]> 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
selftests: add script to stress-test nft packet path vs. control plane
[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/vmalloc.h>
17#include <linux/rhashtable.h>
18#include <linux/netfilter.h>
19#include <linux/netfilter/nfnetlink.h>
20#include <linux/netfilter/nf_tables.h>
21#include <net/netfilter/nf_flow_table.h>
22#include <net/netfilter/nf_tables_core.h>
23#include <net/netfilter/nf_tables.h>
24#include <net/net_namespace.h>
25#include <net/sock.h>
26
27static LIST_HEAD(nf_tables_expressions);
28static LIST_HEAD(nf_tables_objects);
29static LIST_HEAD(nf_tables_flowtables);
30static LIST_HEAD(nf_tables_destroy_list);
31static DEFINE_SPINLOCK(nf_tables_destroy_list_lock);
32static u64 table_handle;
33
34enum {
35 NFT_VALIDATE_SKIP = 0,
36 NFT_VALIDATE_NEED,
37 NFT_VALIDATE_DO,
38};
39
40static u32 nft_chain_hash(const void *data, u32 len, u32 seed);
41static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed);
42static int nft_chain_hash_cmp(struct rhashtable_compare_arg *, const void *);
43
44static const struct rhashtable_params nft_chain_ht_params = {
45 .head_offset = offsetof(struct nft_chain, rhlhead),
46 .key_offset = offsetof(struct nft_chain, name),
47 .hashfn = nft_chain_hash,
48 .obj_hashfn = nft_chain_hash_obj,
49 .obj_cmpfn = nft_chain_hash_cmp,
50 .locks_mul = 1,
51 .automatic_shrinking = true,
52};
53
54static void nft_validate_state_update(struct net *net, u8 new_validate_state)
55{
56 switch (net->nft.validate_state) {
57 case NFT_VALIDATE_SKIP:
58 WARN_ON_ONCE(new_validate_state == NFT_VALIDATE_DO);
59 break;
60 case NFT_VALIDATE_NEED:
61 break;
62 case NFT_VALIDATE_DO:
63 if (new_validate_state == NFT_VALIDATE_NEED)
64 return;
65 }
66
67 net->nft.validate_state = new_validate_state;
68}
69static void nf_tables_trans_destroy_work(struct work_struct *w);
70static DECLARE_WORK(trans_destroy_work, nf_tables_trans_destroy_work);
71
72static void nft_ctx_init(struct nft_ctx *ctx,
73 struct net *net,
74 const struct sk_buff *skb,
75 const struct nlmsghdr *nlh,
76 u8 family,
77 struct nft_table *table,
78 struct nft_chain *chain,
79 const struct nlattr * const *nla)
80{
81 ctx->net = net;
82 ctx->family = family;
83 ctx->level = 0;
84 ctx->table = table;
85 ctx->chain = chain;
86 ctx->nla = nla;
87 ctx->portid = NETLINK_CB(skb).portid;
88 ctx->report = nlmsg_report(nlh);
89 ctx->seq = nlh->nlmsg_seq;
90}
91
92static struct nft_trans *nft_trans_alloc_gfp(const struct nft_ctx *ctx,
93 int msg_type, u32 size, gfp_t gfp)
94{
95 struct nft_trans *trans;
96
97 trans = kzalloc(sizeof(struct nft_trans) + size, gfp);
98 if (trans == NULL)
99 return NULL;
100
101 trans->msg_type = msg_type;
102 trans->ctx = *ctx;
103
104 return trans;
105}
106
107static struct nft_trans *nft_trans_alloc(const struct nft_ctx *ctx,
108 int msg_type, u32 size)
109{
110 return nft_trans_alloc_gfp(ctx, msg_type, size, GFP_KERNEL);
111}
112
113static void nft_trans_destroy(struct nft_trans *trans)
114{
115 list_del(&trans->list);
116 kfree(trans);
117}
118
119static int nf_tables_register_hook(struct net *net,
120 const struct nft_table *table,
121 struct nft_chain *chain)
122{
123 const struct nft_base_chain *basechain;
124 const struct nf_hook_ops *ops;
125
126 if (table->flags & NFT_TABLE_F_DORMANT ||
127 !nft_is_base_chain(chain))
128 return 0;
129
130 basechain = nft_base_chain(chain);
131 ops = &basechain->ops;
132
133 if (basechain->type->ops_register)
134 return basechain->type->ops_register(net, ops);
135
136 return nf_register_net_hook(net, ops);
137}
138
139static void nf_tables_unregister_hook(struct net *net,
140 const struct nft_table *table,
141 struct nft_chain *chain)
142{
143 const struct nft_base_chain *basechain;
144 const struct nf_hook_ops *ops;
145
146 if (table->flags & NFT_TABLE_F_DORMANT ||
147 !nft_is_base_chain(chain))
148 return;
149 basechain = nft_base_chain(chain);
150 ops = &basechain->ops;
151
152 if (basechain->type->ops_unregister)
153 return basechain->type->ops_unregister(net, ops);
154
155 nf_unregister_net_hook(net, ops);
156}
157
158static int nft_trans_table_add(struct nft_ctx *ctx, int msg_type)
159{
160 struct nft_trans *trans;
161
162 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_table));
163 if (trans == NULL)
164 return -ENOMEM;
165
166 if (msg_type == NFT_MSG_NEWTABLE)
167 nft_activate_next(ctx->net, ctx->table);
168
169 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
170 return 0;
171}
172
173static int nft_deltable(struct nft_ctx *ctx)
174{
175 int err;
176
177 err = nft_trans_table_add(ctx, NFT_MSG_DELTABLE);
178 if (err < 0)
179 return err;
180
181 nft_deactivate_next(ctx->net, ctx->table);
182 return err;
183}
184
185static int nft_trans_chain_add(struct nft_ctx *ctx, int msg_type)
186{
187 struct nft_trans *trans;
188
189 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_chain));
190 if (trans == NULL)
191 return -ENOMEM;
192
193 if (msg_type == NFT_MSG_NEWCHAIN)
194 nft_activate_next(ctx->net, ctx->chain);
195
196 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
197 return 0;
198}
199
200static int nft_delchain(struct nft_ctx *ctx)
201{
202 int err;
203
204 err = nft_trans_chain_add(ctx, NFT_MSG_DELCHAIN);
205 if (err < 0)
206 return err;
207
208 ctx->table->use--;
209 nft_deactivate_next(ctx->net, ctx->chain);
210
211 return err;
212}
213
214/* either expr ops provide both activate/deactivate, or neither */
215static bool nft_expr_check_ops(const struct nft_expr_ops *ops)
216{
217 if (!ops)
218 return true;
219
220 if (WARN_ON_ONCE((!ops->activate ^ !ops->deactivate)))
221 return false;
222
223 return true;
224}
225
226static void nft_rule_expr_activate(const struct nft_ctx *ctx,
227 struct nft_rule *rule)
228{
229 struct nft_expr *expr;
230
231 expr = nft_expr_first(rule);
232 while (expr != nft_expr_last(rule) && expr->ops) {
233 if (expr->ops->activate)
234 expr->ops->activate(ctx, expr);
235
236 expr = nft_expr_next(expr);
237 }
238}
239
240static void nft_rule_expr_deactivate(const struct nft_ctx *ctx,
241 struct nft_rule *rule)
242{
243 struct nft_expr *expr;
244
245 expr = nft_expr_first(rule);
246 while (expr != nft_expr_last(rule) && expr->ops) {
247 if (expr->ops->deactivate)
248 expr->ops->deactivate(ctx, expr);
249
250 expr = nft_expr_next(expr);
251 }
252}
253
254static int
255nf_tables_delrule_deactivate(struct nft_ctx *ctx, struct nft_rule *rule)
256{
257 /* You cannot delete the same rule twice */
258 if (nft_is_active_next(ctx->net, rule)) {
259 nft_deactivate_next(ctx->net, rule);
260 ctx->chain->use--;
261 return 0;
262 }
263 return -ENOENT;
264}
265
266static struct nft_trans *nft_trans_rule_add(struct nft_ctx *ctx, int msg_type,
267 struct nft_rule *rule)
268{
269 struct nft_trans *trans;
270
271 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_rule));
272 if (trans == NULL)
273 return NULL;
274
275 if (msg_type == NFT_MSG_NEWRULE && ctx->nla[NFTA_RULE_ID] != NULL) {
276 nft_trans_rule_id(trans) =
277 ntohl(nla_get_be32(ctx->nla[NFTA_RULE_ID]));
278 }
279 nft_trans_rule(trans) = rule;
280 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
281
282 return trans;
283}
284
285static int nft_delrule(struct nft_ctx *ctx, struct nft_rule *rule)
286{
287 struct nft_trans *trans;
288 int err;
289
290 trans = nft_trans_rule_add(ctx, NFT_MSG_DELRULE, rule);
291 if (trans == NULL)
292 return -ENOMEM;
293
294 err = nf_tables_delrule_deactivate(ctx, rule);
295 if (err < 0) {
296 nft_trans_destroy(trans);
297 return err;
298 }
299 nft_rule_expr_deactivate(ctx, rule);
300
301 return 0;
302}
303
304static int nft_delrule_by_chain(struct nft_ctx *ctx)
305{
306 struct nft_rule *rule;
307 int err;
308
309 list_for_each_entry(rule, &ctx->chain->rules, list) {
310 err = nft_delrule(ctx, rule);
311 if (err < 0)
312 return err;
313 }
314 return 0;
315}
316
317static int nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
318 struct nft_set *set)
319{
320 struct nft_trans *trans;
321
322 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_set));
323 if (trans == NULL)
324 return -ENOMEM;
325
326 if (msg_type == NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID] != NULL) {
327 nft_trans_set_id(trans) =
328 ntohl(nla_get_be32(ctx->nla[NFTA_SET_ID]));
329 nft_activate_next(ctx->net, set);
330 }
331 nft_trans_set(trans) = set;
332 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
333
334 return 0;
335}
336
337static int nft_delset(const struct nft_ctx *ctx, struct nft_set *set)
338{
339 int err;
340
341 err = nft_trans_set_add(ctx, NFT_MSG_DELSET, set);
342 if (err < 0)
343 return err;
344
345 nft_deactivate_next(ctx->net, set);
346 ctx->table->use--;
347
348 return err;
349}
350
351static int nft_trans_obj_add(struct nft_ctx *ctx, int msg_type,
352 struct nft_object *obj)
353{
354 struct nft_trans *trans;
355
356 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_obj));
357 if (trans == NULL)
358 return -ENOMEM;
359
360 if (msg_type == NFT_MSG_NEWOBJ)
361 nft_activate_next(ctx->net, obj);
362
363 nft_trans_obj(trans) = obj;
364 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
365
366 return 0;
367}
368
369static int nft_delobj(struct nft_ctx *ctx, struct nft_object *obj)
370{
371 int err;
372
373 err = nft_trans_obj_add(ctx, NFT_MSG_DELOBJ, obj);
374 if (err < 0)
375 return err;
376
377 nft_deactivate_next(ctx->net, obj);
378 ctx->table->use--;
379
380 return err;
381}
382
383static int nft_trans_flowtable_add(struct nft_ctx *ctx, int msg_type,
384 struct nft_flowtable *flowtable)
385{
386 struct nft_trans *trans;
387
388 trans = nft_trans_alloc(ctx, msg_type,
389 sizeof(struct nft_trans_flowtable));
390 if (trans == NULL)
391 return -ENOMEM;
392
393 if (msg_type == NFT_MSG_NEWFLOWTABLE)
394 nft_activate_next(ctx->net, flowtable);
395
396 nft_trans_flowtable(trans) = flowtable;
397 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
398
399 return 0;
400}
401
402static int nft_delflowtable(struct nft_ctx *ctx,
403 struct nft_flowtable *flowtable)
404{
405 int err;
406
407 err = nft_trans_flowtable_add(ctx, NFT_MSG_DELFLOWTABLE, flowtable);
408 if (err < 0)
409 return err;
410
411 nft_deactivate_next(ctx->net, flowtable);
412 ctx->table->use--;
413
414 return err;
415}
416
417/*
418 * Tables
419 */
420
421static struct nft_table *nft_table_lookup(const struct net *net,
422 const struct nlattr *nla,
423 u8 family, u8 genmask)
424{
425 struct nft_table *table;
426
427 if (nla == NULL)
428 return ERR_PTR(-EINVAL);
429
430 list_for_each_entry_rcu(table, &net->nft.tables, list) {
431 if (!nla_strcmp(nla, table->name) &&
432 table->family == family &&
433 nft_active_genmask(table, genmask))
434 return table;
435 }
436
437 return ERR_PTR(-ENOENT);
438}
439
440static struct nft_table *nft_table_lookup_byhandle(const struct net *net,
441 const struct nlattr *nla,
442 u8 genmask)
443{
444 struct nft_table *table;
445
446 list_for_each_entry(table, &net->nft.tables, list) {
447 if (be64_to_cpu(nla_get_be64(nla)) == table->handle &&
448 nft_active_genmask(table, genmask))
449 return table;
450 }
451
452 return ERR_PTR(-ENOENT);
453}
454
455static inline u64 nf_tables_alloc_handle(struct nft_table *table)
456{
457 return ++table->hgenerator;
458}
459
460static const struct nft_chain_type *chain_type[NFPROTO_NUMPROTO][NFT_CHAIN_T_MAX];
461
462static const struct nft_chain_type *
463__nf_tables_chain_type_lookup(const struct nlattr *nla, u8 family)
464{
465 int i;
466
467 for (i = 0; i < NFT_CHAIN_T_MAX; i++) {
468 if (chain_type[family][i] != NULL &&
469 !nla_strcmp(nla, chain_type[family][i]->name))
470 return chain_type[family][i];
471 }
472 return NULL;
473}
474
475/*
476 * Loading a module requires dropping mutex that guards the
477 * transaction.
478 * We first need to abort any pending transactions as once
479 * mutex is unlocked a different client could start a new
480 * transaction. It must not see any 'future generation'
481 * changes * as these changes will never happen.
482 */
483#ifdef CONFIG_MODULES
484static int __nf_tables_abort(struct net *net);
485
486static void nft_request_module(struct net *net, const char *fmt, ...)
487{
488 char module_name[MODULE_NAME_LEN];
489 va_list args;
490 int ret;
491
492 __nf_tables_abort(net);
493
494 va_start(args, fmt);
495 ret = vsnprintf(module_name, MODULE_NAME_LEN, fmt, args);
496 va_end(args);
497 if (WARN(ret >= MODULE_NAME_LEN, "truncated: '%s' (len %d)", module_name, ret))
498 return;
499
500 mutex_unlock(&net->nft.commit_mutex);
501 request_module("%s", module_name);
502 mutex_lock(&net->nft.commit_mutex);
503}
504#endif
505
506static void lockdep_nfnl_nft_mutex_not_held(void)
507{
508#ifdef CONFIG_PROVE_LOCKING
509 WARN_ON_ONCE(lockdep_nfnl_is_held(NFNL_SUBSYS_NFTABLES));
510#endif
511}
512
513static const struct nft_chain_type *
514nf_tables_chain_type_lookup(struct net *net, const struct nlattr *nla,
515 u8 family, bool autoload)
516{
517 const struct nft_chain_type *type;
518
519 type = __nf_tables_chain_type_lookup(nla, family);
520 if (type != NULL)
521 return type;
522
523 lockdep_nfnl_nft_mutex_not_held();
524#ifdef CONFIG_MODULES
525 if (autoload) {
526 nft_request_module(net, "nft-chain-%u-%.*s", family,
527 nla_len(nla), (const char *)nla_data(nla));
528 type = __nf_tables_chain_type_lookup(nla, family);
529 if (type != NULL)
530 return ERR_PTR(-EAGAIN);
531 }
532#endif
533 return ERR_PTR(-ENOENT);
534}
535
536static const struct nla_policy nft_table_policy[NFTA_TABLE_MAX + 1] = {
537 [NFTA_TABLE_NAME] = { .type = NLA_STRING,
538 .len = NFT_TABLE_MAXNAMELEN - 1 },
539 [NFTA_TABLE_FLAGS] = { .type = NLA_U32 },
540 [NFTA_TABLE_HANDLE] = { .type = NLA_U64 },
541};
542
543static int nf_tables_fill_table_info(struct sk_buff *skb, struct net *net,
544 u32 portid, u32 seq, int event, u32 flags,
545 int family, const struct nft_table *table)
546{
547 struct nlmsghdr *nlh;
548 struct nfgenmsg *nfmsg;
549
550 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
551 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
552 if (nlh == NULL)
553 goto nla_put_failure;
554
555 nfmsg = nlmsg_data(nlh);
556 nfmsg->nfgen_family = family;
557 nfmsg->version = NFNETLINK_V0;
558 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
559
560 if (nla_put_string(skb, NFTA_TABLE_NAME, table->name) ||
561 nla_put_be32(skb, NFTA_TABLE_FLAGS, htonl(table->flags)) ||
562 nla_put_be32(skb, NFTA_TABLE_USE, htonl(table->use)) ||
563 nla_put_be64(skb, NFTA_TABLE_HANDLE, cpu_to_be64(table->handle),
564 NFTA_TABLE_PAD))
565 goto nla_put_failure;
566
567 nlmsg_end(skb, nlh);
568 return 0;
569
570nla_put_failure:
571 nlmsg_trim(skb, nlh);
572 return -1;
573}
574
575static void nf_tables_table_notify(const struct nft_ctx *ctx, int event)
576{
577 struct sk_buff *skb;
578 int err;
579
580 if (!ctx->report &&
581 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
582 return;
583
584 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
585 if (skb == NULL)
586 goto err;
587
588 err = nf_tables_fill_table_info(skb, ctx->net, ctx->portid, ctx->seq,
589 event, 0, ctx->family, ctx->table);
590 if (err < 0) {
591 kfree_skb(skb);
592 goto err;
593 }
594
595 nfnetlink_send(skb, ctx->net, ctx->portid, NFNLGRP_NFTABLES,
596 ctx->report, GFP_KERNEL);
597 return;
598err:
599 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
600}
601
602static int nf_tables_dump_tables(struct sk_buff *skb,
603 struct netlink_callback *cb)
604{
605 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
606 const struct nft_table *table;
607 unsigned int idx = 0, s_idx = cb->args[0];
608 struct net *net = sock_net(skb->sk);
609 int family = nfmsg->nfgen_family;
610
611 rcu_read_lock();
612 cb->seq = net->nft.base_seq;
613
614 list_for_each_entry_rcu(table, &net->nft.tables, list) {
615 if (family != NFPROTO_UNSPEC && family != table->family)
616 continue;
617
618 if (idx < s_idx)
619 goto cont;
620 if (idx > s_idx)
621 memset(&cb->args[1], 0,
622 sizeof(cb->args) - sizeof(cb->args[0]));
623 if (!nft_is_active(net, table))
624 continue;
625 if (nf_tables_fill_table_info(skb, net,
626 NETLINK_CB(cb->skb).portid,
627 cb->nlh->nlmsg_seq,
628 NFT_MSG_NEWTABLE, NLM_F_MULTI,
629 table->family, table) < 0)
630 goto done;
631
632 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
633cont:
634 idx++;
635 }
636done:
637 rcu_read_unlock();
638 cb->args[0] = idx;
639 return skb->len;
640}
641
642static int nft_netlink_dump_start_rcu(struct sock *nlsk, struct sk_buff *skb,
643 const struct nlmsghdr *nlh,
644 struct netlink_dump_control *c)
645{
646 int err;
647
648 if (!try_module_get(THIS_MODULE))
649 return -EINVAL;
650
651 rcu_read_unlock();
652 err = netlink_dump_start(nlsk, skb, nlh, c);
653 rcu_read_lock();
654 module_put(THIS_MODULE);
655
656 return err;
657}
658
659/* called with rcu_read_lock held */
660static int nf_tables_gettable(struct net *net, struct sock *nlsk,
661 struct sk_buff *skb, const struct nlmsghdr *nlh,
662 const struct nlattr * const nla[],
663 struct netlink_ext_ack *extack)
664{
665 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
666 u8 genmask = nft_genmask_cur(net);
667 const struct nft_table *table;
668 struct sk_buff *skb2;
669 int family = nfmsg->nfgen_family;
670 int err;
671
672 if (nlh->nlmsg_flags & NLM_F_DUMP) {
673 struct netlink_dump_control c = {
674 .dump = nf_tables_dump_tables,
675 .module = THIS_MODULE,
676 };
677
678 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
679 }
680
681 table = nft_table_lookup(net, nla[NFTA_TABLE_NAME], family, genmask);
682 if (IS_ERR(table)) {
683 NL_SET_BAD_ATTR(extack, nla[NFTA_TABLE_NAME]);
684 return PTR_ERR(table);
685 }
686
687 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
688 if (!skb2)
689 return -ENOMEM;
690
691 err = nf_tables_fill_table_info(skb2, net, NETLINK_CB(skb).portid,
692 nlh->nlmsg_seq, NFT_MSG_NEWTABLE, 0,
693 family, table);
694 if (err < 0)
695 goto err;
696
697 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
698
699err:
700 kfree_skb(skb2);
701 return err;
702}
703
704static void nft_table_disable(struct net *net, struct nft_table *table, u32 cnt)
705{
706 struct nft_chain *chain;
707 u32 i = 0;
708
709 list_for_each_entry(chain, &table->chains, list) {
710 if (!nft_is_active_next(net, chain))
711 continue;
712 if (!nft_is_base_chain(chain))
713 continue;
714
715 if (cnt && i++ == cnt)
716 break;
717
718 nf_unregister_net_hook(net, &nft_base_chain(chain)->ops);
719 }
720}
721
722static int nf_tables_table_enable(struct net *net, struct nft_table *table)
723{
724 struct nft_chain *chain;
725 int err, i = 0;
726
727 list_for_each_entry(chain, &table->chains, list) {
728 if (!nft_is_active_next(net, chain))
729 continue;
730 if (!nft_is_base_chain(chain))
731 continue;
732
733 err = nf_register_net_hook(net, &nft_base_chain(chain)->ops);
734 if (err < 0)
735 goto err;
736
737 i++;
738 }
739 return 0;
740err:
741 if (i)
742 nft_table_disable(net, table, i);
743 return err;
744}
745
746static void nf_tables_table_disable(struct net *net, struct nft_table *table)
747{
748 nft_table_disable(net, table, 0);
749}
750
751static int nf_tables_updtable(struct nft_ctx *ctx)
752{
753 struct nft_trans *trans;
754 u32 flags;
755 int ret = 0;
756
757 if (!ctx->nla[NFTA_TABLE_FLAGS])
758 return 0;
759
760 flags = ntohl(nla_get_be32(ctx->nla[NFTA_TABLE_FLAGS]));
761 if (flags & ~NFT_TABLE_F_DORMANT)
762 return -EINVAL;
763
764 if (flags == ctx->table->flags)
765 return 0;
766
767 trans = nft_trans_alloc(ctx, NFT_MSG_NEWTABLE,
768 sizeof(struct nft_trans_table));
769 if (trans == NULL)
770 return -ENOMEM;
771
772 if ((flags & NFT_TABLE_F_DORMANT) &&
773 !(ctx->table->flags & NFT_TABLE_F_DORMANT)) {
774 nft_trans_table_enable(trans) = false;
775 } else if (!(flags & NFT_TABLE_F_DORMANT) &&
776 ctx->table->flags & NFT_TABLE_F_DORMANT) {
777 ret = nf_tables_table_enable(ctx->net, ctx->table);
778 if (ret >= 0) {
779 ctx->table->flags &= ~NFT_TABLE_F_DORMANT;
780 nft_trans_table_enable(trans) = true;
781 }
782 }
783 if (ret < 0)
784 goto err;
785
786 nft_trans_table_update(trans) = true;
787 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
788 return 0;
789err:
790 nft_trans_destroy(trans);
791 return ret;
792}
793
794static u32 nft_chain_hash(const void *data, u32 len, u32 seed)
795{
796 const char *name = data;
797
798 return jhash(name, strlen(name), seed);
799}
800
801static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed)
802{
803 const struct nft_chain *chain = data;
804
805 return nft_chain_hash(chain->name, 0, seed);
806}
807
808static int nft_chain_hash_cmp(struct rhashtable_compare_arg *arg,
809 const void *ptr)
810{
811 const struct nft_chain *chain = ptr;
812 const char *name = arg->key;
813
814 return strcmp(chain->name, name);
815}
816
817static int nf_tables_newtable(struct net *net, struct sock *nlsk,
818 struct sk_buff *skb, const struct nlmsghdr *nlh,
819 const struct nlattr * const nla[],
820 struct netlink_ext_ack *extack)
821{
822 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
823 u8 genmask = nft_genmask_next(net);
824 int family = nfmsg->nfgen_family;
825 const struct nlattr *attr;
826 struct nft_table *table;
827 u32 flags = 0;
828 struct nft_ctx ctx;
829 int err;
830
831 lockdep_assert_held(&net->nft.commit_mutex);
832 attr = nla[NFTA_TABLE_NAME];
833 table = nft_table_lookup(net, attr, family, genmask);
834 if (IS_ERR(table)) {
835 if (PTR_ERR(table) != -ENOENT)
836 return PTR_ERR(table);
837 } else {
838 if (nlh->nlmsg_flags & NLM_F_EXCL) {
839 NL_SET_BAD_ATTR(extack, attr);
840 return -EEXIST;
841 }
842 if (nlh->nlmsg_flags & NLM_F_REPLACE)
843 return -EOPNOTSUPP;
844
845 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
846 return nf_tables_updtable(&ctx);
847 }
848
849 if (nla[NFTA_TABLE_FLAGS]) {
850 flags = ntohl(nla_get_be32(nla[NFTA_TABLE_FLAGS]));
851 if (flags & ~NFT_TABLE_F_DORMANT)
852 return -EINVAL;
853 }
854
855 err = -ENOMEM;
856 table = kzalloc(sizeof(*table), GFP_KERNEL);
857 if (table == NULL)
858 goto err_kzalloc;
859
860 table->name = nla_strdup(attr, GFP_KERNEL);
861 if (table->name == NULL)
862 goto err_strdup;
863
864 err = rhltable_init(&table->chains_ht, &nft_chain_ht_params);
865 if (err)
866 goto err_chain_ht;
867
868 INIT_LIST_HEAD(&table->chains);
869 INIT_LIST_HEAD(&table->sets);
870 INIT_LIST_HEAD(&table->objects);
871 INIT_LIST_HEAD(&table->flowtables);
872 table->family = family;
873 table->flags = flags;
874 table->handle = ++table_handle;
875
876 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
877 err = nft_trans_table_add(&ctx, NFT_MSG_NEWTABLE);
878 if (err < 0)
879 goto err_trans;
880
881 list_add_tail_rcu(&table->list, &net->nft.tables);
882 return 0;
883err_trans:
884 rhltable_destroy(&table->chains_ht);
885err_chain_ht:
886 kfree(table->name);
887err_strdup:
888 kfree(table);
889err_kzalloc:
890 return err;
891}
892
893static int nft_flush_table(struct nft_ctx *ctx)
894{
895 struct nft_flowtable *flowtable, *nft;
896 struct nft_chain *chain, *nc;
897 struct nft_object *obj, *ne;
898 struct nft_set *set, *ns;
899 int err;
900
901 list_for_each_entry(chain, &ctx->table->chains, list) {
902 if (!nft_is_active_next(ctx->net, chain))
903 continue;
904
905 ctx->chain = chain;
906
907 err = nft_delrule_by_chain(ctx);
908 if (err < 0)
909 goto out;
910 }
911
912 list_for_each_entry_safe(set, ns, &ctx->table->sets, list) {
913 if (!nft_is_active_next(ctx->net, set))
914 continue;
915
916 if (nft_set_is_anonymous(set) &&
917 !list_empty(&set->bindings))
918 continue;
919
920 err = nft_delset(ctx, set);
921 if (err < 0)
922 goto out;
923 }
924
925 list_for_each_entry_safe(flowtable, nft, &ctx->table->flowtables, list) {
926 err = nft_delflowtable(ctx, flowtable);
927 if (err < 0)
928 goto out;
929 }
930
931 list_for_each_entry_safe(obj, ne, &ctx->table->objects, list) {
932 err = nft_delobj(ctx, obj);
933 if (err < 0)
934 goto out;
935 }
936
937 list_for_each_entry_safe(chain, nc, &ctx->table->chains, list) {
938 if (!nft_is_active_next(ctx->net, chain))
939 continue;
940
941 ctx->chain = chain;
942
943 err = nft_delchain(ctx);
944 if (err < 0)
945 goto out;
946 }
947
948 err = nft_deltable(ctx);
949out:
950 return err;
951}
952
953static int nft_flush(struct nft_ctx *ctx, int family)
954{
955 struct nft_table *table, *nt;
956 const struct nlattr * const *nla = ctx->nla;
957 int err = 0;
958
959 list_for_each_entry_safe(table, nt, &ctx->net->nft.tables, list) {
960 if (family != AF_UNSPEC && table->family != family)
961 continue;
962
963 ctx->family = table->family;
964
965 if (!nft_is_active_next(ctx->net, table))
966 continue;
967
968 if (nla[NFTA_TABLE_NAME] &&
969 nla_strcmp(nla[NFTA_TABLE_NAME], table->name) != 0)
970 continue;
971
972 ctx->table = table;
973
974 err = nft_flush_table(ctx);
975 if (err < 0)
976 goto out;
977 }
978out:
979 return err;
980}
981
982static int nf_tables_deltable(struct net *net, struct sock *nlsk,
983 struct sk_buff *skb, const struct nlmsghdr *nlh,
984 const struct nlattr * const nla[],
985 struct netlink_ext_ack *extack)
986{
987 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
988 u8 genmask = nft_genmask_next(net);
989 int family = nfmsg->nfgen_family;
990 const struct nlattr *attr;
991 struct nft_table *table;
992 struct nft_ctx ctx;
993
994 nft_ctx_init(&ctx, net, skb, nlh, 0, NULL, NULL, nla);
995 if (family == AF_UNSPEC ||
996 (!nla[NFTA_TABLE_NAME] && !nla[NFTA_TABLE_HANDLE]))
