFT_MSG_NEWRULE:
9165                 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
9166                 break;
9167         case NFT_MSG_NEWSET:
9168                 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
9169                 break;
9170         case NFT_MSG_NEWSETELEM:
9171                 nft_set_elem_destroy(nft_trans_elem_set(trans),
9172                                      nft_trans_elem(trans).priv, true);
9173                 break;
9174         case NFT_MSG_NEWOBJ:
9175                 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
9176                 break;
9177         case NFT_MSG_NEWFLOWTABLE:
9178                 if (nft_trans_flowtable_update(trans))
9179                         nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
9180                 else
9181                         nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
9182                 break;
9183         }
9184         kfree(trans);
9185 }
9186
9187 static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)
9188 {
9189         struct nftables_pernet *nft_net = nft_pernet(net);
9190         struct nft_trans *trans, *next;
9191         struct nft_trans_elem *te;
9192
9193         if (action == NFNL_ABORT_VALIDATE &&
9194             nf_tables_validate(net) < 0)
9195                 return -EAGAIN;
9196
9197         list_for_each_entry_safe_reverse(trans, next, &nft_net->commit_list,
9198                                          list) {
9199                 switch (trans->msg_type) {
9200                 case NFT_MSG_NEWTABLE:
9201                         if (nft_trans_table_update(trans)) {
9202                                 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
9203                                         nft_trans_destroy(trans);
9204                                         break;
9205                                 }
9206                                 if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_DORMANT) {
9207                                         nf_tables_table_disable(net, trans->ctx.table);
9208                                         trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
9209                                 } else if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_AWAKEN) {
9210                                         trans->ctx.table->flags &= ~NFT_TABLE_F_DORMANT;
9211                                 }
9212                                 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
9213                                 nft_trans_destroy(trans);
9214                         } else {
9215                                 list_del_rcu(&trans->ctx.table->list);
9216                         }
9217                         break;
9218                 case NFT_MSG_DELTABLE:
9219                         nft_clear(trans->ctx.net, trans->ctx.table);
9220                         nft_trans_destroy(trans);
9221                         break;
9222                 case NFT_MSG_NEWCHAIN:
9223                         if (nft_trans_chain_update(trans)) {
9224                                 free_percpu(nft_trans_chain_stats(trans));
9225                                 kfree(nft_trans_chain_name(trans));
9226                                 nft_trans_destroy(trans);
9227                         } else {
9228                                 if (nft_chain_is_bound(trans->ctx.chain)) {
9229                                         nft_trans_destroy(trans);
9230                                         break;
9231                                 }
9232                                 trans->ctx.table->use--;
9233                                 nft_chain_del(trans->ctx.chain);
9234                                 nf_tables_unregister_hook(trans->ctx.net,
9235                                                           trans->ctx.table,
9236                                                           trans->ctx.chain);
9237                         }
9238                         break;
9239                 case NFT_MSG_DELCHAIN:
9240                         trans->ctx.table->use++;
9241                         nft_clear(trans->ctx.net, trans->ctx.chain);
9242                         nft_trans_destroy(trans);
9243                         break;
9244                 case NFT_MSG_NEWRULE:
9245                         trans->ctx.chain->use--;
9246                         list_del_rcu(&nft_trans_rule(trans)->list);
9247                         nft_rule_expr_deactivate(&trans->ctx,
9248                                                  nft_trans_rule(trans),
9249                                                  NFT_TRANS_ABORT);
9250                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9251                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9252                         break;
9253                 case NFT_MSG_DELRULE:
9254                         trans->ctx.chain->use++;
9255                         nft_clear(trans->ctx.net, nft_trans_rule(trans));
9256                         nft_rule_expr_activate(&trans->ctx, nft_trans_rule(trans));
9257                         if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
9258                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
9259
9260                         nft_trans_destroy(trans);
9261                         break;
9262                 case NFT_MSG_NEWSET:
9263                         trans->ctx.table->use--;
9264                         if (nft_trans_set_bound(trans)) {
9265                                 nft_trans_destroy(trans);
9266                                 break;
9267                         }
9268                         list_del_rcu(&nft_trans_set(trans)->list);
9269                         break;
9270                 case NFT_MSG_DELSET:
9271                         trans->ctx.table->use++;
9272                         nft_clear(trans->ctx.net, nft_trans_set(trans));
9273                         nft_trans_destroy(trans);
9274                         break;
9275                 case NFT_MSG_NEWSETELEM:
9276                         if (nft_trans_elem_set_bound(trans)) {
9277                                 nft_trans_destroy(trans);
9278                                 break;
9279                         }
