]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - net/netfilter/nf_tables_api.c
Merge tag 'trace-v5.15-rc6-3' of git://git.kernel.org/pub/scm/linux/kernel/git/rosted...
[mirror_ubuntu-jammy-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 unsigned int nf_tables_net_id __read_mostly;
29
30 static LIST_HEAD(nf_tables_expressions);
31 static LIST_HEAD(nf_tables_objects);
32 static LIST_HEAD(nf_tables_flowtables);
33 static LIST_HEAD(nf_tables_destroy_list);
34 static DEFINE_SPINLOCK(nf_tables_destroy_list_lock);
35 static u64 table_handle;
36
37 enum {
38 NFT_VALIDATE_SKIP = 0,
39 NFT_VALIDATE_NEED,
40 NFT_VALIDATE_DO,
41 };
42
43 static struct rhltable nft_objname_ht;
44
45 static u32 nft_chain_hash(const void *data, u32 len, u32 seed);
46 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed);
47 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *, const void *);
48
49 static u32 nft_objname_hash(const void *data, u32 len, u32 seed);
50 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed);
51 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *, const void *);
52
53 static const struct rhashtable_params nft_chain_ht_params = {
54 .head_offset = offsetof(struct nft_chain, rhlhead),
55 .key_offset = offsetof(struct nft_chain, name),
56 .hashfn = nft_chain_hash,
57 .obj_hashfn = nft_chain_hash_obj,
58 .obj_cmpfn = nft_chain_hash_cmp,
59 .automatic_shrinking = true,
60 };
61
62 static const struct rhashtable_params nft_objname_ht_params = {
63 .head_offset = offsetof(struct nft_object, rhlhead),
64 .key_offset = offsetof(struct nft_object, key),
65 .hashfn = nft_objname_hash,
66 .obj_hashfn = nft_objname_hash_obj,
67 .obj_cmpfn = nft_objname_hash_cmp,
68 .automatic_shrinking = true,
69 };
70
71 struct nft_audit_data {
72 struct nft_table *table;
73 int entries;
74 int op;
75 struct list_head list;
76 };
77
78 static const u8 nft2audit_op[NFT_MSG_MAX] = { // enum nf_tables_msg_types
79 [NFT_MSG_NEWTABLE] = AUDIT_NFT_OP_TABLE_REGISTER,
80 [NFT_MSG_GETTABLE] = AUDIT_NFT_OP_INVALID,
81 [NFT_MSG_DELTABLE] = AUDIT_NFT_OP_TABLE_UNREGISTER,
82 [NFT_MSG_NEWCHAIN] = AUDIT_NFT_OP_CHAIN_REGISTER,
83 [NFT_MSG_GETCHAIN] = AUDIT_NFT_OP_INVALID,
84 [NFT_MSG_DELCHAIN] = AUDIT_NFT_OP_CHAIN_UNREGISTER,
85 [NFT_MSG_NEWRULE] = AUDIT_NFT_OP_RULE_REGISTER,
86 [NFT_MSG_GETRULE] = AUDIT_NFT_OP_INVALID,
87 [NFT_MSG_DELRULE] = AUDIT_NFT_OP_RULE_UNREGISTER,
88 [NFT_MSG_NEWSET] = AUDIT_NFT_OP_SET_REGISTER,
89 [NFT_MSG_GETSET] = AUDIT_NFT_OP_INVALID,
90 [NFT_MSG_DELSET] = AUDIT_NFT_OP_SET_UNREGISTER,
91 [NFT_MSG_NEWSETELEM] = AUDIT_NFT_OP_SETELEM_REGISTER,
92 [NFT_MSG_GETSETELEM] = AUDIT_NFT_OP_INVALID,
93 [NFT_MSG_DELSETELEM] = AUDIT_NFT_OP_SETELEM_UNREGISTER,
94 [NFT_MSG_NEWGEN] = AUDIT_NFT_OP_GEN_REGISTER,
95 [NFT_MSG_GETGEN] = AUDIT_NFT_OP_INVALID,
96 [NFT_MSG_TRACE] = AUDIT_NFT_OP_INVALID,
97 [NFT_MSG_NEWOBJ] = AUDIT_NFT_OP_OBJ_REGISTER,
98 [NFT_MSG_GETOBJ] = AUDIT_NFT_OP_INVALID,
99 [NFT_MSG_DELOBJ] = AUDIT_NFT_OP_OBJ_UNREGISTER,
100 [NFT_MSG_GETOBJ_RESET] = AUDIT_NFT_OP_OBJ_RESET,
101 [NFT_MSG_NEWFLOWTABLE] = AUDIT_NFT_OP_FLOWTABLE_REGISTER,
102 [NFT_MSG_GETFLOWTABLE] = AUDIT_NFT_OP_INVALID,
103 [NFT_MSG_DELFLOWTABLE] = AUDIT_NFT_OP_FLOWTABLE_UNREGISTER,
104 };
105
106 static void nft_validate_state_update(struct net *net, u8 new_validate_state)
107 {
108 struct nftables_pernet *nft_net = nft_pernet(net);
109
110 switch (nft_net->validate_state) {
111 case NFT_VALIDATE_SKIP:
112 WARN_ON_ONCE(new_validate_state == NFT_VALIDATE_DO);
113 break;
114 case NFT_VALIDATE_NEED:
115 break;
116 case NFT_VALIDATE_DO:
117 if (new_validate_state == NFT_VALIDATE_NEED)
118 return;
119 }
120
121 nft_net->validate_state = new_validate_state;
122 }
123 static void nf_tables_trans_destroy_work(struct work_struct *w);
124 static DECLARE_WORK(trans_destroy_work, nf_tables_trans_destroy_work);
125
126 static void nft_ctx_init(struct nft_ctx *ctx,
127 struct net *net,
128 const struct sk_buff *skb,
129 const struct nlmsghdr *nlh,
130 u8 family,
131 struct nft_table *table,
132 struct nft_chain *chain,
133 const struct nlattr * const *nla)
134 {
135 ctx->net = net;
136 ctx->family = family;
137 ctx->level = 0;
138 ctx->table = table;
139 ctx->chain = chain;
140 ctx->nla = nla;
141 ctx->portid = NETLINK_CB(skb).portid;
142 ctx->report = nlmsg_report(nlh);
143 ctx->flags = nlh->nlmsg_flags;
144 ctx->seq = nlh->nlmsg_seq;
145 }
146
147 static struct nft_trans *nft_trans_alloc_gfp(const struct nft_ctx *ctx,
148 int msg_type, u32 size, gfp_t gfp)
149 {
150 struct nft_trans *trans;
151
152 trans = kzalloc(sizeof(struct nft_trans) + size, gfp);
153 if (trans == NULL)
154 return NULL;
155
156 trans->msg_type = msg_type;
157 trans->ctx = *ctx;
158
159 return trans;
160 }
161
162 static struct nft_trans *nft_trans_alloc(const struct nft_ctx *ctx,
163 int msg_type, u32 size)
164 {
165 return nft_trans_alloc_gfp(ctx, msg_type, size, GFP_KERNEL);
166 }
167
168 static void nft_trans_destroy(struct nft_trans *trans)
169 {
170 list_del(&trans->list);
171 kfree(trans);
172 }
173
174 static void nft_set_trans_bind(const struct nft_ctx *ctx, struct nft_set *set)
175 {
176 struct nftables_pernet *nft_net;
177 struct net *net = ctx->net;
178 struct nft_trans *trans;
179
180 if (!nft_set_is_anonymous(set))
181 return;
182
183 nft_net = nft_pernet(net);
184 list_for_each_entry_reverse(trans, &nft_net->commit_list, list) {
185 switch (trans->msg_type) {
186 case NFT_MSG_NEWSET:
187 if (nft_trans_set(trans) == set)
188 nft_trans_set_bound(trans) = true;
189 break;
190 case NFT_MSG_NEWSETELEM:
191 if (nft_trans_elem_set(trans) == set)
192 nft_trans_elem_set_bound(trans) = true;
193 break;
194 }
195 }
196 }
197
198 static int nft_netdev_register_hooks(struct net *net,
199 struct list_head *hook_list)
200 {
201 struct nft_hook *hook;
202 int err, j;
203
204 j = 0;
205 list_for_each_entry(hook, hook_list, list) {
206 err = nf_register_net_hook(net, &hook->ops);
207 if (err < 0)
208 goto err_register;
209
210 j++;
211 }
212 return 0;
213
214 err_register:
215 list_for_each_entry(hook, hook_list, list) {
216 if (j-- <= 0)
217 break;
218
219 nf_unregister_net_hook(net, &hook->ops);
220 }
221 return err;
222 }
223
224 static void nft_netdev_unregister_hooks(struct net *net,
225 struct list_head *hook_list)
226 {
227 struct nft_hook *hook;
228
229 list_for_each_entry(hook, hook_list, list)
230 nf_unregister_net_hook(net, &hook->ops);
231 }
232
233 static int nf_tables_register_hook(struct net *net,
234 const struct nft_table *table,
235 struct nft_chain *chain)
236 {
237 struct nft_base_chain *basechain;
238 const struct nf_hook_ops *ops;
239
240 if (table->flags & NFT_TABLE_F_DORMANT ||
241 !nft_is_base_chain(chain))
242 return 0;
243
244 basechain = nft_base_chain(chain);
245 ops = &basechain->ops;
246
247 if (basechain->type->ops_register)
248 return basechain->type->ops_register(net, ops);
249
250 if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
251 return nft_netdev_register_hooks(net, &basechain->hook_list);
252
253 return nf_register_net_hook(net, &basechain->ops);
254 }
255
256 static void nf_tables_unregister_hook(struct net *net,
257 const struct nft_table *table,
258 struct nft_chain *chain)
259 {
260 struct nft_base_chain *basechain;
261 const struct nf_hook_ops *ops;
262
263 if (table->flags & NFT_TABLE_F_DORMANT ||
264 !nft_is_base_chain(chain))
265 return;
266 basechain = nft_base_chain(chain);
267 ops = &basechain->ops;
268
269 if (basechain->type->ops_unregister)
270 return basechain->type->ops_unregister(net, ops);
271
272 if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
273 nft_netdev_unregister_hooks(net, &basechain->hook_list);
274 else
275 nf_unregister_net_hook(net, &basechain->ops);
276 }
277
278 static void nft_trans_commit_list_add_tail(struct net *net, struct nft_trans *trans)
279 {
280 struct nftables_pernet *nft_net = nft_pernet(net);
281
282 list_add_tail(&trans->list, &nft_net->commit_list);
283 }
284
285 static int nft_trans_table_add(struct nft_ctx *ctx, int msg_type)
286 {
287 struct nft_trans *trans;
288
289 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_table));
290 if (trans == NULL)
291 return -ENOMEM;
292
293 if (msg_type == NFT_MSG_NEWTABLE)
294 nft_activate_next(ctx->net, ctx->table);
295
296 nft_trans_commit_list_add_tail(ctx->net, trans);
297 return 0;
298 }
299
300 static int nft_deltable(struct nft_ctx *ctx)
