]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - net/netfilter/nf_tables_api.c
Merge tag 'iommu-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux
[mirror_ubuntu-hirsute-kernel.git] / net / netfilter / nf_tables_api.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (c) 2007-2009 Patrick McHardy <kaber@trash.net>
4 *
5 * Development of this code funded by Astaro AG (http://www.astaro.com/)
6 */
7
8 #include <linux/module.h>
9 #include <linux/init.h>
10 #include <linux/list.h>
11 #include <linux/skbuff.h>
12 #include <linux/netlink.h>
13 #include <linux/vmalloc.h>
14 #include <linux/rhashtable.h>
15 #include <linux/audit.h>
16 #include <linux/netfilter.h>
17 #include <linux/netfilter/nfnetlink.h>
18 #include <linux/netfilter/nf_tables.h>
19 #include <net/netfilter/nf_flow_table.h>
20 #include <net/netfilter/nf_tables_core.h>
21 #include <net/netfilter/nf_tables.h>
22 #include <net/netfilter/nf_tables_offload.h>
23 #include <net/net_namespace.h>
24 #include <net/sock.h>
25
26 #define NFT_MODULE_AUTOLOAD_LIMIT (MODULE_NAME_LEN - sizeof("nft-expr-255-"))
27
28 static LIST_HEAD(nf_tables_expressions);
29 static LIST_HEAD(nf_tables_objects);
30 static LIST_HEAD(nf_tables_flowtables);
31 static LIST_HEAD(nf_tables_destroy_list);
32 static DEFINE_SPINLOCK(nf_tables_destroy_list_lock);
33 static u64 table_handle;
34
35 enum {
36 NFT_VALIDATE_SKIP = 0,
37 NFT_VALIDATE_NEED,
38 NFT_VALIDATE_DO,
39 };
40
41 static struct rhltable nft_objname_ht;
42
43 static u32 nft_chain_hash(const void *data, u32 len, u32 seed);
44 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed);
45 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *, const void *);
46
47 static u32 nft_objname_hash(const void *data, u32 len, u32 seed);
48 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed);
49 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *, const void *);
50
51 static const struct rhashtable_params nft_chain_ht_params = {
52 .head_offset = offsetof(struct nft_chain, rhlhead),
53 .key_offset = offsetof(struct nft_chain, name),
54 .hashfn = nft_chain_hash,
55 .obj_hashfn = nft_chain_hash_obj,
56 .obj_cmpfn = nft_chain_hash_cmp,
57 .automatic_shrinking = true,
58 };
59
60 static const struct rhashtable_params nft_objname_ht_params = {
61 .head_offset = offsetof(struct nft_object, rhlhead),
62 .key_offset = offsetof(struct nft_object, key),
63 .hashfn = nft_objname_hash,
64 .obj_hashfn = nft_objname_hash_obj,
65 .obj_cmpfn = nft_objname_hash_cmp,
66 .automatic_shrinking = true,
67 };
68
69 static void nft_validate_state_update(struct net *net, u8 new_validate_state)
70 {
71 switch (net->nft.validate_state) {
72 case NFT_VALIDATE_SKIP:
73 WARN_ON_ONCE(new_validate_state == NFT_VALIDATE_DO);
74 break;
75 case NFT_VALIDATE_NEED:
76 break;
77 case NFT_VALIDATE_DO:
78 if (new_validate_state == NFT_VALIDATE_NEED)
79 return;
80 }
81
82 net->nft.validate_state = new_validate_state;
83 }
84 static void nf_tables_trans_destroy_work(struct work_struct *w);
85 static DECLARE_WORK(trans_destroy_work, nf_tables_trans_destroy_work);
86
87 static void nft_ctx_init(struct nft_ctx *ctx,
88 struct net *net,
89 const struct sk_buff *skb,
90 const struct nlmsghdr *nlh,
91 u8 family,
92 struct nft_table *table,
93 struct nft_chain *chain,
94 const struct nlattr * const *nla)
95 {
96 ctx->net = net;
97 ctx->family = family;
98 ctx->level = 0;
99 ctx->table = table;
100 ctx->chain = chain;
101 ctx->nla = nla;
102 ctx->portid = NETLINK_CB(skb).portid;
103 ctx->report = nlmsg_report(nlh);
104 ctx->flags = nlh->nlmsg_flags;
105 ctx->seq = nlh->nlmsg_seq;
106 }
107
108 static struct nft_trans *nft_trans_alloc_gfp(const struct nft_ctx *ctx,
109 int msg_type, u32 size, gfp_t gfp)
110 {
111 struct nft_trans *trans;
112
113 trans = kzalloc(sizeof(struct nft_trans) + size, gfp);
114 if (trans == NULL)
115 return NULL;
116
117 trans->msg_type = msg_type;
118 trans->ctx = *ctx;
119
120 return trans;
121 }
122
123 static struct nft_trans *nft_trans_alloc(const struct nft_ctx *ctx,
124 int msg_type, u32 size)
125 {
126 return nft_trans_alloc_gfp(ctx, msg_type, size, GFP_KERNEL);
127 }
128
129 static void nft_trans_destroy(struct nft_trans *trans)
130 {
131 list_del(&trans->list);
132 kfree(trans);
133 }
134
135 static void nft_set_trans_bind(const struct nft_ctx *ctx, struct nft_set *set)
136 {
137 struct net *net = ctx->net;
138 struct nft_trans *trans;
139
140 if (!nft_set_is_anonymous(set))
141 return;
142
143 list_for_each_entry_reverse(trans, &net->nft.commit_list, list) {
144 switch (trans->msg_type) {
145 case NFT_MSG_NEWSET:
146 if (nft_trans_set(trans) == set)
147 nft_trans_set_bound(trans) = true;
148 break;
149 case NFT_MSG_NEWSETELEM:
150 if (nft_trans_elem_set(trans) == set)
151 nft_trans_elem_set_bound(trans) = true;
152 break;
153 }
154 }
155 }
156
157 static int nft_netdev_register_hooks(struct net *net,
158 struct list_head *hook_list)
159 {
160 struct nft_hook *hook;
161 int err, j;
162
163 j = 0;
164 list_for_each_entry(hook, hook_list, list) {
165 err = nf_register_net_hook(net, &hook->ops);
166 if (err < 0)
167 goto err_register;
168
169 j++;
170 }
171 return 0;
172
173 err_register:
174 list_for_each_entry(hook, hook_list, list) {
175 if (j-- <= 0)
176 break;
177
178 nf_unregister_net_hook(net, &hook->ops);
179 }
180 return err;
181 }
182
183 static void nft_netdev_unregister_hooks(struct net *net,
184 struct list_head *hook_list)
185 {
186 struct nft_hook *hook;
187
188 list_for_each_entry(hook, hook_list, list)
189 nf_unregister_net_hook(net, &hook->ops);
190 }
191
192 static int nf_tables_register_hook(struct net *net,
193 const struct nft_table *table,
194 struct nft_chain *chain)
195 {
196 struct nft_base_chain *basechain;
197 const struct nf_hook_ops *ops;
198
199 if (table->flags & NFT_TABLE_F_DORMANT ||
200 !nft_is_base_chain(chain))
201 return 0;
202
203 basechain = nft_base_chain(chain);
204 ops = &basechain->ops;
205
206 if (basechain->type->ops_register)
207 return basechain->type->ops_register(net, ops);
208
209 if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
210 return nft_netdev_register_hooks(net, &basechain->hook_list);
211
212 return nf_register_net_hook(net, &basechain->ops);
213 }
214
215 static void nf_tables_unregister_hook(struct net *net,
216 const struct nft_table *table,
217 struct nft_chain *chain)
218 {
219 struct nft_base_chain *basechain;
220 const struct nf_hook_ops *ops;
221
222 if (table->flags & NFT_TABLE_F_DORMANT ||
223 !nft_is_base_chain(chain))
224 return;
225 basechain = nft_base_chain(chain);
226 ops = &basechain->ops;
227
228 if (basechain->type->ops_unregister)
229 return basechain->type->ops_unregister(net, ops);
230
231 if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
232 nft_netdev_unregister_hooks(net, &basechain->hook_list);
233 else
234 nf_unregister_net_hook(net, &basechain->ops);
235 }
236
237 static int nft_trans_table_add(struct nft_ctx *ctx, int msg_type)
238 {
239 struct nft_trans *trans;
240
241 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_table));
242 if (trans == NULL)
243 return -ENOMEM;
244
245 if (msg_type == NFT_MSG_NEWTABLE)
246 nft_activate_next(ctx->net, ctx->table);
247
248 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
249 return 0;
250 }
251
252 static int nft_deltable(struct nft_ctx *ctx)
253 {
254 int err;
255
256 err = nft_trans_table_add(ctx, NFT_MSG_DELTABLE);
257 if (err < 0)
258 return err;
259
260 nft_deactivate_next(ctx->net, ctx->table);
261 return err;
262 }
263
264 static struct nft_trans *nft_trans_chain_add(struct nft_ctx *ctx, int msg_type)
265 {
266 struct nft_trans *trans;
267
268 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_chain));
269 if (trans == NULL)
270 return ERR_PTR(-ENOMEM);
271
272 if (msg_type == NFT_MSG_NEWCHAIN) {
273 nft_activate_next(ctx->net, ctx->chain);
274
275 if (ctx->nla[NFTA_CHAIN_ID]) {
276 nft_trans_chain_id(trans) =
277 ntohl(nla_get_be32(ctx->nla[NFTA_CHAIN_ID]));
278 }
279 }
280
281 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
282 return trans;
283 }
284
285 static int nft_delchain(struct nft_ctx *ctx)
286 {
287 struct nft_trans *trans;
288
289 trans = nft_trans_chain_add(ctx, NFT_MSG_DELCHAIN);
290 if (IS_ERR(trans))
291 return PTR_ERR(trans);
292
293 ctx->table->use--;
294 nft_deactivate_next(ctx->net, ctx->chain);
295
296 return 0;
297 }
298
299 static void nft_rule_expr_activate(const struct nft_ctx *ctx,
300 struct nft_rule *rule)
301 {
302 struct nft_expr *expr;
303
304 expr = nft_expr_first(rule);
305 while (nft_expr_more(rule, expr)) {
306 if (expr->ops->activate)
307 expr->ops->activate(ctx, expr);
308
309 expr = nft_expr_next(expr);
310 }
311 }
312
313 static void nft_rule_expr_deactivate(const struct nft_ctx *ctx,
314 struct nft_rule *rule,
315 enum nft_trans_phase phase)
316 {
317 struct nft_expr *expr;
318
319 expr = nft_expr_first(rule);
320 while (nft_expr_more(rule, expr)) {
321 if (expr->ops->deactivate)
322 expr->ops->deactivate(ctx, expr, phase);
323
324 expr = nft_expr_next(expr);
325 }
326 }
327
328 static int
329 nf_tables_delrule_deactivate(struct nft_ctx *ctx, struct nft_rule *rule)
330 {
331 /* You cannot delete the same rule twice */
332 if (nft_is_active_next(ctx->net, rule)) {
333 nft_deactivate_next(ctx->net, rule);
334 ctx->chain->use--;
335 return 0;
336 }
337 return -ENOENT;
338 }
339
340 static struct nft_trans *nft_trans_rule_add(struct nft_ctx *ctx, int msg_type,
341 struct nft_rule *rule)
342 {
343 struct nft_trans *trans;
344
345 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_rule));
346 if (trans == NULL)
347 return NULL;
348
349 if (msg_type == NFT_MSG_NEWRULE && ctx->nla[NFTA_RULE_ID] != NULL) {
350 nft_trans_rule_id(trans) =
351 ntohl(nla_get_be32(ctx->nla[NFTA_RULE_ID]));
352 }
353 nft_trans_rule(trans) = rule;
354 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
355
356 return trans;
357 }
358
359 static int nft_delrule(struct nft_ctx *ctx, struct nft_rule *rule)
360 {
361 struct nft_flow_rule *flow;
362 struct nft_trans *trans;
363 int err;
364
365 trans = nft_trans_rule_add(ctx, NFT_MSG_DELRULE, rule);
366 if (trans == NULL)
367 return -ENOMEM;
368
369 if (ctx->chain->flags & NFT_CHAIN_HW_OFFLOAD) {
370 flow = nft_flow_rule_create(ctx->net, rule);
371 if (IS_ERR(flow)) {
372 nft_trans_destroy(trans);
373 return PTR_ERR(flow);
374 }
375
376 nft_trans_flow_rule(trans) = flow;
377 }
378
379 err = nf_tables_delrule_deactivate(ctx, rule);
380 if (err < 0) {
381 nft_trans_destroy(trans);
382 return err;
383 }
384 nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_PREPARE);
385
386 return 0;
387 }
388
389 static int nft_delrule_by_chain(struct nft_ctx *ctx)
390 {
391 struct nft_rule *rule;
392 int err;
393
394 list_for_each_entry(rule, &ctx->chain->rules, list) {
395 if (!nft_is_active_next(ctx->net, rule))
396 continue;
397
398 err = nft_delrule(ctx, rule);
399 if (err < 0)
400 return err;
401 }
402 return 0;
403 }
404
405 static int nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
406 struct nft_set *set)
407 {
408 struct nft_trans *trans;
409
410 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_set));
411 if (trans == NULL)
412 return -ENOMEM;
413
414 if (msg_type == NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID] != NULL) {
415 nft_trans_set_id(trans) =
416 ntohl(nla_get_be32(ctx->nla[NFTA_SET_ID]));
417 nft_activate_next(ctx->net, set);
418 }
419 nft_trans_set(trans) = set;
420 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
421
422 return 0;
423 }
424
425 static int nft_delset(const struct nft_ctx *ctx, struct nft_set *set)
426 {
427 int err;
428
429 err = nft_trans_set_add(ctx, NFT_MSG_DELSET, set);
430 if (err < 0)
431 return err;
432
433 nft_deactivate_next(ctx->net, set);
434 ctx->table->use--;
435
436 return err;
437 }
438
439 static int nft_trans_obj_add(struct nft_ctx *ctx, int msg_type,
440 struct nft_object *obj)
441 {
442 struct nft_trans *trans;
443
444 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_obj));
445 if (trans == NULL)
446 return -ENOMEM;
447
448 if (msg_type == NFT_MSG_NEWOBJ)
449 nft_activate_next(ctx->net, obj);
450
451 nft_trans_obj(trans) = obj;
452 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
453
454 return 0;
455 }
456
457 static int nft_delobj(struct nft_ctx *ctx, struct nft_object *obj)
458 {
459 int err;
460
461 err = nft_trans_obj_add(ctx, NFT_MSG_DELOBJ, obj);
462 if (err < 0)
463 return err;
464
465 nft_deactivate_next(ctx->net, obj);
466 ctx->table->use--;
467
468 return err;
469 }
470
471 static int nft_trans_flowtable_add(struct nft_ctx *ctx, int msg_type,
472 struct nft_flowtable *flowtable)
473 {
474 struct nft_trans *trans;
475
476 trans = nft_trans_alloc(ctx, msg_type,
477 sizeof(struct nft_trans_flowtable));
478 if (trans == NULL)
479 return -ENOMEM;
480
481 if (msg_type == NFT_MSG_NEWFLOWTABLE)
482 nft_activate_next(ctx->net, flowtable);
483
484 nft_trans_flowtable(trans) = flowtable;
485 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
486
487 return 0;
488 }
489
490 static int nft_delflowtable(struct nft_ctx *ctx,
491 struct nft_flowtable *flowtable)
492 {
493 int err;
494
495 err = nft_trans_flowtable_add(ctx, NFT_MSG_DELFLOWTABLE, flowtable);
496 if (err < 0)
497 return err;
498
499 nft_deactivate_next(ctx->net, flowtable);
500 ctx->table->use--;
501
502 return err;
503 }
504
505 /*
506 * Tables
507 */
508
509 static struct nft_table *nft_table_lookup(const struct net *net,
510 const struct nlattr *nla,
511 u8 family, u8 genmask)
512 {
513 struct nft_table *table;
514
515 if (nla == NULL)
516 return ERR_PTR(-EINVAL);
517
518 list_for_each_entry_rcu(table, &net->nft.tables, list,
519 lockdep_is_held(&net->nft.commit_mutex)) {
520 if (!nla_strcmp(nla, table->name) &&
521 table->family == family &&
522 nft_active_genmask(table, genmask))
523 return table;
524 }
525
526 return ERR_PTR(-ENOENT);
527 }
528
529 static struct nft_table *nft_table_lookup_byhandle(const struct net *net,
530 const struct nlattr *nla,
531 u8 genmask)
532 {
533 struct nft_table *table;
534
535 list_for_each_entry(table, &net->nft.tables, list) {
536 if (be64_to_cpu(nla_get_be64(nla)) == table->handle &&
537 nft_active_genmask(table, genmask))
538 return table;
539 }
540
541 return ERR_PTR(-ENOENT);
542 }
543
544 static inline u64 nf_tables_alloc_handle(struct nft_table *table)
545 {
546 return ++table->hgenerator;
547 }
548
549 static const struct nft_chain_type *chain_type[NFPROTO_NUMPROTO][NFT_CHAIN_T_MAX];
550
551 static const struct nft_chain_type *
552 __nft_chain_type_get(u8 family, enum nft_chain_types type)
553 {
554 if (family >= NFPROTO_NUMPROTO ||
555 type >= NFT_CHAIN_T_MAX)
556 return NULL;
557
558 return chain_type[family][type];
559 }
560
561 static const struct nft_chain_type *
562 __nf_tables_chain_type_lookup(const struct nlattr *nla, u8 family)
563 {
564 const struct nft_chain_type *type;
565 int i;
566
567 for (i = 0; i < NFT_CHAIN_T_MAX; i++) {
568 type = __nft_chain_type_get(family, i);
569 if (!type)
570 continue;
571 if (!nla_strcmp(nla, type->name))
572 return type;
573 }
574 return NULL;
575 }
576
577 struct nft_module_request {
578 struct list_head list;
579 char module[MODULE_NAME_LEN];
580 bool done;
581 };
582
583 #ifdef CONFIG_MODULES
584 static int nft_request_module(struct net *net, const char *fmt, ...)
585 {
586 char module_name[MODULE_NAME_LEN];
587 struct nft_module_request *req;
588 va_list args;
589 int ret;
590
591 va_start(args, fmt);
592 ret = vsnprintf(module_name, MODULE_NAME_LEN, fmt, args);
593 va_end(args);
594 if (ret >= MODULE_NAME_LEN)
595 return 0;
596
597 list_for_each_entry(req, &net->nft.module_list, list) {
598 if (!strcmp(req->module, module_name)) {
599 if (req->done)
600 return 0;
601
602 /* A request to load this module already exists. */
603 return -EAGAIN;
604 }
605 }
606
607 req = kmalloc(sizeof(*req), GFP_KERNEL);
608 if (!req)
609 return -ENOMEM;
610
611 req->done = false;
612 strlcpy(req->module, module_name, MODULE_NAME_LEN);
613 list_add_tail(&req->list, &net->nft.module_list);
614
615 return -EAGAIN;
616 }
617 #endif
618
619 static void lockdep_nfnl_nft_mutex_not_held(void)
620 {
621 #ifdef CONFIG_PROVE_LOCKING
622 if (debug_locks)
623 WARN_ON_ONCE(lockdep_nfnl_is_held(NFNL_SUBSYS_NFTABLES));
624 #endif
625 }
626
627 static const struct nft_chain_type *
628 nf_tables_chain_type_lookup(struct net *net, const struct nlattr *nla,
629 u8 family, bool autoload)
630 {
631 const struct nft_chain_type *type;
632
633 type = __nf_tables_chain_type_lookup(nla, family);
634 if (type != NULL)
635 return type;
636
637 lockdep_nfnl_nft_mutex_not_held();
638 #ifdef CONFIG_MODULES
639 if (autoload) {
640 if (nft_request_module(net, "nft-chain-%u-%.*s", family,
641 nla_len(nla),
642 (const char *)nla_data(nla)) == -EAGAIN)
643 return ERR_PTR(-EAGAIN);
644 }
645 #endif
646 return ERR_PTR(-ENOENT);
647 }
648
649 static const struct nla_policy nft_table_policy[NFTA_TABLE_MAX + 1] = {
650 [NFTA_TABLE_NAME] = { .type = NLA_STRING,
651 .len = NFT_TABLE_MAXNAMELEN - 1 },
652 [NFTA_TABLE_FLAGS] = { .type = NLA_U32 },
653 [NFTA_TABLE_HANDLE] = { .type = NLA_U64 },
654 [NFTA_TABLE_USERDATA] = { .type = NLA_BINARY,
655 .len = NFT_USERDATA_MAXLEN }
656 };
657
658 static int nf_tables_fill_table_info(struct sk_buff *skb, struct net *net,
659 u32 portid, u32 seq, int event, u32 flags,
660 int family, const struct nft_table *table)
661 {
662 struct nlmsghdr *nlh;
663 struct nfgenmsg *nfmsg;
664
665 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
666 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
667 if (nlh == NULL)
668 goto nla_put_failure;
669
670 nfmsg = nlmsg_data(nlh);
671 nfmsg->nfgen_family = family;
672 nfmsg->version = NFNETLINK_V0;
673 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
674
675 if (nla_put_string(skb, NFTA_TABLE_NAME, table->name) ||
676 nla_put_be32(skb, NFTA_TABLE_FLAGS, htonl(table->flags)) ||
677 nla_put_be32(skb, NFTA_TABLE_USE, htonl(table->use)) ||
678 nla_put_be64(skb, NFTA_TABLE_HANDLE, cpu_to_be64(table->handle),
679 NFTA_TABLE_PAD))
680 goto nla_put_failure;
681
682 if (table->udata) {
683 if (nla_put(skb, NFTA_TABLE_USERDATA, table->udlen, table->udata))
684 goto nla_put_failure;
685 }
686
687 nlmsg_end(skb, nlh);
688 return 0;
689
690 nla_put_failure:
691 nlmsg_trim(skb, nlh);
692 return -1;
693 }
694
695 struct nftnl_skb_parms {
696 bool report;
697 };
698 #define NFT_CB(skb) (*(struct nftnl_skb_parms*)&((skb)->cb))
699
700 static void nft_notify_enqueue(struct sk_buff *skb, bool report,
701 struct list_head *notify_list)
702 {
703 NFT_CB(skb).report = report;
704 list_add_tail(&skb->list, notify_list);
705 }
706
707 static void nf_tables_table_notify(const struct nft_ctx *ctx, int event)
708 {
709 struct sk_buff *skb;
710 int err;
711 char *buf = kasprintf(GFP_KERNEL, "%s:%llu;?:0",
712 ctx->table->name, ctx->table->handle);
713
714 audit_log_nfcfg(buf,
715 ctx->family,
716 ctx->table->use,
717 event == NFT_MSG_NEWTABLE ?
718 AUDIT_NFT_OP_TABLE_REGISTER :
719 AUDIT_NFT_OP_TABLE_UNREGISTER,
720 GFP_KERNEL);
721 kfree(buf);
722
723 if (!ctx->report &&
724 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
725 return;
726
727 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
728 if (skb == NULL)
729 goto err;
730
731 err = nf_tables_fill_table_info(skb, ctx->net, ctx->portid, ctx->seq,
732 event, 0, ctx->family, ctx->table);
733 if (err < 0) {
734 kfree_skb(skb);
735 goto err;
736 }
737
738 nft_notify_enqueue(skb, ctx->report, &ctx->net->nft.notify_list);
739 return;
740 err:
741 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
742 }
743
744 static int nf_tables_dump_tables(struct sk_buff *skb,
745 struct netlink_callback *cb)
746 {
747 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
748 const struct nft_table *table;
749 unsigned int idx = 0, s_idx = cb->args[0];
750 struct net *net = sock_net(skb->sk);
751 int family = nfmsg->nfgen_family;
752
753 rcu_read_lock();
754 cb->seq = net->nft.base_seq;
755
756 list_for_each_entry_rcu(table, &net->nft.tables, list) {
757 if (family != NFPROTO_UNSPEC && family != table->family)
758 continue;
759
760 if (idx < s_idx)
761 goto cont;
762 if (idx > s_idx)
763 memset(&cb->args[1], 0,
764 sizeof(cb->args) - sizeof(cb->args[0]));
765 if (!nft_is_active(net, table))
766 continue;
767 if (nf_tables_fill_table_info(skb, net,
768 NETLINK_CB(cb->skb).portid,
769 cb->nlh->nlmsg_seq,
770 NFT_MSG_NEWTABLE, NLM_F_MULTI,
771 table->family, table) < 0)
772 goto done;
773
774 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
775 cont:
776 idx++;
777 }
778 done:
779 rcu_read_unlock();
780 cb->args[0] = idx;
781 return skb->len;
782 }
783
784 static int nft_netlink_dump_start_rcu(struct sock *nlsk, struct sk_buff *skb,
785 const struct nlmsghdr *nlh,
786 struct netlink_dump_control *c)
787 {
788 int err;
789
790 if (!try_module_get(THIS_MODULE))
791 return -EINVAL;
792
793 rcu_read_unlock();
794 err = netlink_dump_start(nlsk, skb, nlh, c);
795 rcu_read_lock();
796 module_put(THIS_MODULE);
797
798 return err;
799 }
800
801 /* called with rcu_read_lock held */
802 static int nf_tables_gettable(struct net *net, struct sock *nlsk,
803 struct sk_buff *skb, const struct nlmsghdr *nlh,
804 const struct nlattr * const nla[],
805 struct netlink_ext_ack *extack)
806 {
807 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
808 u8 genmask = nft_genmask_cur(net);
809 const struct nft_table *table;
810 struct sk_buff *skb2;
811 int family = nfmsg->nfgen_family;
812 int err;
813
814 if (nlh->nlmsg_flags & NLM_F_DUMP) {
815 struct netlink_dump_control c = {
816 .dump = nf_tables_dump_tables,
817 .module = THIS_MODULE,
818 };
819
820 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
821 }
822
823 table = nft_table_lookup(net, nla[NFTA_TABLE_NAME], family, genmask);
824 if (IS_ERR(table)) {
825 NL_SET_BAD_ATTR(extack, nla[NFTA_TABLE_NAME]);
826 return PTR_ERR(table);
827 }
828
829 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
830 if (!skb2)
831 return -ENOMEM;
832
833 err = nf_tables_fill_table_info(skb2, net, NETLINK_CB(skb).portid,
834 nlh->nlmsg_seq, NFT_MSG_NEWTABLE, 0,
835 family, table);
836 if (err < 0)
837 goto err_fill_table_info;
838
839 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
840
841 err_fill_table_info:
842 kfree_skb(skb2);
843 return err;
844 }
845
846 static void nft_table_disable(struct net *net, struct nft_table *table, u32 cnt)
847 {
848 struct nft_chain *chain;
849 u32 i = 0;
850
851 list_for_each_entry(chain, &table->chains, list) {
852 if (!nft_is_active_next(net, chain))
853 continue;
854 if (!nft_is_base_chain(chain))
855 continue;
856
857 if (cnt && i++ == cnt)
858 break;
859
860 nf_tables_unregister_hook(net, table, chain);
861 }
862 }
863
864 static int nf_tables_table_enable(struct net *net, struct nft_table *table)
865 {
866 struct nft_chain *chain;
867 int err, i = 0;
868
869 list_for_each_entry(chain, &table->chains, list) {
870 if (!nft_is_active_next(net, chain))
871 continue;
872 if (!nft_is_base_chain(chain))
873 continue;
874
875 err = nf_tables_register_hook(net, table, chain);
876 if (err < 0)
877 goto err_register_hooks;
878
879 i++;
880 }
881 return 0;
882
883 err_register_hooks:
884 if (i)
885 nft_table_disable(net, table, i);
886 return err;
887 }
888
889 static void nf_tables_table_disable(struct net *net, struct nft_table *table)
890 {
891 nft_table_disable(net, table, 0);
892 }
893
894 static int nf_tables_updtable(struct nft_ctx *ctx)
895 {
896 struct nft_trans *trans;
897 u32 flags;
898 int ret = 0;
899
900 if (!ctx->nla[NFTA_TABLE_FLAGS])
901 return 0;
902
903 flags = ntohl(nla_get_be32(ctx->nla[NFTA_TABLE_FLAGS]));
904 if (flags & ~NFT_TABLE_F_DORMANT)
905 return -EINVAL;
906
907 if (flags == ctx->table->flags)
908 return 0;
909
910 trans = nft_trans_alloc(ctx, NFT_MSG_NEWTABLE,
911 sizeof(struct nft_trans_table));
912 if (trans == NULL)
913 return -ENOMEM;
914
915 if ((flags & NFT_TABLE_F_DORMANT) &&
916 !(ctx->table->flags & NFT_TABLE_F_DORMANT)) {
917 nft_trans_table_enable(trans) = false;
918 } else if (!(flags & NFT_TABLE_F_DORMANT) &&
919 ctx->table->flags & NFT_TABLE_F_DORMANT) {
920 ctx->table->flags &= ~NFT_TABLE_F_DORMANT;
921 ret = nf_tables_table_enable(ctx->net, ctx->table);
922 if (ret >= 0)
923 nft_trans_table_enable(trans) = true;
924 else
925 ctx->table->flags |= NFT_TABLE_F_DORMANT;
926 }
927 if (ret < 0)
928 goto err;
929
930 nft_trans_table_update(trans) = true;
931 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
932 return 0;
933 err:
934 nft_trans_destroy(trans);
935 return ret;
936 }
937
938 static u32 nft_chain_hash(const void *data, u32 len, u32 seed)
939 {
940 const char *name = data;
941
942 return jhash(name, strlen(name), seed);
943 }
944
945 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed)
946 {
947 const struct nft_chain *chain = data;
948
949 return nft_chain_hash(chain->name, 0, seed);
950 }
951
952 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *arg,
953 const void *ptr)
954 {
955 const struct nft_chain *chain = ptr;
956 const char *name = arg->key;
957
958 return strcmp(chain->name, name);
959 }
960
961 static u32 nft_objname_hash(const void *data, u32 len, u32 seed)
962 {
963 const struct nft_object_hash_key *k = data;
964
965 seed ^= hash_ptr(k->table, 32);
966
967 return jhash(k->name, strlen(k->name), seed);
968 }
969
970 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed)
971 {
972 const struct nft_object *obj = data;
973
974 return nft_objname_hash(&obj->key, 0, seed);
975 }
976
977 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *arg,
978 const void *ptr)
979 {
980 const struct nft_object_hash_key *k = arg->key;
981 const struct nft_object *obj = ptr;
982
983 if (obj->key.table != k->table)
984 return -1;
985
986 return strcmp(obj->key.name, k->name);
987 }
988
989 static int nf_tables_newtable(struct net *net, struct sock *nlsk,
990 struct sk_buff *skb, const struct nlmsghdr *nlh,
991 const struct nlattr * const nla[],
992 struct netlink_ext_ack *extack)
993 {
994 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
995 u8 genmask = nft_genmask_next(net);
996 int family = nfmsg->nfgen_family;
997 const struct nlattr *attr;
998 struct nft_table *table;
999 struct nft_ctx ctx;
1000 u32 flags = 0;
1001 int err;
1002
1003 lockdep_assert_held(&net->nft.commit_mutex);
1004 attr = nla[NFTA_TABLE_NAME];
1005 table = nft_table_lookup(net, attr, family, genmask);
1006 if (IS_ERR(table)) {
1007 if (PTR_ERR(table) != -ENOENT)
1008 return PTR_ERR(table);
1009 } else {
1010 if (nlh->nlmsg_flags & NLM_F_EXCL) {
1011 NL_SET_BAD_ATTR(extack, attr);
1012 return -EEXIST;
1013 }
1014 if (nlh->nlmsg_flags & NLM_F_REPLACE)
1015 return -EOPNOTSUPP;
1016
1017 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
1018 return nf_tables_updtable(&ctx);
1019 }
1020
1021 if (nla[NFTA_TABLE_FLAGS]) {
1022 flags = ntohl(nla_get_be32(nla[NFTA_TABLE_FLAGS]));
1023 if (flags & ~NFT_TABLE_F_DORMANT)
1024 return -EINVAL;
1025 }
1026
1027 err = -ENOMEM;
1028 table = kzalloc(sizeof(*table), GFP_KERNEL);
1029 if (table == NULL)
1030 goto err_kzalloc;
1031
1032 table->name = nla_strdup(attr, GFP_KERNEL);
1033 if (table->name == NULL)
1034 goto err_strdup;
1035
1036 if (nla[NFTA_TABLE_USERDATA]) {
1037 table->udata = nla_memdup(nla[NFTA_TABLE_USERDATA], GFP_KERNEL);
1038 if (table->udata == NULL)
1039 goto err_table_udata;
1040
1041 table->udlen = nla_len(nla[NFTA_TABLE_USERDATA]);
1042 }
1043
1044 err = rhltable_init(&table->chains_ht, &nft_chain_ht_params);
1045 if (err)
1046 goto err_chain_ht;
1047
1048 INIT_LIST_HEAD(&table->chains);
1049 INIT_LIST_HEAD(&table->sets);
1050 INIT_LIST_HEAD(&table->objects);
1051 INIT_LIST_HEAD(&table->flowtables);
1052 table->family = family;
1053 table->flags = flags;
1054 table->handle = ++table_handle;
1055
1056 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
1057 err = nft_trans_table_add(&ctx, NFT_MSG_NEWTABLE);
1058 if (err < 0)
1059 goto err_trans;
1060
1061 list_add_tail_rcu(&table->list, &net->nft.tables);
1062 return 0;
1063 err_trans:
1064 rhltable_destroy(&table->chains_ht);
1065 err_chain_ht:
1066 kfree(table->udata);
1067 err_table_udata:
1068 kfree(table->name);
1069 err_strdup:
1070 kfree(table);
1071 err_kzalloc:
1072 return err;
1073 }
1074
1075 static int nft_flush_table(struct nft_ctx *ctx)
1076 {
1077 struct nft_flowtable *flowtable, *nft;
1078 struct nft_chain *chain, *nc;
1079 struct nft_object *obj, *ne;
1080 struct nft_set *set, *ns;
1081 int err;
1082
1083 list_for_each_entry(chain, &ctx->table->chains, list) {
1084 if (!nft_is_active_next(ctx->net, chain))
1085 continue;
1086
1087 if (nft_chain_is_bound(chain))
1088 continue;
1089
1090 ctx->chain = chain;
1091
1092 err = nft_delrule_by_chain(ctx);
1093 if (err < 0)
1094 goto out;
1095 }
1096
1097 list_for_each_entry_safe(set, ns, &ctx->table->sets, list) {
1098 if (!nft_is_active_next(ctx->net, set))
1099 continue;
1100
1101 if (nft_set_is_anonymous(set) &&
1102 !list_empty(&set->bindings))
1103 continue;
1104
1105 err = nft_delset(ctx, set);
1106 if (err < 0)
1107 goto out;
1108 }
1109
1110 list_for_each_entry_safe(flowtable, nft, &ctx->table->flowtables, list) {
1111 if (!nft_is_active_next(ctx->net, flowtable))
1112 continue;
1113
1114 err = nft_delflowtable(ctx, flowtable);
1115 if (err < 0)
1116 goto out;
1117 }
1118
1119 list_for_each_entry_safe(obj, ne, &ctx->table->objects, list) {
1120 if (!nft_is_active_next(ctx->net, obj))
1121 continue;
1122
1123 err = nft_delobj(ctx, obj);
1124 if (err < 0)
1125 goto out;
1126 }
1127
1128 list_for_each_entry_safe(chain, nc, &ctx->table->chains, list) {
1129 if (!nft_is_active_next(ctx->net, chain))
1130 continue;
1131
1132 if (nft_chain_is_bound(chain))
1133 continue;
1134
1135 ctx->chain = chain;
1136
1137 err = nft_delchain(ctx);
1138 if (err < 0)
1139 goto out;
1140 }
1141
1142 err = nft_deltable(ctx);
1143 out:
1144 return err;
1145 }
1146
1147 static int nft_flush(struct nft_ctx *ctx, int family)
1148 {
1149 struct nft_table *table, *nt;
1150 const struct nlattr * const *nla = ctx->nla;
1151 int err = 0;
1152
1153 list_for_each_entry_safe(table, nt, &ctx->net->nft.tables, list) {
1154 if (family != AF_UNSPEC && table->family != family)
1155 continue;
1156
1157 ctx->family = table->family;
1158
1159 if (!nft_is_active_next(ctx->net, table))
1160 continue;
1161
1162 if (nla[NFTA_TABLE_NAME] &&
1163 nla_strcmp(nla[NFTA_TABLE_NAME], table->name) != 0)
1164 continue;
1165
1166 ctx->table = table;
1167
1168 err = nft_flush_table(ctx);
1169 if (err < 0)
1170 goto out;
1171 }
1172 out:
1173 return err;
1174 }
1175
1176 static int nf_tables_deltable(struct net *net, struct sock *nlsk,
1177 struct sk_buff *skb, const struct nlmsghdr *nlh,
1178 const struct nlattr * const nla[],
1179 struct netlink_ext_ack *extack)
1180 {
1181 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1182 u8 genmask = nft_genmask_next(net);
1183 int family = nfmsg->nfgen_family;
1184 const struct nlattr *attr;
1185 struct nft_table *table;
1186 struct nft_ctx ctx;
1187
1188 nft_ctx_init(&ctx, net, skb, nlh, 0, NULL, NULL, nla);
1189 if (family == AF_UNSPEC ||
1190 (!nla[NFTA_TABLE_NAME] && !nla[NFTA_TABLE_HANDLE]))
1191 return nft_flush(&ctx, family);
1192
1193 if (nla[NFTA_TABLE_HANDLE]) {
1194 attr = nla[NFTA_TABLE_HANDLE];
1195 table = nft_table_lookup_byhandle(net, attr, genmask);
1196 } else {
1197 attr = nla[NFTA_TABLE_NAME];
1198 table = nft_table_lookup(net, attr, family, genmask);
1199 }
1200
1201 if (IS_ERR(table)) {
1202 NL_SET_BAD_ATTR(extack, attr);
1203 return PTR_ERR(table);
1204 }
1205
1206 if (nlh->nlmsg_flags & NLM_F_NONREC &&
1207 table->use > 0)
1208 return -EBUSY;
1209
1210 ctx.family = family;
1211 ctx.table = table;
1212
1213 return nft_flush_table(&ctx);
1214 }
1215
1216 static void nf_tables_table_destroy(struct nft_ctx *ctx)
1217 {
1218 if (WARN_ON(ctx->table->use > 0))
1219 return;
1220
1221 rhltable_destroy(&ctx->table->chains_ht);
1222 kfree(ctx->table->name);
1223 kfree(ctx->table->udata);
1224 kfree(ctx->table);
1225 }
1226
1227 void nft_register_chain_type(const struct nft_chain_type *ctype)
1228 {
1229 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1230 if (WARN_ON(__nft_chain_type_get(ctype->family, ctype->type))) {
1231 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1232 return;
1233 }
1234 chain_type[ctype->family][ctype->type] = ctype;
1235 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1236 }
1237 EXPORT_SYMBOL_GPL(nft_register_chain_type);
1238
1239 void nft_unregister_chain_type(const struct nft_chain_type *ctype)
1240 {
1241 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1242 chain_type[ctype->family][ctype->type] = NULL;
1243 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1244 }