997 return nft_flush(&ctx, family);
998
999 if (nla[NFTA_TABLE_HANDLE]) {
1000 attr = nla[NFTA_TABLE_HANDLE];
1001 table = nft_table_lookup_byhandle(net, attr, genmask);
1002 } else {
1003 attr = nla[NFTA_TABLE_NAME];
1004 table = nft_table_lookup(net, attr, family, genmask);
1005 }
1006
1007 if (IS_ERR(table)) {
1008 NL_SET_BAD_ATTR(extack, attr);
1009 return PTR_ERR(table);
1010 }
1011
1012 if (nlh->nlmsg_flags & NLM_F_NONREC &&
1013 table->use > 0)
1014 return -EBUSY;
1015
1016 ctx.family = family;
1017 ctx.table = table;
1018
1019 return nft_flush_table(&ctx);
1020}
1021
1022static void nf_tables_table_destroy(struct nft_ctx *ctx)
1023{
1024 if (WARN_ON(ctx->table->use > 0))
1025 return;
1026
1027 rhltable_destroy(&ctx->table->chains_ht);
1028 kfree(ctx->table->name);
1029 kfree(ctx->table);
1030}
1031
1032void nft_register_chain_type(const struct nft_chain_type *ctype)
1033{
1034 if (WARN_ON(ctype->family >= NFPROTO_NUMPROTO))
1035 return;
1036
1037 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1038 if (WARN_ON(chain_type[ctype->family][ctype->type] != NULL)) {
1039 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1040 return;
1041 }
1042 chain_type[ctype->family][ctype->type] = ctype;
1043 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1044}
1045EXPORT_SYMBOL_GPL(nft_register_chain_type);
1046
1047void nft_unregister_chain_type(const struct nft_chain_type *ctype)
1048{
1049 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1050 chain_type[ctype->family][ctype->type] = NULL;
1051 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1052}
1053EXPORT_SYMBOL_GPL(nft_unregister_chain_type);
1054
1055/*
1056 * Chains
1057 */
1058
1059static struct nft_chain *
1060nft_chain_lookup_byhandle(const struct nft_table *table, u64 handle, u8 genmask)
1061{
1062 struct nft_chain *chain;
1063
1064 list_for_each_entry(chain, &table->chains, list) {
1065 if (chain->handle == handle &&
1066 nft_active_genmask(chain, genmask))
1067 return chain;
1068 }
1069
1070 return ERR_PTR(-ENOENT);
1071}
1072
1073static bool lockdep_commit_lock_is_held(struct net *net)
1074{
1075#ifdef CONFIG_PROVE_LOCKING
1076 return lockdep_is_held(&net->nft.commit_mutex);
1077#else
1078 return true;
1079#endif
1080}
1081
1082static struct nft_chain *nft_chain_lookup(struct net *net,
1083 struct nft_table *table,
1084 const struct nlattr *nla, u8 genmask)
1085{
1086 char search[NFT_CHAIN_MAXNAMELEN + 1];
1087 struct rhlist_head *tmp, *list;
1088 struct nft_chain *chain;
1089
1090 if (nla == NULL)
1091 return ERR_PTR(-EINVAL);
1092
1093 nla_strlcpy(search, nla, sizeof(search));
1094
1095 WARN_ON(!rcu_read_lock_held() &&
1096 !lockdep_commit_lock_is_held(net));
1097
1098 chain = ERR_PTR(-ENOENT);
1099 rcu_read_lock();
1100 list = rhltable_lookup(&table->chains_ht, search, nft_chain_ht_params);
1101 if (!list)
1102 goto out_unlock;
1103
1104 rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
1105 if (nft_active_genmask(chain, genmask))
1106 goto out_unlock;
1107 }
1108 chain = ERR_PTR(-ENOENT);
1109out_unlock:
1110 rcu_read_unlock();
1111 return chain;
1112}
1113
1114static const struct nla_policy nft_chain_policy[NFTA_CHAIN_MAX + 1] = {
1115 [NFTA_CHAIN_TABLE] = { .type = NLA_STRING,
1116 .len = NFT_TABLE_MAXNAMELEN - 1 },
1117 [NFTA_CHAIN_HANDLE] = { .type = NLA_U64 },
1118 [NFTA_CHAIN_NAME] = { .type = NLA_STRING,
1119 .len = NFT_CHAIN_MAXNAMELEN - 1 },
1120 [NFTA_CHAIN_HOOK] = { .type = NLA_NESTED },
1121 [NFTA_CHAIN_POLICY] = { .type = NLA_U32 },
1122 [NFTA_CHAIN_TYPE] = { .type = NLA_STRING },
1123 [NFTA_CHAIN_COUNTERS] = { .type = NLA_NESTED },
1124};
1125
1126static const struct nla_policy nft_hook_policy[NFTA_HOOK_MAX + 1] = {
1127 [NFTA_HOOK_HOOKNUM] = { .type = NLA_U32 },
1128 [NFTA_HOOK_PRIORITY] = { .type = NLA_U32 },
1129 [NFTA_HOOK_DEV] = { .type = NLA_STRING,
1130 .len = IFNAMSIZ - 1 },
1131};
1132
1133static int nft_dump_stats(struct sk_buff *skb, struct nft_stats __percpu *stats)
1134{
1135 struct nft_stats *cpu_stats, total;
1136 struct nlattr *nest;
1137 unsigned int seq;
1138 u64 pkts, bytes;
1139 int cpu;
1140
1141 memset(&total, 0, sizeof(total));
1142 for_each_possible_cpu(cpu) {
1143 cpu_stats = per_cpu_ptr(stats, cpu);
1144 do {
1145 seq = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1146 pkts = cpu_stats->pkts;
1147 bytes = cpu_stats->bytes;
1148 } while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, seq));
1149 total.pkts += pkts;
1150 total.bytes += bytes;
1151 }
1152 nest = nla_nest_start(skb, NFTA_CHAIN_COUNTERS);
1153 if (nest == NULL)
1154 goto nla_put_failure;
1155
1156 if (nla_put_be64(skb, NFTA_COUNTER_PACKETS, cpu_to_be64(total.pkts),
1157 NFTA_COUNTER_PAD) ||
1158 nla_put_be64(skb, NFTA_COUNTER_BYTES, cpu_to_be64(total.bytes),
1159 NFTA_COUNTER_PAD))
1160 goto nla_put_failure;
1161
1162 nla_nest_end(skb, nest);
1163 return 0;
1164
1165nla_put_failure:
1166 return -ENOSPC;
1167}
1168
1169static int nf_tables_fill_chain_info(struct sk_buff *skb, struct net *net,
1170 u32 portid, u32 seq, int event, u32 flags,
1171 int family, const struct nft_table *table,
1172 const struct nft_chain *chain)
1173{
1174 struct nlmsghdr *nlh;
1175 struct nfgenmsg *nfmsg;
1176
1177 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
1178 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
1179 if (nlh == NULL)
1180 goto nla_put_failure;
1181
1182 nfmsg = nlmsg_data(nlh);
1183 nfmsg->nfgen_family = family;
1184 nfmsg->version = NFNETLINK_V0;
1185 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
1186
1187 if (nla_put_string(skb, NFTA_CHAIN_TABLE, table->name))
1188 goto nla_put_failure;
1189 if (nla_put_be64(skb, NFTA_CHAIN_HANDLE, cpu_to_be64(chain->handle),
1190 NFTA_CHAIN_PAD))
1191 goto nla_put_failure;
1192 if (nla_put_string(skb, NFTA_CHAIN_NAME, chain->name))
1193 goto nla_put_failure;
1194
1195 if (nft_is_base_chain(chain)) {
1196 const struct nft_base_chain *basechain = nft_base_chain(chain);
1197 const struct nf_hook_ops *ops = &basechain->ops;
1198 struct nlattr *nest;
1199
1200 nest = nla_nest_start(skb, NFTA_CHAIN_HOOK);
1201 if (nest == NULL)
1202 goto nla_put_failure;
1203 if (nla_put_be32(skb, NFTA_HOOK_HOOKNUM, htonl(ops->hooknum)))
1204 goto nla_put_failure;
1205 if (nla_put_be32(skb, NFTA_HOOK_PRIORITY, htonl(ops->priority)))
1206 goto nla_put_failure;
1207 if (basechain->dev_name[0] &&
1208 nla_put_string(skb, NFTA_HOOK_DEV, basechain->dev_name))
1209 goto nla_put_failure;
1210 nla_nest_end(skb, nest);
1211
1212 if (nla_put_be32(skb, NFTA_CHAIN_POLICY,
1213 htonl(basechain->policy)))
1214 goto nla_put_failure;
1215
1216 if (nla_put_string(skb, NFTA_CHAIN_TYPE, basechain->type->name))
1217 goto nla_put_failure;
1218
1219 if (basechain->stats && nft_dump_stats(skb, basechain->stats))
1220 goto nla_put_failure;
1221 }
1222
1223 if (nla_put_be32(skb, NFTA_CHAIN_USE, htonl(chain->use)))
1224 goto nla_put_failure;
1225
1226 nlmsg_end(skb, nlh);
1227 return 0;
1228
1229nla_put_failure:
1230 nlmsg_trim(skb, nlh);
1231 return -1;
1232}
1233
1234static void nf_tables_chain_notify(const struct nft_ctx *ctx, int event)
1235{
1236 struct sk_buff *skb;
1237 int err;
1238
1239 if (!ctx->report &&
1240 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1241 return;
1242
1243 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1244 if (skb == NULL)
1245 goto err;
1246
1247 err = nf_tables_fill_chain_info(skb, ctx->net, ctx->portid, ctx->seq,
1248 event, 0, ctx->family, ctx->table,
1249 ctx->chain);
1250 if (err < 0) {
1251 kfree_skb(skb);
1252 goto err;
1253 }
1254
1255 nfnetlink_send(skb, ctx->net, ctx->portid, NFNLGRP_NFTABLES,
1256 ctx->report, GFP_KERNEL);
1257 return;
1258err:
1259 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1260}
1261
1262static int nf_tables_dump_chains(struct sk_buff *skb,
1263 struct netlink_callback *cb)
1264{
1265 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1266 const struct nft_table *table;
1267 const struct nft_chain *chain;
1268 unsigned int idx = 0, s_idx = cb->args[0];
1269 struct net *net = sock_net(skb->sk);
1270 int family = nfmsg->nfgen_family;
1271
1272 rcu_read_lock();
1273 cb->seq = net->nft.base_seq;
1274
1275 list_for_each_entry_rcu(table, &net->nft.tables, list) {
1276 if (family != NFPROTO_UNSPEC && family != table->family)
1277 continue;
1278
1279 list_for_each_entry_rcu(chain, &table->chains, list) {
1280 if (idx < s_idx)
1281 goto cont;
1282 if (idx > s_idx)
1283 memset(&cb->args[1], 0,
1284 sizeof(cb->args) - sizeof(cb->args[0]));
1285 if (!nft_is_active(net, chain))
1286 continue;
1287 if (nf_tables_fill_chain_info(skb, net,
1288 NETLINK_CB(cb->skb).portid,
1289 cb->nlh->nlmsg_seq,
1290 NFT_MSG_NEWCHAIN,
1291 NLM_F_MULTI,
1292 table->family, table,
1293 chain) < 0)
1294 goto done;
1295
1296 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1297cont:
1298 idx++;
1299 }
1300 }
1301done:
1302 rcu_read_unlock();
1303 cb->args[0] = idx;
1304 return skb->len;
1305}
1306
1307/* called with rcu_read_lock held */
1308static int nf_tables_getchain(struct net *net, struct sock *nlsk,
1309 struct sk_buff *skb, const struct nlmsghdr *nlh,
1310 const struct nlattr * const nla[],
1311 struct netlink_ext_ack *extack)
1312{
1313 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1314 u8 genmask = nft_genmask_cur(net);
1315 const struct nft_chain *chain;
1316 struct nft_table *table;
1317 struct sk_buff *skb2;
1318 int family = nfmsg->nfgen_family;
1319 int err;
1320
1321 if (nlh->nlmsg_flags & NLM_F_DUMP) {
1322 struct netlink_dump_control c = {
1323 .dump = nf_tables_dump_chains,
1324 .module = THIS_MODULE,
1325 };
1326
1327 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
1328 }
1329
1330 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask);
1331 if (IS_ERR(table)) {
1332 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
1333 return PTR_ERR(table);
1334 }
1335
1336 chain = nft_chain_lookup(net, table, nla[NFTA_CHAIN_NAME], genmask);
1337 if (IS_ERR(chain)) {
1338 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
1339 return PTR_ERR(chain);
1340 }
1341
1342 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1343 if (!skb2)
1344 return -ENOMEM;
1345
1346 err = nf_tables_fill_chain_info(skb2, net, NETLINK_CB(skb).portid,
1347 nlh->nlmsg_seq, NFT_MSG_NEWCHAIN, 0,
1348 family, table, chain);
1349 if (err < 0)
1350 goto err;
1351
1352 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
1353
1354err:
1355 kfree_skb(skb2);
1356 return err;
1357}
1358
1359static const struct nla_policy nft_counter_policy[NFTA_COUNTER_MAX + 1] = {
1360 [NFTA_COUNTER_PACKETS] = { .type = NLA_U64 },
1361 [NFTA_COUNTER_BYTES] = { .type = NLA_U64 },
1362};
1363
1364static struct nft_stats __percpu *nft_stats_alloc(const struct nlattr *attr)
1365{
1366 struct nlattr *tb[NFTA_COUNTER_MAX+1];
1367 struct nft_stats __percpu *newstats;
1368 struct nft_stats *stats;
1369 int err;
1370
1371 err = nla_parse_nested(tb, NFTA_COUNTER_MAX, attr, nft_counter_policy,
1372 NULL);
1373 if (err < 0)
1374 return ERR_PTR(err);
1375
1376 if (!tb[NFTA_COUNTER_BYTES] || !tb[NFTA_COUNTER_PACKETS])
1377 return ERR_PTR(-EINVAL);
1378
1379 newstats = netdev_alloc_pcpu_stats(struct nft_stats);
1380 if (newstats == NULL)
1381 return ERR_PTR(-ENOMEM);
1382
1383 /* Restore old counters on this cpu, no problem. Per-cpu statistics
1384 * are not exposed to userspace.
1385 */
1386 preempt_disable();
1387 stats = this_cpu_ptr(newstats);
1388 stats->bytes = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_BYTES]));
1389 stats->pkts = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_PACKETS]));
1390 preempt_enable();
1391
1392 return newstats;
1393}
1394
1395static void nft_chain_stats_replace(struct nft_base_chain *chain,
1396 struct nft_stats __percpu *newstats)
1397{
1398 struct nft_stats __percpu *oldstats;
1399
1400 if (newstats == NULL)
1401 return;
1402
1403 if (chain->stats) {
1404 oldstats = nfnl_dereference(chain->stats, NFNL_SUBSYS_NFTABLES);
1405 rcu_assign_pointer(chain->stats, newstats);
1406 synchronize_rcu();
1407 free_percpu(oldstats);
1408 } else {
1409 rcu_assign_pointer(chain->stats, newstats);
1410 static_branch_inc(&nft_counters_enabled);
1411 }
1412}
1413
1414static void nf_tables_chain_free_chain_rules(struct nft_chain *chain)
1415{
1416 struct nft_rule **g0 = rcu_dereference_raw(chain->rules_gen_0);
1417 struct nft_rule **g1 = rcu_dereference_raw(chain->rules_gen_1);
1418
1419 if (g0 != g1)
1420 kvfree(g1);
1421 kvfree(g0);
1422
1423 /* should be NULL either via abort or via successful commit */
1424 WARN_ON_ONCE(chain->rules_next);
1425 kvfree(chain->rules_next);
1426}
1427
1428static void nf_tables_chain_destroy(struct nft_ctx *ctx)
1429{
1430 struct nft_chain *chain = ctx->chain;
1431
1432 if (WARN_ON(chain->use > 0))
1433 return;
1434
1435 /* no concurrent access possible anymore */
1436 nf_tables_chain_free_chain_rules(chain);
1437
1438 if (nft_is_base_chain(chain)) {
1439 struct nft_base_chain *basechain = nft_base_chain(chain);
1440
1441 module_put(basechain->type->owner);
1442 free_percpu(basechain->stats);
1443 if (basechain->stats)
1444 static_branch_dec(&nft_counters_enabled);
1445 kfree(chain->name);
1446 kfree(basechain);
1447 } else {
1448 kfree(chain->name);
1449 kfree(chain);
1450 }
1451}
1452
1453struct nft_chain_hook {
1454 u32 num;
1455 s32 priority;
1456 const struct nft_chain_type *type;
1457 struct net_device *dev;
1458};
1459
1460static int nft_chain_parse_hook(struct net *net,
1461 const struct nlattr * const nla[],
1462 struct nft_chain_hook *hook, u8 family,
1463 bool autoload)
1464{
1465 struct nlattr *ha[NFTA_HOOK_MAX + 1];
1466 const struct nft_chain_type *type;
1467 struct net_device *dev;
1468 int err;
1469
1470 lockdep_assert_held(&net->nft.commit_mutex);
1471 lockdep_nfnl_nft_mutex_not_held();
1472
1473 err = nla_parse_nested(ha, NFTA_HOOK_MAX, nla[NFTA_CHAIN_HOOK],
1474 nft_hook_policy, NULL);
1475 if (err < 0)
1476 return err;
1477
1478 if (ha[NFTA_HOOK_HOOKNUM] == NULL ||
1479 ha[NFTA_HOOK_PRIORITY] == NULL)
1480 return -EINVAL;
1481
1482 hook->num = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
1483 hook->priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
1484
1485 type = chain_type[family][NFT_CHAIN_T_DEFAULT];
1486 if (nla[NFTA_CHAIN_TYPE]) {
1487 type = nf_tables_chain_type_lookup(net, nla[NFTA_CHAIN_TYPE],
1488 family, autoload);
1489 if (IS_ERR(type))
1490 return PTR_ERR(type);
1491 }
1492 if (!(type->hook_mask & (1 << hook->num)))
1493 return -EOPNOTSUPP;
1494
1495 if (type->type == NFT_CHAIN_T_NAT &&
1496 hook->priority <= NF_IP_PRI_CONNTRACK)
1497 return -EOPNOTSUPP;
1498
1499 if (!try_module_get(type->owner))
1500 return -ENOENT;
1501
1502 hook->type = type;
1503
1504 hook->dev = NULL;
1505 if (family == NFPROTO_NETDEV) {
1506 char ifname[IFNAMSIZ];
1507
1508 if (!ha[NFTA_HOOK_DEV]) {
1509 module_put(type->owner);
1510 return -EOPNOTSUPP;
1511 }
1512
1513 nla_strlcpy(ifname, ha[NFTA_HOOK_DEV], IFNAMSIZ);
1514 dev = __dev_get_by_name(net, ifname);
1515 if (!dev) {
1516 module_put(type->owner);
1517 return -ENOENT;
1518 }
1519 hook->dev = dev;
1520 } else if (ha[NFTA_HOOK_DEV]) {
1521 module_put(type->owner);
1522 return -EOPNOTSUPP;
1523 }
1524
1525 return 0;
1526}
1527
1528static void nft_chain_release_hook(struct nft_chain_hook *hook)
1529{
1530 module_put(hook->type->owner);
1531}
1532
1533struct nft_rules_old {
1534 struct rcu_head h;
1535 struct nft_rule **start;
1536};
1537
1538static struct nft_rule **nf_tables_chain_alloc_rules(const struct nft_chain *chain,
1539 unsigned int alloc)
1540{
1541 if (alloc > INT_MAX)
1542 return NULL;
1543
1544 alloc += 1; /* NULL, ends rules */
1545 if (sizeof(struct nft_rule *) > INT_MAX / alloc)
1546 return NULL;
1547
1548 alloc *= sizeof(struct nft_rule *);
1549 alloc += sizeof(struct nft_rules_old);
1550
1551 return kvmalloc(alloc, GFP_KERNEL);
1552}
1553
1554static int nf_tables_addchain(struct nft_ctx *ctx, u8 family, u8 genmask,
1555 u8 policy)
1556{
1557 const struct nlattr * const *nla = ctx->nla;
1558 struct nft_table *table = ctx->table;
1559 struct nft_base_chain *basechain;
1560 struct nft_stats __percpu *stats;
1561 struct net *net = ctx->net;
1562 struct nft_chain *chain;
1563 struct nft_rule **rules;
1564 int err;
1565
1566 if (table->use == UINT_MAX)
1567 return -EOVERFLOW;
1568
1569 if (nla[NFTA_CHAIN_HOOK]) {
1570 struct nft_chain_hook hook;
1571 struct nf_hook_ops *ops;
1572
1573 err = nft_chain_parse_hook(net, nla, &hook, family, true);
1574 if (err < 0)
1575 return err;
1576
1577 basechain = kzalloc(sizeof(*basechain), GFP_KERNEL);
1578 if (basechain == NULL) {
1579 nft_chain_release_hook(&hook);
1580 return -ENOMEM;
1581 }
1582
1583 if (hook.dev != NULL)
1584 strncpy(basechain->dev_name, hook.dev->name, IFNAMSIZ);
1585
1586 if (nla[NFTA_CHAIN_COUNTERS]) {
1587 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
1588 if (IS_ERR(stats)) {
1589 nft_chain_release_hook(&hook);
1590 kfree(basechain);
1591 return PTR_ERR(stats);
1592 }
1593 basechain->stats = stats;
1594 static_branch_inc(&nft_counters_enabled);
1595 }
1596
1597 basechain->type = hook.type;
1598 chain = &basechain->chain;
1599
1600 ops = &basechain->ops;
1601 ops->pf = family;
1602 ops->hooknum = hook.num;
1603 ops->priority = hook.priority;
1604 ops->priv = chain;
1605 ops->hook = hook.type->hooks[ops->hooknum];
1606 ops->dev = hook.dev;
1607
1608 chain->flags |= NFT_BASE_CHAIN;
1609 basechain->policy = policy;
1610 } else {
1611 chain = kzalloc(sizeof(*chain), GFP_KERNEL);
1612 if (chain == NULL)
1613 return -ENOMEM;
1614 }
1615 ctx->chain = chain;
1616
1617 INIT_LIST_HEAD(&chain->rules);
1618 chain->handle = nf_tables_alloc_handle(table);
1619 chain->table = table;
1620 chain->name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL);
1621 if (!chain->name) {
1622 err = -ENOMEM;
1623 goto err1;
1624 }
1625
1626 rules = nf_tables_chain_alloc_rules(chain, 0);
1627 if (!rules) {
1628 err = -ENOMEM;
1629 goto err1;
1630 }
1631
1632 *rules = NULL;
1633 rcu_assign_pointer(chain->rules_gen_0, rules);
1634 rcu_assign_pointer(chain->rules_gen_1, rules);
1635
1636 err = nf_tables_register_hook(net, table, chain);
1637 if (err < 0)
1638 goto err1;
1639
1640 err = rhltable_insert_key(&table->chains_ht, chain->name,
1641 &chain->rhlhead, nft_chain_ht_params);
1642 if (err)
1643 goto err2;
1644
1645 err = nft_trans_chain_add(ctx, NFT_MSG_NEWCHAIN);
1646 if (err < 0) {
1647 rhltable_remove(&table->chains_ht, &chain->rhlhead,
1648 nft_chain_ht_params);
1649 goto err2;
1650 }
1651
1652 table->use++;
1653 list_add_tail_rcu(&chain->list, &table->chains);
1654
1655 return 0;
1656err2:
1657 nf_tables_unregister_hook(net, table, chain);
1658err1:
1659 nf_tables_chain_destroy(ctx);
1660
1661 return err;
1662}
1663
1664static int nf_tables_updchain(struct nft_ctx *ctx, u8 genmask, u8 policy)
1665{
1666 const struct nlattr * const *nla = ctx->nla;
1667 struct nft_table *table = ctx->table;
1668 struct nft_chain *chain = ctx->chain;
1669 struct nft_base_chain *basechain;
1670 struct nft_stats *stats = NULL;
1671 struct nft_chain_hook hook;
1672 struct nf_hook_ops *ops;
1673 struct nft_trans *trans;
1674 int err;
1675
1676 if (nla[NFTA_CHAIN_HOOK]) {
1677 if (!nft_is_base_chain(chain))
1678 return -EBUSY;
1679
1680 err = nft_chain_parse_hook(ctx->net, nla, &hook, ctx->family,
1681 false);
1682 if (err < 0)
1683 return err;
1684
1685 basechain = nft_base_chain(chain);
1686 if (basechain->type != hook.type) {
1687 nft_chain_release_hook(&hook);
1688 return -EBUSY;
1689 }
1690
1691 ops = &basechain->ops;
1692 if (ops->hooknum != hook.num ||
1693 ops->priority != hook.priority ||
1694 ops->dev != hook.dev) {
1695 nft_chain_release_hook(&hook);
1696 return -EBUSY;
1697 }
1698 nft_chain_release_hook(&hook);
1699 }
1700
1701 if (nla[NFTA_CHAIN_HANDLE] &&
1702 nla[NFTA_CHAIN_NAME]) {
1703 struct nft_chain *chain2;
1704
1705 chain2 = nft_chain_lookup(ctx->net, table,
1706 nla[NFTA_CHAIN_NAME], genmask);
1707 if (!IS_ERR(chain2))
1708 return -EEXIST;
1709 }
1710
1711 if (nla[NFTA_CHAIN_COUNTERS]) {
1712 if (!nft_is_base_chain(chain))
1713 return -EOPNOTSUPP;
1714
1715 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
1716 if (IS_ERR(stats))
1717 return PTR_ERR(stats);
1718 }
1719
1720 err = -ENOMEM;
1721 trans = nft_trans_alloc(ctx, NFT_MSG_NEWCHAIN,
1722 sizeof(struct nft_trans_chain));
1723 if (trans == NULL)
1724 goto err;
1725
1726 nft_trans_chain_stats(trans) = stats;
1727 nft_trans_chain_update(trans) = true;
1728
1729 if (nla[NFTA_CHAIN_POLICY])
1730 nft_trans_chain_policy(trans) = policy;
1731 else
1732 nft_trans_chain_policy(trans) = -1;
1733
1734 if (nla[NFTA_CHAIN_HANDLE] &&
1735 nla[NFTA_CHAIN_NAME]) {
1736 struct nft_trans *tmp;
1737 char *name;
1738
1739 err = -ENOMEM;
1740 name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL);
1741 if (!name)
1742 goto err;
1743
1744 err = -EEXIST;
1745 list_for_each_entry(tmp, &ctx->net->nft.commit_list, list) {
1746 if (tmp->msg_type == NFT_MSG_NEWCHAIN &&
1747 tmp->ctx.table == table &&
1748 nft_trans_chain_update(tmp) &&
1749 nft_trans_chain_name(tmp) &&
1750 strcmp(name, nft_trans_chain_name(tmp)) == 0) {
1751 kfree(name);
1752 goto err;
1753 }
1754 }
1755
1756 nft_trans_chain_name(trans) = name;
1757 }
1758 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
1759
1760 return 0;
1761err:
1762 free_percpu(stats);
1763 kfree(trans);
1764 return err;
1765}
1766
1767static int nf_tables_newchain(struct net *net, struct sock *nlsk,
1768 struct sk_buff *skb, const struct nlmsghdr *nlh,
1769 const struct nlattr * const nla[],
1770 struct netlink_ext_ack *extack)
1771{
1772 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1773 u8 genmask = nft_genmask_next(net);
1774 int family = nfmsg->nfgen_family;
1775 const struct nlattr *attr;
1776 struct nft_table *table;
1777 struct nft_chain *chain;
1778 u8 policy = NF_ACCEPT;
1779 struct nft_ctx ctx;
1780 u64 handle = 0;
1781
1782 lockdep_assert_held(&net->nft.commit_mutex);
1783
1784 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask);
1785 if (IS_ERR(table)) {
1786 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
1787 return PTR_ERR(table);
1788 }
1789
1790 chain = NULL;
1791 attr = nla[NFTA_CHAIN_NAME];
1792
1793 if (nla[NFTA_CHAIN_HANDLE]) {
1794 handle = be64_to_cpu(nla_get_be64(nla[NFTA_CHAIN_HANDLE]));
1795 chain = nft_chain_lookup_byhandle(table, handle, genmask);
1796 if (IS_ERR(chain)) {
1797 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_HANDLE]);
1798 return PTR_ERR(chain);
1799 }
1800 attr = nla[NFTA_CHAIN_HANDLE];
1801 } else {
1802 chain = nft_chain_lookup(net, table, attr, genmask);
1803 if (IS_ERR(chain)) {
1804 if (PTR_ERR(chain) != -ENOENT) {
1805 NL_SET_BAD_ATTR(extack, attr);
1806 return PTR_ERR(chain);
1807 }
1808 chain = NULL;
1809 }
1810 }
1811
1812 if (nla[NFTA_CHAIN_POLICY]) {
1813 if (chain != NULL &&
1814 !nft_is_base_chain(chain)) {
1815 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
1816 return -EOPNOTSUPP;
1817 }
1818
1819 if (chain == NULL &&
1820 nla[NFTA_CHAIN_HOOK] == NULL) {
1821 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
1822 return -EOPNOTSUPP;
1823 }
1824
1825 policy = ntohl(nla_get_be32(nla[NFTA_CHAIN_POLICY]));
1826 switch (policy) {
1827 case NF_DROP:
1828 case NF_ACCEPT:
1829 break;
1830 default:
1831 return -EINVAL;
1832 }
1833 }
1834
1835 nft_ctx_init(&ctx, net, skb, nlh, family, table, chain, nla);
1836
1837 if (chain != NULL) {
1838 if (nlh->nlmsg_flags & NLM_F_EXCL) {
1839 NL_SET_BAD_ATTR(extack, attr);
1840 return -EEXIST;
1841 }
1842 if (nlh->nlmsg_flags & NLM_F_REPLACE)
1843 return -EOPNOTSUPP;
1844
1845 return nf_tables_updchain(&ctx, genmask, policy);
1846 }
1847
1848 return nf_tables_addchain(&ctx, family, genmask, policy);
1849}
1850
1851static int nf_tables_delchain(struct net *net, struct sock *nlsk,
1852 struct sk_buff *skb, const struct nlmsghdr *nlh,
1853 const struct nlattr * const nla[],
1854 struct netlink_ext_ack *extack)
1855{
1856 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1857 u8 genmask = nft_genmask_next(net);
1858 int family = nfmsg->nfgen_family;
1859 const struct nlattr *attr;
1860 struct nft_table *table;
1861 struct nft_chain *chain;
1862 struct nft_rule *rule;
1863 struct nft_ctx ctx;
1864 u64 handle;
1865 u32 use;
1866 int err;
1867
1868 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask);
1869 if (IS_ERR(table)) {
1870 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
1871 return PTR_ERR(table);
1872 }
1873
1874 if (nla[NFTA_CHAIN_HANDLE]) {
1875 attr = nla[NFTA_CHAIN_HANDLE];
1876 handle = be64_to_cpu(nla_get_be64(attr));
1877 chain = nft_chain_lookup_byhandle(table, handle, genmask);
1878 } else {
1879 attr = nla[NFTA_CHAIN_NAME];
1880 chain = nft_chain_lookup(net, table, attr, genmask);
1881 }
1882 if (IS_ERR(chain)) {
1883 NL_SET_BAD_ATTR(extack, attr);
1884 return PTR_ERR(chain);
1885 }
1886
1887 if (nlh->nlmsg_flags & NLM_F_NONREC &&
1888 chain->use > 0)
1889 return -EBUSY;
1890
1891 nft_ctx_init(&ctx, net, skb, nlh, family, table, chain, nla);
1892
1893 use = chain->use;
1894 list_for_each_entry(rule, &chain->rules, list) {
1895 if (!nft_is_active_next(net, rule))
1896 continue;
1897 use--;
1898
1899 err = nft_delrule(&ctx, rule);
1900 if (err < 0)
1901 return err;
1902 }
1903
1904 /* There are rules and elements that are still holding references to us,
1905 * we cannot do a recursive removal in this case.