301 {
302 int err;
303
304 err = nft_trans_table_add(ctx, NFT_MSG_DELTABLE);
305 if (err < 0)
306 return err;
307
308 nft_deactivate_next(ctx->net, ctx->table);
309 return err;
310 }
311
312 static struct nft_trans *nft_trans_chain_add(struct nft_ctx *ctx, int msg_type)
313 {
314 struct nft_trans *trans;
315
316 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_chain));
317 if (trans == NULL)
318 return ERR_PTR(-ENOMEM);
319
320 if (msg_type == NFT_MSG_NEWCHAIN) {
321 nft_activate_next(ctx->net, ctx->chain);
322
323 if (ctx->nla[NFTA_CHAIN_ID]) {
324 nft_trans_chain_id(trans) =
325 ntohl(nla_get_be32(ctx->nla[NFTA_CHAIN_ID]));
326 }
327 }
328
329 nft_trans_commit_list_add_tail(ctx->net, trans);
330 return trans;
331 }
332
333 static int nft_delchain(struct nft_ctx *ctx)
334 {
335 struct nft_trans *trans;
336
337 trans = nft_trans_chain_add(ctx, NFT_MSG_DELCHAIN);
338 if (IS_ERR(trans))
339 return PTR_ERR(trans);
340
341 ctx->table->use--;
342 nft_deactivate_next(ctx->net, ctx->chain);
343
344 return 0;
345 }
346
347 static void nft_rule_expr_activate(const struct nft_ctx *ctx,
348 struct nft_rule *rule)
349 {
350 struct nft_expr *expr;
351
352 expr = nft_expr_first(rule);
353 while (nft_expr_more(rule, expr)) {
354 if (expr->ops->activate)
355 expr->ops->activate(ctx, expr);
356
357 expr = nft_expr_next(expr);
358 }
359 }
360
361 static void nft_rule_expr_deactivate(const struct nft_ctx *ctx,
362 struct nft_rule *rule,
363 enum nft_trans_phase phase)
364 {
365 struct nft_expr *expr;
366
367 expr = nft_expr_first(rule);
368 while (nft_expr_more(rule, expr)) {
369 if (expr->ops->deactivate)
370 expr->ops->deactivate(ctx, expr, phase);
371
372 expr = nft_expr_next(expr);
373 }
374 }
375
376 static int
377 nf_tables_delrule_deactivate(struct nft_ctx *ctx, struct nft_rule *rule)
378 {
379 /* You cannot delete the same rule twice */
380 if (nft_is_active_next(ctx->net, rule)) {
381 nft_deactivate_next(ctx->net, rule);
382 ctx->chain->use--;
383 return 0;
384 }
385 return -ENOENT;
386 }
387
388 static struct nft_trans *nft_trans_rule_add(struct nft_ctx *ctx, int msg_type,
389 struct nft_rule *rule)
390 {
391 struct nft_trans *trans;
392
393 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_rule));
394 if (trans == NULL)
395 return NULL;
396
397 if (msg_type == NFT_MSG_NEWRULE && ctx->nla[NFTA_RULE_ID] != NULL) {
398 nft_trans_rule_id(trans) =
399 ntohl(nla_get_be32(ctx->nla[NFTA_RULE_ID]));
400 }
401 nft_trans_rule(trans) = rule;
402 nft_trans_commit_list_add_tail(ctx->net, trans);
403
404 return trans;
405 }
406
407 static int nft_delrule(struct nft_ctx *ctx, struct nft_rule *rule)
408 {
409 struct nft_flow_rule *flow;
410 struct nft_trans *trans;
411 int err;
412
413 trans = nft_trans_rule_add(ctx, NFT_MSG_DELRULE, rule);
414 if (trans == NULL)
415 return -ENOMEM;
416
417 if (ctx->chain->flags & NFT_CHAIN_HW_OFFLOAD) {
418 flow = nft_flow_rule_create(ctx->net, rule);
419 if (IS_ERR(flow)) {
420 nft_trans_destroy(trans);
421 return PTR_ERR(flow);
422 }
423
424 nft_trans_flow_rule(trans) = flow;
425 }
426
427 err = nf_tables_delrule_deactivate(ctx, rule);
428 if (err < 0) {
429 nft_trans_destroy(trans);
430 return err;
431 }
432 nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_PREPARE);
433
434 return 0;
435 }
436
437 static int nft_delrule_by_chain(struct nft_ctx *ctx)
438 {
439 struct nft_rule *rule;
440 int err;
441
442 list_for_each_entry(rule, &ctx->chain->rules, list) {
443 if (!nft_is_active_next(ctx->net, rule))
444 continue;
445
446 err = nft_delrule(ctx, rule);
447 if (err < 0)
448 return err;
449 }
450 return 0;
451 }
452
453 static int nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
454 struct nft_set *set)
455 {
456 struct nft_trans *trans;
457
458 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_set));
459 if (trans == NULL)
460 return -ENOMEM;
461
462 if (msg_type == NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID] != NULL) {
463 nft_trans_set_id(trans) =
464 ntohl(nla_get_be32(ctx->nla[NFTA_SET_ID]));
465 nft_activate_next(ctx->net, set);
466 }
467 nft_trans_set(trans) = set;
468 nft_trans_commit_list_add_tail(ctx->net, trans);
469
470 return 0;
471 }
472
473 static int nft_delset(const struct nft_ctx *ctx, struct nft_set *set)
474 {
475 int err;
476
477 err = nft_trans_set_add(ctx, NFT_MSG_DELSET, set);
478 if (err < 0)
479 return err;
480
481 nft_deactivate_next(ctx->net, set);
482 ctx->table->use--;
483
484 return err;
485 }
486
487 static int nft_trans_obj_add(struct nft_ctx *ctx, int msg_type,
488 struct nft_object *obj)
489 {
490 struct nft_trans *trans;
491
492 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_obj));
493 if (trans == NULL)
494 return -ENOMEM;
495
496 if (msg_type == NFT_MSG_NEWOBJ)
497 nft_activate_next(ctx->net, obj);
498
499 nft_trans_obj(trans) = obj;
500 nft_trans_commit_list_add_tail(ctx->net, trans);
501
502 return 0;
503 }
504
505 static int nft_delobj(struct nft_ctx *ctx, struct nft_object *obj)
506 {
507 int err;
508
509 err = nft_trans_obj_add(ctx, NFT_MSG_DELOBJ, obj);
510 if (err < 0)
511 return err;
512
513 nft_deactivate_next(ctx->net, obj);
514 ctx->table->use--;
515
516 return err;
517 }
518
519 static int nft_trans_flowtable_add(struct nft_ctx *ctx, int msg_type,
520 struct nft_flowtable *flowtable)
521 {
522 struct nft_trans *trans;
523
524 trans = nft_trans_alloc(ctx, msg_type,
525 sizeof(struct nft_trans_flowtable));
526 if (trans == NULL)
527 return -ENOMEM;
528
529 if (msg_type == NFT_MSG_NEWFLOWTABLE)
530 nft_activate_next(ctx->net, flowtable);
531
532 nft_trans_flowtable(trans) = flowtable;
533 nft_trans_commit_list_add_tail(ctx->net, trans);
534
535 return 0;
536 }
537
538 static int nft_delflowtable(struct nft_ctx *ctx,
539 struct nft_flowtable *flowtable)
540 {
541 int err;
542
543 err = nft_trans_flowtable_add(ctx, NFT_MSG_DELFLOWTABLE, flowtable);
544 if (err < 0)
545 return err;
546
547 nft_deactivate_next(ctx->net, flowtable);
548 ctx->table->use--;
549
550 return err;
551 }
552
553 /*
554 * Tables
555 */
556
557 static struct nft_table *nft_table_lookup(const struct net *net,
558 const struct nlattr *nla,
559 u8 family, u8 genmask, u32 nlpid)
560 {
561 struct nftables_pernet *nft_net;
562 struct nft_table *table;
563
564 if (nla == NULL)
565 return ERR_PTR(-EINVAL);
566
567 nft_net = nft_pernet(net);
568 list_for_each_entry_rcu(table, &nft_net->tables, list,
569 lockdep_is_held(&nft_net->commit_mutex)) {
570 if (!nla_strcmp(nla, table->name) &&
571 table->family == family &&
572 nft_active_genmask(table, genmask)) {
573 if (nft_table_has_owner(table) &&
574 nlpid && table->nlpid != nlpid)
575 return ERR_PTR(-EPERM);
576
577 return table;
578 }
579 }
580
581 return ERR_PTR(-ENOENT);
582 }
583
584 static struct nft_table *nft_table_lookup_byhandle(const struct net *net,
585 const struct nlattr *nla,
586 u8 genmask, u32 nlpid)
587 {
588 struct nftables_pernet *nft_net;
589 struct nft_table *table;
590
591 nft_net = nft_pernet(net);
592 list_for_each_entry(table, &nft_net->tables, list) {
593 if (be64_to_cpu(nla_get_be64(nla)) == table->handle &&
594 nft_active_genmask(table, genmask)) {
595 if (nft_table_has_owner(table) &&
596 nlpid && table->nlpid != nlpid)
597 return ERR_PTR(-EPERM);
598
599 return table;
600 }
601 }
602
603 return ERR_PTR(-ENOENT);
604 }
605
606 static inline u64 nf_tables_alloc_handle(struct nft_table *table)
607 {
608 return ++table->hgenerator;
609 }
610
611 static const struct nft_chain_type *chain_type[NFPROTO_NUMPROTO][NFT_CHAIN_T_MAX];
612
613 static const struct nft_chain_type *
614 __nft_chain_type_get(u8 family, enum nft_chain_types type)
615 {
616 if (family >= NFPROTO_NUMPROTO ||
617 type >= NFT_CHAIN_T_MAX)
618 return NULL;
619
620 return chain_type[family][type];
621 }
622
623 static const struct nft_chain_type *
624 __nf_tables_chain_type_lookup(const struct nlattr *nla, u8 family)
625 {
626 const struct nft_chain_type *type;
627 int i;
628
629 for (i = 0; i < NFT_CHAIN_T_MAX; i++) {
630 type = __nft_chain_type_get(family, i);
631 if (!type)
632 continue;
633 if (!nla_strcmp(nla, type->name))
634 return type;
635 }
636 return NULL;
637 }
638
639 struct nft_module_request {
640 struct list_head list;
641 char module[MODULE_NAME_LEN];
642 bool done;
643 };
644
645 #ifdef CONFIG_MODULES
646 __printf(2, 3) int nft_request_module(struct net *net, const char *fmt,
647 ...)