1245 EXPORT_SYMBOL_GPL(nft_unregister_chain_type);
1246
1247 /*
1248 * Chains
1249 */
1250
1251 static struct nft_chain *
1252 nft_chain_lookup_byhandle(const struct nft_table *table, u64 handle, u8 genmask)
1253 {
1254 struct nft_chain *chain;
1255
1256 list_for_each_entry(chain, &table->chains, list) {
1257 if (chain->handle == handle &&
1258 nft_active_genmask(chain, genmask))
1259 return chain;
1260 }
1261
1262 return ERR_PTR(-ENOENT);
1263 }
1264
1265 static bool lockdep_commit_lock_is_held(const struct net *net)
1266 {
1267 #ifdef CONFIG_PROVE_LOCKING
1268 return lockdep_is_held(&net->nft.commit_mutex);
1269 #else
1270 return true;
1271 #endif
1272 }
1273
1274 static struct nft_chain *nft_chain_lookup(struct net *net,
1275 struct nft_table *table,
1276 const struct nlattr *nla, u8 genmask)
1277 {
1278 char search[NFT_CHAIN_MAXNAMELEN + 1];
1279 struct rhlist_head *tmp, *list;
1280 struct nft_chain *chain;
1281
1282 if (nla == NULL)
1283 return ERR_PTR(-EINVAL);
1284
1285 nla_strlcpy(search, nla, sizeof(search));
1286
1287 WARN_ON(!rcu_read_lock_held() &&
1288 !lockdep_commit_lock_is_held(net));
1289
1290 chain = ERR_PTR(-ENOENT);
1291 rcu_read_lock();
1292 list = rhltable_lookup(&table->chains_ht, search, nft_chain_ht_params);
1293 if (!list)
1294 goto out_unlock;
1295
1296 rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
1297 if (nft_active_genmask(chain, genmask))
1298 goto out_unlock;
1299 }
1300 chain = ERR_PTR(-ENOENT);
1301 out_unlock:
1302 rcu_read_unlock();
1303 return chain;
1304 }
1305
1306 static const struct nla_policy nft_chain_policy[NFTA_CHAIN_MAX + 1] = {
1307 [NFTA_CHAIN_TABLE] = { .type = NLA_STRING,
1308 .len = NFT_TABLE_MAXNAMELEN - 1 },
1309 [NFTA_CHAIN_HANDLE] = { .type = NLA_U64 },
1310 [NFTA_CHAIN_NAME] = { .type = NLA_STRING,
1311 .len = NFT_CHAIN_MAXNAMELEN - 1 },
1312 [NFTA_CHAIN_HOOK] = { .type = NLA_NESTED },
1313 [NFTA_CHAIN_POLICY] = { .type = NLA_U32 },
1314 [NFTA_CHAIN_TYPE] = { .type = NLA_STRING,
1315 .len = NFT_MODULE_AUTOLOAD_LIMIT },
1316 [NFTA_CHAIN_COUNTERS] = { .type = NLA_NESTED },
1317 [NFTA_CHAIN_FLAGS] = { .type = NLA_U32 },
1318 [NFTA_CHAIN_ID] = { .type = NLA_U32 },
1319 [NFTA_CHAIN_USERDATA] = { .type = NLA_BINARY,
1320 .len = NFT_USERDATA_MAXLEN },
1321 };
1322
1323 static const struct nla_policy nft_hook_policy[NFTA_HOOK_MAX + 1] = {
1324 [NFTA_HOOK_HOOKNUM] = { .type = NLA_U32 },
1325 [NFTA_HOOK_PRIORITY] = { .type = NLA_U32 },
1326 [NFTA_HOOK_DEV] = { .type = NLA_STRING,
1327 .len = IFNAMSIZ - 1 },
1328 };
1329
1330 static int nft_dump_stats(struct sk_buff *skb, struct nft_stats __percpu *stats)
1331 {
1332 struct nft_stats *cpu_stats, total;
1333 struct nlattr *nest;
1334 unsigned int seq;
1335 u64 pkts, bytes;
1336 int cpu;
1337
1338 if (!stats)
1339 return 0;
1340
1341 memset(&total, 0, sizeof(total));
1342 for_each_possible_cpu(cpu) {
1343 cpu_stats = per_cpu_ptr(stats, cpu);
1344 do {
1345 seq = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1346 pkts = cpu_stats->pkts;
1347 bytes = cpu_stats->bytes;
1348 } while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, seq));
1349 total.pkts += pkts;
1350 total.bytes += bytes;
1351 }
1352 nest = nla_nest_start_noflag(skb, NFTA_CHAIN_COUNTERS);
1353 if (nest == NULL)
1354 goto nla_put_failure;
1355
1356 if (nla_put_be64(skb, NFTA_COUNTER_PACKETS, cpu_to_be64(total.pkts),
1357 NFTA_COUNTER_PAD) ||
1358 nla_put_be64(skb, NFTA_COUNTER_BYTES, cpu_to_be64(total.bytes),
1359 NFTA_COUNTER_PAD))
1360 goto nla_put_failure;
1361
1362 nla_nest_end(skb, nest);
1363 return 0;
1364
1365 nla_put_failure:
1366 return -ENOSPC;
1367 }
1368
1369 static int nft_dump_basechain_hook(struct sk_buff *skb, int family,
1370 const struct nft_base_chain *basechain)
1371 {
1372 const struct nf_hook_ops *ops = &basechain->ops;
1373 struct nft_hook *hook, *first = NULL;
1374 struct nlattr *nest, *nest_devs;
1375 int n = 0;
1376
1377 nest = nla_nest_start_noflag(skb, NFTA_CHAIN_HOOK);
1378 if (nest == NULL)
1379 goto nla_put_failure;
1380 if (nla_put_be32(skb, NFTA_HOOK_HOOKNUM, htonl(ops->hooknum)))
1381 goto nla_put_failure;
1382 if (nla_put_be32(skb, NFTA_HOOK_PRIORITY, htonl(ops->priority)))
1383 goto nla_put_failure;
1384
1385 if (nft_base_chain_netdev(family, ops->hooknum)) {
1386 nest_devs = nla_nest_start_noflag(skb, NFTA_HOOK_DEVS);
1387 list_for_each_entry(hook, &basechain->hook_list, list) {
1388 if (!first)
1389 first = hook;
1390
1391 if (nla_put_string(skb, NFTA_DEVICE_NAME,
1392 hook->ops.dev->name))
1393 goto nla_put_failure;
1394 n++;
1395 }
1396 nla_nest_end(skb, nest_devs);
1397
1398 if (n == 1 &&
1399 nla_put_string(skb, NFTA_HOOK_DEV, first->ops.dev->name))
1400 goto nla_put_failure;
1401 }
1402 nla_nest_end(skb, nest);
1403
1404 return 0;
1405 nla_put_failure:
1406 return -1;
1407 }
1408
1409 static int nf_tables_fill_chain_info(struct sk_buff *skb, struct net *net,
1410 u32 portid, u32 seq, int event, u32 flags,
1411 int family, const struct nft_table *table,
1412 const struct nft_chain *chain)
1413 {
1414 struct nlmsghdr *nlh;
1415 struct nfgenmsg *nfmsg;
1416
1417 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
1418 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
1419 if (nlh == NULL)
1420 goto nla_put_failure;
1421
1422 nfmsg = nlmsg_data(nlh);
1423 nfmsg->nfgen_family = family;
1424 nfmsg->version = NFNETLINK_V0;
1425 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
1426
1427 if (nla_put_string(skb, NFTA_CHAIN_TABLE, table->name))
1428 goto nla_put_failure;
1429 if (nla_put_be64(skb, NFTA_CHAIN_HANDLE, cpu_to_be64(chain->handle),
1430 NFTA_CHAIN_PAD))
1431 goto nla_put_failure;
1432 if (nla_put_string(skb, NFTA_CHAIN_NAME, chain->name))
1433 goto nla_put_failure;
1434
1435 if (nft_is_base_chain(chain)) {
1436 const struct nft_base_chain *basechain = nft_base_chain(chain);
1437 struct nft_stats __percpu *stats;
1438
1439 if (nft_dump_basechain_hook(skb, family, basechain))
1440 goto nla_put_failure;
1441
1442 if (nla_put_be32(skb, NFTA_CHAIN_POLICY,
1443 htonl(basechain->policy)))
1444 goto nla_put_failure;
1445
1446 if (nla_put_string(skb, NFTA_CHAIN_TYPE, basechain->type->name))
1447 goto nla_put_failure;
1448
1449 stats = rcu_dereference_check(basechain->stats,
1450 lockdep_commit_lock_is_held(net));
1451 if (nft_dump_stats(skb, stats))
1452 goto nla_put_failure;
1453 }
1454
1455 if (chain->flags &&
1456 nla_put_be32(skb, NFTA_CHAIN_FLAGS, htonl(chain->flags)))
1457 goto nla_put_failure;
1458
1459 if (nla_put_be32(skb, NFTA_CHAIN_USE, htonl(chain->use)))
1460 goto nla_put_failure;
1461
1462 if (chain->udata &&
1463 nla_put(skb, NFTA_CHAIN_USERDATA, chain->udlen, chain->udata))
1464 goto nla_put_failure;
1465
1466 nlmsg_end(skb, nlh);
1467 return 0;
1468
1469 nla_put_failure:
1470 nlmsg_trim(skb, nlh);
1471 return -1;
1472 }
1473
1474 static void nf_tables_chain_notify(const struct nft_ctx *ctx, int event)
1475 {
1476 struct sk_buff *skb;
1477 int err;
1478 char *buf = kasprintf(GFP_KERNEL, "%s:%llu;%s:%llu",
1479 ctx->table->name, ctx->table->handle,
1480 ctx->chain->name, ctx->chain->handle);
1481
1482 audit_log_nfcfg(buf,
1483 ctx->family,
1484 ctx->chain->use,
1485 event == NFT_MSG_NEWCHAIN ?
1486 AUDIT_NFT_OP_CHAIN_REGISTER :
1487 AUDIT_NFT_OP_CHAIN_UNREGISTER,
1488 GFP_KERNEL);
1489 kfree(buf);
1490
1491 if (!ctx->report &&
1492 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1493 return;
1494
1495 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1496 if (skb == NULL)
1497 goto err;
1498
1499 err = nf_tables_fill_chain_info(skb, ctx->net, ctx->portid, ctx->seq,
1500 event, 0, ctx->family, ctx->table,
1501 ctx->chain);
1502 if (err < 0) {
1503 kfree_skb(skb);
1504 goto err;
1505 }
1506
1507 nft_notify_enqueue(skb, ctx->report, &ctx->net->nft.notify_list);
1508 return;
1509 err:
1510 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1511 }
1512
1513 static int nf_tables_dump_chains(struct sk_buff *skb,
1514 struct netlink_callback *cb)
1515 {
1516 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1517 const struct nft_table *table;
1518 const struct nft_chain *chain;
1519 unsigned int idx = 0, s_idx = cb->args[0];
1520 struct net *net = sock_net(skb->sk);
1521 int family = nfmsg->nfgen_family;
1522
1523 rcu_read_lock();
1524 cb->seq = net->nft.base_seq;
1525
1526 list_for_each_entry_rcu(table, &net->nft.tables, list) {
1527 if (family != NFPROTO_UNSPEC && family != table->family)
1528 continue;
1529
1530 list_for_each_entry_rcu(chain, &table->chains, list) {
1531 if (idx < s_idx)
1532 goto cont;
1533 if (idx > s_idx)
1534 memset(&cb->args[1], 0,
1535 sizeof(cb->args) - sizeof(cb->args[0]));
1536 if (!nft_is_active(net, chain))
1537 continue;
1538 if (nf_tables_fill_chain_info(skb, net,
1539 NETLINK_CB(cb->skb).portid,
1540 cb->nlh->nlmsg_seq,
1541 NFT_MSG_NEWCHAIN,
1542 NLM_F_MULTI,
1543 table->family, table,
1544 chain) < 0)
1545 goto done;
1546
1547 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1548 cont:
1549 idx++;
1550 }
1551 }
1552 done:
1553 rcu_read_unlock();
1554 cb->args[0] = idx;
1555 return skb->len;
1556 }
1557
1558 /* called with rcu_read_lock held */
1559 static int nf_tables_getchain(struct net *net, struct sock *nlsk,
1560 struct sk_buff *skb, const struct nlmsghdr *nlh,
1561 const struct nlattr * const nla[],
1562 struct netlink_ext_ack *extack)
1563 {
1564 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1565 u8 genmask = nft_genmask_cur(net);
1566 const struct nft_chain *chain;
1567 struct nft_table *table;
1568 struct sk_buff *skb2;
1569 int family = nfmsg->nfgen_family;
1570 int err;
1571
1572 if (nlh->nlmsg_flags & NLM_F_DUMP) {
1573 struct netlink_dump_control c = {
1574 .dump = nf_tables_dump_chains,
1575 .module = THIS_MODULE,
1576 };
1577
1578 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
1579 }
1580
1581 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask);
1582 if (IS_ERR(table)) {
1583 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
1584 return PTR_ERR(table);
1585 }
1586
1587 chain = nft_chain_lookup(net, table, nla[NFTA_CHAIN_NAME], genmask);
1588 if (IS_ERR(chain)) {
1589 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
1590 return PTR_ERR(chain);
1591 }
1592
1593 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1594 if (!skb2)
1595 return -ENOMEM;
1596
1597 err = nf_tables_fill_chain_info(skb2, net, NETLINK_CB(skb).portid,
1598 nlh->nlmsg_seq, NFT_MSG_NEWCHAIN, 0,
1599 family, table, chain);
1600 if (err < 0)
1601 goto err_fill_chain_info;
1602
1603 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
1604
1605 err_fill_chain_info:
1606 kfree_skb(skb2);
1607 return err;
1608 }
1609
1610 static const struct nla_policy nft_counter_policy[NFTA_COUNTER_MAX + 1] = {
1611 [NFTA_COUNTER_PACKETS] = { .type = NLA_U64 },
1612 [NFTA_COUNTER_BYTES] = { .type = NLA_U64 },
1613 };
1614
1615 static struct nft_stats __percpu *nft_stats_alloc(const struct nlattr *attr)
1616 {
1617 struct nlattr *tb[NFTA_COUNTER_MAX+1];
1618 struct nft_stats __percpu *newstats;
1619 struct nft_stats *stats;
1620 int err;
1621
1622 err = nla_parse_nested_deprecated(tb, NFTA_COUNTER_MAX, attr,
1623 nft_counter_policy, NULL);
1624 if (err < 0)
1625 return ERR_PTR(err);
1626
1627 if (!tb[NFTA_COUNTER_BYTES] || !tb[NFTA_COUNTER_PACKETS])
1628 return ERR_PTR(-EINVAL);
1629
1630 newstats = netdev_alloc_pcpu_stats(struct nft_stats);
1631 if (newstats == NULL)
1632 return ERR_PTR(-ENOMEM);
1633
1634 /* Restore old counters on this cpu, no problem. Per-cpu statistics
1635 * are not exposed to userspace.
1636 */
1637 preempt_disable();
1638 stats = this_cpu_ptr(newstats);
1639 stats->bytes = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_BYTES]));
1640 stats->pkts = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_PACKETS]));
1641 preempt_enable();
1642
1643 return newstats;
1644 }
1645
1646 static void nft_chain_stats_replace(struct nft_trans *trans)
1647 {
1648 struct nft_base_chain *chain = nft_base_chain(trans->ctx.chain);
1649
1650 if (!nft_trans_chain_stats(trans))
1651 return;
1652
1653 nft_trans_chain_stats(trans) =
1654 rcu_replace_pointer(chain->stats, nft_trans_chain_stats(trans),
1655 lockdep_commit_lock_is_held(trans->ctx.net));
1656
1657 if (!nft_trans_chain_stats(trans))
1658 static_branch_inc(&nft_counters_enabled);
1659 }
1660
1661 static void nf_tables_chain_free_chain_rules(struct nft_chain *chain)
1662 {
1663 struct nft_rule **g0 = rcu_dereference_raw(chain->rules_gen_0);
1664 struct nft_rule **g1 = rcu_dereference_raw(chain->rules_gen_1);
1665
1666 if (g0 != g1)
1667 kvfree(g1);
1668 kvfree(g0);
1669
1670 /* should be NULL either via abort or via successful commit */
1671 WARN_ON_ONCE(chain->rules_next);
1672 kvfree(chain->rules_next);
1673 }
1674
1675 void nf_tables_chain_destroy(struct nft_ctx *ctx)
1676 {
1677 struct nft_chain *chain = ctx->chain;
1678 struct nft_hook *hook, *next;
1679
1680 if (WARN_ON(chain->use > 0))
1681 return;
1682
1683 /* no concurrent access possible anymore */
1684 nf_tables_chain_free_chain_rules(chain);
1685
1686 if (nft_is_base_chain(chain)) {
1687 struct nft_base_chain *basechain = nft_base_chain(chain);
1688
1689 if (nft_base_chain_netdev(ctx->family, basechain->ops.hooknum)) {
1690 list_for_each_entry_safe(hook, next,
1691 &basechain->hook_list, list) {
1692 list_del_rcu(&hook->list);
1693 kfree_rcu(hook, rcu);
1694 }
1695 }
1696 module_put(basechain->type->owner);
1697 if (rcu_access_pointer(basechain->stats)) {
1698 static_branch_dec(&nft_counters_enabled);
1699 free_percpu(rcu_dereference_raw(basechain->stats));
1700 }
1701 kfree(chain->name);
1702 kfree(chain->udata);
1703 kfree(basechain);
1704 } else {
1705 kfree(chain->name);
1706 kfree(chain->udata);
1707 kfree(chain);
1708 }
1709 }
1710
1711 static struct nft_hook *nft_netdev_hook_alloc(struct net *net,
1712 const struct nlattr *attr)
1713 {
1714 struct net_device *dev;
1715 char ifname[IFNAMSIZ];
1716 struct nft_hook *hook;
1717 int err;
1718
1719 hook = kmalloc(sizeof(struct nft_hook), GFP_KERNEL);
1720 if (!hook) {
1721 err = -ENOMEM;
1722 goto err_hook_alloc;
1723 }
1724
1725 nla_strlcpy(ifname, attr, IFNAMSIZ);
1726 dev = __dev_get_by_name(net, ifname);
1727 if (!dev) {
1728 err = -ENOENT;
1729 goto err_hook_dev;
1730 }
1731 hook->ops.dev = dev;
1732 hook->inactive = false;
1733
1734 return hook;
1735
1736 err_hook_dev:
1737 kfree(hook);
1738 err_hook_alloc:
1739 return ERR_PTR(err);
1740 }
1741
1742 static struct nft_hook *nft_hook_list_find(struct list_head *hook_list,
1743 const struct nft_hook *this)
1744 {
1745 struct nft_hook *hook;
1746
1747 list_for_each_entry(hook, hook_list, list) {
1748 if (this->ops.dev == hook->ops.dev)
1749 return hook;
1750 }
1751
1752 return NULL;
1753 }
1754
1755 static int nf_tables_parse_netdev_hooks(struct net *net,
1756 const struct nlattr *attr,
1757 struct list_head *hook_list)
1758 {
1759 struct nft_hook *hook, *next;
1760 const struct nlattr *tmp;
1761 int rem, n = 0, err;
1762
1763 nla_for_each_nested(tmp, attr, rem) {
1764 if (nla_type(tmp) != NFTA_DEVICE_NAME) {
1765 err = -EINVAL;
1766 goto err_hook;
1767 }
1768
1769 hook = nft_netdev_hook_alloc(net, tmp);
1770 if (IS_ERR(hook)) {
1771 err = PTR_ERR(hook);
1772 goto err_hook;
1773 }
1774 if (nft_hook_list_find(hook_list, hook)) {
1775 kfree(hook);
1776 err = -EEXIST;
1777 goto err_hook;
1778 }
1779 list_add_tail(&hook->list, hook_list);
1780 n++;
1781
1782 if (n == NFT_NETDEVICE_MAX) {
1783 err = -EFBIG;
1784 goto err_hook;
1785 }
1786 }
1787
1788 return 0;
1789
1790 err_hook:
1791 list_for_each_entry_safe(hook, next, hook_list, list) {
1792 list_del(&hook->list);
1793 kfree(hook);
1794 }
1795 return err;
1796 }
1797
1798 struct nft_chain_hook {
1799 u32 num;
1800 s32 priority;
1801 const struct nft_chain_type *type;
1802 struct list_head list;
1803 };
1804
1805 static int nft_chain_parse_netdev(struct net *net,
1806 struct nlattr *tb[],
1807 struct list_head *hook_list)
1808 {
1809 struct nft_hook *hook;
1810 int err;
1811
1812 if (tb[NFTA_HOOK_DEV]) {
1813 hook = nft_netdev_hook_alloc(net, tb[NFTA_HOOK_DEV]);
1814 if (IS_ERR(hook))
1815 return PTR_ERR(hook);
1816
1817 list_add_tail(&hook->list, hook_list);
1818 } else if (tb[NFTA_HOOK_DEVS]) {
1819 err = nf_tables_parse_netdev_hooks(net, tb[NFTA_HOOK_DEVS],
1820 hook_list);
1821 if (err < 0)
1822 return err;
1823
1824 if (list_empty(hook_list))
1825 return -EINVAL;
1826 } else {
1827 return -EINVAL;
1828 }
1829
1830 return 0;
1831 }
1832
1833 static int nft_chain_parse_hook(struct net *net,
1834 const struct nlattr * const nla[],
1835 struct nft_chain_hook *hook, u8 family,
1836 bool autoload)
1837 {
1838 struct nlattr *ha[NFTA_HOOK_MAX + 1];
1839 const struct nft_chain_type *type;
1840 int err;
1841
1842 lockdep_assert_held(&net->nft.commit_mutex);
1843 lockdep_nfnl_nft_mutex_not_held();
1844
1845 err = nla_parse_nested_deprecated(ha, NFTA_HOOK_MAX,
1846 nla[NFTA_CHAIN_HOOK],
1847 nft_hook_policy, NULL);
1848 if (err < 0)
1849 return err;
1850
1851 if (ha[NFTA_HOOK_HOOKNUM] == NULL ||
1852 ha[NFTA_HOOK_PRIORITY] == NULL)
1853 return -EINVAL;
1854
1855 hook->num = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
1856 hook->priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
1857
1858 type = __nft_chain_type_get(family, NFT_CHAIN_T_DEFAULT);
1859 if (!type)
1860 return -EOPNOTSUPP;
1861
1862 if (nla[NFTA_CHAIN_TYPE]) {
1863 type = nf_tables_chain_type_lookup(net, nla[NFTA_CHAIN_TYPE],
1864 family, autoload);
1865 if (IS_ERR(type))
1866 return PTR_ERR(type);
1867 }
1868 if (hook->num >= NFT_MAX_HOOKS || !(type->hook_mask & (1 << hook->num)))
1869 return -EOPNOTSUPP;
1870
1871 if (type->type == NFT_CHAIN_T_NAT &&
1872 hook->priority <= NF_IP_PRI_CONNTRACK)
1873 return -EOPNOTSUPP;
1874
1875 if (!try_module_get(type->owner))
1876 return -ENOENT;
1877
1878 hook->type = type;
1879
1880 INIT_LIST_HEAD(&hook->list);
1881 if (nft_base_chain_netdev(family, hook->num)) {
1882 err = nft_chain_parse_netdev(net, ha, &hook->list);
1883 if (err < 0) {
1884 module_put(type->owner);
1885 return err;
1886 }
1887 } else if (ha[NFTA_HOOK_DEV] || ha[NFTA_HOOK_DEVS]) {
1888 module_put(type->owner);
1889 return -EOPNOTSUPP;
1890 }
1891
1892 return 0;
1893 }
1894
1895 static void nft_chain_release_hook(struct nft_chain_hook *hook)
1896 {
1897 struct nft_hook *h, *next;
1898
1899 list_for_each_entry_safe(h, next, &hook->list, list) {
1900 list_del(&h->list);
1901 kfree(h);
1902 }
1903 module_put(hook->type->owner);
1904 }
1905
1906 struct nft_rules_old {
1907 struct rcu_head h;
1908 struct nft_rule **start;
1909 };
1910
1911 static struct nft_rule **nf_tables_chain_alloc_rules(const struct nft_chain *chain,
1912 unsigned int alloc)
1913 {
1914 if (alloc > INT_MAX)
1915 return NULL;
1916
1917 alloc += 1; /* NULL, ends rules */
1918 if (sizeof(struct nft_rule *) > INT_MAX / alloc)
1919 return NULL;
1920
1921 alloc *= sizeof(struct nft_rule *);
1922 alloc += sizeof(struct nft_rules_old);
1923
1924 return kvmalloc(alloc, GFP_KERNEL);
1925 }
1926
1927 static void nft_basechain_hook_init(struct nf_hook_ops *ops, u8 family,
1928 const struct nft_chain_hook *hook,
1929 struct nft_chain *chain)
1930 {
1931 ops->pf = family;
1932 ops->hooknum = hook->num;
1933 ops->priority = hook->priority;
1934 ops->priv = chain;
1935 ops->hook = hook->type->hooks[ops->hooknum];
1936 }
1937
1938 static int nft_basechain_init(struct nft_base_chain *basechain, u8 family,
1939 struct nft_chain_hook *hook, u32 flags)
1940 {
1941 struct nft_chain *chain;
1942 struct nft_hook *h;
1943
1944 basechain->type = hook->type;
1945 INIT_LIST_HEAD(&basechain->hook_list);
1946 chain = &basechain->chain;
1947
1948 if (nft_base_chain_netdev(family, hook->num)) {
1949 list_splice_init(&hook->list, &basechain->hook_list);
1950 list_for_each_entry(h, &basechain->hook_list, list)
1951 nft_basechain_hook_init(&h->ops, family, hook, chain);
1952
1953 basechain->ops.hooknum = hook->num;
1954 basechain->ops.priority = hook->priority;
1955 } else {
1956 nft_basechain_hook_init(&basechain->ops, family, hook, chain);
1957 }
1958
1959 chain->flags |= NFT_CHAIN_BASE | flags;
1960 basechain->policy = NF_ACCEPT;
1961 if (chain->flags & NFT_CHAIN_HW_OFFLOAD &&
1962 nft_chain_offload_priority(basechain) < 0)
1963 return -EOPNOTSUPP;
1964
1965 flow_block_init(&basechain->flow_block);
1966
1967 return 0;
1968 }
1969
1970 static int nft_chain_add(struct nft_table *table, struct nft_chain *chain)
1971 {
1972 int err;
1973
1974 err = rhltable_insert_key(&table->chains_ht, chain->name,
1975 &chain->rhlhead, nft_chain_ht_params);
1976 if (err)
1977 return err;
1978
1979 list_add_tail_rcu(&chain->list, &table->chains);
1980
1981 return 0;
1982 }
1983
1984 static u64 chain_id;
1985
1986 static int nf_tables_addchain(struct nft_ctx *ctx, u8 family, u8 genmask,
1987 u8 policy, u32 flags)
1988 {
1989 const struct nlattr * const *nla = ctx->nla;
1990 struct nft_table *table = ctx->table;
1991 struct nft_base_chain *basechain;
1992 struct nft_stats __percpu *stats;
1993 struct net *net = ctx->net;
1994 char name[NFT_NAME_MAXLEN];
1995 struct nft_trans *trans;
1996 struct nft_chain *chain;
1997 struct nft_rule **rules;
1998 int err;
1999
2000 if (table->use == UINT_MAX)
2001 return -EOVERFLOW;
2002
2003 if (nla[NFTA_CHAIN_HOOK]) {
2004 struct nft_chain_hook hook;
2005
2006 if (flags & NFT_CHAIN_BINDING)
2007 return -EOPNOTSUPP;
2008
2009 err = nft_chain_parse_hook(net, nla, &hook, family, true);
2010 if (err < 0)
2011 return err;
2012
2013 basechain = kzalloc(sizeof(*basechain), GFP_KERNEL);
2014 if (basechain == NULL) {
2015 nft_chain_release_hook(&hook);
2016 return -ENOMEM;
2017 }
2018 chain = &basechain->chain;
2019
2020 if (nla[NFTA_CHAIN_COUNTERS]) {
2021 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2022 if (IS_ERR(stats)) {
2023 nft_chain_release_hook(&hook);
2024 kfree(basechain);
2025 return PTR_ERR(stats);
2026 }
2027 rcu_assign_pointer(basechain->stats, stats);
2028 static_branch_inc(&nft_counters_enabled);
2029 }
2030
2031 err = nft_basechain_init(basechain, family, &hook, flags);
2032 if (err < 0) {
2033 nft_chain_release_hook(&hook);
2034 kfree(basechain);
2035 return err;
2036 }
2037 } else {
2038 if (flags & NFT_CHAIN_BASE)
2039 return -EINVAL;
2040 if (flags & NFT_CHAIN_HW_OFFLOAD)
2041 return -EOPNOTSUPP;
2042
2043 chain = kzalloc(sizeof(*chain), GFP_KERNEL);
2044 if (chain == NULL)
2045 return -ENOMEM;
2046
2047 chain->flags = flags;
2048 }
2049 ctx->chain = chain;
2050
2051 INIT_LIST_HEAD(&chain->rules);
2052 chain->handle = nf_tables_alloc_handle(table);
2053 chain->table = table;
2054
2055 if (nla[NFTA_CHAIN_NAME]) {
2056 chain->name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL);
2057 } else {
2058 if (!(flags & NFT_CHAIN_BINDING)) {
2059 err = -EINVAL;
2060 goto err_destroy_chain;
2061 }
2062
2063 snprintf(name, sizeof(name), "__chain%llu", ++chain_id);
2064 chain->name = kstrdup(name, GFP_KERNEL);
2065 }
2066
2067 if (!chain->name) {
2068 err = -ENOMEM;
2069 goto err_destroy_chain;
2070 }
2071
2072 if (nla[NFTA_CHAIN_USERDATA]) {
2073 chain->udata = nla_memdup(nla[NFTA_CHAIN_USERDATA], GFP_KERNEL);
2074 if (chain->udata == NULL) {
2075 err = -ENOMEM;
2076 goto err_destroy_chain;
2077 }
2078 chain->udlen = nla_len(nla[NFTA_CHAIN_USERDATA]);
2079 }
2080
2081 rules = nf_tables_chain_alloc_rules(chain, 0);
2082 if (!rules) {
2083 err = -ENOMEM;
2084 goto err_destroy_chain;
2085 }
2086
2087 *rules = NULL;
2088 rcu_assign_pointer(chain->rules_gen_0, rules);
2089 rcu_assign_pointer(chain->rules_gen_1, rules);
2090
2091 err = nf_tables_register_hook(net, table, chain);
2092 if (err < 0)
2093 goto err_destroy_chain;
2094
2095 trans = nft_trans_chain_add(ctx, NFT_MSG_NEWCHAIN);
2096 if (IS_ERR(trans)) {
2097 err = PTR_ERR(trans);
2098 goto err_unregister_hook;
2099 }
2100
2101 nft_trans_chain_policy(trans) = NFT_CHAIN_POLICY_UNSET;
2102 if (nft_is_base_chain(chain))
2103 nft_trans_chain_policy(trans) = policy;
2104
2105 err = nft_chain_add(table, chain);
2106 if (err < 0) {
2107 nft_trans_destroy(trans);
2108 goto err_unregister_hook;
2109 }
2110
2111 table->use++;
2112
2113 return 0;
2114 err_unregister_hook:
2115 nf_tables_unregister_hook(net, table, chain);
2116 err_destroy_chain:
2117 nf_tables_chain_destroy(ctx);
2118
2119 return err;
2120 }
2121
2122 static bool nft_hook_list_equal(struct list_head *hook_list1,
2123 struct list_head *hook_list2)
2124 {
2125 struct nft_hook *hook;
2126 int n = 0, m = 0;
2127
2128 n = 0;
2129 list_for_each_entry(hook, hook_list2, list) {
2130 if (!nft_hook_list_find(hook_list1, hook))
2131 return false;
2132
2133 n++;
2134 }
2135 list_for_each_entry(hook, hook_list1, list)
2136 m++;
2137
2138 return n == m;
2139 }
2140
2141 static int nf_tables_updchain(struct nft_ctx *ctx, u8 genmask, u8 policy,
2142 u32 flags, const struct nlattr *attr,
2143 struct netlink_ext_ack *extack)
2144 {
2145 const struct nlattr * const *nla = ctx->nla;
2146 struct nft_table *table = ctx->table;
2147 struct nft_chain *chain = ctx->chain;
2148 struct nft_base_chain *basechain;
2149 struct nft_stats *stats = NULL;
2150 struct nft_chain_hook hook;
2151 struct nf_hook_ops *ops;
2152 struct nft_trans *trans;
2153 int err;
2154
2155 if (chain->flags ^ flags)
2156 return -EOPNOTSUPP;
2157
2158 if (nla[NFTA_CHAIN_HOOK]) {
2159 if (!nft_is_base_chain(chain)) {
2160 NL_SET_BAD_ATTR(extack, attr);
2161 return -EEXIST;
2162 }
2163 err = nft_chain_parse_hook(ctx->net, nla, &hook, ctx->family,
2164 false);
2165 if (err < 0)
2166 return err;
2167
2168 basechain = nft_base_chain(chain);
2169 if (basechain->type != hook.type) {
2170 nft_chain_release_hook(&hook);
2171 NL_SET_BAD_ATTR(extack, attr);
2172 return -EEXIST;
2173 }
2174
2175 if (nft_base_chain_netdev(ctx->family, hook.num)) {
2176 if (!nft_hook_list_equal(&basechain->hook_list,
2177 &hook.list)) {
2178 nft_chain_release_hook(&hook);
2179 NL_SET_BAD_ATTR(extack, attr);
2180 return -EEXIST;
2181 }
2182 } else {
2183 ops = &basechain->ops;
2184 if (ops->hooknum != hook.num ||
2185 ops->priority != hook.priority) {
2186 nft_chain_release_hook(&hook);
2187 NL_SET_BAD_ATTR(extack, attr);
2188 return -EEXIST;
2189 }
2190 }
2191 nft_chain_release_hook(&hook);
2192 }
2193
2194 if (nla[NFTA_CHAIN_HANDLE] &&
2195 nla[NFTA_CHAIN_NAME]) {
2196 struct nft_chain *chain2;
2197
2198 chain2 = nft_chain_lookup(ctx->net, table,
2199 nla[NFTA_CHAIN_NAME], genmask);
2200 if (!IS_ERR(chain2)) {
2201 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2202 return -EEXIST;
2203 }
2204 }
2205
2206 if (nla[NFTA_CHAIN_COUNTERS]) {
2207 if (!nft_is_base_chain(chain))
2208 return -EOPNOTSUPP;
2209
2210 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2211 if (IS_ERR(stats))
2212 return PTR_ERR(stats);
2213 }
2214
2215 err = -ENOMEM;
2216 trans = nft_trans_alloc(ctx, NFT_MSG_NEWCHAIN,
2217 sizeof(struct nft_trans_chain));
2218 if (trans == NULL)
2219 goto err;
2220
2221 nft_trans_chain_stats(trans) = stats;
2222 nft_trans_chain_update(trans) = true;
2223
2224 if (nla[NFTA_CHAIN_POLICY])
2225 nft_trans_chain_policy(trans) = policy;
2226 else
2227 nft_trans_chain_policy(trans) = -1;
2228
2229 if (nla[NFTA_CHAIN_HANDLE] &&
2230 nla[NFTA_CHAIN_NAME]) {
2231 struct nft_trans *tmp;
2232 char *name;
2233
2234 err = -ENOMEM;
2235 name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL);
2236 if (!name)
2237 goto err;
2238
2239 err = -EEXIST;
2240 list_for_each_entry(tmp, &ctx->net->nft.commit_list, list) {
2241 if (tmp->msg_type == NFT_MSG_NEWCHAIN &&
2242 tmp->ctx.table == table &&
2243 nft_trans_chain_update(tmp) &&
2244 nft_trans_chain_name(tmp) &&
2245 strcmp(name, nft_trans_chain_name(tmp)) == 0) {
2246 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2247 kfree(name);
2248 goto err;
2249 }
2250 }
2251
2252 nft_trans_chain_name(trans) = name;
2253 }
2254 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
2255
2256 return 0;
2257 err:
2258 free_percpu(stats);
2259 kfree(trans);
2260 return err;
2261 }
2262
2263 static struct nft_chain *nft_chain_lookup_byid(const struct net *net,
2264 const struct nlattr *nla)
2265 {
2266 u32 id = ntohl(nla_get_be32(nla));
2267 struct nft_trans *trans;
2268
2269 list_for_each_entry(trans, &net->nft.commit_list, list) {
2270 struct nft_chain *chain = trans->ctx.chain;
2271
2272 if (trans->msg_type == NFT_MSG_NEWCHAIN &&
2273 id == nft_trans_chain_id(trans))
2274 return chain;
2275 }
2276 return ERR_PTR(-ENOENT);
2277 }
2278
2279 static int nf_tables_newchain(struct net *net, struct sock *nlsk,
2280 struct sk_buff *skb, const struct nlmsghdr *nlh,
2281 const struct nlattr * const nla[],
2282 struct netlink_ext_ack *extack)
2283 {
2284 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2285 u8 genmask = nft_genmask_next(net);
2286 int family = nfmsg->nfgen_family;
2287 struct nft_chain *chain = NULL;
2288 const struct nlattr *attr;
2289 struct nft_table *table;
2290 u8 policy = NF_ACCEPT;
2291 struct nft_ctx ctx;
2292 u64 handle = 0;
2293 u32 flags = 0;
2294
2295 lockdep_assert_held(&net->nft.commit_mutex);
2296
2297 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask);
2298 if (IS_ERR(table)) {
2299 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2300 return PTR_ERR(table);
2301 }
2302
2303 chain = NULL;
2304 attr = nla[NFTA_CHAIN_NAME];
2305
2306 if (nla[NFTA_CHAIN_HANDLE]) {
2307 handle = be64_to_cpu(nla_get_be64(nla[NFTA_CHAIN_HANDLE]));
2308 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2309 if (IS_ERR(chain)) {
2310 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_HANDLE]);
2311 return PTR_ERR(chain);
2312 }
2313 attr = nla[NFTA_CHAIN_HANDLE];
2314 } else if (nla[NFTA_CHAIN_NAME]) {
2315 chain = nft_chain_lookup(net, table, attr, genmask);
2316 if (IS_ERR(chain)) {
2317 if (PTR_ERR(chain) != -ENOENT) {
2318 NL_SET_BAD_ATTR(extack, attr);
2319 return PTR_ERR(chain);
2320 }
2321 chain = NULL;
2322 }
2323 } else if (!nla[NFTA_CHAIN_ID]) {
2324 return -EINVAL;
2325 }
2326
2327 if (nla[NFTA_CHAIN_POLICY]) {
2328 if (chain != NULL &&
2329 !nft_is_base_chain(chain)) {
2330 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2331 return -EOPNOTSUPP;
2332 }
2333
2334 if (chain == NULL &&
2335 nla[NFTA_CHAIN_HOOK] == NULL) {
2336 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2337 return -EOPNOTSUPP;
2338 }
2339
2340 policy = ntohl(nla_get_be32(nla[NFTA_CHAIN_POLICY]));
2341 switch (policy) {
2342 case NF_DROP:
2343 case NF_ACCEPT:
2344 break;
2345 default:
2346 return -EINVAL;
2347 }
2348 }
2349
2350 if (nla[NFTA_CHAIN_FLAGS])
2351 flags = ntohl(nla_get_be32(nla[NFTA_CHAIN_FLAGS]));
2352 else if (chain)
2353 flags = chain->flags;
2354
2355 if (flags & ~NFT_CHAIN_FLAGS)
2356 return -EOPNOTSUPP;
2357
2358 nft_ctx_init(&ctx, net, skb, nlh, family, table, chain, nla);
2359
2360 if (chain != NULL) {
2361 if (nlh->nlmsg_flags & NLM_F_EXCL) {
2362 NL_SET_BAD_ATTR(extack, attr);
2363 return -EEXIST;
2364 }
2365 if (nlh->nlmsg_flags & NLM_F_REPLACE)
2366 return -EOPNOTSUPP;
2367
2368 flags |= chain->flags & NFT_CHAIN_BASE;
2369 return nf_tables_updchain(&ctx, genmask, policy, flags, attr,
2370 extack);
2371 }
2372
2373 return nf_tables_addchain(&ctx, family, genmask, policy, flags);
2374 }
2375
2376 static int nf_tables_delchain(struct net *net, struct sock *nlsk,
2377 struct sk_buff *skb, const struct nlmsghdr *nlh,
2378 const struct nlattr * const nla[],
2379 struct netlink_ext_ack *extack)
2380 {
2381 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2382 u8 genmask = nft_genmask_next(net);
2383 int family = nfmsg->nfgen_family;
2384 const struct nlattr *attr;
2385 struct nft_table *table;
2386 struct nft_chain *chain;
2387 struct nft_rule *rule;
2388 struct nft_ctx ctx;
2389 u64 handle;
2390 u32 use;
2391 int err;
2392
2393 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask);
2394 if (IS_ERR(table)) {
2395 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2396 return PTR_ERR(table);
2397 }
2398
2399 if (nla[NFTA_CHAIN_HANDLE]) {
2400 attr = nla[NFTA_CHAIN_HANDLE];
2401 handle = be64_to_cpu(nla_get_be64(attr));
2402 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2403 } else {
2404 attr = nla[NFTA_CHAIN_NAME];
2405 chain = nft_chain_lookup(net, table, attr, genmask);
2406 }
2407 if (IS_ERR(chain)) {
2408 NL_SET_BAD_ATTR(extack, attr);
2409 return PTR_ERR(chain);
2410 }
2411
2412 if (nlh->nlmsg_flags & NLM_F_NONREC &&
2413 chain->use > 0)
2414 return -EBUSY;
2415
2416 nft_ctx_init(&ctx, net, skb, nlh, family, table, chain, nla);
2417
2418 use = chain->use;
2419 list_for_each_entry(rule, &chain->rules, list) {
2420 if (!nft_is_active_next(net, rule))
2421 continue;
2422 use--;
2423
2424 err = nft_delrule(&ctx, rule);
2425 if (err < 0)
2426 return err;
2427 }
2428
2429 /* There are rules and elements that are still holding references to us,
2430 * we cannot do a recursive removal in this case.