9280                         te = (struct nft_trans_elem *)trans->data;
9281                         nft_setelem_remove(net, te->set, &te->elem);
9282                         if (!nft_setelem_is_catchall(te->set, &te->elem))
9283                                 atomic_dec(&te->set->nelems);
9284                         break;
9285                 case NFT_MSG_DELSETELEM:
9286                         te = (struct nft_trans_elem *)trans->data;
9287
9288                         nft_setelem_data_activate(net, te->set, &te->elem);
9289                         nft_setelem_activate(net, te->set, &te->elem);
9290                         if (!nft_setelem_is_catchall(te->set, &te->elem))
9291                                 te->set->ndeact--;
9292
9293                         nft_trans_destroy(trans);
9294                         break;
9295                 case NFT_MSG_NEWOBJ:
9296                         if (nft_trans_obj_update(trans)) {
9297                                 nft_obj_destroy(&trans->ctx, nft_trans_obj_newobj(trans));
9298                                 nft_trans_destroy(trans);
9299                         } else {
9300                                 trans->ctx.table->use--;
9301                                 nft_obj_del(nft_trans_obj(trans));
9302                         }
9303                         break;
9304                 case NFT_MSG_DELOBJ:
9305                         trans->ctx.table->use++;
9306                         nft_clear(trans->ctx.net, nft_trans_obj(trans));
9307                         nft_trans_destroy(trans);
9308                         break;
9309                 case NFT_MSG_NEWFLOWTABLE:
9310                         if (nft_trans_flowtable_update(trans)) {
9311                                 nft_unregister_flowtable_net_hooks(net,
9312                                                 &nft_trans_flowtable_hooks(trans));
9313                         } else {
9314                                 trans->ctx.table->use--;
9315                                 list_del_rcu(&nft_trans_flowtable(trans)->list);
9316                                 nft_unregister_flowtable_net_hooks(net,
9317                                                 &nft_trans_flowtable(trans)->hook_list);
9318                         }
9319                         break;
9320                 case NFT_MSG_DELFLOWTABLE:
9321                         if (nft_trans_flowtable_update(trans)) {
9322                                 list_splice(&nft_trans_flowtable_hooks(trans),
9323                                             &nft_trans_flowtable(trans)->hook_list);
9324                         } else {
9325                                 trans->ctx.table->use++;
9326                                 nft_clear(trans->ctx.net, nft_trans_flowtable(trans));
9327                         }
9328                         nft_trans_destroy(trans);
9329                         break;
9330                 }
9331         }
9332
9333         synchronize_rcu();
9334
9335         list_for_each_entry_safe_reverse(trans, next,
9336                                          &nft_net->commit_list, list) {
9337                 list_del(&trans->list);
9338                 nf_tables_abort_release(trans);
9339         }
9340
9341         if (action == NFNL_ABORT_AUTOLOAD)
9342                 nf_tables_module_autoload(net);
9343         else
9344                 nf_tables_module_autoload_cleanup(net);
9345
9346         return 0;
9347 }
9348
9349 static void nf_tables_cleanup(struct net *net)
9350 {
9351         nft_validate_state_update(net, NFT_VALIDATE_SKIP);
9352 }
9353
9354 static int nf_tables_abort(struct net *net, struct sk_buff *skb,
9355                            enum nfnl_abort_action action)
9356 {
9357         struct nftables_pernet *nft_net = nft_pernet(net);
9358         int ret = __nf_tables_abort(net, action);
9359
9360         mutex_unlock(&nft_net->commit_mutex);
9361
9362         return ret;
9363 }
9364
9365 static bool nf_tables_valid_genid(struct net *net, u32 genid)
9366 {
9367         struct nftables_pernet *nft_net = nft_pernet(net);
9368         bool genid_ok;
9369
9370         mutex_lock(&nft_net->commit_mutex);
9371
9372         genid_ok = genid == 0 || nft_net->base_seq == genid;
9373         if (!genid_ok)
9374                 mutex_unlock(&nft_net->commit_mutex);
9375
9376         /* else, commit mutex has to be released by commit or abort function */
9377         return genid_ok;
9378 }
9379
9380 static const struct nfnetlink_subsystem nf_tables_subsys = {
9381         .name           = "nf_tables",
9382         .subsys_id      = NFNL_SUBSYS_NFTABLES,
9383         .cb_count       = NFT_MSG_MAX,
9384         .cb             = nf_tables_cb,
9385         .commit         = nf_tables_commit,
9386         .abort          = nf_tables_abort,
9387         .cleanup        = nf_tables_cleanup,
9388         .valid_genid    = nf_tables_valid_genid,
9389         .owner          = THIS_MODULE,
9390 };
9391
9392 int nft_chain_validate_dependency(const struct nft_chain *chain,
9393                                   enum nft_chain_types type)
9394 {
9395         const struct nft_base_chain *basechain;
9396
9397         if (nft_is_base_chain(chain)) {
9398                 basechain = nft_base_chain(chain);
9399                 if (basechain->type->type != type)
9400                         return -EOPNOTSUPP;
9401         }
9402         return 0;
9403 }
9404 EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
9405
9406 int nft_chain_validate_hooks(const struct nft_chain *chain,
9407                              unsigned int hook_flags)
9408 {
9409         struct nft_base_chain *basechain;
9410
9411         if (nft_is_base_chain(chain)) {
9412                 basechain = nft_base_chain(chain);
9413
9414                 if ((1 << basechain->ops.hooknum) & hook_flags)
9415                         return 0;
9416
9417                 return -EOPNOTSUPP;
9418         }
9419
9420         return 0;
9421 }
9422 EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
9423
9424 /*
9425  * Loop detection - walk through the ruleset beginning at the destination chain
9426  * of a new jump until either the source chain is reached (loop) or all
9427  * reachable chains have been traversed.