648 {
649 char module_name[MODULE_NAME_LEN];
650 struct nftables_pernet *nft_net;
651 struct nft_module_request *req;
652 va_list args;
653 int ret;
654
655 va_start(args, fmt);
656 ret = vsnprintf(module_name, MODULE_NAME_LEN, fmt, args);
657 va_end(args);
658 if (ret >= MODULE_NAME_LEN)
659 return 0;
660
661 nft_net = nft_pernet(net);
662 list_for_each_entry(req, &nft_net->module_list, list) {
663 if (!strcmp(req->module, module_name)) {
664 if (req->done)
665 return 0;
666
667 /* A request to load this module already exists. */
668 return -EAGAIN;
669 }
670 }
671
672 req = kmalloc(sizeof(*req), GFP_KERNEL);
673 if (!req)
674 return -ENOMEM;
675
676 req->done = false;
677 strlcpy(req->module, module_name, MODULE_NAME_LEN);
678 list_add_tail(&req->list, &nft_net->module_list);
679
680 return -EAGAIN;
681 }
682 EXPORT_SYMBOL_GPL(nft_request_module);
683 #endif
684
685 static void lockdep_nfnl_nft_mutex_not_held(void)
686 {
687 #ifdef CONFIG_PROVE_LOCKING
688 if (debug_locks)
689 WARN_ON_ONCE(lockdep_nfnl_is_held(NFNL_SUBSYS_NFTABLES));
690 #endif
691 }
692
693 static const struct nft_chain_type *
694 nf_tables_chain_type_lookup(struct net *net, const struct nlattr *nla,
695 u8 family, bool autoload)
696 {
697 const struct nft_chain_type *type;
698
699 type = __nf_tables_chain_type_lookup(nla, family);
700 if (type != NULL)
701 return type;
702
703 lockdep_nfnl_nft_mutex_not_held();
704 #ifdef CONFIG_MODULES
705 if (autoload) {
706 if (nft_request_module(net, "nft-chain-%u-%.*s", family,
707 nla_len(nla),
708 (const char *)nla_data(nla)) == -EAGAIN)
709 return ERR_PTR(-EAGAIN);
710 }
711 #endif
712 return ERR_PTR(-ENOENT);
713 }
714
715 static __be16 nft_base_seq(const struct net *net)
716 {
717 struct nftables_pernet *nft_net = nft_pernet(net);
718
719 return htons(nft_net->base_seq & 0xffff);
720 }
721
722 static const struct nla_policy nft_table_policy[NFTA_TABLE_MAX + 1] = {
723 [NFTA_TABLE_NAME] = { .type = NLA_STRING,
724 .len = NFT_TABLE_MAXNAMELEN - 1 },
725 [NFTA_TABLE_FLAGS] = { .type = NLA_U32 },
726 [NFTA_TABLE_HANDLE] = { .type = NLA_U64 },
727 [NFTA_TABLE_USERDATA] = { .type = NLA_BINARY,
728 .len = NFT_USERDATA_MAXLEN }
729 };
730
731 static int nf_tables_fill_table_info(struct sk_buff *skb, struct net *net,
732 u32 portid, u32 seq, int event, u32 flags,
733 int family, const struct nft_table *table)
734 {
735 struct nlmsghdr *nlh;
736
737 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
738 nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
739 NFNETLINK_V0, nft_base_seq(net));
740 if (!nlh)
741 goto nla_put_failure;
742
743 if (nla_put_string(skb, NFTA_TABLE_NAME, table->name) ||
744 nla_put_be32(skb, NFTA_TABLE_FLAGS,
745 htonl(table->flags & NFT_TABLE_F_MASK)) ||
746 nla_put_be32(skb, NFTA_TABLE_USE, htonl(table->use)) ||
747 nla_put_be64(skb, NFTA_TABLE_HANDLE, cpu_to_be64(table->handle),
748 NFTA_TABLE_PAD))
749 goto nla_put_failure;
750 if (nft_table_has_owner(table) &&
751 nla_put_be32(skb, NFTA_TABLE_OWNER, htonl(table->nlpid)))
752 goto nla_put_failure;
753
754 if (table->udata) {
755 if (nla_put(skb, NFTA_TABLE_USERDATA, table->udlen, table->udata))
756 goto nla_put_failure;
757 }
758
759 nlmsg_end(skb, nlh);
760 return 0;
761
762 nla_put_failure:
763 nlmsg_trim(skb, nlh);
764 return -1;
765 }
766
767 struct nftnl_skb_parms {
768 bool report;
769 };
770 #define NFT_CB(skb) (*(struct nftnl_skb_parms*)&((skb)->cb))
771
772 static void nft_notify_enqueue(struct sk_buff *skb, bool report,
773 struct list_head *notify_list)
774 {
775 NFT_CB(skb).report = report;
776 list_add_tail(&skb->list, notify_list);
777 }
778
779 static void nf_tables_table_notify(const struct nft_ctx *ctx, int event)
780 {
781 struct nftables_pernet *nft_net;
782 struct sk_buff *skb;
783 u16 flags = 0;
784 int err;
785
786 if (!ctx->report &&
787 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
788 return;
789
790 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
791 if (skb == NULL)
792 goto err;
793
794 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
795 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
796
797 err = nf_tables_fill_table_info(skb, ctx->net, ctx->portid, ctx->seq,
798 event, flags, ctx->family, ctx->table);
799 if (err < 0) {
800 kfree_skb(skb);
801 goto err;
802 }
803
804 nft_net = nft_pernet(ctx->net);
805 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
806 return;
807 err:
808 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
809 }
810
811 static int nf_tables_dump_tables(struct sk_buff *skb,
812 struct netlink_callback *cb)
813 {
814 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
815 struct nftables_pernet *nft_net;
816 const struct nft_table *table;
817 unsigned int idx = 0, s_idx = cb->args[0];
818 struct net *net = sock_net(skb->sk);
819 int family = nfmsg->nfgen_family;
820
821 rcu_read_lock();
822 nft_net = nft_pernet(net);
823 cb->seq = nft_net->base_seq;
824
825 list_for_each_entry_rcu(table, &nft_net->tables, list) {
826 if (family != NFPROTO_UNSPEC && family != table->family)
827 continue;
828
829 if (idx < s_idx)
830 goto cont;
831 if (idx > s_idx)
832 memset(&cb->args[1], 0,
833 sizeof(cb->args) - sizeof(cb->args[0]));
834 if (!nft_is_active(net, table))
835 continue;
836 if (nf_tables_fill_table_info(skb, net,
837 NETLINK_CB(cb->skb).portid,
838 cb->nlh->nlmsg_seq,
839 NFT_MSG_NEWTABLE, NLM_F_MULTI,
840 table->family, table) < 0)
841 goto done;
842
843 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
844 cont:
845 idx++;
846 }
847 done:
848 rcu_read_unlock();
849 cb->args[0] = idx;
850 return skb->len;
851 }
852
853 static int nft_netlink_dump_start_rcu(struct sock *nlsk, struct sk_buff *skb,
854 const struct nlmsghdr *nlh,
855 struct netlink_dump_control *c)
856 {
857 int err;
858
859 if (!try_module_get(THIS_MODULE))
860 return -EINVAL;
861
862 rcu_read_unlock();
863 err = netlink_dump_start(nlsk, skb, nlh, c);
864 rcu_read_lock();
865 module_put(THIS_MODULE);
866
867 return err;
868 }
869
870 /* called with rcu_read_lock held */
871 static int nf_tables_gettable(struct sk_buff *skb, const struct nfnl_info *info,
872 const struct nlattr * const nla[])
873 {
874 struct netlink_ext_ack *extack = info->extack;
875 u8 genmask = nft_genmask_cur(info->net);
876 u8 family = info->nfmsg->nfgen_family;
877 const struct nft_table *table;
878 struct net *net = info->net;
879 struct sk_buff *skb2;
880 int err;
881
882 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
883 struct netlink_dump_control c = {
884 .dump = nf_tables_dump_tables,
885 .module = THIS_MODULE,
886 };
887
888 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
889 }
890
891 table = nft_table_lookup(net, nla[NFTA_TABLE_NAME], family, genmask, 0);
892 if (IS_ERR(table)) {
893 NL_SET_BAD_ATTR(extack, nla[NFTA_TABLE_NAME]);
894 return PTR_ERR(table);
895 }
896
897 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
898 if (!skb2)
899 return -ENOMEM;
900
901 err = nf_tables_fill_table_info(skb2, net, NETLINK_CB(skb).portid,
902 info->nlh->nlmsg_seq, NFT_MSG_NEWTABLE,
903 0, family, table);
904 if (err < 0)
905 goto err_fill_table_info;
906
907 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
908
909 err_fill_table_info:
910 kfree_skb(skb2);
911 return err;
912 }
913
914 static void nft_table_disable(struct net *net, struct nft_table *table, u32 cnt)
915 {
916 struct nft_chain *chain;
917 u32 i = 0;
918
919 list_for_each_entry(chain, &table->chains, list) {
920 if (!nft_is_active_next(net, chain))
921 continue;
922 if (!nft_is_base_chain(chain))
923 continue;
924
925 if (cnt && i++ == cnt)
926 break;
927
928 nf_tables_unregister_hook(net, table, chain);
929 }
930 }
931
932 static int nf_tables_table_enable(struct net *net, struct nft_table *table)
933 {
934 struct nft_chain *chain;
935 int err, i = 0;
936
937 list_for_each_entry(chain, &table->chains, list) {
938 if (!nft_is_active_next(net, chain))
939 continue;
940 if (!nft_is_base_chain(chain))
941 continue;
942
943 err = nf_tables_register_hook(net, table, chain);
944 if (err < 0)
945 goto err_register_hooks;
946
947 i++;
948 }
949 return 0;
950
951 err_register_hooks:
952 if (i)
953 nft_table_disable(net, table, i);
954 return err;
955 }
956
957 static void nf_tables_table_disable(struct net *net, struct nft_table *table)
958 {
959 table->flags &= ~NFT_TABLE_F_DORMANT;
960 nft_table_disable(net, table, 0);
961 table->flags |= NFT_TABLE_F_DORMANT;
962 }
963
964 #define __NFT_TABLE_F_INTERNAL (NFT_TABLE_F_MASK + 1)
965 #define __NFT_TABLE_F_WAS_DORMANT (__NFT_TABLE_F_INTERNAL << 0)
966 #define __NFT_TABLE_F_WAS_AWAKEN (__NFT_TABLE_F_INTERNAL << 1)
967 #define __NFT_TABLE_F_UPDATE (__NFT_TABLE_F_WAS_DORMANT | \
968 __NFT_TABLE_F_WAS_AWAKEN)
969
970 static int nf_tables_updtable(struct nft_ctx *ctx)
971 {
972 struct nft_trans *trans;
973 u32 flags;
974 int ret;
975
976 if (!ctx->nla[NFTA_TABLE_FLAGS])
977 return 0;
978
979 flags = ntohl(nla_get_be32(ctx->nla[NFTA_TABLE_FLAGS]));
980 if (flags & ~NFT_TABLE_F_MASK)
981 return -EOPNOTSUPP;
982
983 if (flags == ctx->table->flags)
984 return 0;
985
986 if ((nft_table_has_owner(ctx->table) &&
987 !(flags & NFT_TABLE_F_OWNER)) ||
988 (!nft_table_has_owner(ctx->table) &&
989 flags & NFT_TABLE_F_OWNER))
990 return -EOPNOTSUPP;
991
992 trans = nft_trans_alloc(ctx, NFT_MSG_NEWTABLE,
993 sizeof(struct nft_trans_table));
994 if (trans == NULL)
995 return -ENOMEM;
996
997 if ((flags & NFT_TABLE_F_DORMANT) &&
998 !(ctx->table->flags & NFT_TABLE_F_DORMANT)) {
999 ctx->table->flags |= NFT_TABLE_F_DORMANT;
1000 if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE))
1001 ctx->table->flags |= __NFT_TABLE_F_WAS_AWAKEN;
1002 } else if (!(flags & NFT_TABLE_F_DORMANT) &&
1003 ctx->table->flags & NFT_TABLE_F_DORMANT) {
1004 ctx->table->flags &= ~NFT_TABLE_F_DORMANT;
1005 if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE)) {
1006 ret = nf_tables_table_enable(ctx->net, ctx->table);
1007 if (ret < 0)
1008 goto err_register_hooks;
1009
1010 ctx->table->flags |= __NFT_TABLE_F_WAS_DORMANT;
1011 }
1012 }
1013