1906 */
1907 if (use > 0) {
1908 NL_SET_BAD_ATTR(extack, attr);
1909 return -EBUSY;
1910 }
1911
1912 return nft_delchain(&ctx);
1913}
1914
1915/*
1916 * Expressions
1917 */
1918
1919/**
1920 * nft_register_expr - register nf_tables expr type
1921 * @ops: expr type
1922 *
1923 * Registers the expr type for use with nf_tables. Returns zero on
1924 * success or a negative errno code otherwise.
1925 */
1926int nft_register_expr(struct nft_expr_type *type)
1927{
1928 if (!nft_expr_check_ops(type->ops))
1929 return -EINVAL;
1930
1931 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1932 if (type->family == NFPROTO_UNSPEC)
1933 list_add_tail_rcu(&type->list, &nf_tables_expressions);
1934 else
1935 list_add_rcu(&type->list, &nf_tables_expressions);
1936 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1937 return 0;
1938}
1939EXPORT_SYMBOL_GPL(nft_register_expr);
1940
1941/**
1942 * nft_unregister_expr - unregister nf_tables expr type
1943 * @ops: expr type
1944 *
1945 * Unregisters the expr typefor use with nf_tables.
1946 */
1947void nft_unregister_expr(struct nft_expr_type *type)
1948{
1949 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1950 list_del_rcu(&type->list);
1951 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1952}
1953EXPORT_SYMBOL_GPL(nft_unregister_expr);
1954
1955static const struct nft_expr_type *__nft_expr_type_get(u8 family,
1956 struct nlattr *nla)
1957{
1958 const struct nft_expr_type *type;
1959
1960 list_for_each_entry(type, &nf_tables_expressions, list) {
1961 if (!nla_strcmp(nla, type->name) &&
1962 (!type->family || type->family == family))
1963 return type;
1964 }
1965 return NULL;
1966}
1967
1968static const struct nft_expr_type *nft_expr_type_get(struct net *net,
1969 u8 family,
1970 struct nlattr *nla)
1971{
1972 const struct nft_expr_type *type;
1973
1974 if (nla == NULL)
1975 return ERR_PTR(-EINVAL);
1976
1977 type = __nft_expr_type_get(family, nla);
1978 if (type != NULL && try_module_get(type->owner))
1979 return type;
1980
1981 lockdep_nfnl_nft_mutex_not_held();
1982#ifdef CONFIG_MODULES
1983 if (type == NULL) {
1984 nft_request_module(net, "nft-expr-%u-%.*s", family,
1985 nla_len(nla), (char *)nla_data(nla));
1986 if (__nft_expr_type_get(family, nla))
1987 return ERR_PTR(-EAGAIN);
1988
1989 nft_request_module(net, "nft-expr-%.*s",
1990 nla_len(nla), (char *)nla_data(nla));
1991 if (__nft_expr_type_get(family, nla))
1992 return ERR_PTR(-EAGAIN);
1993 }
1994#endif
1995 return ERR_PTR(-ENOENT);
1996}
1997
1998static const struct nla_policy nft_expr_policy[NFTA_EXPR_MAX + 1] = {
1999 [NFTA_EXPR_NAME] = { .type = NLA_STRING },
2000 [NFTA_EXPR_DATA] = { .type = NLA_NESTED },
2001};
2002
2003static int nf_tables_fill_expr_info(struct sk_buff *skb,
2004 const struct nft_expr *expr)
2005{
2006 if (nla_put_string(skb, NFTA_EXPR_NAME, expr->ops->type->name))
2007 goto nla_put_failure;
2008
2009 if (expr->ops->dump) {
2010 struct nlattr *data = nla_nest_start(skb, NFTA_EXPR_DATA);
2011 if (data == NULL)
2012 goto nla_put_failure;
2013 if (expr->ops->dump(skb, expr) < 0)
2014 goto nla_put_failure;
2015 nla_nest_end(skb, data);
2016 }
2017
2018 return skb->len;
2019
2020nla_put_failure:
2021 return -1;
2022};
2023
2024int nft_expr_dump(struct sk_buff *skb, unsigned int attr,
2025 const struct nft_expr *expr)
2026{
2027 struct nlattr *nest;
2028
2029 nest = nla_nest_start(skb, attr);
2030 if (!nest)
2031 goto nla_put_failure;
2032 if (nf_tables_fill_expr_info(skb, expr) < 0)
2033 goto nla_put_failure;
2034 nla_nest_end(skb, nest);
2035 return 0;
2036
2037nla_put_failure:
2038 return -1;
2039}
2040
2041struct nft_expr_info {
2042 const struct nft_expr_ops *ops;
2043 struct nlattr *tb[NFT_EXPR_MAXATTR + 1];
2044};
2045
2046static int nf_tables_expr_parse(const struct nft_ctx *ctx,
2047 const struct nlattr *nla,
2048 struct nft_expr_info *info)
2049{
2050 const struct nft_expr_type *type;
2051 const struct nft_expr_ops *ops;
2052 struct nlattr *tb[NFTA_EXPR_MAX + 1];
2053 int err;
2054
2055 err = nla_parse_nested(tb, NFTA_EXPR_MAX, nla, nft_expr_policy, NULL);
2056 if (err < 0)
2057 return err;
2058
2059 type = nft_expr_type_get(ctx->net, ctx->family, tb[NFTA_EXPR_NAME]);
2060 if (IS_ERR(type))
2061 return PTR_ERR(type);
2062
2063 if (tb[NFTA_EXPR_DATA]) {
2064 err = nla_parse_nested(info->tb, type->maxattr,
2065 tb[NFTA_EXPR_DATA], type->policy, NULL);
2066 if (err < 0)
2067 goto err1;
2068 } else
2069 memset(info->tb, 0, sizeof(info->tb[0]) * (type->maxattr + 1));
2070
2071 if (type->select_ops != NULL) {
2072 ops = type->select_ops(ctx,
2073 (const struct nlattr * const *)info->tb);
2074 if (IS_ERR(ops)) {
2075 err = PTR_ERR(ops);
2076 goto err1;
2077 }
2078 if (!nft_expr_check_ops(ops)) {
2079 err = -EINVAL;
2080 goto err1;
2081 }
2082 } else
2083 ops = type->ops;
2084
2085 info->ops = ops;
2086 return 0;
2087
2088err1:
2089 module_put(type->owner);
2090 return err;
2091}
2092
2093static int nf_tables_newexpr(const struct nft_ctx *ctx,
2094 const struct nft_expr_info *info,
2095 struct nft_expr *expr)
2096{
2097 const struct nft_expr_ops *ops = info->ops;
2098 int err;
2099
2100 expr->ops = ops;
2101 if (ops->init) {
2102 err = ops->init(ctx, expr, (const struct nlattr **)info->tb);
2103 if (err < 0)
2104 goto err1;
2105 }
2106
2107 return 0;
2108err1:
2109 expr->ops = NULL;
2110 return err;
2111}
2112
2113static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
2114 struct nft_expr *expr)
2115{
2116 if (expr->ops->destroy)
2117 expr->ops->destroy(ctx, expr);
2118 module_put(expr->ops->type->owner);
2119}
2120
2121struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
2122 const struct nlattr *nla)
2123{
2124 struct nft_expr_info info;
2125 struct nft_expr *expr;
2126 int err;
2127
2128 err = nf_tables_expr_parse(ctx, nla, &info);
2129 if (err < 0)
2130 goto err1;
2131
2132 err = -ENOMEM;
2133 expr = kzalloc(info.ops->size, GFP_KERNEL);
2134 if (expr == NULL)
2135 goto err2;
2136
2137 err = nf_tables_newexpr(ctx, &info, expr);
2138 if (err < 0)
2139 goto err3;
2140
2141 return expr;
2142err3:
2143 kfree(expr);
2144err2:
2145 module_put(info.ops->type->owner);
2146err1:
2147 return ERR_PTR(err);
2148}
2149
2150void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
2151{
2152 nf_tables_expr_destroy(ctx, expr);
2153 kfree(expr);
2154}
2155
2156/*
2157 * Rules
2158 */
2159
2160static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
2161 u64 handle)
2162{
2163 struct nft_rule *rule;
2164
2165 // FIXME: this sucks
2166 list_for_each_entry_rcu(rule, &chain->rules, list) {
2167 if (handle == rule->handle)
2168 return rule;
2169 }
2170
2171 return ERR_PTR(-ENOENT);
2172}
2173
2174static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
2175 const struct nlattr *nla)
2176{
2177 if (nla == NULL)
2178 return ERR_PTR(-EINVAL);
2179
2180 return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
2181}
2182
2183static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
2184 [NFTA_RULE_TABLE] = { .type = NLA_STRING,
2185 .len = NFT_TABLE_MAXNAMELEN - 1 },
2186 [NFTA_RULE_CHAIN] = { .type = NLA_STRING,
2187 .len = NFT_CHAIN_MAXNAMELEN - 1 },
2188 [NFTA_RULE_HANDLE] = { .type = NLA_U64 },
2189 [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
2190 [NFTA_RULE_COMPAT] = { .type = NLA_NESTED },
2191 [NFTA_RULE_POSITION] = { .type = NLA_U64 },
2192 [NFTA_RULE_USERDATA] = { .type = NLA_BINARY,
2193 .len = NFT_USERDATA_MAXLEN },
2194 [NFTA_RULE_ID] = { .type = NLA_U32 },
2195};
2196
2197static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
2198 u32 portid, u32 seq, int event,
2199 u32 flags, int family,
2200 const struct nft_table *table,
2201 const struct nft_chain *chain,
2202 const struct nft_rule *rule)
2203{
2204 struct nlmsghdr *nlh;
2205 struct nfgenmsg *nfmsg;
2206 const struct nft_expr *expr, *next;
2207 struct nlattr *list;
2208 const struct nft_rule *prule;
2209 u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
2210
2211 nlh = nlmsg_put(skb, portid, seq, type, sizeof(struct nfgenmsg), flags);
2212 if (nlh == NULL)
2213 goto nla_put_failure;
2214
2215 nfmsg = nlmsg_data(nlh);
2216 nfmsg->nfgen_family = family;
2217 nfmsg->version = NFNETLINK_V0;
2218 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
2219
2220 if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
2221 goto nla_put_failure;
2222 if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
2223 goto nla_put_failure;
2224 if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
2225 NFTA_RULE_PAD))
2226 goto nla_put_failure;
2227
2228 if ((event != NFT_MSG_DELRULE) && (rule->list.prev != &chain->rules)) {
2229 prule = list_prev_entry(rule, list);
2230 if (nla_put_be64(skb, NFTA_RULE_POSITION,
2231 cpu_to_be64(prule->handle),
2232 NFTA_RULE_PAD))
2233 goto nla_put_failure;
2234 }
2235
2236 list = nla_nest_start(skb, NFTA_RULE_EXPRESSIONS);
2237 if (list == NULL)
2238 goto nla_put_failure;
2239 nft_rule_for_each_expr(expr, next, rule) {
2240 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
2241 goto nla_put_failure;
2242 }
2243 nla_nest_end(skb, list);
2244
2245 if (rule->udata) {
2246 struct nft_userdata *udata = nft_userdata(rule);
2247 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
2248 udata->data) < 0)
2249 goto nla_put_failure;
2250 }
2251
2252 nlmsg_end(skb, nlh);
2253 return 0;
2254
2255nla_put_failure:
2256 nlmsg_trim(skb, nlh);
2257 return -1;
2258}
2259
2260static void nf_tables_rule_notify(const struct nft_ctx *ctx,
2261 const struct nft_rule *rule, int event)
2262{
2263 struct sk_buff *skb;
2264 int err;
2265
2266 if (!ctx->report &&
2267 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
2268 return;
2269
2270 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
2271 if (skb == NULL)
2272 goto err;
2273
2274 err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
2275 event, 0, ctx->family, ctx->table,
2276 ctx->chain, rule);
2277 if (err < 0) {
2278 kfree_skb(skb);
2279 goto err;
2280 }
2281
2282 nfnetlink_send(skb, ctx->net, ctx->portid, NFNLGRP_NFTABLES,
2283 ctx->report, GFP_KERNEL);
2284 return;
2285err:
2286 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
2287}
2288
2289struct nft_rule_dump_ctx {
2290 char *table;
2291 char *chain;
2292};
2293
2294static int nf_tables_dump_rules(struct sk_buff *skb,
2295 struct netlink_callback *cb)
2296{
2297 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2298 const struct nft_rule_dump_ctx *ctx = cb->data;
2299 const struct nft_table *table;
2300 const struct nft_chain *chain;
2301 const struct nft_rule *rule;
2302 unsigned int idx = 0, s_idx = cb->args[0];
2303 struct net *net = sock_net(skb->sk);
2304 int family = nfmsg->nfgen_family;
2305
2306 rcu_read_lock();
2307 cb->seq = net->nft.base_seq;
2308
2309 list_for_each_entry_rcu(table, &net->nft.tables, list) {
2310 if (family != NFPROTO_UNSPEC && family != table->family)
2311 continue;
2312
2313 if (ctx && ctx->table && strcmp(ctx->table, table->name) != 0)
2314 continue;
2315
2316 list_for_each_entry_rcu(chain, &table->chains, list) {
2317 if (ctx && ctx->chain &&
2318 strcmp(ctx->chain, chain->name) != 0)
2319 continue;
2320
2321 list_for_each_entry_rcu(rule, &chain->rules, list) {
2322 if (!nft_is_active(net, rule))
2323 goto cont;
2324 if (idx < s_idx)
2325 goto cont;
2326 if (idx > s_idx)
2327 memset(&cb->args[1], 0,
2328 sizeof(cb->args) - sizeof(cb->args[0]));
2329 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
2330 cb->nlh->nlmsg_seq,
2331 NFT_MSG_NEWRULE,
2332 NLM_F_MULTI | NLM_F_APPEND,
2333 table->family,
2334 table, chain, rule) < 0)
2335 goto done;
2336
2337 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
2338cont:
2339 idx++;
2340 }
2341 }
2342 }
2343done:
2344 rcu_read_unlock();
2345
2346 cb->args[0] = idx;
2347 return skb->len;
2348}
2349
2350static int nf_tables_dump_rules_start(struct netlink_callback *cb)
2351{
2352 const struct nlattr * const *nla = cb->data;
2353 struct nft_rule_dump_ctx *ctx = NULL;
2354
2355 if (nla[NFTA_RULE_TABLE] || nla[NFTA_RULE_CHAIN]) {
2356 ctx = kzalloc(sizeof(*ctx), GFP_ATOMIC);
2357 if (!ctx)
2358 return -ENOMEM;
2359
2360 if (nla[NFTA_RULE_TABLE]) {
2361 ctx->table = nla_strdup(nla[NFTA_RULE_TABLE],
2362 GFP_ATOMIC);
2363 if (!ctx->table) {
2364 kfree(ctx);
2365 return -ENOMEM;
2366 }
2367 }
2368 if (nla[NFTA_RULE_CHAIN]) {
2369 ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN],
2370 GFP_ATOMIC);
2371 if (!ctx->chain) {
2372 kfree(ctx->table);
2373 kfree(ctx);
2374 return -ENOMEM;
2375 }
2376 }
2377 }
2378
2379 cb->data = ctx;
2380 return 0;
2381}
2382
2383static int nf_tables_dump_rules_done(struct netlink_callback *cb)
2384{
2385 struct nft_rule_dump_ctx *ctx = cb->data;
2386
2387 if (ctx) {
2388 kfree(ctx->table);
2389 kfree(ctx->chain);
2390 kfree(ctx);
2391 }
2392 return 0;
2393}
2394
2395/* called with rcu_read_lock held */
2396static int nf_tables_getrule(struct net *net, struct sock *nlsk,
2397 struct sk_buff *skb, const struct nlmsghdr *nlh,
2398 const struct nlattr * const nla[],
2399 struct netlink_ext_ack *extack)
2400{
2401 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2402 u8 genmask = nft_genmask_cur(net);
2403 const struct nft_chain *chain;
2404 const struct nft_rule *rule;
2405 struct nft_table *table;
2406 struct sk_buff *skb2;
2407 int family = nfmsg->nfgen_family;
2408 int err;
2409
2410 if (nlh->nlmsg_flags & NLM_F_DUMP) {
2411 struct netlink_dump_control c = {
2412 .start= nf_tables_dump_rules_start,
2413 .dump = nf_tables_dump_rules,
2414 .done = nf_tables_dump_rules_done,
2415 .module = THIS_MODULE,
2416 .data = (void *)nla,
2417 };
2418
2419 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
2420 }
2421
2422 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask);
2423 if (IS_ERR(table)) {
2424 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
2425 return PTR_ERR(table);
2426 }
2427
2428 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
2429 if (IS_ERR(chain)) {
2430 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
2431 return PTR_ERR(chain);
2432 }
2433
2434 rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
2435 if (IS_ERR(rule)) {
2436 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
2437 return PTR_ERR(rule);
2438 }
2439
2440 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
2441 if (!skb2)
2442 return -ENOMEM;
2443
2444 err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
2445 nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
2446 family, table, chain, rule);
2447 if (err < 0)
2448 goto err;
2449
2450 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
2451
2452err:
2453 kfree_skb(skb2);
2454 return err;
2455}
2456
2457static void nf_tables_rule_destroy(const struct nft_ctx *ctx,
2458 struct nft_rule *rule)
2459{
2460 struct nft_expr *expr;
2461
2462 /*
2463 * Careful: some expressions might not be initialized in case this
2464 * is called on error from nf_tables_newrule().
2465 */
2466 expr = nft_expr_first(rule);
2467 while (expr != nft_expr_last(rule) && expr->ops) {
2468 nf_tables_expr_destroy(ctx, expr);
2469 expr = nft_expr_next(expr);
2470 }
2471 kfree(rule);
2472}
2473
2474static void nf_tables_rule_release(const struct nft_ctx *ctx,
2475 struct nft_rule *rule)
2476{
2477 nft_rule_expr_deactivate(ctx, rule);
2478 nf_tables_rule_destroy(ctx, rule);
2479}
2480
2481int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
2482{
2483 struct nft_expr *expr, *last;
2484 const struct nft_data *data;
2485 struct nft_rule *rule;
2486 int err;
2487
2488 if (ctx->level == NFT_JUMP_STACK_SIZE)
2489 return -EMLINK;
2490
2491 list_for_each_entry(rule, &chain->rules, list) {
2492 if (!nft_is_active_next(ctx->net, rule))
2493 continue;
2494
2495 nft_rule_for_each_expr(expr, last, rule) {
2496 if (!expr->ops->validate)
2497 continue;
2498
2499 err = expr->ops->validate(ctx, expr, &data);
2500 if (err < 0)
2501 return err;
2502 }
2503 }
2504
2505 return 0;
2506}
2507EXPORT_SYMBOL_GPL(nft_chain_validate);
2508
2509static int nft_table_validate(struct net *net, const struct nft_table *table)
2510{
2511 struct nft_chain *chain;
2512 struct nft_ctx ctx = {
2513 .net = net,
2514 .family = table->family,
2515 };
2516 int err;
2517
2518 list_for_each_entry(chain, &table->chains, list) {
2519 if (!nft_is_base_chain(chain))
2520 continue;
2521
2522 ctx.chain = chain;
2523 err = nft_chain_validate(&ctx, chain);
2524 if (err < 0)
2525 return err;
2526 }
2527
2528 return 0;
2529}
2530
2531#define NFT_RULE_MAXEXPRS 128
2532
2533static int nf_tables_newrule(struct net *net, struct sock *nlsk,
2534 struct sk_buff *skb, const struct nlmsghdr *nlh,
2535 const struct nlattr * const nla[],
2536 struct netlink_ext_ack *extack)
2537{
2538 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2539 u8 genmask = nft_genmask_next(net);
2540 struct nft_expr_info *info = NULL;
2541 int family = nfmsg->nfgen_family;
2542 struct nft_table *table;
2543 struct nft_chain *chain;
2544 struct nft_rule *rule, *old_rule = NULL;
2545 struct nft_userdata *udata;
2546 struct nft_trans *trans = NULL;
2547 struct nft_expr *expr;
2548 struct nft_ctx ctx;
2549 struct nlattr *tmp;
2550 unsigned int size, i, n, ulen = 0, usize = 0;
2551 int err, rem;
2552 u64 handle, pos_handle;
2553
2554 lockdep_assert_held(&net->nft.commit_mutex);
2555
2556 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask);
2557 if (IS_ERR(table)) {
2558 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
2559 return PTR_ERR(table);
2560 }
2561
2562 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
2563 if (IS_ERR(chain)) {
2564 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
2565 return PTR_ERR(chain);
2566 }
2567
2568 if (nla[NFTA_RULE_HANDLE]) {
2569 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
2570 rule = __nft_rule_lookup(chain, handle);
2571 if (IS_ERR(rule)) {
2572 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
2573 return PTR_ERR(rule);
2574 }
2575
2576 if (nlh->nlmsg_flags & NLM_F_EXCL) {
2577 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
2578 return -EEXIST;
2579 }
2580 if (nlh->nlmsg_flags & NLM_F_REPLACE)
2581 old_rule = rule;
2582 else
2583 return -EOPNOTSUPP;
2584 } else {
2585 if (!(nlh->nlmsg_flags & NLM_F_CREATE) ||
2586 nlh->nlmsg_flags & NLM_F_REPLACE)
2587 return -EINVAL;
2588 handle = nf_tables_alloc_handle(table);
2589
2590 if (chain->use == UINT_MAX)
2591 return -EOVERFLOW;
2592 }
2593
2594 if (nla[NFTA_RULE_POSITION]) {
2595 if (!(nlh->nlmsg_flags & NLM_F_CREATE))
2596 return -EOPNOTSUPP;
2597
2598 pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
2599 old_rule = __nft_rule_lookup(chain, pos_handle);
2600 if (IS_ERR(old_rule)) {
2601 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
2602 return PTR_ERR(old_rule);
2603 }
2604 }
2605
2606 nft_ctx_init(&ctx, net, skb, nlh, family, table, chain, nla);
2607
2608 n = 0;
2609 size = 0;
2610 if (nla[NFTA_RULE_EXPRESSIONS]) {
2611 info = kvmalloc_array(NFT_RULE_MAXEXPRS,
2612 sizeof(struct nft_expr_info),
2613 GFP_KERNEL);
2614 if (!info)
2615 return -ENOMEM;
2616
2617 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
2618 err = -EINVAL;
2619 if (nla_type(tmp) != NFTA_LIST_ELEM)
2620 goto err1;
2621 if (n == NFT_RULE_MAXEXPRS)
2622 goto err1;
2623 err = nf_tables_expr_parse(&ctx, tmp, &info[n]);
2624 if (err < 0)
2625 goto err1;
2626 size += info[n].ops->size;
2627 n++;
2628 }
2629 }
2630 /* Check for overflow of dlen field */
2631 err = -EFBIG;
2632 if (size >= 1 << 12)
2633 goto err1;
2634
2635 if (nla[NFTA_RULE_USERDATA]) {
2636 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
2637 if (ulen > 0)
2638 usize = sizeof(struct nft_userdata) + ulen;
2639 }
2640
2641 err = -ENOMEM;
2642 rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL);
2643 if (rule == NULL)
2644 goto err1;
2645
2646 nft_activate_next(net, rule);
2647
2648 rule->handle = handle;
2649 rule->dlen = size;
2650 rule->udata = ulen ? 1 : 0;
2651
2652 if (ulen) {
2653 udata = nft_userdata(rule);
2654 udata->len = ulen - 1;
2655 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
2656 }
2657
2658 expr = nft_expr_first(rule);
2659 for (i = 0; i < n; i++) {
2660 err = nf_tables_newexpr(&ctx, &info[i], expr);
2661 if (err < 0)
2662 goto err2;
2663
2664 if (info[i].ops->validate)
2665 nft_validate_state_update(net, NFT_VALIDATE_NEED);
2666
2667 info[i].ops = NULL;
2668 expr = nft_expr_next(expr);
2669 }
2670
2671 if (nlh->nlmsg_flags & NLM_F_REPLACE) {
2672 if (!nft_is_active_next(net, old_rule)) {
2673 err = -ENOENT;
2674 goto err2;
2675 }
2676 trans = nft_trans_rule_add(&ctx, NFT_MSG_DELRULE,
2677 old_rule);
2678 if (trans == NULL) {
2679 err = -ENOMEM;
2680 goto err2;
2681 }
2682 nft_deactivate_next(net, old_rule);
2683 chain->use--;
2684
2685 if (nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule) == NULL) {
2686 err = -ENOMEM;
2687 goto err2;
2688 }
2689
2690 list_add_tail_rcu(&rule->list, &old_rule->list);
2691 } else {
2692 if (nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule) == NULL) {
2693 err = -ENOMEM;
2694 goto err2;
2695 }
2696
2697 if (nlh->nlmsg_flags & NLM_F_APPEND) {
2698 if (old_rule)
2699 list_add_rcu(&rule->list, &old_rule->list);
2700 else
2701 list_add_tail_rcu(&rule->list, &chain->rules);
2702 } else {
2703 if (old_rule)
2704 list_add_tail_rcu(&rule->list, &old_rule->list);
2705 else
2706 list_add_rcu(&rule->list, &chain->rules);
2707 }
2708 }
2709 kvfree(info);
2710 chain->use++;
2711
2712 if (net->nft.validate_state == NFT_VALIDATE_DO)
2713 return nft_table_validate(net, table);
2714
2715 return 0;
2716err2:
2717 nf_tables_rule_release(&ctx, rule);
2718err1:
2719 for (i = 0; i < n; i++) {
2720 if (info[i].ops != NULL)
2721 module_put(info[i].ops->type->owner);
2722 }
2723 kvfree(info);
2724 return err;
2725}
2726
2727static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
2728 const struct nlattr *nla)
2729{
2730 u32 id = ntohl(nla_get_be32(nla));
2731 struct nft_trans *trans;
2732
2733 list_for_each_entry(trans, &net->nft.commit_list, list) {
2734 struct nft_rule *rule = nft_trans_rule(trans);
2735
2736 if (trans->msg_type == NFT_MSG_NEWRULE &&
2737 id == nft_trans_rule_id(trans))
2738 return rule;
2739 }
2740 return ERR_PTR(-ENOENT);
2741}
2742
2743static int nf_tables_delrule(struct net *net, struct sock *nlsk,
2744 struct sk_buff *skb, const struct nlmsghdr *nlh,
2745 const struct nlattr * const nla[],
2746 struct netlink_ext_ack *extack)
2747{
2748 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2749 u8 genmask = nft_genmask_next(net);
2750 struct nft_table *table;
2751 struct nft_chain *chain = NULL;
2752 struct nft_rule *rule;
2753 int family = nfmsg->nfgen_family, err = 0;
2754 struct nft_ctx ctx;
2755
2756 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask);
2757 if (IS_ERR(table)) {
2758 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
2759 return PTR_ERR(table);
2760 }
2761
2762 if (nla[NFTA_RULE_CHAIN]) {
2763 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
2764 genmask);
2765 if (IS_ERR(chain)) {
2766 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
2767 return PTR_ERR(chain);
2768 }
2769 }
2770
2771 nft_ctx_init(&ctx, net, skb, nlh, family, table, chain, nla);
2772
2773 if (chain) {
2774 if (nla[NFTA_RULE_HANDLE]) {
2775 rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
2776 if (IS_ERR(rule)) {
2777 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
2778 return PTR_ERR(rule);
2779 }
2780
2781 err = nft_delrule(&ctx, rule);
2782 } else if (nla[NFTA_RULE_ID]) {
2783 rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_ID]);
2784 if (IS_ERR(rule)) {
2785 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
2786 return PTR_ERR(rule);
2787 }
2788
2789 err = nft_delrule(&ctx, rule);
2790 } else {
2791 err = nft_delrule_by_chain(&ctx);
2792 }
2793 } else {
2794 list_for_each_entry(chain, &table->chains, list) {
2795 if (!nft_is_active_next(net, chain))
2796 continue;
2797
2798 ctx.chain = chain;
2799 err = nft_delrule_by_chain(&ctx);
2800 if (err < 0)
2801 break;
2802 }
2803 }
2804
2805 return err;
2806}
2807
2808/*
2809 * Sets
2810 */
2811
2812static LIST_HEAD(nf_tables_set_types);
2813
2814int nft_register_set(struct nft_set_type *type)
2815{
2816 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2817 list_add_tail_rcu(&type->list, &nf_tables_set_types);
2818 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2819 return 0;
2820}
2821EXPORT_SYMBOL_GPL(nft_register_set);
2822
2823void nft_unregister_set(struct nft_set_type *type)
2824{
2825 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2826 list_del_rcu(&type->list);
2827 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2828}
2829EXPORT_SYMBOL_GPL(nft_unregister_set);
2830
2831#define NFT_SET_FEATURES (NFT_SET_INTERVAL | NFT_SET_MAP | \
2832 NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
2833 NFT_SET_EVAL)
2834
2835static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
2836{
2837 return (flags & type->features) == (flags & NFT_SET_FEATURES);
2838}
2839
2840/*
2841 * Select a set implementation based on the data characteristics and the
2842 * given policy. The total memory use might not be known if no size is
2843 * given, in that case the amount of memory per element is used.