2431 */
2432 if (use > 0) {
2433 NL_SET_BAD_ATTR(extack, attr);
2434 return -EBUSY;
2435 }
2436
2437 return nft_delchain(&ctx);
2438 }
2439
2440 /*
2441 * Expressions
2442 */
2443
2444 /**
2445 * nft_register_expr - register nf_tables expr type
2446 * @type: expr type
2447 *
2448 * Registers the expr type for use with nf_tables. Returns zero on
2449 * success or a negative errno code otherwise.
2450 */
2451 int nft_register_expr(struct nft_expr_type *type)
2452 {
2453 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2454 if (type->family == NFPROTO_UNSPEC)
2455 list_add_tail_rcu(&type->list, &nf_tables_expressions);
2456 else
2457 list_add_rcu(&type->list, &nf_tables_expressions);
2458 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2459 return 0;
2460 }
2461 EXPORT_SYMBOL_GPL(nft_register_expr);
2462
2463 /**
2464 * nft_unregister_expr - unregister nf_tables expr type
2465 * @type: expr type
2466 *
2467 * Unregisters the expr typefor use with nf_tables.
2468 */
2469 void nft_unregister_expr(struct nft_expr_type *type)
2470 {
2471 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2472 list_del_rcu(&type->list);
2473 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2474 }
2475 EXPORT_SYMBOL_GPL(nft_unregister_expr);
2476
2477 static const struct nft_expr_type *__nft_expr_type_get(u8 family,
2478 struct nlattr *nla)
2479 {
2480 const struct nft_expr_type *type, *candidate = NULL;
2481
2482 list_for_each_entry(type, &nf_tables_expressions, list) {
2483 if (!nla_strcmp(nla, type->name)) {
2484 if (!type->family && !candidate)
2485 candidate = type;
2486 else if (type->family == family)
2487 candidate = type;
2488 }
2489 }
2490 return candidate;
2491 }
2492
2493 #ifdef CONFIG_MODULES
2494 static int nft_expr_type_request_module(struct net *net, u8 family,
2495 struct nlattr *nla)
2496 {
2497 if (nft_request_module(net, "nft-expr-%u-%.*s", family,
2498 nla_len(nla), (char *)nla_data(nla)) == -EAGAIN)
2499 return -EAGAIN;
2500
2501 return 0;
2502 }
2503 #endif
2504
2505 static const struct nft_expr_type *nft_expr_type_get(struct net *net,
2506 u8 family,
2507 struct nlattr *nla)
2508 {
2509 const struct nft_expr_type *type;
2510
2511 if (nla == NULL)
2512 return ERR_PTR(-EINVAL);
2513
2514 type = __nft_expr_type_get(family, nla);
2515 if (type != NULL && try_module_get(type->owner))
2516 return type;
2517
2518 lockdep_nfnl_nft_mutex_not_held();
2519 #ifdef CONFIG_MODULES
2520 if (type == NULL) {
2521 if (nft_expr_type_request_module(net, family, nla) == -EAGAIN)
2522 return ERR_PTR(-EAGAIN);
2523
2524 if (nft_request_module(net, "nft-expr-%.*s",
2525 nla_len(nla),
2526 (char *)nla_data(nla)) == -EAGAIN)
2527 return ERR_PTR(-EAGAIN);
2528 }
2529 #endif
2530 return ERR_PTR(-ENOENT);
2531 }
2532
2533 static const struct nla_policy nft_expr_policy[NFTA_EXPR_MAX + 1] = {
2534 [NFTA_EXPR_NAME] = { .type = NLA_STRING,
2535 .len = NFT_MODULE_AUTOLOAD_LIMIT },
2536 [NFTA_EXPR_DATA] = { .type = NLA_NESTED },
2537 };
2538
2539 static int nf_tables_fill_expr_info(struct sk_buff *skb,
2540 const struct nft_expr *expr)
2541 {
2542 if (nla_put_string(skb, NFTA_EXPR_NAME, expr->ops->type->name))
2543 goto nla_put_failure;
2544
2545 if (expr->ops->dump) {
2546 struct nlattr *data = nla_nest_start_noflag(skb,
2547 NFTA_EXPR_DATA);
2548 if (data == NULL)
2549 goto nla_put_failure;
2550 if (expr->ops->dump(skb, expr) < 0)
2551 goto nla_put_failure;
2552 nla_nest_end(skb, data);
2553 }
2554
2555 return skb->len;
2556
2557 nla_put_failure:
2558 return -1;
2559 };
2560
2561 int nft_expr_dump(struct sk_buff *skb, unsigned int attr,
2562 const struct nft_expr *expr)
2563 {
2564 struct nlattr *nest;
2565
2566 nest = nla_nest_start_noflag(skb, attr);
2567 if (!nest)
2568 goto nla_put_failure;
2569 if (nf_tables_fill_expr_info(skb, expr) < 0)
2570 goto nla_put_failure;
2571 nla_nest_end(skb, nest);
2572 return 0;
2573
2574 nla_put_failure:
2575 return -1;
2576 }
2577
2578 struct nft_expr_info {
2579 const struct nft_expr_ops *ops;
2580 const struct nlattr *attr;
2581 struct nlattr *tb[NFT_EXPR_MAXATTR + 1];
2582 };
2583
2584 static int nf_tables_expr_parse(const struct nft_ctx *ctx,
2585 const struct nlattr *nla,
2586 struct nft_expr_info *info)
2587 {
2588 const struct nft_expr_type *type;
2589 const struct nft_expr_ops *ops;
2590 struct nlattr *tb[NFTA_EXPR_MAX + 1];
2591 int err;
2592
2593 err = nla_parse_nested_deprecated(tb, NFTA_EXPR_MAX, nla,
2594 nft_expr_policy, NULL);
2595 if (err < 0)
2596 return err;
2597
2598 type = nft_expr_type_get(ctx->net, ctx->family, tb[NFTA_EXPR_NAME]);
2599 if (IS_ERR(type))
2600 return PTR_ERR(type);
2601
2602 if (tb[NFTA_EXPR_DATA]) {
2603 err = nla_parse_nested_deprecated(info->tb, type->maxattr,
2604 tb[NFTA_EXPR_DATA],
2605 type->policy, NULL);
2606 if (err < 0)
2607 goto err1;
2608 } else
2609 memset(info->tb, 0, sizeof(info->tb[0]) * (type->maxattr + 1));
2610
2611 if (type->select_ops != NULL) {
2612 ops = type->select_ops(ctx,
2613 (const struct nlattr * const *)info->tb);
2614 if (IS_ERR(ops)) {
2615 err = PTR_ERR(ops);
2616 #ifdef CONFIG_MODULES
2617 if (err == -EAGAIN)
2618 if (nft_expr_type_request_module(ctx->net,
2619 ctx->family,
2620 tb[NFTA_EXPR_NAME]) != -EAGAIN)
2621 err = -ENOENT;
2622 #endif
2623 goto err1;
2624 }
2625 } else
2626 ops = type->ops;
2627
2628 info->attr = nla;
2629 info->ops = ops;
2630
2631 return 0;
2632
2633 err1:
2634 module_put(type->owner);
2635 return err;
2636 }
2637
2638 static int nf_tables_newexpr(const struct nft_ctx *ctx,
2639 const struct nft_expr_info *info,
2640 struct nft_expr *expr)
2641 {
2642 const struct nft_expr_ops *ops = info->ops;
2643 int err;
2644
2645 expr->ops = ops;
2646 if (ops->init) {
2647 err = ops->init(ctx, expr, (const struct nlattr **)info->tb);
2648 if (err < 0)
2649 goto err1;
2650 }
2651
2652 return 0;
2653 err1:
2654 expr->ops = NULL;
2655 return err;
2656 }
2657
2658 static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
2659 struct nft_expr *expr)
2660 {
2661 const struct nft_expr_type *type = expr->ops->type;
2662
2663 if (expr->ops->destroy)
2664 expr->ops->destroy(ctx, expr);
2665 module_put(type->owner);
2666 }
2667
2668 static struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
2669 const struct nlattr *nla)
2670 {
2671 struct nft_expr_info info;
2672 struct nft_expr *expr;
2673 struct module *owner;
2674 int err;
2675
2676 err = nf_tables_expr_parse(ctx, nla, &info);
2677 if (err < 0)
2678 goto err1;
2679
2680 err = -ENOMEM;
2681 expr = kzalloc(info.ops->size, GFP_KERNEL);
2682 if (expr == NULL)
2683 goto err2;
2684
2685 err = nf_tables_newexpr(ctx, &info, expr);
2686 if (err < 0)
2687 goto err3;
2688
2689 return expr;
2690 err3:
2691 kfree(expr);
2692 err2:
2693 owner = info.ops->type->owner;
2694 if (info.ops->type->release_ops)
2695 info.ops->type->release_ops(info.ops);
2696
2697 module_put(owner);
2698 err1:
2699 return ERR_PTR(err);
2700 }
2701
2702 int nft_expr_clone(struct nft_expr *dst, struct nft_expr *src)
2703 {
2704 int err;
2705
2706 if (src->ops->clone) {
2707 dst->ops = src->ops;
2708 err = src->ops->clone(dst, src);
2709 if (err < 0)
2710 return err;
2711 } else {
2712 memcpy(dst, src, src->ops->size);
2713 }
2714
2715 __module_get(src->ops->type->owner);
2716
2717 return 0;
2718 }
2719
2720 void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
2721 {
2722 nf_tables_expr_destroy(ctx, expr);
2723 kfree(expr);
2724 }
2725
2726 /*
2727 * Rules
2728 */
2729
2730 static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
2731 u64 handle)
2732 {
2733 struct nft_rule *rule;
2734
2735 // FIXME: this sucks
2736 list_for_each_entry_rcu(rule, &chain->rules, list) {
2737 if (handle == rule->handle)
2738 return rule;
2739 }
2740
2741 return ERR_PTR(-ENOENT);
2742 }
2743
2744 static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
2745 const struct nlattr *nla)
2746 {
2747 if (nla == NULL)
2748 return ERR_PTR(-EINVAL);
2749
2750 return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
2751 }
2752
2753 static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
2754 [NFTA_RULE_TABLE] = { .type = NLA_STRING,
2755 .len = NFT_TABLE_MAXNAMELEN - 1 },
2756 [NFTA_RULE_CHAIN] = { .type = NLA_STRING,
2757 .len = NFT_CHAIN_MAXNAMELEN - 1 },
2758 [NFTA_RULE_HANDLE] = { .type = NLA_U64 },
2759 [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
2760 [NFTA_RULE_COMPAT] = { .type = NLA_NESTED },
2761 [NFTA_RULE_POSITION] = { .type = NLA_U64 },
2762 [NFTA_RULE_USERDATA] = { .type = NLA_BINARY,
2763 .len = NFT_USERDATA_MAXLEN },
2764 [NFTA_RULE_ID] = { .type = NLA_U32 },
2765 [NFTA_RULE_POSITION_ID] = { .type = NLA_U32 },
2766 [NFTA_RULE_CHAIN_ID] = { .type = NLA_U32 },
2767 };
2768
2769 static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
2770 u32 portid, u32 seq, int event,
2771 u32 flags, int family,
2772 const struct nft_table *table,
2773 const struct nft_chain *chain,
2774 const struct nft_rule *rule,
2775 const struct nft_rule *prule)
2776 {
2777 struct nlmsghdr *nlh;
2778 struct nfgenmsg *nfmsg;
2779 const struct nft_expr *expr, *next;
2780 struct nlattr *list;
2781 u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
2782
2783 nlh = nlmsg_put(skb, portid, seq, type, sizeof(struct nfgenmsg), flags);
2784 if (nlh == NULL)
2785 goto nla_put_failure;
2786
2787 nfmsg = nlmsg_data(nlh);
2788 nfmsg->nfgen_family = family;
2789 nfmsg->version = NFNETLINK_V0;
2790 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
2791
2792 if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
2793 goto nla_put_failure;
2794 if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
2795 goto nla_put_failure;
2796 if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
2797 NFTA_RULE_PAD))
2798 goto nla_put_failure;
2799
2800 if (event != NFT_MSG_DELRULE && prule) {
2801 if (nla_put_be64(skb, NFTA_RULE_POSITION,
2802 cpu_to_be64(prule->handle),
2803 NFTA_RULE_PAD))
2804 goto nla_put_failure;
2805 }
2806
2807 list = nla_nest_start_noflag(skb, NFTA_RULE_EXPRESSIONS);
2808 if (list == NULL)
2809 goto nla_put_failure;
2810 nft_rule_for_each_expr(expr, next, rule) {
2811 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
2812 goto nla_put_failure;
2813 }
2814 nla_nest_end(skb, list);
2815
2816 if (rule->udata) {
2817 struct nft_userdata *udata = nft_userdata(rule);
2818 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
2819 udata->data) < 0)
2820 goto nla_put_failure;
2821 }
2822
2823 nlmsg_end(skb, nlh);
2824 return 0;
2825
2826 nla_put_failure:
2827 nlmsg_trim(skb, nlh);
2828 return -1;
2829 }
2830
2831 static void nf_tables_rule_notify(const struct nft_ctx *ctx,
2832 const struct nft_rule *rule, int event)
2833 {
2834 struct sk_buff *skb;
2835 int err;
2836 char *buf = kasprintf(GFP_KERNEL, "%s:%llu;%s:%llu",
2837 ctx->table->name, ctx->table->handle,
2838 ctx->chain->name, ctx->chain->handle);
2839
2840 audit_log_nfcfg(buf,
2841 ctx->family,
2842 rule->handle,
2843 event == NFT_MSG_NEWRULE ?
2844 AUDIT_NFT_OP_RULE_REGISTER :
2845 AUDIT_NFT_OP_RULE_UNREGISTER,
2846 GFP_KERNEL);
2847 kfree(buf);
2848
2849 if (!ctx->report &&
2850 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
2851 return;
2852
2853 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
2854 if (skb == NULL)
2855 goto err;
2856
2857 err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
2858 event, 0, ctx->family, ctx->table,
2859 ctx->chain, rule, NULL);
2860 if (err < 0) {
2861 kfree_skb(skb);
2862 goto err;
2863 }
2864
2865 nft_notify_enqueue(skb, ctx->report, &ctx->net->nft.notify_list);
2866 return;
2867 err:
2868 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
2869 }
2870
2871 struct nft_rule_dump_ctx {
2872 char *table;
2873 char *chain;
2874 };
2875
2876 static int __nf_tables_dump_rules(struct sk_buff *skb,
2877 unsigned int *idx,
2878 struct netlink_callback *cb,
2879 const struct nft_table *table,
2880 const struct nft_chain *chain)
2881 {
2882 struct net *net = sock_net(skb->sk);
2883 const struct nft_rule *rule, *prule;
2884 unsigned int s_idx = cb->args[0];
2885
2886 prule = NULL;
2887 list_for_each_entry_rcu(rule, &chain->rules, list) {
2888 if (!nft_is_active(net, rule))
2889 goto cont_skip;
2890 if (*idx < s_idx)
2891 goto cont;
2892 if (*idx > s_idx) {
2893 memset(&cb->args[1], 0,
2894 sizeof(cb->args) - sizeof(cb->args[0]));
2895 }
2896 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
2897 cb->nlh->nlmsg_seq,
2898 NFT_MSG_NEWRULE,
2899 NLM_F_MULTI | NLM_F_APPEND,
2900 table->family,
2901 table, chain, rule, prule) < 0)
2902 return 1;
2903
2904 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
2905 cont:
2906 prule = rule;
2907 cont_skip:
2908 (*idx)++;
2909 }
2910 return 0;
2911 }
2912
2913 static int nf_tables_dump_rules(struct sk_buff *skb,
2914 struct netlink_callback *cb)
2915 {
2916 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2917 const struct nft_rule_dump_ctx *ctx = cb->data;
2918 struct nft_table *table;
2919 const struct nft_chain *chain;
2920 unsigned int idx = 0;
2921 struct net *net = sock_net(skb->sk);
2922 int family = nfmsg->nfgen_family;
2923
2924 rcu_read_lock();
2925 cb->seq = net->nft.base_seq;
2926
2927 list_for_each_entry_rcu(table, &net->nft.tables, list) {
2928 if (family != NFPROTO_UNSPEC && family != table->family)
2929 continue;
2930
2931 if (ctx && ctx->table && strcmp(ctx->table, table->name) != 0)
2932 continue;
2933
2934 if (ctx && ctx->table && ctx->chain) {
2935 struct rhlist_head *list, *tmp;
2936
2937 list = rhltable_lookup(&table->chains_ht, ctx->chain,
2938 nft_chain_ht_params);
2939 if (!list)
2940 goto done;
2941
2942 rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
2943 if (!nft_is_active(net, chain))
2944 continue;
2945 __nf_tables_dump_rules(skb, &idx,
2946 cb, table, chain);
2947 break;
2948 }
2949 goto done;
2950 }
2951
2952 list_for_each_entry_rcu(chain, &table->chains, list) {
2953 if (__nf_tables_dump_rules(skb, &idx, cb, table, chain))
2954 goto done;
2955 }
2956
2957 if (ctx && ctx->table)
2958 break;
2959 }
2960 done:
2961 rcu_read_unlock();
2962
2963 cb->args[0] = idx;
2964 return skb->len;
2965 }
2966
2967 static int nf_tables_dump_rules_start(struct netlink_callback *cb)
2968 {
2969 const struct nlattr * const *nla = cb->data;
2970 struct nft_rule_dump_ctx *ctx = NULL;
2971
2972 if (nla[NFTA_RULE_TABLE] || nla[NFTA_RULE_CHAIN]) {
2973 ctx = kzalloc(sizeof(*ctx), GFP_ATOMIC);
2974 if (!ctx)
2975 return -ENOMEM;
2976
2977 if (nla[NFTA_RULE_TABLE]) {
2978 ctx->table = nla_strdup(nla[NFTA_RULE_TABLE],
2979 GFP_ATOMIC);
2980 if (!ctx->table) {
2981 kfree(ctx);
2982 return -ENOMEM;
2983 }
2984 }
2985 if (nla[NFTA_RULE_CHAIN]) {
2986 ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN],
2987 GFP_ATOMIC);
2988 if (!ctx->chain) {
2989 kfree(ctx->table);
2990 kfree(ctx);
2991 return -ENOMEM;
2992 }
2993 }
2994 }
2995
2996 cb->data = ctx;
2997 return 0;
2998 }
2999
3000 static int nf_tables_dump_rules_done(struct netlink_callback *cb)
3001 {
3002 struct nft_rule_dump_ctx *ctx = cb->data;
3003
3004 if (ctx) {
3005 kfree(ctx->table);
3006 kfree(ctx->chain);
3007 kfree(ctx);
3008 }
3009 return 0;
3010 }
3011
3012 /* called with rcu_read_lock held */
3013 static int nf_tables_getrule(struct net *net, struct sock *nlsk,
3014 struct sk_buff *skb, const struct nlmsghdr *nlh,
3015 const struct nlattr * const nla[],
3016 struct netlink_ext_ack *extack)
3017 {
3018 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3019 u8 genmask = nft_genmask_cur(net);
3020 const struct nft_chain *chain;
3021 const struct nft_rule *rule;
3022 struct nft_table *table;
3023 struct sk_buff *skb2;
3024 int family = nfmsg->nfgen_family;
3025 int err;
3026
3027 if (nlh->nlmsg_flags & NLM_F_DUMP) {
3028 struct netlink_dump_control c = {
3029 .start= nf_tables_dump_rules_start,
3030 .dump = nf_tables_dump_rules,
3031 .done = nf_tables_dump_rules_done,
3032 .module = THIS_MODULE,
3033 .data = (void *)nla,
3034 };
3035
3036 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
3037 }
3038
3039 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask);
3040 if (IS_ERR(table)) {
3041 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3042 return PTR_ERR(table);
3043 }
3044
3045 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
3046 if (IS_ERR(chain)) {
3047 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3048 return PTR_ERR(chain);
3049 }
3050
3051 rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3052 if (IS_ERR(rule)) {
3053 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3054 return PTR_ERR(rule);
3055 }
3056
3057 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3058 if (!skb2)
3059 return -ENOMEM;
3060
3061 err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
3062 nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
3063 family, table, chain, rule, NULL);
3064 if (err < 0)
3065 goto err_fill_rule_info;
3066
3067 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
3068
3069 err_fill_rule_info:
3070 kfree_skb(skb2);
3071 return err;
3072 }
3073
3074 static void nf_tables_rule_destroy(const struct nft_ctx *ctx,
3075 struct nft_rule *rule)
3076 {
3077 struct nft_expr *expr, *next;
3078
3079 /*
3080 * Careful: some expressions might not be initialized in case this
3081 * is called on error from nf_tables_newrule().
3082 */
3083 expr = nft_expr_first(rule);
3084 while (nft_expr_more(rule, expr)) {
3085 next = nft_expr_next(expr);
3086 nf_tables_expr_destroy(ctx, expr);
3087 expr = next;
3088 }
3089 kfree(rule);
3090 }
3091
3092 void nf_tables_rule_release(const struct nft_ctx *ctx, struct nft_rule *rule)
3093 {
3094 nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_RELEASE);
3095 nf_tables_rule_destroy(ctx, rule);
3096 }
3097
3098 int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
3099 {
3100 struct nft_expr *expr, *last;
3101 const struct nft_data *data;
3102 struct nft_rule *rule;
3103 int err;
3104
3105 if (ctx->level == NFT_JUMP_STACK_SIZE)
3106 return -EMLINK;
3107
3108 list_for_each_entry(rule, &chain->rules, list) {
3109 if (!nft_is_active_next(ctx->net, rule))
3110 continue;
3111
3112 nft_rule_for_each_expr(expr, last, rule) {
3113 if (!expr->ops->validate)
3114 continue;
3115
3116 err = expr->ops->validate(ctx, expr, &data);
3117 if (err < 0)
3118 return err;
3119 }
3120 }
3121
3122 return 0;
3123 }
3124 EXPORT_SYMBOL_GPL(nft_chain_validate);
3125
3126 static int nft_table_validate(struct net *net, const struct nft_table *table)
3127 {
3128 struct nft_chain *chain;
3129 struct nft_ctx ctx = {
3130 .net = net,
3131 .family = table->family,
3132 };
3133 int err;
3134
3135 list_for_each_entry(chain, &table->chains, list) {
3136 if (!nft_is_base_chain(chain))
3137 continue;
3138
3139 ctx.chain = chain;
3140 err = nft_chain_validate(&ctx, chain);
3141 if (err < 0)
3142 return err;
3143 }
3144
3145 return 0;
3146 }
3147
3148 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3149 const struct nlattr *nla);
3150
3151 #define NFT_RULE_MAXEXPRS 128
3152
3153 static int nf_tables_newrule(struct net *net, struct sock *nlsk,
3154 struct sk_buff *skb, const struct nlmsghdr *nlh,
3155 const struct nlattr * const nla[],
3156 struct netlink_ext_ack *extack)
3157 {
3158 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3159 u8 genmask = nft_genmask_next(net);
3160 struct nft_expr_info *info = NULL;
3161 int family = nfmsg->nfgen_family;
3162 struct nft_flow_rule *flow;
3163 struct nft_table *table;
3164 struct nft_chain *chain;
3165 struct nft_rule *rule, *old_rule = NULL;
3166 struct nft_userdata *udata;
3167 struct nft_trans *trans = NULL;
3168 struct nft_expr *expr;
3169 struct nft_ctx ctx;
3170 struct nlattr *tmp;
3171 unsigned int size, i, n, ulen = 0, usize = 0;
3172 int err, rem;
3173 u64 handle, pos_handle;
3174
3175 lockdep_assert_held(&net->nft.commit_mutex);
3176
3177 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask);
3178 if (IS_ERR(table)) {
3179 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3180 return PTR_ERR(table);
3181 }
3182
3183 if (nla[NFTA_RULE_CHAIN]) {
3184 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3185 genmask);
3186 if (IS_ERR(chain)) {
3187 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3188 return PTR_ERR(chain);
3189 }
3190 if (nft_chain_is_bound(chain))
3191 return -EOPNOTSUPP;
3192
3193 } else if (nla[NFTA_RULE_CHAIN_ID]) {
3194 chain = nft_chain_lookup_byid(net, nla[NFTA_RULE_CHAIN_ID]);
3195 if (IS_ERR(chain)) {
3196 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN_ID]);
3197 return PTR_ERR(chain);
3198 }
3199 } else {
3200 return -EINVAL;
3201 }
3202
3203 if (nla[NFTA_RULE_HANDLE]) {
3204 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
3205 rule = __nft_rule_lookup(chain, handle);
3206 if (IS_ERR(rule)) {
3207 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3208 return PTR_ERR(rule);
3209 }
3210
3211 if (nlh->nlmsg_flags & NLM_F_EXCL) {
3212 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3213 return -EEXIST;
3214 }
3215 if (nlh->nlmsg_flags & NLM_F_REPLACE)
3216 old_rule = rule;
3217 else
3218 return -EOPNOTSUPP;
3219 } else {
3220 if (!(nlh->nlmsg_flags & NLM_F_CREATE) ||
3221 nlh->nlmsg_flags & NLM_F_REPLACE)
3222 return -EINVAL;
3223 handle = nf_tables_alloc_handle(table);
3224
3225 if (chain->use == UINT_MAX)
3226 return -EOVERFLOW;
3227
3228 if (nla[NFTA_RULE_POSITION]) {
3229 pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
3230 old_rule = __nft_rule_lookup(chain, pos_handle);
3231 if (IS_ERR(old_rule)) {
3232 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
3233 return PTR_ERR(old_rule);
3234 }
3235 } else if (nla[NFTA_RULE_POSITION_ID]) {
3236 old_rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_POSITION_ID]);
3237 if (IS_ERR(old_rule)) {
3238 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION_ID]);
3239 return PTR_ERR(old_rule);
3240 }
3241 }
3242 }
3243
3244 nft_ctx_init(&ctx, net, skb, nlh, family, table, chain, nla);
3245
3246 n = 0;
3247 size = 0;
3248 if (nla[NFTA_RULE_EXPRESSIONS]) {
3249 info = kvmalloc_array(NFT_RULE_MAXEXPRS,
3250 sizeof(struct nft_expr_info),
3251 GFP_KERNEL);
3252 if (!info)
3253 return -ENOMEM;
3254
3255 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
3256 err = -EINVAL;
3257 if (nla_type(tmp) != NFTA_LIST_ELEM)
3258 goto err1;
3259 if (n == NFT_RULE_MAXEXPRS)
3260 goto err1;
3261 err = nf_tables_expr_parse(&ctx, tmp, &info[n]);
3262 if (err < 0)
3263 goto err1;
3264 size += info[n].ops->size;
3265 n++;
3266 }
3267 }
3268 /* Check for overflow of dlen field */
3269 err = -EFBIG;
3270 if (size >= 1 << 12)
3271 goto err1;
3272
3273 if (nla[NFTA_RULE_USERDATA]) {
3274 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
3275 if (ulen > 0)
3276 usize = sizeof(struct nft_userdata) + ulen;
3277 }
3278
3279 err = -ENOMEM;
3280 rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL);
3281 if (rule == NULL)
3282 goto err1;
3283
3284 nft_activate_next(net, rule);
3285
3286 rule->handle = handle;
3287 rule->dlen = size;
3288 rule->udata = ulen ? 1 : 0;
3289
3290 if (ulen) {
3291 udata = nft_userdata(rule);
3292 udata->len = ulen - 1;
3293 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
3294 }
3295
3296 expr = nft_expr_first(rule);
3297 for (i = 0; i < n; i++) {
3298 err = nf_tables_newexpr(&ctx, &info[i], expr);
3299 if (err < 0) {
3300 NL_SET_BAD_ATTR(extack, info[i].attr);
3301 goto err2;
3302 }
3303
3304 if (info[i].ops->validate)
3305 nft_validate_state_update(net, NFT_VALIDATE_NEED);
3306
3307 info[i].ops = NULL;
3308 expr = nft_expr_next(expr);
3309 }
3310
3311 if (nlh->nlmsg_flags & NLM_F_REPLACE) {
3312 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3313 if (trans == NULL) {
3314 err = -ENOMEM;
3315 goto err2;
3316 }
3317 err = nft_delrule(&ctx, old_rule);
3318 if (err < 0) {
3319 nft_trans_destroy(trans);
3320 goto err2;
3321 }
3322
3323 list_add_tail_rcu(&rule->list, &old_rule->list);
3324 } else {
3325 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3326 if (!trans) {
3327 err = -ENOMEM;
3328 goto err2;
3329 }
3330
3331 if (nlh->nlmsg_flags & NLM_F_APPEND) {
3332 if (old_rule)
3333 list_add_rcu(&rule->list, &old_rule->list);
3334 else
3335 list_add_tail_rcu(&rule->list, &chain->rules);
3336 } else {
3337 if (old_rule)
3338 list_add_tail_rcu(&rule->list, &old_rule->list);
3339 else
3340 list_add_rcu(&rule->list, &chain->rules);
3341 }
3342 }
3343 kvfree(info);
3344 chain->use++;
3345
3346 if (net->nft.validate_state == NFT_VALIDATE_DO)
3347 return nft_table_validate(net, table);
3348
3349 if (chain->flags & NFT_CHAIN_HW_OFFLOAD) {
3350 flow = nft_flow_rule_create(net, rule);
3351 if (IS_ERR(flow))
3352 return PTR_ERR(flow);
3353
3354 nft_trans_flow_rule(trans) = flow;
3355 }
3356
3357 return 0;
3358 err2:
3359 nf_tables_rule_release(&ctx, rule);
3360 err1:
3361 for (i = 0; i < n; i++) {
3362 if (info[i].ops) {
3363 module_put(info[i].ops->type->owner);
3364 if (info[i].ops->type->release_ops)
3365 info[i].ops->type->release_ops(info[i].ops);
3366 }
3367 }
3368 kvfree(info);
3369 return err;
3370 }
3371
3372 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3373 const struct nlattr *nla)
3374 {
3375 u32 id = ntohl(nla_get_be32(nla));
3376 struct nft_trans *trans;
3377
3378 list_for_each_entry(trans, &net->nft.commit_list, list) {
3379 struct nft_rule *rule = nft_trans_rule(trans);
3380
3381 if (trans->msg_type == NFT_MSG_NEWRULE &&
3382 id == nft_trans_rule_id(trans))
3383 return rule;
3384 }
3385 return ERR_PTR(-ENOENT);
3386 }
3387
3388 static int nf_tables_delrule(struct net *net, struct sock *nlsk,
3389 struct sk_buff *skb, const struct nlmsghdr *nlh,
3390 const struct nlattr * const nla[],
3391 struct netlink_ext_ack *extack)
3392 {
3393 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3394 u8 genmask = nft_genmask_next(net);
3395 struct nft_table *table;
3396 struct nft_chain *chain = NULL;
3397 struct nft_rule *rule;
3398 int family = nfmsg->nfgen_family, err = 0;
3399 struct nft_ctx ctx;
3400
3401 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask);
3402 if (IS_ERR(table)) {
3403 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3404 return PTR_ERR(table);
3405 }
3406
3407 if (nla[NFTA_RULE_CHAIN]) {
3408 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3409 genmask);
3410 if (IS_ERR(chain)) {
3411 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3412 return PTR_ERR(chain);
3413 }
3414 if (nft_chain_is_bound(chain))
3415 return -EOPNOTSUPP;
3416 }
3417
3418 nft_ctx_init(&ctx, net, skb, nlh, family, table, chain, nla);
3419
3420 if (chain) {
3421 if (nla[NFTA_RULE_HANDLE]) {
3422 rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3423 if (IS_ERR(rule)) {
3424 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3425 return PTR_ERR(rule);
3426 }
3427
3428 err = nft_delrule(&ctx, rule);
3429 } else if (nla[NFTA_RULE_ID]) {
3430 rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_ID]);
3431 if (IS_ERR(rule)) {
3432 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
3433 return PTR_ERR(rule);
3434 }
3435
3436 err = nft_delrule(&ctx, rule);
3437 } else {
3438 err = nft_delrule_by_chain(&ctx);
3439 }
3440 } else {
3441 list_for_each_entry(chain, &table->chains, list) {
3442 if (!nft_is_active_next(net, chain))
3443 continue;
3444
3445 ctx.chain = chain;
3446 err = nft_delrule_by_chain(&ctx);
3447 if (err < 0)
3448 break;
3449 }
3450 }
3451
3452 return err;
3453 }
3454
3455 /*
3456 * Sets
3457 */
3458 static const struct nft_set_type *nft_set_types[] = {
3459 &nft_set_hash_fast_type,
3460 &nft_set_hash_type,
3461 &nft_set_rhash_type,
3462 &nft_set_bitmap_type,
3463 &nft_set_rbtree_type,
3464 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
3465 &nft_set_pipapo_avx2_type,
3466 #endif
3467 &nft_set_pipapo_type,
3468 };
3469
3470 #define NFT_SET_FEATURES (NFT_SET_INTERVAL | NFT_SET_MAP | \
3471 NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
3472 NFT_SET_EVAL)
3473
3474 static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
3475 {
3476 return (flags & type->features) == (flags & NFT_SET_FEATURES);
3477 }
3478
3479 /*
3480 * Select a set implementation based on the data characteristics and the
3481 * given policy. The total memory use might not be known if no size is
3482 * given, in that case the amount of memory per element is used.