1014 nft_trans_table_update(trans) = true;
1015 nft_trans_commit_list_add_tail(ctx->net, trans);
1016
1017 return 0;
1018
1019 err_register_hooks:
1020 nft_trans_destroy(trans);
1021 return ret;
1022 }
1023
1024 static u32 nft_chain_hash(const void *data, u32 len, u32 seed)
1025 {
1026 const char *name = data;
1027
1028 return jhash(name, strlen(name), seed);
1029 }
1030
1031 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed)
1032 {
1033 const struct nft_chain *chain = data;
1034
1035 return nft_chain_hash(chain->name, 0, seed);
1036 }
1037
1038 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *arg,
1039 const void *ptr)
1040 {
1041 const struct nft_chain *chain = ptr;
1042 const char *name = arg->key;
1043
1044 return strcmp(chain->name, name);
1045 }
1046
1047 static u32 nft_objname_hash(const void *data, u32 len, u32 seed)
1048 {
1049 const struct nft_object_hash_key *k = data;
1050
1051 seed ^= hash_ptr(k->table, 32);
1052
1053 return jhash(k->name, strlen(k->name), seed);
1054 }
1055
1056 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed)
1057 {
1058 const struct nft_object *obj = data;
1059
1060 return nft_objname_hash(&obj->key, 0, seed);
1061 }
1062
1063 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *arg,
1064 const void *ptr)
1065 {
1066 const struct nft_object_hash_key *k = arg->key;
1067 const struct nft_object *obj = ptr;
1068
1069 if (obj->key.table != k->table)
1070 return -1;
1071
1072 return strcmp(obj->key.name, k->name);
1073 }
1074
1075 static int nf_tables_newtable(struct sk_buff *skb, const struct nfnl_info *info,
1076 const struct nlattr * const nla[])
1077 {
1078 struct nftables_pernet *nft_net = nft_pernet(info->net);
1079 struct netlink_ext_ack *extack = info->extack;
1080 u8 genmask = nft_genmask_next(info->net);
1081 u8 family = info->nfmsg->nfgen_family;
1082 struct net *net = info->net;
1083 const struct nlattr *attr;
1084 struct nft_table *table;
1085 struct nft_ctx ctx;
1086 u32 flags = 0;
1087 int err;
1088
1089 lockdep_assert_held(&nft_net->commit_mutex);
1090 attr = nla[NFTA_TABLE_NAME];
1091 table = nft_table_lookup(net, attr, family, genmask,
1092 NETLINK_CB(skb).portid);
1093 if (IS_ERR(table)) {
1094 if (PTR_ERR(table) != -ENOENT)
1095 return PTR_ERR(table);
1096 } else {
1097 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
1098 NL_SET_BAD_ATTR(extack, attr);
1099 return -EEXIST;
1100 }
1101 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
1102 return -EOPNOTSUPP;
1103
1104 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1105
1106 return nf_tables_updtable(&ctx);
1107 }
1108
1109 if (nla[NFTA_TABLE_FLAGS]) {
1110 flags = ntohl(nla_get_be32(nla[NFTA_TABLE_FLAGS]));
1111 if (flags & ~NFT_TABLE_F_MASK)
1112 return -EOPNOTSUPP;
1113 }
1114
1115 err = -ENOMEM;
1116 table = kzalloc(sizeof(*table), GFP_KERNEL);
1117 if (table == NULL)
1118 goto err_kzalloc;
1119
1120 table->name = nla_strdup(attr, GFP_KERNEL);
1121 if (table->name == NULL)
1122 goto err_strdup;
1123
1124 if (nla[NFTA_TABLE_USERDATA]) {
1125 table->udata = nla_memdup(nla[NFTA_TABLE_USERDATA], GFP_KERNEL);
1126 if (table->udata == NULL)
1127 goto err_table_udata;
1128
1129 table->udlen = nla_len(nla[NFTA_TABLE_USERDATA]);
1130 }
1131
1132 err = rhltable_init(&table->chains_ht, &nft_chain_ht_params);
1133 if (err)
1134 goto err_chain_ht;
1135
1136 INIT_LIST_HEAD(&table->chains);
1137 INIT_LIST_HEAD(&table->sets);
1138 INIT_LIST_HEAD(&table->objects);
1139 INIT_LIST_HEAD(&table->flowtables);
1140 table->family = family;
1141 table->flags = flags;
1142 table->handle = ++table_handle;
1143 if (table->flags & NFT_TABLE_F_OWNER)
1144 table->nlpid = NETLINK_CB(skb).portid;
1145
1146 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1147 err = nft_trans_table_add(&ctx, NFT_MSG_NEWTABLE);
1148 if (err < 0)
1149 goto err_trans;
1150
1151 list_add_tail_rcu(&table->list, &nft_net->tables);
1152 return 0;
1153 err_trans:
1154 rhltable_destroy(&table->chains_ht);
1155 err_chain_ht:
1156 kfree(table->udata);
1157 err_table_udata:
1158 kfree(table->name);
1159 err_strdup:
1160 kfree(table);
1161 err_kzalloc:
1162 return err;
1163 }
1164
1165 static int nft_flush_table(struct nft_ctx *ctx)
1166 {
1167 struct nft_flowtable *flowtable, *nft;
1168 struct nft_chain *chain, *nc;
1169 struct nft_object *obj, *ne;
1170 struct nft_set *set, *ns;
1171 int err;
1172
1173 list_for_each_entry(chain, &ctx->table->chains, list) {
1174 if (!nft_is_active_next(ctx->net, chain))
1175 continue;
1176
1177 if (nft_chain_is_bound(chain))
1178 continue;
1179
1180 ctx->chain = chain;
1181
1182 err = nft_delrule_by_chain(ctx);
1183 if (err < 0)
1184 goto out;
1185 }
1186
1187 list_for_each_entry_safe(set, ns, &ctx->table->sets, list) {
1188 if (!nft_is_active_next(ctx->net, set))
1189 continue;
1190
1191 if (nft_set_is_anonymous(set) &&
1192 !list_empty(&set->bindings))
1193 continue;
1194
1195 err = nft_delset(ctx, set);
1196 if (err < 0)
1197 goto out;
1198 }
1199
1200 list_for_each_entry_safe(flowtable, nft, &ctx->table->flowtables, list) {
1201 if (!nft_is_active_next(ctx->net, flowtable))
1202 continue;
1203
1204 err = nft_delflowtable(ctx, flowtable);
1205 if (err < 0)
1206 goto out;
1207 }
1208
1209 list_for_each_entry_safe(obj, ne, &ctx->table->objects, list) {
1210 if (!nft_is_active_next(ctx->net, obj))
1211 continue;
1212
1213 err = nft_delobj(ctx, obj);
1214 if (err < 0)
1215 goto out;
1216 }
1217
1218 list_for_each_entry_safe(chain, nc, &ctx->table->chains, list) {
1219 if (!nft_is_active_next(ctx->net, chain))
1220 continue;
1221
1222 if (nft_chain_is_bound(chain))
1223 continue;
1224
1225 ctx->chain = chain;
1226
1227 err = nft_delchain(ctx);
1228 if (err < 0)
1229 goto out;
1230 }
1231
1232 err = nft_deltable(ctx);
1233 out:
1234 return err;
1235 }
1236
1237 static int nft_flush(struct nft_ctx *ctx, int family)
1238 {
1239 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
1240 const struct nlattr * const *nla = ctx->nla;
1241 struct nft_table *table, *nt;
1242 int err = 0;
1243
1244 list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
1245 if (family != AF_UNSPEC && table->family != family)
1246 continue;
1247
1248 ctx->family = table->family;
1249
1250 if (!nft_is_active_next(ctx->net, table))
1251 continue;
1252
1253 if (nft_table_has_owner(table) && table->nlpid != ctx->portid)
1254 continue;
1255
1256 if (nla[NFTA_TABLE_NAME] &&
1257 nla_strcmp(nla[NFTA_TABLE_NAME], table->name) != 0)
1258 continue;
1259
1260 ctx->table = table;
1261
1262 err = nft_flush_table(ctx);
1263 if (err < 0)
1264 goto out;
1265 }
1266 out:
1267 return err;
1268 }
1269
1270 static int nf_tables_deltable(struct sk_buff *skb, const struct nfnl_info *info,
1271 const struct nlattr * const nla[])
1272 {
1273 struct netlink_ext_ack *extack = info->extack;
1274 u8 genmask = nft_genmask_next(info->net);
1275 u8 family = info->nfmsg->nfgen_family;
1276 struct net *net = info->net;
1277 const struct nlattr *attr;
1278 struct nft_table *table;
1279 struct nft_ctx ctx;
1280
1281 nft_ctx_init(&ctx, net, skb, info->nlh, 0, NULL, NULL, nla);
1282 if (family == AF_UNSPEC ||
1283 (!nla[NFTA_TABLE_NAME] && !nla[NFTA_TABLE_HANDLE]))
1284 return nft_flush(&ctx, family);
1285
1286 if (nla[NFTA_TABLE_HANDLE]) {
1287 attr = nla[NFTA_TABLE_HANDLE];
1288 table = nft_table_lookup_byhandle(net, attr, genmask,
1289 NETLINK_CB(skb).portid);
1290 } else {
1291 attr = nla[NFTA_TABLE_NAME];
1292 table = nft_table_lookup(net, attr, family, genmask,
1293 NETLINK_CB(skb).portid);
1294 }
1295
1296 if (IS_ERR(table)) {
1297 NL_SET_BAD_ATTR(extack, attr);
1298 return PTR_ERR(table);
1299 }
1300
1301 if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
1302 table->use > 0)
1303 return -EBUSY;
1304
1305 ctx.family = family;
1306 ctx.table = table;
1307
1308 return nft_flush_table(&ctx);
1309 }
1310
1311 static void nf_tables_table_destroy(struct nft_ctx *ctx)
1312 {
1313 if (WARN_ON(ctx->table->use > 0))
1314 return;
1315
1316 rhltable_destroy(&ctx->table->chains_ht);
1317 kfree(ctx->table->name);
1318 kfree(ctx->table->udata);
1319 kfree(ctx->table);
1320 }
1321
1322 void nft_register_chain_type(const struct nft_chain_type *ctype)
1323 {
1324 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1325 if (WARN_ON(__nft_chain_type_get(ctype->family, ctype->type))) {
1326 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1327 return;
1328 }
1329 chain_type[ctype->family][ctype->type] = ctype;
1330 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1331 }
1332 EXPORT_SYMBOL_GPL(nft_register_chain_type);
1333
1334 void nft_unregister_chain_type(const struct nft_chain_type *ctype)
1335 {
1336 nfnl_lock(NFNL_SUBSYS_NFTABLES);
1337 chain_type[ctype->family][ctype->type] = NULL;
1338 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1339 }
1340 EXPORT_SYMBOL_GPL(nft_unregister_chain_type);
1341
1342 /*
1343 * Chains
1344 */
1345
1346 static struct nft_chain *
1347 nft_chain_lookup_byhandle(const struct nft_table *table, u64 handle, u8 genmask)
1348 {
1349 struct nft_chain *chain;
1350
1351 list_for_each_entry(chain, &table->chains, list) {
1352 if (chain->handle == handle &&
1353 nft_active_genmask(chain, genmask))
1354 return chain;
1355 }
1356
1357 return ERR_PTR(-ENOENT);
1358 }
1359
1360 static bool lockdep_commit_lock_is_held(const struct net *net)
1361 {
1362 #ifdef CONFIG_PROVE_LOCKING
1363 struct nftables_pernet *nft_net = nft_pernet(net);
1364
1365 return lockdep_is_held(&nft_net->commit_mutex);
1366 #else
1367 return true;
1368 #endif
1369 }
1370
1371 static struct nft_chain *nft_chain_lookup(struct net *net,
1372 struct nft_table *table,
1373 const struct nlattr *nla, u8 genmask)
1374 {
1375 char search[NFT_CHAIN_MAXNAMELEN + 1];
1376 struct rhlist_head *tmp, *list;
1377 struct nft_chain *chain;
1378
1379 if (nla == NULL)
1380 return ERR_PTR(-EINVAL);
1381
1382 nla_strscpy(search, nla, sizeof(search));
1383
1384 WARN_ON(!rcu_read_lock_held() &&
1385 !lockdep_commit_lock_is_held(net));
1386
1387 chain = ERR_PTR(-ENOENT);
1388 rcu_read_lock();
1389 list = rhltable_lookup(&table->chains_ht, search, nft_chain_ht_params);
