]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blame_incremental - net/netfilter/nf_tables_api.c
projects / mirror_ubuntu-jammy-kernel.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
netfilter: nf_tables: mark stateful expressions
[mirror_ubuntu-jammy-kernel.git] / net / netfilter / nf_tables_api.c
... / ...
CommitLineData
1/*
2 * Copyright (c) 2007-2009 Patrick McHardy <kaber@trash.net>
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License version 2 as
6 * published by the Free Software Foundation.
7 *
8 * Development of this code funded by Astaro AG (http://www.astaro.com/)
9 */
10
11#include <linux/module.h>
12#include <linux/init.h>
13#include <linux/list.h>
14#include <linux/skbuff.h>
15#include <linux/netlink.h>
16#include <linux/netfilter.h>
17#include <linux/netfilter/nfnetlink.h>
18#include <linux/netfilter/nf_tables.h>
19#include <net/netfilter/nf_tables_core.h>
20#include <net/netfilter/nf_tables.h>
21#include <net/net_namespace.h>
22#include <net/sock.h>
23
24static LIST_HEAD(nf_tables_expressions);
25
26/**
27 * nft_register_afinfo - register nf_tables address family info
28 *
29 * @afi: address family info to register
30 *
31 * Register the address family for use with nf_tables. Returns zero on
32 * success or a negative errno code otherwise.
33 */
34int nft_register_afinfo(struct net *net, struct nft_af_info *afi)
35{
36 INIT_LIST_HEAD(&afi->tables);
37 nfnl_lock(NFNL_SUBSYS_NFTABLES);
38 list_add_tail_rcu(&afi->list, &net->nft.af_info);
39 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
40 return 0;
41}
42EXPORT_SYMBOL_GPL(nft_register_afinfo);
43
44/**
45 * nft_unregister_afinfo - unregister nf_tables address family info
46 *
47 * @afi: address family info to unregister
48 *
49 * Unregister the address family for use with nf_tables.
50 */
51void nft_unregister_afinfo(struct nft_af_info *afi)
52{
53 nfnl_lock(NFNL_SUBSYS_NFTABLES);
54 list_del_rcu(&afi->list);
55 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
56}
57EXPORT_SYMBOL_GPL(nft_unregister_afinfo);
58
59static struct nft_af_info *nft_afinfo_lookup(struct net *net, int family)
60{
61 struct nft_af_info *afi;
62
63 list_for_each_entry(afi, &net->nft.af_info, list) {
64 if (afi->family == family)
65 return afi;
66 }
67 return NULL;
68}
69
70static struct nft_af_info *
71nf_tables_afinfo_lookup(struct net *net, int family, bool autoload)
72{
73 struct nft_af_info *afi;
74
75 afi = nft_afinfo_lookup(net, family);
76 if (afi != NULL)
77 return afi;
78#ifdef CONFIG_MODULES
79 if (autoload) {
80 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
81 request_module("nft-afinfo-%u", family);
82 nfnl_lock(NFNL_SUBSYS_NFTABLES);
83 afi = nft_afinfo_lookup(net, family);
84 if (afi != NULL)
85 return ERR_PTR(-EAGAIN);
86 }
87#endif
88 return ERR_PTR(-EAFNOSUPPORT);
89}
90
91static void nft_ctx_init(struct nft_ctx *ctx,
92 const struct sk_buff *skb,
93 const struct nlmsghdr *nlh,
94 struct nft_af_info *afi,
95 struct nft_table *table,
96 struct nft_chain *chain,
97 const struct nlattr * const *nla)
98{
99 ctx->net = sock_net(skb->sk);
100 ctx->afi = afi;
101 ctx->table = table;
102 ctx->chain = chain;
103 ctx->nla = nla;
104 ctx->portid = NETLINK_CB(skb).portid;
105 ctx->report = nlmsg_report(nlh);
106 ctx->seq = nlh->nlmsg_seq;
107}
108
109static struct nft_trans *nft_trans_alloc(struct nft_ctx *ctx, int msg_type,
110 u32 size)
111{
112 struct nft_trans *trans;
113
114 trans = kzalloc(sizeof(struct nft_trans) + size, GFP_KERNEL);
115 if (trans == NULL)
116 return NULL;
117
118 trans->msg_type = msg_type;
119 trans->ctx = *ctx;
120
121 return trans;
122}
123
124static void nft_trans_destroy(struct nft_trans *trans)
125{
126 list_del(&trans->list);
127 kfree(trans);
128}
129
130static void nf_tables_unregister_hooks(const struct nft_table *table,
131 const struct nft_chain *chain,
132 unsigned int hook_nops)
133{
134 if (!(table->flags & NFT_TABLE_F_DORMANT) &&
135 chain->flags & NFT_BASE_CHAIN)
136 nf_unregister_hooks(nft_base_chain(chain)->ops, hook_nops);
137}
138
139/* Internal table flags */
140#define NFT_TABLE_INACTIVE (1 << 15)
141
142static int nft_trans_table_add(struct nft_ctx *ctx, int msg_type)
143{
144 struct nft_trans *trans;
145
146 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_table));
147 if (trans == NULL)
148 return -ENOMEM;
149
150 if (msg_type == NFT_MSG_NEWTABLE)
151 ctx->table->flags |= NFT_TABLE_INACTIVE;
152
153 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
154 return 0;
155}
156
157static int nft_deltable(struct nft_ctx *ctx)
158{
159 int err;
160
161 err = nft_trans_table_add(ctx, NFT_MSG_DELTABLE);
162 if (err < 0)
163 return err;
164
165 list_del_rcu(&ctx->table->list);
166 return err;
167}
168
169static int nft_trans_chain_add(struct nft_ctx *ctx, int msg_type)
170{
171 struct nft_trans *trans;
172
173 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_chain));
174 if (trans == NULL)
175 return -ENOMEM;
176
177 if (msg_type == NFT_MSG_NEWCHAIN)
178 ctx->chain->flags |= NFT_CHAIN_INACTIVE;
179
180 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
181 return 0;
182}
183
184static int nft_delchain(struct nft_ctx *ctx)
185{
186 int err;
187
188 err = nft_trans_chain_add(ctx, NFT_MSG_DELCHAIN);
189 if (err < 0)
190 return err;
191
192 ctx->table->use--;
193 list_del_rcu(&ctx->chain->list);
194
195 return err;
196}
197
198static inline bool
199nft_rule_is_active(struct net *net, const struct nft_rule *rule)
200{
201 return (rule->genmask & nft_genmask_cur(net)) == 0;
202}
203
204static inline int
205nft_rule_is_active_next(struct net *net, const struct nft_rule *rule)
206{
207 return (rule->genmask & nft_genmask_next(net)) == 0;
208}
209
210static inline void
211nft_rule_activate_next(struct net *net, struct nft_rule *rule)
212{
213 /* Now inactive, will be active in the future */
214 rule->genmask = nft_genmask_cur(net);
215}
216
217static inline void
218nft_rule_deactivate_next(struct net *net, struct nft_rule *rule)
219{
220 rule->genmask = nft_genmask_next(net);
221}
222
223static inline void nft_rule_clear(struct net *net, struct nft_rule *rule)
224{
225 rule->genmask &= ~nft_genmask_next(net);
226}
227
228static int
229nf_tables_delrule_deactivate(struct nft_ctx *ctx, struct nft_rule *rule)
230{
231 /* You cannot delete the same rule twice */
232 if (nft_rule_is_active_next(ctx->net, rule)) {
233 nft_rule_deactivate_next(ctx->net, rule);
234 ctx->chain->use--;
235 return 0;
236 }
237 return -ENOENT;
238}
239
240static struct nft_trans *nft_trans_rule_add(struct nft_ctx *ctx, int msg_type,
241 struct nft_rule *rule)
242{
243 struct nft_trans *trans;
244
245 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_rule));
246 if (trans == NULL)
247 return NULL;
248
249 nft_trans_rule(trans) = rule;
250 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
251
252 return trans;
253}
254
255static int nft_delrule(struct nft_ctx *ctx, struct nft_rule *rule)
256{
257 struct nft_trans *trans;
258 int err;
259
260 trans = nft_trans_rule_add(ctx, NFT_MSG_DELRULE, rule);
261 if (trans == NULL)
262 return -ENOMEM;
263
264 err = nf_tables_delrule_deactivate(ctx, rule);
265 if (err < 0) {
266 nft_trans_destroy(trans);
267 return err;
268 }
269
270 return 0;
271}
272
273static int nft_delrule_by_chain(struct nft_ctx *ctx)
274{
275 struct nft_rule *rule;
276 int err;
277
278 list_for_each_entry(rule, &ctx->chain->rules, list) {
279 err = nft_delrule(ctx, rule);
280 if (err < 0)
281 return err;
282 }
283 return 0;
284}
285
286/* Internal set flag */
287#define NFT_SET_INACTIVE (1 << 15)
288
289static int nft_trans_set_add(struct nft_ctx *ctx, int msg_type,
290 struct nft_set *set)
291{
292 struct nft_trans *trans;
293
294 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_set));
295 if (trans == NULL)
296 return -ENOMEM;
297
298 if (msg_type == NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID] != NULL) {
299 nft_trans_set_id(trans) =
300 ntohl(nla_get_be32(ctx->nla[NFTA_SET_ID]));
301 set->flags |= NFT_SET_INACTIVE;
302 }
303 nft_trans_set(trans) = set;
304 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
305
306 return 0;
307}
308
309static int nft_delset(struct nft_ctx *ctx, struct nft_set *set)
310{
311 int err;
312
313 err = nft_trans_set_add(ctx, NFT_MSG_DELSET, set);
314 if (err < 0)
315 return err;
316
317 list_del_rcu(&set->list);
318 ctx->table->use--;
319
320 return err;
321}
322
323/*
324 * Tables
325 */
326
327static struct nft_table *nft_table_lookup(const struct nft_af_info *afi,
328 const struct nlattr *nla)
329{
330 struct nft_table *table;
331
332 list_for_each_entry(table, &afi->tables, list) {
333 if (!nla_strcmp(nla, table->name))
334 return table;
335 }
336 return NULL;
337}
338
339static struct nft_table *nf_tables_table_lookup(const struct nft_af_info *afi,
340 const struct nlattr *nla)
341{
342 struct nft_table *table;
343
344 if (nla == NULL)
345 return ERR_PTR(-EINVAL);
346
347 table = nft_table_lookup(afi, nla);
348 if (table != NULL)
349 return table;
350
351 return ERR_PTR(-ENOENT);
352}
353
354static inline u64 nf_tables_alloc_handle(struct nft_table *table)
355{
356 return ++table->hgenerator;
357}
358
359static const struct nf_chain_type *chain_type[AF_MAX][NFT_CHAIN_T_MAX];
360
361static const struct nf_chain_type *
362__nf_tables_chain_type_lookup(int family, const struct nlattr *nla)
363{
364 int i;
365
366 for (i = 0; i < NFT_CHAIN_T_MAX; i++) {
367 if (chain_type[family][i] != NULL &&
368 !nla_strcmp(nla, chain_type[family][i]->name))
369 return chain_type[family][i];
370 }
371 return NULL;
372}
373
374static const struct nf_chain_type *
375nf_tables_chain_type_lookup(const struct nft_af_info *afi,
376 const struct nlattr *nla,
377 bool autoload)
378{
379 const struct nf_chain_type *type;
380
381 type = __nf_tables_chain_type_lookup(afi->family, nla);
382 if (type != NULL)
383 return type;
384#ifdef CONFIG_MODULES
385 if (autoload) {
386 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
387 request_module("nft-chain-%u-%.*s", afi->family,
388 nla_len(nla), (const char *)nla_data(nla));
389 nfnl_lock(NFNL_SUBSYS_NFTABLES);
390 type = __nf_tables_chain_type_lookup(afi->family, nla);
391 if (type != NULL)
392 return ERR_PTR(-EAGAIN);
393 }
394#endif
395 return ERR_PTR(-ENOENT);
396}
397
398static const struct nla_policy nft_table_policy[NFTA_TABLE_MAX + 1] = {
399 [NFTA_TABLE_NAME] = { .type = NLA_STRING,
400 .len = NFT_TABLE_MAXNAMELEN - 1 },
401 [NFTA_TABLE_FLAGS] = { .type = NLA_U32 },
402};
403
404static int nf_tables_fill_table_info(struct sk_buff *skb, struct net *net,
405 u32 portid, u32 seq, int event, u32 flags,
406 int family, const struct nft_table *table)
407{
408 struct nlmsghdr *nlh;
409 struct nfgenmsg *nfmsg;
410
411 event |= NFNL_SUBSYS_NFTABLES << 8;
412 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
413 if (nlh == NULL)
414 goto nla_put_failure;
415
416 nfmsg = nlmsg_data(nlh);
417 nfmsg->nfgen_family = family;
418 nfmsg->version = NFNETLINK_V0;
419 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
420
421 if (nla_put_string(skb, NFTA_TABLE_NAME, table->name) ||
422 nla_put_be32(skb, NFTA_TABLE_FLAGS, htonl(table->flags)) ||
423 nla_put_be32(skb, NFTA_TABLE_USE, htonl(table->use)))
424 goto nla_put_failure;
425
426 nlmsg_end(skb, nlh);
427 return 0;
428
429nla_put_failure:
430 nlmsg_trim(skb, nlh);
431 return -1;
432}
433
434static int nf_tables_table_notify(const struct nft_ctx *ctx, int event)
435{
436 struct sk_buff *skb;
437 int err;
438
439 if (!ctx->report &&
440 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
441 return 0;
442
443 err = -ENOBUFS;
444 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
445 if (skb == NULL)
446 goto err;
447
448 err = nf_tables_fill_table_info(skb, ctx->net, ctx->portid, ctx->seq,
449 event, 0, ctx->afi->family, ctx->table);
450 if (err < 0) {
451 kfree_skb(skb);
452 goto err;
453 }
454
455 err = nfnetlink_send(skb, ctx->net, ctx->portid, NFNLGRP_NFTABLES,
456 ctx->report, GFP_KERNEL);
457err:
458 if (err < 0) {
459 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES,
460 err);
461 }
462 return err;
463}
464
465static int nf_tables_dump_tables(struct sk_buff *skb,
466 struct netlink_callback *cb)
467{
468 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
469 const struct nft_af_info *afi;
470 const struct nft_table *table;
471 unsigned int idx = 0, s_idx = cb->args[0];
472 struct net *net = sock_net(skb->sk);
473 int family = nfmsg->nfgen_family;
474
475 rcu_read_lock();
476 cb->seq = net->nft.base_seq;
477
478 list_for_each_entry_rcu(afi, &net->nft.af_info, list) {
479 if (family != NFPROTO_UNSPEC && family != afi->family)
480 continue;
481
482 list_for_each_entry_rcu(table, &afi->tables, list) {
483 if (idx < s_idx)
484 goto cont;
485 if (idx > s_idx)
486 memset(&cb->args[1], 0,
487 sizeof(cb->args) - sizeof(cb->args[0]));
488 if (nf_tables_fill_table_info(skb, net,
489 NETLINK_CB(cb->skb).portid,
490 cb->nlh->nlmsg_seq,
491 NFT_MSG_NEWTABLE,
492 NLM_F_MULTI,
493 afi->family, table) < 0)
494 goto done;
495
496 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
497cont:
498 idx++;
499 }
500 }
501done:
502 rcu_read_unlock();
503 cb->args[0] = idx;
504 return skb->len;
505}
506
507static int nf_tables_gettable(struct sock *nlsk, struct sk_buff *skb,
508 const struct nlmsghdr *nlh,
509 const struct nlattr * const nla[])
510{
511 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
512 const struct nft_af_info *afi;
513 const struct nft_table *table;
514 struct sk_buff *skb2;
515 struct net *net = sock_net(skb->sk);
516 int family = nfmsg->nfgen_family;
517 int err;
518
519 if (nlh->nlmsg_flags & NLM_F_DUMP) {
520 struct netlink_dump_control c = {
521 .dump = nf_tables_dump_tables,
522 };
523 return netlink_dump_start(nlsk, skb, nlh, &c);
524 }
525
526 afi = nf_tables_afinfo_lookup(net, family, false);
527 if (IS_ERR(afi))
528 return PTR_ERR(afi);
529
530 table = nf_tables_table_lookup(afi, nla[NFTA_TABLE_NAME]);
531 if (IS_ERR(table))
532 return PTR_ERR(table);
533 if (table->flags & NFT_TABLE_INACTIVE)
534 return -ENOENT;
535
536 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
537 if (!skb2)
538 return -ENOMEM;
539
540 err = nf_tables_fill_table_info(skb2, net, NETLINK_CB(skb).portid,
541 nlh->nlmsg_seq, NFT_MSG_NEWTABLE, 0,
542 family, table);
543 if (err < 0)
544 goto err;
545
546 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
547
548err:
549 kfree_skb(skb2);
550 return err;
551}
552
553static int nf_tables_table_enable(const struct nft_af_info *afi,
554 struct nft_table *table)
555{
556 struct nft_chain *chain;
557 int err, i = 0;
558
559 list_for_each_entry(chain, &table->chains, list) {
560 if (!(chain->flags & NFT_BASE_CHAIN))
561 continue;
562
563 err = nf_register_hooks(nft_base_chain(chain)->ops, afi->nops);
564 if (err < 0)
565 goto err;
566
567 i++;
568 }
569 return 0;
570err:
571 list_for_each_entry(chain, &table->chains, list) {
572 if (!(chain->flags & NFT_BASE_CHAIN))
573 continue;
574
575 if (i-- <= 0)
576 break;
577
578 nf_unregister_hooks(nft_base_chain(chain)->ops, afi->nops);