9428  *
9429  * The loop check is performed whenever a new jump verdict is added to an
9430  * expression or verdict map or a verdict map is bound to a new chain.
9431  */
9432
9433 static int nf_tables_check_loops(const struct nft_ctx *ctx,
9434                                  const struct nft_chain *chain);
9435
9436 static int nft_check_loops(const struct nft_ctx *ctx,
9437                            const struct nft_set_ext *ext)
9438 {
9439         const struct nft_data *data;
9440         int ret;
9441
9442         data = nft_set_ext_data(ext);
9443         switch (data->verdict.code) {
9444         case NFT_JUMP:
9445         case NFT_GOTO:
9446                 ret = nf_tables_check_loops(ctx, data->verdict.chain);
9447                 break;
9448         default:
9449                 ret = 0;
9450                 break;
9451         }
9452
9453         return ret;
9454 }
9455
9456 static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
9457                                         struct nft_set *set,
9458                                         const struct nft_set_iter *iter,
9459                                         struct nft_set_elem *elem)
9460 {
9461         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
9462
9463         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
9464             *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
9465                 return 0;
9466
9467         return nft_check_loops(ctx, ext);
9468 }
9469
9470 static int nft_set_catchall_loops(const struct nft_ctx *ctx,
9471                                   struct nft_set *set)
9472 {
9473         u8 genmask = nft_genmask_next(ctx->net);
9474         struct nft_set_elem_catchall *catchall;
9475         struct nft_set_ext *ext;
9476         int ret = 0;
9477
9478         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
9479                 ext = nft_set_elem_ext(set, catchall->elem);
9480                 if (!nft_set_elem_active(ext, genmask))
9481                         continue;
9482
9483                 ret = nft_check_loops(ctx, ext);
9484                 if (ret < 0)
9485                         return ret;
9486         }
9487
9488         return ret;
9489 }
9490
9491 static int nf_tables_check_loops(const struct nft_ctx *ctx,
9492                                  const struct nft_chain *chain)
9493 {
9494         const struct nft_rule *rule;
9495         const struct nft_expr *expr, *last;
9496         struct nft_set *set;
9497         struct nft_set_binding *binding;
9498         struct nft_set_iter iter;
9499
9500         if (ctx->chain == chain)
9501                 return -ELOOP;
9502
9503         list_for_each_entry(rule, &chain->rules, list) {
9504                 nft_rule_for_each_expr(expr, last, rule) {
9505                         struct nft_immediate_expr *priv;
9506                         const struct nft_data *data;
9507                         int err;
9508
9509                         if (strcmp(expr->ops->type->name, "immediate"))
9510                                 continue;
9511
9512                         priv = nft_expr_priv(expr);
9513                         if (priv->dreg != NFT_REG_VERDICT)
9514                                 continue;
9515
9516                         data = &priv->data;
9517                         switch (data->verdict.code) {
9518                         case NFT_JUMP:
9519                         case NFT_GOTO:
9520                                 err = nf_tables_check_loops(ctx,
9521                                                         data->verdict.chain);
9522                                 if (err < 0)
9523                                         return err;
9524                                 break;
9525                         default:
9526                                 break;
9527                         }
9528                 }
9529         }
9530
9531         list_for_each_entry(set, &ctx->table->sets, list) {
9532                 if (!nft_is_active_next(ctx->net, set))
9533                         continue;
9534                 if (!(set->flags & NFT_SET_MAP) ||
9535                     set->dtype != NFT_DATA_VERDICT)
9536                         continue;
9537
9538                 list_for_each_entry(binding, &set->bindings, list) {
9539                         if (!(binding->flags & NFT_SET_MAP) ||
9540                             binding->chain != chain)
9541                                 continue;
9542
9543                         iter.genmask    = nft_genmask_next(ctx->net);
9544                         iter.skip       = 0;
9545                         iter.count      = 0;
9546                         iter.err        = 0;
9547                         iter.fn         = nf_tables_loop_check_setelem;
9548
9549                         set->ops->walk(ctx, set, &iter);
9550                         if (!iter.err)
9551                                 iter.err = nft_set_catchall_loops(ctx, set);
9552
9553                         if (iter.err < 0)
9554                                 return iter.err;
9555                 }
9556         }
9557
9558         return 0;
9559 }
9560
9561 /**
9562  *      nft_parse_u32_check - fetch u32 attribute and check for maximum value
9563  *
9564  *      @attr: netlink attribute to fetch value from
9565  *      @max: maximum value to be stored in dest
9566  *      @dest: pointer to the variable
9567  *
9568  *      Parse, check and store a given u32 netlink attribute into variable.