1390 if (!list)
1391 goto out_unlock;
1392
1393 rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
1394 if (nft_active_genmask(chain, genmask))
1395 goto out_unlock;
1396 }
1397 chain = ERR_PTR(-ENOENT);
1398 out_unlock:
1399 rcu_read_unlock();
1400 return chain;
1401 }
1402
1403 static const struct nla_policy nft_chain_policy[NFTA_CHAIN_MAX + 1] = {
1404 [NFTA_CHAIN_TABLE] = { .type = NLA_STRING,
1405 .len = NFT_TABLE_MAXNAMELEN - 1 },
1406 [NFTA_CHAIN_HANDLE] = { .type = NLA_U64 },
1407 [NFTA_CHAIN_NAME] = { .type = NLA_STRING,
1408 .len = NFT_CHAIN_MAXNAMELEN - 1 },
1409 [NFTA_CHAIN_HOOK] = { .type = NLA_NESTED },
1410 [NFTA_CHAIN_POLICY] = { .type = NLA_U32 },
1411 [NFTA_CHAIN_TYPE] = { .type = NLA_STRING,
1412 .len = NFT_MODULE_AUTOLOAD_LIMIT },
1413 [NFTA_CHAIN_COUNTERS] = { .type = NLA_NESTED },
1414 [NFTA_CHAIN_FLAGS] = { .type = NLA_U32 },
1415 [NFTA_CHAIN_ID] = { .type = NLA_U32 },
1416 [NFTA_CHAIN_USERDATA] = { .type = NLA_BINARY,
1417 .len = NFT_USERDATA_MAXLEN },
1418 };
1419
1420 static const struct nla_policy nft_hook_policy[NFTA_HOOK_MAX + 1] = {
1421 [NFTA_HOOK_HOOKNUM] = { .type = NLA_U32 },
1422 [NFTA_HOOK_PRIORITY] = { .type = NLA_U32 },
1423 [NFTA_HOOK_DEV] = { .type = NLA_STRING,
1424 .len = IFNAMSIZ - 1 },
1425 };
1426
1427 static int nft_dump_stats(struct sk_buff *skb, struct nft_stats __percpu *stats)
1428 {
1429 struct nft_stats *cpu_stats, total;
1430 struct nlattr *nest;
1431 unsigned int seq;
1432 u64 pkts, bytes;
1433 int cpu;
1434
1435 if (!stats)
1436 return 0;
1437
1438 memset(&total, 0, sizeof(total));
1439 for_each_possible_cpu(cpu) {
1440 cpu_stats = per_cpu_ptr(stats, cpu);
1441 do {
1442 seq = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
1443 pkts = cpu_stats->pkts;
1444 bytes = cpu_stats->bytes;
1445 } while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, seq));
1446 total.pkts += pkts;
1447 total.bytes += bytes;
1448 }
1449 nest = nla_nest_start_noflag(skb, NFTA_CHAIN_COUNTERS);
1450 if (nest == NULL)
1451 goto nla_put_failure;
1452
1453 if (nla_put_be64(skb, NFTA_COUNTER_PACKETS, cpu_to_be64(total.pkts),
1454 NFTA_COUNTER_PAD) ||
1455 nla_put_be64(skb, NFTA_COUNTER_BYTES, cpu_to_be64(total.bytes),
1456 NFTA_COUNTER_PAD))
1457 goto nla_put_failure;
1458
1459 nla_nest_end(skb, nest);
1460 return 0;
1461
1462 nla_put_failure:
1463 return -ENOSPC;
1464 }
1465
1466 static int nft_dump_basechain_hook(struct sk_buff *skb, int family,
1467 const struct nft_base_chain *basechain)
1468 {
1469 const struct nf_hook_ops *ops = &basechain->ops;
1470 struct nft_hook *hook, *first = NULL;
1471 struct nlattr *nest, *nest_devs;
1472 int n = 0;
1473
1474 nest = nla_nest_start_noflag(skb, NFTA_CHAIN_HOOK);
1475 if (nest == NULL)
1476 goto nla_put_failure;
1477 if (nla_put_be32(skb, NFTA_HOOK_HOOKNUM, htonl(ops->hooknum)))
1478 goto nla_put_failure;
1479 if (nla_put_be32(skb, NFTA_HOOK_PRIORITY, htonl(ops->priority)))
1480 goto nla_put_failure;
1481
1482 if (nft_base_chain_netdev(family, ops->hooknum)) {
1483 nest_devs = nla_nest_start_noflag(skb, NFTA_HOOK_DEVS);
1484 list_for_each_entry(hook, &basechain->hook_list, list) {
1485 if (!first)
1486 first = hook;
1487
1488 if (nla_put_string(skb, NFTA_DEVICE_NAME,
1489 hook->ops.dev->name))
1490 goto nla_put_failure;
1491 n++;
1492 }
1493 nla_nest_end(skb, nest_devs);
1494
1495 if (n == 1 &&
1496 nla_put_string(skb, NFTA_HOOK_DEV, first->ops.dev->name))
1497 goto nla_put_failure;
1498 }
1499 nla_nest_end(skb, nest);
1500
1501 return 0;
1502 nla_put_failure:
1503 return -1;
1504 }
1505
1506 static int nf_tables_fill_chain_info(struct sk_buff *skb, struct net *net,
1507 u32 portid, u32 seq, int event, u32 flags,
1508 int family, const struct nft_table *table,
1509 const struct nft_chain *chain)
1510 {
1511 struct nlmsghdr *nlh;
1512
1513 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
1514 nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
1515 NFNETLINK_V0, nft_base_seq(net));
1516 if (!nlh)
1517 goto nla_put_failure;
1518
1519 if (nla_put_string(skb, NFTA_CHAIN_TABLE, table->name))
1520 goto nla_put_failure;
1521 if (nla_put_be64(skb, NFTA_CHAIN_HANDLE, cpu_to_be64(chain->handle),
1522 NFTA_CHAIN_PAD))
1523 goto nla_put_failure;
1524 if (nla_put_string(skb, NFTA_CHAIN_NAME, chain->name))
1525 goto nla_put_failure;
1526
1527 if (nft_is_base_chain(chain)) {
1528 const struct nft_base_chain *basechain = nft_base_chain(chain);
1529 struct nft_stats __percpu *stats;
1530
1531 if (nft_dump_basechain_hook(skb, family, basechain))
1532 goto nla_put_failure;
1533
1534 if (nla_put_be32(skb, NFTA_CHAIN_POLICY,
1535 htonl(basechain->policy)))
1536 goto nla_put_failure;
1537
1538 if (nla_put_string(skb, NFTA_CHAIN_TYPE, basechain->type->name))
1539 goto nla_put_failure;
1540
1541 stats = rcu_dereference_check(basechain->stats,
1542 lockdep_commit_lock_is_held(net));
1543 if (nft_dump_stats(skb, stats))
1544 goto nla_put_failure;
1545 }
1546
1547 if (chain->flags &&
1548 nla_put_be32(skb, NFTA_CHAIN_FLAGS, htonl(chain->flags)))
1549 goto nla_put_failure;
1550
1551 if (nla_put_be32(skb, NFTA_CHAIN_USE, htonl(chain->use)))
1552 goto nla_put_failure;
1553
1554 if (chain->udata &&
1555 nla_put(skb, NFTA_CHAIN_USERDATA, chain->udlen, chain->udata))
1556 goto nla_put_failure;
1557
1558 nlmsg_end(skb, nlh);
1559 return 0;
1560
1561 nla_put_failure:
1562 nlmsg_trim(skb, nlh);
1563 return -1;
1564 }
1565
1566 static void nf_tables_chain_notify(const struct nft_ctx *ctx, int event)
1567 {
1568 struct nftables_pernet *nft_net;
1569 struct sk_buff *skb;
1570 u16 flags = 0;
1571 int err;
1572
1573 if (!ctx->report &&
1574 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1575 return;
1576
1577 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1578 if (skb == NULL)
1579 goto err;
1580
1581 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
1582 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
1583
1584 err = nf_tables_fill_chain_info(skb, ctx->net, ctx->portid, ctx->seq,
1585 event, flags, ctx->family, ctx->table,
1586 ctx->chain);
1587 if (err < 0) {
1588 kfree_skb(skb);
1589 goto err;
1590 }
1591
1592 nft_net = nft_pernet(ctx->net);
1593 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
1594 return;
1595 err:
1596 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1597 }
1598
1599 static int nf_tables_dump_chains(struct sk_buff *skb,
1600 struct netlink_callback *cb)
1601 {
1602 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1603 unsigned int idx = 0, s_idx = cb->args[0];
1604 struct net *net = sock_net(skb->sk);
1605 int family = nfmsg->nfgen_family;
1606 struct nftables_pernet *nft_net;
1607 const struct nft_table *table;
1608 const struct nft_chain *chain;
1609
1610 rcu_read_lock();
1611 nft_net = nft_pernet(net);
1612 cb->seq = nft_net->base_seq;
1613
1614 list_for_each_entry_rcu(table, &nft_net->tables, list) {
1615 if (family != NFPROTO_UNSPEC && family != table->family)
1616 continue;
1617
1618 list_for_each_entry_rcu(chain, &table->chains, list) {
1619 if (idx < s_idx)
1620 goto cont;
1621 if (idx > s_idx)
1622 memset(&cb->args[1], 0,
1623 sizeof(cb->args) - sizeof(cb->args[0]));
1624 if (!nft_is_active(net, chain))
1625 continue;
1626 if (nf_tables_fill_chain_info(skb, net,
1627 NETLINK_CB(cb->skb).portid,
1628 cb->nlh->nlmsg_seq,
1629 NFT_MSG_NEWCHAIN,
1630 NLM_F_MULTI,
1631 table->family, table,
1632 chain) < 0)
1633 goto done;
1634
1635 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1636 cont:
1637 idx++;
1638 }
1639 }
1640 done:
1641 rcu_read_unlock();
1642 cb->args[0] = idx;
1643 return skb->len;
1644 }
1645
1646 /* called with rcu_read_lock held */
1647 static int nf_tables_getchain(struct sk_buff *skb, const struct nfnl_info *info,
1648 const struct nlattr * const nla[])
1649 {
1650 struct netlink_ext_ack *extack = info->extack;
1651 u8 genmask = nft_genmask_cur(info->net);
1652 u8 family = info->nfmsg->nfgen_family;
1653 const struct nft_chain *chain;
1654 struct net *net = info->net;
1655 struct nft_table *table;
1656 struct sk_buff *skb2;
1657 int err;
1658
1659 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1660 struct netlink_dump_control c = {
1661 .dump = nf_tables_dump_chains,
1662 .module = THIS_MODULE,
1663 };
1664
1665 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
1666 }
1667
1668 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask, 0);
1669 if (IS_ERR(table)) {
1670 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
1671 return PTR_ERR(table);
1672 }
1673
1674 chain = nft_chain_lookup(net, table, nla[NFTA_CHAIN_NAME], genmask);
1675 if (IS_ERR(chain)) {
1676 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
1677 return PTR_ERR(chain);
1678 }
1679
1680 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1681 if (!skb2)
1682 return -ENOMEM;
1683
1684 err = nf_tables_fill_chain_info(skb2, net, NETLINK_CB(skb).portid,
1685 info->nlh->nlmsg_seq, NFT_MSG_NEWCHAIN,
1686 0, family, table, chain);
1687 if (err < 0)
1688 goto err_fill_chain_info;
1689
1690 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
1691
1692 err_fill_chain_info:
1693 kfree_skb(skb2);
1694 return err;
1695 }
1696
1697 static const struct nla_policy nft_counter_policy[NFTA_COUNTER_MAX + 1] = {
1698 [NFTA_COUNTER_PACKETS] = { .type = NLA_U64 },
1699 [NFTA_COUNTER_BYTES] = { .type = NLA_U64 },
1700 };
1701
1702 static struct nft_stats __percpu *nft_stats_alloc(const struct nlattr *attr)
1703 {
1704 struct nlattr *tb[NFTA_COUNTER_MAX+1];
1705 struct nft_stats __percpu *newstats;
1706 struct nft_stats *stats;
1707 int err;
1708
1709 err = nla_parse_nested_deprecated(tb, NFTA_COUNTER_MAX, attr,
1710 nft_counter_policy, NULL);
1711 if (err < 0)
1712 return ERR_PTR(err);
1713
1714 if (!tb[NFTA_COUNTER_BYTES] || !tb[NFTA_COUNTER_PACKETS])
1715 return ERR_PTR(-EINVAL);
1716
1717 newstats = netdev_alloc_pcpu_stats(struct nft_stats);
1718 if (newstats == NULL)
1719 return ERR_PTR(-ENOMEM);
1720
1721 /* Restore old counters on this cpu, no problem. Per-cpu statistics
1722 * are not exposed to userspace.