579 }
580 return err;
581}
582
583static void nf_tables_table_disable(const struct nft_af_info *afi,
584 struct nft_table *table)
585{
586 struct nft_chain *chain;
587
588 list_for_each_entry(chain, &table->chains, list) {
589 if (chain->flags & NFT_BASE_CHAIN)
590 nf_unregister_hooks(nft_base_chain(chain)->ops,
591 afi->nops);
592 }
593}
594
595static int nf_tables_updtable(struct nft_ctx *ctx)
596{
597 struct nft_trans *trans;
598 u32 flags;
599 int ret = 0;
600
601 if (!ctx->nla[NFTA_TABLE_FLAGS])
602 return 0;
603
604 flags = ntohl(nla_get_be32(ctx->nla[NFTA_TABLE_FLAGS]));
605 if (flags & ~NFT_TABLE_F_DORMANT)
606 return -EINVAL;
607
608 if (flags == ctx->table->flags)
609 return 0;
610
611 trans = nft_trans_alloc(ctx, NFT_MSG_NEWTABLE,
612 sizeof(struct nft_trans_table));
613 if (trans == NULL)
614 return -ENOMEM;
615
616 if ((flags & NFT_TABLE_F_DORMANT) &&
617 !(ctx->table->flags & NFT_TABLE_F_DORMANT)) {
618 nft_trans_table_enable(trans) = false;
619 } else if (!(flags & NFT_TABLE_F_DORMANT) &&
620 ctx->table->flags & NFT_TABLE_F_DORMANT) {
621 ret = nf_tables_table_enable(ctx->afi, ctx->table);
622 if (ret >= 0) {
623 ctx->table->flags &= ~NFT_TABLE_F_DORMANT;
624 nft_trans_table_enable(trans) = true;
625 }
626 }
627 if (ret < 0)
628 goto err;
629
630 nft_trans_table_update(trans) = true;
631 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
632 return 0;
633err:
634 nft_trans_destroy(trans);
635 return ret;
636}
637
638static int nf_tables_newtable(struct sock *nlsk, struct sk_buff *skb,
639 const struct nlmsghdr *nlh,
640 const struct nlattr * const nla[])
641{
642 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
643 const struct nlattr *name;
644 struct nft_af_info *afi;
645 struct nft_table *table;
646 struct net *net = sock_net(skb->sk);
647 int family = nfmsg->nfgen_family;
648 u32 flags = 0;
649 struct nft_ctx ctx;
650 int err;
651
652 afi = nf_tables_afinfo_lookup(net, family, true);
653 if (IS_ERR(afi))
654 return PTR_ERR(afi);
655
656 name = nla[NFTA_TABLE_NAME];
657 table = nf_tables_table_lookup(afi, name);
658 if (IS_ERR(table)) {
659 if (PTR_ERR(table) != -ENOENT)
660 return PTR_ERR(table);
661 table = NULL;
662 }
663
664 if (table != NULL) {
665 if (table->flags & NFT_TABLE_INACTIVE)
666 return -ENOENT;
667 if (nlh->nlmsg_flags & NLM_F_EXCL)
668 return -EEXIST;
669 if (nlh->nlmsg_flags & NLM_F_REPLACE)
670 return -EOPNOTSUPP;
671
672 nft_ctx_init(&ctx, skb, nlh, afi, table, NULL, nla);
673 return nf_tables_updtable(&ctx);
674 }
675
676 if (nla[NFTA_TABLE_FLAGS]) {
677 flags = ntohl(nla_get_be32(nla[NFTA_TABLE_FLAGS]));
678 if (flags & ~NFT_TABLE_F_DORMANT)
679 return -EINVAL;
680 }
681
682 if (!try_module_get(afi->owner))
683 return -EAFNOSUPPORT;
684
685 err = -ENOMEM;
686 table = kzalloc(sizeof(*table), GFP_KERNEL);
687 if (table == NULL)
688 goto err1;
689
690 nla_strlcpy(table->name, name, NFT_TABLE_MAXNAMELEN);
691 INIT_LIST_HEAD(&table->chains);
692 INIT_LIST_HEAD(&table->sets);
693 table->flags = flags;
694
695 nft_ctx_init(&ctx, skb, nlh, afi, table, NULL, nla);
696 err = nft_trans_table_add(&ctx, NFT_MSG_NEWTABLE);
697 if (err < 0)
698 goto err2;
699
700 list_add_tail_rcu(&table->list, &afi->tables);
701 return 0;
702err2:
703 kfree(table);
704err1:
705 module_put(afi->owner);
706 return err;
707}
708
709static int nft_flush_table(struct nft_ctx *ctx)
710{
711 int err;
712 struct nft_chain *chain, *nc;
713 struct nft_set *set, *ns;
714
715 list_for_each_entry(chain, &ctx->table->chains, list) {
716 ctx->chain = chain;
717
718 err = nft_delrule_by_chain(ctx);
719 if (err < 0)
720 goto out;
721 }
722
723 list_for_each_entry_safe(set, ns, &ctx->table->sets, list) {
724 if (set->flags & NFT_SET_ANONYMOUS &&
725 !list_empty(&set->bindings))
726 continue;
727
728 err = nft_delset(ctx, set);
729 if (err < 0)
730 goto out;
731 }
732
733 list_for_each_entry_safe(chain, nc, &ctx->table->chains, list) {
734 ctx->chain = chain;
735
736 err = nft_delchain(ctx);
737 if (err < 0)
738 goto out;
739 }
740
741 err = nft_deltable(ctx);
742out:
743 return err;
744}
745
746static int nft_flush(struct nft_ctx *ctx, int family)
747{
748 struct nft_af_info *afi;
749 struct nft_table *table, *nt;
750 const struct nlattr * const *nla = ctx->nla;
751 int err = 0;
752
753 list_for_each_entry(afi, &ctx->net->nft.af_info, list) {
754 if (family != AF_UNSPEC && afi->family != family)
755 continue;
756
757 ctx->afi = afi;
758 list_for_each_entry_safe(table, nt, &afi->tables, list) {
759 if (nla[NFTA_TABLE_NAME] &&
760 nla_strcmp(nla[NFTA_TABLE_NAME], table->name) != 0)
761 continue;
762
763 ctx->table = table;
764
765 err = nft_flush_table(ctx);
766 if (err < 0)
767 goto out;
768 }
769 }
770out:
771 return err;
772}
773
774static int nf_tables_deltable(struct sock *nlsk, struct sk_buff *skb,
775 const struct nlmsghdr *nlh,
776 const struct nlattr * const nla[])
777{
778 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
779 struct nft_af_info *afi;
780 struct nft_table *table;
781 struct net *net = sock_net(skb->sk);
782 int family = nfmsg->nfgen_family;
783 struct nft_ctx ctx;
784
785 nft_ctx_init(&ctx, skb, nlh, NULL, NULL, NULL, nla);
786 if (family == AF_UNSPEC || nla[NFTA_TABLE_NAME] == NULL)
787 return nft_flush(&ctx, family);
788
789 afi = nf_tables_afinfo_lookup(net, family, false);
790 if (IS_ERR(afi))
791 return PTR_ERR(afi);
792
793 table = nf_tables_table_lookup(afi, nla[NFTA_TABLE_NAME]);
794 if (IS_ERR(table))
795 return PTR_ERR(table);
796 if (table->flags & NFT_TABLE_INACTIVE)
797 return -ENOENT;
798
799 ctx.afi = afi;
800 ctx.table = table;
801
802 return nft_flush_table(&ctx);
803}
804
805static void nf_tables_table_destroy(struct nft_ctx *ctx)
806{
807 BUG_ON(ctx->table->use > 0);
808
809 kfree(ctx->table);
810 module_put(ctx->afi->owner);
811}
812
813int nft_register_chain_type(const struct nf_chain_type *ctype)
814{
815 int err = 0;
816
817 nfnl_lock(NFNL_SUBSYS_NFTABLES);
818 if (chain_type[ctype->family][ctype->type] != NULL) {
819 err = -EBUSY;
820 goto out;
821 }
822 chain_type[ctype->family][ctype->type] = ctype;
823out:
824 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
825 return err;
826}
827EXPORT_SYMBOL_GPL(nft_register_chain_type);
828
829void nft_unregister_chain_type(const struct nf_chain_type *ctype)
830{
831 nfnl_lock(NFNL_SUBSYS_NFTABLES);
832 chain_type[ctype->family][ctype->type] = NULL;
833 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
834}
835EXPORT_SYMBOL_GPL(nft_unregister_chain_type);
836
837/*
838 * Chains
839 */
840
841static struct nft_chain *
842nf_tables_chain_lookup_byhandle(const struct nft_table *table, u64 handle)
843{
844 struct nft_chain *chain;
845
846 list_for_each_entry(chain, &table->chains, list) {
847 if (chain->handle == handle)
848 return chain;
849 }
850
851 return ERR_PTR(-ENOENT);
852}
853
854static struct nft_chain *nf_tables_chain_lookup(const struct nft_table *table,
855 const struct nlattr *nla)
856{
857 struct nft_chain *chain;
858
859 if (nla == NULL)
860 return ERR_PTR(-EINVAL);
861
862 list_for_each_entry(chain, &table->chains, list) {
863 if (!nla_strcmp(nla, chain->name))
864 return chain;
865 }
866
867 return ERR_PTR(-ENOENT);
868}
869
870static const struct nla_policy nft_chain_policy[NFTA_CHAIN_MAX + 1] = {
871 [NFTA_CHAIN_TABLE] = { .type = NLA_STRING },
872 [NFTA_CHAIN_HANDLE] = { .type = NLA_U64 },
873 [NFTA_CHAIN_NAME] = { .type = NLA_STRING,
874 .len = NFT_CHAIN_MAXNAMELEN - 1 },
875 [NFTA_CHAIN_HOOK] = { .type = NLA_NESTED },
876 [NFTA_CHAIN_POLICY] = { .type = NLA_U32 },
877 [NFTA_CHAIN_TYPE] = { .type = NLA_STRING },
878 [NFTA_CHAIN_COUNTERS] = { .type = NLA_NESTED },
879};
880
881static const struct nla_policy nft_hook_policy[NFTA_HOOK_MAX + 1] = {
882 [NFTA_HOOK_HOOKNUM] = { .type = NLA_U32 },
883 [NFTA_HOOK_PRIORITY] = { .type = NLA_U32 },
884};
885
886static int nft_dump_stats(struct sk_buff *skb, struct nft_stats __percpu *stats)
887{
888 struct nft_stats *cpu_stats, total;
889 struct nlattr *nest;
890 unsigned int seq;
891 u64 pkts, bytes;
892 int cpu;
893
894 memset(&total, 0, sizeof(total));
895 for_each_possible_cpu(cpu) {
896 cpu_stats = per_cpu_ptr(stats, cpu);
897 do {
898 seq = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
899 pkts = cpu_stats->pkts;
900 bytes = cpu_stats->bytes;
901 } while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, seq));
902 total.pkts += pkts;
903 total.bytes += bytes;
904 }
905 nest = nla_nest_start(skb, NFTA_CHAIN_COUNTERS);
906 if (nest == NULL)
907 goto nla_put_failure;
908
909 if (nla_put_be64(skb, NFTA_COUNTER_PACKETS, cpu_to_be64(total.pkts)) ||
910 nla_put_be64(skb, NFTA_COUNTER_BYTES, cpu_to_be64(total.bytes)))
911 goto nla_put_failure;
912
913 nla_nest_end(skb, nest);
914 return 0;
915
916nla_put_failure:
917 return -ENOSPC;
918}
919
920static int nf_tables_fill_chain_info(struct sk_buff *skb, struct net *net,
921 u32 portid, u32 seq, int event, u32 flags,
922 int family, const struct nft_table *table,
923 const struct nft_chain *chain)
924{
925 struct nlmsghdr *nlh;
926 struct nfgenmsg *nfmsg;
927
928 event |= NFNL_SUBSYS_NFTABLES << 8;
929 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
930 if (nlh == NULL)
931 goto nla_put_failure;
932
933 nfmsg = nlmsg_data(nlh);
934 nfmsg->nfgen_family = family;
935 nfmsg->version = NFNETLINK_V0;
936 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
937
938 if (nla_put_string(skb, NFTA_CHAIN_TABLE, table->name))
939 goto nla_put_failure;
940 if (nla_put_be64(skb, NFTA_CHAIN_HANDLE, cpu_to_be64(chain->handle)))
941 goto nla_put_failure;
942 if (nla_put_string(skb, NFTA_CHAIN_NAME, chain->name))
943 goto nla_put_failure;
944
945 if (chain->flags & NFT_BASE_CHAIN) {
946 const struct nft_base_chain *basechain = nft_base_chain(chain);
947 const struct nf_hook_ops *ops = &basechain->ops[0];
948 struct nlattr *nest;
949
950 nest = nla_nest_start(skb, NFTA_CHAIN_HOOK);
951 if (nest == NULL)
952 goto nla_put_failure;
953 if (nla_put_be32(skb, NFTA_HOOK_HOOKNUM, htonl(ops->hooknum)))
954 goto nla_put_failure;
955 if (nla_put_be32(skb, NFTA_HOOK_PRIORITY, htonl(ops->priority)))
956 goto nla_put_failure;
957 nla_nest_end(skb, nest);
958
959 if (nla_put_be32(skb, NFTA_CHAIN_POLICY,
960 htonl(basechain->policy)))
961 goto nla_put_failure;
962
963 if (nla_put_string(skb, NFTA_CHAIN_TYPE, basechain->type->name))
964 goto nla_put_failure;
965
966 if (nft_dump_stats(skb, nft_base_chain(chain)->stats))
967 goto nla_put_failure;
968 }
969
970 if (nla_put_be32(skb, NFTA_CHAIN_USE, htonl(chain->use)))
971 goto nla_put_failure;
972
973 nlmsg_end(skb, nlh);
974 return 0;
975
976nla_put_failure:
977 nlmsg_trim(skb, nlh);
978 return -1;
979}
980
981static int nf_tables_chain_notify(const struct nft_ctx *ctx, int event)
982{
983 struct sk_buff *skb;
984 int err;
985
986 if (!ctx->report &&
987 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
988 return 0;
989
990 err = -ENOBUFS;
991 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
992 if (skb == NULL)
993 goto err;
994
995 err = nf_tables_fill_chain_info(skb, ctx->net, ctx->portid, ctx->seq,
996 event, 0, ctx->afi->family, ctx->table,
997 ctx->chain);
998 if (err < 0) {
999 kfree_skb(skb);
1000 goto err;
1001 }
1002
1003 err = nfnetlink_send(skb, ctx->net, ctx->portid, NFNLGRP_NFTABLES,
1004 ctx->report, GFP_KERNEL);
1005err:
1006 if (err < 0) {
1007 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES,
1008 err);
1009 }
1010 return err;
1011}
1012
1013static int nf_tables_dump_chains(struct sk_buff *skb,
1014 struct netlink_callback *cb)
1015{
1016 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1017 const struct nft_af_info *afi;
1018 const struct nft_table *table;
1019 const struct nft_chain *chain;
1020 unsigned int idx = 0, s_idx = cb->args[0];
1021 struct net *net = sock_net(skb->sk);
1022 int family = nfmsg->nfgen_family;
1023
1024 rcu_read_lock();
1025 cb->seq = net->nft.base_seq;
1026
1027 list_for_each_entry_rcu(afi, &net->nft.af_info, list) {
1028 if (family != NFPROTO_UNSPEC && family != afi->family)
1029 continue;
1030
1031 list_for_each_entry_rcu(table, &afi->tables, list) {
1032 list_for_each_entry_rcu(chain, &table->chains, list) {
1033 if (idx < s_idx)
1034 goto cont;
1035 if (idx > s_idx)
1036 memset(&cb->args[1], 0,
1037 sizeof(cb->args) - sizeof(cb->args[0]));
1038 if (nf_tables_fill_chain_info(skb, net,
1039 NETLINK_CB(cb->skb).portid,
1040 cb->nlh->nlmsg_seq,
1041 NFT_MSG_NEWCHAIN,
1042 NLM_F_MULTI,
1043 afi->family, table, chain) < 0)
1044 goto done;
1045
1046 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1047cont:
1048 idx++;
1049 }
1050 }
1051 }
1052done:
1053 rcu_read_unlock();
1054 cb->args[0] = idx;
1055 return skb->len;
1056}
1057
1058static int nf_tables_getchain(struct sock *nlsk, struct sk_buff *skb,
1059 const struct nlmsghdr *nlh,
1060 const struct nlattr * const nla[])
1061{
1062 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1063 const struct nft_af_info *afi;
1064 const struct nft_table *table;
1065 const struct nft_chain *chain;
1066 struct sk_buff *skb2;
1067 struct net *net = sock_net(skb->sk);
1068 int family = nfmsg->nfgen_family;
1069 int err;
1070
1071 if (nlh->nlmsg_flags & NLM_F_DUMP) {
1072 struct netlink_dump_control c = {
1073 .dump = nf_tables_dump_chains,
1074 };
1075 return netlink_dump_start(nlsk, skb, nlh, &c);
1076 }
1077
1078 afi = nf_tables_afinfo_lookup(net, family, false);
1079 if (IS_ERR(afi))
1080 return PTR_ERR(afi);
1081
1082 table = nf_tables_table_lookup(afi, nla[NFTA_CHAIN_TABLE]);
1083 if (IS_ERR(table))
1084 return PTR_ERR(table);
1085 if (table->flags & NFT_TABLE_INACTIVE)
1086 return -ENOENT;
1087
1088 chain = nf_tables_chain_lookup(table, nla[NFTA_CHAIN_NAME]);
1089 if (IS_ERR(chain))
1090 return PTR_ERR(chain);
1091 if (chain->flags & NFT_CHAIN_INACTIVE)
1092 return -ENOENT;
1093
1094 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
1095 if (!skb2)
1096 return -ENOMEM;
1097
1098 err = nf_tables_fill_chain_info(skb2, net, NETLINK_CB(skb).portid,
1099 nlh->nlmsg_seq, NFT_MSG_NEWCHAIN, 0,
1100 family, table, chain);
1101 if (err < 0)
1102 goto err;
1103
1104 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
1105
1106err:
1107 kfree_skb(skb2);
1108 return err;
1109}
1110
1111static const struct nla_policy nft_counter_policy[NFTA_COUNTER_MAX + 1] = {
1112 [NFTA_COUNTER_PACKETS] = { .type = NLA_U64 },
1113 [NFTA_COUNTER_BYTES] = { .type = NLA_U64 },
1114};
1115
1116static struct nft_stats __percpu *nft_stats_alloc(const struct nlattr *attr)
1117{
1118 struct nlattr *tb[NFTA_COUNTER_MAX+1];
1119 struct nft_stats __percpu *newstats;
1120 struct nft_stats *stats;
1121 int err;
1122
1123 err = nla_parse_nested(tb, NFTA_COUNTER_MAX, attr, nft_counter_policy);
1124 if (err < 0)
1125 return ERR_PTR(err);
1126
1127 if (!tb[NFTA_COUNTER_BYTES] || !tb[NFTA_COUNTER_PACKETS])
1128 return ERR_PTR(-EINVAL);
1129
1130 newstats = netdev_alloc_pcpu_stats(struct nft_stats);
1131 if (newstats == NULL)
1132 return ERR_PTR(-ENOMEM);
1133
1134 /* Restore old counters on this cpu, no problem. Per-cpu statistics
1135 * are not exposed to userspace.