3483 */
3484 static const struct nft_set_ops *
3485 nft_select_set_ops(const struct nft_ctx *ctx,
3486 const struct nlattr * const nla[],
3487 const struct nft_set_desc *desc,
3488 enum nft_set_policies policy)
3489 {
3490 const struct nft_set_ops *ops, *bops;
3491 struct nft_set_estimate est, best;
3492 const struct nft_set_type *type;
3493 u32 flags = 0;
3494 int i;
3495
3496 lockdep_assert_held(&ctx->net->nft.commit_mutex);
3497 lockdep_nfnl_nft_mutex_not_held();
3498
3499 if (nla[NFTA_SET_FLAGS] != NULL)
3500 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
3501
3502 bops = NULL;
3503 best.size = ~0;
3504 best.lookup = ~0;
3505 best.space = ~0;
3506
3507 for (i = 0; i < ARRAY_SIZE(nft_set_types); i++) {
3508 type = nft_set_types[i];
3509 ops = &type->ops;
3510
3511 if (!nft_set_ops_candidate(type, flags))
3512 continue;
3513 if (!ops->estimate(desc, flags, &est))
3514 continue;
3515
3516 switch (policy) {
3517 case NFT_SET_POL_PERFORMANCE:
3518 if (est.lookup < best.lookup)
3519 break;
3520 if (est.lookup == best.lookup &&
3521 est.space < best.space)
3522 break;
3523 continue;
3524 case NFT_SET_POL_MEMORY:
3525 if (!desc->size) {
3526 if (est.space < best.space)
3527 break;
3528 if (est.space == best.space &&
3529 est.lookup < best.lookup)
3530 break;
3531 } else if (est.size < best.size || !bops) {
3532 break;
3533 }
3534 continue;
3535 default:
3536 break;
3537 }
3538
3539 bops = ops;
3540 best = est;
3541 }
3542
3543 if (bops != NULL)
3544 return bops;
3545
3546 return ERR_PTR(-EOPNOTSUPP);
3547 }
3548
3549 static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
3550 [NFTA_SET_TABLE] = { .type = NLA_STRING,
3551 .len = NFT_TABLE_MAXNAMELEN - 1 },
3552 [NFTA_SET_NAME] = { .type = NLA_STRING,
3553 .len = NFT_SET_MAXNAMELEN - 1 },
3554 [NFTA_SET_FLAGS] = { .type = NLA_U32 },
3555 [NFTA_SET_KEY_TYPE] = { .type = NLA_U32 },
3556 [NFTA_SET_KEY_LEN] = { .type = NLA_U32 },
3557 [NFTA_SET_DATA_TYPE] = { .type = NLA_U32 },
3558 [NFTA_SET_DATA_LEN] = { .type = NLA_U32 },
3559 [NFTA_SET_POLICY] = { .type = NLA_U32 },
3560 [NFTA_SET_DESC] = { .type = NLA_NESTED },
3561 [NFTA_SET_ID] = { .type = NLA_U32 },
3562 [NFTA_SET_TIMEOUT] = { .type = NLA_U64 },
3563 [NFTA_SET_GC_INTERVAL] = { .type = NLA_U32 },
3564 [NFTA_SET_USERDATA] = { .type = NLA_BINARY,
3565 .len = NFT_USERDATA_MAXLEN },
3566 [NFTA_SET_OBJ_TYPE] = { .type = NLA_U32 },
3567 [NFTA_SET_HANDLE] = { .type = NLA_U64 },
3568 [NFTA_SET_EXPR] = { .type = NLA_NESTED },
3569 };
3570
3571 static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
3572 [NFTA_SET_DESC_SIZE] = { .type = NLA_U32 },
3573 [NFTA_SET_DESC_CONCAT] = { .type = NLA_NESTED },
3574 };
3575
3576 static int nft_ctx_init_from_setattr(struct nft_ctx *ctx, struct net *net,
3577 const struct sk_buff *skb,
3578 const struct nlmsghdr *nlh,
3579 const struct nlattr * const nla[],
3580 struct netlink_ext_ack *extack,
3581 u8 genmask)
3582 {
3583 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3584 int family = nfmsg->nfgen_family;
3585 struct nft_table *table = NULL;
3586
3587 if (nla[NFTA_SET_TABLE] != NULL) {
3588 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
3589 genmask);
3590 if (IS_ERR(table)) {
3591 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
3592 return PTR_ERR(table);
3593 }
3594 }
3595
3596 nft_ctx_init(ctx, net, skb, nlh, family, table, NULL, nla);
3597 return 0;
3598 }
3599
3600 static struct nft_set *nft_set_lookup(const struct nft_table *table,
3601 const struct nlattr *nla, u8 genmask)
3602 {
3603 struct nft_set *set;
3604
3605 if (nla == NULL)
3606 return ERR_PTR(-EINVAL);
3607
3608 list_for_each_entry_rcu(set, &table->sets, list) {
3609 if (!nla_strcmp(nla, set->name) &&
3610 nft_active_genmask(set, genmask))
3611 return set;
3612 }
3613 return ERR_PTR(-ENOENT);
3614 }
3615
3616 static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
3617 const struct nlattr *nla,
3618 u8 genmask)
3619 {
3620 struct nft_set *set;
3621
3622 list_for_each_entry(set, &table->sets, list) {
3623 if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
3624 nft_active_genmask(set, genmask))
3625 return set;
3626 }
3627 return ERR_PTR(-ENOENT);
3628 }
3629
3630 static struct nft_set *nft_set_lookup_byid(const struct net *net,
3631 const struct nlattr *nla, u8 genmask)
3632 {
3633 struct nft_trans *trans;
3634 u32 id = ntohl(nla_get_be32(nla));
3635
3636 list_for_each_entry(trans, &net->nft.commit_list, list) {
3637 if (trans->msg_type == NFT_MSG_NEWSET) {
3638 struct nft_set *set = nft_trans_set(trans);
3639
3640 if (id == nft_trans_set_id(trans) &&
3641 nft_active_genmask(set, genmask))
3642 return set;
3643 }
3644 }
3645 return ERR_PTR(-ENOENT);
3646 }
3647
3648 struct nft_set *nft_set_lookup_global(const struct net *net,
3649 const struct nft_table *table,
3650 const struct nlattr *nla_set_name,
3651 const struct nlattr *nla_set_id,
3652 u8 genmask)
3653 {
3654 struct nft_set *set;
3655
3656 set = nft_set_lookup(table, nla_set_name, genmask);
3657 if (IS_ERR(set)) {
3658 if (!nla_set_id)
3659 return set;
3660
3661 set = nft_set_lookup_byid(net, nla_set_id, genmask);
3662 }
3663 return set;
3664 }
3665 EXPORT_SYMBOL_GPL(nft_set_lookup_global);
3666
3667 static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
3668 const char *name)
3669 {
3670 const struct nft_set *i;
3671 const char *p;
3672 unsigned long *inuse;
3673 unsigned int n = 0, min = 0;
3674
3675 p = strchr(name, '%');
3676 if (p != NULL) {
3677 if (p[1] != 'd' || strchr(p + 2, '%'))
3678 return -EINVAL;
3679
3680 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
3681 if (inuse == NULL)
3682 return -ENOMEM;
3683 cont:
3684 list_for_each_entry(i, &ctx->table->sets, list) {
3685 int tmp;
3686
3687 if (!nft_is_active_next(ctx->net, set))
3688 continue;
3689 if (!sscanf(i->name, name, &tmp))
3690 continue;
3691 if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
3692 continue;
3693
3694 set_bit(tmp - min, inuse);
3695 }
3696
3697 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
3698 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
3699 min += BITS_PER_BYTE * PAGE_SIZE;
3700 memset(inuse, 0, PAGE_SIZE);
3701 goto cont;
3702 }
3703 free_page((unsigned long)inuse);
3704 }
3705
3706 set->name = kasprintf(GFP_KERNEL, name, min + n);
3707 if (!set->name)
3708 return -ENOMEM;
3709
3710 list_for_each_entry(i, &ctx->table->sets, list) {
3711 if (!nft_is_active_next(ctx->net, i))
3712 continue;
3713 if (!strcmp(set->name, i->name)) {
3714 kfree(set->name);
3715 set->name = NULL;
3716 return -ENFILE;
3717 }
3718 }
3719 return 0;
3720 }
3721
3722 static int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
3723 {
3724 u64 ms = be64_to_cpu(nla_get_be64(nla));
3725 u64 max = (u64)(~((u64)0));
3726
3727 max = div_u64(max, NSEC_PER_MSEC);
3728 if (ms >= max)
3729 return -ERANGE;
3730
3731 ms *= NSEC_PER_MSEC;
3732 *result = nsecs_to_jiffies64(ms);
3733 return 0;
3734 }
3735
3736 static __be64 nf_jiffies64_to_msecs(u64 input)
3737 {
3738 return cpu_to_be64(jiffies64_to_msecs(input));
3739 }
3740
3741 static int nf_tables_fill_set_concat(struct sk_buff *skb,
3742 const struct nft_set *set)
3743 {
3744 struct nlattr *concat, *field;
3745 int i;
3746
3747 concat = nla_nest_start_noflag(skb, NFTA_SET_DESC_CONCAT);
3748 if (!concat)
3749 return -ENOMEM;
3750
3751 for (i = 0; i < set->field_count; i++) {
3752 field = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
3753 if (!field)
3754 return -ENOMEM;
3755
3756 if (nla_put_be32(skb, NFTA_SET_FIELD_LEN,
3757 htonl(set->field_len[i])))
3758 return -ENOMEM;
3759
3760 nla_nest_end(skb, field);
3761 }
3762
3763 nla_nest_end(skb, concat);
3764
3765 return 0;
3766 }
3767
3768 static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
3769 const struct nft_set *set, u16 event, u16 flags)
3770 {
3771 struct nfgenmsg *nfmsg;
3772 struct nlmsghdr *nlh;
3773 u32 portid = ctx->portid;
3774 struct nlattr *nest;
3775 u32 seq = ctx->seq;
3776
3777 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3778 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
3779 flags);
3780 if (nlh == NULL)
3781 goto nla_put_failure;
3782
3783 nfmsg = nlmsg_data(nlh);
3784 nfmsg->nfgen_family = ctx->family;
3785 nfmsg->version = NFNETLINK_V0;
3786 nfmsg->res_id = htons(ctx->net->nft.base_seq & 0xffff);
3787
3788 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
3789 goto nla_put_failure;
3790 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
3791 goto nla_put_failure;
3792 if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
3793 NFTA_SET_PAD))
3794 goto nla_put_failure;
3795 if (set->flags != 0)
3796 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
3797 goto nla_put_failure;
3798
3799 if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
3800 goto nla_put_failure;
3801 if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
3802 goto nla_put_failure;
3803 if (set->flags & NFT_SET_MAP) {
3804 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
3805 goto nla_put_failure;
3806 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
3807 goto nla_put_failure;
3808 }
3809 if (set->flags & NFT_SET_OBJECT &&
3810 nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
3811 goto nla_put_failure;
3812
3813 if (set->timeout &&
3814 nla_put_be64(skb, NFTA_SET_TIMEOUT,
3815 nf_jiffies64_to_msecs(set->timeout),
3816 NFTA_SET_PAD))
3817 goto nla_put_failure;
3818 if (set->gc_int &&
3819 nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(set->gc_int)))
3820 goto nla_put_failure;
3821
3822 if (set->policy != NFT_SET_POL_PERFORMANCE) {
3823 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
3824 goto nla_put_failure;
3825 }
3826
3827 if (set->udata &&
3828 nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
3829 goto nla_put_failure;
3830
3831 nest = nla_nest_start_noflag(skb, NFTA_SET_DESC);
3832 if (!nest)
3833 goto nla_put_failure;
3834 if (set->size &&
3835 nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
3836 goto nla_put_failure;
3837
3838 if (set->field_count > 1 &&
3839 nf_tables_fill_set_concat(skb, set))
3840 goto nla_put_failure;
3841
3842 nla_nest_end(skb, nest);
3843
3844 if (set->expr) {
3845 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPR);
3846 if (nf_tables_fill_expr_info(skb, set->expr) < 0)
3847 goto nla_put_failure;
3848
3849 nla_nest_end(skb, nest);
3850 }
3851
3852 nlmsg_end(skb, nlh);
3853 return 0;
3854
3855 nla_put_failure:
3856 nlmsg_trim(skb, nlh);
3857 return -1;
3858 }
3859
3860 static void nf_tables_set_notify(const struct nft_ctx *ctx,
3861 const struct nft_set *set, int event,
3862 gfp_t gfp_flags)
3863 {
3864 struct sk_buff *skb;
3865 u32 portid = ctx->portid;
3866 int err;
3867 char *buf = kasprintf(gfp_flags, "%s:%llu;%s:%llu",
3868 ctx->table->name, ctx->table->handle,
3869 set->name, set->handle);
3870
3871 audit_log_nfcfg(buf,
3872 ctx->family,
3873 set->field_count,
3874 event == NFT_MSG_NEWSET ?
3875 AUDIT_NFT_OP_SET_REGISTER :
3876 AUDIT_NFT_OP_SET_UNREGISTER,
3877 gfp_flags);
3878 kfree(buf);
3879
3880 if (!ctx->report &&
3881 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3882 return;
3883
3884 skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
3885 if (skb == NULL)
3886 goto err;
3887
3888 err = nf_tables_fill_set(skb, ctx, set, event, 0);
3889 if (err < 0) {
3890 kfree_skb(skb);
3891 goto err;
3892 }
3893
3894 nft_notify_enqueue(skb, ctx->report, &ctx->net->nft.notify_list);
3895 return;
3896 err:
3897 nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
3898 }
3899
3900 static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
3901 {
3902 const struct nft_set *set;
3903 unsigned int idx, s_idx = cb->args[0];
3904 struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
3905 struct net *net = sock_net(skb->sk);
3906 struct nft_ctx *ctx = cb->data, ctx_set;
3907
3908 if (cb->args[1])
3909 return skb->len;
3910
3911 rcu_read_lock();
3912 cb->seq = net->nft.base_seq;
3913
3914 list_for_each_entry_rcu(table, &net->nft.tables, list) {
3915 if (ctx->family != NFPROTO_UNSPEC &&
3916 ctx->family != table->family)
3917 continue;
3918
3919 if (ctx->table && ctx->table != table)
3920 continue;
3921
3922 if (cur_table) {
3923 if (cur_table != table)
3924 continue;
3925
3926 cur_table = NULL;
3927 }
3928 idx = 0;
3929 list_for_each_entry_rcu(set, &table->sets, list) {
3930 if (idx < s_idx)
3931 goto cont;
3932 if (!nft_is_active(net, set))
3933 goto cont;
3934
3935 ctx_set = *ctx;
3936 ctx_set.table = table;
3937 ctx_set.family = table->family;
3938
3939 if (nf_tables_fill_set(skb, &ctx_set, set,
3940 NFT_MSG_NEWSET,
3941 NLM_F_MULTI) < 0) {
3942 cb->args[0] = idx;
3943 cb->args[2] = (unsigned long) table;
3944 goto done;
3945 }
3946 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
3947 cont:
3948 idx++;
3949 }
3950 if (s_idx)
3951 s_idx = 0;
3952 }
3953 cb->args[1] = 1;
3954 done:
3955 rcu_read_unlock();
3956 return skb->len;
3957 }
3958
3959 static int nf_tables_dump_sets_start(struct netlink_callback *cb)
3960 {
3961 struct nft_ctx *ctx_dump = NULL;
3962
3963 ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
3964 if (ctx_dump == NULL)
3965 return -ENOMEM;
3966
3967 cb->data = ctx_dump;
3968 return 0;
3969 }
3970
3971 static int nf_tables_dump_sets_done(struct netlink_callback *cb)
3972 {
3973 kfree(cb->data);
3974 return 0;
3975 }
3976
3977 /* called with rcu_read_lock held */
3978 static int nf_tables_getset(struct net *net, struct sock *nlsk,
3979 struct sk_buff *skb, const struct nlmsghdr *nlh,
3980 const struct nlattr * const nla[],
3981 struct netlink_ext_ack *extack)
3982 {
3983 u8 genmask = nft_genmask_cur(net);
3984 const struct nft_set *set;
3985 struct nft_ctx ctx;
3986 struct sk_buff *skb2;
3987 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3988 int err;
3989
3990 /* Verify existence before starting dump */
3991 err = nft_ctx_init_from_setattr(&ctx, net, skb, nlh, nla, extack,
3992 genmask);
3993 if (err < 0)
3994 return err;
3995
3996 if (nlh->nlmsg_flags & NLM_F_DUMP) {
3997 struct netlink_dump_control c = {
3998 .start = nf_tables_dump_sets_start,
3999 .dump = nf_tables_dump_sets,
4000 .done = nf_tables_dump_sets_done,
4001 .data = &ctx,
4002 .module = THIS_MODULE,
4003 };
4004
4005 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
4006 }
4007
4008 /* Only accept unspec with dump */
4009 if (nfmsg->nfgen_family == NFPROTO_UNSPEC)
4010 return -EAFNOSUPPORT;
4011 if (!nla[NFTA_SET_TABLE])
4012 return -EINVAL;
4013
4014 set = nft_set_lookup(ctx.table, nla[NFTA_SET_NAME], genmask);
4015 if (IS_ERR(set))
4016 return PTR_ERR(set);
4017
4018 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
4019 if (skb2 == NULL)
4020 return -ENOMEM;
4021
4022 err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
4023 if (err < 0)
4024 goto err_fill_set_info;
4025
4026 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
4027
4028 err_fill_set_info:
4029 kfree_skb(skb2);
4030 return err;
4031 }
4032
4033 static const struct nla_policy nft_concat_policy[NFTA_SET_FIELD_MAX + 1] = {
4034 [NFTA_SET_FIELD_LEN] = { .type = NLA_U32 },
4035 };
4036
4037 static int nft_set_desc_concat_parse(const struct nlattr *attr,
4038 struct nft_set_desc *desc)
4039 {
4040 struct nlattr *tb[NFTA_SET_FIELD_MAX + 1];
4041 u32 len;
4042 int err;
4043
4044 err = nla_parse_nested_deprecated(tb, NFTA_SET_FIELD_MAX, attr,
4045 nft_concat_policy, NULL);
4046 if (err < 0)
4047 return err;
4048
4049 if (!tb[NFTA_SET_FIELD_LEN])
4050 return -EINVAL;
4051
4052 len = ntohl(nla_get_be32(tb[NFTA_SET_FIELD_LEN]));
4053
4054 if (len * BITS_PER_BYTE / 32 > NFT_REG32_COUNT)
4055 return -E2BIG;
4056
4057 desc->field_len[desc->field_count++] = len;
4058
4059 return 0;
4060 }
4061
4062 static int nft_set_desc_concat(struct nft_set_desc *desc,
4063 const struct nlattr *nla)
4064 {
4065 struct nlattr *attr;
4066 int rem, err;
4067
4068 nla_for_each_nested(attr, nla, rem) {
4069 if (nla_type(attr) != NFTA_LIST_ELEM)
4070 return -EINVAL;
4071
4072 err = nft_set_desc_concat_parse(attr, desc);
4073 if (err < 0)
4074 return err;
4075 }
4076
4077 return 0;
4078 }
4079
4080 static int nf_tables_set_desc_parse(struct nft_set_desc *desc,
4081 const struct nlattr *nla)
4082 {
4083 struct nlattr *da[NFTA_SET_DESC_MAX + 1];
4084 int err;
4085
4086 err = nla_parse_nested_deprecated(da, NFTA_SET_DESC_MAX, nla,
4087 nft_set_desc_policy, NULL);
4088 if (err < 0)
4089 return err;
4090
4091 if (da[NFTA_SET_DESC_SIZE] != NULL)
4092 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
4093 if (da[NFTA_SET_DESC_CONCAT])
4094 err = nft_set_desc_concat(desc, da[NFTA_SET_DESC_CONCAT]);
4095
4096 return err;
4097 }
4098
4099 static int nf_tables_newset(struct net *net, struct sock *nlsk,
4100 struct sk_buff *skb, const struct nlmsghdr *nlh,
4101 const struct nlattr * const nla[],
4102 struct netlink_ext_ack *extack)
4103 {
4104 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
4105 u8 genmask = nft_genmask_next(net);
4106 int family = nfmsg->nfgen_family;
4107 const struct nft_set_ops *ops;
4108 struct nft_expr *expr = NULL;
4109 struct nft_table *table;
4110 struct nft_set *set;
4111 struct nft_ctx ctx;
4112 char *name;
4113 u64 size;
4114 u64 timeout;
4115 u32 ktype, dtype, flags, policy, gc_int, objtype;
4116 struct nft_set_desc desc;
4117 unsigned char *udata;
4118 u16 udlen;
4119 int err;
4120 int i;
4121
4122 if (nla[NFTA_SET_TABLE] == NULL ||
4123 nla[NFTA_SET_NAME] == NULL ||
4124 nla[NFTA_SET_KEY_LEN] == NULL ||
4125 nla[NFTA_SET_ID] == NULL)
4126 return -EINVAL;
4127
4128 memset(&desc, 0, sizeof(desc));
4129
4130 ktype = NFT_DATA_VALUE;
4131 if (nla[NFTA_SET_KEY_TYPE] != NULL) {
4132 ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
4133 if ((ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
4134 return -EINVAL;
4135 }
4136
4137 desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
4138 if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
4139 return -EINVAL;
4140
4141 flags = 0;
4142 if (nla[NFTA_SET_FLAGS] != NULL) {
4143 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
4144 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
4145 NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
4146 NFT_SET_MAP | NFT_SET_EVAL |
4147 NFT_SET_OBJECT | NFT_SET_CONCAT))
4148 return -EOPNOTSUPP;
4149 /* Only one of these operations is supported */
4150 if ((flags & (NFT_SET_MAP | NFT_SET_OBJECT)) ==
4151 (NFT_SET_MAP | NFT_SET_OBJECT))
4152 return -EOPNOTSUPP;
4153 if ((flags & (NFT_SET_EVAL | NFT_SET_OBJECT)) ==
4154 (NFT_SET_EVAL | NFT_SET_OBJECT))
4155 return -EOPNOTSUPP;
4156 }
4157
4158 dtype = 0;
4159 if (nla[NFTA_SET_DATA_TYPE] != NULL) {
4160 if (!(flags & NFT_SET_MAP))
4161 return -EINVAL;
4162
4163 dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
4164 if ((dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
4165 dtype != NFT_DATA_VERDICT)
4166 return -EINVAL;
4167
4168 if (dtype != NFT_DATA_VERDICT) {
4169 if (nla[NFTA_SET_DATA_LEN] == NULL)
4170 return -EINVAL;
4171 desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
4172 if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
4173 return -EINVAL;
4174 } else
4175 desc.dlen = sizeof(struct nft_verdict);
4176 } else if (flags & NFT_SET_MAP)
4177 return -EINVAL;
4178
4179 if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
4180 if (!(flags & NFT_SET_OBJECT))
4181 return -EINVAL;
4182
4183 objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
4184 if (objtype == NFT_OBJECT_UNSPEC ||
4185 objtype > NFT_OBJECT_MAX)
4186 return -EOPNOTSUPP;
4187 } else if (flags & NFT_SET_OBJECT)
4188 return -EINVAL;
4189 else
4190 objtype = NFT_OBJECT_UNSPEC;
4191
4192 timeout = 0;
4193 if (nla[NFTA_SET_TIMEOUT] != NULL) {
4194 if (!(flags & NFT_SET_TIMEOUT))
4195 return -EINVAL;
4196
4197 err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &timeout);
4198 if (err)
4199 return err;
4200 }
4201 gc_int = 0;
4202 if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
4203 if (!(flags & NFT_SET_TIMEOUT))
4204 return -EINVAL;
4205 gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
4206 }
4207
4208 policy = NFT_SET_POL_PERFORMANCE;
4209 if (nla[NFTA_SET_POLICY] != NULL)
4210 policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
4211
4212 if (nla[NFTA_SET_DESC] != NULL) {
4213 err = nf_tables_set_desc_parse(&desc, nla[NFTA_SET_DESC]);
4214 if (err < 0)
4215 return err;
4216 }
4217
4218 if (nla[NFTA_SET_EXPR])
4219 desc.expr = true;
4220
4221 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask);
4222 if (IS_ERR(table)) {
4223 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4224 return PTR_ERR(table);
4225 }
4226
4227 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
4228
4229 set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4230 if (IS_ERR(set)) {
4231 if (PTR_ERR(set) != -ENOENT) {
4232 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4233 return PTR_ERR(set);
4234 }
4235 } else {
4236 if (nlh->nlmsg_flags & NLM_F_EXCL) {
4237 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4238 return -EEXIST;
4239 }
4240 if (nlh->nlmsg_flags & NLM_F_REPLACE)
4241 return -EOPNOTSUPP;
4242
4243 return 0;
4244 }
4245
4246 if (!(nlh->nlmsg_flags & NLM_F_CREATE))
4247 return -ENOENT;
4248
4249 ops = nft_select_set_ops(&ctx, nla, &desc, policy);
4250 if (IS_ERR(ops))
4251 return PTR_ERR(ops);
4252
4253 udlen = 0;
4254 if (nla[NFTA_SET_USERDATA])
4255 udlen = nla_len(nla[NFTA_SET_USERDATA]);
4256
4257 size = 0;
4258 if (ops->privsize != NULL)
4259 size = ops->privsize(nla, &desc);
4260
4261 set = kvzalloc(sizeof(*set) + size + udlen, GFP_KERNEL);
4262 if (!set)
4263 return -ENOMEM;
4264
4265 name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL);
4266 if (!name) {
4267 err = -ENOMEM;
4268 goto err_set_name;
4269 }
4270
4271 err = nf_tables_set_alloc_name(&ctx, set, name);
4272 kfree(name);
4273 if (err < 0)
4274 goto err_set_alloc_name;
4275
4276 if (nla[NFTA_SET_EXPR]) {
4277 expr = nft_set_elem_expr_alloc(&ctx, set, nla[NFTA_SET_EXPR]);
4278 if (IS_ERR(expr)) {
4279 err = PTR_ERR(expr);
4280 goto err_set_alloc_name;
4281 }
4282 }
4283
4284 udata = NULL;
4285 if (udlen) {
4286 udata = set->data + size;
4287 nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
4288 }
4289
4290 INIT_LIST_HEAD(&set->bindings);
4291 set->table = table;
4292 write_pnet(&set->net, net);
4293 set->ops = ops;
4294 set->ktype = ktype;
4295 set->klen = desc.klen;
4296 set->dtype = dtype;
4297 set->objtype = objtype;
4298 set->dlen = desc.dlen;
4299 set->expr = expr;
4300 set->flags = flags;
4301 set->size = desc.size;
4302 set->policy = policy;
4303 set->udlen = udlen;
4304 set->udata = udata;
4305 set->timeout = timeout;
4306 set->gc_int = gc_int;
4307 set->handle = nf_tables_alloc_handle(table);
4308
4309 set->field_count = desc.field_count;
4310 for (i = 0; i < desc.field_count; i++)
4311 set->field_len[i] = desc.field_len[i];
4312
4313 err = ops->init(set, &desc, nla);
4314 if (err < 0)
4315 goto err_set_init;
4316
4317 err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
4318 if (err < 0)
4319 goto err_set_trans;
4320
4321 list_add_tail_rcu(&set->list, &table->sets);
4322 table->use++;
4323 return 0;
4324
4325 err_set_trans:
4326 ops->destroy(set);
4327 err_set_init:
4328 if (expr)
4329 nft_expr_destroy(&ctx, expr);
4330 err_set_alloc_name:
4331 kfree(set->name);
4332 err_set_name:
4333 kvfree(set);
4334 return err;
4335 }
4336
4337 static void nft_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
4338 {
4339 if (WARN_ON(set->use > 0))
4340 return;
4341
4342 if (set->expr)
4343 nft_expr_destroy(ctx, set->expr);
4344
4345 set->ops->destroy(set);
4346 kfree(set->name);
4347 kvfree(set);
4348 }
4349
4350 static int nf_tables_delset(struct net *net, struct sock *nlsk,
4351 struct sk_buff *skb, const struct nlmsghdr *nlh,
4352 const struct nlattr * const nla[],
4353 struct netlink_ext_ack *extack)
4354 {
4355 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
4356 u8 genmask = nft_genmask_next(net);
4357 const struct nlattr *attr;
4358 struct nft_set *set;
4359 struct nft_ctx ctx;
4360 int err;
4361
4362 if (nfmsg->nfgen_family == NFPROTO_UNSPEC)
4363 return -EAFNOSUPPORT;
4364 if (nla[NFTA_SET_TABLE] == NULL)
4365 return -EINVAL;
4366
4367 err = nft_ctx_init_from_setattr(&ctx, net, skb, nlh, nla, extack,
4368 genmask);
4369 if (err < 0)
4370 return err;
4371
4372 if (nla[NFTA_SET_HANDLE]) {
4373 attr = nla[NFTA_SET_HANDLE];
4374 set = nft_set_lookup_byhandle(ctx.table, attr, genmask);
4375 } else {
4376 attr = nla[NFTA_SET_NAME];
4377 set = nft_set_lookup(ctx.table, attr, genmask);
4378 }
4379
4380 if (IS_ERR(set)) {
4381 NL_SET_BAD_ATTR(extack, attr);
4382 return PTR_ERR(set);
4383 }
4384 if (set->use ||
4385 (nlh->nlmsg_flags & NLM_F_NONREC && atomic_read(&set->nelems) > 0)) {
4386 NL_SET_BAD_ATTR(extack, attr);
4387 return -EBUSY;
4388 }
4389
4390 return nft_delset(&ctx, set);
4391 }
4392
4393 static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
4394 struct nft_set *set,
4395 const struct nft_set_iter *iter,
4396 struct nft_set_elem *elem)
4397 {
4398 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4399 enum nft_registers dreg;
4400
4401 dreg = nft_type_to_reg(set->dtype);
4402 return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
4403 set->dtype == NFT_DATA_VERDICT ?
4404 NFT_DATA_VERDICT : NFT_DATA_VALUE,
4405 set->dlen);
4406 }
4407
4408 int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
4409 struct nft_set_binding *binding)
4410 {
4411 struct nft_set_binding *i;
4412 struct nft_set_iter iter;
4413
4414 if (set->use == UINT_MAX)
4415 return -EOVERFLOW;
4416
4417 if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
4418 return -EBUSY;
4419
4420 if (binding->flags & NFT_SET_MAP) {
4421 /* If the set is already bound to the same chain all
4422 * jumps are already validated for that chain.