2844 */
2845static const struct nft_set_ops *
2846nft_select_set_ops(const struct nft_ctx *ctx,
2847 const struct nlattr * const nla[],
2848 const struct nft_set_desc *desc,
2849 enum nft_set_policies policy)
2850{
2851 const struct nft_set_ops *ops, *bops;
2852 struct nft_set_estimate est, best;
2853 const struct nft_set_type *type;
2854 u32 flags = 0;
2855
2856 lockdep_assert_held(&ctx->net->nft.commit_mutex);
2857 lockdep_nfnl_nft_mutex_not_held();
2858#ifdef CONFIG_MODULES
2859 if (list_empty(&nf_tables_set_types)) {
2860 nft_request_module(ctx->net, "nft-set");
2861 if (!list_empty(&nf_tables_set_types))
2862 return ERR_PTR(-EAGAIN);
2863 }
2864#endif
2865 if (nla[NFTA_SET_FLAGS] != NULL)
2866 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
2867
2868 bops = NULL;
2869 best.size = ~0;
2870 best.lookup = ~0;
2871 best.space = ~0;
2872
2873 list_for_each_entry(type, &nf_tables_set_types, list) {
2874 ops = &type->ops;
2875
2876 if (!nft_set_ops_candidate(type, flags))
2877 continue;
2878 if (!ops->estimate(desc, flags, &est))
2879 continue;
2880
2881 switch (policy) {
2882 case NFT_SET_POL_PERFORMANCE:
2883 if (est.lookup < best.lookup)
2884 break;
2885 if (est.lookup == best.lookup &&
2886 est.space < best.space)
2887 break;
2888 continue;
2889 case NFT_SET_POL_MEMORY:
2890 if (!desc->size) {
2891 if (est.space < best.space)
2892 break;
2893 if (est.space == best.space &&
2894 est.lookup < best.lookup)
2895 break;
2896 } else if (est.size < best.size || !bops) {
2897 break;
2898 }
2899 continue;
2900 default:
2901 break;
2902 }
2903
2904 if (!try_module_get(type->owner))
2905 continue;
2906 if (bops != NULL)
2907 module_put(to_set_type(bops)->owner);
2908
2909 bops = ops;
2910 best = est;
2911 }
2912
2913 if (bops != NULL)
2914 return bops;
2915
2916 return ERR_PTR(-EOPNOTSUPP);
2917}
2918
2919static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
2920 [NFTA_SET_TABLE] = { .type = NLA_STRING,
2921 .len = NFT_TABLE_MAXNAMELEN - 1 },
2922 [NFTA_SET_NAME] = { .type = NLA_STRING,
2923 .len = NFT_SET_MAXNAMELEN - 1 },
2924 [NFTA_SET_FLAGS] = { .type = NLA_U32 },
2925 [NFTA_SET_KEY_TYPE] = { .type = NLA_U32 },
2926 [NFTA_SET_KEY_LEN] = { .type = NLA_U32 },
2927 [NFTA_SET_DATA_TYPE] = { .type = NLA_U32 },
2928 [NFTA_SET_DATA_LEN] = { .type = NLA_U32 },
2929 [NFTA_SET_POLICY] = { .type = NLA_U32 },
2930 [NFTA_SET_DESC] = { .type = NLA_NESTED },
2931 [NFTA_SET_ID] = { .type = NLA_U32 },
2932 [NFTA_SET_TIMEOUT] = { .type = NLA_U64 },
2933 [NFTA_SET_GC_INTERVAL] = { .type = NLA_U32 },
2934 [NFTA_SET_USERDATA] = { .type = NLA_BINARY,
2935 .len = NFT_USERDATA_MAXLEN },
2936 [NFTA_SET_OBJ_TYPE] = { .type = NLA_U32 },
2937 [NFTA_SET_HANDLE] = { .type = NLA_U64 },
2938};
2939
2940static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
2941 [NFTA_SET_DESC_SIZE] = { .type = NLA_U32 },
2942};
2943
2944static int nft_ctx_init_from_setattr(struct nft_ctx *ctx, struct net *net,
2945 const struct sk_buff *skb,
2946 const struct nlmsghdr *nlh,
2947 const struct nlattr * const nla[],
2948 struct netlink_ext_ack *extack,
2949 u8 genmask)
2950{
2951 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2952 int family = nfmsg->nfgen_family;
2953 struct nft_table *table = NULL;
2954
2955 if (nla[NFTA_SET_TABLE] != NULL) {
2956 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
2957 genmask);
2958 if (IS_ERR(table)) {
2959 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
2960 return PTR_ERR(table);
2961 }
2962 }
2963
2964 nft_ctx_init(ctx, net, skb, nlh, family, table, NULL, nla);
2965 return 0;
2966}
2967
2968static struct nft_set *nft_set_lookup(const struct nft_table *table,
2969 const struct nlattr *nla, u8 genmask)
2970{
2971 struct nft_set *set;
2972
2973 if (nla == NULL)
2974 return ERR_PTR(-EINVAL);
2975
2976 list_for_each_entry_rcu(set, &table->sets, list) {
2977 if (!nla_strcmp(nla, set->name) &&
2978 nft_active_genmask(set, genmask))
2979 return set;
2980 }
2981 return ERR_PTR(-ENOENT);
2982}
2983
2984static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
2985 const struct nlattr *nla,
2986 u8 genmask)
2987{
2988 struct nft_set *set;
2989
2990 list_for_each_entry(set, &table->sets, list) {
2991 if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
2992 nft_active_genmask(set, genmask))
2993 return set;
2994 }
2995 return ERR_PTR(-ENOENT);
2996}
2997
2998static struct nft_set *nft_set_lookup_byid(const struct net *net,
2999 const struct nlattr *nla, u8 genmask)
3000{
3001 struct nft_trans *trans;
3002 u32 id = ntohl(nla_get_be32(nla));
3003
3004 list_for_each_entry(trans, &net->nft.commit_list, list) {
3005 if (trans->msg_type == NFT_MSG_NEWSET) {
3006 struct nft_set *set = nft_trans_set(trans);
3007
3008 if (id == nft_trans_set_id(trans) &&
3009 nft_active_genmask(set, genmask))
3010 return set;
3011 }
3012 }
3013 return ERR_PTR(-ENOENT);
3014}
3015
3016struct nft_set *nft_set_lookup_global(const struct net *net,
3017 const struct nft_table *table,
3018 const struct nlattr *nla_set_name,
3019 const struct nlattr *nla_set_id,
3020 u8 genmask)
3021{
3022 struct nft_set *set;
3023
3024 set = nft_set_lookup(table, nla_set_name, genmask);
3025 if (IS_ERR(set)) {
3026 if (!nla_set_id)
3027 return set;
3028
3029 set = nft_set_lookup_byid(net, nla_set_id, genmask);
3030 }
3031 return set;
3032}
3033EXPORT_SYMBOL_GPL(nft_set_lookup_global);
3034
3035static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
3036 const char *name)
3037{
3038 const struct nft_set *i;
3039 const char *p;
3040 unsigned long *inuse;
3041 unsigned int n = 0, min = 0;
3042
3043 p = strchr(name, '%');
3044 if (p != NULL) {
3045 if (p[1] != 'd' || strchr(p + 2, '%'))
3046 return -EINVAL;
3047
3048 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
3049 if (inuse == NULL)
3050 return -ENOMEM;
3051cont:
3052 list_for_each_entry(i, &ctx->table->sets, list) {
3053 int tmp;
3054
3055 if (!nft_is_active_next(ctx->net, set))
3056 continue;
3057 if (!sscanf(i->name, name, &tmp))
3058 continue;
3059 if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
3060 continue;
3061
3062 set_bit(tmp - min, inuse);
3063 }
3064
3065 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
3066 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
3067 min += BITS_PER_BYTE * PAGE_SIZE;
3068 memset(inuse, 0, PAGE_SIZE);
3069 goto cont;
3070 }
3071 free_page((unsigned long)inuse);
3072 }
3073
3074 set->name = kasprintf(GFP_KERNEL, name, min + n);
3075 if (!set->name)
3076 return -ENOMEM;
3077
3078 list_for_each_entry(i, &ctx->table->sets, list) {
3079 if (!nft_is_active_next(ctx->net, i))
3080 continue;
3081 if (!strcmp(set->name, i->name)) {
3082 kfree(set->name);
3083 return -ENFILE;
3084 }
3085 }
3086 return 0;
3087}
3088
3089static int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
3090{
3091 u64 ms = be64_to_cpu(nla_get_be64(nla));
3092 u64 max = (u64)(~((u64)0));
3093
3094 max = div_u64(max, NSEC_PER_MSEC);
3095 if (ms >= max)
3096 return -ERANGE;
3097
3098 ms *= NSEC_PER_MSEC;
3099 *result = nsecs_to_jiffies64(ms);
3100 return 0;
3101}
3102
3103static __be64 nf_jiffies64_to_msecs(u64 input)
3104{
3105 u64 ms = jiffies64_to_nsecs(input);
3106
3107 return cpu_to_be64(div_u64(ms, NSEC_PER_MSEC));
3108}
3109
3110static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
3111 const struct nft_set *set, u16 event, u16 flags)
3112{
3113 struct nfgenmsg *nfmsg;
3114 struct nlmsghdr *nlh;
3115 struct nlattr *desc;
3116 u32 portid = ctx->portid;
3117 u32 seq = ctx->seq;
3118
3119 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3120 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
3121 flags);
3122 if (nlh == NULL)
3123 goto nla_put_failure;
3124
3125 nfmsg = nlmsg_data(nlh);
3126 nfmsg->nfgen_family = ctx->family;
3127 nfmsg->version = NFNETLINK_V0;
3128 nfmsg->res_id = htons(ctx->net->nft.base_seq & 0xffff);
3129
3130 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
3131 goto nla_put_failure;
3132 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
3133 goto nla_put_failure;
3134 if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
3135 NFTA_SET_PAD))
3136 goto nla_put_failure;
3137 if (set->flags != 0)
3138 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
3139 goto nla_put_failure;
3140
3141 if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
3142 goto nla_put_failure;
3143 if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
3144 goto nla_put_failure;
3145 if (set->flags & NFT_SET_MAP) {
3146 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
3147 goto nla_put_failure;
3148 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
3149 goto nla_put_failure;
3150 }
3151 if (set->flags & NFT_SET_OBJECT &&
3152 nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
3153 goto nla_put_failure;
3154
3155 if (set->timeout &&
3156 nla_put_be64(skb, NFTA_SET_TIMEOUT,
3157 nf_jiffies64_to_msecs(set->timeout),
3158 NFTA_SET_PAD))
3159 goto nla_put_failure;
3160 if (set->gc_int &&
3161 nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(set->gc_int)))
3162 goto nla_put_failure;
3163
3164 if (set->policy != NFT_SET_POL_PERFORMANCE) {
3165 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
3166 goto nla_put_failure;
3167 }
3168
3169 if (nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
3170 goto nla_put_failure;
3171
3172 desc = nla_nest_start(skb, NFTA_SET_DESC);
3173 if (desc == NULL)
3174 goto nla_put_failure;
3175 if (set->size &&
3176 nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
3177 goto nla_put_failure;
3178 nla_nest_end(skb, desc);
3179
3180 nlmsg_end(skb, nlh);
3181 return 0;
3182
3183nla_put_failure:
3184 nlmsg_trim(skb, nlh);
3185 return -1;
3186}
3187
3188static void nf_tables_set_notify(const struct nft_ctx *ctx,
3189 const struct nft_set *set, int event,
3190 gfp_t gfp_flags)
3191{
3192 struct sk_buff *skb;
3193 u32 portid = ctx->portid;
3194 int err;
3195
3196 if (!ctx->report &&
3197 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3198 return;
3199
3200 skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
3201 if (skb == NULL)
3202 goto err;
3203
3204 err = nf_tables_fill_set(skb, ctx, set, event, 0);
3205 if (err < 0) {
3206 kfree_skb(skb);
3207 goto err;
3208 }
3209
3210 nfnetlink_send(skb, ctx->net, portid, NFNLGRP_NFTABLES, ctx->report,
3211 gfp_flags);
3212 return;
3213err:
3214 nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
3215}
3216
3217static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
3218{
3219 const struct nft_set *set;
3220 unsigned int idx, s_idx = cb->args[0];
3221 struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
3222 struct net *net = sock_net(skb->sk);
3223 struct nft_ctx *ctx = cb->data, ctx_set;
3224
3225 if (cb->args[1])
3226 return skb->len;
3227
3228 rcu_read_lock();
3229 cb->seq = net->nft.base_seq;
3230
3231 list_for_each_entry_rcu(table, &net->nft.tables, list) {
3232 if (ctx->family != NFPROTO_UNSPEC &&
3233 ctx->family != table->family)
3234 continue;
3235
3236 if (ctx->table && ctx->table != table)
3237 continue;
3238
3239 if (cur_table) {
3240 if (cur_table != table)
3241 continue;
3242
3243 cur_table = NULL;
3244 }
3245 idx = 0;
3246 list_for_each_entry_rcu(set, &table->sets, list) {
3247 if (idx < s_idx)
3248 goto cont;
3249 if (!nft_is_active(net, set))
3250 goto cont;
3251
3252 ctx_set = *ctx;
3253 ctx_set.table = table;
3254 ctx_set.family = table->family;
3255
3256 if (nf_tables_fill_set(skb, &ctx_set, set,
3257 NFT_MSG_NEWSET,
3258 NLM_F_MULTI) < 0) {
3259 cb->args[0] = idx;
3260 cb->args[2] = (unsigned long) table;
3261 goto done;
3262 }
3263 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
3264cont:
3265 idx++;
3266 }
3267 if (s_idx)
3268 s_idx = 0;
3269 }
3270 cb->args[1] = 1;
3271done:
3272 rcu_read_unlock();
3273 return skb->len;
3274}
3275
3276static int nf_tables_dump_sets_start(struct netlink_callback *cb)
3277{
3278 struct nft_ctx *ctx_dump = NULL;
3279
3280 ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
3281 if (ctx_dump == NULL)
3282 return -ENOMEM;
3283
3284 cb->data = ctx_dump;
3285 return 0;
3286}
3287
3288static int nf_tables_dump_sets_done(struct netlink_callback *cb)
3289{
3290 kfree(cb->data);
3291 return 0;
3292}
3293
3294/* called with rcu_read_lock held */
3295static int nf_tables_getset(struct net *net, struct sock *nlsk,
3296 struct sk_buff *skb, const struct nlmsghdr *nlh,
3297 const struct nlattr * const nla[],
3298 struct netlink_ext_ack *extack)
3299{
3300 u8 genmask = nft_genmask_cur(net);
3301 const struct nft_set *set;
3302 struct nft_ctx ctx;
3303 struct sk_buff *skb2;
3304 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3305 int err;
3306
3307 /* Verify existence before starting dump */
3308 err = nft_ctx_init_from_setattr(&ctx, net, skb, nlh, nla, extack,
3309 genmask);
3310 if (err < 0)
3311 return err;
3312
3313 if (nlh->nlmsg_flags & NLM_F_DUMP) {
3314 struct netlink_dump_control c = {
3315 .start = nf_tables_dump_sets_start,
3316 .dump = nf_tables_dump_sets,
3317 .done = nf_tables_dump_sets_done,
3318 .data = &ctx,
3319 .module = THIS_MODULE,
3320 };
3321
3322 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
3323 }
3324
3325 /* Only accept unspec with dump */
3326 if (nfmsg->nfgen_family == NFPROTO_UNSPEC)
3327 return -EAFNOSUPPORT;
3328 if (!nla[NFTA_SET_TABLE])
3329 return -EINVAL;
3330
3331 set = nft_set_lookup(ctx.table, nla[NFTA_SET_NAME], genmask);
3332 if (IS_ERR(set))
3333 return PTR_ERR(set);
3334
3335 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3336 if (skb2 == NULL)
3337 return -ENOMEM;
3338
3339 err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
3340 if (err < 0)
3341 goto err;
3342
3343 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
3344
3345err:
3346 kfree_skb(skb2);
3347 return err;
3348}
3349
3350static int nf_tables_set_desc_parse(const struct nft_ctx *ctx,
3351 struct nft_set_desc *desc,
3352 const struct nlattr *nla)
3353{
3354 struct nlattr *da[NFTA_SET_DESC_MAX + 1];
3355 int err;
3356
3357 err = nla_parse_nested(da, NFTA_SET_DESC_MAX, nla,
3358 nft_set_desc_policy, NULL);
3359 if (err < 0)
3360 return err;
3361
3362 if (da[NFTA_SET_DESC_SIZE] != NULL)
3363 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
3364
3365 return 0;
3366}
3367
3368static int nf_tables_newset(struct net *net, struct sock *nlsk,
3369 struct sk_buff *skb, const struct nlmsghdr *nlh,
3370 const struct nlattr * const nla[],
3371 struct netlink_ext_ack *extack)
3372{
3373 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3374 u8 genmask = nft_genmask_next(net);
3375 int family = nfmsg->nfgen_family;
3376 const struct nft_set_ops *ops;
3377 struct nft_table *table;
3378 struct nft_set *set;
3379 struct nft_ctx ctx;
3380 char *name;
3381 u64 size;
3382 u64 timeout;
3383 u32 ktype, dtype, flags, policy, gc_int, objtype;
3384 struct nft_set_desc desc;
3385 unsigned char *udata;
3386 u16 udlen;
3387 int err;
3388
3389 if (nla[NFTA_SET_TABLE] == NULL ||
3390 nla[NFTA_SET_NAME] == NULL ||
3391 nla[NFTA_SET_KEY_LEN] == NULL ||
3392 nla[NFTA_SET_ID] == NULL)
3393 return -EINVAL;
3394
3395 memset(&desc, 0, sizeof(desc));
3396
3397 ktype = NFT_DATA_VALUE;
3398 if (nla[NFTA_SET_KEY_TYPE] != NULL) {
3399 ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
3400 if ((ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
3401 return -EINVAL;
3402 }
3403
3404 desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
3405 if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
3406 return -EINVAL;
3407
3408 flags = 0;
3409 if (nla[NFTA_SET_FLAGS] != NULL) {
3410 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
3411 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
3412 NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
3413 NFT_SET_MAP | NFT_SET_EVAL |
3414 NFT_SET_OBJECT))
3415 return -EINVAL;
3416 /* Only one of these operations is supported */
3417 if ((flags & (NFT_SET_MAP | NFT_SET_EVAL | NFT_SET_OBJECT)) ==
3418 (NFT_SET_MAP | NFT_SET_EVAL | NFT_SET_OBJECT))
3419 return -EOPNOTSUPP;
3420 }
3421
3422 dtype = 0;
3423 if (nla[NFTA_SET_DATA_TYPE] != NULL) {
3424 if (!(flags & NFT_SET_MAP))
3425 return -EINVAL;
3426
3427 dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
3428 if ((dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
3429 dtype != NFT_DATA_VERDICT)
3430 return -EINVAL;
3431
3432 if (dtype != NFT_DATA_VERDICT) {
3433 if (nla[NFTA_SET_DATA_LEN] == NULL)
3434 return -EINVAL;
3435 desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
3436 if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
3437 return -EINVAL;
3438 } else
3439 desc.dlen = sizeof(struct nft_verdict);
3440 } else if (flags & NFT_SET_MAP)
3441 return -EINVAL;
3442
3443 if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
3444 if (!(flags & NFT_SET_OBJECT))
3445 return -EINVAL;
3446
3447 objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
3448 if (objtype == NFT_OBJECT_UNSPEC ||
3449 objtype > NFT_OBJECT_MAX)
3450 return -EINVAL;
3451 } else if (flags & NFT_SET_OBJECT)
3452 return -EINVAL;
3453 else
3454 objtype = NFT_OBJECT_UNSPEC;
3455
3456 timeout = 0;
3457 if (nla[NFTA_SET_TIMEOUT] != NULL) {
3458 if (!(flags & NFT_SET_TIMEOUT))
3459 return -EINVAL;
3460
3461 err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &timeout);
3462 if (err)
3463 return err;
3464 }
3465 gc_int = 0;
3466 if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
3467 if (!(flags & NFT_SET_TIMEOUT))
3468 return -EINVAL;
3469 gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
3470 }
3471
3472 policy = NFT_SET_POL_PERFORMANCE;
3473 if (nla[NFTA_SET_POLICY] != NULL)
3474 policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
3475
3476 if (nla[NFTA_SET_DESC] != NULL) {
3477 err = nf_tables_set_desc_parse(&ctx, &desc, nla[NFTA_SET_DESC]);
3478 if (err < 0)
3479 return err;
3480 }
3481
3482 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask);
3483 if (IS_ERR(table)) {
3484 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
3485 return PTR_ERR(table);
3486 }
3487
3488 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
3489
3490 set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
3491 if (IS_ERR(set)) {
3492 if (PTR_ERR(set) != -ENOENT) {
3493 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
3494 return PTR_ERR(set);
3495 }
3496 } else {
3497 if (nlh->nlmsg_flags & NLM_F_EXCL) {
3498 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
3499 return -EEXIST;
3500 }
3501 if (nlh->nlmsg_flags & NLM_F_REPLACE)
3502 return -EOPNOTSUPP;
3503
3504 return 0;
3505 }
3506
3507 if (!(nlh->nlmsg_flags & NLM_F_CREATE))
3508 return -ENOENT;
3509
3510 ops = nft_select_set_ops(&ctx, nla, &desc, policy);
3511 if (IS_ERR(ops))
3512 return PTR_ERR(ops);
3513
3514 udlen = 0;
3515 if (nla[NFTA_SET_USERDATA])
3516 udlen = nla_len(nla[NFTA_SET_USERDATA]);
3517
3518 size = 0;
3519 if (ops->privsize != NULL)
3520 size = ops->privsize(nla, &desc);
3521
3522 set = kvzalloc(sizeof(*set) + size + udlen, GFP_KERNEL);
3523 if (!set) {
3524 err = -ENOMEM;
3525 goto err1;
3526 }
3527
3528 name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL);
3529 if (!name) {
3530 err = -ENOMEM;
3531 goto err2;
3532 }
3533
3534 err = nf_tables_set_alloc_name(&ctx, set, name);
3535 kfree(name);
3536 if (err < 0)
3537 goto err2;
3538
3539 udata = NULL;
3540 if (udlen) {
3541 udata = set->data + size;
3542 nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
3543 }
3544
3545 INIT_LIST_HEAD(&set->bindings);
3546 set->table = table;
3547 write_pnet(&set->net, net);
3548 set->ops = ops;
3549 set->ktype = ktype;
3550 set->klen = desc.klen;
3551 set->dtype = dtype;
3552 set->objtype = objtype;
3553 set->dlen = desc.dlen;
3554 set->flags = flags;
3555 set->size = desc.size;
3556 set->policy = policy;
3557 set->udlen = udlen;
3558 set->udata = udata;
3559 set->timeout = timeout;
3560 set->gc_int = gc_int;
3561 set->handle = nf_tables_alloc_handle(table);
3562
3563 err = ops->init(set, &desc, nla);
3564 if (err < 0)
3565 goto err3;
3566
3567 err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
3568 if (err < 0)
3569 goto err4;
3570
3571 list_add_tail_rcu(&set->list, &table->sets);
3572 table->use++;
3573 return 0;
3574
3575err4:
3576 ops->destroy(set);
3577err3:
3578 kfree(set->name);
3579err2:
3580 kvfree(set);
3581err1:
3582 module_put(to_set_type(ops)->owner);
3583 return err;
3584}
3585
3586static void nft_set_destroy(struct nft_set *set)
3587{
3588 set->ops->destroy(set);
3589 module_put(to_set_type(set->ops)->owner);
3590 kfree(set->name);
3591 kvfree(set);
3592}
3593
3594static int nf_tables_delset(struct net *net, struct sock *nlsk,
3595 struct sk_buff *skb, const struct nlmsghdr *nlh,
3596 const struct nlattr * const nla[],
3597 struct netlink_ext_ack *extack)
3598{
3599 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3600 u8 genmask = nft_genmask_next(net);
3601 const struct nlattr *attr;
3602 struct nft_set *set;
3603 struct nft_ctx ctx;
3604 int err;
3605
3606 if (nfmsg->nfgen_family == NFPROTO_UNSPEC)
3607 return -EAFNOSUPPORT;
3608 if (nla[NFTA_SET_TABLE] == NULL)
3609 return -EINVAL;
3610
3611 err = nft_ctx_init_from_setattr(&ctx, net, skb, nlh, nla, extack,
3612 genmask);
3613 if (err < 0)
3614 return err;
3615
3616 if (nla[NFTA_SET_HANDLE]) {
3617 attr = nla[NFTA_SET_HANDLE];
3618 set = nft_set_lookup_byhandle(ctx.table, attr, genmask);
3619 } else {
3620 attr = nla[NFTA_SET_NAME];
3621 set = nft_set_lookup(ctx.table, attr, genmask);
3622 }
3623
3624 if (IS_ERR(set)) {
3625 NL_SET_BAD_ATTR(extack, attr);
3626 return PTR_ERR(set);
3627 }
3628 if (!list_empty(&set->bindings) ||
3629 (nlh->nlmsg_flags & NLM_F_NONREC && atomic_read(&set->nelems) > 0)) {
3630 NL_SET_BAD_ATTR(extack, attr);
3631 return -EBUSY;
3632 }
3633
3634 return nft_delset(&ctx, set);
3635}
3636
3637static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
3638 struct nft_set *set,
3639 const struct nft_set_iter *iter,
3640 struct nft_set_elem *elem)
3641{
3642 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
3643 enum nft_registers dreg;
3644
3645 dreg = nft_type_to_reg(set->dtype);
3646 return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
3647 set->dtype == NFT_DATA_VERDICT ?
3648 NFT_DATA_VERDICT : NFT_DATA_VALUE,
3649 set->dlen);
3650}
3651
3652int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
3653 struct nft_set_binding *binding)
3654{
3655 struct nft_set_binding *i;
3656 struct nft_set_iter iter;
3657
3658 if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
3659 return -EBUSY;
3660
3661 if (binding->flags & NFT_SET_MAP) {
3662 /* If the set is already bound to the same chain all
3663 * jumps are already validated for that chain.