1723 */
1724 preempt_disable();
1725 stats = this_cpu_ptr(newstats);
1726 stats->bytes = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_BYTES]));
1727 stats->pkts = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_PACKETS]));
1728 preempt_enable();
1729
1730 return newstats;
1731 }
1732
1733 static void nft_chain_stats_replace(struct nft_trans *trans)
1734 {
1735 struct nft_base_chain *chain = nft_base_chain(trans->ctx.chain);
1736
1737 if (!nft_trans_chain_stats(trans))
1738 return;
1739
1740 nft_trans_chain_stats(trans) =
1741 rcu_replace_pointer(chain->stats, nft_trans_chain_stats(trans),
1742 lockdep_commit_lock_is_held(trans->ctx.net));
1743
1744 if (!nft_trans_chain_stats(trans))
1745 static_branch_inc(&nft_counters_enabled);
1746 }
1747
1748 static void nf_tables_chain_free_chain_rules(struct nft_chain *chain)
1749 {
1750 struct nft_rule **g0 = rcu_dereference_raw(chain->rules_gen_0);
1751 struct nft_rule **g1 = rcu_dereference_raw(chain->rules_gen_1);
1752
1753 if (g0 != g1)
1754 kvfree(g1);
1755 kvfree(g0);
1756
1757 /* should be NULL either via abort or via successful commit */
1758 WARN_ON_ONCE(chain->rules_next);
1759 kvfree(chain->rules_next);
1760 }
1761
1762 void nf_tables_chain_destroy(struct nft_ctx *ctx)
1763 {
1764 struct nft_chain *chain = ctx->chain;
1765 struct nft_hook *hook, *next;
1766
1767 if (WARN_ON(chain->use > 0))
1768 return;
1769
1770 /* no concurrent access possible anymore */
1771 nf_tables_chain_free_chain_rules(chain);
1772
1773 if (nft_is_base_chain(chain)) {
1774 struct nft_base_chain *basechain = nft_base_chain(chain);
1775
1776 if (nft_base_chain_netdev(ctx->family, basechain->ops.hooknum)) {
1777 list_for_each_entry_safe(hook, next,
1778 &basechain->hook_list, list) {
1779 list_del_rcu(&hook->list);
1780 kfree_rcu(hook, rcu);
1781 }
1782 }
1783 module_put(basechain->type->owner);
1784 if (rcu_access_pointer(basechain->stats)) {
1785 static_branch_dec(&nft_counters_enabled);
1786 free_percpu(rcu_dereference_raw(basechain->stats));
1787 }
1788 kfree(chain->name);
1789 kfree(chain->udata);
1790 kfree(basechain);
1791 } else {
1792 kfree(chain->name);
1793 kfree(chain->udata);
1794 kfree(chain);
1795 }
1796 }
1797
1798 static struct nft_hook *nft_netdev_hook_alloc(struct net *net,
1799 const struct nlattr *attr)
1800 {
1801 struct net_device *dev;
1802 char ifname[IFNAMSIZ];
1803 struct nft_hook *hook;
1804 int err;
1805
1806 hook = kmalloc(sizeof(struct nft_hook), GFP_KERNEL);
1807 if (!hook) {
1808 err = -ENOMEM;
1809 goto err_hook_alloc;
1810 }
1811
1812 nla_strscpy(ifname, attr, IFNAMSIZ);
1813 /* nf_tables_netdev_event() is called under rtnl_mutex, this is
1814 * indirectly serializing all the other holders of the commit_mutex with
1815 * the rtnl_mutex.
1816 */
1817 dev = __dev_get_by_name(net, ifname);
1818 if (!dev) {
1819 err = -ENOENT;
1820 goto err_hook_dev;
1821 }
1822 hook->ops.dev = dev;
1823 hook->inactive = false;
1824
1825 return hook;
1826
1827 err_hook_dev:
1828 kfree(hook);
1829 err_hook_alloc:
1830 return ERR_PTR(err);
1831 }
1832
1833 static struct nft_hook *nft_hook_list_find(struct list_head *hook_list,
1834 const struct nft_hook *this)
1835 {
1836 struct nft_hook *hook;
1837
1838 list_for_each_entry(hook, hook_list, list) {
1839 if (this->ops.dev == hook->ops.dev)
1840 return hook;
1841 }
1842
1843 return NULL;
1844 }
1845
1846 static int nf_tables_parse_netdev_hooks(struct net *net,
1847 const struct nlattr *attr,
1848 struct list_head *hook_list)
1849 {
1850 struct nft_hook *hook, *next;
1851 const struct nlattr *tmp;
1852 int rem, n = 0, err;
1853
1854 nla_for_each_nested(tmp, attr, rem) {
1855 if (nla_type(tmp) != NFTA_DEVICE_NAME) {
1856 err = -EINVAL;
1857 goto err_hook;
1858 }
1859
1860 hook = nft_netdev_hook_alloc(net, tmp);
1861 if (IS_ERR(hook)) {
1862 err = PTR_ERR(hook);
1863 goto err_hook;
1864 }
1865 if (nft_hook_list_find(hook_list, hook)) {
1866 kfree(hook);
1867 err = -EEXIST;
1868 goto err_hook;
1869 }
1870 list_add_tail(&hook->list, hook_list);
1871 n++;
1872
1873 if (n == NFT_NETDEVICE_MAX) {
1874 err = -EFBIG;
1875 goto err_hook;
1876 }
1877 }
1878
1879 return 0;
1880
1881 err_hook:
1882 list_for_each_entry_safe(hook, next, hook_list, list) {
1883 list_del(&hook->list);
1884 kfree(hook);
1885 }
1886 return err;
1887 }
1888
1889 struct nft_chain_hook {
1890 u32 num;
1891 s32 priority;
1892 const struct nft_chain_type *type;
1893 struct list_head list;
1894 };
1895
1896 static int nft_chain_parse_netdev(struct net *net,
1897 struct nlattr *tb[],
1898 struct list_head *hook_list)
1899 {
1900 struct nft_hook *hook;
1901 int err;
1902
1903 if (tb[NFTA_HOOK_DEV]) {
1904 hook = nft_netdev_hook_alloc(net, tb[NFTA_HOOK_DEV]);
1905 if (IS_ERR(hook))
1906 return PTR_ERR(hook);
1907
1908 list_add_tail(&hook->list, hook_list);
1909 } else if (tb[NFTA_HOOK_DEVS]) {
1910 err = nf_tables_parse_netdev_hooks(net, tb[NFTA_HOOK_DEVS],
1911 hook_list);
1912 if (err < 0)
1913 return err;
1914
1915 if (list_empty(hook_list))
1916 return -EINVAL;
1917 } else {
1918 return -EINVAL;
1919 }
1920
1921 return 0;
1922 }
1923
1924 static int nft_chain_parse_hook(struct net *net,
1925 const struct nlattr * const nla[],
1926 struct nft_chain_hook *hook, u8 family,
1927 struct netlink_ext_ack *extack, bool autoload)
1928 {
1929 struct nftables_pernet *nft_net = nft_pernet(net);
1930 struct nlattr *ha[NFTA_HOOK_MAX + 1];
1931 const struct nft_chain_type *type;
1932 int err;
1933
1934 lockdep_assert_held(&nft_net->commit_mutex);
1935 lockdep_nfnl_nft_mutex_not_held();
1936
1937 err = nla_parse_nested_deprecated(ha, NFTA_HOOK_MAX,
1938 nla[NFTA_CHAIN_HOOK],
1939 nft_hook_policy, NULL);
1940 if (err < 0)
1941 return err;
1942
1943 if (ha[NFTA_HOOK_HOOKNUM] == NULL ||
1944 ha[NFTA_HOOK_PRIORITY] == NULL)
1945 return -EINVAL;
1946
1947 hook->num = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
1948 hook->priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
1949
1950 type = __nft_chain_type_get(family, NFT_CHAIN_T_DEFAULT);
1951 if (!type)
1952 return -EOPNOTSUPP;
1953
1954 if (nla[NFTA_CHAIN_TYPE]) {
1955 type = nf_tables_chain_type_lookup(net, nla[NFTA_CHAIN_TYPE],
1956 family, autoload);
1957 if (IS_ERR(type)) {
1958 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
1959 return PTR_ERR(type);
1960 }
1961 }
1962 if (hook->num >= NFT_MAX_HOOKS || !(type->hook_mask & (1 << hook->num)))
1963 return -EOPNOTSUPP;
1964
1965 if (type->type == NFT_CHAIN_T_NAT &&
1966 hook->priority <= NF_IP_PRI_CONNTRACK)
1967 return -EOPNOTSUPP;
1968
1969 if (!try_module_get(type->owner)) {
1970 if (nla[NFTA_CHAIN_TYPE])
1971 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
1972 return -ENOENT;
1973 }
1974
1975 hook->type = type;
1976
1977 INIT_LIST_HEAD(&hook->list);
1978 if (nft_base_chain_netdev(family, hook->num)) {
1979 err = nft_chain_parse_netdev(net, ha, &hook->list);
1980 if (err < 0) {
1981 module_put(type->owner);
1982 return err;
1983 }
1984 } else if (ha[NFTA_HOOK_DEV] || ha[NFTA_HOOK_DEVS]) {
1985 module_put(type->owner);
1986 return -EOPNOTSUPP;
1987 }
1988
1989 return 0;
1990 }
1991
1992 static void nft_chain_release_hook(struct nft_chain_hook *hook)
1993 {
1994 struct nft_hook *h, *next;
1995
1996 list_for_each_entry_safe(h, next, &hook->list, list) {
1997 list_del(&h->list);
1998 kfree(h);
1999 }
2000 module_put(hook->type->owner);
2001 }
2002
2003 struct nft_rules_old {
2004 struct rcu_head h;
2005 struct nft_rule **start;
2006 };
2007
2008 static struct nft_rule **nf_tables_chain_alloc_rules(const struct nft_chain *chain,
2009 unsigned int alloc)
2010 {
2011 if (alloc > INT_MAX)
2012 return NULL;
2013
2014 alloc += 1; /* NULL, ends rules */
2015 if (sizeof(struct nft_rule *) > INT_MAX / alloc)
2016 return NULL;
2017
2018 alloc *= sizeof(struct nft_rule *);
2019 alloc += sizeof(struct nft_rules_old);
2020
2021 return kvmalloc(alloc, GFP_KERNEL);
2022 }
2023
2024 static void nft_basechain_hook_init(struct nf_hook_ops *ops, u8 family,
2025 const struct nft_chain_hook *hook,
2026 struct nft_chain *chain)
2027 {
2028 ops->pf = family;
2029 ops->hooknum = hook->num;
2030 ops->priority = hook->priority;
2031 ops->priv = chain;
2032 ops->hook = hook->type->hooks[ops->hooknum];
2033 ops->hook_ops_type = NF_HOOK_OP_NF_TABLES;
2034 }
2035