4423 */
4424 list_for_each_entry(i, &set->bindings, list) {
4425 if (i->flags & NFT_SET_MAP &&
4426 i->chain == binding->chain)
4427 goto bind;
4428 }
4429
4430 iter.genmask = nft_genmask_next(ctx->net);
4431 iter.skip = 0;
4432 iter.count = 0;
4433 iter.err = 0;
4434 iter.fn = nf_tables_bind_check_setelem;
4435
4436 set->ops->walk(ctx, set, &iter);
4437 if (iter.err < 0)
4438 return iter.err;
4439 }
4440 bind:
4441 binding->chain = ctx->chain;
4442 list_add_tail_rcu(&binding->list, &set->bindings);
4443 nft_set_trans_bind(ctx, set);
4444 set->use++;
4445
4446 return 0;
4447 }
4448 EXPORT_SYMBOL_GPL(nf_tables_bind_set);
4449
4450 static void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
4451 struct nft_set_binding *binding, bool event)
4452 {
4453 list_del_rcu(&binding->list);
4454
4455 if (list_empty(&set->bindings) && nft_set_is_anonymous(set)) {
4456 list_del_rcu(&set->list);
4457 if (event)
4458 nf_tables_set_notify(ctx, set, NFT_MSG_DELSET,
4459 GFP_KERNEL);
4460 }
4461 }
4462
4463 void nf_tables_deactivate_set(const struct nft_ctx *ctx, struct nft_set *set,
4464 struct nft_set_binding *binding,
4465 enum nft_trans_phase phase)
4466 {
4467 switch (phase) {
4468 case NFT_TRANS_PREPARE:
4469 set->use--;
4470 return;
4471 case NFT_TRANS_ABORT:
4472 case NFT_TRANS_RELEASE:
4473 set->use--;
4474 fallthrough;
4475 default:
4476 nf_tables_unbind_set(ctx, set, binding,
4477 phase == NFT_TRANS_COMMIT);
4478 }
4479 }
4480 EXPORT_SYMBOL_GPL(nf_tables_deactivate_set);
4481
4482 void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
4483 {
4484 if (list_empty(&set->bindings) && nft_set_is_anonymous(set))
4485 nft_set_destroy(ctx, set);
4486 }
4487 EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
4488
4489 const struct nft_set_ext_type nft_set_ext_types[] = {
4490 [NFT_SET_EXT_KEY] = {
4491 .align = __alignof__(u32),
4492 },
4493 [NFT_SET_EXT_DATA] = {
4494 .align = __alignof__(u32),
4495 },
4496 [NFT_SET_EXT_EXPR] = {
4497 .align = __alignof__(struct nft_expr),
4498 },
4499 [NFT_SET_EXT_OBJREF] = {
4500 .len = sizeof(struct nft_object *),
4501 .align = __alignof__(struct nft_object *),
4502 },
4503 [NFT_SET_EXT_FLAGS] = {
4504 .len = sizeof(u8),
4505 .align = __alignof__(u8),
4506 },
4507 [NFT_SET_EXT_TIMEOUT] = {
4508 .len = sizeof(u64),
4509 .align = __alignof__(u64),
4510 },
4511 [NFT_SET_EXT_EXPIRATION] = {
4512 .len = sizeof(u64),
4513 .align = __alignof__(u64),
4514 },
4515 [NFT_SET_EXT_USERDATA] = {
4516 .len = sizeof(struct nft_userdata),
4517 .align = __alignof__(struct nft_userdata),
4518 },
4519 [NFT_SET_EXT_KEY_END] = {
4520 .align = __alignof__(u32),
4521 },
4522 };
4523
4524 /*
4525 * Set elements
4526 */
4527
4528 static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
4529 [NFTA_SET_ELEM_KEY] = { .type = NLA_NESTED },
4530 [NFTA_SET_ELEM_DATA] = { .type = NLA_NESTED },
4531 [NFTA_SET_ELEM_FLAGS] = { .type = NLA_U32 },
4532 [NFTA_SET_ELEM_TIMEOUT] = { .type = NLA_U64 },
4533 [NFTA_SET_ELEM_EXPIRATION] = { .type = NLA_U64 },
4534 [NFTA_SET_ELEM_USERDATA] = { .type = NLA_BINARY,
4535 .len = NFT_USERDATA_MAXLEN },
4536 [NFTA_SET_ELEM_EXPR] = { .type = NLA_NESTED },
4537 [NFTA_SET_ELEM_OBJREF] = { .type = NLA_STRING,
4538 .len = NFT_OBJ_MAXNAMELEN - 1 },
4539 [NFTA_SET_ELEM_KEY_END] = { .type = NLA_NESTED },
4540 };
4541
4542 static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
4543 [NFTA_SET_ELEM_LIST_TABLE] = { .type = NLA_STRING,
4544 .len = NFT_TABLE_MAXNAMELEN - 1 },
4545 [NFTA_SET_ELEM_LIST_SET] = { .type = NLA_STRING,
4546 .len = NFT_SET_MAXNAMELEN - 1 },
4547 [NFTA_SET_ELEM_LIST_ELEMENTS] = { .type = NLA_NESTED },
4548 [NFTA_SET_ELEM_LIST_SET_ID] = { .type = NLA_U32 },
4549 };
4550
4551 static int nft_ctx_init_from_elemattr(struct nft_ctx *ctx, struct net *net,
4552 const struct sk_buff *skb,
4553 const struct nlmsghdr *nlh,
4554 const struct nlattr * const nla[],
4555 struct netlink_ext_ack *extack,
4556 u8 genmask)
4557 {
4558 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
4559 int family = nfmsg->nfgen_family;
4560 struct nft_table *table;
4561
4562 table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
4563 genmask);
4564 if (IS_ERR(table)) {
4565 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
4566 return PTR_ERR(table);
4567 }
4568
4569 nft_ctx_init(ctx, net, skb, nlh, family, table, NULL, nla);
4570 return 0;
4571 }
4572
4573 static int nf_tables_fill_setelem(struct sk_buff *skb,
4574 const struct nft_set *set,
4575 const struct nft_set_elem *elem)
4576 {
4577 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4578 unsigned char *b = skb_tail_pointer(skb);
4579 struct nlattr *nest;
4580
4581 nest = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4582 if (nest == NULL)
4583 goto nla_put_failure;
4584
4585 if (nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
4586 NFT_DATA_VALUE, set->klen) < 0)
4587 goto nla_put_failure;
4588
4589 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
4590 nft_data_dump(skb, NFTA_SET_ELEM_KEY_END, nft_set_ext_key_end(ext),
4591 NFT_DATA_VALUE, set->klen) < 0)
4592 goto nla_put_failure;
4593
4594 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
4595 nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
4596 set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
4597 set->dlen) < 0)
4598 goto nla_put_failure;
4599
4600 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPR) &&
4601 nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, nft_set_ext_expr(ext)) < 0)
4602 goto nla_put_failure;
4603
4604 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
4605 nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
4606 (*nft_set_ext_obj(ext))->key.name) < 0)
4607 goto nla_put_failure;
4608
4609 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
4610 nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
4611 htonl(*nft_set_ext_flags(ext))))
4612 goto nla_put_failure;
4613
4614 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
4615 nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
4616 nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
4617 NFTA_SET_ELEM_PAD))
4618 goto nla_put_failure;
4619
4620 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
4621 u64 expires, now = get_jiffies_64();
4622
4623 expires = *nft_set_ext_expiration(ext);
4624 if (time_before64(now, expires))
4625 expires -= now;
4626 else
4627 expires = 0;
4628
4629 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
4630 nf_jiffies64_to_msecs(expires),
4631 NFTA_SET_ELEM_PAD))
4632 goto nla_put_failure;
4633 }
4634
4635 if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
4636 struct nft_userdata *udata;
4637
4638 udata = nft_set_ext_userdata(ext);
4639 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
4640 udata->len + 1, udata->data))
4641 goto nla_put_failure;
4642 }
4643
4644 nla_nest_end(skb, nest);
4645 return 0;
4646
4647 nla_put_failure:
4648 nlmsg_trim(skb, b);
4649 return -EMSGSIZE;
4650 }
4651
4652 struct nft_set_dump_args {
4653 const struct netlink_callback *cb;
4654 struct nft_set_iter iter;
4655 struct sk_buff *skb;
4656 };
4657
4658 static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
4659 struct nft_set *set,
4660 const struct nft_set_iter *iter,
4661 struct nft_set_elem *elem)
4662 {
4663 struct nft_set_dump_args *args;
4664
4665 args = container_of(iter, struct nft_set_dump_args, iter);
4666 return nf_tables_fill_setelem(args->skb, set, elem);
4667 }
4668
4669 struct nft_set_dump_ctx {
4670 const struct nft_set *set;
4671 struct nft_ctx ctx;
4672 };
4673
4674 static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
4675 {
4676 struct nft_set_dump_ctx *dump_ctx = cb->data;
4677 struct net *net = sock_net(skb->sk);
4678 struct nft_table *table;
4679 struct nft_set *set;
4680 struct nft_set_dump_args args;
4681 bool set_found = false;
4682 struct nfgenmsg *nfmsg;
4683 struct nlmsghdr *nlh;
4684 struct nlattr *nest;
4685 u32 portid, seq;
4686 int event;
4687
4688 rcu_read_lock();
4689 list_for_each_entry_rcu(table, &net->nft.tables, list) {
4690 if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
4691 dump_ctx->ctx.family != table->family)
4692 continue;
4693
4694 if (table != dump_ctx->ctx.table)
4695 continue;
4696
4697 list_for_each_entry_rcu(set, &table->sets, list) {
4698 if (set == dump_ctx->set) {
4699 set_found = true;
4700 break;
4701 }
4702 }
4703 break;
4704 }
4705
4706 if (!set_found) {
4707 rcu_read_unlock();
4708 return -ENOENT;
4709 }
4710
4711 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
4712 portid = NETLINK_CB(cb->skb).portid;
4713 seq = cb->nlh->nlmsg_seq;
4714
4715 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
4716 NLM_F_MULTI);
4717 if (nlh == NULL)
4718 goto nla_put_failure;
4719
4720 nfmsg = nlmsg_data(nlh);
4721 nfmsg->nfgen_family = table->family;
4722 nfmsg->version = NFNETLINK_V0;
4723 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
4724
4725 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
4726 goto nla_put_failure;
4727 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
4728 goto nla_put_failure;
4729
4730 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
4731 if (nest == NULL)
4732 goto nla_put_failure;
4733
4734 args.cb = cb;
4735 args.skb = skb;
4736 args.iter.genmask = nft_genmask_cur(net);
4737 args.iter.skip = cb->args[0];
4738 args.iter.count = 0;
4739 args.iter.err = 0;
4740 args.iter.fn = nf_tables_dump_setelem;
4741 set->ops->walk(&dump_ctx->ctx, set, &args.iter);
4742 rcu_read_unlock();
4743
4744 nla_nest_end(skb, nest);
4745 nlmsg_end(skb, nlh);
4746
4747 if (args.iter.err && args.iter.err != -EMSGSIZE)
4748 return args.iter.err;
4749 if (args.iter.count == cb->args[0])
4750 return 0;
4751
4752 cb->args[0] = args.iter.count;
4753 return skb->len;
4754
4755 nla_put_failure:
4756 rcu_read_unlock();
4757 return -ENOSPC;
4758 }
4759
4760 static int nf_tables_dump_set_start(struct netlink_callback *cb)
4761 {
4762 struct nft_set_dump_ctx *dump_ctx = cb->data;
4763
4764 cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
4765
4766 return cb->data ? 0 : -ENOMEM;
4767 }
4768
4769 static int nf_tables_dump_set_done(struct netlink_callback *cb)
4770 {
4771 kfree(cb->data);
4772 return 0;
4773 }
4774
4775 static int nf_tables_fill_setelem_info(struct sk_buff *skb,
4776 const struct nft_ctx *ctx, u32 seq,
4777 u32 portid, int event, u16 flags,
4778 const struct nft_set *set,
4779 const struct nft_set_elem *elem)
4780 {
4781 struct nfgenmsg *nfmsg;
4782 struct nlmsghdr *nlh;
4783 struct nlattr *nest;
4784 int err;
4785
4786 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
4787 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg),
4788 flags);
4789 if (nlh == NULL)
4790 goto nla_put_failure;
4791
4792 nfmsg = nlmsg_data(nlh);
4793 nfmsg->nfgen_family = ctx->family;
4794 nfmsg->version = NFNETLINK_V0;
4795 nfmsg->res_id = htons(ctx->net->nft.base_seq & 0xffff);
4796
4797 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
4798 goto nla_put_failure;
4799 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
4800 goto nla_put_failure;
4801
4802 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
4803 if (nest == NULL)
4804 goto nla_put_failure;
4805
4806 err = nf_tables_fill_setelem(skb, set, elem);
4807 if (err < 0)
4808 goto nla_put_failure;
4809
4810 nla_nest_end(skb, nest);
4811
4812 nlmsg_end(skb, nlh);
4813 return 0;
4814
4815 nla_put_failure:
4816 nlmsg_trim(skb, nlh);
4817 return -1;
4818 }
4819
4820 static int nft_setelem_parse_flags(const struct nft_set *set,
4821 const struct nlattr *attr, u32 *flags)
4822 {
4823 if (attr == NULL)
4824 return 0;
4825
4826 *flags = ntohl(nla_get_be32(attr));
4827 if (*flags & ~NFT_SET_ELEM_INTERVAL_END)
4828 return -EINVAL;
4829 if (!(set->flags & NFT_SET_INTERVAL) &&
4830 *flags & NFT_SET_ELEM_INTERVAL_END)
4831 return -EINVAL;
4832
4833 return 0;
4834 }
4835
4836 static int nft_setelem_parse_key(struct nft_ctx *ctx, struct nft_set *set,
4837 struct nft_data *key, struct nlattr *attr)
4838 {
4839 struct nft_data_desc desc;
4840 int err;
4841
4842 err = nft_data_init(ctx, key, NFT_DATA_VALUE_MAXLEN, &desc, attr);
4843 if (err < 0)
4844 return err;
4845
4846 if (desc.type != NFT_DATA_VALUE || desc.len != set->klen) {
4847 nft_data_release(key, desc.type);
4848 return -EINVAL;
4849 }
4850
4851 return 0;
4852 }
4853
4854 static int nft_setelem_parse_data(struct nft_ctx *ctx, struct nft_set *set,
4855 struct nft_data_desc *desc,
4856 struct nft_data *data,
4857 struct nlattr *attr)
4858 {
4859 int err;
4860
4861 err = nft_data_init(ctx, data, NFT_DATA_VALUE_MAXLEN, desc, attr);
4862 if (err < 0)
4863 return err;
4864
4865 if (desc->type != NFT_DATA_VERDICT && desc->len != set->dlen) {
4866 nft_data_release(data, desc->type);
4867 return -EINVAL;
4868 }
4869
4870 return 0;
4871 }
4872
4873 static int nft_get_set_elem(struct nft_ctx *ctx, struct nft_set *set,
4874 const struct nlattr *attr)
4875 {
4876 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
4877 struct nft_set_elem elem;
4878 struct sk_buff *skb;
4879 uint32_t flags = 0;
4880 void *priv;
4881 int err;
4882
4883 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
4884 nft_set_elem_policy, NULL);
4885 if (err < 0)
4886 return err;
4887
4888 if (!nla[NFTA_SET_ELEM_KEY])
4889 return -EINVAL;
4890
4891 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
4892 if (err < 0)
4893 return err;
4894
4895 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
4896 nla[NFTA_SET_ELEM_KEY]);
4897 if (err < 0)
4898 return err;
4899
4900 if (nla[NFTA_SET_ELEM_KEY_END]) {
4901 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
4902 nla[NFTA_SET_ELEM_KEY_END]);
4903 if (err < 0)
4904 return err;
4905 }
4906
4907 priv = set->ops->get(ctx->net, set, &elem, flags);
4908 if (IS_ERR(priv))
4909 return PTR_ERR(priv);
4910
4911 elem.priv = priv;
4912
4913 err = -ENOMEM;
4914 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
4915 if (skb == NULL)
4916 return err;
4917
4918 err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
4919 NFT_MSG_NEWSETELEM, 0, set, &elem);
4920 if (err < 0)
4921 goto err_fill_setelem;
4922
4923 return nfnetlink_unicast(skb, ctx->net, ctx->portid);
4924
4925 err_fill_setelem:
4926 kfree_skb(skb);
4927 return err;
4928 }
4929
4930 /* called with rcu_read_lock held */
4931 static int nf_tables_getsetelem(struct net *net, struct sock *nlsk,
4932 struct sk_buff *skb, const struct nlmsghdr *nlh,
4933 const struct nlattr * const nla[],
4934 struct netlink_ext_ack *extack)
4935 {
4936 u8 genmask = nft_genmask_cur(net);
4937 struct nft_set *set;
4938 struct nlattr *attr;
4939 struct nft_ctx ctx;
4940 int rem, err = 0;
4941
4942 err = nft_ctx_init_from_elemattr(&ctx, net, skb, nlh, nla, extack,
4943 genmask);
4944 if (err < 0)
4945 return err;
4946
4947 set = nft_set_lookup(ctx.table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
4948 if (IS_ERR(set))
4949 return PTR_ERR(set);
4950
4951 if (nlh->nlmsg_flags & NLM_F_DUMP) {
4952 struct netlink_dump_control c = {
4953 .start = nf_tables_dump_set_start,
4954 .dump = nf_tables_dump_set,
4955 .done = nf_tables_dump_set_done,
4956 .module = THIS_MODULE,
4957 };
4958 struct nft_set_dump_ctx dump_ctx = {
4959 .set = set,
4960 .ctx = ctx,
4961 };
4962
4963 c.data = &dump_ctx;
4964 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
4965 }
4966
4967 if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
4968 return -EINVAL;
4969
4970 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
4971 err = nft_get_set_elem(&ctx, set, attr);
4972 if (err < 0)
4973 break;
4974 }
4975
4976 return err;
4977 }
4978
4979 static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
4980 const struct nft_set *set,
4981 const struct nft_set_elem *elem,
4982 int event, u16 flags)
4983 {
4984 struct net *net = ctx->net;
4985 u32 portid = ctx->portid;
4986 struct sk_buff *skb;
4987 int err;
4988 char *buf = kasprintf(GFP_KERNEL, "%s:%llu;%s:%llu",
4989 ctx->table->name, ctx->table->handle,
4990 set->name, set->handle);
4991
4992 audit_log_nfcfg(buf,
4993 ctx->family,
4994 set->handle,
4995 event == NFT_MSG_NEWSETELEM ?
4996 AUDIT_NFT_OP_SETELEM_REGISTER :
4997 AUDIT_NFT_OP_SETELEM_UNREGISTER,
4998 GFP_KERNEL);
4999 kfree(buf);
5000
5001 if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5002 return;
5003
5004 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5005 if (skb == NULL)
5006 goto err;
5007
5008 err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
5009 set, elem);
5010 if (err < 0) {
5011 kfree_skb(skb);
5012 goto err;
5013 }
5014
5015 nft_notify_enqueue(skb, ctx->report, &ctx->net->nft.notify_list);
5016 return;
5017 err:
5018 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
5019 }
5020
5021 static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
5022 int msg_type,
5023 struct nft_set *set)
5024 {
5025 struct nft_trans *trans;
5026
5027 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
5028 if (trans == NULL)
5029 return NULL;
5030
5031 nft_trans_elem_set(trans) = set;
5032 return trans;
5033 }
5034
5035 struct nft_expr *nft_set_elem_expr_alloc(const struct nft_ctx *ctx,
5036 const struct nft_set *set,
5037 const struct nlattr *attr)
5038 {
5039 struct nft_expr *expr;
5040 int err;
5041
5042 expr = nft_expr_init(ctx, attr);
5043 if (IS_ERR(expr))
5044 return expr;
5045
5046 err = -EOPNOTSUPP;
5047 if (!(expr->ops->type->flags & NFT_EXPR_STATEFUL))
5048 goto err_set_elem_expr;
5049
5050 if (expr->ops->type->flags & NFT_EXPR_GC) {
5051 if (set->flags & NFT_SET_TIMEOUT)
5052 goto err_set_elem_expr;
5053 if (!set->ops->gc_init)
5054 goto err_set_elem_expr;
5055 set->ops->gc_init(set);
5056 }
5057
5058 return expr;
5059
5060 err_set_elem_expr:
5061 nft_expr_destroy(ctx, expr);
5062 return ERR_PTR(err);
5063 }
5064
5065 void *nft_set_elem_init(const struct nft_set *set,
5066 const struct nft_set_ext_tmpl *tmpl,
5067 const u32 *key, const u32 *key_end,
5068 const u32 *data, u64 timeout, u64 expiration, gfp_t gfp)
5069 {
5070 struct nft_set_ext *ext;
5071 void *elem;
5072
5073 elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
5074 if (elem == NULL)
5075 return NULL;
5076
5077 ext = nft_set_elem_ext(set, elem);
5078 nft_set_ext_init(ext, tmpl);
5079
5080 memcpy(nft_set_ext_key(ext), key, set->klen);
5081 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END))
5082 memcpy(nft_set_ext_key_end(ext), key_end, set->klen);
5083 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5084 memcpy(nft_set_ext_data(ext), data, set->dlen);
5085 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5086 *nft_set_ext_expiration(ext) = get_jiffies_64() + expiration;
5087 if (expiration == 0)
5088 *nft_set_ext_expiration(ext) += timeout;
5089 }
5090 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
5091 *nft_set_ext_timeout(ext) = timeout;
5092
5093 return elem;
5094 }
5095
5096 static void nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5097 struct nft_expr *expr)
5098 {
5099 if (expr->ops->destroy_clone) {
5100 expr->ops->destroy_clone(ctx, expr);
5101 module_put(expr->ops->type->owner);
5102 } else {
5103 nf_tables_expr_destroy(ctx, expr);
5104 }
5105 }
5106
5107 void nft_set_elem_destroy(const struct nft_set *set, void *elem,
5108 bool destroy_expr)
5109 {
5110 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5111 struct nft_ctx ctx = {
5112 .net = read_pnet(&set->net),
5113 .family = set->table->family,
5114 };
5115
5116 nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
5117 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5118 nft_data_release(nft_set_ext_data(ext), set->dtype);
5119 if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPR))
5120 nft_set_elem_expr_destroy(&ctx, nft_set_ext_expr(ext));
5121
5122 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
5123 (*nft_set_ext_obj(ext))->use--;
5124 kfree(elem);
5125 }
5126 EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
5127
5128 /* Only called from commit path, nft_set_elem_deactivate() already deals with
5129 * the refcounting from the preparation phase.
5130 */
5131 static void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
5132 const struct nft_set *set, void *elem)
5133 {
5134 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5135
5136 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPR))
5137 nft_set_elem_expr_destroy(ctx, nft_set_ext_expr(ext));
5138
5139 kfree(elem);
5140 }
5141
5142 static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5143 const struct nlattr *attr, u32 nlmsg_flags)
5144 {
5145 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5146 u8 genmask = nft_genmask_next(ctx->net);
5147 struct nft_set_ext_tmpl tmpl;
5148 struct nft_set_ext *ext, *ext2;
5149 struct nft_set_elem elem;
5150 struct nft_set_binding *binding;
5151 struct nft_object *obj = NULL;
5152 struct nft_expr *expr = NULL;
5153 struct nft_userdata *udata;
5154 struct nft_data_desc desc;
5155 enum nft_registers dreg;
5156 struct nft_trans *trans;
5157 u32 flags = 0;
5158 u64 timeout;
5159 u64 expiration;
5160 u8 ulen;
5161 int err;
5162
5163 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5164 nft_set_elem_policy, NULL);
5165 if (err < 0)
5166 return err;
5167
5168 if (nla[NFTA_SET_ELEM_KEY] == NULL)
5169 return -EINVAL;
5170
5171 nft_set_ext_prepare(&tmpl);
5172
5173 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5174 if (err < 0)
5175 return err;
5176 if (flags != 0)
5177 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
5178
5179 if (set->flags & NFT_SET_MAP) {
5180 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
5181 !(flags & NFT_SET_ELEM_INTERVAL_END))
5182 return -EINVAL;
5183 } else {
5184 if (nla[NFTA_SET_ELEM_DATA] != NULL)
5185 return -EINVAL;
5186 }
5187
5188 if ((flags & NFT_SET_ELEM_INTERVAL_END) &&
5189 (nla[NFTA_SET_ELEM_DATA] ||
5190 nla[NFTA_SET_ELEM_OBJREF] ||
5191 nla[NFTA_SET_ELEM_TIMEOUT] ||
5192 nla[NFTA_SET_ELEM_EXPIRATION] ||
5193 nla[NFTA_SET_ELEM_USERDATA] ||
5194 nla[NFTA_SET_ELEM_EXPR]))
5195 return -EINVAL;
5196
5197 timeout = 0;
5198 if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
5199 if (!(set->flags & NFT_SET_TIMEOUT))
5200 return -EINVAL;
5201 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
5202 &timeout);
5203 if (err)
5204 return err;
5205 } else if (set->flags & NFT_SET_TIMEOUT) {
5206 timeout = set->timeout;
5207 }
5208
5209 expiration = 0;
5210 if (nla[NFTA_SET_ELEM_EXPIRATION] != NULL) {
5211 if (!(set->flags & NFT_SET_TIMEOUT))
5212 return -EINVAL;
5213 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_EXPIRATION],
5214 &expiration);
5215 if (err)
5216 return err;
5217 }
5218
5219 if (nla[NFTA_SET_ELEM_EXPR] != NULL) {
5220 expr = nft_set_elem_expr_alloc(ctx, set,
5221 nla[NFTA_SET_ELEM_EXPR]);
5222 if (IS_ERR(expr))
5223 return PTR_ERR(expr);
5224
5225 err = -EOPNOTSUPP;
5226 if (set->expr && set->expr->ops != expr->ops)
5227 goto err_set_elem_expr;
5228 } else if (set->expr) {
5229 expr = kzalloc(set->expr->ops->size, GFP_KERNEL);
5230 if (!expr)
5231 return -ENOMEM;
5232
5233 err = nft_expr_clone(expr, set->expr);
5234 if (err < 0)
5235 goto err_set_elem_expr;
5236 }
5237
5238 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5239 nla[NFTA_SET_ELEM_KEY]);
5240 if (err < 0)
5241 goto err_set_elem_expr;
5242
5243 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
5244
5245 if (nla[NFTA_SET_ELEM_KEY_END]) {
5246 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5247 nla[NFTA_SET_ELEM_KEY_END]);
5248 if (err < 0)
5249 goto err_parse_key;
5250
5251 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
5252 }
5253
5254 if (timeout > 0) {
5255 nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
5256 if (timeout != set->timeout)
5257 nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
5258 }
5259
5260 if (expr)
5261 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_EXPR,
5262 expr->ops->size);
5263
5264 if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
5265 if (!(set->flags & NFT_SET_OBJECT)) {
5266 err = -EINVAL;
5267 goto err_parse_key_end;
5268 }
5269 obj = nft_obj_lookup(ctx->net, ctx->table,
5270 nla[NFTA_SET_ELEM_OBJREF],
5271 set->objtype, genmask);
5272 if (IS_ERR(obj)) {
5273 err = PTR_ERR(obj);
5274 goto err_parse_key_end;
5275 }
5276 nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
5277 }
5278
5279 if (nla[NFTA_SET_ELEM_DATA] != NULL) {
5280 err = nft_setelem_parse_data(ctx, set, &desc, &elem.data.val,
5281 nla[NFTA_SET_ELEM_DATA]);
5282 if (err < 0)
5283 goto err_parse_key_end;
5284
5285 dreg = nft_type_to_reg(set->dtype);
5286 list_for_each_entry(binding, &set->bindings, list) {
5287 struct nft_ctx bind_ctx = {
5288 .net = ctx->net,
5289 .family = ctx->family,
5290 .table = ctx->table,
5291 .chain = (struct nft_chain *)binding->chain,
5292 };
5293
5294 if (!(binding->flags & NFT_SET_MAP))
5295 continue;
5296
5297 err = nft_validate_register_store(&bind_ctx, dreg,
5298 &elem.data.val,
5299 desc.type, desc.len);
5300 if (err < 0)
5301 goto err_parse_data;
5302
5303 if (desc.type == NFT_DATA_VERDICT &&
5304 (elem.data.val.verdict.code == NFT_GOTO ||
5305 elem.data.val.verdict.code == NFT_JUMP))
5306 nft_validate_state_update(ctx->net,
5307 NFT_VALIDATE_NEED);
5308 }
5309
5310 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, desc.len);
5311 }
5312
5313 /* The full maximum length of userdata can exceed the maximum
5314 * offset value (U8_MAX) for following extensions, therefor it
5315 * must be the last extension added.
5316 */
5317 ulen = 0;
5318 if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
5319 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
5320 if (ulen > 0)
5321 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
5322 ulen);
5323 }
5324
5325 err = -ENOMEM;
5326 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
5327 elem.key_end.val.data, elem.data.val.data,
5328 timeout, expiration, GFP_KERNEL);
5329 if (elem.priv == NULL)
5330 goto err_parse_data;
5331
5332 ext = nft_set_elem_ext(set, elem.priv);
5333 if (flags)
5334 *nft_set_ext_flags(ext) = flags;
5335 if (ulen > 0) {
5336 udata = nft_set_ext_userdata(ext);
5337 udata->len = ulen - 1;
5338 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
5339 }
5340 if (obj) {
5341 *nft_set_ext_obj(ext) = obj;
5342 obj->use++;
5343 }
5344 if (expr) {
5345 memcpy(nft_set_ext_expr(ext), expr, expr->ops->size);
5346 kfree(expr);
5347 expr = NULL;
5348 }
5349
5350 trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
5351 if (trans == NULL)
5352 goto err_trans;
5353
5354 ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
5355 err = set->ops->insert(ctx->net, set, &elem, &ext2);
5356 if (err) {
5357 if (err == -EEXIST) {
5358 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
5359 nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
5360 nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
5361 nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF))
5362 goto err_element_clash;
5363 if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5364 nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
5365 memcmp(nft_set_ext_data(ext),
5366 nft_set_ext_data(ext2), set->dlen) != 0) ||
5367 (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
5368 nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
5369 *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
5370 goto err_element_clash;
5371 else if (!(nlmsg_flags & NLM_F_EXCL))
5372 err = 0;
5373 } else if (err == -ENOTEMPTY) {
5374 /* ENOTEMPTY reports overlapping between this element
5375 * and an existing one.
5376 */
5377 err = -EEXIST;
5378 }
5379 goto err_element_clash;
5380 }
5381
5382 if (set->size &&
5383 !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact)) {
5384 err = -ENFILE;
5385 goto err_set_full;
5386 }
5387
5388 nft_trans_elem(trans) = elem;
5389 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
5390 return 0;
5391
5392 err_set_full:
5393 set->ops->remove(ctx->net, set, &elem);
5394 err_element_clash:
5395 kfree(trans);
5396 err_trans:
5397 if (obj)
5398 obj->use--;
5399
5400 nf_tables_set_elem_destroy(ctx, set, elem.priv);
5401 err_parse_data:
5402 if (nla[NFTA_SET_ELEM_DATA] != NULL)
5403 nft_data_release(&elem.data.val, desc.type);
5404 err_parse_key_end:
5405 nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
5406 err_parse_key:
5407 nft_data_release(&elem.key.val, NFT_DATA_VALUE);
5408 err_set_elem_expr:
5409 if (expr != NULL)
5410 nft_expr_destroy(ctx, expr);
5411
5412 return err;
5413 }
5414
5415 static int nf_tables_newsetelem(struct net *net, struct sock *nlsk,
5416 struct sk_buff *skb, const struct nlmsghdr *nlh,
5417 const struct nlattr * const nla[],
5418 struct netlink_ext_ack *extack)
5419 {
5420 u8 genmask = nft_genmask_next(net);
5421 const struct nlattr *attr;
5422 struct nft_set *set;
5423 struct nft_ctx ctx;
5424 int rem, err;
5425
5426 if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
5427 return -EINVAL;
5428
5429 err = nft_ctx_init_from_elemattr(&ctx, net, skb, nlh, nla, extack,
5430 genmask);
5431 if (err < 0)
5432 return err;
5433
5434 set = nft_set_lookup_global(net, ctx.table, nla[NFTA_SET_ELEM_LIST_SET],
5435 nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
5436 if (IS_ERR(set))
5437 return PTR_ERR(set);
5438
5439 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
5440 return -EBUSY;
5441
5442 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
5443 err = nft_add_set_elem(&ctx, set, attr, nlh->nlmsg_flags);
5444 if (err < 0)
5445 return err;
5446 }
5447
5448 if (net->nft.validate_state == NFT_VALIDATE_DO)
5449 return nft_table_validate(net, ctx.table);
5450
5451 return 0;
5452 }
5453
5454 /**
5455 * nft_data_hold - hold a nft_data item
5456 *
5457 * @data: struct nft_data to release
5458 * @type: type of data
5459 *
5460 * Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
5461 * NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
5462 * NFT_GOTO verdicts. This function must be called on active data objects
5463 * from the second phase of the commit protocol.
5464 */
5465 void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
5466 {
5467 struct nft_chain *chain;
5468 struct nft_rule *rule;
5469
5470 if (type == NFT_DATA_VERDICT) {
5471 switch (data->verdict.code) {
5472 case NFT_JUMP:
5473 case NFT_GOTO:
5474 chain = data->verdict.chain;
5475 chain->use++;
5476
5477 if (!nft_chain_is_bound(chain))
5478 break;
5479
5480 chain->table->use++;
5481 list_for_each_entry(rule, &chain->rules, list)
5482 chain->use++;
5483
5484 nft_chain_add(chain->table, chain);
5485 break;
5486 }
5487 }
5488 }
5489
5490 static void nft_set_elem_activate(const struct net *net,
5491 const struct nft_set *set,
5492 struct nft_set_elem *elem)
5493 {
5494 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5495
5496 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5497 nft_data_hold(nft_set_ext_data(ext), set->dtype);
5498 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
5499 (*nft_set_ext_obj(ext))->use++;
5500 }
5501
5502 static void nft_set_elem_deactivate(const struct net *net,
5503 const struct nft_set *set,
5504 struct nft_set_elem *elem)
5505 {
5506 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5507
5508 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5509 nft_data_release(nft_set_ext_data(ext), set->dtype);
5510 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
5511 (*nft_set_ext_obj(ext))->use--;
5512 }
5513
5514 static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
5515 const struct nlattr *attr)
5516 {
5517 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5518 struct nft_set_ext_tmpl tmpl;
5519 struct nft_set_elem elem;
5520 struct nft_set_ext *ext;
5521 struct nft_trans *trans;
5522 u32 flags = 0;
5523 void *priv;
5524 int err;
5525
5526 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5527 nft_set_elem_policy, NULL);
5528 if (err < 0)
5529 return err;
5530
5531 if (nla[NFTA_SET_ELEM_KEY] == NULL)
5532 return -EINVAL;
5533
5534 nft_set_ext_prepare(&tmpl);
5535
5536 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5537 if (err < 0)
5538 return err;
5539 if (flags != 0)
5540 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
5541
5542 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5543 nla[NFTA_SET_ELEM_KEY]);
5544 if (err < 0)
5545 return err;
5546
5547 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
5548
5549 if (nla[NFTA_SET_ELEM_KEY_END]) {
5550 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5551 nla[NFTA_SET_ELEM_KEY_END]);
5552 if (err < 0)
5553 return err;
5554
5555 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
5556 }
5557
5558 err = -ENOMEM;
5559 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
5560 elem.key_end.val.data, NULL, 0, 0,
5561 GFP_KERNEL);
5562 if (elem.priv == NULL)
5563 goto fail_elem;
5564
5565 ext = nft_set_elem_ext(set, elem.priv);
5566 if (flags)
5567 *nft_set_ext_flags(ext) = flags;
5568
5569 trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
5570 if (trans == NULL)
5571 goto fail_trans;
5572
5573 priv = set->ops->deactivate(ctx->net, set, &elem);
5574 if (priv == NULL) {
5575 err = -ENOENT;
5576 goto fail_ops;
5577 }
5578 kfree(elem.priv);
5579 elem.priv = priv;
5580
5581 nft_set_elem_deactivate(ctx->net, set, &elem);
5582
5583 nft_trans_elem(trans) = elem;
5584 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
5585 return 0;
5586
5587 fail_ops:
5588 kfree(trans);
5589 fail_trans:
5590 kfree(elem.priv);
5591 fail_elem:
5592 nft_data_release(&elem.key.val, NFT_DATA_VALUE);
5593 return err;
5594 }
5595
5596 static int nft_flush_set(const struct nft_ctx *ctx,
5597 struct nft_set *set,
5598 const struct nft_set_iter *iter,
5599 struct nft_set_elem *elem)
5600 {
5601 struct nft_trans *trans;
5602 int err;
5603
5604 trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
5605 sizeof(struct nft_trans_elem), GFP_ATOMIC);
5606 if (!trans)
5607 return -ENOMEM;
5608
5609 if (!set->ops->flush(ctx->net, set, elem->priv)) {
5610 err = -ENOENT;
5611 goto err1;
5612 }
5613 set->ndeact++;
5614
5615 nft_set_elem_deactivate(ctx->net, set, elem);
5616 nft_trans_elem_set(trans) = set;
5617 nft_trans_elem(trans) = *elem;
5618 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
5619
5620 return 0;
5621 err1:
5622 kfree(trans);
5623 return err;
5624 }
5625
5626 static int nf_tables_delsetelem(struct net *net, struct sock *nlsk,
5627 struct sk_buff *skb, const struct nlmsghdr *nlh,
5628 const struct nlattr * const nla[],
5629 struct netlink_ext_ack *extack)
5630 {
5631 u8 genmask = nft_genmask_next(net);
5632 const struct nlattr *attr;
5633 struct nft_set *set;
5634 struct nft_ctx ctx;
5635 int rem, err = 0;
5636
5637 err = nft_ctx_init_from_elemattr(&ctx, net, skb, nlh, nla, extack,
5638 genmask);
5639 if (err < 0)
5640 return err;
5641
5642 set = nft_set_lookup(ctx.table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
5643 if (IS_ERR(set))
5644 return PTR_ERR(set);
5645 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
5646 return -EBUSY;
5647
5648 if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL) {
5649 struct nft_set_iter iter = {
5650 .genmask = genmask,
5651 .fn = nft_flush_set,
5652 };
5653 set->ops->walk(&ctx, set, &iter);
5654
5655 return iter.err;
5656 }
5657
5658 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
5659 err = nft_del_setelem(&ctx, set, attr);
5660 if (err < 0)
5661 break;
5662
5663 set->ndeact++;
5664 }
5665 return err;
5666 }
5667
5668 void nft_set_gc_batch_release(struct rcu_head *rcu)
5669 {
5670 struct nft_set_gc_batch *gcb;
5671 unsigned int i;
5672
5673 gcb = container_of(rcu, struct nft_set_gc_batch, head.rcu);
5674 for (i = 0; i < gcb->head.cnt; i++)
5675 nft_set_elem_destroy(gcb->head.set, gcb->elems[i], true);
5676 kfree(gcb);
5677 }
5678
5679 struct nft_set_gc_batch *nft_set_gc_batch_alloc(const struct nft_set *set,
5680 gfp_t gfp)
5681 {
5682 struct nft_set_gc_batch *gcb;
5683
5684 gcb = kzalloc(sizeof(*gcb), gfp);
5685 if (gcb == NULL)
5686 return gcb;
5687 gcb->head.set = set;
5688 return gcb;
5689 }
5690
5691 /*
5692 * Stateful objects
5693 */
5694
5695 /**
5696 * nft_register_obj- register nf_tables stateful object type
5697 * @obj_type: object type
5698 *
5699 * Registers the object type for use with nf_tables. Returns zero on
5700 * success or a negative errno code otherwise.