3664 */
3665 list_for_each_entry(i, &set->bindings, list) {
3666 if (i->flags & NFT_SET_MAP &&
3667 i->chain == binding->chain)
3668 goto bind;
3669 }
3670
3671 iter.genmask = nft_genmask_next(ctx->net);
3672 iter.skip = 0;
3673 iter.count = 0;
3674 iter.err = 0;
3675 iter.fn = nf_tables_bind_check_setelem;
3676
3677 set->ops->walk(ctx, set, &iter);
3678 if (iter.err < 0)
3679 return iter.err;
3680 }
3681bind:
3682 binding->chain = ctx->chain;
3683 list_add_tail_rcu(&binding->list, &set->bindings);
3684 return 0;
3685}
3686EXPORT_SYMBOL_GPL(nf_tables_bind_set);
3687
3688void nf_tables_rebind_set(const struct nft_ctx *ctx, struct nft_set *set,
3689 struct nft_set_binding *binding)
3690{
3691 if (list_empty(&set->bindings) && nft_set_is_anonymous(set) &&
3692 nft_is_active(ctx->net, set))
3693 list_add_tail_rcu(&set->list, &ctx->table->sets);
3694
3695 list_add_tail_rcu(&binding->list, &set->bindings);
3696}
3697EXPORT_SYMBOL_GPL(nf_tables_rebind_set);
3698
3699void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
3700 struct nft_set_binding *binding)
3701{
3702 list_del_rcu(&binding->list);
3703
3704 if (list_empty(&set->bindings) && nft_set_is_anonymous(set) &&
3705 nft_is_active(ctx->net, set))
3706 list_del_rcu(&set->list);
3707}
3708EXPORT_SYMBOL_GPL(nf_tables_unbind_set);
3709
3710void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
3711{
3712 if (list_empty(&set->bindings) && nft_set_is_anonymous(set) &&
3713 nft_is_active(ctx->net, set)) {
3714 nf_tables_set_notify(ctx, set, NFT_MSG_DELSET, GFP_ATOMIC);
3715 nft_set_destroy(set);
3716 }
3717}
3718EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
3719
3720const struct nft_set_ext_type nft_set_ext_types[] = {
3721 [NFT_SET_EXT_KEY] = {
3722 .align = __alignof__(u32),
3723 },
3724 [NFT_SET_EXT_DATA] = {
3725 .align = __alignof__(u32),
3726 },
3727 [NFT_SET_EXT_EXPR] = {
3728 .align = __alignof__(struct nft_expr),
3729 },
3730 [NFT_SET_EXT_OBJREF] = {
3731 .len = sizeof(struct nft_object *),
3732 .align = __alignof__(struct nft_object *),
3733 },
3734 [NFT_SET_EXT_FLAGS] = {
3735 .len = sizeof(u8),
3736 .align = __alignof__(u8),
3737 },
3738 [NFT_SET_EXT_TIMEOUT] = {
3739 .len = sizeof(u64),
3740 .align = __alignof__(u64),
3741 },
3742 [NFT_SET_EXT_EXPIRATION] = {
3743 .len = sizeof(u64),
3744 .align = __alignof__(u64),
3745 },
3746 [NFT_SET_EXT_USERDATA] = {
3747 .len = sizeof(struct nft_userdata),
3748 .align = __alignof__(struct nft_userdata),
3749 },
3750};
3751EXPORT_SYMBOL_GPL(nft_set_ext_types);
3752
3753/*
3754 * Set elements
3755 */
3756
3757static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
3758 [NFTA_SET_ELEM_KEY] = { .type = NLA_NESTED },
3759 [NFTA_SET_ELEM_DATA] = { .type = NLA_NESTED },
3760 [NFTA_SET_ELEM_FLAGS] = { .type = NLA_U32 },
3761 [NFTA_SET_ELEM_TIMEOUT] = { .type = NLA_U64 },
3762 [NFTA_SET_ELEM_USERDATA] = { .type = NLA_BINARY,
3763 .len = NFT_USERDATA_MAXLEN },
3764 [NFTA_SET_ELEM_EXPR] = { .type = NLA_NESTED },
3765 [NFTA_SET_ELEM_OBJREF] = { .type = NLA_STRING },
3766};
3767
3768static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
3769 [NFTA_SET_ELEM_LIST_TABLE] = { .type = NLA_STRING,
3770 .len = NFT_TABLE_MAXNAMELEN - 1 },
3771 [NFTA_SET_ELEM_LIST_SET] = { .type = NLA_STRING,
3772 .len = NFT_SET_MAXNAMELEN - 1 },
3773 [NFTA_SET_ELEM_LIST_ELEMENTS] = { .type = NLA_NESTED },
3774 [NFTA_SET_ELEM_LIST_SET_ID] = { .type = NLA_U32 },
3775};
3776
3777static int nft_ctx_init_from_elemattr(struct nft_ctx *ctx, struct net *net,
3778 const struct sk_buff *skb,
3779 const struct nlmsghdr *nlh,
3780 const struct nlattr * const nla[],
3781 struct netlink_ext_ack *extack,
3782 u8 genmask)
3783{
3784 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3785 int family = nfmsg->nfgen_family;
3786 struct nft_table *table;
3787
3788 table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
3789 genmask);
3790 if (IS_ERR(table)) {
3791 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
3792 return PTR_ERR(table);
3793 }
3794
3795 nft_ctx_init(ctx, net, skb, nlh, family, table, NULL, nla);
3796 return 0;
3797}
3798
3799static int nf_tables_fill_setelem(struct sk_buff *skb,
3800 const struct nft_set *set,
3801 const struct nft_set_elem *elem)
3802{
3803 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
3804 unsigned char *b = skb_tail_pointer(skb);
3805 struct nlattr *nest;
3806
3807 nest = nla_nest_start(skb, NFTA_LIST_ELEM);
3808 if (nest == NULL)
3809 goto nla_put_failure;
3810
3811 if (nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
3812 NFT_DATA_VALUE, set->klen) < 0)
3813 goto nla_put_failure;
3814
3815 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
3816 nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
3817 set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
3818 set->dlen) < 0)
3819 goto nla_put_failure;
3820
3821 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPR) &&
3822 nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, nft_set_ext_expr(ext)) < 0)
3823 goto nla_put_failure;
3824
3825 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
3826 nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
3827 (*nft_set_ext_obj(ext))->name) < 0)
3828 goto nla_put_failure;
3829
3830 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
3831 nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
3832 htonl(*nft_set_ext_flags(ext))))
3833 goto nla_put_failure;
3834
3835 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
3836 nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
3837 nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
3838 NFTA_SET_ELEM_PAD))
3839 goto nla_put_failure;
3840
3841 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
3842 u64 expires, now = get_jiffies_64();
3843
3844 expires = *nft_set_ext_expiration(ext);
3845 if (time_before64(now, expires))
3846 expires -= now;
3847 else
3848 expires = 0;
3849
3850 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
3851 nf_jiffies64_to_msecs(expires),
3852 NFTA_SET_ELEM_PAD))
3853 goto nla_put_failure;
3854 }
3855
3856 if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
3857 struct nft_userdata *udata;
3858
3859 udata = nft_set_ext_userdata(ext);
3860 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
3861 udata->len + 1, udata->data))
3862 goto nla_put_failure;
3863 }
3864
3865 nla_nest_end(skb, nest);
3866 return 0;
3867
3868nla_put_failure:
3869 nlmsg_trim(skb, b);
3870 return -EMSGSIZE;
3871}
3872
3873struct nft_set_dump_args {
3874 const struct netlink_callback *cb;
3875 struct nft_set_iter iter;
3876 struct sk_buff *skb;
3877};
3878
3879static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
3880 struct nft_set *set,
3881 const struct nft_set_iter *iter,
3882 struct nft_set_elem *elem)
3883{
3884 struct nft_set_dump_args *args;
3885
3886 args = container_of(iter, struct nft_set_dump_args, iter);
3887 return nf_tables_fill_setelem(args->skb, set, elem);
3888}
3889
3890struct nft_set_dump_ctx {
3891 const struct nft_set *set;
3892 struct nft_ctx ctx;
3893};
3894
3895static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
3896{
3897 struct nft_set_dump_ctx *dump_ctx = cb->data;
3898 struct net *net = sock_net(skb->sk);
3899 struct nft_table *table;
3900 struct nft_set *set;
3901 struct nft_set_dump_args args;
3902 bool set_found = false;
3903 struct nfgenmsg *nfmsg;
3904 struct nlmsghdr *nlh;
3905 struct nlattr *nest;
3906 u32 portid, seq;
3907 int event;
3908
3909 rcu_read_lock();
3910 list_for_each_entry_rcu(table, &net->nft.tables, list) {
3911 if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
3912 dump_ctx->ctx.family != table->family)
3913 continue;
3914
3915 if (table != dump_ctx->ctx.table)
3916 continue;
3917
3918 list_for_each_entry_rcu(set, &table->sets, list) {
3919 if (set == dump_ctx->set) {
3920 set_found = true;
3921 break;
3922 }
3923 }
3924 break;
3925 }
3926
3927 if (!set_found) {
3928 rcu_read_unlock();
3929 return -ENOENT;
3930 }
3931
3932 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
3933 portid = NETLINK_CB(cb->skb).portid;
3934 seq = cb->nlh->nlmsg_seq;
3935
3936 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
3937 NLM_F_MULTI);
3938 if (nlh == NULL)
3939 goto nla_put_failure;
3940
3941 nfmsg = nlmsg_data(nlh);
3942 nfmsg->nfgen_family = table->family;
3943 nfmsg->version = NFNETLINK_V0;
3944 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
3945
3946 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
3947 goto nla_put_failure;
3948 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
3949 goto nla_put_failure;
3950
3951 nest = nla_nest_start(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
3952 if (nest == NULL)
3953 goto nla_put_failure;
3954
3955 args.cb = cb;
3956 args.skb = skb;
3957 args.iter.genmask = nft_genmask_cur(net);
3958 args.iter.skip = cb->args[0];
3959 args.iter.count = 0;
3960 args.iter.err = 0;
3961 args.iter.fn = nf_tables_dump_setelem;
3962 set->ops->walk(&dump_ctx->ctx, set, &args.iter);
3963 rcu_read_unlock();
3964
3965 nla_nest_end(skb, nest);
3966 nlmsg_end(skb, nlh);
3967
3968 if (args.iter.err && args.iter.err != -EMSGSIZE)
3969 return args.iter.err;
3970 if (args.iter.count == cb->args[0])
3971 return 0;
3972
3973 cb->args[0] = args.iter.count;
3974 return skb->len;
3975
3976nla_put_failure:
3977 rcu_read_unlock();
3978 return -ENOSPC;
3979}
3980
3981static int nf_tables_dump_set_start(struct netlink_callback *cb)
3982{
3983 struct nft_set_dump_ctx *dump_ctx = cb->data;
3984
3985 cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
3986
3987 return cb->data ? 0 : -ENOMEM;
3988}
3989
3990static int nf_tables_dump_set_done(struct netlink_callback *cb)
3991{
3992 kfree(cb->data);
3993 return 0;
3994}
3995
3996static int nf_tables_fill_setelem_info(struct sk_buff *skb,
3997 const struct nft_ctx *ctx, u32 seq,
3998 u32 portid, int event, u16 flags,
3999 const struct nft_set *set,
4000 const struct nft_set_elem *elem)
4001{
4002 struct nfgenmsg *nfmsg;
4003 struct nlmsghdr *nlh;
4004 struct nlattr *nest;
4005 int err;
4006
4007 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
4008 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
4009 flags);
4010 if (nlh == NULL)
4011 goto nla_put_failure;
4012
4013 nfmsg = nlmsg_data(nlh);
4014 nfmsg->nfgen_family = ctx->family;
4015 nfmsg->version = NFNETLINK_V0;
4016 nfmsg->res_id = htons(ctx->net->nft.base_seq & 0xffff);
4017
4018 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
4019 goto nla_put_failure;
4020 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
4021 goto nla_put_failure;
4022
4023 nest = nla_nest_start(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
4024 if (nest == NULL)
4025 goto nla_put_failure;
4026
4027 err = nf_tables_fill_setelem(skb, set, elem);
4028 if (err < 0)
4029 goto nla_put_failure;
4030
4031 nla_nest_end(skb, nest);
4032
4033 nlmsg_end(skb, nlh);
4034 return 0;
4035
4036nla_put_failure:
4037 nlmsg_trim(skb, nlh);
4038 return -1;
4039}
4040
4041static int nft_setelem_parse_flags(const struct nft_set *set,
4042 const struct nlattr *attr, u32 *flags)
4043{
4044 if (attr == NULL)
4045 return 0;
4046
4047 *flags = ntohl(nla_get_be32(attr));
4048 if (*flags & ~NFT_SET_ELEM_INTERVAL_END)
4049 return -EINVAL;
4050 if (!(set->flags & NFT_SET_INTERVAL) &&
4051 *flags & NFT_SET_ELEM_INTERVAL_END)
4052 return -EINVAL;
4053
4054 return 0;
4055}
4056
4057static int nft_get_set_elem(struct nft_ctx *ctx, struct nft_set *set,
4058 const struct nlattr *attr)
4059{
4060 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
4061 struct nft_data_desc desc;
4062 struct nft_set_elem elem;
4063 struct sk_buff *skb;
4064 uint32_t flags = 0;
4065 void *priv;
4066 int err;
4067
4068 err = nla_parse_nested(nla, NFTA_SET_ELEM_MAX, attr,
4069 nft_set_elem_policy, NULL);
4070 if (err < 0)
4071 return err;
4072
4073 if (!nla[NFTA_SET_ELEM_KEY])
4074 return -EINVAL;
4075
4076 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
4077 if (err < 0)
4078 return err;
4079
4080 err = nft_data_init(ctx, &elem.key.val, sizeof(elem.key), &desc,
4081 nla[NFTA_SET_ELEM_KEY]);
4082 if (err < 0)
4083 return err;
4084
4085 err = -EINVAL;
4086 if (desc.type != NFT_DATA_VALUE || desc.len != set->klen)
4087 return err;
4088
4089 priv = set->ops->get(ctx->net, set, &elem, flags);
4090 if (IS_ERR(priv))
4091 return PTR_ERR(priv);
4092
4093 elem.priv = priv;
4094
4095 err = -ENOMEM;
4096 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
4097 if (skb == NULL)
4098 goto err1;
4099
4100 err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
4101 NFT_MSG_NEWSETELEM, 0, set, &elem);
4102 if (err < 0)
4103 goto err2;
4104
4105 err = nfnetlink_unicast(skb, ctx->net, ctx->portid, MSG_DONTWAIT);
4106 /* This avoids a loop in nfnetlink. */
4107 if (err < 0)
4108 goto err1;
4109
4110 return 0;
4111err2:
4112 kfree_skb(skb);
4113err1:
4114 /* this avoids a loop in nfnetlink. */
4115 return err == -EAGAIN ? -ENOBUFS : err;
4116}
4117
4118/* called with rcu_read_lock held */
4119static int nf_tables_getsetelem(struct net *net, struct sock *nlsk,
4120 struct sk_buff *skb, const struct nlmsghdr *nlh,
4121 const struct nlattr * const nla[],
4122 struct netlink_ext_ack *extack)
4123{
4124 u8 genmask = nft_genmask_cur(net);
4125 struct nft_set *set;
4126 struct nlattr *attr;
4127 struct nft_ctx ctx;
4128 int rem, err = 0;
4129
4130 err = nft_ctx_init_from_elemattr(&ctx, net, skb, nlh, nla, extack,
4131 genmask);
4132 if (err < 0)
4133 return err;
4134
4135 set = nft_set_lookup(ctx.table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
4136 if (IS_ERR(set))
4137 return PTR_ERR(set);
4138
4139 if (nlh->nlmsg_flags & NLM_F_DUMP) {
4140 struct netlink_dump_control c = {
4141 .start = nf_tables_dump_set_start,
4142 .dump = nf_tables_dump_set,
4143 .done = nf_tables_dump_set_done,
4144 .module = THIS_MODULE,
4145 };
4146 struct nft_set_dump_ctx dump_ctx = {
4147 .set = set,
4148 .ctx = ctx,
4149 };
4150
4151 c.data = &dump_ctx;
4152 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
4153 }
4154
4155 if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
4156 return -EINVAL;
4157
4158 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
4159 err = nft_get_set_elem(&ctx, set, attr);
4160 if (err < 0)
4161 break;
4162 }
4163
4164 return err;
4165}
4166
4167static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
4168 const struct nft_set *set,
4169 const struct nft_set_elem *elem,
4170 int event, u16 flags)
4171{
4172 struct net *net = ctx->net;
4173 u32 portid = ctx->portid;
4174 struct sk_buff *skb;
4175 int err;
4176
4177 if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
4178 return;
4179
4180 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
4181 if (skb == NULL)
4182 goto err;
4183
4184 err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
4185 set, elem);
4186 if (err < 0) {
4187 kfree_skb(skb);
4188 goto err;
4189 }
4190
4191 nfnetlink_send(skb, net, portid, NFNLGRP_NFTABLES, ctx->report,
4192 GFP_KERNEL);
4193 return;
4194err:
4195 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
4196}
4197
4198static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
4199 int msg_type,
4200 struct nft_set *set)
4201{
4202 struct nft_trans *trans;
4203
4204 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
4205 if (trans == NULL)
4206 return NULL;
4207
4208 nft_trans_elem_set(trans) = set;
4209 return trans;
4210}
4211
4212void *nft_set_elem_init(const struct nft_set *set,
4213 const struct nft_set_ext_tmpl *tmpl,
4214 const u32 *key, const u32 *data,
4215 u64 timeout, gfp_t gfp)
4216{
4217 struct nft_set_ext *ext;
4218 void *elem;
4219
4220 elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
4221 if (elem == NULL)
4222 return NULL;
4223
4224 ext = nft_set_elem_ext(set, elem);
4225 nft_set_ext_init(ext, tmpl);
4226
4227 memcpy(nft_set_ext_key(ext), key, set->klen);
4228 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
4229 memcpy(nft_set_ext_data(ext), data, set->dlen);
4230 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION))
4231 *nft_set_ext_expiration(ext) =
4232 get_jiffies_64() + timeout;
4233 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
4234 *nft_set_ext_timeout(ext) = timeout;
4235
4236 return elem;
4237}
4238
4239void nft_set_elem_destroy(const struct nft_set *set, void *elem,
4240 bool destroy_expr)
4241{
4242 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
4243 struct nft_ctx ctx = {
4244 .net = read_pnet(&set->net),
4245 .family = set->table->family,
4246 };
4247
4248 nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
4249 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
4250 nft_data_release(nft_set_ext_data(ext), set->dtype);
4251 if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPR)) {
4252 struct nft_expr *expr = nft_set_ext_expr(ext);
4253
4254 if (expr->ops->destroy_clone) {
4255 expr->ops->destroy_clone(&ctx, expr);
4256 module_put(expr->ops->type->owner);
4257 } else {
4258 nf_tables_expr_destroy(&ctx, expr);
4259 }
4260 }
4261 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
4262 (*nft_set_ext_obj(ext))->use--;
4263 kfree(elem);
4264}
4265EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
4266
4267/* Only called from commit path, nft_set_elem_deactivate() already deals with
4268 * the refcounting from the preparation phase.
4269 */
4270static void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
4271 const struct nft_set *set, void *elem)
4272{
4273 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
4274
4275 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPR))
4276 nf_tables_expr_destroy(ctx, nft_set_ext_expr(ext));
4277 kfree(elem);
4278}
4279
4280static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
4281 const struct nlattr *attr, u32 nlmsg_flags)
4282{
4283 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
4284 u8 genmask = nft_genmask_next(ctx->net);
4285 struct nft_data_desc d1, d2;
4286 struct nft_set_ext_tmpl tmpl;
4287 struct nft_set_ext *ext, *ext2;
4288 struct nft_set_elem elem;
4289 struct nft_set_binding *binding;
4290 struct nft_object *obj = NULL;
4291 struct nft_userdata *udata;
4292 struct nft_data data;
4293 enum nft_registers dreg;
4294 struct nft_trans *trans;
4295 u32 flags = 0;
4296 u64 timeout;
4297 u8 ulen;
4298 int err;
4299
4300 err = nla_parse_nested(nla, NFTA_SET_ELEM_MAX, attr,
4301 nft_set_elem_policy, NULL);
4302 if (err < 0)
4303 return err;
4304
4305 if (nla[NFTA_SET_ELEM_KEY] == NULL)
4306 return -EINVAL;
4307
4308 nft_set_ext_prepare(&tmpl);
4309
4310 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
4311 if (err < 0)
4312 return err;
4313 if (flags != 0)
4314 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
4315
4316 if (set->flags & NFT_SET_MAP) {
4317 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
4318 !(flags & NFT_SET_ELEM_INTERVAL_END))
4319 return -EINVAL;
4320 if (nla[NFTA_SET_ELEM_DATA] != NULL &&
4321 flags & NFT_SET_ELEM_INTERVAL_END)
4322 return -EINVAL;
4323 } else {
4324 if (nla[NFTA_SET_ELEM_DATA] != NULL)
4325 return -EINVAL;
4326 }
4327
4328 timeout = 0;
4329 if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
4330 if (!(set->flags & NFT_SET_TIMEOUT))
4331 return -EINVAL;
4332 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
4333 &timeout);
4334 if (err)
4335 return err;
4336 } else if (set->flags & NFT_SET_TIMEOUT) {
4337 timeout = set->timeout;
4338 }
4339
4340 err = nft_data_init(ctx, &elem.key.val, sizeof(elem.key), &d1,
4341 nla[NFTA_SET_ELEM_KEY]);
4342 if (err < 0)
4343 goto err1;
4344 err = -EINVAL;
4345 if (d1.type != NFT_DATA_VALUE || d1.len != set->klen)
4346 goto err2;
4347
4348 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, d1.len);
4349 if (timeout > 0) {
4350 nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
4351 if (timeout != set->timeout)
4352 nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
4353 }
4354
4355 if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
4356 if (!(set->flags & NFT_SET_OBJECT)) {
4357 err = -EINVAL;
4358 goto err2;
4359 }
4360 obj = nft_obj_lookup(ctx->table, nla[NFTA_SET_ELEM_OBJREF],
4361 set->objtype, genmask);
4362 if (IS_ERR(obj)) {
4363 err = PTR_ERR(obj);
4364 goto err2;
4365 }
4366 nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
4367 }
4368
4369 if (nla[NFTA_SET_ELEM_DATA] != NULL) {
4370 err = nft_data_init(ctx, &data, sizeof(data), &d2,
4371 nla[NFTA_SET_ELEM_DATA]);
4372 if (err < 0)
4373 goto err2;
4374
4375 err = -EINVAL;
4376 if (set->dtype != NFT_DATA_VERDICT && d2.len != set->dlen)
4377 goto err3;
4378
4379 dreg = nft_type_to_reg(set->dtype);
4380 list_for_each_entry(binding, &set->bindings, list) {
4381 struct nft_ctx bind_ctx = {
4382 .net = ctx->net,
4383 .family = ctx->family,
4384 .table = ctx->table,
4385 .chain = (struct nft_chain *)binding->chain,
4386 };
4387
4388 if (!(binding->flags & NFT_SET_MAP))
4389 continue;
4390
4391 err = nft_validate_register_store(&bind_ctx, dreg,
4392 &data,
4393 d2.type, d2.len);
4394 if (err < 0)
4395 goto err3;
4396
4397 if (d2.type == NFT_DATA_VERDICT &&
4398 (data.verdict.code == NFT_GOTO ||
4399 data.verdict.code == NFT_JUMP))
4400 nft_validate_state_update(ctx->net,
4401 NFT_VALIDATE_NEED);
4402 }
4403
4404 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, d2.len);
4405 }
4406
4407 /* The full maximum length of userdata can exceed the maximum
4408 * offset value (U8_MAX) for following extensions, therefor it
4409 * must be the last extension added.
4410 */
4411 ulen = 0;
4412 if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
4413 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
4414 if (ulen > 0)
4415 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
4416 ulen);
4417 }
4418
4419 err = -ENOMEM;
4420 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data, data.data,
4421 timeout, GFP_KERNEL);
4422 if (elem.priv == NULL)
4423 goto err3;
4424
4425 ext = nft_set_elem_ext(set, elem.priv);
4426 if (flags)
4427 *nft_set_ext_flags(ext) = flags;
4428 if (ulen > 0) {
4429 udata = nft_set_ext_userdata(ext);
4430 udata->len = ulen - 1;
4431 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
4432 }
4433 if (obj) {
4434 *nft_set_ext_obj(ext) = obj;
4435 obj->use++;
4436 }
4437
4438 trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
4439 if (trans == NULL)
4440 goto err4;
4441
4442 ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
4443 err = set->ops->insert(ctx->net, set, &elem, &ext2);
4444 if (err) {
4445 if (err == -EEXIST) {
4446 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
4447 nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
4448 nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
4449 nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF)) {
4450 err = -EBUSY;
4451 goto err5;
4452 }
4453 if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
4454 nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
4455 memcmp(nft_set_ext_data(ext),
4456 nft_set_ext_data(ext2), set->dlen) != 0) ||
4457 (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
4458 nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
4459 *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
4460 err = -EBUSY;
4461 else if (!(nlmsg_flags & NLM_F_EXCL))
4462 err = 0;
4463 }
4464 goto err5;
4465 }
4466
4467 if (set->size &&
4468 !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact)) {
4469 err = -ENFILE;
4470 goto err6;
4471 }
4472
4473 nft_trans_elem(trans) = elem;
4474 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
4475 return 0;
4476
4477err6:
4478 set->ops->remove(ctx->net, set, &elem);
4479err5:
4480 kfree(trans);
4481err4:
4482 kfree(elem.priv);
4483err3:
4484 if (nla[NFTA_SET_ELEM_DATA] != NULL)
4485 nft_data_release(&data, d2.type);
4486err2:
4487 nft_data_release(&elem.key.val, d1.type);
4488err1:
4489 return err;
4490}
4491
4492static int nf_tables_newsetelem(struct net *net, struct sock *nlsk,
4493 struct sk_buff *skb, const struct nlmsghdr *nlh,
4494 const struct nlattr * const nla[],
4495 struct netlink_ext_ack *extack)
4496{
4497 u8 genmask = nft_genmask_next(net);
4498 const struct nlattr *attr;
4499 struct nft_set *set;
4500 struct nft_ctx ctx;
4501 int rem, err;
4502
4503 if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
4504 return -EINVAL;
4505
4506 err = nft_ctx_init_from_elemattr(&ctx, net, skb, nlh, nla, extack,
4507 genmask);
4508 if (err < 0)
4509 return err;
4510
4511 set = nft_set_lookup_global(net, ctx.table, nla[NFTA_SET_ELEM_LIST_SET],
4512 nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
4513 if (IS_ERR(set))
4514 return PTR_ERR(set);
4515
4516 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
4517 return -EBUSY;
4518
4519 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
4520 err = nft_add_set_elem(&ctx, set, attr, nlh->nlmsg_flags);
4521 if (err < 0)
4522 return err;
4523 }
4524
4525 if (net->nft.validate_state == NFT_VALIDATE_DO)
4526 return nft_table_validate(net, ctx.table);
4527
4528 return 0;
4529}
4530
4531/**
4532 * nft_data_hold - hold a nft_data item
4533 *
4534 * @data: struct nft_data to release
4535 * @type: type of data
4536 *
4537 * Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
4538 * NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
4539 * NFT_GOTO verdicts. This function must be called on active data objects
4540 * from the second phase of the commit protocol.