9569  *      This function returns -ERANGE if the value goes over maximum value.
9570  *      Otherwise a 0 is returned and the attribute value is stored in the
9571  *      destination variable.
9572  */
9573 int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
9574 {
9575         u32 val;
9576
9577         val = ntohl(nla_get_be32(attr));
9578         if (val > max)
9579                 return -ERANGE;
9580
9581         *dest = val;
9582         return 0;
9583 }
9584 EXPORT_SYMBOL_GPL(nft_parse_u32_check);
9585
9586 static int nft_parse_register(const struct nlattr *attr, u32 *preg)
9587 {
9588         unsigned int reg;
9589
9590         reg = ntohl(nla_get_be32(attr));
9591         switch (reg) {
9592         case NFT_REG_VERDICT...NFT_REG_4:
9593                 *preg = reg * NFT_REG_SIZE / NFT_REG32_SIZE;
9594                 break;
9595         case NFT_REG32_00...NFT_REG32_15:
9596                 *preg = reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
9597                 break;
9598         default:
9599                 return -ERANGE;
9600         }
9601
9602         return 0;
9603 }
9604
9605 /**
9606  *      nft_dump_register - dump a register value to a netlink attribute
9607  *
9608  *      @skb: socket buffer
9609  *      @attr: attribute number
9610  *      @reg: register number
9611  *
9612  *      Construct a netlink attribute containing the register number. For
9613  *      compatibility reasons, register numbers being a multiple of 4 are
9614  *      translated to the corresponding 128 bit register numbers.
9615  */
9616 int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
9617 {
9618         if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
9619                 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
9620         else
9621                 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
9622
9623         return nla_put_be32(skb, attr, htonl(reg));
9624 }
9625 EXPORT_SYMBOL_GPL(nft_dump_register);
9626
9627 static int nft_validate_register_load(enum nft_registers reg, unsigned int len)
9628 {
9629         if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9630                 return -EINVAL;
9631         if (len == 0)
9632                 return -EINVAL;
9633         if (reg * NFT_REG32_SIZE + len > sizeof_field(struct nft_regs, data))
9634                 return -ERANGE;
9635
9636         return 0;
9637 }
9638
9639 int nft_parse_register_load(const struct nlattr *attr, u8 *sreg, u32 len)
9640 {
9641         u32 reg;
9642         int err;
9643
9644         err = nft_parse_register(attr, &reg);
9645         if (err < 0)
9646                 return err;
9647
9648         err = nft_validate_register_load(reg, len);
9649         if (err < 0)
9650                 return err;
9651
9652         *sreg = reg;
9653         return 0;
9654 }
9655 EXPORT_SYMBOL_GPL(nft_parse_register_load);
9656
9657 static int nft_validate_register_store(const struct nft_ctx *ctx,
9658                                        enum nft_registers reg,
9659                                        const struct nft_data *data,
9660                                        enum nft_data_types type,
9661                                        unsigned int len)
9662 {
9663         int err;
9664
9665         switch (reg) {
9666         case NFT_REG_VERDICT:
9667                 if (type != NFT_DATA_VERDICT)
9668                         return -EINVAL;
9669
9670                 if (data != NULL &&
9671                     (data->verdict.code == NFT_GOTO ||
9672                      data->verdict.code == NFT_JUMP)) {
9673                         err = nf_tables_check_loops(ctx, data->verdict.chain);
9674                         if (err < 0)
9675                                 return err;
9676                 }
9677
9678                 return 0;
9679         default:
9680                 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9681                         return -EINVAL;
9682                 if (len == 0)
9683                         return -EINVAL;
9684                 if (reg * NFT_REG32_SIZE + len >
9685                     sizeof_field(struct nft_regs, data))
9686                         return -ERANGE;
9687
9688                 if (data != NULL && type != NFT_DATA_VALUE)
9689                         return -EINVAL;
9690                 return 0;
9691         }
9692 }
9693
9694 int nft_parse_register_store(const struct nft_ctx *ctx,
9695                              const struct nlattr *attr, u8 *dreg,
9696                              const struct nft_data *data,
9697                              enum nft_data_types type, unsigned int len)
9698 {
9699         int err;
9700         u32 reg;
9701
9702         err = nft_parse_register(attr, &reg);
9703         if (err < 0)
9704                 return err;
9705
9706         err = nft_validate_register_store(ctx, reg, data, type, len);
9707         if (err < 0)
9708                 return err;
9709
9710         *dreg = reg;
9711         return 0;
9712 }
9713 EXPORT_SYMBOL_GPL(nft_parse_register_store);
9714
9715 static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
9716         [NFTA_VERDICT_CODE]     = { .type = NLA_U32 },
9717         [NFTA_VERDICT_CHAIN]    = { .type = NLA_STRING,
9718                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
9719         [NFTA_VERDICT_CHAIN_ID] = { .type = NLA_U32 },
9720 };
9721
9722 static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