2036 static int nft_basechain_init(struct nft_base_chain *basechain, u8 family,
2037 struct nft_chain_hook *hook, u32 flags)
2038 {
2039 struct nft_chain *chain;
2040 struct nft_hook *h;
2041
2042 basechain->type = hook->type;
2043 INIT_LIST_HEAD(&basechain->hook_list);
2044 chain = &basechain->chain;
2045
2046 if (nft_base_chain_netdev(family, hook->num)) {
2047 list_splice_init(&hook->list, &basechain->hook_list);
2048 list_for_each_entry(h, &basechain->hook_list, list)
2049 nft_basechain_hook_init(&h->ops, family, hook, chain);
2050
2051 basechain->ops.hooknum = hook->num;
2052 basechain->ops.priority = hook->priority;
2053 } else {
2054 nft_basechain_hook_init(&basechain->ops, family, hook, chain);
2055 }
2056
2057 chain->flags |= NFT_CHAIN_BASE | flags;
2058 basechain->policy = NF_ACCEPT;
2059 if (chain->flags & NFT_CHAIN_HW_OFFLOAD &&
2060 nft_chain_offload_priority(basechain) < 0)
2061 return -EOPNOTSUPP;
2062
2063 flow_block_init(&basechain->flow_block);
2064
2065 return 0;
2066 }
2067
2068 static int nft_chain_add(struct nft_table *table, struct nft_chain *chain)
2069 {
2070 int err;
2071
2072 err = rhltable_insert_key(&table->chains_ht, chain->name,
2073 &chain->rhlhead, nft_chain_ht_params);
2074 if (err)
2075 return err;
2076
2077 list_add_tail_rcu(&chain->list, &table->chains);
2078
2079 return 0;
2080 }
2081
2082 static u64 chain_id;
2083
2084 static int nf_tables_addchain(struct nft_ctx *ctx, u8 family, u8 genmask,
2085 u8 policy, u32 flags,
2086 struct netlink_ext_ack *extack)
2087 {
2088 const struct nlattr * const *nla = ctx->nla;
2089 struct nft_table *table = ctx->table;
2090 struct nft_base_chain *basechain;
2091 struct nft_stats __percpu *stats;
2092 struct net *net = ctx->net;
2093 char name[NFT_NAME_MAXLEN];
2094 struct nft_trans *trans;
2095 struct nft_chain *chain;
2096 struct nft_rule **rules;
2097 int err;
2098
2099 if (table->use == UINT_MAX)
2100 return -EOVERFLOW;
2101
2102 if (nla[NFTA_CHAIN_HOOK]) {
2103 struct nft_chain_hook hook;
2104
2105 if (flags & NFT_CHAIN_BINDING)
2106 return -EOPNOTSUPP;
2107
2108 err = nft_chain_parse_hook(net, nla, &hook, family, extack,
2109 true);
2110 if (err < 0)
2111 return err;
2112
2113 basechain = kzalloc(sizeof(*basechain), GFP_KERNEL);
2114 if (basechain == NULL) {
2115 nft_chain_release_hook(&hook);
2116 return -ENOMEM;
2117 }
2118 chain = &basechain->chain;
2119
2120 if (nla[NFTA_CHAIN_COUNTERS]) {
2121 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2122 if (IS_ERR(stats)) {
2123 nft_chain_release_hook(&hook);
2124 kfree(basechain);
2125 return PTR_ERR(stats);
2126 }
2127 rcu_assign_pointer(basechain->stats, stats);
2128 static_branch_inc(&nft_counters_enabled);
2129 }
2130
2131 err = nft_basechain_init(basechain, family, &hook, flags);
2132 if (err < 0) {
2133 nft_chain_release_hook(&hook);
2134 kfree(basechain);
2135 return err;
2136 }
2137 } else {
2138 if (flags & NFT_CHAIN_BASE)
2139 return -EINVAL;
2140 if (flags & NFT_CHAIN_HW_OFFLOAD)
2141 return -EOPNOTSUPP;
2142
2143 chain = kzalloc(sizeof(*chain), GFP_KERNEL);
2144 if (chain == NULL)
2145 return -ENOMEM;
2146
2147 chain->flags = flags;
2148 }
2149 ctx->chain = chain;
2150
2151 INIT_LIST_HEAD(&chain->rules);
2152 chain->handle = nf_tables_alloc_handle(table);
2153 chain->table = table;
2154
2155 if (nla[NFTA_CHAIN_NAME]) {
2156 chain->name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL);
2157 } else {
2158 if (!(flags & NFT_CHAIN_BINDING)) {
2159 err = -EINVAL;
2160 goto err_destroy_chain;
2161 }
2162
2163 snprintf(name, sizeof(name), "__chain%llu", ++chain_id);
2164 chain->name = kstrdup(name, GFP_KERNEL);
2165 }
2166
2167 if (!chain->name) {
2168 err = -ENOMEM;
2169 goto err_destroy_chain;
2170 }
2171
2172 if (nla[NFTA_CHAIN_USERDATA]) {
2173 chain->udata = nla_memdup(nla[NFTA_CHAIN_USERDATA], GFP_KERNEL);
2174 if (chain->udata == NULL) {
2175 err = -ENOMEM;
2176 goto err_destroy_chain;
2177 }
2178 chain->udlen = nla_len(nla[NFTA_CHAIN_USERDATA]);
2179 }
2180
2181 rules = nf_tables_chain_alloc_rules(chain, 0);
2182 if (!rules) {
2183 err = -ENOMEM;
2184 goto err_destroy_chain;
2185 }
2186
2187 *rules = NULL;
2188 rcu_assign_pointer(chain->rules_gen_0, rules);
2189 rcu_assign_pointer(chain->rules_gen_1, rules);
2190
2191 err = nf_tables_register_hook(net, table, chain);
2192 if (err < 0)
2193 goto err_destroy_chain;
2194
2195 trans = nft_trans_chain_add(ctx, NFT_MSG_NEWCHAIN);
2196 if (IS_ERR(trans)) {
2197 err = PTR_ERR(trans);
2198 goto err_unregister_hook;
2199 }
2200
2201 nft_trans_chain_policy(trans) = NFT_CHAIN_POLICY_UNSET;
2202 if (nft_is_base_chain(chain))
2203 nft_trans_chain_policy(trans) = policy;
2204
2205 err = nft_chain_add(table, chain);
2206 if (err < 0) {
2207 nft_trans_destroy(trans);
2208 goto err_unregister_hook;
2209 }
2210
2211 table->use++;
2212
2213 return 0;
2214 err_unregister_hook:
2215 nf_tables_unregister_hook(net, table, chain);
2216 err_destroy_chain:
2217 nf_tables_chain_destroy(ctx);
2218
2219 return err;
2220 }
2221
2222 static bool nft_hook_list_equal(struct list_head *hook_list1,
2223 struct list_head *hook_list2)
2224 {
2225 struct nft_hook *hook;
2226 int n = 0, m = 0;
2227
2228 n = 0;
2229 list_for_each_entry(hook, hook_list2, list) {
2230 if (!nft_hook_list_find(hook_list1, hook))
2231 return false;
2232
2233 n++;
2234 }
2235 list_for_each_entry(hook, hook_list1, list)
2236 m++;
2237
2238 return n == m;
2239 }
2240
2241 static int nf_tables_updchain(struct nft_ctx *ctx, u8 genmask, u8 policy,
2242 u32 flags, const struct nlattr *attr,
2243 struct netlink_ext_ack *extack)
2244 {
2245 const struct nlattr * const *nla = ctx->nla;
2246 struct nft_table *table = ctx->table;
2247 struct nft_chain *chain = ctx->chain;
2248 struct nft_base_chain *basechain;
2249 struct nft_stats *stats = NULL;
2250 struct nft_chain_hook hook;
2251 struct nf_hook_ops *ops;
2252 struct nft_trans *trans;
2253 int err;
2254
2255 if (chain->flags ^ flags)
2256 return -EOPNOTSUPP;
2257
2258 if (nla[NFTA_CHAIN_HOOK]) {
2259 if (!nft_is_base_chain(chain)) {
2260 NL_SET_BAD_ATTR(extack, attr);
2261 return -EEXIST;
2262 }
2263 err = nft_chain_parse_hook(ctx->net, nla, &hook, ctx->family,
2264 extack, false);
2265 if (err < 0)
2266 return err;
2267
2268 basechain = nft_base_chain(chain);
2269 if (basechain->type != hook.type) {
2270 nft_chain_release_hook(&hook);
2271 NL_SET_BAD_ATTR(extack, attr);
2272 return -EEXIST;
2273 }
2274
2275 if (nft_base_chain_netdev(ctx->family, hook.num)) {
2276 if (!nft_hook_list_equal(&basechain->hook_list,
2277 &hook.list)) {
2278 nft_chain_release_hook(&hook);
2279 NL_SET_BAD_ATTR(extack, attr);
2280 return -EEXIST;
2281 }
2282 } else {
2283 ops = &basechain->ops;
2284 if (ops->hooknum != hook.num ||
2285 ops->priority != hook.priority) {
2286 nft_chain_release_hook(&hook);
2287 NL_SET_BAD_ATTR(extack, attr);
2288 return -EEXIST;
2289 }
2290 }
2291 nft_chain_release_hook(&hook);
2292 }
2293
2294 if (nla[NFTA_CHAIN_HANDLE] &&
2295 nla[NFTA_CHAIN_NAME]) {
2296 struct nft_chain *chain2;
2297
2298 chain2 = nft_chain_lookup(ctx->net, table,
2299 nla[NFTA_CHAIN_NAME], genmask);
2300 if (!IS_ERR(chain2)) {
2301 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2302 return -EEXIST;
2303 }
2304 }
2305
2306 if (nla[NFTA_CHAIN_COUNTERS]) {
2307 if (!nft_is_base_chain(chain))
2308 return -EOPNOTSUPP;
2309
2310 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2311 if (IS_ERR(stats))
2312 return PTR_ERR(stats);
2313 }
2314
2315 err = -ENOMEM;
2316 trans = nft_trans_alloc(ctx, NFT_MSG_NEWCHAIN,
2317 sizeof(struct nft_trans_chain));
2318 if (trans == NULL)
2319 goto err;
2320
2321 nft_trans_chain_stats(trans) = stats;
2322 nft_trans_chain_update(trans) = true;
2323
2324 if (nla[NFTA_CHAIN_POLICY])
2325 nft_trans_chain_policy(trans) = policy;
2326 else
2327 nft_trans_chain_policy(trans) = -1;
2328
2329 if (nla[NFTA_CHAIN_HANDLE] &&
2330 nla[NFTA_CHAIN_NAME]) {
2331 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
2332 struct nft_trans *tmp;
2333 char *name;
2334
2335 err = -ENOMEM;
2336 name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL);
2337 if (!name)
2338 goto err;
2339
2340 err = -EEXIST;
2341 list_for_each_entry(tmp, &nft_net->commit_list, list) {
2342 if (tmp->msg_type == NFT_MSG_NEWCHAIN &&
2343 tmp->ctx.table == table &&
2344 nft_trans_chain_update(tmp) &&