4541 */
4542void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
4543{
4544 if (type == NFT_DATA_VERDICT) {
4545 switch (data->verdict.code) {
4546 case NFT_JUMP:
4547 case NFT_GOTO:
4548 data->verdict.chain->use++;
4549 break;
4550 }
4551 }
4552}
4553
4554static void nft_set_elem_activate(const struct net *net,
4555 const struct nft_set *set,
4556 struct nft_set_elem *elem)
4557{
4558 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4559
4560 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
4561 nft_data_hold(nft_set_ext_data(ext), set->dtype);
4562 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
4563 (*nft_set_ext_obj(ext))->use++;
4564}
4565
4566static void nft_set_elem_deactivate(const struct net *net,
4567 const struct nft_set *set,
4568 struct nft_set_elem *elem)
4569{
4570 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4571
4572 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
4573 nft_data_release(nft_set_ext_data(ext), set->dtype);
4574 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
4575 (*nft_set_ext_obj(ext))->use--;
4576}
4577
4578static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
4579 const struct nlattr *attr)
4580{
4581 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
4582 struct nft_set_ext_tmpl tmpl;
4583 struct nft_data_desc desc;
4584 struct nft_set_elem elem;
4585 struct nft_set_ext *ext;
4586 struct nft_trans *trans;
4587 u32 flags = 0;
4588 void *priv;
4589 int err;
4590
4591 err = nla_parse_nested(nla, NFTA_SET_ELEM_MAX, attr,
4592 nft_set_elem_policy, NULL);
4593 if (err < 0)
4594 goto err1;
4595
4596 err = -EINVAL;
4597 if (nla[NFTA_SET_ELEM_KEY] == NULL)
4598 goto err1;
4599
4600 nft_set_ext_prepare(&tmpl);
4601
4602 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
4603 if (err < 0)
4604 return err;
4605 if (flags != 0)
4606 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
4607
4608 err = nft_data_init(ctx, &elem.key.val, sizeof(elem.key), &desc,
4609 nla[NFTA_SET_ELEM_KEY]);
4610 if (err < 0)
4611 goto err1;
4612
4613 err = -EINVAL;
4614 if (desc.type != NFT_DATA_VALUE || desc.len != set->klen)
4615 goto err2;
4616
4617 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, desc.len);
4618
4619 err = -ENOMEM;
4620 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data, NULL, 0,
4621 GFP_KERNEL);
4622 if (elem.priv == NULL)
4623 goto err2;
4624
4625 ext = nft_set_elem_ext(set, elem.priv);
4626 if (flags)
4627 *nft_set_ext_flags(ext) = flags;
4628
4629 trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
4630 if (trans == NULL) {
4631 err = -ENOMEM;
4632 goto err3;
4633 }
4634
4635 priv = set->ops->deactivate(ctx->net, set, &elem);
4636 if (priv == NULL) {
4637 err = -ENOENT;
4638 goto err4;
4639 }
4640 kfree(elem.priv);
4641 elem.priv = priv;
4642
4643 nft_set_elem_deactivate(ctx->net, set, &elem);
4644
4645 nft_trans_elem(trans) = elem;
4646 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
4647 return 0;
4648
4649err4:
4650 kfree(trans);
4651err3:
4652 kfree(elem.priv);
4653err2:
4654 nft_data_release(&elem.key.val, desc.type);
4655err1:
4656 return err;
4657}
4658
4659static int nft_flush_set(const struct nft_ctx *ctx,
4660 struct nft_set *set,
4661 const struct nft_set_iter *iter,
4662 struct nft_set_elem *elem)
4663{
4664 struct nft_trans *trans;
4665 int err;
4666
4667 trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
4668 sizeof(struct nft_trans_elem), GFP_ATOMIC);
4669 if (!trans)
4670 return -ENOMEM;
4671
4672 if (!set->ops->flush(ctx->net, set, elem->priv)) {
4673 err = -ENOENT;
4674 goto err1;
4675 }
4676 set->ndeact++;
4677
4678 nft_set_elem_deactivate(ctx->net, set, elem);
4679 nft_trans_elem_set(trans) = set;
4680 nft_trans_elem(trans) = *elem;
4681 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
4682
4683 return 0;
4684err1:
4685 kfree(trans);
4686 return err;
4687}
4688
4689static int nf_tables_delsetelem(struct net *net, struct sock *nlsk,
4690 struct sk_buff *skb, const struct nlmsghdr *nlh,
4691 const struct nlattr * const nla[],
4692 struct netlink_ext_ack *extack)
4693{
4694 u8 genmask = nft_genmask_next(net);
4695 const struct nlattr *attr;
4696 struct nft_set *set;
4697 struct nft_ctx ctx;
4698 int rem, err = 0;
4699
4700 err = nft_ctx_init_from_elemattr(&ctx, net, skb, nlh, nla, extack,
4701 genmask);
4702 if (err < 0)
4703 return err;
4704
4705 set = nft_set_lookup(ctx.table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
4706 if (IS_ERR(set))
4707 return PTR_ERR(set);
4708 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
4709 return -EBUSY;
4710
4711 if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL) {
4712 struct nft_set_iter iter = {
4713 .genmask = genmask,
4714 .fn = nft_flush_set,
4715 };
4716 set->ops->walk(&ctx, set, &iter);
4717
4718 return iter.err;
4719 }
4720
4721 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
4722 err = nft_del_setelem(&ctx, set, attr);
4723 if (err < 0)
4724 break;
4725
4726 set->ndeact++;
4727 }
4728 return err;
4729}
4730
4731void nft_set_gc_batch_release(struct rcu_head *rcu)
4732{
4733 struct nft_set_gc_batch *gcb;
4734 unsigned int i;
4735
4736 gcb = container_of(rcu, struct nft_set_gc_batch, head.rcu);
4737 for (i = 0; i < gcb->head.cnt; i++)
4738 nft_set_elem_destroy(gcb->head.set, gcb->elems[i], true);
4739 kfree(gcb);
4740}
4741EXPORT_SYMBOL_GPL(nft_set_gc_batch_release);
4742
4743struct nft_set_gc_batch *nft_set_gc_batch_alloc(const struct nft_set *set,
4744 gfp_t gfp)
4745{
4746 struct nft_set_gc_batch *gcb;
4747
4748 gcb = kzalloc(sizeof(*gcb), gfp);
4749 if (gcb == NULL)
4750 return gcb;
4751 gcb->head.set = set;
4752 return gcb;
4753}
4754EXPORT_SYMBOL_GPL(nft_set_gc_batch_alloc);
4755
4756/*
4757 * Stateful objects
4758 */
4759
4760/**
4761 * nft_register_obj- register nf_tables stateful object type
4762 * @obj: object type
4763 *
4764 * Registers the object type for use with nf_tables. Returns zero on
4765 * success or a negative errno code otherwise.
4766 */
4767int nft_register_obj(struct nft_object_type *obj_type)
4768{
4769 if (obj_type->type == NFT_OBJECT_UNSPEC)
4770 return -EINVAL;
4771
4772 nfnl_lock(NFNL_SUBSYS_NFTABLES);
4773 list_add_rcu(&obj_type->list, &nf_tables_objects);
4774 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
4775 return 0;
4776}
4777EXPORT_SYMBOL_GPL(nft_register_obj);
4778
4779/**
4780 * nft_unregister_obj - unregister nf_tables object type
4781 * @obj: object type
4782 *
4783 * Unregisters the object type for use with nf_tables.
4784 */
4785void nft_unregister_obj(struct nft_object_type *obj_type)
4786{
4787 nfnl_lock(NFNL_SUBSYS_NFTABLES);
4788 list_del_rcu(&obj_type->list);
4789 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
4790}
4791EXPORT_SYMBOL_GPL(nft_unregister_obj);
4792
4793struct nft_object *nft_obj_lookup(const struct nft_table *table,
4794 const struct nlattr *nla, u32 objtype,
4795 u8 genmask)
4796{
4797 struct nft_object *obj;
4798
4799 list_for_each_entry_rcu(obj, &table->objects, list) {
4800 if (!nla_strcmp(nla, obj->name) &&
4801 objtype == obj->ops->type->type &&
4802 nft_active_genmask(obj, genmask))
4803 return obj;
4804 }
4805 return ERR_PTR(-ENOENT);
4806}
4807EXPORT_SYMBOL_GPL(nft_obj_lookup);
4808
4809static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
4810 const struct nlattr *nla,
4811 u32 objtype, u8 genmask)
4812{
4813 struct nft_object *obj;
4814
4815 list_for_each_entry(obj, &table->objects, list) {
4816 if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
4817 objtype == obj->ops->type->type &&
4818 nft_active_genmask(obj, genmask))
4819 return obj;
4820 }
4821 return ERR_PTR(-ENOENT);
4822}
4823
4824static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
4825 [NFTA_OBJ_TABLE] = { .type = NLA_STRING,
4826 .len = NFT_TABLE_MAXNAMELEN - 1 },
4827 [NFTA_OBJ_NAME] = { .type = NLA_STRING,
4828 .len = NFT_OBJ_MAXNAMELEN - 1 },
4829 [NFTA_OBJ_TYPE] = { .type = NLA_U32 },
4830 [NFTA_OBJ_DATA] = { .type = NLA_NESTED },
4831 [NFTA_OBJ_HANDLE] = { .type = NLA_U64},
4832};
4833
4834static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
4835 const struct nft_object_type *type,
4836 const struct nlattr *attr)
4837{
4838 struct nlattr **tb;
4839 const struct nft_object_ops *ops;
4840 struct nft_object *obj;
4841 int err = -ENOMEM;
4842
4843 tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
4844 if (!tb)
4845 goto err1;
4846
4847 if (attr) {
4848 err = nla_parse_nested(tb, type->maxattr, attr, type->policy,
4849 NULL);
4850 if (err < 0)
4851 goto err2;
4852 } else {
4853 memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
4854 }
4855
4856 if (type->select_ops) {
4857 ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
4858 if (IS_ERR(ops)) {
4859 err = PTR_ERR(ops);
4860 goto err2;
4861 }
4862 } else {
4863 ops = type->ops;
4864 }
4865
4866 err = -ENOMEM;
4867 obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL);
4868 if (!obj)
4869 goto err2;
4870
4871 err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
4872 if (err < 0)
4873 goto err3;
4874
4875 obj->ops = ops;
4876
4877 kfree(tb);
4878 return obj;
4879err3:
4880 kfree(obj);
4881err2:
4882 kfree(tb);
4883err1:
4884 return ERR_PTR(err);
4885}
4886
4887static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
4888 struct nft_object *obj, bool reset)
4889{
4890 struct nlattr *nest;
4891
4892 nest = nla_nest_start(skb, attr);
4893 if (!nest)
4894 goto nla_put_failure;
4895 if (obj->ops->dump(skb, obj, reset) < 0)
4896 goto nla_put_failure;
4897 nla_nest_end(skb, nest);
4898 return 0;
4899
4900nla_put_failure:
4901 return -1;
4902}
4903
4904static const struct nft_object_type *__nft_obj_type_get(u32 objtype)
4905{
4906 const struct nft_object_type *type;
4907
4908 list_for_each_entry(type, &nf_tables_objects, list) {
4909 if (objtype == type->type)
4910 return type;
4911 }
4912 return NULL;
4913}
4914
4915static const struct nft_object_type *
4916nft_obj_type_get(struct net *net, u32 objtype)
4917{
4918 const struct nft_object_type *type;
4919
4920 type = __nft_obj_type_get(objtype);
4921 if (type != NULL && try_module_get(type->owner))
4922 return type;
4923
4924 lockdep_nfnl_nft_mutex_not_held();
4925#ifdef CONFIG_MODULES
4926 if (type == NULL) {
4927 nft_request_module(net, "nft-obj-%u", objtype);
4928 if (__nft_obj_type_get(objtype))
4929 return ERR_PTR(-EAGAIN);
4930 }
4931#endif
4932 return ERR_PTR(-ENOENT);
4933}
4934
4935static int nf_tables_newobj(struct net *net, struct sock *nlsk,
4936 struct sk_buff *skb, const struct nlmsghdr *nlh,
4937 const struct nlattr * const nla[],
4938 struct netlink_ext_ack *extack)
4939{
4940 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
4941 const struct nft_object_type *type;
4942 u8 genmask = nft_genmask_next(net);
4943 int family = nfmsg->nfgen_family;
4944 struct nft_table *table;
4945 struct nft_object *obj;
4946 struct nft_ctx ctx;
4947 u32 objtype;
4948 int err;
4949
4950 if (!nla[NFTA_OBJ_TYPE] ||
4951 !nla[NFTA_OBJ_NAME] ||
4952 !nla[NFTA_OBJ_DATA])
4953 return -EINVAL;
4954
4955 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask);
4956 if (IS_ERR(table)) {
4957 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
4958 return PTR_ERR(table);
4959 }
4960
4961 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
4962 obj = nft_obj_lookup(table, nla[NFTA_OBJ_NAME], objtype, genmask);
4963 if (IS_ERR(obj)) {
4964 err = PTR_ERR(obj);
4965 if (err != -ENOENT) {
4966 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
4967 return err;
4968 }
4969 } else {
4970 if (nlh->nlmsg_flags & NLM_F_EXCL) {
4971 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
4972 return -EEXIST;
4973 }
4974 return 0;
4975 }
4976
4977 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
4978
4979 type = nft_obj_type_get(net, objtype);
4980 if (IS_ERR(type))
4981 return PTR_ERR(type);
4982
4983 obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
4984 if (IS_ERR(obj)) {
4985 err = PTR_ERR(obj);
4986 goto err1;
4987 }
4988 obj->table = table;
4989 obj->handle = nf_tables_alloc_handle(table);
4990
4991 obj->name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL);
4992 if (!obj->name) {
4993 err = -ENOMEM;
4994 goto err2;
4995 }
4996
4997 err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
4998 if (err < 0)
4999 goto err3;
5000
5001 list_add_tail_rcu(&obj->list, &table->objects);
5002 table->use++;
5003 return 0;
5004err3:
5005 kfree(obj->name);
5006err2:
5007 if (obj->ops->destroy)
5008 obj->ops->destroy(&ctx, obj);
5009 kfree(obj);
5010err1:
5011 module_put(type->owner);
5012 return err;
5013}
5014
5015static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
5016 u32 portid, u32 seq, int event, u32 flags,
5017 int family, const struct nft_table *table,
5018 struct nft_object *obj, bool reset)
5019{
5020 struct nfgenmsg *nfmsg;
5021 struct nlmsghdr *nlh;
5022
5023 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
5024 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
5025 if (nlh == NULL)
5026 goto nla_put_failure;
5027
5028 nfmsg = nlmsg_data(nlh);
5029 nfmsg->nfgen_family = family;
5030 nfmsg->version = NFNETLINK_V0;
5031 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
5032
5033 if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
5034 nla_put_string(skb, NFTA_OBJ_NAME, obj->name) ||
5035 nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
5036 nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
5037 nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset) ||
5038 nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
5039 NFTA_OBJ_PAD))
5040 goto nla_put_failure;
5041
5042 nlmsg_end(skb, nlh);
5043 return 0;
5044
5045nla_put_failure:
5046 nlmsg_trim(skb, nlh);
5047 return -1;
5048}
5049
5050struct nft_obj_filter {
5051 char *table;
5052 u32 type;
5053};
5054
5055static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
5056{
5057 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
5058 const struct nft_table *table;
5059 unsigned int idx = 0, s_idx = cb->args[0];
5060 struct nft_obj_filter *filter = cb->data;
5061 struct net *net = sock_net(skb->sk);
5062 int family = nfmsg->nfgen_family;
5063 struct nft_object *obj;
5064 bool reset = false;
5065
5066 if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
5067 reset = true;
5068
5069 rcu_read_lock();
5070 cb->seq = net->nft.base_seq;
5071
5072 list_for_each_entry_rcu(table, &net->nft.tables, list) {
5073 if (family != NFPROTO_UNSPEC && family != table->family)
5074 continue;
5075
5076 list_for_each_entry_rcu(obj, &table->objects, list) {
5077 if (!nft_is_active(net, obj))
5078 goto cont;
5079 if (idx < s_idx)
5080 goto cont;
5081 if (idx > s_idx)
5082 memset(&cb->args[1], 0,
5083 sizeof(cb->args) - sizeof(cb->args[0]));
5084 if (filter && filter->table &&
5085 strcmp(filter->table, table->name))
5086 goto cont;
5087 if (filter &&
5088 filter->type != NFT_OBJECT_UNSPEC &&
5089 obj->ops->type->type != filter->type)
5090 goto cont;
5091
5092 if (nf_tables_fill_obj_info(skb, net, NETLINK_CB(cb->skb).portid,
5093 cb->nlh->nlmsg_seq,
5094 NFT_MSG_NEWOBJ,
5095 NLM_F_MULTI | NLM_F_APPEND,
5096 table->family, table,
5097 obj, reset) < 0)
5098 goto done;
5099
5100 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
5101cont:
5102 idx++;
5103 }
5104 }
5105done:
5106 rcu_read_unlock();
5107
5108 cb->args[0] = idx;
5109 return skb->len;
5110}
5111
5112static int nf_tables_dump_obj_start(struct netlink_callback *cb)
5113{
5114 const struct nlattr * const *nla = cb->data;
5115 struct nft_obj_filter *filter = NULL;
5116
5117 if (nla[NFTA_OBJ_TABLE] || nla[NFTA_OBJ_TYPE]) {
5118 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
5119 if (!filter)
5120 return -ENOMEM;
5121
5122 if (nla[NFTA_OBJ_TABLE]) {
5123 filter->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
5124 if (!filter->table) {
5125 kfree(filter);
5126 return -ENOMEM;
5127 }
5128 }
5129
5130 if (nla[NFTA_OBJ_TYPE])
5131 filter->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
5132 }
5133
5134 cb->data = filter;
5135 return 0;
5136}
5137
5138static int nf_tables_dump_obj_done(struct netlink_callback *cb)
5139{
5140 struct nft_obj_filter *filter = cb->data;
5141
5142 if (filter) {
5143 kfree(filter->table);
5144 kfree(filter);
5145 }
5146
5147 return 0;
5148}
5149
5150/* called with rcu_read_lock held */
5151static int nf_tables_getobj(struct net *net, struct sock *nlsk,
5152 struct sk_buff *skb, const struct nlmsghdr *nlh,
5153 const struct nlattr * const nla[],
5154 struct netlink_ext_ack *extack)
5155{
5156 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
5157 u8 genmask = nft_genmask_cur(net);
5158 int family = nfmsg->nfgen_family;
5159 const struct nft_table *table;
5160 struct nft_object *obj;
5161 struct sk_buff *skb2;
5162 bool reset = false;
5163 u32 objtype;
5164 int err;
5165
5166 if (nlh->nlmsg_flags & NLM_F_DUMP) {
5167 struct netlink_dump_control c = {
5168 .start = nf_tables_dump_obj_start,
5169 .dump = nf_tables_dump_obj,
5170 .done = nf_tables_dump_obj_done,
5171 .module = THIS_MODULE,
5172 .data = (void *)nla,
5173 };
5174
5175 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
5176 }
5177
5178 if (!nla[NFTA_OBJ_NAME] ||
5179 !nla[NFTA_OBJ_TYPE])
5180 return -EINVAL;
5181
5182 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask);
5183 if (IS_ERR(table)) {
5184 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
5185 return PTR_ERR(table);
5186 }
5187
5188 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
5189 obj = nft_obj_lookup(table, nla[NFTA_OBJ_NAME], objtype, genmask);
5190 if (IS_ERR(obj)) {
5191 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
5192 return PTR_ERR(obj);
5193 }
5194
5195 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
5196 if (!skb2)
5197 return -ENOMEM;
5198
5199 if (NFNL_MSG_TYPE(nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
5200 reset = true;
5201
5202 err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
5203 nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
5204 family, table, obj, reset);
5205 if (err < 0)
5206 goto err;
5207
5208 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
5209err:
5210 kfree_skb(skb2);
5211 return err;
5212}
5213
5214static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
5215{
5216 if (obj->ops->destroy)
5217 obj->ops->destroy(ctx, obj);
5218
5219 module_put(obj->ops->type->owner);
5220 kfree(obj->name);
5221 kfree(obj);
5222}
5223
5224static int nf_tables_delobj(struct net *net, struct sock *nlsk,
5225 struct sk_buff *skb, const struct nlmsghdr *nlh,
5226 const struct nlattr * const nla[],
5227 struct netlink_ext_ack *extack)
5228{
5229 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
5230 u8 genmask = nft_genmask_next(net);
5231 int family = nfmsg->nfgen_family;
5232 const struct nlattr *attr;
5233 struct nft_table *table;
5234 struct nft_object *obj;
5235 struct nft_ctx ctx;
5236 u32 objtype;
5237
5238 if (!nla[NFTA_OBJ_TYPE] ||
5239 (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
5240 return -EINVAL;
5241
5242 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask);
5243 if (IS_ERR(table)) {
5244 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
5245 return PTR_ERR(table);
5246 }
5247
5248 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
5249 if (nla[NFTA_OBJ_HANDLE]) {
5250 attr = nla[NFTA_OBJ_HANDLE];
5251 obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
5252 } else {
5253 attr = nla[NFTA_OBJ_NAME];
5254 obj = nft_obj_lookup(table, attr, objtype, genmask);
5255 }
5256
5257 if (IS_ERR(obj)) {
5258 NL_SET_BAD_ATTR(extack, attr);
5259 return PTR_ERR(obj);
5260 }
5261 if (obj->use > 0) {
5262 NL_SET_BAD_ATTR(extack, attr);
5263 return -EBUSY;
5264 }
5265
5266 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
5267
5268 return nft_delobj(&ctx, obj);
5269}
5270
5271void nft_obj_notify(struct net *net, struct nft_table *table,
5272 struct nft_object *obj, u32 portid, u32 seq, int event,
5273 int family, int report, gfp_t gfp)
5274{
5275 struct sk_buff *skb;
5276 int err;
5277
5278 if (!report &&
5279 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5280 return;
5281
5282 skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
5283 if (skb == NULL)
5284 goto err;
5285
5286 err = nf_tables_fill_obj_info(skb, net, portid, seq, event, 0, family,
5287 table, obj, false);
5288 if (err < 0) {
5289 kfree_skb(skb);
5290 goto err;
5291 }
5292
5293 nfnetlink_send(skb, net, portid, NFNLGRP_NFTABLES, report, gfp);
5294 return;
5295err:
5296 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
5297}
5298EXPORT_SYMBOL_GPL(nft_obj_notify);
5299
5300static void nf_tables_obj_notify(const struct nft_ctx *ctx,
5301 struct nft_object *obj, int event)
5302{
5303 nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid, ctx->seq, event,
5304 ctx->family, ctx->report, GFP_KERNEL);
5305}
5306
5307/*
5308 * Flow tables
5309 */
5310void nft_register_flowtable_type(struct nf_flowtable_type *type)
5311{
5312 nfnl_lock(NFNL_SUBSYS_NFTABLES);
5313 list_add_tail_rcu(&type->list, &nf_tables_flowtables);
5314 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
5315}
5316EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
5317
5318void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
5319{
5320 nfnl_lock(NFNL_SUBSYS_NFTABLES);
5321 list_del_rcu(&type->list);
5322 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
5323}
5324EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
5325
5326static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
5327 [NFTA_FLOWTABLE_TABLE] = { .type = NLA_STRING,
5328 .len = NFT_NAME_MAXLEN - 1 },
5329 [NFTA_FLOWTABLE_NAME] = { .type = NLA_STRING,
5330 .len = NFT_NAME_MAXLEN - 1 },
5331 [NFTA_FLOWTABLE_HOOK] = { .type = NLA_NESTED },
5332 [NFTA_FLOWTABLE_HANDLE] = { .type = NLA_U64 },
5333};
5334
5335struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
5336 const struct nlattr *nla, u8 genmask)
5337{
5338 struct nft_flowtable *flowtable;
5339
5340 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
5341 if (!nla_strcmp(nla, flowtable->name) &&
5342 nft_active_genmask(flowtable, genmask))
5343 return flowtable;
5344 }
5345 return ERR_PTR(-ENOENT);
5346}
5347EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
5348
5349static struct nft_flowtable *
5350nft_flowtable_lookup_byhandle(const struct nft_table *table,
5351 const struct nlattr *nla, u8 genmask)
5352{
5353 struct nft_flowtable *flowtable;
5354
5355 list_for_each_entry(flowtable, &table->flowtables, list) {
5356 if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
5357 nft_active_genmask(flowtable, genmask))
5358 return flowtable;
5359 }
5360 return ERR_PTR(-ENOENT);
5361}
5362
5363static int nf_tables_parse_devices(const struct nft_ctx *ctx,
5364 const struct nlattr *attr,
5365 struct net_device *dev_array[], int *len)
5366{
5367 const struct nlattr *tmp;
5368 struct net_device *dev;
5369 char ifname[IFNAMSIZ];
5370 int rem, n = 0, err;
5371
5372 nla_for_each_nested(tmp, attr, rem) {
5373 if (nla_type(tmp) != NFTA_DEVICE_NAME) {
5374 err = -EINVAL;
5375 goto err1;
5376 }
5377
5378 nla_strlcpy(ifname, tmp, IFNAMSIZ);
5379 dev = __dev_get_by_name(ctx->net, ifname);
5380 if (!dev) {
5381 err = -ENOENT;
5382 goto err1;
5383 }
5384
5385 dev_array[n++] = dev;
5386 if (n == NFT_FLOWTABLE_DEVICE_MAX) {
5387 err = -EFBIG;
5388 goto err1;
5389 }
5390 }
5391 if (!len)
5392 return -EINVAL;
5393
5394 err = 0;
5395err1:
5396 *len = n;
5397 return err;
5398}
5399
5400static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
5401 [NFTA_FLOWTABLE_HOOK_NUM] = { .type = NLA_U32 },
5402 [NFTA_FLOWTABLE_HOOK_PRIORITY] = { .type = NLA_U32 },
5403 [NFTA_FLOWTABLE_HOOK_DEVS] = { .type = NLA_NESTED },
5404};
5405
5406static int nf_tables_flowtable_parse_hook(const struct nft_ctx *ctx,
5407 const struct nlattr *attr,
5408 struct nft_flowtable *flowtable)
5409{
5410 struct net_device *dev_array[NFT_FLOWTABLE_DEVICE_MAX];
5411 struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
5412 struct nf_hook_ops *ops;
5413 int hooknum, priority;
5414 int err, n = 0, i;
5415
5416 err = nla_parse_nested(tb, NFTA_FLOWTABLE_HOOK_MAX, attr,
5417 nft_flowtable_hook_policy, NULL);
5418 if (err < 0)
5419 return err;
5420
5421 if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
5422 !tb[NFTA_FLOWTABLE_HOOK_PRIORITY] ||
5423 !tb[NFTA_FLOWTABLE_HOOK_DEVS])
5424 return -EINVAL;
5425
5426 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
5427 if (hooknum != NF_NETDEV_INGRESS)
5428 return -EINVAL;
5429
5430 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
5431
5432 err = nf_tables_parse_devices(ctx, tb[NFTA_FLOWTABLE_HOOK_DEVS],
5433 dev_array, &n);
5434 if (err < 0)
5435 return err;
5436
5437 ops = kcalloc(n, sizeof(struct nf_hook_ops), GFP_KERNEL);
5438 if (!ops)
5439 return -ENOMEM;
5440
5441 flowtable->hooknum = hooknum;
5442 flowtable->priority = priority;
5443 flowtable->ops = ops;
5444 flowtable->ops_len = n;
5445
5446 for (i = 0; i < n; i++) {
5447 flowtable->ops[i].pf = NFPROTO_NETDEV;
5448 flowtable->ops[i].hooknum = hooknum;
5449 flowtable->ops[i].priority = priority;
5450 flowtable->ops[i].priv = &flowtable->data;
5451 flowtable->ops[i].hook = flowtable->data.type->hook;
5452 flowtable->ops[i].dev = dev_array[i];
5453 }
5454
5455 return err;
5456}
5457
5458static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
5459{
5460 const struct nf_flowtable_type *type;
5461
5462 list_for_each_entry(type, &nf_tables_flowtables, list) {
5463 if (family == type->family)
5464 return type;
5465 }
5466 return NULL;
5467}
5468
5469static const struct nf_flowtable_type *
5470nft_flowtable_type_get(struct net *net, u8 family)
5471{
5472 const struct nf_flowtable_type *type;
5473
5474 type = __nft_flowtable_type_get(family);
5475 if (type != NULL && try_module_get(type->owner))
5476 return type;
5477
5478 lockdep_nfnl_nft_mutex_not_held();
5479#ifdef CONFIG_MODULES
5480 if (type == NULL) {
5481 nft_request_module(net, "nf-flowtable-%u", family);
5482 if (__nft_flowtable_type_get(family))
5483 return ERR_PTR(-EAGAIN);
5484 }
5485#endif
5486 return ERR_PTR(-ENOENT);
5487}
5488
5489static void nft_unregister_flowtable_net_hooks(struct net *net,
5490 struct nft_flowtable *flowtable)
5491{
5492 int i;
5493
5494 for (i = 0; i < flowtable->ops_len; i++) {
5495 if (!flowtable->ops[i].dev)
5496 continue;
5497
5498 nf_unregister_net_hook(net, &flowtable->ops[i]);
5499 }
5500}
5501
5502static int nf_tables_newflowtable(struct net *net, struct sock *nlsk,
5503 struct sk_buff *skb,
5504 const struct nlmsghdr *nlh,
5505 const struct nlattr * const nla[],
5506 struct netlink_ext_ack *extack)
5507{
5508 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
5509 const struct nf_flowtable_type *type;
5510 struct nft_flowtable *flowtable, *ft;
5511 u8 genmask = nft_genmask_next(net);
5512 int family = nfmsg->nfgen_family;
5513 struct nft_table *table;
5514 struct nft_ctx ctx;
5515 int err, i, k;
5516
5517 if (!nla[NFTA_FLOWTABLE_TABLE] ||
5518 !nla[NFTA_FLOWTABLE_NAME] ||
5519 !nla[NFTA_FLOWTABLE_HOOK])
5520 return -EINVAL;
5521
5522 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
5523 genmask);
5524 if (IS_ERR(table)) {
5525 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