9723                             struct nft_data_desc *desc, const struct nlattr *nla)
9724 {
9725         u8 genmask = nft_genmask_next(ctx->net);
9726         struct nlattr *tb[NFTA_VERDICT_MAX + 1];
9727         struct nft_chain *chain;
9728         int err;
9729
9730         err = nla_parse_nested_deprecated(tb, NFTA_VERDICT_MAX, nla,
9731                                           nft_verdict_policy, NULL);
9732         if (err < 0)
9733                 return err;
9734
9735         if (!tb[NFTA_VERDICT_CODE])
9736                 return -EINVAL;
9737         data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
9738
9739         switch (data->verdict.code) {
9740         default:
9741                 switch (data->verdict.code & NF_VERDICT_MASK) {
9742                 case NF_ACCEPT:
9743                 case NF_DROP:
9744                 case NF_QUEUE:
9745                         break;
9746                 default:
9747                         return -EINVAL;
9748                 }
9749                 fallthrough;
9750         case NFT_CONTINUE:
9751         case NFT_BREAK:
9752         case NFT_RETURN:
9753                 break;
9754         case NFT_JUMP:
9755         case NFT_GOTO:
9756                 if (tb[NFTA_VERDICT_CHAIN]) {
9757                         chain = nft_chain_lookup(ctx->net, ctx->table,
9758                                                  tb[NFTA_VERDICT_CHAIN],
9759                                                  genmask);
9760                 } else if (tb[NFTA_VERDICT_CHAIN_ID]) {
9761                         chain = nft_chain_lookup_byid(ctx->net, ctx->table,
9762                                                       tb[NFTA_VERDICT_CHAIN_ID]);
9763                         if (IS_ERR(chain))
9764                                 return PTR_ERR(chain);
9765                 } else {
9766                         return -EINVAL;
9767                 }
9768
9769                 if (IS_ERR(chain))
9770                         return PTR_ERR(chain);
9771                 if (nft_is_base_chain(chain))
9772                         return -EOPNOTSUPP;
9773                 if (nft_chain_is_bound(chain))
9774                         return -EINVAL;
9775                 if (desc->flags & NFT_DATA_DESC_SETELEM &&
9776                     chain->flags & NFT_CHAIN_BINDING)
9777                         return -EINVAL;
9778
9779                 chain->use++;
9780                 data->verdict.chain = chain;
9781                 break;
9782         }
9783
9784         desc->len = sizeof(data->verdict);
9785
9786         return 0;
9787 }
9788
9789 static void nft_verdict_uninit(const struct nft_data *data)
9790 {
9791         struct nft_chain *chain;
9792         struct nft_rule *rule;
9793
9794         switch (data->verdict.code) {
9795         case NFT_JUMP:
9796         case NFT_GOTO:
9797                 chain = data->verdict.chain;
9798                 chain->use--;
9799
9800                 if (!nft_chain_is_bound(chain))
9801                         break;
9802
9803                 chain->table->use--;
9804                 list_for_each_entry(rule, &chain->rules, list)
9805                         chain->use--;
9806
9807                 nft_chain_del(chain);
9808                 break;
9809         }
9810 }
9811
9812 int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
9813 {
9814         struct nlattr *nest;
9815
9816         nest = nla_nest_start_noflag(skb, type);
9817         if (!nest)
9818                 goto nla_put_failure;
9819
9820         if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
9821                 goto nla_put_failure;
9822
9823         switch (v->code) {
9824         case NFT_JUMP:
9825         case NFT_GOTO:
9826                 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
9827                                    v->chain->name))
9828                         goto nla_put_failure;
9829         }
9830         nla_nest_end(skb, nest);
9831         return 0;
9832
9833 nla_put_failure:
9834         return -1;
9835 }
9836
9837 static int nft_value_init(const struct nft_ctx *ctx,
9838                           struct nft_data *data, struct nft_data_desc *desc,
9839                           const struct nlattr *nla)
9840 {
9841         unsigned int len;
9842
9843         len = nla_len(nla);
9844         if (len == 0)
9845                 return -EINVAL;
9846         if (len > desc->size)
9847                 return -EOVERFLOW;
9848         if (desc->len) {
9849                 if (len != desc->len)
9850                         return -EINVAL;
9851         } else {
9852                 desc->len = len;
9853         }
9854
9855         nla_memcpy(data->data, nla, len);
9856
9857         return 0;
9858 }
9859
9860 static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
9861                           unsigned int len)
9862 {
9863         return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
9864 }
9865
9866 static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
9867         [NFTA_DATA_VALUE]       = { .type = NLA_BINARY },
9868         [NFTA_DATA_VERDICT]     = { .type = NLA_NESTED },
9869 };
9870
9871 /**
9872  *      nft_data_init - parse nf_tables data netlink attributes
9873  *
9874  *      @ctx: context of the expression using the data
9875  *      @data: destination struct nft_data
9876  *      @desc: data description
9877  *      @nla: netlink attribute containing data
9878  *
9879  *      Parse the netlink data attributes and initialize a struct nft_data.