1136 */
1137 preempt_disable();
1138 stats = this_cpu_ptr(newstats);
1139 stats->bytes = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_BYTES]));
1140 stats->pkts = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_PACKETS]));
1141 preempt_enable();
1142
1143 return newstats;
1144}
1145
1146static void nft_chain_stats_replace(struct nft_base_chain *chain,
1147 struct nft_stats __percpu *newstats)
1148{
1149 if (newstats == NULL)
1150 return;
1151
1152 if (chain->stats) {
1153 struct nft_stats __percpu *oldstats =
1154 nft_dereference(chain->stats);
1155
1156 rcu_assign_pointer(chain->stats, newstats);
1157 synchronize_rcu();
1158 free_percpu(oldstats);
1159 } else
1160 rcu_assign_pointer(chain->stats, newstats);
1161}
1162
1163static void nf_tables_chain_destroy(struct nft_chain *chain)
1164{
1165 BUG_ON(chain->use > 0);
1166
1167 if (chain->flags & NFT_BASE_CHAIN) {
1168 module_put(nft_base_chain(chain)->type->owner);
1169 free_percpu(nft_base_chain(chain)->stats);
1170 kfree(nft_base_chain(chain));
1171 } else {
1172 kfree(chain);
1173 }
1174}
1175
1176static int nf_tables_newchain(struct sock *nlsk, struct sk_buff *skb,
1177 const struct nlmsghdr *nlh,
1178 const struct nlattr * const nla[])
1179{
1180 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1181 const struct nlattr * uninitialized_var(name);
1182 struct nft_af_info *afi;
1183 struct nft_table *table;
1184 struct nft_chain *chain;
1185 struct nft_base_chain *basechain = NULL;
1186 struct nlattr *ha[NFTA_HOOK_MAX + 1];
1187 struct net *net = sock_net(skb->sk);
1188 int family = nfmsg->nfgen_family;
1189 u8 policy = NF_ACCEPT;
1190 u64 handle = 0;
1191 unsigned int i;
1192 struct nft_stats __percpu *stats;
1193 int err;
1194 bool create;
1195 struct nft_ctx ctx;
1196
1197 create = nlh->nlmsg_flags & NLM_F_CREATE ? true : false;
1198
1199 afi = nf_tables_afinfo_lookup(net, family, true);
1200 if (IS_ERR(afi))
1201 return PTR_ERR(afi);
1202
1203 table = nf_tables_table_lookup(afi, nla[NFTA_CHAIN_TABLE]);
1204 if (IS_ERR(table))
1205 return PTR_ERR(table);
1206
1207 chain = NULL;
1208 name = nla[NFTA_CHAIN_NAME];
1209
1210 if (nla[NFTA_CHAIN_HANDLE]) {
1211 handle = be64_to_cpu(nla_get_be64(nla[NFTA_CHAIN_HANDLE]));
1212 chain = nf_tables_chain_lookup_byhandle(table, handle);
1213 if (IS_ERR(chain))
1214 return PTR_ERR(chain);
1215 } else {
1216 chain = nf_tables_chain_lookup(table, name);
1217 if (IS_ERR(chain)) {
1218 if (PTR_ERR(chain) != -ENOENT)
1219 return PTR_ERR(chain);
1220 chain = NULL;
1221 }
1222 }
1223
1224 if (nla[NFTA_CHAIN_POLICY]) {
1225 if ((chain != NULL &&
1226 !(chain->flags & NFT_BASE_CHAIN)))
1227 return -EOPNOTSUPP;
1228
1229 if (chain == NULL &&
1230 nla[NFTA_CHAIN_HOOK] == NULL)
1231 return -EOPNOTSUPP;
1232
1233 policy = ntohl(nla_get_be32(nla[NFTA_CHAIN_POLICY]));
1234 switch (policy) {
1235 case NF_DROP:
1236 case NF_ACCEPT:
1237 break;
1238 default:
1239 return -EINVAL;
1240 }
1241 }
1242
1243 if (chain != NULL) {
1244 struct nft_stats *stats = NULL;
1245 struct nft_trans *trans;
1246
1247 if (chain->flags & NFT_CHAIN_INACTIVE)
1248 return -ENOENT;
1249 if (nlh->nlmsg_flags & NLM_F_EXCL)
1250 return -EEXIST;
1251 if (nlh->nlmsg_flags & NLM_F_REPLACE)
1252 return -EOPNOTSUPP;
1253
1254 if (nla[NFTA_CHAIN_HANDLE] && name &&
1255 !IS_ERR(nf_tables_chain_lookup(table, nla[NFTA_CHAIN_NAME])))
1256 return -EEXIST;
1257
1258 if (nla[NFTA_CHAIN_COUNTERS]) {
1259 if (!(chain->flags & NFT_BASE_CHAIN))
1260 return -EOPNOTSUPP;
1261
1262 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
1263 if (IS_ERR(stats))
1264 return PTR_ERR(stats);
1265 }
1266
1267 nft_ctx_init(&ctx, skb, nlh, afi, table, chain, nla);
1268 trans = nft_trans_alloc(&ctx, NFT_MSG_NEWCHAIN,
1269 sizeof(struct nft_trans_chain));
1270 if (trans == NULL) {
1271 free_percpu(stats);
1272 return -ENOMEM;
1273 }
1274
1275 nft_trans_chain_stats(trans) = stats;
1276 nft_trans_chain_update(trans) = true;
1277
1278 if (nla[NFTA_CHAIN_POLICY])
1279 nft_trans_chain_policy(trans) = policy;
1280 else
1281 nft_trans_chain_policy(trans) = -1;
1282
1283 if (nla[NFTA_CHAIN_HANDLE] && name) {
1284 nla_strlcpy(nft_trans_chain_name(trans), name,
1285 NFT_CHAIN_MAXNAMELEN);
1286 }
1287 list_add_tail(&trans->list, &net->nft.commit_list);
1288 return 0;
1289 }
1290
1291 if (table->use == UINT_MAX)
1292 return -EOVERFLOW;
1293
1294 if (nla[NFTA_CHAIN_HOOK]) {
1295 const struct nf_chain_type *type;
1296 struct nf_hook_ops *ops;
1297 nf_hookfn *hookfn;
1298 u32 hooknum, priority;
1299
1300 type = chain_type[family][NFT_CHAIN_T_DEFAULT];
1301 if (nla[NFTA_CHAIN_TYPE]) {
1302 type = nf_tables_chain_type_lookup(afi,
1303 nla[NFTA_CHAIN_TYPE],
1304 create);
1305 if (IS_ERR(type))
1306 return PTR_ERR(type);
1307 }
1308
1309 err = nla_parse_nested(ha, NFTA_HOOK_MAX, nla[NFTA_CHAIN_HOOK],
1310 nft_hook_policy);
1311 if (err < 0)
1312 return err;
1313 if (ha[NFTA_HOOK_HOOKNUM] == NULL ||
1314 ha[NFTA_HOOK_PRIORITY] == NULL)
1315 return -EINVAL;
1316
1317 hooknum = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
1318 if (hooknum >= afi->nhooks)
1319 return -EINVAL;
1320 priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
1321
1322 if (!(type->hook_mask & (1 << hooknum)))
1323 return -EOPNOTSUPP;
1324 if (!try_module_get(type->owner))
1325 return -ENOENT;
1326 hookfn = type->hooks[hooknum];
1327
1328 basechain = kzalloc(sizeof(*basechain), GFP_KERNEL);
1329 if (basechain == NULL) {
1330 module_put(type->owner);
1331 return -ENOMEM;
1332 }
1333
1334 if (nla[NFTA_CHAIN_COUNTERS]) {
1335 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
1336 if (IS_ERR(stats)) {
1337 module_put(type->owner);
1338 kfree(basechain);
1339 return PTR_ERR(stats);
1340 }
1341 basechain->stats = stats;
1342 } else {
1343 stats = netdev_alloc_pcpu_stats(struct nft_stats);
1344 if (stats == NULL) {
1345 module_put(type->owner);
1346 kfree(basechain);
1347 return -ENOMEM;
1348 }
1349 rcu_assign_pointer(basechain->stats, stats);
1350 }
1351
1352 write_pnet(&basechain->pnet, net);
1353 basechain->type = type;
1354 chain = &basechain->chain;
1355
1356 for (i = 0; i < afi->nops; i++) {
1357 ops = &basechain->ops[i];
1358 ops->pf = family;
1359 ops->owner = afi->owner;
1360 ops->hooknum = hooknum;
1361 ops->priority = priority;
1362 ops->priv = chain;
1363 ops->hook = afi->hooks[ops->hooknum];
1364 if (hookfn)
1365 ops->hook = hookfn;
1366 if (afi->hook_ops_init)
1367 afi->hook_ops_init(ops, i);
1368 }
1369
1370 chain->flags |= NFT_BASE_CHAIN;
1371 basechain->policy = policy;
1372 } else {
1373 chain = kzalloc(sizeof(*chain), GFP_KERNEL);
1374 if (chain == NULL)
1375 return -ENOMEM;
1376 }
1377
1378 INIT_LIST_HEAD(&chain->rules);
1379 chain->handle = nf_tables_alloc_handle(table);
1380 chain->table = table;
1381 nla_strlcpy(chain->name, name, NFT_CHAIN_MAXNAMELEN);
1382
1383 if (!(table->flags & NFT_TABLE_F_DORMANT) &&
1384 chain->flags & NFT_BASE_CHAIN) {
1385 err = nf_register_hooks(nft_base_chain(chain)->ops, afi->nops);
1386 if (err < 0)
1387 goto err1;
1388 }
1389
1390 nft_ctx_init(&ctx, skb, nlh, afi, table, chain, nla);
1391 err = nft_trans_chain_add(&ctx, NFT_MSG_NEWCHAIN);
1392 if (err < 0)
1393 goto err2;
1394
1395 table->use++;
1396 list_add_tail_rcu(&chain->list, &table->chains);
1397 return 0;
1398err2:
1399 nf_tables_unregister_hooks(table, chain, afi->nops);
1400err1:
1401 nf_tables_chain_destroy(chain);
1402 return err;
1403}
1404
1405static int nf_tables_delchain(struct sock *nlsk, struct sk_buff *skb,
1406 const struct nlmsghdr *nlh,
1407 const struct nlattr * const nla[])
1408{
1409 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1410 struct nft_af_info *afi;
1411 struct nft_table *table;
1412 struct nft_chain *chain;
1413 struct net *net = sock_net(skb->sk);
1414 int family = nfmsg->nfgen_family;
1415 struct nft_ctx ctx;
1416
1417 afi = nf_tables_afinfo_lookup(net, family, false);
1418 if (IS_ERR(afi))
1419 return PTR_ERR(afi);
1420
1421 table = nf_tables_table_lookup(afi, nla[NFTA_CHAIN_TABLE]);
1422 if (IS_ERR(table))
1423 return PTR_ERR(table);
1424 if (table->flags & NFT_TABLE_INACTIVE)
1425 return -ENOENT;
1426
1427 chain = nf_tables_chain_lookup(table, nla[NFTA_CHAIN_NAME]);
1428 if (IS_ERR(chain))
1429 return PTR_ERR(chain);
1430 if (chain->flags & NFT_CHAIN_INACTIVE)
1431 return -ENOENT;
1432 if (chain->use > 0)
1433 return -EBUSY;
1434
1435 nft_ctx_init(&ctx, skb, nlh, afi, table, chain, nla);
1436
1437 return nft_delchain(&ctx);
1438}
1439
1440/*
1441 * Expressions
1442 */
1443
1444/**
1445 * nft_register_expr - register nf_tables expr type
1446 * @ops: expr type
1447 *
1448 * Registers the expr type for use with nf_tables. Returns zero on
1449 * success or a negative errno code otherwise.
1450 */
1451int nft_register_expr(struct nft_expr_type *type)
1452{
1453 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1454 if (type->family == NFPROTO_UNSPEC)
1455 list_add_tail_rcu(&type->list, &nf_tables_expressions);
1456 else
1457 list_add_rcu(&type->list, &nf_tables_expressions);
1458 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1459 return 0;
1460}
1461EXPORT_SYMBOL_GPL(nft_register_expr);
1462
1463/**
1464 * nft_unregister_expr - unregister nf_tables expr type
1465 * @ops: expr type
1466 *
1467 * Unregisters the expr typefor use with nf_tables.
1468 */
1469void nft_unregister_expr(struct nft_expr_type *type)
1470{
1471 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1472 list_del_rcu(&type->list);
1473 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1474}
1475EXPORT_SYMBOL_GPL(nft_unregister_expr);
1476
1477static const struct nft_expr_type *__nft_expr_type_get(u8 family,
1478 struct nlattr *nla)
1479{
1480 const struct nft_expr_type *type;
1481
1482 list_for_each_entry(type, &nf_tables_expressions, list) {
1483 if (!nla_strcmp(nla, type->name) &&
1484 (!type->family || type->family == family))
1485 return type;
1486 }
1487 return NULL;
1488}
1489
1490static const struct nft_expr_type *nft_expr_type_get(u8 family,
1491 struct nlattr *nla)
1492{
1493 const struct nft_expr_type *type;
1494
1495 if (nla == NULL)
1496 return ERR_PTR(-EINVAL);
1497
1498 type = __nft_expr_type_get(family, nla);
1499 if (type != NULL && try_module_get(type->owner))
1500 return type;
1501
1502#ifdef CONFIG_MODULES
1503 if (type == NULL) {
1504 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1505 request_module("nft-expr-%u-%.*s", family,
1506 nla_len(nla), (char *)nla_data(nla));
1507 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1508 if (__nft_expr_type_get(family, nla))
1509 return ERR_PTR(-EAGAIN);
1510
1511 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1512 request_module("nft-expr-%.*s",
1513 nla_len(nla), (char *)nla_data(nla));
1514 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1515 if (__nft_expr_type_get(family, nla))
1516 return ERR_PTR(-EAGAIN);
1517 }
1518#endif
1519 return ERR_PTR(-ENOENT);
1520}
1521
1522static const struct nla_policy nft_expr_policy[NFTA_EXPR_MAX + 1] = {
1523 [NFTA_EXPR_NAME] = { .type = NLA_STRING },
1524 [NFTA_EXPR_DATA] = { .type = NLA_NESTED },
1525};
1526
1527static int nf_tables_fill_expr_info(struct sk_buff *skb,
1528 const struct nft_expr *expr)
1529{
1530 if (nla_put_string(skb, NFTA_EXPR_NAME, expr->ops->type->name))
1531 goto nla_put_failure;
1532
1533 if (expr->ops->dump) {
1534 struct nlattr *data = nla_nest_start(skb, NFTA_EXPR_DATA);
1535 if (data == NULL)
1536 goto nla_put_failure;
1537 if (expr->ops->dump(skb, expr) < 0)
1538 goto nla_put_failure;
1539 nla_nest_end(skb, data);
1540 }
1541
1542 return skb->len;
1543
1544nla_put_failure:
1545 return -1;
1546};
1547
1548int nft_expr_dump(struct sk_buff *skb, unsigned int attr,
1549 const struct nft_expr *expr)
1550{
1551 struct nlattr *nest;
1552
1553 nest = nla_nest_start(skb, attr);
1554 if (!nest)
1555 goto nla_put_failure;
1556 if (nf_tables_fill_expr_info(skb, expr) < 0)
1557 goto nla_put_failure;
1558 nla_nest_end(skb, nest);
1559 return 0;
1560
1561nla_put_failure:
1562 return -1;
1563}
1564
1565struct nft_expr_info {
1566 const struct nft_expr_ops *ops;
1567 struct nlattr *tb[NFT_EXPR_MAXATTR + 1];
1568};
1569
1570static int nf_tables_expr_parse(const struct nft_ctx *ctx,
1571 const struct nlattr *nla,
1572 struct nft_expr_info *info)
1573{
1574 const struct nft_expr_type *type;
1575 const struct nft_expr_ops *ops;
1576 struct nlattr *tb[NFTA_EXPR_MAX + 1];
1577 int err;
1578
1579 err = nla_parse_nested(tb, NFTA_EXPR_MAX, nla, nft_expr_policy);
1580 if (err < 0)
1581 return err;
1582
1583 type = nft_expr_type_get(ctx->afi->family, tb[NFTA_EXPR_NAME]);
1584 if (IS_ERR(type))
1585 return PTR_ERR(type);
1586
1587 if (tb[NFTA_EXPR_DATA]) {
1588 err = nla_parse_nested(info->tb, type->maxattr,
1589 tb[NFTA_EXPR_DATA], type->policy);
1590 if (err < 0)
1591 goto err1;
1592 } else
1593 memset(info->tb, 0, sizeof(info->tb[0]) * (type->maxattr + 1));
1594
1595 if (type->select_ops != NULL) {
1596 ops = type->select_ops(ctx,
1597 (const struct nlattr * const *)info->tb);
1598 if (IS_ERR(ops)) {
1599 err = PTR_ERR(ops);
1600 goto err1;
1601 }
1602 } else
1603 ops = type->ops;
1604
1605 info->ops = ops;
1606 return 0;
1607
1608err1:
1609 module_put(type->owner);
1610 return err;
1611}
1612
1613static int nf_tables_newexpr(const struct nft_ctx *ctx,
1614 const struct nft_expr_info *info,
1615 struct nft_expr *expr)
1616{
1617 const struct nft_expr_ops *ops = info->ops;
1618 int err;
1619
1620 expr->ops = ops;
1621 if (ops->init) {
1622 err = ops->init(ctx, expr, (const struct nlattr **)info->tb);
1623 if (err < 0)
1624 goto err1;
1625 }
1626
1627 return 0;
1628
1629err1:
1630 expr->ops = NULL;
1631 return err;
1632}
1633
1634static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
1635 struct nft_expr *expr)
1636{
1637 if (expr->ops->destroy)
1638 expr->ops->destroy(ctx, expr);
1639 module_put(expr->ops->type->owner);
1640}
1641
1642struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
1643 const struct nlattr *nla)
1644{
1645 struct nft_expr_info info;
1646 struct nft_expr *expr;
1647 int err;
1648
1649 err = nf_tables_expr_parse(ctx, nla, &info);
1650 if (err < 0)
1651 goto err1;
1652
1653 err = -ENOMEM;
1654 expr = kzalloc(info.ops->size, GFP_KERNEL);
1655 if (expr == NULL)
1656 goto err2;
1657
1658 err = nf_tables_newexpr(ctx, &info, expr);
1659 if (err < 0)
1660 goto err2;
1661
1662 return expr;
1663err2:
1664 module_put(info.ops->type->owner);
1665err1:
1666 return ERR_PTR(err);
1667}
1668