5701 */
5702 int nft_register_obj(struct nft_object_type *obj_type)
5703 {
5704 if (obj_type->type == NFT_OBJECT_UNSPEC)
5705 return -EINVAL;
5706
5707 nfnl_lock(NFNL_SUBSYS_NFTABLES);
5708 list_add_rcu(&obj_type->list, &nf_tables_objects);
5709 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
5710 return 0;
5711 }
5712 EXPORT_SYMBOL_GPL(nft_register_obj);
5713
5714 /**
5715 * nft_unregister_obj - unregister nf_tables object type
5716 * @obj_type: object type
5717 *
5718 * Unregisters the object type for use with nf_tables.
5719 */
5720 void nft_unregister_obj(struct nft_object_type *obj_type)
5721 {
5722 nfnl_lock(NFNL_SUBSYS_NFTABLES);
5723 list_del_rcu(&obj_type->list);
5724 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
5725 }
5726 EXPORT_SYMBOL_GPL(nft_unregister_obj);
5727
5728 struct nft_object *nft_obj_lookup(const struct net *net,
5729 const struct nft_table *table,
5730 const struct nlattr *nla, u32 objtype,
5731 u8 genmask)
5732 {
5733 struct nft_object_hash_key k = { .table = table };
5734 char search[NFT_OBJ_MAXNAMELEN];
5735 struct rhlist_head *tmp, *list;
5736 struct nft_object *obj;
5737
5738 nla_strlcpy(search, nla, sizeof(search));
5739 k.name = search;
5740
5741 WARN_ON_ONCE(!rcu_read_lock_held() &&
5742 !lockdep_commit_lock_is_held(net));
5743
5744 rcu_read_lock();
5745 list = rhltable_lookup(&nft_objname_ht, &k, nft_objname_ht_params);
5746 if (!list)
5747 goto out;
5748
5749 rhl_for_each_entry_rcu(obj, tmp, list, rhlhead) {
5750 if (objtype == obj->ops->type->type &&
5751 nft_active_genmask(obj, genmask)) {
5752 rcu_read_unlock();
5753 return obj;
5754 }
5755 }
5756 out:
5757 rcu_read_unlock();
5758 return ERR_PTR(-ENOENT);
5759 }
5760 EXPORT_SYMBOL_GPL(nft_obj_lookup);
5761
5762 static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
5763 const struct nlattr *nla,
5764 u32 objtype, u8 genmask)
5765 {
5766 struct nft_object *obj;
5767
5768 list_for_each_entry(obj, &table->objects, list) {
5769 if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
5770 objtype == obj->ops->type->type &&
5771 nft_active_genmask(obj, genmask))
5772 return obj;
5773 }
5774 return ERR_PTR(-ENOENT);
5775 }
5776
5777 static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
5778 [NFTA_OBJ_TABLE] = { .type = NLA_STRING,
5779 .len = NFT_TABLE_MAXNAMELEN - 1 },
5780 [NFTA_OBJ_NAME] = { .type = NLA_STRING,
5781 .len = NFT_OBJ_MAXNAMELEN - 1 },
5782 [NFTA_OBJ_TYPE] = { .type = NLA_U32 },
5783 [NFTA_OBJ_DATA] = { .type = NLA_NESTED },
5784 [NFTA_OBJ_HANDLE] = { .type = NLA_U64},
5785 [NFTA_OBJ_USERDATA] = { .type = NLA_BINARY,
5786 .len = NFT_USERDATA_MAXLEN },
5787 };
5788
5789 static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
5790 const struct nft_object_type *type,
5791 const struct nlattr *attr)
5792 {
5793 struct nlattr **tb;
5794 const struct nft_object_ops *ops;
5795 struct nft_object *obj;
5796 int err = -ENOMEM;
5797
5798 tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
5799 if (!tb)
5800 goto err1;
5801
5802 if (attr) {
5803 err = nla_parse_nested_deprecated(tb, type->maxattr, attr,
5804 type->policy, NULL);
5805 if (err < 0)
5806 goto err2;
5807 } else {
5808 memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
5809 }
5810
5811 if (type->select_ops) {
5812 ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
5813 if (IS_ERR(ops)) {
5814 err = PTR_ERR(ops);
5815 goto err2;
5816 }
5817 } else {
5818 ops = type->ops;
5819 }
5820
5821 err = -ENOMEM;
5822 obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL);
5823 if (!obj)
5824 goto err2;
5825
5826 err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
5827 if (err < 0)
5828 goto err3;
5829
5830 obj->ops = ops;
5831
5832 kfree(tb);
5833 return obj;
5834 err3:
5835 kfree(obj);
5836 err2:
5837 kfree(tb);
5838 err1:
5839 return ERR_PTR(err);
5840 }
5841
5842 static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
5843 struct nft_object *obj, bool reset)
5844 {
5845 struct nlattr *nest;
5846
5847 nest = nla_nest_start_noflag(skb, attr);
5848 if (!nest)
5849 goto nla_put_failure;
5850 if (obj->ops->dump(skb, obj, reset) < 0)
5851 goto nla_put_failure;
5852 nla_nest_end(skb, nest);
5853 return 0;
5854
5855 nla_put_failure:
5856 return -1;
5857 }
5858
5859 static const struct nft_object_type *__nft_obj_type_get(u32 objtype)
5860 {
5861 const struct nft_object_type *type;
5862
5863 list_for_each_entry(type, &nf_tables_objects, list) {
5864 if (objtype == type->type)
5865 return type;
5866 }
5867 return NULL;
5868 }
5869
5870 static const struct nft_object_type *
5871 nft_obj_type_get(struct net *net, u32 objtype)
5872 {
5873 const struct nft_object_type *type;
5874
5875 type = __nft_obj_type_get(objtype);
5876 if (type != NULL && try_module_get(type->owner))
5877 return type;
5878
5879 lockdep_nfnl_nft_mutex_not_held();
5880 #ifdef CONFIG_MODULES
5881 if (type == NULL) {
5882 if (nft_request_module(net, "nft-obj-%u", objtype) == -EAGAIN)
5883 return ERR_PTR(-EAGAIN);
5884 }
5885 #endif
5886 return ERR_PTR(-ENOENT);
5887 }
5888
5889 static int nf_tables_updobj(const struct nft_ctx *ctx,
5890 const struct nft_object_type *type,
5891 const struct nlattr *attr,
5892 struct nft_object *obj)
5893 {
5894 struct nft_object *newobj;
5895 struct nft_trans *trans;
5896 int err;
5897
5898 trans = nft_trans_alloc(ctx, NFT_MSG_NEWOBJ,
5899 sizeof(struct nft_trans_obj));
5900 if (!trans)
5901 return -ENOMEM;
5902
5903 newobj = nft_obj_init(ctx, type, attr);
5904 if (IS_ERR(newobj)) {
5905 err = PTR_ERR(newobj);
5906 goto err_free_trans;
5907 }
5908
5909 nft_trans_obj(trans) = obj;
5910 nft_trans_obj_update(trans) = true;
5911 nft_trans_obj_newobj(trans) = newobj;
5912 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
5913
5914 return 0;
5915
5916 err_free_trans:
5917 kfree(trans);
5918 return err;
5919 }
5920
5921 static int nf_tables_newobj(struct net *net, struct sock *nlsk,
5922 struct sk_buff *skb, const struct nlmsghdr *nlh,
5923 const struct nlattr * const nla[],
5924 struct netlink_ext_ack *extack)
5925 {
5926 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
5927 const struct nft_object_type *type;
5928 u8 genmask = nft_genmask_next(net);
5929 int family = nfmsg->nfgen_family;
5930 struct nft_table *table;
5931 struct nft_object *obj;
5932 struct nft_ctx ctx;
5933 u32 objtype;
5934 int err;
5935
5936 if (!nla[NFTA_OBJ_TYPE] ||
5937 !nla[NFTA_OBJ_NAME] ||
5938 !nla[NFTA_OBJ_DATA])
5939 return -EINVAL;
5940
5941 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask);
5942 if (IS_ERR(table)) {
5943 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
5944 return PTR_ERR(table);
5945 }
5946
5947 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
5948 obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
5949 if (IS_ERR(obj)) {
5950 err = PTR_ERR(obj);
5951 if (err != -ENOENT) {
5952 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
5953 return err;
5954 }
5955 } else {
5956 if (nlh->nlmsg_flags & NLM_F_EXCL) {
5957 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
5958 return -EEXIST;
5959 }
5960 if (nlh->nlmsg_flags & NLM_F_REPLACE)
5961 return -EOPNOTSUPP;
5962
5963 type = __nft_obj_type_get(objtype);
5964 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
5965
5966 return nf_tables_updobj(&ctx, type, nla[NFTA_OBJ_DATA], obj);
5967 }
5968
5969 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
5970
5971 type = nft_obj_type_get(net, objtype);
5972 if (IS_ERR(type))
5973 return PTR_ERR(type);
5974
5975 obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
5976 if (IS_ERR(obj)) {
5977 err = PTR_ERR(obj);
5978 goto err_init;
5979 }
5980 obj->key.table = table;
5981 obj->handle = nf_tables_alloc_handle(table);
5982
5983 obj->key.name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL);
5984 if (!obj->key.name) {
5985 err = -ENOMEM;
5986 goto err_strdup;
5987 }
5988
5989 if (nla[NFTA_OBJ_USERDATA]) {
5990 obj->udata = nla_memdup(nla[NFTA_OBJ_USERDATA], GFP_KERNEL);
5991 if (obj->udata == NULL)
5992 goto err_userdata;
5993
5994 obj->udlen = nla_len(nla[NFTA_OBJ_USERDATA]);
5995 }
5996
5997 err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
5998 if (err < 0)
5999 goto err_trans;
6000
6001 err = rhltable_insert(&nft_objname_ht, &obj->rhlhead,
6002 nft_objname_ht_params);
6003 if (err < 0)
6004 goto err_obj_ht;
6005
6006 list_add_tail_rcu(&obj->list, &table->objects);
6007 table->use++;
6008 return 0;
6009 err_obj_ht:
6010 /* queued in transaction log */
6011 INIT_LIST_HEAD(&obj->list);
6012 return err;
6013 err_trans:
6014 kfree(obj->key.name);
6015 err_userdata:
6016 kfree(obj->udata);
6017 err_strdup:
6018 if (obj->ops->destroy)
6019 obj->ops->destroy(&ctx, obj);
6020 kfree(obj);
6021 err_init:
6022 module_put(type->owner);
6023 return err;
6024 }
6025
6026 static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
6027 u32 portid, u32 seq, int event, u32 flags,
6028 int family, const struct nft_table *table,
6029 struct nft_object *obj, bool reset)
6030 {
6031 struct nfgenmsg *nfmsg;
6032 struct nlmsghdr *nlh;
6033
6034 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
6035 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
6036 if (nlh == NULL)
6037 goto nla_put_failure;
6038
6039 nfmsg = nlmsg_data(nlh);
6040 nfmsg->nfgen_family = family;
6041 nfmsg->version = NFNETLINK_V0;
6042 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
6043
6044 if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
6045 nla_put_string(skb, NFTA_OBJ_NAME, obj->key.name) ||
6046 nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
6047 nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
6048 nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset) ||
6049 nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
6050 NFTA_OBJ_PAD))
6051 goto nla_put_failure;
6052
6053 if (obj->udata &&
6054 nla_put(skb, NFTA_OBJ_USERDATA, obj->udlen, obj->udata))
6055 goto nla_put_failure;
6056
6057 nlmsg_end(skb, nlh);
6058 return 0;
6059
6060 nla_put_failure:
6061 nlmsg_trim(skb, nlh);
6062 return -1;
6063 }
6064
6065 struct nft_obj_filter {
6066 char *table;
6067 u32 type;
6068 };
6069
6070 static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
6071 {
6072 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
6073 const struct nft_table *table;
6074 unsigned int idx = 0, s_idx = cb->args[0];
6075 struct nft_obj_filter *filter = cb->data;
6076 struct net *net = sock_net(skb->sk);
6077 int family = nfmsg->nfgen_family;
6078 struct nft_object *obj;
6079 bool reset = false;
6080
6081 if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6082 reset = true;
6083
6084 rcu_read_lock();
6085 cb->seq = net->nft.base_seq;
6086
6087 list_for_each_entry_rcu(table, &net->nft.tables, list) {
6088 if (family != NFPROTO_UNSPEC && family != table->family)
6089 continue;
6090
6091 list_for_each_entry_rcu(obj, &table->objects, list) {
6092 if (!nft_is_active(net, obj))
6093 goto cont;
6094 if (idx < s_idx)
6095 goto cont;
6096 if (idx > s_idx)
6097 memset(&cb->args[1], 0,
6098 sizeof(cb->args) - sizeof(cb->args[0]));
6099 if (filter && filter->table &&
6100 strcmp(filter->table, table->name))
6101 goto cont;
6102 if (filter &&
6103 filter->type != NFT_OBJECT_UNSPEC &&
6104 obj->ops->type->type != filter->type)
6105 goto cont;
6106
6107 if (reset) {
6108 char *buf = kasprintf(GFP_ATOMIC,
6109 "%s:%llu;?:0",
6110 table->name,
6111 table->handle);
6112
6113 audit_log_nfcfg(buf,
6114 family,
6115 obj->handle,
6116 AUDIT_NFT_OP_OBJ_RESET,
6117 GFP_ATOMIC);
6118 kfree(buf);
6119 }
6120
6121 if (nf_tables_fill_obj_info(skb, net, NETLINK_CB(cb->skb).portid,
6122 cb->nlh->nlmsg_seq,
6123 NFT_MSG_NEWOBJ,
6124 NLM_F_MULTI | NLM_F_APPEND,
6125 table->family, table,
6126 obj, reset) < 0)
6127 goto done;
6128
6129 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
6130 cont:
6131 idx++;
6132 }
6133 }
6134 done:
6135 rcu_read_unlock();
6136
6137 cb->args[0] = idx;
6138 return skb->len;
6139 }
6140
6141 static int nf_tables_dump_obj_start(struct netlink_callback *cb)
6142 {
6143 const struct nlattr * const *nla = cb->data;
6144 struct nft_obj_filter *filter = NULL;
6145
6146 if (nla[NFTA_OBJ_TABLE] || nla[NFTA_OBJ_TYPE]) {
6147 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
6148 if (!filter)
6149 return -ENOMEM;
6150
6151 if (nla[NFTA_OBJ_TABLE]) {
6152 filter->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
6153 if (!filter->table) {
6154 kfree(filter);
6155 return -ENOMEM;
6156 }
6157 }
6158
6159 if (nla[NFTA_OBJ_TYPE])
6160 filter->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6161 }
6162
6163 cb->data = filter;
6164 return 0;
6165 }
6166
6167 static int nf_tables_dump_obj_done(struct netlink_callback *cb)
6168 {
6169 struct nft_obj_filter *filter = cb->data;
6170
6171 if (filter) {
6172 kfree(filter->table);
6173 kfree(filter);
6174 }
6175
6176 return 0;
6177 }
6178
6179 /* called with rcu_read_lock held */
6180 static int nf_tables_getobj(struct net *net, struct sock *nlsk,
6181 struct sk_buff *skb, const struct nlmsghdr *nlh,
6182 const struct nlattr * const nla[],
6183 struct netlink_ext_ack *extack)
6184 {
6185 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
6186 u8 genmask = nft_genmask_cur(net);
6187 int family = nfmsg->nfgen_family;
6188 const struct nft_table *table;
6189 struct nft_object *obj;
6190 struct sk_buff *skb2;
6191 bool reset = false;
6192 u32 objtype;
6193 int err;
6194
6195 if (nlh->nlmsg_flags & NLM_F_DUMP) {
6196 struct netlink_dump_control c = {
6197 .start = nf_tables_dump_obj_start,
6198 .dump = nf_tables_dump_obj,
6199 .done = nf_tables_dump_obj_done,
6200 .module = THIS_MODULE,
6201 .data = (void *)nla,
6202 };
6203
6204 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
6205 }
6206
6207 if (!nla[NFTA_OBJ_NAME] ||
6208 !nla[NFTA_OBJ_TYPE])
6209 return -EINVAL;
6210
6211 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask);
6212 if (IS_ERR(table)) {
6213 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6214 return PTR_ERR(table);
6215 }
6216
6217 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6218 obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6219 if (IS_ERR(obj)) {
6220 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6221 return PTR_ERR(obj);
6222 }
6223
6224 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
6225 if (!skb2)
6226 return -ENOMEM;
6227
6228 if (NFNL_MSG_TYPE(nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6229 reset = true;
6230
6231 if (reset) {
6232 char *buf = kasprintf(GFP_ATOMIC, "%s:%llu;?:0",
6233 table->name, table->handle);
6234
6235 audit_log_nfcfg(buf,
6236 family,
6237 obj->handle,
6238 AUDIT_NFT_OP_OBJ_RESET,
6239 GFP_ATOMIC);
6240 kfree(buf);
6241 }
6242
6243 err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
6244 nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
6245 family, table, obj, reset);
6246 if (err < 0)
6247 goto err_fill_obj_info;
6248
6249 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
6250
6251 err_fill_obj_info:
6252 kfree_skb(skb2);
6253 return err;
6254 }
6255
6256 static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
6257 {
6258 if (obj->ops->destroy)
6259 obj->ops->destroy(ctx, obj);
6260
6261 module_put(obj->ops->type->owner);
6262 kfree(obj->key.name);
6263 kfree(obj->udata);
6264 kfree(obj);
6265 }
6266
6267 static int nf_tables_delobj(struct net *net, struct sock *nlsk,
6268 struct sk_buff *skb, const struct nlmsghdr *nlh,
6269 const struct nlattr * const nla[],
6270 struct netlink_ext_ack *extack)
6271 {
6272 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
6273 u8 genmask = nft_genmask_next(net);
6274 int family = nfmsg->nfgen_family;
6275 const struct nlattr *attr;
6276 struct nft_table *table;
6277 struct nft_object *obj;
6278 struct nft_ctx ctx;
6279 u32 objtype;
6280
6281 if (!nla[NFTA_OBJ_TYPE] ||
6282 (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
6283 return -EINVAL;
6284
6285 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask);
6286 if (IS_ERR(table)) {
6287 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6288 return PTR_ERR(table);
6289 }
6290
6291 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6292 if (nla[NFTA_OBJ_HANDLE]) {
6293 attr = nla[NFTA_OBJ_HANDLE];
6294 obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
6295 } else {
6296 attr = nla[NFTA_OBJ_NAME];
6297 obj = nft_obj_lookup(net, table, attr, objtype, genmask);
6298 }
6299
6300 if (IS_ERR(obj)) {
6301 NL_SET_BAD_ATTR(extack, attr);
6302 return PTR_ERR(obj);
6303 }
6304 if (obj->use > 0) {
6305 NL_SET_BAD_ATTR(extack, attr);
6306 return -EBUSY;
6307 }
6308
6309 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
6310
6311 return nft_delobj(&ctx, obj);
6312 }
6313
6314 void nft_obj_notify(struct net *net, const struct nft_table *table,
6315 struct nft_object *obj, u32 portid, u32 seq, int event,
6316 int family, int report, gfp_t gfp)
6317 {
6318 struct sk_buff *skb;
6319 int err;
6320 char *buf = kasprintf(gfp, "%s:%llu;?:0",
6321 table->name, table->handle);
6322
6323 audit_log_nfcfg(buf,
6324 family,
6325 obj->handle,
6326 event == NFT_MSG_NEWOBJ ?
6327 AUDIT_NFT_OP_OBJ_REGISTER :
6328 AUDIT_NFT_OP_OBJ_UNREGISTER,
6329 gfp);
6330 kfree(buf);
6331
6332 if (!report &&
6333 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
6334 return;
6335
6336 skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
6337 if (skb == NULL)
6338 goto err;
6339
6340 err = nf_tables_fill_obj_info(skb, net, portid, seq, event, 0, family,
6341 table, obj, false);
6342 if (err < 0) {
6343 kfree_skb(skb);
6344 goto err;
6345 }
6346
6347 nft_notify_enqueue(skb, report, &net->nft.notify_list);
6348 return;
6349 err:
6350 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
6351 }
6352 EXPORT_SYMBOL_GPL(nft_obj_notify);
6353
6354 static void nf_tables_obj_notify(const struct nft_ctx *ctx,
6355 struct nft_object *obj, int event)
6356 {
6357 nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid, ctx->seq, event,
6358 ctx->family, ctx->report, GFP_KERNEL);
6359 }
6360
6361 /*
6362 * Flow tables
6363 */
6364 void nft_register_flowtable_type(struct nf_flowtable_type *type)
6365 {
6366 nfnl_lock(NFNL_SUBSYS_NFTABLES);
6367 list_add_tail_rcu(&type->list, &nf_tables_flowtables);
6368 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6369 }
6370 EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
6371
6372 void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
6373 {
6374 nfnl_lock(NFNL_SUBSYS_NFTABLES);
6375 list_del_rcu(&type->list);
6376 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6377 }
6378 EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
6379
6380 static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
6381 [NFTA_FLOWTABLE_TABLE] = { .type = NLA_STRING,
6382 .len = NFT_NAME_MAXLEN - 1 },
6383 [NFTA_FLOWTABLE_NAME] = { .type = NLA_STRING,
6384 .len = NFT_NAME_MAXLEN - 1 },
6385 [NFTA_FLOWTABLE_HOOK] = { .type = NLA_NESTED },
6386 [NFTA_FLOWTABLE_HANDLE] = { .type = NLA_U64 },
6387 [NFTA_FLOWTABLE_FLAGS] = { .type = NLA_U32 },
6388 };
6389
6390 struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
6391 const struct nlattr *nla, u8 genmask)
6392 {
6393 struct nft_flowtable *flowtable;
6394
6395 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
6396 if (!nla_strcmp(nla, flowtable->name) &&
6397 nft_active_genmask(flowtable, genmask))
6398 return flowtable;
6399 }
6400 return ERR_PTR(-ENOENT);
6401 }
6402 EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
6403
6404 void nf_tables_deactivate_flowtable(const struct nft_ctx *ctx,
6405 struct nft_flowtable *flowtable,
6406 enum nft_trans_phase phase)
6407 {
6408 switch (phase) {
6409 case NFT_TRANS_PREPARE:
6410 case NFT_TRANS_ABORT:
6411 case NFT_TRANS_RELEASE:
6412 flowtable->use--;
6413 fallthrough;
6414 default:
6415 return;
6416 }
6417 }
6418 EXPORT_SYMBOL_GPL(nf_tables_deactivate_flowtable);
6419
6420 static struct nft_flowtable *
6421 nft_flowtable_lookup_byhandle(const struct nft_table *table,
6422 const struct nlattr *nla, u8 genmask)
6423 {
6424 struct nft_flowtable *flowtable;
6425
6426 list_for_each_entry(flowtable, &table->flowtables, list) {
6427 if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
6428 nft_active_genmask(flowtable, genmask))
6429 return flowtable;
6430 }
6431 return ERR_PTR(-ENOENT);
6432 }
6433
6434 struct nft_flowtable_hook {
6435 u32 num;
6436 int priority;
6437 struct list_head list;
6438 };
6439
6440 static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
6441 [NFTA_FLOWTABLE_HOOK_NUM] = { .type = NLA_U32 },
6442 [NFTA_FLOWTABLE_HOOK_PRIORITY] = { .type = NLA_U32 },
6443 [NFTA_FLOWTABLE_HOOK_DEVS] = { .type = NLA_NESTED },
6444 };
6445
6446 static int nft_flowtable_parse_hook(const struct nft_ctx *ctx,
6447 const struct nlattr *attr,
6448 struct nft_flowtable_hook *flowtable_hook,
6449 struct nft_flowtable *flowtable, bool add)
6450 {
6451 struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
6452 struct nft_hook *hook;
6453 int hooknum, priority;
6454 int err;
6455
6456 INIT_LIST_HEAD(&flowtable_hook->list);
6457
6458 err = nla_parse_nested_deprecated(tb, NFTA_FLOWTABLE_HOOK_MAX, attr,
6459 nft_flowtable_hook_policy, NULL);
6460 if (err < 0)
6461 return err;
6462
6463 if (add) {
6464 if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
6465 !tb[NFTA_FLOWTABLE_HOOK_PRIORITY])
6466 return -EINVAL;
6467
6468 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
6469 if (hooknum != NF_NETDEV_INGRESS)
6470 return -EOPNOTSUPP;
6471
6472 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
6473
6474 flowtable_hook->priority = priority;
6475 flowtable_hook->num = hooknum;
6476 } else {
6477 if (tb[NFTA_FLOWTABLE_HOOK_NUM]) {
6478 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
6479 if (hooknum != flowtable->hooknum)
6480 return -EOPNOTSUPP;
6481 }
6482
6483 if (tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
6484 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
6485 if (priority != flowtable->data.priority)
6486 return -EOPNOTSUPP;
6487 }
6488
6489 flowtable_hook->priority = flowtable->data.priority;
6490 flowtable_hook->num = flowtable->hooknum;
6491 }
6492
6493 if (tb[NFTA_FLOWTABLE_HOOK_DEVS]) {
6494 err = nf_tables_parse_netdev_hooks(ctx->net,
6495 tb[NFTA_FLOWTABLE_HOOK_DEVS],
6496 &flowtable_hook->list);
6497 if (err < 0)
6498 return err;
6499 }
6500
6501 list_for_each_entry(hook, &flowtable_hook->list, list) {
6502 hook->ops.pf = NFPROTO_NETDEV;
6503 hook->ops.hooknum = flowtable_hook->num;
6504 hook->ops.priority = flowtable_hook->priority;
6505 hook->ops.priv = &flowtable->data;
6506 hook->ops.hook = flowtable->data.type->hook;
6507 }
6508
6509 return err;
6510 }
6511
6512 static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
6513 {
6514 const struct nf_flowtable_type *type;
6515
6516 list_for_each_entry(type, &nf_tables_flowtables, list) {
6517 if (family == type->family)
6518 return type;
6519 }
6520 return NULL;
6521 }
6522
6523 static const struct nf_flowtable_type *
6524 nft_flowtable_type_get(struct net *net, u8 family)
6525 {
6526 const struct nf_flowtable_type *type;
6527
6528 type = __nft_flowtable_type_get(family);
6529 if (type != NULL && try_module_get(type->owner))
6530 return type;
6531
6532 lockdep_nfnl_nft_mutex_not_held();
6533 #ifdef CONFIG_MODULES
6534 if (type == NULL) {
6535 if (nft_request_module(net, "nf-flowtable-%u", family) == -EAGAIN)
6536 return ERR_PTR(-EAGAIN);
6537 }
6538 #endif
6539 return ERR_PTR(-ENOENT);
6540 }
6541
6542 /* Only called from error and netdev event paths. */
6543 static void nft_unregister_flowtable_hook(struct net *net,
6544 struct nft_flowtable *flowtable,
6545 struct nft_hook *hook)
6546 {
6547 nf_unregister_net_hook(net, &hook->ops);
6548 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
6549 FLOW_BLOCK_UNBIND);
6550 }
6551
6552 static void nft_unregister_flowtable_net_hooks(struct net *net,
6553 struct list_head *hook_list)
6554 {
6555 struct nft_hook *hook;
6556
6557 list_for_each_entry(hook, hook_list, list)
6558 nf_unregister_net_hook(net, &hook->ops);
6559 }
6560
6561 static int nft_register_flowtable_net_hooks(struct net *net,
6562 struct nft_table *table,
6563 struct list_head *hook_list,
6564 struct nft_flowtable *flowtable)
6565 {
6566 struct nft_hook *hook, *hook2, *next;
6567 struct nft_flowtable *ft;
6568 int err, i = 0;
6569
6570 list_for_each_entry(hook, hook_list, list) {
6571 list_for_each_entry(ft, &table->flowtables, list) {
6572 list_for_each_entry(hook2, &ft->hook_list, list) {
6573 if (hook->ops.dev == hook2->ops.dev &&
6574 hook->ops.pf == hook2->ops.pf) {
6575 err = -EEXIST;
6576 goto err_unregister_net_hooks;
6577 }
6578 }
6579 }
6580
6581 err = flowtable->data.type->setup(&flowtable->data,
6582 hook->ops.dev,
6583 FLOW_BLOCK_BIND);
6584 if (err < 0)
6585 goto err_unregister_net_hooks;
6586
6587 err = nf_register_net_hook(net, &hook->ops);
6588 if (err < 0) {
6589 flowtable->data.type->setup(&flowtable->data,
6590 hook->ops.dev,
6591 FLOW_BLOCK_UNBIND);
6592 goto err_unregister_net_hooks;
6593 }
6594
6595 i++;
6596 }
6597
6598 return 0;
6599
6600 err_unregister_net_hooks:
6601 list_for_each_entry_safe(hook, next, hook_list, list) {
6602 if (i-- <= 0)
6603 break;
6604
6605 nft_unregister_flowtable_hook(net, flowtable, hook);
6606 list_del_rcu(&hook->list);
6607 kfree_rcu(hook, rcu);
6608 }
6609
6610 return err;
6611 }
6612
6613 static void nft_flowtable_hooks_destroy(struct list_head *hook_list)
6614 {
6615 struct nft_hook *hook, *next;
6616
6617 list_for_each_entry_safe(hook, next, hook_list, list) {
6618 list_del_rcu(&hook->list);
6619 kfree_rcu(hook, rcu);
6620 }
6621 }
6622
6623 static int nft_flowtable_update(struct nft_ctx *ctx, const struct nlmsghdr *nlh,
6624 struct nft_flowtable *flowtable)
6625 {
6626 const struct nlattr * const *nla = ctx->nla;
6627 struct nft_flowtable_hook flowtable_hook;
6628 struct nft_hook *hook, *next;
6629 struct nft_trans *trans;
6630 bool unregister = false;
6631 int err;
6632
6633 err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
6634 &flowtable_hook, flowtable, false);
6635 if (err < 0)
6636 return err;
6637
6638 list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
6639 if (nft_hook_list_find(&flowtable->hook_list, hook)) {
6640 list_del(&hook->list);
6641 kfree(hook);
6642 }
6643 }
6644
6645 err = nft_register_flowtable_net_hooks(ctx->net, ctx->table,
6646 &flowtable_hook.list, flowtable);
6647 if (err < 0)
6648 goto err_flowtable_update_hook;
6649
6650 trans = nft_trans_alloc(ctx, NFT_MSG_NEWFLOWTABLE,
6651 sizeof(struct nft_trans_flowtable));
6652 if (!trans) {
6653 unregister = true;
6654 err = -ENOMEM;
6655 goto err_flowtable_update_hook;
6656 }
6657
6658 nft_trans_flowtable(trans) = flowtable;
6659 nft_trans_flowtable_update(trans) = true;
6660 INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
6661 list_splice(&flowtable_hook.list, &nft_trans_flowtable_hooks(trans));
6662
6663 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
6664
6665 return 0;
6666
6667 err_flowtable_update_hook:
6668 list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
6669 if (unregister)
6670 nft_unregister_flowtable_hook(ctx->net, flowtable, hook);
6671 list_del_rcu(&hook->list);
6672 kfree_rcu(hook, rcu);
6673 }
6674
6675 return err;
6676
6677 }
6678
6679 static int nf_tables_newflowtable(struct net *net, struct sock *nlsk,
6680 struct sk_buff *skb,
6681 const struct nlmsghdr *nlh,
6682 const struct nlattr * const nla[],
6683 struct netlink_ext_ack *extack)
6684 {
6685 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
6686 struct nft_flowtable_hook flowtable_hook;
6687 const struct nf_flowtable_type *type;
6688 u8 genmask = nft_genmask_next(net);
6689 int family = nfmsg->nfgen_family;
6690 struct nft_flowtable *flowtable;
6691 struct nft_hook *hook, *next;
6692 struct nft_table *table;
6693 struct nft_ctx ctx;
6694 int err;
6695
6696 if (!nla[NFTA_FLOWTABLE_TABLE] ||
6697 !nla[NFTA_FLOWTABLE_NAME] ||
6698 !nla[NFTA_FLOWTABLE_HOOK])
6699 return -EINVAL;
6700
6701 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
6702 genmask);
6703 if (IS_ERR(table)) {
6704 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
6705 return PTR_ERR(table);
6706 }
6707
6708 flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
6709 genmask);
6710 if (IS_ERR(flowtable)) {
6711 err = PTR_ERR(flowtable);
6712 if (err != -ENOENT) {
6713 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
6714 return err;
6715 }
6716 } else {
6717 if (nlh->nlmsg_flags & NLM_F_EXCL) {
6718 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
6719 return -EEXIST;
6720 }
6721
6722 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
6723
6724 return nft_flowtable_update(&ctx, nlh, flowtable);
6725 }
6726
6727 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
6728
6729 flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL);
6730 if (!flowtable)
6731 return -ENOMEM;
6732
6733 flowtable->table = table;
6734 flowtable->handle = nf_tables_alloc_handle(table);
6735 INIT_LIST_HEAD(&flowtable->hook_list);
6736
6737 flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL);
6738 if (!flowtable->name) {
6739 err = -ENOMEM;
6740 goto err1;
6741 }
6742
6743 type = nft_flowtable_type_get(net, family);
6744 if (IS_ERR(type)) {
6745 err = PTR_ERR(type);
6746 goto err2;
6747 }
6748
6749 if (nla[NFTA_FLOWTABLE_FLAGS]) {
6750 flowtable->data.flags =
6751 ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
6752 if (flowtable->data.flags & ~NFT_FLOWTABLE_MASK)
6753 goto err3;
6754 }
6755
6756 write_pnet(&flowtable->data.net, net);
6757 flowtable->data.type = type;
6758 err = type->init(&flowtable->data);
6759 if (err < 0)
6760 goto err3;
6761
6762 err = nft_flowtable_parse_hook(&ctx, nla[NFTA_FLOWTABLE_HOOK],
6763 &flowtable_hook, flowtable, true);
6764 if (err < 0)
6765 goto err4;
6766
6767 list_splice(&flowtable_hook.list, &flowtable->hook_list);
6768 flowtable->data.priority = flowtable_hook.priority;
6769 flowtable->hooknum = flowtable_hook.num;
6770
6771 err = nft_register_flowtable_net_hooks(ctx.net, table,
6772 &flowtable->hook_list,
6773 flowtable);