9880  *      The type and length of data are returned in the data description.
9881  *
9882  *      The caller can indicate that it only wants to accept data of type
9883  *      NFT_DATA_VALUE by passing NULL for the ctx argument.
9884  */
9885 int nft_data_init(const struct nft_ctx *ctx, struct nft_data *data,
9886                   struct nft_data_desc *desc, const struct nlattr *nla)
9887 {
9888         struct nlattr *tb[NFTA_DATA_MAX + 1];
9889         int err;
9890
9891         if (WARN_ON_ONCE(!desc->size))
9892                 return -EINVAL;
9893
9894         err = nla_parse_nested_deprecated(tb, NFTA_DATA_MAX, nla,
9895                                           nft_data_policy, NULL);
9896         if (err < 0)
9897                 return err;
9898
9899         if (tb[NFTA_DATA_VALUE]) {
9900                 if (desc->type != NFT_DATA_VALUE)
9901                         return -EINVAL;
9902
9903                 err = nft_value_init(ctx, data, desc, tb[NFTA_DATA_VALUE]);
9904         } else if (tb[NFTA_DATA_VERDICT] && ctx != NULL) {
9905                 if (desc->type != NFT_DATA_VERDICT)
9906                         return -EINVAL;
9907
9908                 err = nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
9909         } else {
9910                 err = -EINVAL;
9911         }
9912
9913         return err;
9914 }
9915 EXPORT_SYMBOL_GPL(nft_data_init);
9916
9917 /**
9918  *      nft_data_release - release a nft_data item
9919  *
9920  *      @data: struct nft_data to release
9921  *      @type: type of data
9922  *
9923  *      Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
9924  *      all others need to be released by calling this function.
9925  */
9926 void nft_data_release(const struct nft_data *data, enum nft_data_types type)
9927 {
9928         if (type < NFT_DATA_VERDICT)
9929                 return;
9930         switch (type) {
9931         case NFT_DATA_VERDICT:
9932                 return nft_verdict_uninit(data);
9933         default:
9934                 WARN_ON(1);
9935         }
9936 }
9937 EXPORT_SYMBOL_GPL(nft_data_release);
9938
9939 int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
9940                   enum nft_data_types type, unsigned int len)
9941 {
9942         struct nlattr *nest;
9943         int err;
9944
9945         nest = nla_nest_start_noflag(skb, attr);
9946         if (nest == NULL)
9947                 return -1;
9948
9949         switch (type) {
9950         case NFT_DATA_VALUE:
9951                 err = nft_value_dump(skb, data, len);
9952                 break;
9953         case NFT_DATA_VERDICT:
9954                 err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
9955                 break;
9956         default:
9957                 err = -EINVAL;
9958                 WARN_ON(1);
9959         }
9960
9961         nla_nest_end(skb, nest);
9962         return err;
9963 }
9964 EXPORT_SYMBOL_GPL(nft_data_dump);
9965
9966 int __nft_release_basechain(struct nft_ctx *ctx)
9967 {
9968         struct nft_rule *rule, *nr;
9969
9970         if (WARN_ON(!nft_is_base_chain(ctx->chain)))
9971                 return 0;
9972
9973         nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
9974         list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
9975                 list_del(&rule->list);
9976                 ctx->chain->use--;
9977                 nf_tables_rule_release(ctx, rule);
9978         }
9979         nft_chain_del(ctx->chain);
9980         ctx->table->use--;
9981         nf_tables_chain_destroy(ctx);
9982
9983         return 0;
9984 }
9985 EXPORT_SYMBOL_GPL(__nft_release_basechain);
9986
9987 static void __nft_release_hook(struct net *net, struct nft_table *table)
9988 {
9989         struct nft_flowtable *flowtable;
9990         struct nft_chain *chain;
9991
9992         list_for_each_entry(chain, &table->chains, list)
9993                 __nf_tables_unregister_hook(net, table, chain, true);
9994         list_for_each_entry(flowtable, &table->flowtables, list)
9995                 __nft_unregister_flowtable_net_hooks(net, &flowtable->hook_list,
9996                                                      true);
9997 }
9998
9999 static void __nft_release_hooks(struct net *net)
10000 {
10001         struct nftables_pernet *nft_net = nft_pernet(net);
10002         struct nft_table *table;
10003
10004         list_for_each_entry(table, &nft_net->tables, list) {
10005                 if (nft_table_has_owner(table))