1669void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
1670{
1671 nf_tables_expr_destroy(ctx, expr);
1672 kfree(expr);
1673}
1674
1675/*
1676 * Rules
1677 */
1678
1679static struct nft_rule *__nf_tables_rule_lookup(const struct nft_chain *chain,
1680 u64 handle)
1681{
1682 struct nft_rule *rule;
1683
1684 // FIXME: this sucks
1685 list_for_each_entry(rule, &chain->rules, list) {
1686 if (handle == rule->handle)
1687 return rule;
1688 }
1689
1690 return ERR_PTR(-ENOENT);
1691}
1692
1693static struct nft_rule *nf_tables_rule_lookup(const struct nft_chain *chain,
1694 const struct nlattr *nla)
1695{
1696 if (nla == NULL)
1697 return ERR_PTR(-EINVAL);
1698
1699 return __nf_tables_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
1700}
1701
1702static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
1703 [NFTA_RULE_TABLE] = { .type = NLA_STRING },
1704 [NFTA_RULE_CHAIN] = { .type = NLA_STRING,
1705 .len = NFT_CHAIN_MAXNAMELEN - 1 },
1706 [NFTA_RULE_HANDLE] = { .type = NLA_U64 },
1707 [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
1708 [NFTA_RULE_COMPAT] = { .type = NLA_NESTED },
1709 [NFTA_RULE_POSITION] = { .type = NLA_U64 },
1710 [NFTA_RULE_USERDATA] = { .type = NLA_BINARY,
1711 .len = NFT_USERDATA_MAXLEN },
1712};
1713
1714static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
1715 u32 portid, u32 seq, int event,
1716 u32 flags, int family,
1717 const struct nft_table *table,
1718 const struct nft_chain *chain,
1719 const struct nft_rule *rule)
1720{
1721 struct nlmsghdr *nlh;
1722 struct nfgenmsg *nfmsg;
1723 const struct nft_expr *expr, *next;
1724 struct nlattr *list;
1725 const struct nft_rule *prule;
1726 int type = event | NFNL_SUBSYS_NFTABLES << 8;
1727
1728 nlh = nlmsg_put(skb, portid, seq, type, sizeof(struct nfgenmsg),
1729 flags);
1730 if (nlh == NULL)
1731 goto nla_put_failure;
1732
1733 nfmsg = nlmsg_data(nlh);
1734 nfmsg->nfgen_family = family;
1735 nfmsg->version = NFNETLINK_V0;
1736 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
1737
1738 if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
1739 goto nla_put_failure;
1740 if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
1741 goto nla_put_failure;
1742 if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle)))
1743 goto nla_put_failure;
1744
1745 if ((event != NFT_MSG_DELRULE) && (rule->list.prev != &chain->rules)) {
1746 prule = list_entry(rule->list.prev, struct nft_rule, list);
1747 if (nla_put_be64(skb, NFTA_RULE_POSITION,
1748 cpu_to_be64(prule->handle)))
1749 goto nla_put_failure;
1750 }
1751
1752 list = nla_nest_start(skb, NFTA_RULE_EXPRESSIONS);
1753 if (list == NULL)
1754 goto nla_put_failure;
1755 nft_rule_for_each_expr(expr, next, rule) {
1756 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
1757 goto nla_put_failure;
1758 }
1759 nla_nest_end(skb, list);
1760
1761 if (rule->udata) {
1762 struct nft_userdata *udata = nft_userdata(rule);
1763 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
1764 udata->data) < 0)
1765 goto nla_put_failure;
1766 }
1767
1768 nlmsg_end(skb, nlh);
1769 return 0;
1770
1771nla_put_failure:
1772 nlmsg_trim(skb, nlh);
1773 return -1;
1774}
1775
1776static int nf_tables_rule_notify(const struct nft_ctx *ctx,
1777 const struct nft_rule *rule,
1778 int event)
1779{
1780 struct sk_buff *skb;
1781 int err;
1782
1783 if (!ctx->report &&
1784 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1785 return 0;
1786
1787 err = -ENOBUFS;
1788 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1789 if (skb == NULL)
1790 goto err;
1791
1792 err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
1793 event, 0, ctx->afi->family, ctx->table,
1794 ctx->chain, rule);
1795 if (err < 0) {
1796 kfree_skb(skb);
1797 goto err;
1798 }
1799
1800 err = nfnetlink_send(skb, ctx->net, ctx->portid, NFNLGRP_NFTABLES,
1801 ctx->report, GFP_KERNEL);
1802err:
1803 if (err < 0) {
1804 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES,
1805 err);
1806 }
1807 return err;
1808}
1809
1810static int nf_tables_dump_rules(struct sk_buff *skb,
1811 struct netlink_callback *cb)
1812{
1813 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1814 const struct nft_af_info *afi;
1815 const struct nft_table *table;
1816 const struct nft_chain *chain;
1817 const struct nft_rule *rule;
1818 unsigned int idx = 0, s_idx = cb->args[0];
1819 struct net *net = sock_net(skb->sk);
1820 int family = nfmsg->nfgen_family;
1821
1822 rcu_read_lock();
1823 cb->seq = net->nft.base_seq;
1824
1825 list_for_each_entry_rcu(afi, &net->nft.af_info, list) {
1826 if (family != NFPROTO_UNSPEC && family != afi->family)
1827 continue;
1828
1829 list_for_each_entry_rcu(table, &afi->tables, list) {
1830 list_for_each_entry_rcu(chain, &table->chains, list) {
1831 list_for_each_entry_rcu(rule, &chain->rules, list) {
1832 if (!nft_rule_is_active(net, rule))
1833 goto cont;
1834 if (idx < s_idx)
1835 goto cont;
1836 if (idx > s_idx)
1837 memset(&cb->args[1], 0,
1838 sizeof(cb->args) - sizeof(cb->args[0]));
1839 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
1840 cb->nlh->nlmsg_seq,
1841 NFT_MSG_NEWRULE,
1842 NLM_F_MULTI | NLM_F_APPEND,
1843 afi->family, table, chain, rule) < 0)
1844 goto done;
1845
1846 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1847cont:
1848 idx++;
1849 }
1850 }
1851 }
1852 }
1853done:
1854 rcu_read_unlock();
1855
1856 cb->args[0] = idx;
1857 return skb->len;
1858}
1859
1860static int nf_tables_getrule(struct sock *nlsk, struct sk_buff *skb,
1861 const struct nlmsghdr *nlh,
1862 const struct nlattr * const nla[])
1863{
1864 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1865 const struct nft_af_info *afi;
1866 const struct nft_table *table;
1867 const struct nft_chain *chain;
1868 const struct nft_rule *rule;
1869 struct sk_buff *skb2;
1870 struct net *net = sock_net(skb->sk);
1871 int family = nfmsg->nfgen_family;
1872 int err;
1873
1874 if (nlh->nlmsg_flags & NLM_F_DUMP) {
1875 struct netlink_dump_control c = {
1876 .dump = nf_tables_dump_rules,
1877 };
1878 return netlink_dump_start(nlsk, skb, nlh, &c);
1879 }
1880
1881 afi = nf_tables_afinfo_lookup(net, family, false);
1882 if (IS_ERR(afi))
1883 return PTR_ERR(afi);
1884
1885 table = nf_tables_table_lookup(afi, nla[NFTA_RULE_TABLE]);
1886 if (IS_ERR(table))
1887 return PTR_ERR(table);
1888 if (table->flags & NFT_TABLE_INACTIVE)
1889 return -ENOENT;
1890
1891 chain = nf_tables_chain_lookup(table, nla[NFTA_RULE_CHAIN]);
1892 if (IS_ERR(chain))
1893 return PTR_ERR(chain);
1894 if (chain->flags & NFT_CHAIN_INACTIVE)
1895 return -ENOENT;
1896
1897 rule = nf_tables_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
1898 if (IS_ERR(rule))
1899 return PTR_ERR(rule);
1900
1901 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
1902 if (!skb2)
1903 return -ENOMEM;
1904
1905 err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
1906 nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
1907 family, table, chain, rule);
1908 if (err < 0)
1909 goto err;
1910
1911 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
1912
1913err:
1914 kfree_skb(skb2);
1915 return err;
1916}
1917
1918static void nf_tables_rule_destroy(const struct nft_ctx *ctx,
1919 struct nft_rule *rule)
1920{
1921 struct nft_expr *expr;
1922
1923 /*
1924 * Careful: some expressions might not be initialized in case this
1925 * is called on error from nf_tables_newrule().
1926 */
1927 expr = nft_expr_first(rule);
1928 while (expr->ops && expr != nft_expr_last(rule)) {
1929 nf_tables_expr_destroy(ctx, expr);
1930 expr = nft_expr_next(expr);
1931 }
1932 kfree(rule);
1933}
1934
1935#define NFT_RULE_MAXEXPRS 128
1936
1937static struct nft_expr_info *info;
1938
1939static int nf_tables_newrule(struct sock *nlsk, struct sk_buff *skb,
1940 const struct nlmsghdr *nlh,
1941 const struct nlattr * const nla[])
1942{
1943 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1944 struct nft_af_info *afi;
1945 struct net *net = sock_net(skb->sk);
1946 struct nft_table *table;
1947 struct nft_chain *chain;
1948 struct nft_rule *rule, *old_rule = NULL;
1949 struct nft_userdata *udata;
1950 struct nft_trans *trans = NULL;
1951 struct nft_expr *expr;
1952 struct nft_ctx ctx;
1953 struct nlattr *tmp;
1954 unsigned int size, i, n, ulen = 0, usize = 0;
1955 int err, rem;
1956 bool create;
1957 u64 handle, pos_handle;
1958
1959 create = nlh->nlmsg_flags & NLM_F_CREATE ? true : false;
1960
1961 afi = nf_tables_afinfo_lookup(net, nfmsg->nfgen_family, create);
1962 if (IS_ERR(afi))
1963 return PTR_ERR(afi);
1964
1965 table = nf_tables_table_lookup(afi, nla[NFTA_RULE_TABLE]);
1966 if (IS_ERR(table))
1967 return PTR_ERR(table);
1968
1969 chain = nf_tables_chain_lookup(table, nla[NFTA_RULE_CHAIN]);
1970 if (IS_ERR(chain))
1971 return PTR_ERR(chain);
1972
1973 if (nla[NFTA_RULE_HANDLE]) {
1974 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
1975 rule = __nf_tables_rule_lookup(chain, handle);
1976 if (IS_ERR(rule))
1977 return PTR_ERR(rule);
1978
1979 if (nlh->nlmsg_flags & NLM_F_EXCL)
1980 return -EEXIST;
1981 if (nlh->nlmsg_flags & NLM_F_REPLACE)
1982 old_rule = rule;
1983 else
1984 return -EOPNOTSUPP;
1985 } else {
1986 if (!create || nlh->nlmsg_flags & NLM_F_REPLACE)
1987 return -EINVAL;
1988 handle = nf_tables_alloc_handle(table);
1989
1990 if (chain->use == UINT_MAX)
1991 return -EOVERFLOW;
1992 }
1993
1994 if (nla[NFTA_RULE_POSITION]) {
1995 if (!(nlh->nlmsg_flags & NLM_F_CREATE))
1996 return -EOPNOTSUPP;
1997
1998 pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
1999 old_rule = __nf_tables_rule_lookup(chain, pos_handle);
2000 if (IS_ERR(old_rule))
2001 return PTR_ERR(old_rule);
2002 }
2003
2004 nft_ctx_init(&ctx, skb, nlh, afi, table, chain, nla);
2005
2006 n = 0;
2007 size = 0;
2008 if (nla[NFTA_RULE_EXPRESSIONS]) {
2009 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
2010 err = -EINVAL;
2011 if (nla_type(tmp) != NFTA_LIST_ELEM)
2012 goto err1;
2013 if (n == NFT_RULE_MAXEXPRS)
2014 goto err1;
2015 err = nf_tables_expr_parse(&ctx, tmp, &info[n]);
2016 if (err < 0)
2017 goto err1;
2018 size += info[n].ops->size;
2019 n++;
2020 }
2021 }
2022 /* Check for overflow of dlen field */
2023 err = -EFBIG;
2024 if (size >= 1 << 12)
2025 goto err1;
2026
2027 if (nla[NFTA_RULE_USERDATA]) {
2028 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
2029 if (ulen > 0)
2030 usize = sizeof(struct nft_userdata) + ulen;
2031 }
2032
2033 err = -ENOMEM;
2034 rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL);
2035 if (rule == NULL)
2036 goto err1;
2037
2038 nft_rule_activate_next(net, rule);
2039
2040 rule->handle = handle;
2041 rule->dlen = size;
2042 rule->udata = ulen ? 1 : 0;
2043
2044 if (ulen) {
2045 udata = nft_userdata(rule);
2046 udata->len = ulen - 1;
2047 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
2048 }
2049
2050 expr = nft_expr_first(rule);
2051 for (i = 0; i < n; i++) {
2052 err = nf_tables_newexpr(&ctx, &info[i], expr);
2053 if (err < 0)
2054 goto err2;
2055 info[i].ops = NULL;
2056 expr = nft_expr_next(expr);
2057 }
2058
2059 if (nlh->nlmsg_flags & NLM_F_REPLACE) {
2060 if (nft_rule_is_active_next(net, old_rule)) {
2061 trans = nft_trans_rule_add(&ctx, NFT_MSG_DELRULE,
2062 old_rule);
2063 if (trans == NULL) {
2064 err = -ENOMEM;
2065 goto err2;
2066 }
2067 nft_rule_deactivate_next(net, old_rule);
2068 chain->use--;
2069 list_add_tail_rcu(&rule->list, &old_rule->list);
2070 } else {
2071 err = -ENOENT;
2072 goto err2;
2073 }
2074 } else if (nlh->nlmsg_flags & NLM_F_APPEND)
2075 if (old_rule)
2076 list_add_rcu(&rule->list, &old_rule->list);
2077 else
2078 list_add_tail_rcu(&rule->list, &chain->rules);
2079 else {
2080 if (old_rule)
2081 list_add_tail_rcu(&rule->list, &old_rule->list);
2082 else
2083 list_add_rcu(&rule->list, &chain->rules);
2084 }
2085
2086 if (nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule) == NULL) {
2087 err = -ENOMEM;
2088 goto err3;
2089 }
2090 chain->use++;
2091 return 0;
2092
2093err3:
2094 list_del_rcu(&rule->list);
2095err2:
2096 nf_tables_rule_destroy(&ctx, rule);
2097err1:
2098 for (i = 0; i < n; i++) {
2099 if (info[i].ops != NULL)
2100 module_put(info[i].ops->type->owner);
2101 }
2102 return err;
2103}
2104
2105static int nf_tables_delrule(struct sock *nlsk, struct sk_buff *skb,
2106 const struct nlmsghdr *nlh,
2107 const struct nlattr * const nla[])
2108{
2109 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2110 struct nft_af_info *afi;
2111 struct net *net = sock_net(skb->sk);
2112 struct nft_table *table;
2113 struct nft_chain *chain = NULL;
2114 struct nft_rule *rule;
2115 int family = nfmsg->nfgen_family, err = 0;
2116 struct nft_ctx ctx;
2117
2118 afi = nf_tables_afinfo_lookup(net, family, false);
2119 if (IS_ERR(afi))
2120 return PTR_ERR(afi);
2121
2122 table = nf_tables_table_lookup(afi, nla[NFTA_RULE_TABLE]);
2123 if (IS_ERR(table))
2124 return PTR_ERR(table);
2125 if (table->flags & NFT_TABLE_INACTIVE)
2126 return -ENOENT;
2127
2128 if (nla[NFTA_RULE_CHAIN]) {
2129 chain = nf_tables_chain_lookup(table, nla[NFTA_RULE_CHAIN]);
2130 if (IS_ERR(chain))
2131 return PTR_ERR(chain);
2132 }
2133
2134 nft_ctx_init(&ctx, skb, nlh, afi, table, chain, nla);
2135
2136 if (chain) {
2137 if (nla[NFTA_RULE_HANDLE]) {
2138 rule = nf_tables_rule_lookup(chain,
2139 nla[NFTA_RULE_HANDLE]);
2140 if (IS_ERR(rule))
2141 return PTR_ERR(rule);
2142
2143 err = nft_delrule(&ctx, rule);
2144 } else {
2145 err = nft_delrule_by_chain(&ctx);
2146 }
2147 } else {
2148 list_for_each_entry(chain, &table->chains, list) {
2149 ctx.chain = chain;
2150 err = nft_delrule_by_chain(&ctx);
2151 if (err < 0)
2152 break;
2153 }
2154 }
2155
2156 return err;
2157}
2158
2159/*
2160 * Sets
2161 */
2162
2163static LIST_HEAD(nf_tables_set_ops);
2164
2165int nft_register_set(struct nft_set_ops *ops)
2166{
2167 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2168 list_add_tail_rcu(&ops->list, &nf_tables_set_ops);
2169 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2170 return 0;
2171}
2172EXPORT_SYMBOL_GPL(nft_register_set);
2173
2174void nft_unregister_set(struct nft_set_ops *ops)
2175{
2176 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2177 list_del_rcu(&ops->list);
2178 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2179}
2180EXPORT_SYMBOL_GPL(nft_unregister_set);
2181
2182/*
2183 * Select a set implementation based on the data characteristics and the
2184 * given policy. The total memory use might not be known if no size is
2185 * given, in that case the amount of memory per element is used.