6774 if (err < 0) {
6775 nft_flowtable_hooks_destroy(&flowtable->hook_list);
6776 goto err4;
6777 }
6778
6779 err = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
6780 if (err < 0)
6781 goto err5;
6782
6783 list_add_tail_rcu(&flowtable->list, &table->flowtables);
6784 table->use++;
6785
6786 return 0;
6787 err5:
6788 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
6789 nft_unregister_flowtable_hook(net, flowtable, hook);
6790 list_del_rcu(&hook->list);
6791 kfree_rcu(hook, rcu);
6792 }
6793 err4:
6794 flowtable->data.type->free(&flowtable->data);
6795 err3:
6796 module_put(type->owner);
6797 err2:
6798 kfree(flowtable->name);
6799 err1:
6800 kfree(flowtable);
6801 return err;
6802 }
6803
6804 static void nft_flowtable_hook_release(struct nft_flowtable_hook *flowtable_hook)
6805 {
6806 struct nft_hook *this, *next;
6807
6808 list_for_each_entry_safe(this, next, &flowtable_hook->list, list) {
6809 list_del(&this->list);
6810 kfree(this);
6811 }
6812 }
6813
6814 static int nft_delflowtable_hook(struct nft_ctx *ctx,
6815 struct nft_flowtable *flowtable)
6816 {
6817 const struct nlattr * const *nla = ctx->nla;
6818 struct nft_flowtable_hook flowtable_hook;
6819 struct nft_hook *this, *hook;
6820 struct nft_trans *trans;
6821 int err;
6822
6823 err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
6824 &flowtable_hook, flowtable, false);
6825 if (err < 0)
6826 return err;
6827
6828 list_for_each_entry(this, &flowtable_hook.list, list) {
6829 hook = nft_hook_list_find(&flowtable->hook_list, this);
6830 if (!hook) {
6831 err = -ENOENT;
6832 goto err_flowtable_del_hook;
6833 }
6834 hook->inactive = true;
6835 }
6836
6837 trans = nft_trans_alloc(ctx, NFT_MSG_DELFLOWTABLE,
6838 sizeof(struct nft_trans_flowtable));
6839 if (!trans) {
6840 err = -ENOMEM;
6841 goto err_flowtable_del_hook;
6842 }
6843
6844 nft_trans_flowtable(trans) = flowtable;
6845 nft_trans_flowtable_update(trans) = true;
6846 INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
6847 nft_flowtable_hook_release(&flowtable_hook);
6848
6849 list_add_tail(&trans->list, &ctx->net->nft.commit_list);
6850
6851 return 0;
6852
6853 err_flowtable_del_hook:
6854 list_for_each_entry(this, &flowtable_hook.list, list) {
6855 hook = nft_hook_list_find(&flowtable->hook_list, this);
6856 if (!hook)
6857 break;
6858
6859 hook->inactive = false;
6860 }
6861 nft_flowtable_hook_release(&flowtable_hook);
6862
6863 return err;
6864 }
6865
6866 static int nf_tables_delflowtable(struct net *net, struct sock *nlsk,
6867 struct sk_buff *skb,
6868 const struct nlmsghdr *nlh,
6869 const struct nlattr * const nla[],
6870 struct netlink_ext_ack *extack)
6871 {
6872 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
6873 u8 genmask = nft_genmask_next(net);
6874 int family = nfmsg->nfgen_family;
6875 struct nft_flowtable *flowtable;
6876 const struct nlattr *attr;
6877 struct nft_table *table;
6878 struct nft_ctx ctx;
6879
6880 if (!nla[NFTA_FLOWTABLE_TABLE] ||
6881 (!nla[NFTA_FLOWTABLE_NAME] &&
6882 !nla[NFTA_FLOWTABLE_HANDLE]))
6883 return -EINVAL;
6884
6885 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
6886 genmask);
6887 if (IS_ERR(table)) {
6888 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
6889 return PTR_ERR(table);
6890 }
6891
6892 if (nla[NFTA_FLOWTABLE_HANDLE]) {
6893 attr = nla[NFTA_FLOWTABLE_HANDLE];
6894 flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
6895 } else {
6896 attr = nla[NFTA_FLOWTABLE_NAME];
6897 flowtable = nft_flowtable_lookup(table, attr, genmask);
6898 }
6899
6900 if (IS_ERR(flowtable)) {
6901 NL_SET_BAD_ATTR(extack, attr);
6902 return PTR_ERR(flowtable);
6903 }
6904
6905 nft_ctx_init(&ctx, net, skb, nlh, family, table, NULL, nla);
6906
6907 if (nla[NFTA_FLOWTABLE_HOOK])
6908 return nft_delflowtable_hook(&ctx, flowtable);
6909
6910 if (flowtable->use > 0) {
6911 NL_SET_BAD_ATTR(extack, attr);
6912 return -EBUSY;
6913 }
6914
6915 return nft_delflowtable(&ctx, flowtable);
6916 }
6917
6918 static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
6919 u32 portid, u32 seq, int event,
6920 u32 flags, int family,
6921 struct nft_flowtable *flowtable,
6922 struct list_head *hook_list)
6923 {
6924 struct nlattr *nest, *nest_devs;
6925 struct nfgenmsg *nfmsg;
6926 struct nft_hook *hook;
6927 struct nlmsghdr *nlh;
6928
6929 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
6930 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), flags);
6931 if (nlh == NULL)
6932 goto nla_put_failure;
6933
6934 nfmsg = nlmsg_data(nlh);
6935 nfmsg->nfgen_family = family;
6936 nfmsg->version = NFNETLINK_V0;
6937 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
6938
6939 if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
6940 nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
6941 nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
6942 nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
6943 NFTA_FLOWTABLE_PAD) ||
6944 nla_put_be32(skb, NFTA_FLOWTABLE_FLAGS, htonl(flowtable->data.flags)))
6945 goto nla_put_failure;
6946
6947 nest = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK);
6948 if (!nest)
6949 goto nla_put_failure;
6950 if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
6951 nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->data.priority)))
6952 goto nla_put_failure;
6953
6954 nest_devs = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK_DEVS);
6955 if (!nest_devs)
6956 goto nla_put_failure;
6957
6958 list_for_each_entry_rcu(hook, hook_list, list) {
6959 if (nla_put_string(skb, NFTA_DEVICE_NAME, hook->ops.dev->name))
6960 goto nla_put_failure;
6961 }
6962 nla_nest_end(skb, nest_devs);
6963 nla_nest_end(skb, nest);
6964
6965 nlmsg_end(skb, nlh);
6966 return 0;
6967
6968 nla_put_failure:
6969 nlmsg_trim(skb, nlh);
6970 return -1;
6971 }
6972
6973 struct nft_flowtable_filter {
6974 char *table;
6975 };
6976
6977 static int nf_tables_dump_flowtable(struct sk_buff *skb,
6978 struct netlink_callback *cb)
6979 {
6980 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
6981 struct nft_flowtable_filter *filter = cb->data;
6982 unsigned int idx = 0, s_idx = cb->args[0];
6983 struct net *net = sock_net(skb->sk);
6984 int family = nfmsg->nfgen_family;
6985 struct nft_flowtable *flowtable;
6986 const struct nft_table *table;
6987
6988 rcu_read_lock();
6989 cb->seq = net->nft.base_seq;
6990
6991 list_for_each_entry_rcu(table, &net->nft.tables, list) {
6992 if (family != NFPROTO_UNSPEC && family != table->family)
6993 continue;
6994
6995 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
6996 if (!nft_is_active(net, flowtable))
6997 goto cont;
6998 if (idx < s_idx)
6999 goto cont;
7000 if (idx > s_idx)
7001 memset(&cb->args[1], 0,
7002 sizeof(cb->args) - sizeof(cb->args[0]));
7003 if (filter && filter->table &&
7004 strcmp(filter->table, table->name))
7005 goto cont;
7006
7007 if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
7008 cb->nlh->nlmsg_seq,
7009 NFT_MSG_NEWFLOWTABLE,
7010 NLM_F_MULTI | NLM_F_APPEND,
7011 table->family,
7012 flowtable,
7013 &flowtable->hook_list) < 0)
7014 goto done;
7015
7016 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
7017 cont:
7018 idx++;
7019 }
7020 }
7021 done:
7022 rcu_read_unlock();
7023
7024 cb->args[0] = idx;
7025 return skb->len;
7026 }
7027
7028 static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
7029 {
7030 const struct nlattr * const *nla = cb->data;
7031 struct nft_flowtable_filter *filter = NULL;
7032
7033 if (nla[NFTA_FLOWTABLE_TABLE]) {
7034 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
7035 if (!filter)
7036 return -ENOMEM;
7037
7038 filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
7039 GFP_ATOMIC);
7040 if (!filter->table) {
7041 kfree(filter);
7042 return -ENOMEM;
7043 }
7044 }
7045
7046 cb->data = filter;
7047 return 0;
7048 }
7049
7050 static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
7051 {
7052 struct nft_flowtable_filter *filter = cb->data;
7053
7054 if (!filter)
7055 return 0;
7056
7057 kfree(filter->table);
7058 kfree(filter);
7059
7060 return 0;
7061 }
7062
7063 /* called with rcu_read_lock held */
7064 static int nf_tables_getflowtable(struct net *net, struct sock *nlsk,
7065 struct sk_buff *skb,
7066 const struct nlmsghdr *nlh,
7067 const struct nlattr * const nla[],
7068 struct netlink_ext_ack *extack)
7069 {
7070 const struct nfgenmsg *nfmsg = nlmsg_data(nlh);
7071 u8 genmask = nft_genmask_cur(net);
7072 int family = nfmsg->nfgen_family;
7073 struct nft_flowtable *flowtable;
7074 const struct nft_table *table;
7075 struct sk_buff *skb2;
7076 int err;
7077
7078 if (nlh->nlmsg_flags & NLM_F_DUMP) {
7079 struct netlink_dump_control c = {
7080 .start = nf_tables_dump_flowtable_start,
7081 .dump = nf_tables_dump_flowtable,
7082 .done = nf_tables_dump_flowtable_done,
7083 .module = THIS_MODULE,
7084 .data = (void *)nla,
7085 };
7086
7087 return nft_netlink_dump_start_rcu(nlsk, skb, nlh, &c);
7088 }
7089
7090 if (!nla[NFTA_FLOWTABLE_NAME])
7091 return -EINVAL;
7092
7093 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7094 genmask);
7095 if (IS_ERR(table))
7096 return PTR_ERR(table);
7097
7098 flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7099 genmask);
7100 if (IS_ERR(flowtable))
7101 return PTR_ERR(flowtable);
7102
7103 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7104 if (!skb2)
7105 return -ENOMEM;
7106
7107 err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
7108 nlh->nlmsg_seq,
7109 NFT_MSG_NEWFLOWTABLE, 0, family,
7110 flowtable, &flowtable->hook_list);
7111 if (err < 0)
7112 goto err_fill_flowtable_info;
7113
7114 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7115
7116 err_fill_flowtable_info:
7117 kfree_skb(skb2);
7118 return err;
7119 }
7120
7121 static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
7122 struct nft_flowtable *flowtable,
7123 struct list_head *hook_list,
7124 int event)
7125 {
7126 struct sk_buff *skb;
7127 int err;
7128 char *buf = kasprintf(GFP_KERNEL, "%s:%llu;%s:%llu",
7129 flowtable->table->name, flowtable->table->handle,
7130 flowtable->name, flowtable->handle);
7131
7132 audit_log_nfcfg(buf,
7133 ctx->family,
7134 flowtable->hooknum,
7135 event == NFT_MSG_NEWFLOWTABLE ?
7136 AUDIT_NFT_OP_FLOWTABLE_REGISTER :
7137 AUDIT_NFT_OP_FLOWTABLE_UNREGISTER,
7138 GFP_KERNEL);
7139 kfree(buf);
7140
7141 if (!ctx->report &&
7142 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
7143 return;
7144
7145 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
7146 if (skb == NULL)
7147 goto err;
7148
7149 err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
7150 ctx->seq, event, 0,
7151 ctx->family, flowtable, hook_list);
7152 if (err < 0) {
7153 kfree_skb(skb);
7154 goto err;
7155 }
7156
7157 nft_notify_enqueue(skb, ctx->report, &ctx->net->nft.notify_list);
7158 return;
7159 err:
7160 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
7161 }
7162
7163 static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
7164 {
7165 struct nft_hook *hook, *next;
7166
7167 flowtable->data.type->free(&flowtable->data);
7168 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7169 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7170 FLOW_BLOCK_UNBIND);
7171 list_del_rcu(&hook->list);
7172 kfree(hook);
7173 }
7174 kfree(flowtable->name);
7175 module_put(flowtable->data.type->owner);
7176 kfree(flowtable);
7177 }
7178
7179 static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
7180 u32 portid, u32 seq)
7181 {
7182 struct nlmsghdr *nlh;
7183 struct nfgenmsg *nfmsg;
7184 char buf[TASK_COMM_LEN];
7185 int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
7186
7187 nlh = nlmsg_put(skb, portid, seq, event, sizeof(struct nfgenmsg), 0);
7188 if (nlh == NULL)
7189 goto nla_put_failure;
7190
7191 nfmsg = nlmsg_data(nlh);
7192 nfmsg->nfgen_family = AF_UNSPEC;
7193 nfmsg->version = NFNETLINK_V0;
7194 nfmsg->res_id = htons(net->nft.base_seq & 0xffff);
7195
7196 if (nla_put_be32(skb, NFTA_GEN_ID, htonl(net->nft.base_seq)) ||
7197 nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
7198 nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
7199 goto nla_put_failure;
7200
7201 nlmsg_end(skb, nlh);
7202 return 0;
7203
7204 nla_put_failure:
7205 nlmsg_trim(skb, nlh);
7206 return -EMSGSIZE;
7207 }
7208
7209 static void nft_flowtable_event(unsigned long event, struct net_device *dev,
7210 struct nft_flowtable *flowtable)
7211 {
7212 struct nft_hook *hook;
7213
7214 list_for_each_entry(hook, &flowtable->hook_list, list) {
7215 if (hook->ops.dev != dev)
7216 continue;
7217
7218 /* flow_offload_netdev_event() cleans up entries for us. */
7219 nft_unregister_flowtable_hook(dev_net(dev), flowtable, hook);
7220 list_del_rcu(&hook->list);
7221 kfree_rcu(hook, rcu);
7222 break;
7223 }
7224 }
7225
7226 static int nf_tables_flowtable_event(struct notifier_block *this,
7227 unsigned long event, void *ptr)
7228 {
7229 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
7230 struct nft_flowtable *flowtable;
7231 struct nft_table *table;
7232 struct net *net;
7233
7234 if (event != NETDEV_UNREGISTER)
7235 return 0;
7236
7237 net = dev_net(dev);
7238 mutex_lock(&net->nft.commit_mutex);
7239 list_for_each_entry(table, &net->nft.tables, list) {
7240 list_for_each_entry(flowtable, &table->flowtables, list) {
7241 nft_flowtable_event(event, dev, flowtable);
7242 }
7243 }
7244 mutex_unlock(&net->nft.commit_mutex);
7245
7246 return NOTIFY_DONE;
7247 }
7248
7249 static struct notifier_block nf_tables_flowtable_notifier = {
7250 .notifier_call = nf_tables_flowtable_event,
7251 };
7252
7253 static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
7254 int event)
7255 {
7256 struct nlmsghdr *nlh = nlmsg_hdr(skb);
7257 struct sk_buff *skb2;
7258 int err;
7259
7260 audit_log_nfcfg("?:0;?:0", 0, net->nft.base_seq,
7261 AUDIT_NFT_OP_GEN_REGISTER, GFP_KERNEL);
7262
7263 if (!nlmsg_report(nlh) &&
7264 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
7265 return;
7266
7267 skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
7268 if (skb2 == NULL)
7269 goto err;
7270
7271 err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
7272 nlh->nlmsg_seq);
7273 if (err < 0) {
7274 kfree_skb(skb2);
7275 goto err;
7276 }
7277
7278 nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
7279 nlmsg_report(nlh), GFP_KERNEL);
7280 return;
7281 err:
7282 nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
7283 -ENOBUFS);
7284 }
7285
7286 static int nf_tables_getgen(struct net *net, struct sock *nlsk,
7287 struct sk_buff *skb, const struct nlmsghdr *nlh,
7288 const struct nlattr * const nla[],
7289 struct netlink_ext_ack *extack)
7290 {
7291 struct sk_buff *skb2;
7292 int err;
7293
7294 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7295 if (skb2 == NULL)
7296 return -ENOMEM;
7297
7298 err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
7299 nlh->nlmsg_seq);
7300 if (err < 0)
7301 goto err_fill_gen_info;
7302
7303 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7304
7305 err_fill_gen_info:
7306 kfree_skb(skb2);
7307 return err;
7308 }
7309
7310 static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
7311 [NFT_MSG_NEWTABLE] = {
7312 .call_batch = nf_tables_newtable,
7313 .attr_count = NFTA_TABLE_MAX,
7314 .policy = nft_table_policy,
7315 },
7316 [NFT_MSG_GETTABLE] = {
7317 .call_rcu = nf_tables_gettable,
7318 .attr_count = NFTA_TABLE_MAX,
7319 .policy = nft_table_policy,
7320 },
7321 [NFT_MSG_DELTABLE] = {
7322 .call_batch = nf_tables_deltable,
7323 .attr_count = NFTA_TABLE_MAX,
7324 .policy = nft_table_policy,
7325 },
7326 [NFT_MSG_NEWCHAIN] = {
7327 .call_batch = nf_tables_newchain,
7328 .attr_count = NFTA_CHAIN_MAX,
7329 .policy = nft_chain_policy,
7330 },
7331 [NFT_MSG_GETCHAIN] = {
7332 .call_rcu = nf_tables_getchain,
7333 .attr_count = NFTA_CHAIN_MAX,
7334 .policy = nft_chain_policy,
7335 },
7336 [NFT_MSG_DELCHAIN] = {
7337 .call_batch = nf_tables_delchain,
7338 .attr_count = NFTA_CHAIN_MAX,
7339 .policy = nft_chain_policy,
7340 },
7341 [NFT_MSG_NEWRULE] = {
7342 .call_batch = nf_tables_newrule,
7343 .attr_count = NFTA_RULE_MAX,
7344 .policy = nft_rule_policy,
7345 },
7346 [NFT_MSG_GETRULE] = {
7347 .call_rcu = nf_tables_getrule,
7348 .attr_count = NFTA_RULE_MAX,
7349 .policy = nft_rule_policy,
7350 },
7351 [NFT_MSG_DELRULE] = {
7352 .call_batch = nf_tables_delrule,
7353 .attr_count = NFTA_RULE_MAX,
7354 .policy = nft_rule_policy,
7355 },
7356 [NFT_MSG_NEWSET] = {
7357 .call_batch = nf_tables_newset,
7358 .attr_count = NFTA_SET_MAX,
7359 .policy = nft_set_policy,
7360 },
7361 [NFT_MSG_GETSET] = {
7362 .call_rcu = nf_tables_getset,
7363 .attr_count = NFTA_SET_MAX,
7364 .policy = nft_set_policy,
7365 },
7366 [NFT_MSG_DELSET] = {
7367 .call_batch = nf_tables_delset,
7368 .attr_count = NFTA_SET_MAX,
7369 .policy = nft_set_policy,
7370 },
7371 [NFT_MSG_NEWSETELEM] = {
7372 .call_batch = nf_tables_newsetelem,
7373 .attr_count = NFTA_SET_ELEM_LIST_MAX,
7374 .policy = nft_set_elem_list_policy,
7375 },
7376 [NFT_MSG_GETSETELEM] = {
7377 .call_rcu = nf_tables_getsetelem,
7378 .attr_count = NFTA_SET_ELEM_LIST_MAX,
7379 .policy = nft_set_elem_list_policy,
7380 },
7381 [NFT_MSG_DELSETELEM] = {
7382 .call_batch = nf_tables_delsetelem,
7383 .attr_count = NFTA_SET_ELEM_LIST_MAX,
7384 .policy = nft_set_elem_list_policy,
7385 },
7386 [NFT_MSG_GETGEN] = {
7387 .call_rcu = nf_tables_getgen,
7388 },
7389 [NFT_MSG_NEWOBJ] = {
7390 .call_batch = nf_tables_newobj,
7391 .attr_count = NFTA_OBJ_MAX,
7392 .policy = nft_obj_policy,
7393 },
7394 [NFT_MSG_GETOBJ] = {
7395 .call_rcu = nf_tables_getobj,
7396 .attr_count = NFTA_OBJ_MAX,
7397 .policy = nft_obj_policy,
7398 },
7399 [NFT_MSG_DELOBJ] = {
7400 .call_batch = nf_tables_delobj,
7401 .attr_count = NFTA_OBJ_MAX,
7402 .policy = nft_obj_policy,
7403 },
7404 [NFT_MSG_GETOBJ_RESET] = {
7405 .call_rcu = nf_tables_getobj,
7406 .attr_count = NFTA_OBJ_MAX,
7407 .policy = nft_obj_policy,
7408 },
7409 [NFT_MSG_NEWFLOWTABLE] = {
7410 .call_batch = nf_tables_newflowtable,
7411 .attr_count = NFTA_FLOWTABLE_MAX,
7412 .policy = nft_flowtable_policy,
7413 },
7414 [NFT_MSG_GETFLOWTABLE] = {
7415 .call_rcu = nf_tables_getflowtable,
7416 .attr_count = NFTA_FLOWTABLE_MAX,
7417 .policy = nft_flowtable_policy,
7418 },
7419 [NFT_MSG_DELFLOWTABLE] = {
7420 .call_batch = nf_tables_delflowtable,
7421 .attr_count = NFTA_FLOWTABLE_MAX,
7422 .policy = nft_flowtable_policy,
7423 },
7424 };
7425
7426 static int nf_tables_validate(struct net *net)
7427 {
7428 struct nft_table *table;
7429
7430 switch (net->nft.validate_state) {
7431 case NFT_VALIDATE_SKIP:
7432 break;
7433 case NFT_VALIDATE_NEED:
7434 nft_validate_state_update(net, NFT_VALIDATE_DO);
7435 fallthrough;
7436 case NFT_VALIDATE_DO:
7437 list_for_each_entry(table, &net->nft.tables, list) {
7438 if (nft_table_validate(net, table) < 0)
7439 return -EAGAIN;
7440 }
7441 break;
7442 }
7443
7444 return 0;
7445 }
7446
7447 /* a drop policy has to be deferred until all rules have been activated,
7448 * otherwise a large ruleset that contains a drop-policy base chain will
7449 * cause all packets to get dropped until the full transaction has been
7450 * processed.
7451 *
7452 * We defer the drop policy until the transaction has been finalized.
7453 */
7454 static void nft_chain_commit_drop_policy(struct nft_trans *trans)
7455 {
7456 struct nft_base_chain *basechain;
7457
7458 if (nft_trans_chain_policy(trans) != NF_DROP)
7459 return;
7460
7461 if (!nft_is_base_chain(trans->ctx.chain))
7462 return;
7463
7464 basechain = nft_base_chain(trans->ctx.chain);
7465 basechain->policy = NF_DROP;
7466 }
7467
7468 static void nft_chain_commit_update(struct nft_trans *trans)
7469 {
7470 struct nft_base_chain *basechain;
7471
7472 if (nft_trans_chain_name(trans)) {
7473 rhltable_remove(&trans->ctx.table->chains_ht,
7474 &trans->ctx.chain->rhlhead,
7475 nft_chain_ht_params);
7476 swap(trans->ctx.chain->name, nft_trans_chain_name(trans));
7477 rhltable_insert_key(&trans->ctx.table->chains_ht,
7478 trans->ctx.chain->name,
7479 &trans->ctx.chain->rhlhead,
7480 nft_chain_ht_params);
7481 }
7482
7483 if (!nft_is_base_chain(trans->ctx.chain))
7484 return;
7485
7486 nft_chain_stats_replace(trans);
7487
7488 basechain = nft_base_chain(trans->ctx.chain);
7489
7490 switch (nft_trans_chain_policy(trans)) {
7491 case NF_DROP:
7492 case NF_ACCEPT:
7493 basechain->policy = nft_trans_chain_policy(trans);
7494 break;
7495 }
7496 }
7497
7498 static void nft_obj_commit_update(struct nft_trans *trans)
7499 {
7500 struct nft_object *newobj;
7501 struct nft_object *obj;
7502
7503 obj = nft_trans_obj(trans);
7504 newobj = nft_trans_obj_newobj(trans);
7505
7506 if (obj->ops->update)
7507 obj->ops->update(obj, newobj);
7508
7509 kfree(newobj);
7510 }
7511
7512 static void nft_commit_release(struct nft_trans *trans)
7513 {
7514 switch (trans->msg_type) {
7515 case NFT_MSG_DELTABLE:
7516 nf_tables_table_destroy(&trans->ctx);
7517 break;
7518 case NFT_MSG_NEWCHAIN:
7519 free_percpu(nft_trans_chain_stats(trans));
7520 kfree(nft_trans_chain_name(trans));
7521 break;
7522 case NFT_MSG_DELCHAIN:
7523 nf_tables_chain_destroy(&trans->ctx);
7524 break;
7525 case NFT_MSG_DELRULE:
7526 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
7527 break;
7528 case NFT_MSG_DELSET:
7529 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
7530 break;
7531 case NFT_MSG_DELSETELEM:
7532 nf_tables_set_elem_destroy(&trans->ctx,
7533 nft_trans_elem_set(trans),
7534 nft_trans_elem(trans).priv);
7535 break;
7536 case NFT_MSG_DELOBJ:
7537 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
7538 break;
7539 case NFT_MSG_DELFLOWTABLE:
7540 if (nft_trans_flowtable_update(trans))
7541 nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
7542 else
7543 nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
7544 break;
7545 }
7546
7547 if (trans->put_net)
7548 put_net(trans->ctx.net);
7549
7550 kfree(trans);
7551 }
7552
7553 static void nf_tables_trans_destroy_work(struct work_struct *w)
7554 {
7555 struct nft_trans *trans, *next;
7556 LIST_HEAD(head);
7557
7558 spin_lock(&nf_tables_destroy_list_lock);
7559 list_splice_init(&nf_tables_destroy_list, &head);
7560 spin_unlock(&nf_tables_destroy_list_lock);
7561
7562 if (list_empty(&head))
7563 return;
7564
7565 synchronize_rcu();
7566
7567 list_for_each_entry_safe(trans, next, &head, list) {
7568 list_del(&trans->list);
7569 nft_commit_release(trans);
7570 }
7571 }
7572
7573 void nf_tables_trans_destroy_flush_work(void)
7574 {
7575 flush_work(&trans_destroy_work);
7576 }
7577 EXPORT_SYMBOL_GPL(nf_tables_trans_destroy_flush_work);
7578
7579 static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
7580 {
7581 struct nft_rule *rule;
7582 unsigned int alloc = 0;
7583 int i;
7584
7585 /* already handled or inactive chain? */
7586 if (chain->rules_next || !nft_is_active_next(net, chain))
7587 return 0;
7588
7589 rule = list_entry(&chain->rules, struct nft_rule, list);
7590 i = 0;
7591
7592 list_for_each_entry_continue(rule, &chain->rules, list) {
7593 if (nft_is_active_next(net, rule))
7594 alloc++;
7595 }
7596
7597 chain->rules_next = nf_tables_chain_alloc_rules(chain, alloc);
7598 if (!chain->rules_next)
7599 return -ENOMEM;
7600
7601 list_for_each_entry_continue(rule, &chain->rules, list) {
7602 if (nft_is_active_next(net, rule))
7603 chain->rules_next[i++] = rule;
7604 }
7605
7606 chain->rules_next[i] = NULL;
7607 return 0;
7608 }
7609
7610 static void nf_tables_commit_chain_prepare_cancel(struct net *net)
7611 {
7612 struct nft_trans *trans, *next;
7613
7614 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
7615 struct nft_chain *chain = trans->ctx.chain;
7616
7617 if (trans->msg_type == NFT_MSG_NEWRULE ||
7618 trans->msg_type == NFT_MSG_DELRULE) {
7619 kvfree(chain->rules_next);
7620 chain->rules_next = NULL;
7621 }
7622 }
7623 }
7624
7625 static void __nf_tables_commit_chain_free_rules_old(struct rcu_head *h)
7626 {
7627 struct nft_rules_old *o = container_of(h, struct nft_rules_old, h);
7628
7629 kvfree(o->start);
7630 }
7631
7632 static void nf_tables_commit_chain_free_rules_old(struct nft_rule **rules)
7633 {
7634 struct nft_rule **r = rules;
7635 struct nft_rules_old *old;
7636
7637 while (*r)
7638 r++;
7639
7640 r++; /* rcu_head is after end marker */
7641 old = (void *) r;
7642 old->start = rules;
7643
7644 call_rcu(&old->h, __nf_tables_commit_chain_free_rules_old);
7645 }
7646
7647 static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
7648 {
7649 struct nft_rule **g0, **g1;
7650 bool next_genbit;
7651
7652 next_genbit = nft_gencursor_next(net);
7653
7654 g0 = rcu_dereference_protected(chain->rules_gen_0,
7655 lockdep_commit_lock_is_held(net));
7656 g1 = rcu_dereference_protected(chain->rules_gen_1,
7657 lockdep_commit_lock_is_held(net));
7658
7659 /* No changes to this chain? */
7660 if (chain->rules_next == NULL) {
7661 /* chain had no change in last or next generation */
7662 if (g0 == g1)
7663 return;
7664 /*
7665 * chain had no change in this generation; make sure next
7666 * one uses same rules as current generation.
7667 */
7668 if (next_genbit) {
7669 rcu_assign_pointer(chain->rules_gen_1, g0);
7670 nf_tables_commit_chain_free_rules_old(g1);
7671 } else {
7672 rcu_assign_pointer(chain->rules_gen_0, g1);
7673 nf_tables_commit_chain_free_rules_old(g0);
7674 }
7675
7676 return;
7677 }
7678
7679 if (next_genbit)
7680 rcu_assign_pointer(chain->rules_gen_1, chain->rules_next);
7681 else
7682 rcu_assign_pointer(chain->rules_gen_0, chain->rules_next);
7683
7684 chain->rules_next = NULL;
7685
7686 if (g0 == g1)
7687 return;
7688
7689 if (next_genbit)
7690 nf_tables_commit_chain_free_rules_old(g1);
7691 else
7692 nf_tables_commit_chain_free_rules_old(g0);
7693 }
7694
7695 static void nft_obj_del(struct nft_object *obj)
7696 {
7697 rhltable_remove(&nft_objname_ht, &obj->rhlhead, nft_objname_ht_params);
7698 list_del_rcu(&obj->list);
7699 }
7700
7701 void nft_chain_del(struct nft_chain *chain)
7702 {
7703 struct nft_table *table = chain->table;
7704
7705 WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
7706 nft_chain_ht_params));
7707 list_del_rcu(&chain->list);
7708 }
7709
7710 static void nft_flowtable_hooks_del(struct nft_flowtable *flowtable,
7711 struct list_head *hook_list)
7712 {
7713 struct nft_hook *hook, *next;
7714
7715 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7716 if (hook->inactive)
7717 list_move(&hook->list, hook_list);
7718 }
7719 }
7720
7721 static void nf_tables_module_autoload_cleanup(struct net *net)
7722 {
7723 struct nft_module_request *req, *next;
7724
7725 WARN_ON_ONCE(!list_empty(&net->nft.commit_list));
7726 list_for_each_entry_safe(req, next, &net->nft.module_list, list) {
7727 WARN_ON_ONCE(!req->done);
7728 list_del(&req->list);
7729 kfree(req);
7730 }
7731 }
7732
7733 static void nf_tables_commit_release(struct net *net)
7734 {
7735 struct nft_trans *trans;
7736
7737 /* all side effects have to be made visible.
7738 * For example, if a chain named 'foo' has been deleted, a
7739 * new transaction must not find it anymore.
7740 *
7741 * Memory reclaim happens asynchronously from work queue
7742 * to prevent expensive synchronize_rcu() in commit phase.
7743 */
7744 if (list_empty(&net->nft.commit_list)) {
7745 nf_tables_module_autoload_cleanup(net);
7746 mutex_unlock(&net->nft.commit_mutex);
7747 return;
7748 }
7749
7750 trans = list_last_entry(&net->nft.commit_list,
7751 struct nft_trans, list);
7752 get_net(trans->ctx.net);
7753 WARN_ON_ONCE(trans->put_net);
7754
7755 trans->put_net = true;
7756 spin_lock(&nf_tables_destroy_list_lock);
7757 list_splice_tail_init(&net->nft.commit_list, &nf_tables_destroy_list);
7758 spin_unlock(&nf_tables_destroy_list_lock);
7759
7760 nf_tables_module_autoload_cleanup(net);
7761 schedule_work(&trans_destroy_work);
7762
7763 mutex_unlock(&net->nft.commit_mutex);
7764 }
7765
7766 static void nft_commit_notify(struct net *net, u32 portid)
7767 {
7768 struct sk_buff *batch_skb = NULL, *nskb, *skb;
7769 unsigned char *data;
7770 int len;
7771
7772 list_for_each_entry_safe(skb, nskb, &net->nft.notify_list, list) {
7773 if (!batch_skb) {
7774 new_batch:
7775 batch_skb = skb;
7776 len = NLMSG_GOODSIZE - skb->len;
7777 list_del(&skb->list);
7778 continue;
7779 }
7780 len -= skb->len;
7781 if (len > 0 && NFT_CB(skb).report == NFT_CB(batch_skb).report) {
7782 data = skb_put(batch_skb, skb->len);
7783 memcpy(data, skb->data, skb->len);
7784 list_del(&skb->list);
7785 kfree_skb(skb);
7786 continue;
7787 }
7788 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
7789 NFT_CB(batch_skb).report, GFP_KERNEL);
7790 goto new_batch;
7791 }
7792
7793 if (batch_skb) {
7794 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
7795 NFT_CB(batch_skb).report, GFP_KERNEL);
7796 }
7797
7798 WARN_ON_ONCE(!list_empty(&net->nft.notify_list));
7799 }
7800
7801 static int nf_tables_commit(struct net *net, struct sk_buff *skb)
7802 {
7803 struct nft_trans *trans, *next;
7804 struct nft_trans_elem *te;
7805 struct nft_chain *chain;
7806 struct nft_table *table;
7807 int err;
7808
7809 if (list_empty(&net->nft.commit_list)) {
7810 mutex_unlock(&net->nft.commit_mutex);
7811 return 0;
7812 }
7813
7814 /* 0. Validate ruleset, otherwise roll back for error reporting. */
7815 if (nf_tables_validate(net) < 0)
7816 return -EAGAIN;
7817
7818 err = nft_flow_rule_offload_commit(net);
7819 if (err < 0)
7820 return err;
7821
7822 /* 1. Allocate space for next generation rules_gen_X[] */
7823 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
7824 int ret;
7825
7826 if (trans->msg_type == NFT_MSG_NEWRULE ||
7827 trans->msg_type == NFT_MSG_DELRULE) {
7828 chain = trans->ctx.chain;
7829
7830 ret = nf_tables_commit_chain_prepare(net, chain);
7831 if (ret < 0) {
7832 nf_tables_commit_chain_prepare_cancel(net);
7833 return ret;
7834 }
7835 }
7836 }
7837
7838 /* step 2. Make rules_gen_X visible to packet path */
7839 list_for_each_entry(table, &net->nft.tables, list) {
7840 list_for_each_entry(chain, &table->chains, list)
7841 nf_tables_commit_chain(net, chain);
7842 }
7843
7844 /*
7845 * Bump generation counter, invalidate any dump in progress.
7846 * Cannot fail after this point.