10006                         continue;
10007
10008                 __nft_release_hook(net, table);
10009         }
10010 }
10011
10012 static void __nft_release_table(struct net *net, struct nft_table *table)
10013 {
10014         struct nft_flowtable *flowtable, *nf;
10015         struct nft_chain *chain, *nc;
10016         struct nft_object *obj, *ne;
10017         struct nft_rule *rule, *nr;
10018         struct nft_set *set, *ns;
10019         struct nft_ctx ctx = {
10020                 .net    = net,
10021                 .family = NFPROTO_NETDEV,
10022         };
10023
10024         ctx.family = table->family;
10025         ctx.table = table;
10026         list_for_each_entry(chain, &table->chains, list) {
10027                 ctx.chain = chain;
10028                 list_for_each_entry_safe(rule, nr, &chain->rules, list) {
10029                         list_del(&rule->list);
10030                         chain->use--;
10031                         nf_tables_rule_release(&ctx, rule);
10032                 }
10033         }
10034         list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
10035                 list_del(&flowtable->list);
10036                 table->use--;
10037                 nf_tables_flowtable_destroy(flowtable);
10038         }
10039         list_for_each_entry_safe(set, ns, &table->sets, list) {
10040                 list_del(&set->list);
10041                 table->use--;
10042                 nft_set_destroy(&ctx, set);
10043         }
10044         list_for_each_entry_safe(obj, ne, &table->objects, list) {
10045                 nft_obj_del(obj);
10046                 table->use--;
10047                 nft_obj_destroy(&ctx, obj);
10048         }
10049         list_for_each_entry_safe(chain, nc, &table->chains, list) {
10050                 ctx.chain = chain;
10051                 nft_chain_del(chain);
10052                 table->use--;
10053                 nf_tables_chain_destroy(&ctx);
10054         }
10055         nf_tables_table_destroy(&ctx);
10056 }
10057
10058 static void __nft_release_tables(struct net *net)
10059 {
10060         struct nftables_pernet *nft_net = nft_pernet(net);
10061         struct nft_table *table, *nt;
10062
10063         list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
10064                 if (nft_table_has_owner(table))
10065                         continue;
10066
10067                 list_del(&table->list);
10068
10069                 __nft_release_table(net, table);
10070         }
10071 }
10072
10073 static int nft_rcv_nl_event(struct notifier_block *this, unsigned long event,
10074                             void *ptr)
10075 {
10076         struct nft_table *table, *to_delete[8];
10077         struct nftables_pernet *nft_net;
10078         struct netlink_notify *n = ptr;
10079         struct net *net = n->net;
10080         unsigned int deleted;
10081         bool restart = false;
10082
10083         if (event != NETLINK_URELEASE || n->protocol != NETLINK_NETFILTER)
10084                 return NOTIFY_DONE;
10085
10086         nft_net = nft_pernet(net);
10087         deleted = 0;
10088         mutex_lock(&nft_net->commit_mutex);
10089         if (!list_empty(&nf_tables_destroy_list))
10090                 rcu_barrier();
10091 again:
10092         list_for_each_entry(table, &nft_net->tables, list) {
10093                 if (nft_table_has_owner(table) &&
10094                     n->portid == table->nlpid) {
10095                         __nft_release_hook(net, table);
10096                         list_del_rcu(&table->list);
10097                         to_delete[deleted++] = table;
10098                         if (deleted >= ARRAY_SIZE(to_delete))
10099                                 break;
10100                 }
10101         }
10102         if (deleted) {
10103                 restart = deleted >= ARRAY_SIZE(to_delete);
10104                 synchronize_rcu();
10105                 while (deleted)
10106                         __nft_release_table(net, to_delete[--deleted]);
10107
10108                 if (restart)
10109                         goto again;
10110         }
10111         mutex_unlock(&nft_net->commit_mutex);
10112
10113         return NOTIFY_DONE;
10114 }
10115
10116 static struct notifier_block nft_nl_notifier = {
10117         .notifier_call  = nft_rcv_nl_event,
10118 };
10119
10120 static int __net_init nf_tables_init_net(struct net *net)
10121 {
10122         struct nftables_pernet *nft_net = nft_pernet(net);