2186 */
2187static const struct nft_set_ops *
2188nft_select_set_ops(const struct nlattr * const nla[],
2189 const struct nft_set_desc *desc,
2190 enum nft_set_policies policy)
2191{
2192 const struct nft_set_ops *ops, *bops;
2193 struct nft_set_estimate est, best;
2194 u32 features;
2195
2196#ifdef CONFIG_MODULES
2197 if (list_empty(&nf_tables_set_ops)) {
2198 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2199 request_module("nft-set");
2200 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2201 if (!list_empty(&nf_tables_set_ops))
2202 return ERR_PTR(-EAGAIN);
2203 }
2204#endif
2205 features = 0;
2206 if (nla[NFTA_SET_FLAGS] != NULL) {
2207 features = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
2208 features &= NFT_SET_INTERVAL | NFT_SET_MAP | NFT_SET_TIMEOUT;
2209 }
2210
2211 bops = NULL;
2212 best.size = ~0;
2213 best.class = ~0;
2214
2215 list_for_each_entry(ops, &nf_tables_set_ops, list) {
2216 if ((ops->features & features) != features)
2217 continue;
2218 if (!ops->estimate(desc, features, &est))
2219 continue;
2220
2221 switch (policy) {
2222 case NFT_SET_POL_PERFORMANCE:
2223 if (est.class < best.class)
2224 break;
2225 if (est.class == best.class && est.size < best.size)
2226 break;
2227 continue;
2228 case NFT_SET_POL_MEMORY:
2229 if (est.size < best.size)
2230 break;
2231 if (est.size == best.size && est.class < best.class)
2232 break;
2233 continue;
2234 default:
2235 break;
2236 }
2237
2238 if (!try_module_get(ops->owner))
2239 continue;
2240 if (bops != NULL)
2241 module_put(bops->owner);
2242
2243 bops = ops;
2244 best = est;
2245 }
2246
2247 if (bops != NULL)
2248 return bops;
2249
2250 return ERR_PTR(-EOPNOTSUPP);
2251}
2252
2253static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
2254 [NFTA_SET_TABLE] = { .type = NLA_STRING },
2255 [NFTA_SET_NAME] = { .type = NLA_STRING,
2256 .len = IFNAMSIZ - 1 },
2257 [NFTA_SET_FLAGS] = { .type = NLA_U32 },
2258 [NFTA_SET_KEY_TYPE] = { .type = NLA_U32 },
2259 [NFTA_SET_KEY_LEN] = { .type = NLA_U32 },
2260 [NFTA_SET_DATA_TYPE] = { .type = NLA_U32 },
2261 [NFTA_SET_DATA_LEN] = { .type = NLA_U32 },
2262 [NFTA_SET_POLICY] = { .type = NLA_U32 },
2263 [NFTA_SET_DESC] = { .type = NLA_NESTED },
2264 [NFTA_SET_ID] = { .type = NLA_U32 },
2265 [NFTA_SET_TIMEOUT] = { .type = NLA_U64 },
2266 [NFTA_SET_GC_INTERVAL] = { .type = NLA_U32 },
2267};
2268
2269static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
2270 [NFTA_SET_DESC_SIZE] = { .type = NLA_U32 },
2271};
2272
2273static int nft_ctx_init_from_setattr(struct nft_ctx *ctx,
2274 const struct sk_buff *skb,
2275 const struct nlmsghdr *nlh,
2276 const struct nlattr * const nla[])
2277{
2278 struct net *net = sock_net(skb->sk);
2279 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2280 struct nft_af_info *afi = NULL;
2281 struct nft_table *table = NULL;
2282
2283 if (nfmsg->nfgen_family != NFPROTO_UNSPEC) {
2284 afi = nf_tables_afinfo_lookup(net, nfmsg->nfgen_family, false);
2285 if (IS_ERR(afi))
2286 return PTR_ERR(afi);
2287 }
2288
2289 if (nla[NFTA_SET_TABLE] != NULL) {
2290 if (afi == NULL)
2291 return -EAFNOSUPPORT;
2292
2293 table = nf_tables_table_lookup(afi, nla[NFTA_SET_TABLE]);
2294 if (IS_ERR(table))
2295 return PTR_ERR(table);
2296 if (table->flags & NFT_TABLE_INACTIVE)
2297 return -ENOENT;
2298 }
2299
2300 nft_ctx_init(ctx, skb, nlh, afi, table, NULL, nla);
2301 return 0;
2302}
2303
2304struct nft_set *nf_tables_set_lookup(const struct nft_table *table,
2305 const struct nlattr *nla)
2306{
2307 struct nft_set *set;
2308
2309 if (nla == NULL)
2310 return ERR_PTR(-EINVAL);
2311
2312 list_for_each_entry(set, &table->sets, list) {
2313 if (!nla_strcmp(nla, set->name))
2314 return set;
2315 }
2316 return ERR_PTR(-ENOENT);
2317}
2318
2319struct nft_set *nf_tables_set_lookup_byid(const struct net *net,
2320 const struct nlattr *nla)
2321{
2322 struct nft_trans *trans;
2323 u32 id = ntohl(nla_get_be32(nla));
2324
2325 list_for_each_entry(trans, &net->nft.commit_list, list) {
2326 if (trans->msg_type == NFT_MSG_NEWSET &&
2327 id == nft_trans_set_id(trans))
2328 return nft_trans_set(trans);
2329 }
2330 return ERR_PTR(-ENOENT);
2331}
2332
2333static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
2334 const char *name)
2335{
2336 const struct nft_set *i;
2337 const char *p;
2338 unsigned long *inuse;
2339 unsigned int n = 0, min = 0;
2340
2341 p = strnchr(name, IFNAMSIZ, '%');
2342 if (p != NULL) {
2343 if (p[1] != 'd' || strchr(p + 2, '%'))
2344 return -EINVAL;
2345
2346 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
2347 if (inuse == NULL)
2348 return -ENOMEM;
2349cont:
2350 list_for_each_entry(i, &ctx->table->sets, list) {
2351 int tmp;
2352
2353 if (!sscanf(i->name, name, &tmp))
2354 continue;
2355 if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
2356 continue;
2357
2358 set_bit(tmp - min, inuse);
2359 }
2360
2361 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
2362 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
2363 min += BITS_PER_BYTE * PAGE_SIZE;
2364 memset(inuse, 0, PAGE_SIZE);
2365 goto cont;
2366 }
2367 free_page((unsigned long)inuse);
2368 }
2369
2370 snprintf(set->name, sizeof(set->name), name, min + n);
2371 list_for_each_entry(i, &ctx->table->sets, list) {
2372 if (!strcmp(set->name, i->name))
2373 return -ENFILE;
2374 }
2375 return 0;
2376}
2377
2378static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
2379 const struct nft_set *set, u16 event, u16 flags)
2380{
2381 struct nfgenmsg *nfmsg;
2382 struct nlmsghdr *nlh;
2383 struct nlattr *desc;
2384 u32 portid = ctx->portid;
2385 u32 seq = ctx->seq;
2386
2387 event |= NFNL_SUBSYS_NFTABLES << 8;
2388 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
2389 flags);
2390 if (nlh == NULL)
2391 goto nla_put_failure;
2392
2393 nfmsg = nlmsg_data(nlh);
2394 nfmsg->nfgen_family = ctx->afi->family;
2395 nfmsg->version = NFNETLINK_V0;
2396 nfmsg->res_id = htons(ctx->net->nft.base_seq & 0xffff);
2397
2398 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
2399 goto nla_put_failure;
2400 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
2401 goto nla_put_failure;
2402 if (set->flags != 0)
2403 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
2404 goto nla_put_failure;
2405
2406 if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
2407 goto nla_put_failure;
2408 if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
2409 goto nla_put_failure;
2410 if (set->flags & NFT_SET_MAP) {
2411 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
2412 goto nla_put_failure;
2413 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
2414 goto nla_put_failure;
2415 }
2416
2417 if (set->timeout &&
2418 nla_put_be64(skb, NFTA_SET_TIMEOUT, cpu_to_be64(set->timeout)))
2419 goto nla_put_failure;
2420 if (set->gc_int &&
2421 nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(set->gc_int)))
2422 goto nla_put_failure;
2423
2424 if (set->policy != NFT_SET_POL_PERFORMANCE) {
2425 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
2426 goto nla_put_failure;
2427 }
2428
2429 desc = nla_nest_start(skb, NFTA_SET_DESC);
2430 if (desc == NULL)
2431 goto nla_put_failure;
2432 if (set->size &&
2433 nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
2434 goto nla_put_failure;
2435 nla_nest_end(skb, desc);
2436
2437 nlmsg_end(skb, nlh);
2438 return 0;
2439
2440nla_put_failure:
2441 nlmsg_trim(skb, nlh);
2442 return -1;
2443}
2444
2445static int nf_tables_set_notify(const struct nft_ctx *ctx,
2446 const struct nft_set *set,
2447 int event, gfp_t gfp_flags)
2448{
2449 struct sk_buff *skb;
2450 u32 portid = ctx->portid;
2451 int err;
2452
2453 if (!ctx->report &&
2454 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
2455 return 0;
2456
2457 err = -ENOBUFS;
2458 skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
2459 if (skb == NULL)
2460 goto err;
2461
2462 err = nf_tables_fill_set(skb, ctx, set, event, 0);
2463 if (err < 0) {
2464 kfree_skb(skb);
2465 goto err;
2466 }
2467
2468 err = nfnetlink_send(skb, ctx->net, portid, NFNLGRP_NFTABLES,
2469 ctx->report, gfp_flags);
2470err:
2471 if (err < 0)
2472 nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, err);
2473 return err;
2474}
2475
2476static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
2477{
2478 const struct nft_set *set;
2479 unsigned int idx, s_idx = cb->args[0];
2480 struct nft_af_info *afi;
2481 struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
2482 struct net *net = sock_net(skb->sk);
2483 int cur_family = cb->args[3];
2484 struct nft_ctx *ctx = cb->data, ctx_set;
2485
2486 if (cb->args[1])
2487 return skb->len;
2488
2489 rcu_read_lock();
2490 cb->seq = net->nft.base_seq;
2491
2492 list_for_each_entry_rcu(afi, &net->nft.af_info, list) {
2493 if (ctx->afi && ctx->afi != afi)
2494 continue;
2495
2496 if (cur_family) {
2497 if (afi->family != cur_family)
2498 continue;
2499
2500 cur_family = 0;
2501 }
2502 list_for_each_entry_rcu(table, &afi->tables, list) {
2503 if (ctx->table && ctx->table != table)
2504 continue;
2505
2506 if (cur_table) {
2507 if (cur_table != table)
2508 continue;
2509
2510 cur_table = NULL;
2511 }
2512 idx = 0;
2513 list_for_each_entry_rcu(set, &table->sets, list) {
2514 if (idx < s_idx)
2515 goto cont;
2516
2517 ctx_set = *ctx;
2518 ctx_set.table = table;
2519 ctx_set.afi = afi;
2520 if (nf_tables_fill_set(skb, &ctx_set, set,
2521 NFT_MSG_NEWSET,
2522 NLM_F_MULTI) < 0) {
2523 cb->args[0] = idx;
2524 cb->args[2] = (unsigned long) table;
2525 cb->args[3] = afi->family;
2526 goto done;
2527 }
2528 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
2529cont:
2530 idx++;
2531 }
2532 if (s_idx)
2533 s_idx = 0;
2534 }
2535 }
2536 cb->args[1] = 1;
2537done:
2538 rcu_read_unlock();
2539 return skb->len;
2540}
2541
2542static int nf_tables_dump_sets_done(struct netlink_callback *cb)
2543{
2544 kfree(cb->data);
2545 return 0;
2546}
2547
2548static int nf_tables_getset(struct sock *nlsk, struct sk_buff *skb,
2549 const struct nlmsghdr *nlh,
2550 const struct nlattr * const nla[])
2551{
2552 const struct nft_set *set;
2553 struct nft_ctx ctx;
2554 struct sk_buff *skb2;
2555 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2556 int err;
2557
2558 /* Verify existence before starting dump */
2559 err = nft_ctx_init_from_setattr(&ctx, skb, nlh, nla);
2560 if (err < 0)
2561 return err;
2562
2563 if (nlh->nlmsg_flags & NLM_F_DUMP) {
2564 struct netlink_dump_control c = {
2565 .dump = nf_tables_dump_sets,
2566 .done = nf_tables_dump_sets_done,
2567 };
2568 struct nft_ctx *ctx_dump;
2569
2570 ctx_dump = kmalloc(sizeof(*ctx_dump), GFP_KERNEL);
2571 if (ctx_dump == NULL)
2572 return -ENOMEM;
2573
2574 *ctx_dump = ctx;
2575 c.data = ctx_dump;
2576
2577 return netlink_dump_start(nlsk, skb, nlh, &c);
2578 }
2579
2580 /* Only accept unspec with dump */
2581 if (nfmsg->nfgen_family == NFPROTO_UNSPEC)
2582 return -EAFNOSUPPORT;
2583
2584 set = nf_tables_set_lookup(ctx.table, nla[NFTA_SET_NAME]);
2585 if (IS_ERR(set))
2586 return PTR_ERR(set);
2587 if (set->flags & NFT_SET_INACTIVE)
2588 return -ENOENT;
2589
2590 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
2591 if (skb2 == NULL)
2592 return -ENOMEM;
2593
2594 err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
2595 if (err < 0)
2596 goto err;
2597
2598 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
2599
2600err:
2601 kfree_skb(skb2);
2602 return err;
2603}
2604
2605static int nf_tables_set_desc_parse(const struct nft_ctx *ctx,
2606 struct nft_set_desc *desc,
2607 const struct nlattr *nla)
2608{
2609 struct nlattr *da[NFTA_SET_DESC_MAX + 1];
2610 int err;
2611
2612 err = nla_parse_nested(da, NFTA_SET_DESC_MAX, nla, nft_set_desc_policy);
2613 if (err < 0)
2614 return err;
2615
2616 if (da[NFTA_SET_DESC_SIZE] != NULL)
2617 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
2618
2619 return 0;
2620}
2621
2622static int nf_tables_newset(struct sock *nlsk, struct sk_buff *skb,
2623 const struct nlmsghdr *nlh,
2624 const struct nlattr * const nla[])
2625{
2626 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2627 const struct nft_set_ops *ops;
2628 struct nft_af_info *afi;
2629 struct net *net = sock_net(skb->sk);
2630 struct nft_table *table;
2631 struct nft_set *set;
2632 struct nft_ctx ctx;
2633 char name[IFNAMSIZ];
2634 unsigned int size;
2635 bool create;
2636 u64 timeout;
2637 u32 ktype, dtype, flags, policy, gc_int;
2638 struct nft_set_desc desc;
2639 int err;
2640
2641 if (nla[NFTA_SET_TABLE] == NULL ||
2642 nla[NFTA_SET_NAME] == NULL ||
2643 nla[NFTA_SET_KEY_LEN] == NULL ||
2644 nla[NFTA_SET_ID] == NULL)
2645 return -EINVAL;
2646
2647 memset(&desc, 0, sizeof(desc));
2648
2649 ktype = NFT_DATA_VALUE;
2650 if (nla[NFTA_SET_KEY_TYPE] != NULL) {
2651 ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
2652 if ((ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
2653 return -EINVAL;
2654 }
2655
2656 desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
2657 if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
2658 return -EINVAL;
2659
2660 flags = 0;
2661 if (nla[NFTA_SET_FLAGS] != NULL) {
2662 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
2663 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
2664 NFT_SET_INTERVAL | NFT_SET_MAP |
2665 NFT_SET_TIMEOUT))
2666 return -EINVAL;
2667 }
2668
2669 dtype = 0;
2670 if (nla[NFTA_SET_DATA_TYPE] != NULL) {
2671 if (!(flags & NFT_SET_MAP))
2672 return -EINVAL;
2673
2674 dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
2675 if ((dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
2676 dtype != NFT_DATA_VERDICT)
2677 return -EINVAL;
2678
2679 if (dtype != NFT_DATA_VERDICT) {
2680 if (nla[NFTA_SET_DATA_LEN] == NULL)
2681 return -EINVAL;
2682 desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
2683 if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
2684 return -EINVAL;
2685 } else
2686 desc.dlen = sizeof(struct nft_verdict);
2687 } else if (flags & NFT_SET_MAP)
2688 return -EINVAL;
2689
2690 timeout = 0;
2691 if (nla[NFTA_SET_TIMEOUT] != NULL) {
2692 if (!(flags & NFT_SET_TIMEOUT))
2693 return -EINVAL;
2694 timeout = be64_to_cpu(nla_get_be64(nla[NFTA_SET_TIMEOUT]));
2695 }
2696 gc_int = 0;
2697 if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
2698 if (!(flags & NFT_SET_TIMEOUT))
2699 return -EINVAL;
2700 gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
2701 }
2702
2703 policy = NFT_SET_POL_PERFORMANCE;
2704 if (nla[NFTA_SET_POLICY] != NULL)
2705 policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
2706
2707 if (nla[NFTA_SET_DESC] != NULL) {
2708 err = nf_tables_set_desc_parse(&ctx, &desc, nla[NFTA_SET_DESC]);
2709 if (err < 0)
2710 return err;
2711 }
2712
2713 create = nlh->nlmsg_flags & NLM_F_CREATE ? true : false;
2714
2715 afi = nf_tables_afinfo_lookup(net, nfmsg->nfgen_family, create);
2716 if (IS_ERR(afi))
2717 return PTR_ERR(afi);
2718
2719 table = nf_tables_table_lookup(afi, nla[NFTA_SET_TABLE]);
2720 if (IS_ERR(table))
2721 return PTR_ERR(table);
2722
2723 nft_ctx_init(&ctx, skb, nlh, afi, table, NULL, nla);
2724
2725 set = nf_tables_set_lookup(table, nla[NFTA_SET_NAME]);
2726 if (IS_ERR(set)) {
2727 if (PTR_ERR(set) != -ENOENT)
2728 return PTR_ERR(set);
2729 set = NULL;
2730 }
2731
2732 if (set != NULL) {
2733 if (nlh->nlmsg_flags & NLM_F_EXCL)
2734 return -EEXIST;
2735 if (nlh->nlmsg_flags & NLM_F_REPLACE)
2736 return -EOPNOTSUPP;
2737 return 0;
2738 }
2739
2740 if (!(nlh->nlmsg_flags & NLM_F_CREATE))
2741 return -ENOENT;
2742
2743 ops = nft_select_set_ops(nla, &desc, policy);
2744 if (IS_ERR(ops))
2745 return PTR_ERR(ops);
2746
2747 size = 0;
2748 if (ops->privsize != NULL)
2749 size = ops->privsize(nla);
2750
2751 err = -ENOMEM;
2752 set = kzalloc(sizeof(*set) + size, GFP_KERNEL);
2753 if (set == NULL)
2754 goto err1;
2755
2756 nla_strlcpy(name, nla[NFTA_SET_NAME], sizeof(set->name));
2757 err = nf_tables_set_alloc_name(&ctx, set, name);
2758 if (err < 0)
2759 goto err2;
2760
2761 INIT_LIST_HEAD(&set->bindings);
2762 write_pnet(&set->pnet, net);
2763 set->ops = ops;
2764 set->ktype = ktype;
2765 set->klen = desc.klen;
2766 set->dtype = dtype;
2767 set->dlen = desc.dlen;
2768 set->flags = flags;
2769 set->size = desc.size;
2770 set->policy = policy;
2771 set->timeout = timeout;
2772 set->gc_int = gc_int;
2773
2774 err = ops->init(set, &desc, nla);
2775 if (err < 0)
2776 goto err2;
2777
2778 err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
2779 if (err < 0)
2780 goto err2;
2781
2782 list_add_tail_rcu(&set->list, &table->sets);
2783 table->use++;
2784 return 0;
2785
2786err2:
2787 kfree(set);
2788err1:
2789 module_put(ops->owner);
2790 return err;
2791}
2792
2793static void nft_set_destroy(struct nft_set *set)
2794{
2795 set->ops->destroy(set);
2796 module_put(set->ops->owner);
2797 kfree(set);
2798}
2799
2800static void nf_tables_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
2801{
2802 list_del_rcu(&set->list);
2803 nf_tables_set_notify(ctx, set, NFT_MSG_DELSET, GFP_ATOMIC);
2804 nft_set_destroy(set);
2805}
2806
2807static int nf_tables_delset(struct sock *nlsk, struct sk_buff *skb,
2808 const struct nlmsghdr *nlh,
2809 const struct nlattr * const nla[])
2810{
2811 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2812 struct nft_set *set;
2813 struct nft_ctx ctx;
2814 int err;
2815
2816 if (nfmsg->nfgen_family == NFPROTO_UNSPEC)
2817 return -EAFNOSUPPORT;
2818 if (nla[NFTA_SET_TABLE] == NULL)
2819 return -EINVAL;
2820
2821 err = nft_ctx_init_from_setattr(&ctx, skb, nlh, nla);
2822 if (err < 0)
2823 return err;
2824
2825 set = nf_tables_set_lookup(ctx.table, nla[NFTA_SET_NAME]);
2826 if (IS_ERR(set))
2827 return PTR_ERR(set);
2828 if (set->flags & NFT_SET_INACTIVE)
2829 return -ENOENT;
2830 if (!list_empty(&set->bindings))
2831 return -EBUSY;
2832
2833 return nft_delset(&ctx, set);
2834}
2835
2836static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
2837 const struct nft_set *set,
2838 const struct nft_set_iter *iter,
2839 const struct nft_set_elem *elem)
2840{
2841 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
2842 enum nft_registers dreg;
2843
2844 dreg = nft_type_to_reg(set->dtype);
2845 return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
2846 set->dtype == NFT_DATA_VERDICT ?