5526 return PTR_ERR(table);
5527 }
5528
5529 flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
5530 genmask);
5531 if (IS_ERR(flowtable)) {
5532 err = PTR_ERR(flowtable);
5533 if (err != -ENOENT) {
5534 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
5535 return err;
5536 }
5537 } else {
5538 if (nlh->nlmsg_flags & NLM_F_EXCL) {
5539 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
5540 return -EEXIST;
5541 }
5542
5543 return 0;
5544 }
5545
5546 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
5547
5548 flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL);
5549 if (!flowtable)
5550 return -ENOMEM;
5551
5552 flowtable->table = table;
5553 flowtable->handle = nf_tables_alloc_handle(table);
5554
5555 flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL);
5556 if (!flowtable->name) {
5557 err = -ENOMEM;
5558 goto err1;
5559 }
5560
5561 type = nft_flowtable_type_get(net, family);
5562 if (IS_ERR(type)) {
5563 err = PTR_ERR(type);
5564 goto err2;
5565 }
5566
5567 flowtable->data.type = type;
5568 err = type->init(&flowtable->data);
5569 if (err < 0)
5570 goto err3;
5571
5572 err = nf_tables_flowtable_parse_hook(&ctx, nla[NFTA_FLOWTABLE_HOOK],
5573 flowtable);
5574 if (err < 0)
5575 goto err4;
5576
5577 for (i = 0; i < flowtable->ops_len; i++) {
5578 if (!flowtable->ops[i].dev)
5579 continue;
5580
5581 list_for_each_entry(ft, &table->flowtables, list) {
5582 for (k = 0; k < ft->ops_len; k++) {
5583 if (!ft->ops[k].dev)
5584 continue;
5585
5586 if (flowtable->ops[i].dev == ft->ops[k].dev &&
5587 flowtable->ops[i].pf == ft->ops[k].pf) {
5588 err = -EBUSY;
5589 goto err5;
5590 }
5591 }
5592 }
5593
5594 err = nf_register_net_hook(net, &flowtable->ops[i]);
5595 if (err < 0)
5596 goto err5;
5597 }
5598
5599 err = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
5600 if (err < 0)
5601 goto err6;
5602
5603 list_add_tail_rcu(&flowtable->list, &table->flowtables);
5604 table->use++;
5605
5606 return 0;
5607err6:
5608 i = flowtable->ops_len;
5609err5:
5610 for (k = i - 1; k >= 0; k--)
5611 nf_unregister_net_hook(net, &flowtable->ops[k]);
5612
5613 kfree(flowtable->ops);
5614err4:
5615 flowtable->data.type->free(&flowtable->data);
5616err3:
5617 module_put(type->owner);
5618err2:
5619 kfree(flowtable->name);
5620err1:
5621 kfree(flowtable);
5622 return err;
5623}
5624
5625static int nf_tables_delflowtable(struct net *net, struct sock *nlsk,
5626 struct sk_buff *skb,
5627 const struct nlmsghdr *nlh,
5628 const struct nlattr * const nla[],
5629 struct netlink_ext_ack *extack)
5630{
5631 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
5632 u8 genmask = nft_genmask_next(net);
5633 int family = nfmsg->nfgen_family;
5634 struct nft_flowtable *flowtable;
5635 const struct nlattr *attr;
5636 struct nft_table *table;
5637 struct nft_ctx ctx;
5638
5639 if (!nla[NFTA_FLOWTABLE_TABLE] ||
5640 (!nla[NFTA_FLOWTABLE_NAME] &&
5641 !nla[NFTA_FLOWTABLE_HANDLE]))
5642 return -EINVAL;
5643
5644 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
5645 genmask);
5646 if (IS_ERR(table)) {
5647 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
5648 return PTR_ERR(table);
5649 }
5650
5651 if (nla[NFTA_FLOWTABLE_HANDLE]) {
5652 attr = nla[NFTA_FLOWTABLE_HANDLE];
5653 flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
5654 } else {
5655 attr = nla[NFTA_FLOWTABLE_NAME];
5656 flowtable = nft_flowtable_lookup(table, attr, genmask);
5657 }
5658
5659 if (IS_ERR(flowtable)) {
5660 NL_SET_BAD_ATTR(extack, attr);
5661 return PTR_ERR(flowtable);
5662 }
5663 if (flowtable->use > 0) {
5664 NL_SET_BAD_ATTR(extack, attr);
5665 return -EBUSY;
5666 }
5667
5668 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
5669
5670 return nft_delflowtable(&ctx, flowtable);
5671}
5672
5673static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
5674 u32 portid, u32 seq, int event,
5675 u32 flags, int family,
5676 struct nft_flowtable *flowtable)
5677{
5678 struct nlattr *nest, *nest_devs;
5679 struct nfgenmsg *nfmsg;
5680 struct nlmsghdr *nlh;
5681 int i;
5682
5683 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
5684 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
5685 if (nlh == NULL)
5686 goto nla_put_failure;
5687
5688 nfmsg = nlmsg_data(nlh);
5689 nfmsg->nfgen_family = family;
5690 nfmsg->version = NFNETLINK_V0;
5691 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
5692
5693 if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
5694 nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
5695 nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
5696 nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
5697 NFTA_FLOWTABLE_PAD))
5698 goto nla_put_failure;
5699
5700 nest = nla_nest_start(skb, NFTA_FLOWTABLE_HOOK);
5701 if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
5702 nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->priority)))
5703 goto nla_put_failure;
5704
5705 nest_devs = nla_nest_start(skb, NFTA_FLOWTABLE_HOOK_DEVS);
5706 if (!nest_devs)
5707 goto nla_put_failure;
5708
5709 for (i = 0; i < flowtable->ops_len; i++) {
5710 const struct net_device *dev = READ_ONCE(flowtable->ops[i].dev);
5711
5712 if (dev &&
5713 nla_put_string(skb, NFTA_DEVICE_NAME, dev->name))
5714 goto nla_put_failure;
5715 }
5716 nla_nest_end(skb, nest_devs);
5717 nla_nest_end(skb, nest);
5718
5719 nlmsg_end(skb, nlh);
5720 return 0;
5721
5722nla_put_failure:
5723 nlmsg_trim(skb, nlh);
5724 return -1;
5725}
5726
5727struct nft_flowtable_filter {
5728 char *table;
5729};
5730
5731static int nf_tables_dump_flowtable(struct sk_buff *skb,
5732 struct netlink_callback *cb)
5733{
5734 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
5735 struct nft_flowtable_filter *filter = cb->data;
5736 unsigned int idx = 0, s_idx = cb->args[0];
5737 struct net *net = sock_net(skb->sk);
5738 int family = nfmsg->nfgen_family;
5739 struct nft_flowtable *flowtable;
5740 const struct nft_table *table;
5741
5742 rcu_read_lock();
5743 cb->seq = net->nft.base_seq;
5744
5745 list_for_each_entry_rcu(table, &net->nft.tables, list) {
5746 if (family != NFPROTO_UNSPEC && family != table->family)
5747 continue;
5748
5749 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
5750 if (!nft_is_active(net, flowtable))
5751 goto cont;
5752 if (idx < s_idx)
5753 goto cont;
5754 if (idx > s_idx)
5755 memset(&cb->args[1], 0,
5756 sizeof(cb->args) - sizeof(cb->args[0]));
5757 if (filter && filter->table &&
5758 strcmp(filter->table, table->name))
5759 goto cont;
5760
5761 if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
5762 cb->nlh->nlmsg_seq,
5763 NFT_MSG_NEWFLOWTABLE,
5764 NLM_F_MULTI | NLM_F_APPEND,
5765 table->family, flowtable) < 0)
5766 goto done;
5767
5768 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
5769cont:
5770 idx++;
5771 }
5772 }
5773done:
5774 rcu_read_unlock();
5775
5776 cb->args[0] = idx;
5777 return skb->len;
5778}
5779
5780static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
5781{
5782 const struct nlattr * const *nla = cb->data;
5783 struct nft_flowtable_filter *filter = NULL;
5784
5785 if (nla[NFTA_FLOWTABLE_TABLE]) {
5786 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
5787 if (!filter)
5788 return -ENOMEM;
5789
5790 filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
5791 GFP_ATOMIC);
5792 if (!filter->table) {
5793 kfree(filter);
5794 return -ENOMEM;
5795 }
5796 }
5797
5798 cb->data = filter;
5799 return 0;
5800}
5801
5802static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
5803{
5804 struct nft_flowtable_filter *filter = cb->data;
5805
5806 if (!filter)
5807 return 0;
5808
5809 kfree(filter->table);
5810 kfree(filter);
5811
5812 return 0;
5813}
5814
5815/* called with rcu_read_lock held */
5816static int nf_tables_getflowtable(struct net *net, struct sock *nlsk,
5817 struct sk_buff *skb,
5818 const struct nlmsghdr *nlh,
5819 const struct nlattr * const nla[],
5820 struct netlink_ext_ack *extack)
5821{
5822 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
5823 u8 genmask = nft_genmask_cur(net);
5824 int family = nfmsg->nfgen_family;
5825 struct nft_flowtable *flowtable;
5826 const struct nft_table *table;
5827 struct sk_buff *skb2;
5828 int err;
5829
5830 if (nlh->nlmsg_flags & NLM_F_DUMP) {
5831 struct netlink_dump_control c = {
5832 .start = nf_tables_dump_flowtable_start,
5833 .dump = nf_tables_dump_flowtable,
5834 .done = nf_tables_dump_flowtable_done,
5835 .module = THIS_MODULE,
5836 .data = (void *)nla,
5837 };
5838
5839 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
5840 }
5841
5842 if (!nla[NFTA_FLOWTABLE_NAME])
5843 return -EINVAL;
5844
5845 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
5846 genmask);
5847 if (IS_ERR(table))
5848 return PTR_ERR(table);
5849
5850 flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
5851 genmask);
5852 if (IS_ERR(flowtable))
5853 return PTR_ERR(flowtable);
5854
5855 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
5856 if (!skb2)
5857 return -ENOMEM;
5858
5859 err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
5860 nlh->nlmsg_seq,
5861 NFT_MSG_NEWFLOWTABLE, 0, family,
5862 flowtable);
5863 if (err < 0)
5864 goto err;
5865
5866 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
5867err:
5868 kfree_skb(skb2);
5869 return err;
5870}
5871
5872static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
5873 struct nft_flowtable *flowtable,
5874 int event)
5875{
5876 struct sk_buff *skb;
5877 int err;
5878
5879 if (ctx->report &&
5880 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
5881 return;
5882
5883 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5884 if (skb == NULL)
5885 goto err;
5886
5887 err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
5888 ctx->seq, event, 0,
5889 ctx->family, flowtable);
5890 if (err < 0) {
5891 kfree_skb(skb);
5892 goto err;
5893 }
5894
5895 nfnetlink_send(skb, ctx->net, ctx->portid, NFNLGRP_NFTABLES,
5896 ctx->report, GFP_KERNEL);
5897 return;
5898err:
5899 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
5900}
5901
5902static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
5903{
5904 kfree(flowtable->ops);
5905 kfree(flowtable->name);
5906 flowtable->data.type->free(&flowtable->data);
5907 module_put(flowtable->data.type->owner);
5908 kfree(flowtable);
5909}
5910
5911static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
5912 u32 portid, u32 seq)
5913{
5914 struct nlmsghdr *nlh;
5915 struct nfgenmsg *nfmsg;
5916 char buf[TASK_COMM_LEN];
5917 int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
5918
5919 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), 0);
5920 if (nlh == NULL)
5921 goto nla_put_failure;
5922
5923 nfmsg = nlmsg_data(nlh);
5924 nfmsg->nfgen_family = AF_UNSPEC;
5925 nfmsg->version = NFNETLINK_V0;
5926 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
5927
5928 if (nla_put_be32(skb, NFTA_GEN_ID, htonl(net->nft.base_seq)) ||
5929 nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
5930 nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
5931 goto nla_put_failure;
5932
5933 nlmsg_end(skb, nlh);
5934 return 0;
5935
5936nla_put_failure:
5937 nlmsg_trim(skb, nlh);
5938 return -EMSGSIZE;
5939}
5940
5941static void nft_flowtable_event(unsigned long event, struct net_device *dev,
5942 struct nft_flowtable *flowtable)
5943{
5944 int i;
5945
5946 for (i = 0; i < flowtable->ops_len; i++) {
5947 if (flowtable->ops[i].dev != dev)
5948 continue;
5949
5950 nf_unregister_net_hook(dev_net(dev), &flowtable->ops[i]);
5951 flowtable->ops[i].dev = NULL;
5952 break;
5953 }
5954}
5955
5956static int nf_tables_flowtable_event(struct notifier_block *this,
5957 unsigned long event, void *ptr)
5958{
5959 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
5960 struct nft_flowtable *flowtable;
5961 struct nft_table *table;
5962 struct net *net;
5963
5964 if (event != NETDEV_UNREGISTER)
5965 return 0;
5966
5967 net = dev_net(dev);
5968 mutex_lock(&net->nft.commit_mutex);
5969 list_for_each_entry(table, &net->nft.tables, list) {
5970 list_for_each_entry(flowtable, &table->flowtables, list) {
5971 nft_flowtable_event(event, dev, flowtable);
5972 }
5973 }
5974 mutex_unlock(&net->nft.commit_mutex);
5975
5976 return NOTIFY_DONE;
5977}
5978
5979static struct notifier_block nf_tables_flowtable_notifier = {
5980 .notifier_call = nf_tables_flowtable_event,
5981};
5982
5983static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
5984 int event)
5985{
5986 struct nlmsghdr *nlh = nlmsg_hdr(skb);
5987 struct sk_buff *skb2;
5988 int err;
5989
5990 if (nlmsg_report(nlh) &&
5991 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5992 return;
5993
5994 skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5995 if (skb2 == NULL)
5996 goto err;
5997
5998 err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
5999 nlh->nlmsg_seq);
6000 if (err < 0) {
6001 kfree_skb(skb2);
6002 goto err;
6003 }
6004
6005 nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
6006 nlmsg_report(nlh), GFP_KERNEL);
6007 return;
6008err:
6009 nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
6010 -ENOBUFS);
6011}
6012
6013static int nf_tables_getgen(struct net *net, struct sock *nlsk,
6014 struct sk_buff *skb, const struct nlmsghdr *nlh,
6015 const struct nlattr * const nla[],
6016 struct netlink_ext_ack *extack)
6017{
6018 struct sk_buff *skb2;
6019 int err;
6020
6021 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
6022 if (skb2 == NULL)
6023 return -ENOMEM;
6024
6025 err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
6026 nlh->nlmsg_seq);
6027 if (err < 0)
6028 goto err;
6029
6030 return nlmsg_unicast(nlsk, skb2, NETLINK_CB(skb).portid);
6031err:
6032 kfree_skb(skb2);
6033 return err;
6034}
6035
6036static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
6037 [NFT_MSG_NEWTABLE] = {
6038 .call_batch = nf_tables_newtable,
6039 .attr_count = NFTA_TABLE_MAX,
6040 .policy = nft_table_policy,
6041 },
6042 [NFT_MSG_GETTABLE] = {
6043 .call_rcu = nf_tables_gettable,
6044 .attr_count = NFTA_TABLE_MAX,
6045 .policy = nft_table_policy,
6046 },
6047 [NFT_MSG_DELTABLE] = {
6048 .call_batch = nf_tables_deltable,
6049 .attr_count = NFTA_TABLE_MAX,
6050 .policy = nft_table_policy,
6051 },
6052 [NFT_MSG_NEWCHAIN] = {
6053 .call_batch = nf_tables_newchain,
6054 .attr_count = NFTA_CHAIN_MAX,
6055 .policy = nft_chain_policy,
6056 },
6057 [NFT_MSG_GETCHAIN] = {
6058 .call_rcu = nf_tables_getchain,
6059 .attr_count = NFTA_CHAIN_MAX,
6060 .policy = nft_chain_policy,
6061 },
6062 [NFT_MSG_DELCHAIN] = {
6063 .call_batch = nf_tables_delchain,
6064 .attr_count = NFTA_CHAIN_MAX,
6065 .policy = nft_chain_policy,
6066 },
6067 [NFT_MSG_NEWRULE] = {
6068 .call_batch = nf_tables_newrule,
6069 .attr_count = NFTA_RULE_MAX,
6070 .policy = nft_rule_policy,
6071 },
6072 [NFT_MSG_GETRULE] = {
6073 .call_rcu = nf_tables_getrule,
6074 .attr_count = NFTA_RULE_MAX,
6075 .policy = nft_rule_policy,
6076 },
6077 [NFT_MSG_DELRULE] = {
6078 .call_batch = nf_tables_delrule,
6079 .attr_count = NFTA_RULE_MAX,
6080 .policy = nft_rule_policy,
6081 },
6082 [NFT_MSG_NEWSET] = {
6083 .call_batch = nf_tables_newset,
6084 .attr_count = NFTA_SET_MAX,
6085 .policy = nft_set_policy,
6086 },
6087 [NFT_MSG_GETSET] = {
6088 .call_rcu = nf_tables_getset,
6089 .attr_count = NFTA_SET_MAX,
6090 .policy = nft_set_policy,
6091 },
6092 [NFT_MSG_DELSET] = {
6093 .call_batch = nf_tables_delset,
6094 .attr_count = NFTA_SET_MAX,
6095 .policy = nft_set_policy,
6096 },
6097 [NFT_MSG_NEWSETELEM] = {
6098 .call_batch = nf_tables_newsetelem,
6099 .attr_count = NFTA_SET_ELEM_LIST_MAX,
6100 .policy = nft_set_elem_list_policy,
6101 },
6102 [NFT_MSG_GETSETELEM] = {
6103 .call_rcu = nf_tables_getsetelem,
6104 .attr_count = NFTA_SET_ELEM_LIST_MAX,
6105 .policy = nft_set_elem_list_policy,
6106 },
6107 [NFT_MSG_DELSETELEM] = {
6108 .call_batch = nf_tables_delsetelem,
6109 .attr_count = NFTA_SET_ELEM_LIST_MAX,
6110 .policy = nft_set_elem_list_policy,
6111 },
6112 [NFT_MSG_GETGEN] = {
6113 .call_rcu = nf_tables_getgen,
6114 },
6115 [NFT_MSG_NEWOBJ] = {
6116 .call_batch = nf_tables_newobj,
6117 .attr_count = NFTA_OBJ_MAX,
6118 .policy = nft_obj_policy,
6119 },
6120 [NFT_MSG_GETOBJ] = {
6121 .call_rcu = nf_tables_getobj,
6122 .attr_count = NFTA_OBJ_MAX,
6123 .policy = nft_obj_policy,
6124 },
6125 [NFT_MSG_DELOBJ] = {
6126 .call_batch = nf_tables_delobj,
6127 .attr_count = NFTA_OBJ_MAX,
6128 .policy = nft_obj_policy,
6129 },
6130 [NFT_MSG_GETOBJ_RESET] = {
6131 .call_rcu = nf_tables_getobj,
6132 .attr_count = NFTA_OBJ_MAX,
6133 .policy = nft_obj_policy,
6134 },
6135 [NFT_MSG_NEWFLOWTABLE] = {
6136 .call_batch = nf_tables_newflowtable,
6137 .attr_count = NFTA_FLOWTABLE_MAX,
6138 .policy = nft_flowtable_policy,
6139 },
6140 [NFT_MSG_GETFLOWTABLE] = {
6141 .call_rcu = nf_tables_getflowtable,
6142 .attr_count = NFTA_FLOWTABLE_MAX,
6143 .policy = nft_flowtable_policy,
6144 },
6145 [NFT_MSG_DELFLOWTABLE] = {
6146 .call_batch = nf_tables_delflowtable,
6147 .attr_count = NFTA_FLOWTABLE_MAX,
6148 .policy = nft_flowtable_policy,
6149 },
6150};
6151
6152static int nf_tables_validate(struct net *net)
6153{
6154 struct nft_table *table;
6155
6156 switch (net->nft.validate_state) {
6157 case NFT_VALIDATE_SKIP:
6158 break;
6159 case NFT_VALIDATE_NEED:
6160 nft_validate_state_update(net, NFT_VALIDATE_DO);
6161 /* fall through */
6162 case NFT_VALIDATE_DO:
6163 list_for_each_entry(table, &net->nft.tables, list) {
6164 if (nft_table_validate(net, table) < 0)
6165 return -EAGAIN;
6166 }
6167 break;
6168 }
6169
6170 return 0;
6171}
6172
6173static void nft_chain_commit_update(struct nft_trans *trans)
6174{
6175 struct nft_base_chain *basechain;
6176
6177 if (nft_trans_chain_name(trans)) {
6178 rhltable_remove(&trans->ctx.table->chains_ht,
6179 &trans->ctx.chain->rhlhead,
6180 nft_chain_ht_params);
6181 swap(trans->ctx.chain->name, nft_trans_chain_name(trans));
6182 rhltable_insert_key(&trans->ctx.table->chains_ht,
6183 trans->ctx.chain->name,
6184 &trans->ctx.chain->rhlhead,
6185 nft_chain_ht_params);
6186 }
6187
6188 if (!nft_is_base_chain(trans->ctx.chain))
6189 return;
6190
6191 basechain = nft_base_chain(trans->ctx.chain);
6192 nft_chain_stats_replace(basechain, nft_trans_chain_stats(trans));
6193
6194 switch (nft_trans_chain_policy(trans)) {
6195 case NF_DROP:
6196 case NF_ACCEPT:
6197 basechain->policy = nft_trans_chain_policy(trans);
6198 break;
6199 }
6200}
6201
6202static void nft_commit_release(struct nft_trans *trans)
6203{
6204 switch (trans->msg_type) {
6205 case NFT_MSG_DELTABLE:
6206 nf_tables_table_destroy(&trans->ctx);
6207 break;
6208 case NFT_MSG_NEWCHAIN:
6209 kfree(nft_trans_chain_name(trans));
6210 break;
6211 case NFT_MSG_DELCHAIN:
6212 nf_tables_chain_destroy(&trans->ctx);
6213 break;
6214 case NFT_MSG_DELRULE:
6215 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
6216 break;
6217 case NFT_MSG_DELSET:
6218 nft_set_destroy(nft_trans_set(trans));
6219 break;
6220 case NFT_MSG_DELSETELEM:
6221 nf_tables_set_elem_destroy(&trans->ctx,
6222 nft_trans_elem_set(trans),
6223 nft_trans_elem(trans).priv);
6224 break;
6225 case NFT_MSG_DELOBJ:
6226 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
6227 break;
6228 case NFT_MSG_DELFLOWTABLE:
6229 nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
6230 break;
6231 }
6232
6233 if (trans->put_net)
6234 put_net(trans->ctx.net);
6235
6236 kfree(trans);
6237}
6238
6239static void nf_tables_trans_destroy_work(struct work_struct *w)
6240{
6241 struct nft_trans *trans, *next;
6242 LIST_HEAD(head);
6243
6244 spin_lock(&nf_tables_destroy_list_lock);
6245 list_splice_init(&nf_tables_destroy_list, &head);
6246 spin_unlock(&nf_tables_destroy_list_lock);
6247
6248 if (list_empty(&head))
6249 return;
6250
6251 synchronize_rcu();
6252
6253 list_for_each_entry_safe(trans, next, &head, list) {
6254 list_del(&trans->list);
6255 nft_commit_release(trans);
6256 }
6257}
6258
6259static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
6260{
6261 struct nft_rule *rule;
6262 unsigned int alloc = 0;
6263 int i;
6264
6265 /* already handled or inactive chain? */
6266 if (chain->rules_next || !nft_is_active_next(net, chain))
6267 return 0;
6268
6269 rule = list_entry(&chain->rules, struct nft_rule, list);
6270 i = 0;
6271
6272 list_for_each_entry_continue(rule, &chain->rules, list) {
6273 if (nft_is_active_next(net, rule))
6274 alloc++;
6275 }
6276
6277 chain->rules_next = nf_tables_chain_alloc_rules(chain, alloc);
6278 if (!chain->rules_next)
6279 return -ENOMEM;
6280
6281 list_for_each_entry_continue(rule, &chain->rules, list) {
6282 if (nft_is_active_next(net, rule))
6283 chain->rules_next[i++] = rule;
6284 }
6285
6286 chain->rules_next[i] = NULL;
6287 return 0;
6288}
6289
6290static void nf_tables_commit_chain_prepare_cancel(struct net *net)
6291{
6292 struct nft_trans *trans, *next;
6293
6294 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
6295 struct nft_chain *chain = trans->ctx.chain;
6296
6297 if (trans->msg_type == NFT_MSG_NEWRULE ||
6298 trans->msg_type == NFT_MSG_DELRULE) {
6299 kvfree(chain->rules_next);
6300 chain->rules_next = NULL;
6301 }
6302 }
6303}
6304
6305static void __nf_tables_commit_chain_free_rules_old(struct rcu_head *h)
6306{
6307 struct nft_rules_old *o = container_of(h, struct nft_rules_old, h);
6308
6309 kvfree(o->start);
6310}
6311
6312static void nf_tables_commit_chain_free_rules_old(struct nft_rule **rules)
6313{
6314 struct nft_rule **r = rules;
6315 struct nft_rules_old *old;
6316
6317 while (*r)
6318 r++;
6319
6320 r++; /* rcu_head is after end marker */
6321 old = (void *) r;
6322 old->start = rules;
6323
6324 call_rcu(&old->h, __nf_tables_commit_chain_free_rules_old);
6325}
6326
6327static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
6328{
6329 struct nft_rule **g0, **g1;
6330 bool next_genbit;
6331
6332 next_genbit = nft_gencursor_next(net);
6333
6334 g0 = rcu_dereference_protected(chain->rules_gen_0,
6335 lockdep_commit_lock_is_held(net));
6336 g1 = rcu_dereference_protected(chain->rules_gen_1,
6337 lockdep_commit_lock_is_held(net));
6338
6339 /* No changes to this chain? */
6340 if (chain->rules_next == NULL) {
6341 /* chain had no change in last or next generation */
6342 if (g0 == g1)
6343 return;
6344 /*
6345 * chain had no change in this generation; make sure next
6346 * one uses same rules as current generation.
6347 */
6348 if (next_genbit) {
6349 rcu_assign_pointer(chain->rules_gen_1, g0);
6350 nf_tables_commit_chain_free_rules_old(g1);
6351 } else {
6352 rcu_assign_pointer(chain->rules_gen_0, g1);
6353 nf_tables_commit_chain_free_rules_old(g0);
6354 }
6355
6356 return;
6357 }
6358
6359 if (next_genbit)
6360 rcu_assign_pointer(chain->rules_gen_1, chain->rules_next);
6361 else
6362 rcu_assign_pointer(chain->rules_gen_0, chain->rules_next);
6363
6364 chain->rules_next = NULL;
6365
6366 if (g0 == g1)
6367 return;
6368
6369 if (next_genbit)
6370 nf_tables_commit_chain_free_rules_old(g1);
6371 else
6372 nf_tables_commit_chain_free_rules_old(g0);
6373}
6374
6375static void nft_chain_del(struct nft_chain *chain)
6376{
6377 struct nft_table *table = chain->table;
6378
6379 WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
6380 nft_chain_ht_params));
6381 list_del_rcu(&chain->list);
6382}
6383
6384static void nf_tables_commit_release(struct net *net)
6385{
6386 struct nft_trans *trans;
6387
6388 /* all side effects have to be made visible.
6389 * For example, if a chain named 'foo' has been deleted, a
6390 * new transaction must not find it anymore.
6391 *
6392 * Memory reclaim happens asynchronously from work queue
6393 * to prevent expensive synchronize_rcu() in commit phase.
6394 */
6395 if (list_empty(&net->nft.commit_list)) {
6396 mutex_unlock(&net->nft.commit_mutex);
6397 return;
6398 }
6399
6400 trans = list_last_entry(&net->nft.commit_list,
6401 struct nft_trans, list);
6402 get_net(trans->ctx.net);
6403 WARN_ON_ONCE(trans->put_net);
6404
6405 trans->put_net = true;
6406 spin_lock(&nf_tables_destroy_list_lock);
6407 list_splice_tail_init(&net->nft.commit_list, &nf_tables_destroy_list);
6408 spin_unlock(&nf_tables_destroy_list_lock);
6409
6410 mutex_unlock(&net->nft.commit_mutex);
6411
6412 schedule_work(&trans_destroy_work);
6413}
6414
6415static int nf_tables_commit(struct net *net, struct sk_buff *skb)
6416{
6417 struct nft_trans *trans, *next;
6418 struct nft_trans_elem *te;
6419 struct nft_chain *chain;
6420 struct nft_table *table;
6421
6422 /* 0. Validate ruleset, otherwise roll back for error reporting. */
6423 if (nf_tables_validate(net) < 0)
6424 return -EAGAIN;
6425
6426 /* 1. Allocate space for next generation rules_gen_X[] */
6427 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
6428 int ret;
6429
6430 if (trans->msg_type == NFT_MSG_NEWRULE ||
6431 trans->msg_type == NFT_MSG_DELRULE) {
6432 chain = trans->ctx.chain;
6433
6434 ret = nf_tables_commit_chain_prepare(net, chain);
6435 if (ret < 0) {
6436 nf_tables_commit_chain_prepare_cancel(net);
6437 return ret;
6438 }
6439 }
6440 }
6441
6442 /* step 2. Make rules_gen_X visible to packet path */
6443 list_for_each_entry(table, &net->nft.tables, list) {
6444 list_for_each_entry(chain, &table->chains, list)
6445 nf_tables_commit_chain(net, chain);
6446 }
6447
6448 /*
6449 * Bump generation counter, invalidate any dump in progress.
6450 * Cannot fail after this point.
6451 */
6452 while (++net->nft.base_seq == 0);
6453
6454 /* step 3. Start new generation, rules_gen_X now in use. */
6455 net->nft.gencursor = nft_gencursor_next(net);
6456
6457 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
6458 switch (trans->msg_type) {
6459 case NFT_MSG_NEWTABLE:
6460 if (nft_trans_table_update(trans)) {
6461 if (!nft_trans_table_enable(trans)) {
6462 nf_tables_table_disable(net,
6463 trans->ctx.table);
6464 trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
6465 }
6466 } else {
6467 nft_clear(net, trans->ctx.table);
6468 }
6469 nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
6470 nft_trans_destroy(trans);
6471 break;
6472 case NFT_MSG_DELTABLE:
6473 list_del_rcu(&trans->ctx.table->list);
6474 nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
6475 break;
6476 case NFT_MSG_NEWCHAIN:
6477 if (nft_trans_chain_update(trans)) {
6478 nft_chain_commit_update(trans);
6479 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
6480 /* trans destroyed after rcu grace period */
6481 } else {
6482 nft_clear(net, trans->ctx.chain);
6483 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
6484 nft_trans_destroy(trans);
6485 }
6486 break;
6487 case NFT_MSG_DELCHAIN:
6488 nft_chain_del(trans->ctx.chain);
6489 nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
6490 nf_tables_unregister_hook(trans->ctx.net,
6491 trans->ctx.table,
6492 trans->ctx.chain);
6493 break;
6494 case NFT_MSG_NEWRULE:
6495 nft_clear(trans->ctx.net, nft_trans_rule(trans));
6496 nf_tables_rule_notify(&trans->ctx,
6497 nft_trans_rule(trans),
6498 NFT_MSG_NEWRULE);
6499 nft_trans_destroy(trans);
6500 break;
6501 case NFT_MSG_DELRULE:
6502 list_del_rcu(&nft_trans_rule(trans)->list);
6503 nf_tables_rule_notify(&trans->ctx,
6504 nft_trans_rule(trans),
6505 NFT_MSG_DELRULE);
6506 break;
6507 case NFT_MSG_NEWSET:
6508 nft_clear(net, nft_trans_set(trans));
6509 /* This avoids hitting -EBUSY when deleting the table
6510 * from the transaction.