10123
10124         INIT_LIST_HEAD(&nft_net->tables);
10125         INIT_LIST_HEAD(&nft_net->commit_list);
10126         INIT_LIST_HEAD(&nft_net->module_list);
10127         INIT_LIST_HEAD(&nft_net->notify_list);
10128         mutex_init(&nft_net->commit_mutex);
10129         nft_net->base_seq = 1;
10130         nft_net->validate_state = NFT_VALIDATE_SKIP;
10131
10132         return 0;
10133 }
10134
10135 static void __net_exit nf_tables_pre_exit_net(struct net *net)
10136 {
10137         struct nftables_pernet *nft_net = nft_pernet(net);
10138
10139         mutex_lock(&nft_net->commit_mutex);
10140         __nft_release_hooks(net);
10141         mutex_unlock(&nft_net->commit_mutex);
10142 }
10143
10144 static void __net_exit nf_tables_exit_net(struct net *net)
10145 {
10146         struct nftables_pernet *nft_net = nft_pernet(net);
10147
10148         mutex_lock(&nft_net->commit_mutex);
10149         if (!list_empty(&nft_net->commit_list) ||
10150             !list_empty(&nft_net->module_list))
10151                 __nf_tables_abort(net, NFNL_ABORT_NONE);
10152         __nft_release_tables(net);
10153         mutex_unlock(&nft_net->commit_mutex);
10154         WARN_ON_ONCE(!list_empty(&nft_net->tables));
10155         WARN_ON_ONCE(!list_empty(&nft_net->module_list));
10156         WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
10157 }
10158
10159 static struct pernet_operations nf_tables_net_ops = {
10160         .init           = nf_tables_init_net,
10161         .pre_exit       = nf_tables_pre_exit_net,
10162         .exit           = nf_tables_exit_net,
10163         .id             = &nf_tables_net_id,
10164         .size           = sizeof(struct nftables_pernet),
10165 };
10166
10167 static int __init nf_tables_module_init(void)
10168 {
10169         int err;
10170
10171         err = register_pernet_subsys(&nf_tables_net_ops);
10172         if (err < 0)
10173                 return err;
10174
10175         err = nft_chain_filter_init();
10176         if (err < 0)
10177                 goto err_chain_filter;
10178
10179         err = nf_tables_core_module_init();
10180         if (err < 0)
10181                 goto err_core_module;
10182
10183         err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
10184         if (err < 0)
10185                 goto err_netdev_notifier;
10186
10187         err = rhltable_init(&nft_objname_ht, &nft_objname_ht_params);
10188         if (err < 0)
10189                 goto err_rht_objname;
10190
10191         err = nft_offload_init();
10192         if (err < 0)
10193                 goto err_offload;
10194
10195         err = netlink_register_notifier(&nft_nl_notifier);
10196         if (err < 0)
10197                 goto err_netlink_notifier;
10198
10199         /* must be last */
10200         err = nfnetlink_subsys_register(&nf_tables_subsys);
10201         if (err < 0)
10202                 goto err_nfnl_subsys;
10203
10204         nft_chain_route_init();
10205
10206         return err;
10207
10208 err_nfnl_subsys:
10209         netlink_unregister_notifier(&nft_nl_notifier);
10210 err_netlink_notifier:
10211         nft_offload_exit();
10212 err_offload:
10213         rhltable_destroy(&nft_objname_ht);
10214 err_rht_objname:
10215         unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
10216 err_netdev_notifier:
10217         nf_tables_core_module_exit();
10218 err_core_module:
10219         nft_chain_filter_fini();
10220 err_chain_filter:
10221         unregister_pernet_subsys(&nf_tables_net_ops);
10222         return err;
10223 }
10224
10225 static void __exit nf_tables_module_exit(void)
10226 {
10227         nfnetlink_subsys_unregister(&nf_tables_subsys);
10228         netlink_unregister_notifier(&nft_nl_notifier);
10229         nft_offload_exit();
10230         unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
10231         nft_chain_filter_fini();
10232         nft_chain_route_fini();
10233         unregister_pernet_subsys(&nf_tables_net_ops);
10234         cancel_work_sync(&trans_destroy_work);
10235         rcu_barrier();
10236         rhltable_destroy(&nft_objname_ht);
10237         nf_tables_core_module_exit();
10238 }
10239
10240 module_init(nf_tables_module_init);
10241 module_exit(nf_tables_module_exit);
10242
10243 MODULE_LICENSE("GPL");
10244 MODULE_AUTHOR("Patrick McHardy <[email protected]>");
10245 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);
This page took 0.629422 seconds and 4 git commands to generate.