7847 */
7848 while (++net->nft.base_seq == 0);
7849
7850 /* step 3. Start new generation, rules_gen_X now in use. */
7851 net->nft.gencursor = nft_gencursor_next(net);
7852
7853 list_for_each_entry_safe(trans, next, &net->nft.commit_list, list) {
7854 switch (trans->msg_type) {
7855 case NFT_MSG_NEWTABLE:
7856 if (nft_trans_table_update(trans)) {
7857 if (!nft_trans_table_enable(trans)) {
7858 nf_tables_table_disable(net,
7859 trans->ctx.table);
7860 trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
7861 }
7862 } else {
7863 nft_clear(net, trans->ctx.table);
7864 }
7865 nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
7866 nft_trans_destroy(trans);
7867 break;
7868 case NFT_MSG_DELTABLE:
7869 list_del_rcu(&trans->ctx.table->list);
7870 nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
7871 break;
7872 case NFT_MSG_NEWCHAIN:
7873 if (nft_trans_chain_update(trans)) {
7874 nft_chain_commit_update(trans);
7875 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
7876 /* trans destroyed after rcu grace period */
7877 } else {
7878 nft_chain_commit_drop_policy(trans);
7879 nft_clear(net, trans->ctx.chain);
7880 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
7881 nft_trans_destroy(trans);
7882 }
7883 break;
7884 case NFT_MSG_DELCHAIN:
7885 nft_chain_del(trans->ctx.chain);
7886 nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
7887 nf_tables_unregister_hook(trans->ctx.net,
7888 trans->ctx.table,
7889 trans->ctx.chain);
7890 break;
7891 case NFT_MSG_NEWRULE:
7892 nft_clear(trans->ctx.net, nft_trans_rule(trans));
7893 nf_tables_rule_notify(&trans->ctx,
7894 nft_trans_rule(trans),
7895 NFT_MSG_NEWRULE);
7896 nft_trans_destroy(trans);
7897 break;
7898 case NFT_MSG_DELRULE:
7899 list_del_rcu(&nft_trans_rule(trans)->list);
7900 nf_tables_rule_notify(&trans->ctx,
7901 nft_trans_rule(trans),
7902 NFT_MSG_DELRULE);
7903 nft_rule_expr_deactivate(&trans->ctx,
7904 nft_trans_rule(trans),
7905 NFT_TRANS_COMMIT);
7906 break;
7907 case NFT_MSG_NEWSET:
7908 nft_clear(net, nft_trans_set(trans));
7909 /* This avoids hitting -EBUSY when deleting the table
7910 * from the transaction.
7911 */
7912 if (nft_set_is_anonymous(nft_trans_set(trans)) &&
7913 !list_empty(&nft_trans_set(trans)->bindings))
7914 trans->ctx.table->use--;
7915
7916 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
7917 NFT_MSG_NEWSET, GFP_KERNEL);
7918 nft_trans_destroy(trans);
7919 break;
7920 case NFT_MSG_DELSET:
7921 list_del_rcu(&nft_trans_set(trans)->list);
7922 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
7923 NFT_MSG_DELSET, GFP_KERNEL);
7924 break;
7925 case NFT_MSG_NEWSETELEM:
7926 te = (struct nft_trans_elem *)trans->data;
7927
7928 te->set->ops->activate(net, te->set, &te->elem);
7929 nf_tables_setelem_notify(&trans->ctx, te->set,
7930 &te->elem,
7931 NFT_MSG_NEWSETELEM, 0);
7932 nft_trans_destroy(trans);
7933 break;
7934 case NFT_MSG_DELSETELEM:
7935 te = (struct nft_trans_elem *)trans->data;
7936
7937 nf_tables_setelem_notify(&trans->ctx, te->set,
7938 &te->elem,
7939 NFT_MSG_DELSETELEM, 0);
7940 te->set->ops->remove(net, te->set, &te->elem);
7941 atomic_dec(&te->set->nelems);
7942 te->set->ndeact--;
7943 break;
7944 case NFT_MSG_NEWOBJ:
7945 if (nft_trans_obj_update(trans)) {
7946 nft_obj_commit_update(trans);
7947 nf_tables_obj_notify(&trans->ctx,
7948 nft_trans_obj(trans),
7949 NFT_MSG_NEWOBJ);
7950 } else {
7951 nft_clear(net, nft_trans_obj(trans));
7952 nf_tables_obj_notify(&trans->ctx,
7953 nft_trans_obj(trans),
7954 NFT_MSG_NEWOBJ);
7955 nft_trans_destroy(trans);
7956 }
7957 break;
7958 case NFT_MSG_DELOBJ:
7959 nft_obj_del(nft_trans_obj(trans));
7960 nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
7961 NFT_MSG_DELOBJ);
7962 break;
7963 case NFT_MSG_NEWFLOWTABLE:
7964 if (nft_trans_flowtable_update(trans)) {
7965 nf_tables_flowtable_notify(&trans->ctx,
7966 nft_trans_flowtable(trans),
7967 &nft_trans_flowtable_hooks(trans),
7968 NFT_MSG_NEWFLOWTABLE);
7969 list_splice(&nft_trans_flowtable_hooks(trans),
7970 &nft_trans_flowtable(trans)->hook_list);
7971 } else {
7972 nft_clear(net, nft_trans_flowtable(trans));
7973 nf_tables_flowtable_notify(&trans->ctx,
7974 nft_trans_flowtable(trans),
7975 &nft_trans_flowtable(trans)->hook_list,
7976 NFT_MSG_NEWFLOWTABLE);
7977 }
7978 nft_trans_destroy(trans);
7979 break;
7980 case NFT_MSG_DELFLOWTABLE:
7981 if (nft_trans_flowtable_update(trans)) {
7982 nft_flowtable_hooks_del(nft_trans_flowtable(trans),
7983 &nft_trans_flowtable_hooks(trans));
7984 nf_tables_flowtable_notify(&trans->ctx,
7985 nft_trans_flowtable(trans),
7986 &nft_trans_flowtable_hooks(trans),
7987 NFT_MSG_DELFLOWTABLE);
7988 nft_unregister_flowtable_net_hooks(net,
7989 &nft_trans_flowtable_hooks(trans));
7990 } else {
7991 list_del_rcu(&nft_trans_flowtable(trans)->list);
7992 nf_tables_flowtable_notify(&trans->ctx,
7993 nft_trans_flowtable(trans),
7994 &nft_trans_flowtable(trans)->hook_list,
7995 NFT_MSG_DELFLOWTABLE);
7996 nft_unregister_flowtable_net_hooks(net,
7997 &nft_trans_flowtable(trans)->hook_list);
7998 }
7999 break;
8000 }
8001 }
8002
8003 nft_commit_notify(net, NETLINK_CB(skb).portid);
8004 nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
8005 nf_tables_commit_release(net);
8006
8007 return 0;
8008 }
8009
8010 static void nf_tables_module_autoload(struct net *net)
8011 {
8012 struct nft_module_request *req, *next;
8013 LIST_HEAD(module_list);
8014
8015 list_splice_init(&net->nft.module_list, &module_list);
8016 mutex_unlock(&net->nft.commit_mutex);
8017 list_for_each_entry_safe(req, next, &module_list, list) {
8018 request_module("%s", req->module);
8019 req->done = true;
8020 }
8021 mutex_lock(&net->nft.commit_mutex);
8022 list_splice(&module_list, &net->nft.module_list);
8023 }
8024
8025 static void nf_tables_abort_release(struct nft_trans *trans)
8026 {
8027 switch (trans->msg_type) {
8028 case NFT_MSG_NEWTABLE:
8029 nf_tables_table_destroy(&trans->ctx);
8030 break;
8031 case NFT_MSG_NEWCHAIN:
8032 nf_tables_chain_destroy(&trans->ctx);
8033 break;
8034 case NFT_MSG_NEWRULE:
8035 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8036 break;
8037 case NFT_MSG_NEWSET:
8038 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8039 break;
8040 case NFT_MSG_NEWSETELEM:
8041 nft_set_elem_destroy(nft_trans_elem_set(trans),
8042 nft_trans_elem(trans).priv, true);
8043 break;
8044 case NFT_MSG_NEWOBJ:
8045 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8046 break;
8047 case NFT_MSG_NEWFLOWTABLE:
8048 if (nft_trans_flowtable_update(trans))
8049 nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8050 else
8051 nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8052 break;
8053 }
8054 kfree(trans);
8055 }
8056
8057 static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)
8058 {
8059 struct nft_trans *trans, *next;
8060 struct nft_trans_elem *te;
8061 struct nft_hook *hook;
8062
8063 if (action == NFNL_ABORT_VALIDATE &&
8064 nf_tables_validate(net) < 0)
8065 return -EAGAIN;
8066
8067 list_for_each_entry_safe_reverse(trans, next, &net->nft.commit_list,
8068 list) {
8069 switch (trans->msg_type) {
8070 case NFT_MSG_NEWTABLE:
8071 if (nft_trans_table_update(trans)) {
8072 if (nft_trans_table_enable(trans)) {
8073 nf_tables_table_disable(net,
8074 trans->ctx.table);
8075 trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
8076 }
8077 nft_trans_destroy(trans);
8078 } else {
8079 list_del_rcu(&trans->ctx.table->list);
8080 }
8081 break;
8082 case NFT_MSG_DELTABLE:
8083 nft_clear(trans->ctx.net, trans->ctx.table);
8084 nft_trans_destroy(trans);
8085 break;
8086 case NFT_MSG_NEWCHAIN:
8087 if (nft_trans_chain_update(trans)) {
8088 free_percpu(nft_trans_chain_stats(trans));
8089 kfree(nft_trans_chain_name(trans));
8090 nft_trans_destroy(trans);
8091 } else {
8092 if (nft_chain_is_bound(trans->ctx.chain)) {
8093 nft_trans_destroy(trans);
8094 break;
8095 }
8096 trans->ctx.table->use--;
8097 nft_chain_del(trans->ctx.chain);
8098 nf_tables_unregister_hook(trans->ctx.net,
8099 trans->ctx.table,
8100 trans->ctx.chain);
8101 }
8102 break;
8103 case NFT_MSG_DELCHAIN:
8104 trans->ctx.table->use++;
8105 nft_clear(trans->ctx.net, trans->ctx.chain);
8106 nft_trans_destroy(trans);
8107 break;
8108 case NFT_MSG_NEWRULE:
8109 trans->ctx.chain->use--;
8110 list_del_rcu(&nft_trans_rule(trans)->list);
8111 nft_rule_expr_deactivate(&trans->ctx,
8112 nft_trans_rule(trans),
8113 NFT_TRANS_ABORT);
8114 break;
8115 case NFT_MSG_DELRULE:
8116 trans->ctx.chain->use++;
8117 nft_clear(trans->ctx.net, nft_trans_rule(trans));
8118 nft_rule_expr_activate(&trans->ctx, nft_trans_rule(trans));
8119 nft_trans_destroy(trans);
8120 break;
8121 case NFT_MSG_NEWSET:
8122 trans->ctx.table->use--;
8123 if (nft_trans_set_bound(trans)) {
8124 nft_trans_destroy(trans);
8125 break;
8126 }
8127 list_del_rcu(&nft_trans_set(trans)->list);
8128 break;
8129 case NFT_MSG_DELSET:
8130 trans->ctx.table->use++;
8131 nft_clear(trans->ctx.net, nft_trans_set(trans));
8132 nft_trans_destroy(trans);
8133 break;
8134 case NFT_MSG_NEWSETELEM:
8135 if (nft_trans_elem_set_bound(trans)) {
8136 nft_trans_destroy(trans);
8137 break;
8138 }
8139 te = (struct nft_trans_elem *)trans->data;
8140 te->set->ops->remove(net, te->set, &te->elem);
8141 atomic_dec(&te->set->nelems);
8142 break;
8143 case NFT_MSG_DELSETELEM:
8144 te = (struct nft_trans_elem *)trans->data;
8145
8146 nft_set_elem_activate(net, te->set, &te->elem);
8147 te->set->ops->activate(net, te->set, &te->elem);
8148 te->set->ndeact--;
8149
8150 nft_trans_destroy(trans);
8151 break;
8152 case NFT_MSG_NEWOBJ:
8153 if (nft_trans_obj_update(trans)) {
8154 kfree(nft_trans_obj_newobj(trans));
8155 nft_trans_destroy(trans);
8156 } else {
8157 trans->ctx.table->use--;
8158 nft_obj_del(nft_trans_obj(trans));
8159 }
8160 break;
8161 case NFT_MSG_DELOBJ:
8162 trans->ctx.table->use++;
8163 nft_clear(trans->ctx.net, nft_trans_obj(trans));
8164 nft_trans_destroy(trans);
8165 break;
8166 case NFT_MSG_NEWFLOWTABLE:
8167 if (nft_trans_flowtable_update(trans)) {
8168 nft_unregister_flowtable_net_hooks(net,
8169 &nft_trans_flowtable_hooks(trans));
8170 } else {
8171 trans->ctx.table->use--;
8172 list_del_rcu(&nft_trans_flowtable(trans)->list);
8173 nft_unregister_flowtable_net_hooks(net,
8174 &nft_trans_flowtable(trans)->hook_list);
8175 }
8176 break;
8177 case NFT_MSG_DELFLOWTABLE:
8178 if (nft_trans_flowtable_update(trans)) {
8179 list_for_each_entry(hook, &nft_trans_flowtable(trans)->hook_list, list)
8180 hook->inactive = false;
8181 } else {
8182 trans->ctx.table->use++;
8183 nft_clear(trans->ctx.net, nft_trans_flowtable(trans));
8184 }
8185 nft_trans_destroy(trans);
8186 break;
8187 }
8188 }
8189
8190 synchronize_rcu();
8191
8192 list_for_each_entry_safe_reverse(trans, next,
8193 &net->nft.commit_list, list) {
8194 list_del(&trans->list);
8195 nf_tables_abort_release(trans);
8196 }
8197
8198 if (action == NFNL_ABORT_AUTOLOAD)
8199 nf_tables_module_autoload(net);
8200 else
8201 nf_tables_module_autoload_cleanup(net);
8202
8203 return 0;
8204 }
8205
8206 static void nf_tables_cleanup(struct net *net)
8207 {
8208 nft_validate_state_update(net, NFT_VALIDATE_SKIP);
8209 }
8210
8211 static int nf_tables_abort(struct net *net, struct sk_buff *skb,
8212 enum nfnl_abort_action action)
8213 {
8214 int ret = __nf_tables_abort(net, action);
8215
8216 mutex_unlock(&net->nft.commit_mutex);
8217
8218 return ret;
8219 }
8220
8221 static bool nf_tables_valid_genid(struct net *net, u32 genid)
8222 {
8223 bool genid_ok;
8224
8225 mutex_lock(&net->nft.commit_mutex);
8226
8227 genid_ok = genid == 0 || net->nft.base_seq == genid;
8228 if (!genid_ok)
8229 mutex_unlock(&net->nft.commit_mutex);
8230
8231 /* else, commit mutex has to be released by commit or abort function */
8232 return genid_ok;
8233 }
8234
8235 static const struct nfnetlink_subsystem nf_tables_subsys = {
8236 .name = "nf_tables",
8237 .subsys_id = NFNL_SUBSYS_NFTABLES,
8238 .cb_count = NFT_MSG_MAX,
8239 .cb = nf_tables_cb,
8240 .commit = nf_tables_commit,
8241 .abort = nf_tables_abort,
8242 .cleanup = nf_tables_cleanup,
8243 .valid_genid = nf_tables_valid_genid,
8244 .owner = THIS_MODULE,
8245 };
8246
8247 int nft_chain_validate_dependency(const struct nft_chain *chain,
8248 enum nft_chain_types type)
8249 {
8250 const struct nft_base_chain *basechain;
8251
8252 if (nft_is_base_chain(chain)) {
8253 basechain = nft_base_chain(chain);
8254 if (basechain->type->type != type)
8255 return -EOPNOTSUPP;
8256 }
8257 return 0;
8258 }
8259 EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
8260
8261 int nft_chain_validate_hooks(const struct nft_chain *chain,
8262 unsigned int hook_flags)
8263 {
8264 struct nft_base_chain *basechain;
8265
8266 if (nft_is_base_chain(chain)) {
8267 basechain = nft_base_chain(chain);
8268
8269 if ((1 << basechain->ops.hooknum) & hook_flags)
8270 return 0;
8271
8272 return -EOPNOTSUPP;
8273 }
8274
8275 return 0;
8276 }
8277 EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
8278
8279 /*
8280 * Loop detection - walk through the ruleset beginning at the destination chain
8281 * of a new jump until either the source chain is reached (loop) or all
8282 * reachable chains have been traversed.
8283 *
8284 * The loop check is performed whenever a new jump verdict is added to an
8285 * expression or verdict map or a verdict map is bound to a new chain.
8286 */
8287
8288 static int nf_tables_check_loops(const struct nft_ctx *ctx,
8289 const struct nft_chain *chain);
8290
8291 static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
8292 struct nft_set *set,
8293 const struct nft_set_iter *iter,
8294 struct nft_set_elem *elem)
8295 {
8296 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
8297 const struct nft_data *data;
8298
8299 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
8300 *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
8301 return 0;
8302
8303 data = nft_set_ext_data(ext);
8304 switch (data->verdict.code) {
8305 case NFT_JUMP:
8306 case NFT_GOTO:
8307 return nf_tables_check_loops(ctx, data->verdict.chain);
8308 default:
8309 return 0;
8310 }
8311 }
8312
8313 static int nf_tables_check_loops(const struct nft_ctx *ctx,
8314 const struct nft_chain *chain)
8315 {
8316 const struct nft_rule *rule;
8317 const struct nft_expr *expr, *last;
8318 struct nft_set *set;
8319 struct nft_set_binding *binding;
8320 struct nft_set_iter iter;
8321
8322 if (ctx->chain == chain)
8323 return -ELOOP;
8324
8325 list_for_each_entry(rule, &chain->rules, list) {
8326 nft_rule_for_each_expr(expr, last, rule) {
8327 struct nft_immediate_expr *priv;
8328 const struct nft_data *data;
8329 int err;
8330
8331 if (strcmp(expr->ops->type->name, "immediate"))
8332 continue;
8333
8334 priv = nft_expr_priv(expr);
8335 if (priv->dreg != NFT_REG_VERDICT)
8336 continue;
8337
8338 data = &priv->data;
8339 switch (data->verdict.code) {
8340 case NFT_JUMP:
8341 case NFT_GOTO:
8342 err = nf_tables_check_loops(ctx,
8343 data->verdict.chain);
8344 if (err < 0)
8345 return err;
8346 default:
8347 break;
8348 }
8349 }
8350 }
8351
8352 list_for_each_entry(set, &ctx->table->sets, list) {
8353 if (!nft_is_active_next(ctx->net, set))
8354 continue;
8355 if (!(set->flags & NFT_SET_MAP) ||
8356 set->dtype != NFT_DATA_VERDICT)
8357 continue;
8358
8359 list_for_each_entry(binding, &set->bindings, list) {
8360 if (!(binding->flags & NFT_SET_MAP) ||
8361 binding->chain != chain)
8362 continue;
8363
8364 iter.genmask = nft_genmask_next(ctx->net);
8365 iter.skip = 0;
8366 iter.count = 0;
8367 iter.err = 0;
8368 iter.fn = nf_tables_loop_check_setelem;
8369
8370 set->ops->walk(ctx, set, &iter);
8371 if (iter.err < 0)
8372 return iter.err;
8373 }
8374 }
8375
8376 return 0;
8377 }
8378
8379 /**
8380 * nft_parse_u32_check - fetch u32 attribute and check for maximum value
8381 *
8382 * @attr: netlink attribute to fetch value from
8383 * @max: maximum value to be stored in dest
8384 * @dest: pointer to the variable
8385 *
8386 * Parse, check and store a given u32 netlink attribute into variable.
8387 * This function returns -ERANGE if the value goes over maximum value.
8388 * Otherwise a 0 is returned and the attribute value is stored in the
8389 * destination variable.
8390 */
8391 int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
8392 {
8393 u32 val;
8394
8395 val = ntohl(nla_get_be32(attr));
8396 if (val > max)
8397 return -ERANGE;
8398
8399 *dest = val;
8400 return 0;
8401 }
8402 EXPORT_SYMBOL_GPL(nft_parse_u32_check);
8403
8404 /**
8405 * nft_parse_register - parse a register value from a netlink attribute
8406 *
8407 * @attr: netlink attribute
8408 *
8409 * Parse and translate a register value from a netlink attribute.
8410 * Registers used to be 128 bit wide, these register numbers will be
8411 * mapped to the corresponding 32 bit register numbers.
8412 */
8413 unsigned int nft_parse_register(const struct nlattr *attr)
8414 {
8415 unsigned int reg;
8416
8417 reg = ntohl(nla_get_be32(attr));
8418 switch (reg) {
8419 case NFT_REG_VERDICT...NFT_REG_4:
8420 return reg * NFT_REG_SIZE / NFT_REG32_SIZE;
8421 default:
8422 return reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
8423 }
8424 }
8425 EXPORT_SYMBOL_GPL(nft_parse_register);
8426
8427 /**
8428 * nft_dump_register - dump a register value to a netlink attribute
8429 *
8430 * @skb: socket buffer
8431 * @attr: attribute number
8432 * @reg: register number
8433 *
8434 * Construct a netlink attribute containing the register number. For
8435 * compatibility reasons, register numbers being a multiple of 4 are
8436 * translated to the corresponding 128 bit register numbers.
8437 */
8438 int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
8439 {
8440 if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
8441 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
8442 else
8443 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
8444
8445 return nla_put_be32(skb, attr, htonl(reg));
8446 }
8447 EXPORT_SYMBOL_GPL(nft_dump_register);
8448
8449 /**
8450 * nft_validate_register_load - validate a load from a register
8451 *
8452 * @reg: the register number
8453 * @len: the length of the data
8454 *
8455 * Validate that the input register is one of the general purpose
8456 * registers and that the length of the load is within the bounds.
8457 */
8458 int nft_validate_register_load(enum nft_registers reg, unsigned int len)
8459 {
8460 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
8461 return -EINVAL;
8462 if (len == 0)
8463 return -EINVAL;
8464 if (reg * NFT_REG32_SIZE + len > sizeof_field(struct nft_regs, data))
8465 return -ERANGE;
8466
8467 return 0;
8468 }
8469 EXPORT_SYMBOL_GPL(nft_validate_register_load);
8470
8471 /**
8472 * nft_validate_register_store - validate an expressions' register store
8473 *
8474 * @ctx: context of the expression performing the load
8475 * @reg: the destination register number
8476 * @data: the data to load
8477 * @type: the data type
8478 * @len: the length of the data
8479 *
8480 * Validate that a data load uses the appropriate data type for
8481 * the destination register and the length is within the bounds.
8482 * A value of NULL for the data means that its runtime gathered
8483 * data.
8484 */
8485 int nft_validate_register_store(const struct nft_ctx *ctx,
8486 enum nft_registers reg,
8487 const struct nft_data *data,
8488 enum nft_data_types type, unsigned int len)
8489 {
8490 int err;
8491
8492 switch (reg) {
8493 case NFT_REG_VERDICT:
8494 if (type != NFT_DATA_VERDICT)
8495 return -EINVAL;
8496
8497 if (data != NULL &&
8498 (data->verdict.code == NFT_GOTO ||
8499 data->verdict.code == NFT_JUMP)) {
8500 err = nf_tables_check_loops(ctx, data->verdict.chain);
8501 if (err < 0)
8502 return err;
8503 }
8504
8505 return 0;
8506 default:
8507 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
8508 return -EINVAL;
8509 if (len == 0)
8510 return -EINVAL;
8511 if (reg * NFT_REG32_SIZE + len >
8512 sizeof_field(struct nft_regs, data))
8513 return -ERANGE;
8514
8515 if (data != NULL && type != NFT_DATA_VALUE)
8516 return -EINVAL;
8517 return 0;
8518 }
8519 }
8520 EXPORT_SYMBOL_GPL(nft_validate_register_store);
8521
8522 static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
8523 [NFTA_VERDICT_CODE] = { .type = NLA_U32 },
8524 [NFTA_VERDICT_CHAIN] = { .type = NLA_STRING,
8525 .len = NFT_CHAIN_MAXNAMELEN - 1 },
8526 [NFTA_VERDICT_CHAIN_ID] = { .type = NLA_U32 },
8527 };
8528
8529 static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
8530 struct nft_data_desc *desc, const struct nlattr *nla)
8531 {
8532 u8 genmask = nft_genmask_next(ctx->net);
8533 struct nlattr *tb[NFTA_VERDICT_MAX + 1];
8534 struct nft_chain *chain;
8535 int err;
8536
8537 err = nla_parse_nested_deprecated(tb, NFTA_VERDICT_MAX, nla,
8538 nft_verdict_policy, NULL);
8539 if (err < 0)
8540 return err;
8541
8542 if (!tb[NFTA_VERDICT_CODE])
8543 return -EINVAL;
8544 data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
8545
8546 switch (data->verdict.code) {
8547 default:
8548 switch (data->verdict.code & NF_VERDICT_MASK) {
8549 case NF_ACCEPT:
8550 case NF_DROP:
8551 case NF_QUEUE:
8552 break;
8553 default:
8554 return -EINVAL;
8555 }
8556 fallthrough;
8557 case NFT_CONTINUE:
8558 case NFT_BREAK:
8559 case NFT_RETURN:
8560 break;
8561 case NFT_JUMP:
8562 case NFT_GOTO:
8563 if (tb[NFTA_VERDICT_CHAIN]) {
8564 chain = nft_chain_lookup(ctx->net, ctx->table,
8565 tb[NFTA_VERDICT_CHAIN],
8566 genmask);
8567 } else if (tb[NFTA_VERDICT_CHAIN_ID]) {
8568 chain = nft_chain_lookup_byid(ctx->net,
8569 tb[NFTA_VERDICT_CHAIN_ID]);
8570 if (IS_ERR(chain))
8571 return PTR_ERR(chain);
8572 } else {
8573 return -EINVAL;
8574 }
8575
8576 if (IS_ERR(chain))
8577 return PTR_ERR(chain);
8578 if (nft_is_base_chain(chain))
8579 return -EOPNOTSUPP;
8580
8581 chain->use++;
8582 data->verdict.chain = chain;
8583 break;
8584 }
8585
8586 desc->len = sizeof(data->verdict);
8587 desc->type = NFT_DATA_VERDICT;
8588 return 0;
8589 }
8590
8591 static void nft_verdict_uninit(const struct nft_data *data)
8592 {
8593 struct nft_chain *chain;
8594 struct nft_rule *rule;
8595
8596 switch (data->verdict.code) {
8597 case NFT_JUMP:
8598 case NFT_GOTO:
8599 chain = data->verdict.chain;
8600 chain->use--;
8601
8602 if (!nft_chain_is_bound(chain))
8603 break;
8604
8605 chain->table->use--;
8606 list_for_each_entry(rule, &chain->rules, list)
8607 chain->use--;
8608
8609 nft_chain_del(chain);
8610 break;
8611 }
8612 }
8613
8614 int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
8615 {
8616 struct nlattr *nest;
8617
8618 nest = nla_nest_start_noflag(skb, type);
8619 if (!nest)
8620 goto nla_put_failure;
8621
8622 if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
8623 goto nla_put_failure;
8624
8625 switch (v->code) {
8626 case NFT_JUMP:
8627 case NFT_GOTO:
8628 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
8629 v->chain->name))
8630 goto nla_put_failure;
8631 }
8632 nla_nest_end(skb, nest);
8633 return 0;
8634
8635 nla_put_failure:
8636 return -1;
8637 }
8638
8639 static int nft_value_init(const struct nft_ctx *ctx,
8640 struct nft_data *data, unsigned int size,
8641 struct nft_data_desc *desc, const struct nlattr *nla)
8642 {
8643 unsigned int len;
8644
8645 len = nla_len(nla);
8646 if (len == 0)
8647 return -EINVAL;
8648 if (len > size)
8649 return -EOVERFLOW;
8650
8651 nla_memcpy(data->data, nla, len);
8652 desc->type = NFT_DATA_VALUE;
8653 desc->len = len;
8654 return 0;
8655 }
8656
8657 static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
8658 unsigned int len)
8659 {
8660 return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
8661 }
8662
8663 static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
8664 [NFTA_DATA_VALUE] = { .type = NLA_BINARY },
8665 [NFTA_DATA_VERDICT] = { .type = NLA_NESTED },
8666 };
8667
8668 /**
8669 * nft_data_init - parse nf_tables data netlink attributes
8670 *
8671 * @ctx: context of the expression using the data
8672 * @data: destination struct nft_data
8673 * @size: maximum data length
8674 * @desc: data description
8675 * @nla: netlink attribute containing data
8676 *
8677 * Parse the netlink data attributes and initialize a struct nft_data.
8678 * The type and length of data are returned in the data description.
8679 *
8680 * The caller can indicate that it only wants to accept data of type
8681 * NFT_DATA_VALUE by passing NULL for the ctx argument.
8682 */
8683 int nft_data_init(const struct nft_ctx *ctx,
8684 struct nft_data *data, unsigned int size,
8685 struct nft_data_desc *desc, const struct nlattr *nla)
8686 {
8687 struct nlattr *tb[NFTA_DATA_MAX + 1];
8688 int err;
8689
8690 err = nla_parse_nested_deprecated(tb, NFTA_DATA_MAX, nla,
8691 nft_data_policy, NULL);
8692 if (err < 0)
8693 return err;
8694
8695 if (tb[NFTA_DATA_VALUE])
8696 return nft_value_init(ctx, data, size, desc,
8697 tb[NFTA_DATA_VALUE]);
8698 if (tb[NFTA_DATA_VERDICT] && ctx != NULL)
8699 return nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
8700 return -EINVAL;
8701 }
8702 EXPORT_SYMBOL_GPL(nft_data_init);
8703
8704 /**
8705 * nft_data_release - release a nft_data item
8706 *
8707 * @data: struct nft_data to release
8708 * @type: type of data
8709 *
8710 * Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
8711 * all others need to be released by calling this function.
8712 */
8713 void nft_data_release(const struct nft_data *data, enum nft_data_types type)
8714 {
8715 if (type < NFT_DATA_VERDICT)
8716 return;
8717 switch (type) {
8718 case NFT_DATA_VERDICT:
8719 return nft_verdict_uninit(data);
8720 default:
8721 WARN_ON(1);
8722 }
8723 }
8724 EXPORT_SYMBOL_GPL(nft_data_release);
8725
8726 int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
8727 enum nft_data_types type, unsigned int len)
8728 {
8729 struct nlattr *nest;
8730 int err;
8731
8732 nest = nla_nest_start_noflag(skb, attr);
8733 if (nest == NULL)
8734 return -1;
8735
8736 switch (type) {
8737 case NFT_DATA_VALUE:
8738 err = nft_value_dump(skb, data, len);
8739 break;
8740 case NFT_DATA_VERDICT:
8741 err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
8742 break;
8743 default:
8744 err = -EINVAL;
8745 WARN_ON(1);
8746 }
8747
8748 nla_nest_end(skb, nest);
8749 return err;
8750 }
8751 EXPORT_SYMBOL_GPL(nft_data_dump);
8752
8753 int __nft_release_basechain(struct nft_ctx *ctx)
8754 {
8755 struct nft_rule *rule, *nr;
8756
8757 if (WARN_ON(!nft_is_base_chain(ctx->chain)))
8758 return 0;
8759
8760 nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
8761 list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
8762 list_del(&rule->list);
8763 ctx->chain->use--;
8764 nf_tables_rule_release(ctx, rule);
8765 }
8766 nft_chain_del(ctx->chain);
8767 ctx->table->use--;
8768 nf_tables_chain_destroy(ctx);
8769
8770 return 0;
8771 }
8772 EXPORT_SYMBOL_GPL(__nft_release_basechain);
8773
8774 static void __nft_release_tables(struct net *net)
8775 {
8776 struct nft_flowtable *flowtable, *nf;
8777 struct nft_table *table, *nt;
8778 struct nft_chain *chain, *nc;
8779 struct nft_object *obj, *ne;
8780 struct nft_rule *rule, *nr;
8781 struct nft_set *set, *ns;
8782 struct nft_ctx ctx = {
8783 .net = net,
8784 .family = NFPROTO_NETDEV,
8785 };
8786
8787 list_for_each_entry_safe(table, nt, &net->nft.tables, list) {
8788 ctx.family = table->family;
8789
8790 list_for_each_entry(chain, &table->chains, list)
8791 nf_tables_unregister_hook(net, table, chain);
8792 /* No packets are walking on these chains anymore. */
8793 ctx.table = table;
8794 list_for_each_entry(chain, &table->chains, list) {
8795 ctx.chain = chain;
8796 list_for_each_entry_safe(rule, nr, &chain->rules, list) {
8797 list_del(&rule->list);
8798 chain->use--;
8799 nf_tables_rule_release(&ctx, rule);
8800 }
8801 }
8802 list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
8803 list_del(&flowtable->list);
8804 table->use--;
8805 nf_tables_flowtable_destroy(flowtable);
8806 }
8807 list_for_each_entry_safe(set, ns, &table->sets, list) {
8808 list_del(&set->list);
8809 table->use--;
8810 nft_set_destroy(&ctx, set);
8811 }
8812 list_for_each_entry_safe(obj, ne, &table->objects, list) {
8813 nft_obj_del(obj);
8814 table->use--;
8815 nft_obj_destroy(&ctx, obj);
8816 }
8817 list_for_each_entry_safe(chain, nc, &table->chains, list) {
8818 ctx.chain = chain;
8819 nft_chain_del(chain);
8820 table->use--;
8821 nf_tables_chain_destroy(&ctx);
8822 }
8823 list_del(&table->list);
8824 nf_tables_table_destroy(&ctx);
8825 }
8826 }
8827
8828 static int __net_init nf_tables_init_net(struct net *net)
8829 {
8830 INIT_LIST_HEAD(&net->nft.tables);
8831 INIT_LIST_HEAD(&net->nft.commit_list);
8832 INIT_LIST_HEAD(&net->nft.module_list);
8833 INIT_LIST_HEAD(&net->nft.notify_list);
8834 mutex_init(&net->nft.commit_mutex);
8835 net->nft.base_seq = 1;
8836 net->nft.validate_state = NFT_VALIDATE_SKIP;
8837
8838 return 0;
8839 }
8840
8841 static void __net_exit nf_tables_exit_net(struct net *net)
8842 {
8843 mutex_lock(&net->nft.commit_mutex);
8844 if (!list_empty(&net->nft.commit_list))
8845 __nf_tables_abort(net, NFNL_ABORT_NONE);
8846 __nft_release_tables(net);
8847 mutex_unlock(&net->nft.commit_mutex);
8848 WARN_ON_ONCE(!list_empty(&net->nft.tables));
8849 WARN_ON_ONCE(!list_empty(&net->nft.module_list));
8850 WARN_ON_ONCE(!list_empty(&net->nft.notify_list));
8851 }
8852
8853 static struct pernet_operations nf_tables_net_ops = {
8854 .init = nf_tables_init_net,
8855 .exit = nf_tables_exit_net,
8856 };
8857
8858 static int __init nf_tables_module_init(void)
8859 {
8860 int err;
8861
8862 spin_lock_init(&nf_tables_destroy_list_lock);
8863 err = register_pernet_subsys(&nf_tables_net_ops);
8864 if (err < 0)
8865 return err;
8866
8867 err = nft_chain_filter_init();
8868 if (err < 0)
8869 goto err1;
8870
8871 err = nf_tables_core_module_init();
8872 if (err < 0)
8873 goto err2;
8874
8875 err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
8876 if (err < 0)
8877 goto err3;
8878
8879 err = rhltable_init(&nft_objname_ht, &nft_objname_ht_params);
8880 if (err < 0)
8881 goto err4;
8882
8883 err = nft_offload_init();
8884 if (err < 0)
8885 goto err5;
8886
8887 /* must be last */
8888 err = nfnetlink_subsys_register(&nf_tables_subsys);
8889 if (err < 0)
8890 goto err6;
8891
8892 nft_chain_route_init();
8893
8894 return err;
8895 err6:
8896 nft_offload_exit();
8897 err5:
8898 rhltable_destroy(&nft_objname_ht);
8899 err4:
8900 unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
8901 err3:
8902 nf_tables_core_module_exit();
8903 err2:
8904 nft_chain_filter_fini();
8905 err1:
8906 unregister_pernet_subsys(&nf_tables_net_ops);
8907 return err;
8908 }
8909
8910 static void __exit nf_tables_module_exit(void)
8911 {
8912 nfnetlink_subsys_unregister(&nf_tables_subsys);
8913 nft_offload_exit();
8914 unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
8915 nft_chain_filter_fini();
8916 nft_chain_route_fini();
8917 unregister_pernet_subsys(&nf_tables_net_ops);
8918 cancel_work_sync(&trans_destroy_work);
8919 rcu_barrier();
8920 rhltable_destroy(&nft_objname_ht);
8921 nf_tables_core_module_exit();
8922 }
8923
8924 module_init(nf_tables_module_init);
8925 module_exit(nf_tables_module_exit);
8926
8927 MODULE_LICENSE("GPL");
8928 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
8929 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);