6511 */
6512 if (nft_set_is_anonymous(nft_trans_set(trans)) &&
6513 !list_empty(&nft_trans_set(trans)->bindings))
6514 trans->ctx.table->use--;
6515
6516 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
6517 NFT_MSG_NEWSET, GFP_KERNEL);
6518 nft_trans_destroy(trans);
6519 break;
6520 case NFT_MSG_DELSET:
6521 list_del_rcu(&nft_trans_set(trans)->list);
6522 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
6523 NFT_MSG_DELSET, GFP_KERNEL);
6524 break;
6525 case NFT_MSG_NEWSETELEM:
6526 te = (struct nft_trans_elem *)trans->data;
6527
6528 te->set->ops->activate(net, te->set, &te->elem);
6529 nf_tables_setelem_notify(&trans->ctx, te->set,
6530 &te->elem,
6531 NFT_MSG_NEWSETELEM, 0);
6532 nft_trans_destroy(trans);
6533 break;
6534 case NFT_MSG_DELSETELEM:
6535 te = (struct nft_trans_elem *)trans->data;
6536
6537 nf_tables_setelem_notify(&trans->ctx, te->set,
6538 &te->elem,
6539 NFT_MSG_DELSETELEM, 0);
6540 te->set->ops->remove(net, te->set, &te->elem);
6541 atomic_dec(&te->set->nelems);
6542 te->set->ndeact--;
6543 break;
6544 case NFT_MSG_NEWOBJ:
6545 nft_clear(net, nft_trans_obj(trans));
6546 nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
6547 NFT_MSG_NEWOBJ);
6548 nft_trans_destroy(trans);
6549 break;
6550 case NFT_MSG_DELOBJ:
6551 list_del_rcu(&nft_trans_obj(trans)->list);
6552 nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
6553 NFT_MSG_DELOBJ);
6554 break;
6555 case NFT_MSG_NEWFLOWTABLE:
6556 nft_clear(net, nft_trans_flowtable(trans));
6557 nf_tables_flowtable_notify(&trans->ctx,
6558 nft_trans_flowtable(trans),
6559 NFT_MSG_NEWFLOWTABLE);
6560 nft_trans_destroy(trans);
6561 break;
6562 case NFT_MSG_DELFLOWTABLE:
6563 list_del_rcu(&nft_trans_flowtable(trans)->list);
6564 nf_tables_flowtable_notify(&trans->ctx,
6565 nft_trans_flowtable(trans),
6566 NFT_MSG_DELFLOWTABLE);
6567 nft_unregister_flowtable_net_hooks(net,
6568 nft_trans_flowtable(trans));
6569 break;
6570 }
6571 }
6572
6573 nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
6574 nf_tables_commit_release(net);
6575
6576 return 0;
6577}
6578
6579static void nf_tables_abort_release(struct nft_trans *trans)
6580{
6581 switch (trans->msg_type) {
6582 case NFT_MSG_NEWTABLE:
6583 nf_tables_table_destroy(&trans->ctx);
6584 break;
6585 case NFT_MSG_NEWCHAIN:
6586 nf_tables_chain_destroy(&trans->ctx);
6587 break;
6588 case NFT_MSG_NEWRULE:
6589 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
6590 break;
6591 case NFT_MSG_NEWSET:
6592 nft_set_destroy(nft_trans_set(trans));
6593 break;
6594 case NFT_MSG_NEWSETELEM:
6595 nft_set_elem_destroy(nft_trans_elem_set(trans),
6596 nft_trans_elem(trans).priv, true);
6597 break;
6598 case NFT_MSG_NEWOBJ:
6599 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
6600 break;
6601 case NFT_MSG_NEWFLOWTABLE:
6602 nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
6603 break;
6604 }
6605 kfree(trans);
6606}
6607
6608static int __nf_tables_abort(struct net *net)
6609{
6610 struct nft_trans *trans, *next;
6611 struct nft_trans_elem *te;
6612
6613 list_for_each_entry_safe_reverse(trans, next, &net->nft.commit_list,
6614 list) {
6615 switch (trans->msg_type) {
6616 case NFT_MSG_NEWTABLE:
6617 if (nft_trans_table_update(trans)) {
6618 if (nft_trans_table_enable(trans)) {
6619 nf_tables_table_disable(net,
6620 trans->ctx.table);
6621 trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
6622 }
6623 nft_trans_destroy(trans);
6624 } else {
6625 list_del_rcu(&trans->ctx.table->list);
6626 }
6627 break;
6628 case NFT_MSG_DELTABLE:
6629 nft_clear(trans->ctx.net, trans->ctx.table);
6630 nft_trans_destroy(trans);
6631 break;
6632 case NFT_MSG_NEWCHAIN:
6633 if (nft_trans_chain_update(trans)) {
6634 free_percpu(nft_trans_chain_stats(trans));
6635 kfree(nft_trans_chain_name(trans));
6636 nft_trans_destroy(trans);
6637 } else {
6638 trans->ctx.table->use--;
6639 nft_chain_del(trans->ctx.chain);
6640 nf_tables_unregister_hook(trans->ctx.net,
6641 trans->ctx.table,
6642 trans->ctx.chain);
6643 }
6644 break;
6645 case NFT_MSG_DELCHAIN:
6646 trans->ctx.table->use++;
6647 nft_clear(trans->ctx.net, trans->ctx.chain);
6648 nft_trans_destroy(trans);
6649 break;
6650 case NFT_MSG_NEWRULE:
6651 trans->ctx.chain->use--;
6652 list_del_rcu(&nft_trans_rule(trans)->list);
6653 nft_rule_expr_deactivate(&trans->ctx, nft_trans_rule(trans));
6654 break;
6655 case NFT_MSG_DELRULE:
6656 trans->ctx.chain->use++;
6657 nft_clear(trans->ctx.net, nft_trans_rule(trans));
6658 nft_rule_expr_activate(&trans->ctx, nft_trans_rule(trans));
6659 nft_trans_destroy(trans);
6660 break;
6661 case NFT_MSG_NEWSET:
6662 trans->ctx.table->use--;
6663 list_del_rcu(&nft_trans_set(trans)->list);
6664 break;
6665 case NFT_MSG_DELSET:
6666 trans->ctx.table->use++;
6667 nft_clear(trans->ctx.net, nft_trans_set(trans));
6668 nft_trans_destroy(trans);
6669 break;
6670 case NFT_MSG_NEWSETELEM:
6671 te = (struct nft_trans_elem *)trans->data;
6672
6673 te->set->ops->remove(net, te->set, &te->elem);
6674 atomic_dec(&te->set->nelems);
6675 break;
6676 case NFT_MSG_DELSETELEM:
6677 te = (struct nft_trans_elem *)trans->data;
6678
6679 nft_set_elem_activate(net, te->set, &te->elem);
6680 te->set->ops->activate(net, te->set, &te->elem);
6681 te->set->ndeact--;
6682
6683 nft_trans_destroy(trans);
6684 break;
6685 case NFT_MSG_NEWOBJ:
6686 trans->ctx.table->use--;
6687 list_del_rcu(&nft_trans_obj(trans)->list);
6688 break;
6689 case NFT_MSG_DELOBJ:
6690 trans->ctx.table->use++;
6691 nft_clear(trans->ctx.net, nft_trans_obj(trans));
6692 nft_trans_destroy(trans);
6693 break;
6694 case NFT_MSG_NEWFLOWTABLE:
6695 trans->ctx.table->use--;
6696 list_del_rcu(&nft_trans_flowtable(trans)->list);
6697 nft_unregister_flowtable_net_hooks(net,
6698 nft_trans_flowtable(trans));
6699 break;
6700 case NFT_MSG_DELFLOWTABLE:
6701 trans->ctx.table->use++;
6702 nft_clear(trans->ctx.net, nft_trans_flowtable(trans));
6703 nft_trans_destroy(trans);
6704 break;
6705 }
6706 }
6707
6708 synchronize_rcu();
6709
6710 list_for_each_entry_safe_reverse(trans, next,
6711 &net->nft.commit_list, list) {
6712 list_del(&trans->list);
6713 nf_tables_abort_release(trans);
6714 }
6715
6716 return 0;
6717}
6718
6719static void nf_tables_cleanup(struct net *net)
6720{
6721 nft_validate_state_update(net, NFT_VALIDATE_SKIP);
6722}
6723
6724static int nf_tables_abort(struct net *net, struct sk_buff *skb)
6725{
6726 int ret = __nf_tables_abort(net);
6727
6728 mutex_unlock(&net->nft.commit_mutex);
6729
6730 return ret;
6731}
6732
6733static bool nf_tables_valid_genid(struct net *net, u32 genid)
6734{
6735 bool genid_ok;
6736
6737 mutex_lock(&net->nft.commit_mutex);
6738
6739 genid_ok = genid == 0 || net->nft.base_seq == genid;
6740 if (!genid_ok)
6741 mutex_unlock(&net->nft.commit_mutex);
6742
6743 /* else, commit mutex has to be released by commit or abort function */
6744 return genid_ok;
6745}
6746
6747static const struct nfnetlink_subsystem nf_tables_subsys = {
6748 .name = "nf_tables",
6749 .subsys_id = NFNL_SUBSYS_NFTABLES,
6750 .cb_count = NFT_MSG_MAX,
6751 .cb = nf_tables_cb,
6752 .commit = nf_tables_commit,
6753 .abort = nf_tables_abort,
6754 .cleanup = nf_tables_cleanup,
6755 .valid_genid = nf_tables_valid_genid,
6756 .owner = THIS_MODULE,
6757};
6758
6759int nft_chain_validate_dependency(const struct nft_chain *chain,
6760 enum nft_chain_types type)
6761{
6762 const struct nft_base_chain *basechain;
6763
6764 if (nft_is_base_chain(chain)) {
6765 basechain = nft_base_chain(chain);
6766 if (basechain->type->type != type)
6767 return -EOPNOTSUPP;
6768 }
6769 return 0;
6770}
6771EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
6772
6773int nft_chain_validate_hooks(const struct nft_chain *chain,
6774 unsigned int hook_flags)
6775{
6776 struct nft_base_chain *basechain;
6777
6778 if (nft_is_base_chain(chain)) {
6779 basechain = nft_base_chain(chain);
6780
6781 if ((1 << basechain->ops.hooknum) & hook_flags)
6782 return 0;
6783
6784 return -EOPNOTSUPP;
6785 }
6786
6787 return 0;
6788}
6789EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
6790
6791/*
6792 * Loop detection - walk through the ruleset beginning at the destination chain
6793 * of a new jump until either the source chain is reached (loop) or all
6794 * reachable chains have been traversed.
6795 *
6796 * The loop check is performed whenever a new jump verdict is added to an
6797 * expression or verdict map or a verdict map is bound to a new chain.
6798 */
6799
6800static int nf_tables_check_loops(const struct nft_ctx *ctx,
6801 const struct nft_chain *chain);
6802
6803static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
6804 struct nft_set *set,
6805 const struct nft_set_iter *iter,
6806 struct nft_set_elem *elem)
6807{
6808 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6809 const struct nft_data *data;
6810
6811 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
6812 *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
6813 return 0;
6814
6815 data = nft_set_ext_data(ext);
6816 switch (data->verdict.code) {
6817 case NFT_JUMP:
6818 case NFT_GOTO:
6819 return nf_tables_check_loops(ctx, data->verdict.chain);
6820 default:
6821 return 0;
6822 }
6823}
6824
6825static int nf_tables_check_loops(const struct nft_ctx *ctx,
6826 const struct nft_chain *chain)
6827{
6828 const struct nft_rule *rule;
6829 const struct nft_expr *expr, *last;
6830 struct nft_set *set;
6831 struct nft_set_binding *binding;
6832 struct nft_set_iter iter;
6833
6834 if (ctx->chain == chain)
6835 return -ELOOP;
6836
6837 list_for_each_entry(rule, &chain->rules, list) {
6838 nft_rule_for_each_expr(expr, last, rule) {
6839 struct nft_immediate_expr *priv;
6840 const struct nft_data *data;
6841 int err;
6842
6843 if (strcmp(expr->ops->type->name, "immediate"))
6844 continue;
6845
6846 priv = nft_expr_priv(expr);
6847 if (priv->dreg != NFT_REG_VERDICT)
6848 continue;
6849
6850 data = &priv->data;
6851 switch (data->verdict.code) {
6852 case NFT_JUMP:
6853 case NFT_GOTO:
6854 err = nf_tables_check_loops(ctx,
6855 data->verdict.chain);
6856 if (err < 0)
6857 return err;
6858 default:
6859 break;
6860 }
6861 }
6862 }
6863
6864 list_for_each_entry(set, &ctx->table->sets, list) {
6865 if (!nft_is_active_next(ctx->net, set))
6866 continue;
6867 if (!(set->flags & NFT_SET_MAP) ||
6868 set->dtype != NFT_DATA_VERDICT)
6869 continue;
6870
6871 list_for_each_entry(binding, &set->bindings, list) {
6872 if (!(binding->flags & NFT_SET_MAP) ||
6873 binding->chain != chain)
6874 continue;
6875
6876 iter.genmask = nft_genmask_next(ctx->net);
6877 iter.skip = 0;
6878 iter.count = 0;
6879 iter.err = 0;
6880 iter.fn = nf_tables_loop_check_setelem;
6881
6882 set->ops->walk(ctx, set, &iter);
6883 if (iter.err < 0)
6884 return iter.err;
6885 }
6886 }
6887
6888 return 0;
6889}
6890
6891/**
6892 * nft_parse_u32_check - fetch u32 attribute and check for maximum value
6893 *
6894 * @attr: netlink attribute to fetch value from
6895 * @max: maximum value to be stored in dest
6896 * @dest: pointer to the variable
6897 *
6898 * Parse, check and store a given u32 netlink attribute into variable.
6899 * This function returns -ERANGE if the value goes over maximum value.
6900 * Otherwise a 0 is returned and the attribute value is stored in the
6901 * destination variable.
6902 */
6903int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
6904{
6905 u32 val;
6906
6907 val = ntohl(nla_get_be32(attr));
6908 if (val > max)
6909 return -ERANGE;
6910
6911 *dest = val;
6912 return 0;
6913}
6914EXPORT_SYMBOL_GPL(nft_parse_u32_check);
6915
6916/**
6917 * nft_parse_register - parse a register value from a netlink attribute
6918 *
6919 * @attr: netlink attribute
6920 *
6921 * Parse and translate a register value from a netlink attribute.
6922 * Registers used to be 128 bit wide, these register numbers will be
6923 * mapped to the corresponding 32 bit register numbers.
6924 */
6925unsigned int nft_parse_register(const struct nlattr *attr)
6926{
6927 unsigned int reg;
6928
6929 reg = ntohl(nla_get_be32(attr));
6930 switch (reg) {
6931 case NFT_REG_VERDICT...NFT_REG_4:
6932 return reg * NFT_REG_SIZE / NFT_REG32_SIZE;
6933 default:
6934 return reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
6935 }
6936}
6937EXPORT_SYMBOL_GPL(nft_parse_register);
6938
6939/**
6940 * nft_dump_register - dump a register value to a netlink attribute
6941 *
6942 * @skb: socket buffer
6943 * @attr: attribute number
6944 * @reg: register number
6945 *
6946 * Construct a netlink attribute containing the register number. For
6947 * compatibility reasons, register numbers being a multiple of 4 are
6948 * translated to the corresponding 128 bit register numbers.
6949 */
6950int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
6951{
6952 if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
6953 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
6954 else
6955 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
6956
6957 return nla_put_be32(skb, attr, htonl(reg));
6958}
6959EXPORT_SYMBOL_GPL(nft_dump_register);
6960
6961/**
6962 * nft_validate_register_load - validate a load from a register
6963 *
6964 * @reg: the register number
6965 * @len: the length of the data
6966 *
6967 * Validate that the input register is one of the general purpose
6968 * registers and that the length of the load is within the bounds.
6969 */
6970int nft_validate_register_load(enum nft_registers reg, unsigned int len)
6971{
6972 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
6973 return -EINVAL;
6974 if (len == 0)
6975 return -EINVAL;
6976 if (reg * NFT_REG32_SIZE + len > FIELD_SIZEOF(struct nft_regs, data))
6977 return -ERANGE;
6978
6979 return 0;
6980}
6981EXPORT_SYMBOL_GPL(nft_validate_register_load);
6982
6983/**
6984 * nft_validate_register_store - validate an expressions' register store
6985 *
6986 * @ctx: context of the expression performing the load
6987 * @reg: the destination register number
6988 * @data: the data to load
6989 * @type: the data type
6990 * @len: the length of the data
6991 *
6992 * Validate that a data load uses the appropriate data type for
6993 * the destination register and the length is within the bounds.
6994 * A value of NULL for the data means that its runtime gathered
6995 * data.
6996 */
6997int nft_validate_register_store(const struct nft_ctx *ctx,
6998 enum nft_registers reg,
6999 const struct nft_data *data,
7000 enum nft_data_types type, unsigned int len)
7001{
7002 int err;
7003
7004 switch (reg) {
7005 case NFT_REG_VERDICT:
7006 if (type != NFT_DATA_VERDICT)
7007 return -EINVAL;
7008
7009 if (data != NULL &&
7010 (data->verdict.code == NFT_GOTO ||
7011 data->verdict.code == NFT_JUMP)) {
7012 err = nf_tables_check_loops(ctx, data->verdict.chain);
7013 if (err < 0)
7014 return err;
7015 }
7016
7017 return 0;
7018 default:
7019 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
7020 return -EINVAL;
7021 if (len == 0)
7022 return -EINVAL;
7023 if (reg * NFT_REG32_SIZE + len >
7024 FIELD_SIZEOF(struct nft_regs, data))
7025 return -ERANGE;
7026
7027 if (data != NULL && type != NFT_DATA_VALUE)
7028 return -EINVAL;
7029 return 0;
7030 }
7031}
7032EXPORT_SYMBOL_GPL(nft_validate_register_store);
7033
7034static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
7035 [NFTA_VERDICT_CODE] = { .type = NLA_U32 },
7036 [NFTA_VERDICT_CHAIN] = { .type = NLA_STRING,
7037 .len = NFT_CHAIN_MAXNAMELEN - 1 },
7038};
7039
7040static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
7041 struct nft_data_desc *desc, const struct nlattr *nla)
7042{
7043 u8 genmask = nft_genmask_next(ctx->net);
7044 struct nlattr *tb[NFTA_VERDICT_MAX + 1];
7045 struct nft_chain *chain;
7046 int err;
7047
7048 err = nla_parse_nested(tb, NFTA_VERDICT_MAX, nla, nft_verdict_policy,
7049 NULL);
7050 if (err < 0)
7051 return err;
7052
7053 if (!tb[NFTA_VERDICT_CODE])
7054 return -EINVAL;
7055 data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
7056
7057 switch (data->verdict.code) {
7058 default:
7059 switch (data->verdict.code & NF_VERDICT_MASK) {
7060 case NF_ACCEPT:
7061 case NF_DROP:
7062 case NF_QUEUE:
7063 break;
7064 default:
7065 return -EINVAL;
7066 }
7067 /* fall through */
7068 case NFT_CONTINUE:
7069 case NFT_BREAK:
7070 case NFT_RETURN:
7071 break;
7072 case NFT_JUMP:
7073 case NFT_GOTO:
7074 if (!tb[NFTA_VERDICT_CHAIN])
7075 return -EINVAL;
7076 chain = nft_chain_lookup(ctx->net, ctx->table,
7077 tb[NFTA_VERDICT_CHAIN], genmask);
7078 if (IS_ERR(chain))
7079 return PTR_ERR(chain);
7080 if (nft_is_base_chain(chain))
7081 return -EOPNOTSUPP;
7082
7083 chain->use++;
7084 data->verdict.chain = chain;
7085 break;
7086 }
7087
7088 desc->len = sizeof(data->verdict);
7089 desc->type = NFT_DATA_VERDICT;
7090 return 0;
7091}
7092
7093static void nft_verdict_uninit(const struct nft_data *data)
7094{
7095 switch (data->verdict.code) {
7096 case NFT_JUMP:
7097 case NFT_GOTO:
7098 data->verdict.chain->use--;
7099 break;
7100 }
7101}
7102
7103int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
7104{
7105 struct nlattr *nest;
7106
7107 nest = nla_nest_start(skb, type);
7108 if (!nest)
7109 goto nla_put_failure;
7110
7111 if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
7112 goto nla_put_failure;
7113
7114 switch (v->code) {
7115 case NFT_JUMP:
7116 case NFT_GOTO:
7117 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
7118 v->chain->name))
7119 goto nla_put_failure;
7120 }
7121 nla_nest_end(skb, nest);
7122 return 0;
7123
7124nla_put_failure:
7125 return -1;
7126}
7127
7128static int nft_value_init(const struct nft_ctx *ctx,
7129 struct nft_data *data, unsigned int size,
7130 struct nft_data_desc *desc, const struct nlattr *nla)
7131{
7132 unsigned int len;
7133
7134 len = nla_len(nla);
7135 if (len == 0)
7136 return -EINVAL;
7137 if (len > size)
7138 return -EOVERFLOW;
7139
7140 nla_memcpy(data->data, nla, len);
7141 desc->type = NFT_DATA_VALUE;
7142 desc->len = len;
7143 return 0;
7144}
7145
7146static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
7147 unsigned int len)
7148{
7149 return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
7150}
7151
7152static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
7153 [NFTA_DATA_VALUE] = { .type = NLA_BINARY },
7154 [NFTA_DATA_VERDICT] = { .type = NLA_NESTED },
7155};
7156
7157/**
7158 * nft_data_init - parse nf_tables data netlink attributes
7159 *
7160 * @ctx: context of the expression using the data
7161 * @data: destination struct nft_data
7162 * @size: maximum data length
7163 * @desc: data description
7164 * @nla: netlink attribute containing data
7165 *
7166 * Parse the netlink data attributes and initialize a struct nft_data.
7167 * The type and length of data are returned in the data description.
7168 *
7169 * The caller can indicate that it only wants to accept data of type
7170 * NFT_DATA_VALUE by passing NULL for the ctx argument.
7171 */
7172int nft_data_init(const struct nft_ctx *ctx,
7173 struct nft_data *data, unsigned int size,
7174 struct nft_data_desc *desc, const struct nlattr *nla)
7175{
7176 struct nlattr *tb[NFTA_DATA_MAX + 1];
7177 int err;
7178
7179 err = nla_parse_nested(tb, NFTA_DATA_MAX, nla, nft_data_policy, NULL);
7180 if (err < 0)
7181 return err;
7182
7183 if (tb[NFTA_DATA_VALUE])
7184 return nft_value_init(ctx, data, size, desc,
7185 tb[NFTA_DATA_VALUE]);
7186 if (tb[NFTA_DATA_VERDICT] && ctx != NULL)
7187 return nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
7188 return -EINVAL;
7189}
7190EXPORT_SYMBOL_GPL(nft_data_init);
7191
7192/**
7193 * nft_data_release - release a nft_data item
7194 *
7195 * @data: struct nft_data to release
7196 * @type: type of data
7197 *
7198 * Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
7199 * all others need to be released by calling this function.
7200 */
7201void nft_data_release(const struct nft_data *data, enum nft_data_types type)
7202{
7203 if (type < NFT_DATA_VERDICT)
7204 return;
7205 switch (type) {
7206 case NFT_DATA_VERDICT:
7207 return nft_verdict_uninit(data);
7208 default:
7209 WARN_ON(1);
7210 }
7211}
7212EXPORT_SYMBOL_GPL(nft_data_release);
7213
7214int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
7215 enum nft_data_types type, unsigned int len)
7216{
7217 struct nlattr *nest;
7218 int err;
7219
7220 nest = nla_nest_start(skb, attr);
7221 if (nest == NULL)
7222 return -1;
7223
7224 switch (type) {
7225 case NFT_DATA_VALUE:
7226 err = nft_value_dump(skb, data, len);
7227 break;
7228 case NFT_DATA_VERDICT:
7229 err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
7230 break;
7231 default:
7232 err = -EINVAL;
7233 WARN_ON(1);
7234 }
7235
7236 nla_nest_end(skb, nest);
7237 return err;
7238}
7239EXPORT_SYMBOL_GPL(nft_data_dump);
7240
7241int __nft_release_basechain(struct nft_ctx *ctx)
7242{
7243 struct nft_rule *rule, *nr;
7244
7245 if (WARN_ON(!nft_is_base_chain(ctx->chain)))
7246 return 0;
7247
7248 nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
7249 list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
7250 list_del(&rule->list);
7251 ctx->chain->use--;
7252 nf_tables_rule_release(ctx, rule);
7253 }
7254 nft_chain_del(ctx->chain);
7255 ctx->table->use--;
7256 nf_tables_chain_destroy(ctx);
7257
7258 return 0;
7259}
7260EXPORT_SYMBOL_GPL(__nft_release_basechain);
7261
7262static void __nft_release_tables(struct net *net)
7263{
7264 struct nft_flowtable *flowtable, *nf;
7265 struct nft_table *table, *nt;
7266 struct nft_chain *chain, *nc;
7267 struct nft_object *obj, *ne;
7268 struct nft_rule *rule, *nr;
7269 struct nft_set *set, *ns;
7270 struct nft_ctx ctx = {
7271 .net = net,
7272 .family = NFPROTO_NETDEV,
7273 };
7274
7275 list_for_each_entry_safe(table, nt, &net->nft.tables, list) {
7276 ctx.family = table->family;
7277
7278 list_for_each_entry(chain, &table->chains, list)
7279 nf_tables_unregister_hook(net, table, chain);
7280 /* No packets are walking on these chains anymore. */
7281 ctx.table = table;
7282 list_for_each_entry(chain, &table->chains, list) {
7283 ctx.chain = chain;
7284 list_for_each_entry_safe(rule, nr, &chain->rules, list) {
7285 list_del(&rule->list);
7286 chain->use--;
7287 nf_tables_rule_release(&ctx, rule);
7288 }
7289 }
7290 list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
7291 list_del(&flowtable->list);
7292 table->use--;
7293 nf_tables_flowtable_destroy(flowtable);
7294 }
7295 list_for_each_entry_safe(set, ns, &table->sets, list) {
7296 list_del(&set->list);
7297 table->use--;
7298 nft_set_destroy(set);
7299 }
7300 list_for_each_entry_safe(obj, ne, &table->objects, list) {
7301 list_del(&obj->list);
7302 table->use--;
7303 nft_obj_destroy(&ctx, obj);
7304 }
7305 list_for_each_entry_safe(chain, nc, &table->chains, list) {
7306 ctx.chain = chain;
7307 nft_chain_del(chain);
7308 table->use--;
7309 nf_tables_chain_destroy(&ctx);
7310 }
7311 list_del(&table->list);
7312 nf_tables_table_destroy(&ctx);
7313 }
7314}
7315
7316static int __net_init nf_tables_init_net(struct net *net)
7317{
7318 INIT_LIST_HEAD(&net->nft.tables);
7319 INIT_LIST_HEAD(&net->nft.commit_list);
7320 mutex_init(&net->nft.commit_mutex);
7321 net->nft.base_seq = 1;
7322 net->nft.validate_state = NFT_VALIDATE_SKIP;
7323
7324 return 0;
7325}
7326
7327static void __net_exit nf_tables_exit_net(struct net *net)
7328{
7329 mutex_lock(&net->nft.commit_mutex);
7330 if (!list_empty(&net->nft.commit_list))
7331 __nf_tables_abort(net);
7332 __nft_release_tables(net);
7333 mutex_unlock(&net->nft.commit_mutex);
7334 WARN_ON_ONCE(!list_empty(&net->nft.tables));
7335}
7336
7337static struct pernet_operations nf_tables_net_ops = {
7338 .init = nf_tables_init_net,
7339 .exit = nf_tables_exit_net,
7340};
7341
7342static int __init nf_tables_module_init(void)
7343{
7344 int err;
7345
7346 spin_lock_init(&nf_tables_destroy_list_lock);
7347 err = register_pernet_subsys(&nf_tables_net_ops);
7348 if (err < 0)
7349 return err;
7350
7351 err = nft_chain_filter_init();
7352 if (err < 0)
7353 goto err1;
7354
7355 err = nf_tables_core_module_init();
7356 if (err < 0)
7357 goto err2;
7358
7359 err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
7360 if (err < 0)
7361 goto err3;
7362
7363 /* must be last */
7364 err = nfnetlink_subsys_register(&nf_tables_subsys);
7365 if (err < 0)
7366 goto err4;
7367
7368 return err;
7369err4:
7370 unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
7371err3:
7372 nf_tables_core_module_exit();
7373err2:
7374 nft_chain_filter_fini();
7375err1:
7376 unregister_pernet_subsys(&nf_tables_net_ops);
7377 return err;
7378}
7379
7380static void __exit nf_tables_module_exit(void)
7381{
7382 nfnetlink_subsys_unregister(&nf_tables_subsys);
7383 unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
7384 nft_chain_filter_fini();
7385 unregister_pernet_subsys(&nf_tables_net_ops);
7386 cancel_work_sync(&trans_destroy_work);
7387 rcu_barrier();
7388 nf_tables_core_module_exit();
7389}
7390
7391module_init(nf_tables_module_init);
7392module_exit(nf_tables_module_exit);
7393
7394MODULE_LICENSE("GPL");
7395MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
7396MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);
Ubuntu Jammy Linux kernel mirror