2847 NFT_DATA_VERDICT : NFT_DATA_VALUE,
2848 set->dlen);
2849}
2850
2851int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
2852 struct nft_set_binding *binding)
2853{
2854 struct nft_set_binding *i;
2855 struct nft_set_iter iter;
2856
2857 if (!list_empty(&set->bindings) && set->flags & NFT_SET_ANONYMOUS)
2858 return -EBUSY;
2859
2860 if (binding->flags & NFT_SET_MAP) {
2861 /* If the set is already bound to the same chain all
2862 * jumps are already validated for that chain.
2863 */
2864 list_for_each_entry(i, &set->bindings, list) {
2865 if (binding->flags & NFT_SET_MAP &&
2866 i->chain == binding->chain)
2867 goto bind;
2868 }
2869
2870 iter.skip = 0;
2871 iter.count = 0;
2872 iter.err = 0;
2873 iter.fn = nf_tables_bind_check_setelem;
2874
2875 set->ops->walk(ctx, set, &iter);
2876 if (iter.err < 0) {
2877 /* Destroy anonymous sets if binding fails */
2878 if (set->flags & NFT_SET_ANONYMOUS)
2879 nf_tables_set_destroy(ctx, set);
2880
2881 return iter.err;
2882 }
2883 }
2884bind:
2885 binding->chain = ctx->chain;
2886 list_add_tail_rcu(&binding->list, &set->bindings);
2887 return 0;
2888}
2889
2890void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
2891 struct nft_set_binding *binding)
2892{
2893 list_del_rcu(&binding->list);
2894
2895 if (list_empty(&set->bindings) && set->flags & NFT_SET_ANONYMOUS &&
2896 !(set->flags & NFT_SET_INACTIVE))
2897 nf_tables_set_destroy(ctx, set);
2898}
2899
2900const struct nft_set_ext_type nft_set_ext_types[] = {
2901 [NFT_SET_EXT_KEY] = {
2902 .align = __alignof__(u32),
2903 },
2904 [NFT_SET_EXT_DATA] = {
2905 .align = __alignof__(u32),
2906 },
2907 [NFT_SET_EXT_EXPR] = {
2908 .align = __alignof__(struct nft_expr),
2909 },
2910 [NFT_SET_EXT_FLAGS] = {
2911 .len = sizeof(u8),
2912 .align = __alignof__(u8),
2913 },
2914 [NFT_SET_EXT_TIMEOUT] = {
2915 .len = sizeof(u64),
2916 .align = __alignof__(u64),
2917 },
2918 [NFT_SET_EXT_EXPIRATION] = {
2919 .len = sizeof(unsigned long),
2920 .align = __alignof__(unsigned long),
2921 },
2922 [NFT_SET_EXT_USERDATA] = {
2923 .len = sizeof(struct nft_userdata),
2924 .align = __alignof__(struct nft_userdata),
2925 },
2926};
2927EXPORT_SYMBOL_GPL(nft_set_ext_types);
2928
2929/*
2930 * Set elements
2931 */
2932
2933static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
2934 [NFTA_SET_ELEM_KEY] = { .type = NLA_NESTED },
2935 [NFTA_SET_ELEM_DATA] = { .type = NLA_NESTED },
2936 [NFTA_SET_ELEM_FLAGS] = { .type = NLA_U32 },
2937 [NFTA_SET_ELEM_TIMEOUT] = { .type = NLA_U64 },
2938 [NFTA_SET_ELEM_USERDATA] = { .type = NLA_BINARY,
2939 .len = NFT_USERDATA_MAXLEN },
2940};
2941
2942static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
2943 [NFTA_SET_ELEM_LIST_TABLE] = { .type = NLA_STRING },
2944 [NFTA_SET_ELEM_LIST_SET] = { .type = NLA_STRING },
2945 [NFTA_SET_ELEM_LIST_ELEMENTS] = { .type = NLA_NESTED },
2946 [NFTA_SET_ELEM_LIST_SET_ID] = { .type = NLA_U32 },
2947};
2948
2949static int nft_ctx_init_from_elemattr(struct nft_ctx *ctx,
2950 const struct sk_buff *skb,
2951 const struct nlmsghdr *nlh,
2952 const struct nlattr * const nla[],
2953 bool trans)
2954{
2955 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2956 struct nft_af_info *afi;
2957 struct nft_table *table;
2958 struct net *net = sock_net(skb->sk);
2959
2960 afi = nf_tables_afinfo_lookup(net, nfmsg->nfgen_family, false);
2961 if (IS_ERR(afi))
2962 return PTR_ERR(afi);
2963
2964 table = nf_tables_table_lookup(afi, nla[NFTA_SET_ELEM_LIST_TABLE]);
2965 if (IS_ERR(table))
2966 return PTR_ERR(table);
2967 if (!trans && (table->flags & NFT_TABLE_INACTIVE))
2968 return -ENOENT;
2969
2970 nft_ctx_init(ctx, skb, nlh, afi, table, NULL, nla);
2971 return 0;
2972}
2973
2974static int nf_tables_fill_setelem(struct sk_buff *skb,
2975 const struct nft_set *set,
2976 const struct nft_set_elem *elem)
2977{
2978 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
2979 unsigned char *b = skb_tail_pointer(skb);
2980 struct nlattr *nest;
2981
2982 nest = nla_nest_start(skb, NFTA_LIST_ELEM);
2983 if (nest == NULL)
2984 goto nla_put_failure;
2985
2986 if (nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
2987 NFT_DATA_VALUE, set->klen) < 0)
2988 goto nla_put_failure;
2989
2990 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
2991 nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
2992 set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
2993 set->dlen) < 0)
2994 goto nla_put_failure;
2995
2996 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPR) &&
2997 nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, nft_set_ext_expr(ext)) < 0)
2998 goto nla_put_failure;
2999
3000 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
3001 nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
3002 htonl(*nft_set_ext_flags(ext))))
3003 goto nla_put_failure;
3004
3005 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
3006 nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
3007 cpu_to_be64(*nft_set_ext_timeout(ext))))
3008 goto nla_put_failure;
3009
3010 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
3011 unsigned long expires, now = jiffies;
3012
3013 expires = *nft_set_ext_expiration(ext);
3014 if (time_before(now, expires))
3015 expires -= now;
3016 else
3017 expires = 0;
3018
3019 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
3020 cpu_to_be64(jiffies_to_msecs(expires))))
3021 goto nla_put_failure;
3022 }
3023
3024 if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
3025 struct nft_userdata *udata;
3026
3027 udata = nft_set_ext_userdata(ext);
3028 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
3029 udata->len + 1, udata->data))
3030 goto nla_put_failure;
3031 }
3032
3033 nla_nest_end(skb, nest);
3034 return 0;
3035
3036nla_put_failure:
3037 nlmsg_trim(skb, b);
3038 return -EMSGSIZE;
3039}
3040
3041struct nft_set_dump_args {
3042 const struct netlink_callback *cb;
3043 struct nft_set_iter iter;
3044 struct sk_buff *skb;
3045};
3046
3047static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
3048 const struct nft_set *set,
3049 const struct nft_set_iter *iter,
3050 const struct nft_set_elem *elem)
3051{
3052 struct nft_set_dump_args *args;
3053
3054 args = container_of(iter, struct nft_set_dump_args, iter);
3055 return nf_tables_fill_setelem(args->skb, set, elem);
3056}
3057
3058static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
3059{
3060 const struct nft_set *set;
3061 struct nft_set_dump_args args;
3062 struct nft_ctx ctx;
3063 struct nlattr *nla[NFTA_SET_ELEM_LIST_MAX + 1];
3064 struct nfgenmsg *nfmsg;
3065 struct nlmsghdr *nlh;
3066 struct nlattr *nest;
3067 u32 portid, seq;
3068 int event, err;
3069
3070 err = nlmsg_parse(cb->nlh, sizeof(struct nfgenmsg), nla,
3071 NFTA_SET_ELEM_LIST_MAX, nft_set_elem_list_policy);
3072 if (err < 0)
3073 return err;
3074
3075 err = nft_ctx_init_from_elemattr(&ctx, cb->skb, cb->nlh, (void *)nla,
3076 false);
3077 if (err < 0)
3078 return err;
3079
3080 set = nf_tables_set_lookup(ctx.table, nla[NFTA_SET_ELEM_LIST_SET]);
3081 if (IS_ERR(set))
3082 return PTR_ERR(set);
3083 if (set->flags & NFT_SET_INACTIVE)
3084 return -ENOENT;
3085
3086 event = NFT_MSG_NEWSETELEM;
3087 event |= NFNL_SUBSYS_NFTABLES << 8;
3088 portid = NETLINK_CB(cb->skb).portid;
3089 seq = cb->nlh->nlmsg_seq;
3090
3091 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
3092 NLM_F_MULTI);
3093 if (nlh == NULL)
3094 goto nla_put_failure;
3095
3096 nfmsg = nlmsg_data(nlh);
3097 nfmsg->nfgen_family = ctx.afi->family;
3098 nfmsg->version = NFNETLINK_V0;
3099 nfmsg->res_id = htons(ctx.net->nft.base_seq & 0xffff);
3100
3101 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, ctx.table->name))
3102 goto nla_put_failure;
3103 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
3104 goto nla_put_failure;
3105
3106 nest = nla_nest_start(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
3107 if (nest == NULL)
3108 goto nla_put_failure;
3109
3110 args.cb = cb;
3111 args.skb = skb;
3112 args.iter.skip = cb->args[0];
3113 args.iter.count = 0;
3114 args.iter.err = 0;
3115 args.iter.fn = nf_tables_dump_setelem;
3116 set->ops->walk(&ctx, set, &args.iter);
3117
3118 nla_nest_end(skb, nest);
3119 nlmsg_end(skb, nlh);
3120
3121 if (args.iter.err && args.iter.err != -EMSGSIZE)
3122 return args.iter.err;
3123 if (args.iter.count == cb->args[0])
3124 return 0;
3125
3126 cb->args[0] = args.iter.count;
3127 return skb->len;
3128
3129nla_put_failure:
3130 return -ENOSPC;
3131}
3132
3133static int nf_tables_getsetelem(struct sock *nlsk, struct sk_buff *skb,
3134 const struct nlmsghdr *nlh,
3135 const struct nlattr * const nla[])
3136{
3137 const struct nft_set *set;
3138 struct nft_ctx ctx;
3139 int err;
3140
3141 err = nft_ctx_init_from_elemattr(&ctx, skb, nlh, nla, false);
3142 if (err < 0)
3143 return err;
3144
3145 set = nf_tables_set_lookup(ctx.table, nla[NFTA_SET_ELEM_LIST_SET]);
3146 if (IS_ERR(set))
3147 return PTR_ERR(set);
3148 if (set->flags & NFT_SET_INACTIVE)
3149 return -ENOENT;
3150
3151 if (nlh->nlmsg_flags & NLM_F_DUMP) {
3152 struct netlink_dump_control c = {
3153 .dump = nf_tables_dump_set,
3154 };
3155 return netlink_dump_start(nlsk, skb, nlh, &c);
3156 }
3157 return -EOPNOTSUPP;
3158}
3159
3160static int nf_tables_fill_setelem_info(struct sk_buff *skb,
3161 const struct nft_ctx *ctx, u32 seq,
3162 u32 portid, int event, u16 flags,
3163 const struct nft_set *set,
3164 const struct nft_set_elem *elem)
3165{
3166 struct nfgenmsg *nfmsg;
3167 struct nlmsghdr *nlh;
3168 struct nlattr *nest;
3169 int err;
3170
3171 event |= NFNL_SUBSYS_NFTABLES << 8;
3172 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
3173 flags);
3174 if (nlh == NULL)
3175 goto nla_put_failure;
3176
3177 nfmsg = nlmsg_data(nlh);
3178 nfmsg->nfgen_family = ctx->afi->family;
3179 nfmsg->version = NFNETLINK_V0;
3180 nfmsg->res_id = htons(ctx->net->nft.base_seq & 0xffff);
3181
3182 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
3183 goto nla_put_failure;
3184 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
3185 goto nla_put_failure;
3186
3187 nest = nla_nest_start(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
3188 if (nest == NULL)
3189 goto nla_put_failure;
3190
3191 err = nf_tables_fill_setelem(skb, set, elem);
3192 if (err < 0)
3193 goto nla_put_failure;
3194
3195 nla_nest_end(skb, nest);
3196
3197 nlmsg_end(skb, nlh);
3198 return 0;
3199
3200nla_put_failure:
3201 nlmsg_trim(skb, nlh);
3202 return -1;
3203}
3204
3205static int nf_tables_setelem_notify(const struct nft_ctx *ctx,
3206 const struct nft_set *set,
3207 const struct nft_set_elem *elem,
3208 int event, u16 flags)
3209{
3210 struct net *net = ctx->net;
3211 u32 portid = ctx->portid;
3212 struct sk_buff *skb;
3213 int err;
3214
3215 if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
3216 return 0;
3217
3218 err = -ENOBUFS;
3219 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
3220 if (skb == NULL)
3221 goto err;
3222
3223 err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
3224 set, elem);
3225 if (err < 0) {
3226 kfree_skb(skb);
3227 goto err;
3228 }
3229
3230 err = nfnetlink_send(skb, net, portid, NFNLGRP_NFTABLES, ctx->report,
3231 GFP_KERNEL);
3232err:
3233 if (err < 0)
3234 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, err);
3235 return err;
3236}
3237
3238static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
3239 int msg_type,
3240 struct nft_set *set)
3241{
3242 struct nft_trans *trans;
3243
3244 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
3245 if (trans == NULL)
3246 return NULL;
3247
3248 nft_trans_elem_set(trans) = set;
3249 return trans;
3250}
3251
3252void *nft_set_elem_init(const struct nft_set *set,
3253 const struct nft_set_ext_tmpl *tmpl,
3254 const u32 *key, const u32 *data,
3255 u64 timeout, gfp_t gfp)
3256{
3257 struct nft_set_ext *ext;
3258 void *elem;
3259
3260 elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
3261 if (elem == NULL)
3262 return NULL;
3263
3264 ext = nft_set_elem_ext(set, elem);
3265 nft_set_ext_init(ext, tmpl);
3266
3267 memcpy(nft_set_ext_key(ext), key, set->klen);
3268 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
3269 memcpy(nft_set_ext_data(ext), data, set->dlen);
3270 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION))
3271 *nft_set_ext_expiration(ext) =
3272 jiffies + msecs_to_jiffies(timeout);
3273 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
3274 *nft_set_ext_timeout(ext) = timeout;
3275
3276 return elem;
3277}
3278
3279void nft_set_elem_destroy(const struct nft_set *set, void *elem)
3280{
3281 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
3282
3283 nft_data_uninit(nft_set_ext_key(ext), NFT_DATA_VALUE);
3284 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
3285 nft_data_uninit(nft_set_ext_data(ext), set->dtype);
3286 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPR))
3287 nf_tables_expr_destroy(NULL, nft_set_ext_expr(ext));
3288
3289 kfree(elem);
3290}
3291EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
3292
3293static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
3294 const struct nlattr *attr)
3295{
3296 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
3297 struct nft_data_desc d1, d2;
3298 struct nft_set_ext_tmpl tmpl;
3299 struct nft_set_ext *ext;
3300 struct nft_set_elem elem;
3301 struct nft_set_binding *binding;
3302 struct nft_userdata *udata;
3303 struct nft_data data;
3304 enum nft_registers dreg;
3305 struct nft_trans *trans;
3306 u64 timeout;
3307 u32 flags;
3308 u8 ulen;
3309 int err;
3310
3311 err = nla_parse_nested(nla, NFTA_SET_ELEM_MAX, attr,
3312 nft_set_elem_policy);
3313 if (err < 0)
3314 return err;
3315
3316 if (nla[NFTA_SET_ELEM_KEY] == NULL)
3317 return -EINVAL;
3318
3319 nft_set_ext_prepare(&tmpl);
3320
3321 flags = 0;
3322 if (nla[NFTA_SET_ELEM_FLAGS] != NULL) {
3323 flags = ntohl(nla_get_be32(nla[NFTA_SET_ELEM_FLAGS]));
3324 if (flags & ~NFT_SET_ELEM_INTERVAL_END)
3325 return -EINVAL;
3326 if (!(set->flags & NFT_SET_INTERVAL) &&
3327 flags & NFT_SET_ELEM_INTERVAL_END)
3328 return -EINVAL;
3329 if (flags != 0)
3330 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
3331 }
3332
3333 if (set->flags & NFT_SET_MAP) {
3334 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
3335 !(flags & NFT_SET_ELEM_INTERVAL_END))
3336 return -EINVAL;
3337 if (nla[NFTA_SET_ELEM_DATA] != NULL &&
3338 flags & NFT_SET_ELEM_INTERVAL_END)
3339 return -EINVAL;
3340 } else {
3341 if (nla[NFTA_SET_ELEM_DATA] != NULL)
3342 return -EINVAL;
3343 }
3344
3345 timeout = 0;
3346 if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
3347 if (!(set->flags & NFT_SET_TIMEOUT))
3348 return -EINVAL;
3349 timeout = be64_to_cpu(nla_get_be64(nla[NFTA_SET_ELEM_TIMEOUT]));
3350 } else if (set->flags & NFT_SET_TIMEOUT) {
3351 timeout = set->timeout;
3352 }
3353
3354 err = nft_data_init(ctx, &elem.key.val, sizeof(elem.key), &d1,
3355 nla[NFTA_SET_ELEM_KEY]);
3356 if (err < 0)
3357 goto err1;
3358 err = -EINVAL;
3359 if (d1.type != NFT_DATA_VALUE || d1.len != set->klen)
3360 goto err2;
3361
3362 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, d1.len);
3363 if (timeout > 0) {
3364 nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
3365 if (timeout != set->timeout)
3366 nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
3367 }
3368
3369 if (nla[NFTA_SET_ELEM_DATA] != NULL) {
3370 err = nft_data_init(ctx, &data, sizeof(data), &d2,
3371 nla[NFTA_SET_ELEM_DATA]);
3372 if (err < 0)
3373 goto err2;
3374
3375 err = -EINVAL;
3376 if (set->dtype != NFT_DATA_VERDICT && d2.len != set->dlen)
3377 goto err3;
3378
3379 dreg = nft_type_to_reg(set->dtype);
3380 list_for_each_entry(binding, &set->bindings, list) {
3381 struct nft_ctx bind_ctx = {
3382 .afi = ctx->afi,
3383 .table = ctx->table,
3384 .chain = (struct nft_chain *)binding->chain,
3385 };
3386
3387 if (!(binding->flags & NFT_SET_MAP))
3388 continue;
3389
3390 err = nft_validate_register_store(&bind_ctx, dreg,
3391 &data,
3392 d2.type, d2.len);
3393 if (err < 0)
3394 goto err3;
3395 }
3396
3397 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, d2.len);
3398 }
3399
3400 /* The full maximum length of userdata can exceed the maximum
3401 * offset value (U8_MAX) for following extensions, therefor it
3402 * must be the last extension added.
3403 */
3404 ulen = 0;
3405 if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
3406 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
3407 if (ulen > 0)
3408 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
3409 ulen);
3410 }
3411
3412 err = -ENOMEM;
3413 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data, data.data,
3414 timeout, GFP_KERNEL);
3415 if (elem.priv == NULL)
3416 goto err3;
3417
3418 ext = nft_set_elem_ext(set, elem.priv);
3419 if (flags)
3420 *nft_set_ext_flags(ext) = flags;
3421 if (ulen > 0) {
3422 udata = nft_set_ext_userdata(ext);
3423 udata->len = ulen - 1;
3424 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
3425 }
3426
3427 trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
3428 if (trans == NULL)
3429 goto err4;
3430
3431 ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
3432 err = set->ops->insert(set, &elem);
3433 if (err < 0)
3434 goto err5;
3435
3436 nft_trans_elem(trans) = elem;
3437 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
3438 return 0;
3439
3440err5:
3441 kfree(trans);
3442err4:
3443 kfree(elem.priv);
3444err3:
3445 if (nla[NFTA_SET_ELEM_DATA] != NULL)
3446 nft_data_uninit(&data, d2.type);
3447err2:
3448 nft_data_uninit(&elem.key.val, d1.type);
3449err1:
3450 return err;
3451}
3452
3453static int nf_tables_newsetelem(struct sock *nlsk, struct sk_buff *skb,
3454 const struct nlmsghdr *nlh,
3455 const struct nlattr * const nla[])
3456{
3457 struct net *net = sock_net(skb->sk);
3458 const struct nlattr *attr;
3459 struct nft_set *set;
3460 struct nft_ctx ctx;
3461 int rem, err = 0;
3462
3463 if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
3464 return -EINVAL;
3465
3466 err = nft_ctx_init_from_elemattr(&ctx, skb, nlh, nla, true);
3467 if (err < 0)
3468 return err;
3469
3470 set = nf_tables_set_lookup(ctx.table, nla[NFTA_SET_ELEM_LIST_SET]);
3471 if (IS_ERR(set)) {
3472 if (nla[NFTA_SET_ELEM_LIST_SET_ID]) {
3473 set = nf_tables_set_lookup_byid(net,
3474 nla[NFTA_SET_ELEM_LIST_SET_ID]);
3475 }
3476 if (IS_ERR(set))
3477 return PTR_ERR(set);
3478 }
3479
3480 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
3481 return -EBUSY;
3482
3483 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
3484 if (set->size &&
3485 !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact))
3486 return -ENFILE;
3487
3488 err = nft_add_set_elem(&ctx, set, attr);
3489 if (err < 0) {
3490 atomic_dec(&set->nelems);
3491 break;
3492 }
3493 }
3494 return err;
3495}
3496
3497static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
3498 const struct nlattr *attr)
3499{
3500 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
3501 struct nft_data_desc desc;
3502 struct nft_set_elem elem;
3503 struct nft_trans *trans;
3504 int err;
3505
3506 err = nla_parse_nested(nla, NFTA_SET_ELEM_MAX, attr,
3507 nft_set_elem_policy);
3508 if (err < 0)
3509 goto err1;
3510
3511 err = -EINVAL;
3512 if (nla[NFTA_SET_ELEM_KEY] == NULL)
3513 goto err1;
3514
3515 err = nft_data_init(ctx, &elem.key.val, sizeof(elem.key), &desc,
3516 nla[NFTA_SET_ELEM_KEY]);
3517 if (err < 0)
3518 goto err1;
3519
3520 err = -EINVAL;
3521 if (desc.type != NFT_DATA_VALUE || desc.len != set->klen)
3522 goto err2;
3523
3524 trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
3525 if (trans == NULL) {
3526 err = -ENOMEM;
3527 goto err2;
3528 }
3529
3530 elem.priv = set->ops->deactivate(set, &elem);
3531 if (elem.priv == NULL) {
3532 err = -ENOENT;
3533 goto err3;
3534 }
3535
3536 nft_trans_elem(trans) = elem;
3537 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
3538 return 0;
3539
3540err3:
3541 kfree(trans);
3542err2:
3543 nft_data_uninit(&elem.key.val, desc.type);
3544err1:
3545 return err;
3546}
3547
3548static int nf_tables_delsetelem(struct sock *nlsk, struct sk_buff *skb,
3549 const struct nlmsghdr *nlh,
3550 const struct nlattr * const nla[])
3551{
3552 const struct nlattr *attr;
3553 struct nft_set *set;
3554 struct nft_ctx ctx;
3555 int rem, err = 0;
3556
3557 if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
3558 return -EINVAL;
3559
3560 err = nft_ctx_init_from_elemattr(&ctx, skb, nlh, nla, false);
3561 if (err < 0)
3562 return err;
3563
3564 set = nf_tables_set_lookup(ctx.table, nla[NFTA_SET_ELEM_LIST_SET]);
3565 if (IS_ERR(set))
3566 return PTR_ERR(set);
3567 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
3568 return -EBUSY;
3569
3570 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
3571 err = nft_del_setelem(&ctx, set, attr);
3572 if (err < 0)
3573 break;
3574
3575 set->ndeact++;
3576 }
3577 return err;
3578}
3579
3580void nft_set_gc_batch_release(struct rcu_head *rcu)
3581{
3582 struct nft_set_gc_batch *gcb;
3583 unsigned int i;
3584
3585 gcb = container_of(rcu, struct nft_set_gc_batch, head.rcu);
3586 for (i = 0; i < gcb->head.cnt; i++)
3587 nft_set_elem_destroy(gcb->head.set, gcb->elems[i]);
3588 kfree(gcb);
3589}
3590EXPORT_SYMBOL_GPL(nft_set_gc_batch_release);
3591
3592struct nft_set_gc_batch *nft_set_gc_batch_alloc(const struct nft_set *set,
3593 gfp_t gfp)
3594{
3595 struct nft_set_gc_batch *gcb;
3596
3597 gcb = kzalloc(sizeof(*gcb), gfp);
3598 if (gcb == NULL)
3599 return gcb;
3600 gcb->head.set = set;
3601 return gcb;
3602}
3603EXPORT_SYMBOL_GPL(nft_set_gc_batch_alloc);
3604
3605static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
3606 u32 portid, u32 seq)
3607{
3608 struct nlmsghdr *nlh;
3609 struct nfgenmsg *nfmsg;
3610 int event = (NFNL_SUBSYS_NFTABLES << 8) | NFT_MSG_NEWGEN;
3611
3612 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), 0);
3613 if (nlh == NULL)
3614 goto nla_put_failure;
3615
3616 nfmsg = nlmsg_data(nlh);
3617 nfmsg->nfgen_family = AF_UNSPEC;
3618 nfmsg->version = NFNETLINK_V0;
3619 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
3620
3621 if (nla_put_be32(skb, NFTA_GEN_ID, htonl(net->nft.base_seq)))
3622 goto nla_put_failure;
3623
3624 nlmsg_end(skb, nlh);
3625 return 0;
3626
3627nla_put_failure:
3628 nlmsg_trim(skb, nlh);
3629 return -EMSGSIZE;
3630}
3631
3632static int nf_tables_gen_notify(struct net *net, struct sk_buff *skb, int event)
3633{
3634 struct nlmsghdr *nlh = nlmsg_hdr(skb);
3635 struct sk_buff *skb2;
3636 int err;
3637
3638 if (nlmsg_report(nlh) &&
3639 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
3640 return 0;
3641
3642 err = -ENOBUFS;
3643 skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
3644 if (skb2 == NULL)
3645 goto err;
3646
3647 err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
3648 nlh->nlmsg_seq);
3649 if (err < 0) {
3650 kfree_skb(skb2);
3651 goto err;
3652 }
3653
3654 err = nfnetlink_send(skb2, net, NETLINK_CB(skb).portid,
3655 NFNLGRP_NFTABLES, nlmsg_report(nlh), GFP_KERNEL);
3656err:
3657 if (err < 0) {
3658 nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
3659 err);
3660 }
3661 return err;
3662}
3663
3664static int nf_tables_getgen(struct sock *nlsk, struct sk_buff *skb,
3665 const struct nlmsghdr *nlh,
3666 const struct nlattr * const nla[])
3667{
3668 struct net *net = sock_net(skb->sk);
3669 struct sk_buff *skb2;
3670 int err;
3671
3672 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_KERNEL);
3673 if (skb2 == NULL)
3674 return -ENOMEM;
3675
3676 err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
3677 nlh->nlmsg_seq);
3678 if (err < 0)
3679 goto err;
3680
3681 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
3682err:
3683 kfree_skb(skb2);
3684 return err;
3685}
3686
3687static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
3688 [NFT_MSG_NEWTABLE] = {
3689 .call_batch = nf_tables_newtable,
3690 .attr_count = NFTA_TABLE_MAX,
3691 .policy = nft_table_policy,
3692 },
3693 [NFT_MSG_GETTABLE] = {
3694 .call = nf_tables_gettable,
3695 .attr_count = NFTA_TABLE_MAX,
3696 .policy = nft_table_policy,
3697 },
3698 [NFT_MSG_DELTABLE] = {
3699 .call_batch = nf_tables_deltable,
3700 .attr_count = NFTA_TABLE_MAX,
3701 .policy = nft_table_policy,
3702 },
3703 [NFT_MSG_NEWCHAIN] = {
3704 .call_batch = nf_tables_newchain,
3705 .attr_count = NFTA_CHAIN_MAX,
3706 .policy = nft_chain_policy,
3707 },
3708 [NFT_MSG_GETCHAIN] = {
3709 .call = nf_tables_getchain,
3710 .attr_count = NFTA_CHAIN_MAX,
3711 .policy = nft_chain_policy,
3712 },
3713 [NFT_MSG_DELCHAIN] = {
3714 .call_batch = nf_tables_delchain,
3715 .attr_count = NFTA_CHAIN_MAX,
3716 .policy = nft_chain_policy,
3717 },
3718 [NFT_MSG_NEWRULE] = {
3719 .call_batch = nf_tables_newrule,
3720 .attr_count = NFTA_RULE_MAX,
3721 .policy = nft_rule_policy,
3722 },
3723 [NFT_MSG_GETRULE] = {
3724 .call = nf_tables_getrule,
3725 .attr_count = NFTA_RULE_MAX,
3726 .policy = nft_rule_policy,
3727 },
3728 [NFT_MSG_DELRULE] = {
3729 .call_batch = nf_tables_delrule,
3730 .attr_count = NFTA_RULE_MAX,
3731 .policy = nft_rule_policy,
3732 },
3733 [NFT_MSG_NEWSET] = {
3734 .call_batch = nf_tables_newset,
3735 .attr_count = NFTA_SET_MAX,
3736 .policy = nft_set_policy,
3737 },
3738 [NFT_MSG_GETSET] = {
3739 .call = nf_tables_getset,
3740 .attr_count = NFTA_SET_MAX,
3741 .policy = nft_set_policy,
3742 },
3743 [NFT_MSG_DELSET] = {
3744 .call_batch = nf_tables_delset,
3745 .attr_count = NFTA_SET_MAX,
3746 .policy = nft_set_policy,
3747 },
3748 [NFT_MSG_NEWSETELEM] = {
3749 .call_batch = nf_tables_newsetelem,
3750 .attr_count = NFTA_SET_ELEM_LIST_MAX,
3751 .policy = nft_set_elem_list_policy,
3752 },
3753 [NFT_MSG_GETSETELEM] = {
3754 .call = nf_tables_getsetelem,
3755 .attr_count = NFTA_SET_ELEM_LIST_MAX,
3756 .policy = nft_set_elem_list_policy,
3757 },
3758 [NFT_MSG_DELSETELEM] = {
3759 .call_batch = nf_tables_delsetelem,
3760 .attr_count = NFTA_SET_ELEM_LIST_MAX,
3761 .policy = nft_set_elem_list_policy,
3762 },
3763 [NFT_MSG_GETGEN] = {
3764 .call = nf_tables_getgen,
3765 },
3766};
3767
3768static void nft_chain_commit_update(struct nft_trans *trans)
3769{
3770 struct nft_base_chain *basechain;
3771
3772 if (nft_trans_chain_name(trans)[0])
3773 strcpy(trans->ctx.chain->name, nft_trans_chain_name(trans));
3774
3775 if (!(trans->ctx.chain->flags & NFT_BASE_CHAIN))
3776 return;
3777
3778 basechain = nft_base_chain(trans->ctx.chain);
3779 nft_chain_stats_replace(basechain, nft_trans_chain_stats(trans));
3780
3781 switch (nft_trans_chain_policy(trans)) {
3782 case NF_DROP:
3783 case NF_ACCEPT:
3784 basechain->policy = nft_trans_chain_policy(trans);
3785 break;
3786 }
3787}
3788
3789static void nf_tables_commit_release(struct nft_trans *trans)
3790{
3791 switch (trans->msg_type) {
3792 case NFT_MSG_DELTABLE:
3793 nf_tables_table_destroy(&trans->ctx);
3794 break;
3795 case NFT_MSG_DELCHAIN:
3796 nf_tables_chain_destroy(trans->ctx.chain);
3797 break;
3798 case NFT_MSG_DELRULE:
3799 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
3800 break;
3801 case NFT_MSG_DELSET:
3802 nft_set_destroy(nft_trans_set(trans));
3803 break;
3804 case NFT_MSG_DELSETELEM:
3805 nft_set_elem_destroy(nft_trans_elem_set(trans),
3806 nft_trans_elem(trans).priv);
3807 break;
3808 }
3809 kfree(trans);
3810}
3811
3812static int nf_tables_commit(struct sk_buff *skb)
3813{
3814 struct net *net = sock_net(skb->sk);
3815 struct nft_trans *trans, *next;
3816 struct nft_trans_elem *te;
3817
3818 /* Bump generation counter, invalidate any dump in progress */
3819 while (++net->nft.base_seq == 0);
3820
3821 /* A new generation has just started */
3822 net->nft.gencursor = nft_gencursor_next(net);
3823
3824 /* Make sure all packets have left the previous generation before
3825 * purging old rules.
3826 */
3827 synchronize_rcu();
3828
3829 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
3830 switch (trans->msg_type) {
3831 case NFT_MSG_NEWTABLE:
3832 if (nft_trans_table_update(trans)) {
3833 if (!nft_trans_table_enable(trans)) {
3834 nf_tables_table_disable(trans->ctx.afi,
3835 trans->ctx.table);
3836 trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
3837 }
3838 } else {
3839 trans->ctx.table->flags &= ~NFT_TABLE_INACTIVE;
3840 }
3841 nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
3842 nft_trans_destroy(trans);
3843 break;
3844 case NFT_MSG_DELTABLE:
3845 nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
3846 break;
3847 case NFT_MSG_NEWCHAIN:
3848 if (nft_trans_chain_update(trans))
3849 nft_chain_commit_update(trans);
3850 else
3851 trans->ctx.chain->flags &= ~NFT_CHAIN_INACTIVE;
3852
3853 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
3854 nft_trans_destroy(trans);
3855 break;
3856 case NFT_MSG_DELCHAIN:
3857 nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
3858 nf_tables_unregister_hooks(trans->ctx.table,
3859 trans->ctx.chain,
3860 trans->ctx.afi->nops);
3861 break;
3862 case NFT_MSG_NEWRULE:
3863 nft_rule_clear(trans->ctx.net, nft_trans_rule(trans));
3864 nf_tables_rule_notify(&trans->ctx,
3865 nft_trans_rule(trans),
3866 NFT_MSG_NEWRULE);
3867 nft_trans_destroy(trans);
3868 break;
3869 case NFT_MSG_DELRULE:
3870 list_del_rcu(&nft_trans_rule(trans)->list);
3871 nf_tables_rule_notify(&trans->ctx,
3872 nft_trans_rule(trans),
3873 NFT_MSG_DELRULE);
3874 break;
3875 case NFT_MSG_NEWSET:
3876 nft_trans_set(trans)->flags &= ~NFT_SET_INACTIVE;
3877 /* This avoids hitting -EBUSY when deleting the table
3878 * from the transaction.
3879 */
3880 if (nft_trans_set(trans)->flags & NFT_SET_ANONYMOUS &&
3881 !list_empty(&nft_trans_set(trans)->bindings))
3882 trans->ctx.table->use--;
3883
3884 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
3885 NFT_MSG_NEWSET, GFP_KERNEL);
3886 nft_trans_destroy(trans);
3887 break;
3888 case NFT_MSG_DELSET:
3889 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
3890 NFT_MSG_DELSET, GFP_KERNEL);
3891 break;
3892 case NFT_MSG_NEWSETELEM:
3893 te = (struct nft_trans_elem *)trans->data;
3894
3895 te->set->ops->activate(te->set, &te->elem);
3896 nf_tables_setelem_notify(&trans->ctx, te->set,
3897 &te->elem,
3898 NFT_MSG_NEWSETELEM, 0);
3899 nft_trans_destroy(trans);
3900 break;
3901 case NFT_MSG_DELSETELEM:
3902 te = (struct nft_trans_elem *)trans->data;
3903
3904 nf_tables_setelem_notify(&trans->ctx, te->set,
3905 &te->elem,
3906 NFT_MSG_DELSETELEM, 0);
3907 te->set->ops->remove(te->set, &te->elem);
3908 atomic_dec(&te->set->nelems);
3909 te->set->ndeact--;
3910 break;
3911 }
3912 }
3913
3914 synchronize_rcu();
3915
3916 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
3917 list_del(&trans->list);
3918 nf_tables_commit_release(trans);
3919 }
3920
3921 nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
3922
3923 return 0;
3924}
3925
3926static void nf_tables_abort_release(struct nft_trans *trans)
3927{
3928 switch (trans->msg_type) {
3929 case NFT_MSG_NEWTABLE:
3930 nf_tables_table_destroy(&trans->ctx);
3931 break;
3932 case NFT_MSG_NEWCHAIN:
3933 nf_tables_chain_destroy(trans->ctx.chain);
3934 break;
3935 case NFT_MSG_NEWRULE:
3936 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
3937 break;
3938 case NFT_MSG_NEWSET:
3939 nft_set_destroy(nft_trans_set(trans));
3940 break;
3941 case NFT_MSG_NEWSETELEM:
3942 nft_set_elem_destroy(nft_trans_elem_set(trans),
3943 nft_trans_elem(trans).priv);
3944 break;
3945 }
3946 kfree(trans);
3947}
3948
3949static int nf_tables_abort(struct sk_buff *skb)
3950{
3951 struct net *net = sock_net(skb->sk);
3952 struct nft_trans *trans, *next;
3953 struct nft_trans_elem *te;
3954
3955 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
3956 switch (trans->msg_type) {
3957 case NFT_MSG_NEWTABLE:
3958 if (nft_trans_table_update(trans)) {
3959 if (nft_trans_table_enable(trans)) {
3960 nf_tables_table_disable(trans->ctx.afi,
3961 trans->ctx.table);
3962 trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
3963 }
3964 nft_trans_destroy(trans);
3965 } else {
3966 list_del_rcu(&trans->ctx.table->list);
3967 }
3968 break;
3969 case NFT_MSG_DELTABLE:
3970 list_add_tail_rcu(&trans->ctx.table->list,
3971 &trans->ctx.afi->tables);
3972 nft_trans_destroy(trans);
3973 break;
3974 case NFT_MSG_NEWCHAIN:
3975 if (nft_trans_chain_update(trans)) {
3976 free_percpu(nft_trans_chain_stats(trans));
3977
3978 nft_trans_destroy(trans);
3979 } else {
3980 trans->ctx.table->use--;
3981 list_del_rcu(&trans->ctx.chain->list);
3982 nf_tables_unregister_hooks(trans->ctx.table,
3983 trans->ctx.chain,
3984 trans->ctx.afi->nops);
3985 }
3986 break;
3987 case NFT_MSG_DELCHAIN:
3988 trans->ctx.table->use++;
3989 list_add_tail_rcu(&trans->ctx.chain->list,
3990 &trans->ctx.table->chains);
3991 nft_trans_destroy(trans);
3992 break;
3993 case NFT_MSG_NEWRULE:
3994 trans->ctx.chain->use--;
3995 list_del_rcu(&nft_trans_rule(trans)->list);
3996 break;
3997 case NFT_MSG_DELRULE:
3998 trans->ctx.chain->use++;
3999 nft_rule_clear(trans->ctx.net, nft_trans_rule(trans));
4000 nft_trans_destroy(trans);
4001 break;
4002 case NFT_MSG_NEWSET:
4003 trans->ctx.table->use--;
4004 list_del_rcu(&nft_trans_set(trans)->list);
4005 break;
4006 case NFT_MSG_DELSET:
4007 trans->ctx.table->use++;
4008 list_add_tail_rcu(&nft_trans_set(trans)->list,
4009 &trans->ctx.table->sets);
4010 nft_trans_destroy(trans);
4011 break;
4012 case NFT_MSG_NEWSETELEM:
4013 te = (struct nft_trans_elem *)trans->data;
4014
4015 te->set->ops->remove(te->set, &te->elem);
4016 atomic_dec(&te->set->nelems);
4017 break;
4018 case NFT_MSG_DELSETELEM:
4019 te = (struct nft_trans_elem *)trans->data;
4020
4021 te->set->ops->activate(te->set, &te->elem);
4022 te->set->ndeact--;
4023
4024 nft_trans_destroy(trans);
4025 break;
4026 }
4027 }
4028
4029 synchronize_rcu();
4030
4031 list_for_each_entry_safe_reverse(trans, next,
4032 &net->nft.commit_list, list) {
4033 list_del(&trans->list);
4034 nf_tables_abort_release(trans);
4035 }
4036
4037 return 0;
4038}
4039
4040static const struct nfnetlink_subsystem nf_tables_subsys = {
4041 .name = "nf_tables",
4042 .subsys_id = NFNL_SUBSYS_NFTABLES,
4043 .cb_count = NFT_MSG_MAX,
4044 .cb = nf_tables_cb,
4045 .commit = nf_tables_commit,
4046 .abort = nf_tables_abort,
4047};
4048
4049int nft_chain_validate_dependency(const struct nft_chain *chain,
4050 enum nft_chain_type type)
4051{
4052 const struct nft_base_chain *basechain;
4053
4054 if (chain->flags & NFT_BASE_CHAIN) {
4055 basechain = nft_base_chain(chain);
4056 if (basechain->type->type != type)
4057 return -EOPNOTSUPP;
4058 }
4059 return 0;
4060}
4061EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
4062
4063int nft_chain_validate_hooks(const struct nft_chain *chain,
4064 unsigned int hook_flags)
4065{
4066 struct nft_base_chain *basechain;
4067
4068 if (chain->flags & NFT_BASE_CHAIN) {
4069 basechain = nft_base_chain(chain);
4070
4071 if ((1 << basechain->ops[0].hooknum) & hook_flags)
4072 return 0;
4073
4074 return -EOPNOTSUPP;
4075 }
4076
4077 return 0;
4078}
4079EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
4080
4081/*
4082 * Loop detection - walk through the ruleset beginning at the destination chain
4083 * of a new jump until either the source chain is reached (loop) or all
4084 * reachable chains have been traversed.
4085 *
4086 * The loop check is performed whenever a new jump verdict is added to an
4087 * expression or verdict map or a verdict map is bound to a new chain.
4088 */
4089
4090static int nf_tables_check_loops(const struct nft_ctx *ctx,
4091 const struct nft_chain *chain);
4092
4093static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
4094 const struct nft_set *set,
4095 const struct nft_set_iter *iter,
4096 const struct nft_set_elem *elem)
4097{
4098 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4099 const struct nft_data *data;
4100
4101 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
4102 *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
4103 return 0;
4104
4105 data = nft_set_ext_data(ext);
4106 switch (data->verdict.code) {
4107 case NFT_JUMP:
4108 case NFT_GOTO:
4109 return nf_tables_check_loops(ctx, data->verdict.chain);
4110 default:
4111 return 0;
4112 }
4113}
4114
4115static int nf_tables_check_loops(const struct nft_ctx *ctx,
4116 const struct nft_chain *chain)
4117{
4118 const struct nft_rule *rule;
4119 const struct nft_expr *expr, *last;
4120 const struct nft_set *set;
4121 struct nft_set_binding *binding;
4122 struct nft_set_iter iter;
4123
4124 if (ctx->chain == chain)
4125 return -ELOOP;
4126
4127 list_for_each_entry(rule, &chain->rules, list) {
4128 nft_rule_for_each_expr(expr, last, rule) {
4129 const struct nft_data *data = NULL;
4130 int err;
4131
4132 if (!expr->ops->validate)
4133 continue;
4134
4135 err = expr->ops->validate(ctx, expr, &data);
4136 if (err < 0)
4137 return err;
4138
4139 if (data == NULL)
4140 continue;
4141
4142 switch (data->verdict.code) {
4143 case NFT_JUMP:
4144 case NFT_GOTO:
4145 err = nf_tables_check_loops(ctx,
4146 data->verdict.chain);
4147 if (err < 0)
4148 return err;
4149 default:
4150 break;
4151 }
4152 }
4153 }
4154
4155 list_for_each_entry(set, &ctx->table->sets, list) {
4156 if (!(set->flags & NFT_SET_MAP) ||
4157 set->dtype != NFT_DATA_VERDICT)
4158 continue;
4159
4160 list_for_each_entry(binding, &set->bindings, list) {
4161 if (!(binding->flags & NFT_SET_MAP) ||
4162 binding->chain != chain)
4163 continue;
4164
4165 iter.skip = 0;
4166 iter.count = 0;
4167 iter.err = 0;
4168 iter.fn = nf_tables_loop_check_setelem;
4169
4170 set->ops->walk(ctx, set, &iter);
4171 if (iter.err < 0)
4172 return iter.err;
4173 }
4174 }
4175
4176 return 0;
4177}
4178
4179/**
4180 * nft_parse_register - parse a register value from a netlink attribute
4181 *
4182 * @attr: netlink attribute
4183 *
4184 * Parse and translate a register value from a netlink attribute.
4185 * Registers used to be 128 bit wide, these register numbers will be
4186 * mapped to the corresponding 32 bit register numbers.
4187 */
4188unsigned int nft_parse_register(const struct nlattr *attr)
4189{
4190 unsigned int reg;
4191
4192 reg = ntohl(nla_get_be32(attr));
4193 switch (reg) {
4194 case NFT_REG_VERDICT...NFT_REG_4:
4195 return reg * NFT_REG_SIZE / NFT_REG32_SIZE;
4196 default:
4197 return reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
4198 }
4199}
4200EXPORT_SYMBOL_GPL(nft_parse_register);
4201
4202/**
4203 * nft_dump_register - dump a register value to a netlink attribute
4204 *
4205 * @skb: socket buffer
4206 * @attr: attribute number
4207 * @reg: register number
4208 *
4209 * Construct a netlink attribute containing the register number. For
4210 * compatibility reasons, register numbers being a multiple of 4 are
4211 * translated to the corresponding 128 bit register numbers.
4212 */
4213int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
4214{
4215 if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
4216 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
4217 else
4218 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
4219
4220 return nla_put_be32(skb, attr, htonl(reg));
4221}
4222EXPORT_SYMBOL_GPL(nft_dump_register);
4223
4224/**
4225 * nft_validate_register_load - validate a load from a register
4226 *
4227 * @reg: the register number
4228 * @len: the length of the data
4229 *
4230 * Validate that the input register is one of the general purpose
4231 * registers and that the length of the load is within the bounds.
4232 */
4233int nft_validate_register_load(enum nft_registers reg, unsigned int len)
4234{
4235 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
4236 return -EINVAL;
4237 if (len == 0)
4238 return -EINVAL;
4239 if (reg * NFT_REG32_SIZE + len > FIELD_SIZEOF(struct nft_regs, data))
4240 return -ERANGE;
4241
4242 return 0;
4243}
4244EXPORT_SYMBOL_GPL(nft_validate_register_load);
4245
4246/**
4247 * nft_validate_register_store - validate an expressions' register store
4248 *
4249 * @ctx: context of the expression performing the load
4250 * @reg: the destination register number
4251 * @data: the data to load
4252 * @type: the data type
4253 * @len: the length of the data
4254 *
4255 * Validate that a data load uses the appropriate data type for
4256 * the destination register and the length is within the bounds.
4257 * A value of NULL for the data means that its runtime gathered
4258 * data.
4259 */
4260int nft_validate_register_store(const struct nft_ctx *ctx,
4261 enum nft_registers reg,
4262 const struct nft_data *data,
4263 enum nft_data_types type, unsigned int len)
4264{
4265 int err;
4266
4267 switch (reg) {
4268 case NFT_REG_VERDICT:
4269 if (type != NFT_DATA_VERDICT)
4270 return -EINVAL;
4271
4272 if (data != NULL &&
4273 (data->verdict.code == NFT_GOTO ||
4274 data->verdict.code == NFT_JUMP)) {
4275 err = nf_tables_check_loops(ctx, data->verdict.chain);
4276 if (err < 0)
4277 return err;
4278
4279 if (ctx->chain->level + 1 >
4280 data->verdict.chain->level) {
4281 if (ctx->chain->level + 1 == NFT_JUMP_STACK_SIZE)
4282 return -EMLINK;
4283 data->verdict.chain->level = ctx->chain->level + 1;
4284 }
4285 }
4286
4287 return 0;
4288 default:
4289 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
4290 return -EINVAL;
4291 if (len == 0)
4292 return -EINVAL;
4293 if (reg * NFT_REG32_SIZE + len >
4294 FIELD_SIZEOF(struct nft_regs, data))
4295 return -ERANGE;
4296
4297 if (data != NULL && type != NFT_DATA_VALUE)
4298 return -EINVAL;
4299 return 0;
4300 }
4301}
4302EXPORT_SYMBOL_GPL(nft_validate_register_store);
4303
4304static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
4305 [NFTA_VERDICT_CODE] = { .type = NLA_U32 },
4306 [NFTA_VERDICT_CHAIN] = { .type = NLA_STRING,
4307 .len = NFT_CHAIN_MAXNAMELEN - 1 },
4308};
4309
4310static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
4311 struct nft_data_desc *desc, const struct nlattr *nla)
4312{
4313 struct nlattr *tb[NFTA_VERDICT_MAX + 1];
4314 struct nft_chain *chain;
4315 int err;
4316
4317 err = nla_parse_nested(tb, NFTA_VERDICT_MAX, nla, nft_verdict_policy);
4318 if (err < 0)
4319 return err;
4320
4321 if (!tb[NFTA_VERDICT_CODE])
4322 return -EINVAL;
4323 data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
4324
4325 switch (data->verdict.code) {
4326 default:
4327 switch (data->verdict.code & NF_VERDICT_MASK) {
4328 case NF_ACCEPT:
4329 case NF_DROP:
4330 case NF_QUEUE:
4331 break;
4332 default:
4333 return -EINVAL;
4334 }
4335 /* fall through */
4336 case NFT_CONTINUE:
4337 case NFT_BREAK:
4338 case NFT_RETURN:
4339 desc->len = sizeof(data->verdict);
4340 break;
4341 case NFT_JUMP:
4342 case NFT_GOTO:
4343 if (!tb[NFTA_VERDICT_CHAIN])
4344 return -EINVAL;
4345 chain = nf_tables_chain_lookup(ctx->table,
4346 tb[NFTA_VERDICT_CHAIN]);
4347 if (IS_ERR(chain))
4348 return PTR_ERR(chain);
4349 if (chain->flags & NFT_BASE_CHAIN)
4350 return -EOPNOTSUPP;
4351
4352 chain->use++;
4353 data->verdict.chain = chain;
4354 desc->len = sizeof(data);
4355 break;
4356 }
4357
4358 desc->type = NFT_DATA_VERDICT;
4359 return 0;
4360}
4361
4362static void nft_verdict_uninit(const struct nft_data *data)
4363{
4364 switch (data->verdict.code) {
4365 case NFT_JUMP:
4366 case NFT_GOTO:
4367 data->verdict.chain->use--;
4368 break;
4369 }
4370}
4371
4372static int nft_verdict_dump(struct sk_buff *skb, const struct nft_data *data)
4373{
4374 struct nlattr *nest;
4375
4376 nest = nla_nest_start(skb, NFTA_DATA_VERDICT);
4377 if (!nest)
4378 goto nla_put_failure;
4379
4380 if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(data->verdict.code)))
4381 goto nla_put_failure;
4382
4383 switch (data->verdict.code) {
4384 case NFT_JUMP:
4385 case NFT_GOTO:
4386 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
4387 data->verdict.chain->name))
4388 goto nla_put_failure;
4389 }
4390 nla_nest_end(skb, nest);
4391 return 0;
4392
4393nla_put_failure:
4394 return -1;
4395}
4396
4397static int nft_value_init(const struct nft_ctx *ctx,
4398 struct nft_data *data, unsigned int size,
4399 struct nft_data_desc *desc, const struct nlattr *nla)
4400{
4401 unsigned int len;
4402
4403 len = nla_len(nla);
4404 if (len == 0)
4405 return -EINVAL;
4406 if (len > size)
4407 return -EOVERFLOW;
4408
4409 nla_memcpy(data->data, nla, len);
4410 desc->type = NFT_DATA_VALUE;
4411 desc->len = len;
4412 return 0;
4413}
4414
4415static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
4416 unsigned int len)
4417{
4418 return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
4419}
4420
4421static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
4422 [NFTA_DATA_VALUE] = { .type = NLA_BINARY },
4423 [NFTA_DATA_VERDICT] = { .type = NLA_NESTED },
4424};
4425
4426/**
4427 * nft_data_init - parse nf_tables data netlink attributes
4428 *
4429 * @ctx: context of the expression using the data
4430 * @data: destination struct nft_data
4431 * @size: maximum data length
4432 * @desc: data description
4433 * @nla: netlink attribute containing data
4434 *
4435 * Parse the netlink data attributes and initialize a struct nft_data.
4436 * The type and length of data are returned in the data description.
4437 *
4438 * The caller can indicate that it only wants to accept data of type
4439 * NFT_DATA_VALUE by passing NULL for the ctx argument.
4440 */
4441int nft_data_init(const struct nft_ctx *ctx,
4442 struct nft_data *data, unsigned int size,
4443 struct nft_data_desc *desc, const struct nlattr *nla)
4444{
4445 struct nlattr *tb[NFTA_DATA_MAX + 1];
4446 int err;
4447
4448 err = nla_parse_nested(tb, NFTA_DATA_MAX, nla, nft_data_policy);
4449 if (err < 0)
4450 return err;
4451
4452 if (tb[NFTA_DATA_VALUE])
4453 return nft_value_init(ctx, data, size, desc,
4454 tb[NFTA_DATA_VALUE]);
4455 if (tb[NFTA_DATA_VERDICT] && ctx != NULL)
4456 return nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
4457 return -EINVAL;
4458}
4459EXPORT_SYMBOL_GPL(nft_data_init);
4460
4461/**
4462 * nft_data_uninit - release a nft_data item
4463 *
4464 * @data: struct nft_data to release
4465 * @type: type of data
4466 *
4467 * Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
4468 * all others need to be released by calling this function.
4469 */
4470void nft_data_uninit(const struct nft_data *data, enum nft_data_types type)
4471{
4472 switch (type) {
4473 case NFT_DATA_VALUE:
4474 return;
4475 case NFT_DATA_VERDICT:
4476 return nft_verdict_uninit(data);
4477 default:
4478 WARN_ON(1);
4479 }
4480}
4481EXPORT_SYMBOL_GPL(nft_data_uninit);
4482
4483int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
4484 enum nft_data_types type, unsigned int len)
4485{
4486 struct nlattr *nest;
4487 int err;
4488
4489 nest = nla_nest_start(skb, attr);
4490 if (nest == NULL)
4491 return -1;
4492
4493 switch (type) {
4494 case NFT_DATA_VALUE:
4495 err = nft_value_dump(skb, data, len);
4496 break;
4497 case NFT_DATA_VERDICT:
4498 err = nft_verdict_dump(skb, data);
4499 break;
4500 default:
4501 err = -EINVAL;
4502 WARN_ON(1);
4503 }
4504
4505 nla_nest_end(skb, nest);
4506 return err;
4507}
4508EXPORT_SYMBOL_GPL(nft_data_dump);
4509
4510static int nf_tables_init_net(struct net *net)
4511{
4512 INIT_LIST_HEAD(&net->nft.af_info);
4513 INIT_LIST_HEAD(&net->nft.commit_list);
4514 net->nft.base_seq = 1;
4515 return 0;
4516}
4517
4518static struct pernet_operations nf_tables_net_ops = {
4519 .init = nf_tables_init_net,
4520};
4521
4522static int __init nf_tables_module_init(void)
4523{
4524 int err;
4525
4526 info = kmalloc(sizeof(struct nft_expr_info) * NFT_RULE_MAXEXPRS,
4527 GFP_KERNEL);
4528 if (info == NULL) {
4529 err = -ENOMEM;
4530 goto err1;
4531 }
4532
4533 err = nf_tables_core_module_init();
4534 if (err < 0)
4535 goto err2;
4536
4537 err = nfnetlink_subsys_register(&nf_tables_subsys);
4538 if (err < 0)
4539 goto err3;
4540
4541 pr_info("nf_tables: (c) 2007-2009 Patrick McHardy <kaber@trash.net>\n");
4542 return register_pernet_subsys(&nf_tables_net_ops);
4543err3:
4544 nf_tables_core_module_exit();
4545err2:
4546 kfree(info);
4547err1:
4548 return err;
4549}
4550
4551static void __exit nf_tables_module_exit(void)
4552{
4553 unregister_pernet_subsys(&nf_tables_net_ops);
4554 nfnetlink_subsys_unregister(&nf_tables_subsys);
4555 rcu_barrier();
4556 nf_tables_core_module_exit();
4557 kfree(info);
4558}
4559
4560module_init(nf_tables_module_init);
4561module_exit(nf_tables_module_exit);
4562
4563MODULE_LICENSE("GPL");
4564MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
4565MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);
Ubuntu Jammy Linux kernel mirror