2345 nft_trans_chain_name(tmp) &&
2346 strcmp(name, nft_trans_chain_name(tmp)) == 0) {
2347 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2348 kfree(name);
2349 goto err;
2350 }
2351 }
2352
2353 nft_trans_chain_name(trans) = name;
2354 }
2355 nft_trans_commit_list_add_tail(ctx->net, trans);
2356
2357 return 0;
2358 err:
2359 free_percpu(stats);
2360 kfree(trans);
2361 return err;
2362 }
2363
2364 static struct nft_chain *nft_chain_lookup_byid(const struct net *net,
2365 const struct nlattr *nla)
2366 {
2367 struct nftables_pernet *nft_net = nft_pernet(net);
2368 u32 id = ntohl(nla_get_be32(nla));
2369 struct nft_trans *trans;
2370
2371 list_for_each_entry(trans, &nft_net->commit_list, list) {
2372 struct nft_chain *chain = trans->ctx.chain;
2373
2374 if (trans->msg_type == NFT_MSG_NEWCHAIN &&
2375 id == nft_trans_chain_id(trans))
2376 return chain;
2377 }
2378 return ERR_PTR(-ENOENT);
2379 }
2380
2381 static int nf_tables_newchain(struct sk_buff *skb, const struct nfnl_info *info,
2382 const struct nlattr * const nla[])
2383 {
2384 struct nftables_pernet *nft_net = nft_pernet(info->net);
2385 struct netlink_ext_ack *extack = info->extack;
2386 u8 genmask = nft_genmask_next(info->net);
2387 u8 family = info->nfmsg->nfgen_family;
2388 struct nft_chain *chain = NULL;
2389 struct net *net = info->net;
2390 const struct nlattr *attr;
2391 struct nft_table *table;
2392 u8 policy = NF_ACCEPT;
2393 struct nft_ctx ctx;
2394 u64 handle = 0;
2395 u32 flags = 0;
2396
2397 lockdep_assert_held(&nft_net->commit_mutex);
2398
2399 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2400 NETLINK_CB(skb).portid);
2401 if (IS_ERR(table)) {
2402 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2403 return PTR_ERR(table);
2404 }
2405
2406 chain = NULL;
2407 attr = nla[NFTA_CHAIN_NAME];
2408
2409 if (nla[NFTA_CHAIN_HANDLE]) {
2410 handle = be64_to_cpu(nla_get_be64(nla[NFTA_CHAIN_HANDLE]));
2411 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2412 if (IS_ERR(chain)) {
2413 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_HANDLE]);
2414 return PTR_ERR(chain);
2415 }
2416 attr = nla[NFTA_CHAIN_HANDLE];
2417 } else if (nla[NFTA_CHAIN_NAME]) {
2418 chain = nft_chain_lookup(net, table, attr, genmask);
2419 if (IS_ERR(chain)) {
2420 if (PTR_ERR(chain) != -ENOENT) {
2421 NL_SET_BAD_ATTR(extack, attr);
2422 return PTR_ERR(chain);
2423 }
2424 chain = NULL;
2425 }
2426 } else if (!nla[NFTA_CHAIN_ID]) {
2427 return -EINVAL;
2428 }
2429
2430 if (nla[NFTA_CHAIN_POLICY]) {
2431 if (chain != NULL &&
2432 !nft_is_base_chain(chain)) {
2433 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2434 return -EOPNOTSUPP;
2435 }
2436
2437 if (chain == NULL &&
2438 nla[NFTA_CHAIN_HOOK] == NULL) {
2439 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2440 return -EOPNOTSUPP;
2441 }
2442
2443 policy = ntohl(nla_get_be32(nla[NFTA_CHAIN_POLICY]));
2444 switch (policy) {
2445 case NF_DROP:
2446 case NF_ACCEPT:
2447 break;
2448 default:
2449 return -EINVAL;
2450 }
2451 }
2452
2453 if (nla[NFTA_CHAIN_FLAGS])
2454 flags = ntohl(nla_get_be32(nla[NFTA_CHAIN_FLAGS]));
2455 else if (chain)
2456 flags = chain->flags;
2457
2458 if (flags & ~NFT_CHAIN_FLAGS)
2459 return -EOPNOTSUPP;
2460
2461 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2462
2463 if (chain != NULL) {
2464 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
2465 NL_SET_BAD_ATTR(extack, attr);
2466 return -EEXIST;
2467 }
2468 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
2469 return -EOPNOTSUPP;
2470
2471 flags |= chain->flags & NFT_CHAIN_BASE;
2472 return nf_tables_updchain(&ctx, genmask, policy, flags, attr,
2473 extack);
2474 }
2475
2476 return nf_tables_addchain(&ctx, family, genmask, policy, flags, extack);
2477 }
2478
2479 static int nf_tables_delchain(struct sk_buff *skb, const struct nfnl_info *info,
2480 const struct nlattr * const nla[])
2481 {
2482 struct netlink_ext_ack *extack = info->extack;
2483 u8 genmask = nft_genmask_next(info->net);
2484 u8 family = info->nfmsg->nfgen_family;
2485 struct net *net = info->net;
2486 const struct nlattr *attr;
2487 struct nft_table *table;
2488 struct nft_chain *chain;
2489 struct nft_rule *rule;
2490 struct nft_ctx ctx;
2491 u64 handle;
2492 u32 use;
2493 int err;
2494
2495 table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2496 NETLINK_CB(skb).portid);
2497 if (IS_ERR(table)) {
2498 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2499 return PTR_ERR(table);
2500 }
2501
2502 if (nla[NFTA_CHAIN_HANDLE]) {
2503 attr = nla[NFTA_CHAIN_HANDLE];
2504 handle = be64_to_cpu(nla_get_be64(attr));
2505 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2506 } else {
2507 attr = nla[NFTA_CHAIN_NAME];
2508 chain = nft_chain_lookup(net, table, attr, genmask);
2509 }
2510 if (IS_ERR(chain)) {
2511 NL_SET_BAD_ATTR(extack, attr);
2512 return PTR_ERR(chain);
2513 }
2514
2515 if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
2516 chain->use > 0)
2517 return -EBUSY;
2518
2519 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2520
2521 use = chain->use;
2522 list_for_each_entry(rule, &chain->rules, list) {
2523 if (!nft_is_active_next(net, rule))
2524 continue;
2525 use--;
2526
2527 err = nft_delrule(&ctx, rule);
2528 if (err < 0)
2529 return err;
2530 }
2531
2532 /* There are rules and elements that are still holding references to us,
2533 * we cannot do a recursive removal in this case.
2534 */
2535 if (use > 0) {
2536 NL_SET_BAD_ATTR(extack, attr);
2537 return -EBUSY;
2538 }
2539
2540 return nft_delchain(&ctx);
2541 }
2542
2543 /*
2544 * Expressions
2545 */
2546
2547 /**
2548 * nft_register_expr - register nf_tables expr type
2549 * @type: expr type
2550 *
2551 * Registers the expr type for use with nf_tables. Returns zero on
2552 * success or a negative errno code otherwise.
2553 */
2554 int nft_register_expr(struct nft_expr_type *type)
2555 {
2556 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2557 if (type->family == NFPROTO_UNSPEC)
2558 list_add_tail_rcu(&type->list, &nf_tables_expressions);
2559 else
2560 list_add_rcu(&type->list, &nf_tables_expressions);
2561 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2562 return 0;
2563 }
2564 EXPORT_SYMBOL_GPL(nft_register_expr);
2565
2566 /**
2567 * nft_unregister_expr - unregister nf_tables expr type
2568 * @type: expr type
2569 *
2570 * Unregisters the expr typefor use with nf_tables.
2571 */
2572 void nft_unregister_expr(struct nft_expr_type *type)
2573 {
2574 nfnl_lock(NFNL_SUBSYS_NFTABLES);
2575 list_del_rcu(&type->list);
2576 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
2577 }
2578 EXPORT_SYMBOL_GPL(nft_unregister_expr);
2579
2580 static const struct nft_expr_type *__nft_expr_type_get(u8 family,
2581 struct nlattr *nla)
2582 {
2583 const struct nft_expr_type *type, *candidate = NULL;
2584
2585 list_for_each_entry(type, &nf_tables_expressions, list) {
2586 if (!nla_strcmp(nla, type->name)) {
2587 if (!type->family && !candidate)
2588 candidate = type;
2589 else if (type->family == family)
2590 candidate = type;
2591 }
2592 }
2593 return candidate;
2594 }
2595
2596 #ifdef CONFIG_MODULES
2597 static int nft_expr_type_request_module(struct net *net, u8 family,
2598 struct nlattr *nla)
2599 {
2600 if (nft_request_module(net, "nft-expr-%u-%.*s", family,
2601 nla_len(nla), (char *)nla_data(nla)) == -EAGAIN)
2602 return -EAGAIN;
2603
2604 return 0;
2605 }
2606 #endif
2607
2608 static const struct nft_expr_type *nft_expr_type_get(struct net *net,
2609 u8 family,
2610 struct nlattr *nla)
2611 {
2612 const struct nft_expr_type *type;
2613
2614 if (nla == NULL)
2615 return ERR_PTR(-EINVAL);
2616
2617 type = __nft_expr_type_get(family, nla);
2618 if (type != NULL && try_module_get(type->owner))
2619 return type;
2620
2621 lockdep_nfnl_nft_mutex_not_held();
2622 #ifdef CONFIG_MODULES
2623 if (type == NULL) {
2624 if (nft_expr_type_request_module(net, family, nla) == -EAGAIN)
2625 return ERR_PTR(-EAGAIN);
2626
2627 if (nft_request_module(net, "nft-expr-%.*s",
2628 nla_len(nla),
2629 (char *)nla_data(nla)) == -EAGAIN)
2630 return ERR_PTR(-EAGAIN);
2631 }
2632 #endif
2633 return ERR_PTR(-ENOENT);
2634 }
2635
2636 static const struct nla_policy nft_expr_policy[NFTA_EXPR_MAX + 1] = {
2637 [NFTA_EXPR_NAME] = { .type = NLA_STRING,
2638 .len = NFT_MODULE_AUTOLOAD_LIMIT },
2639 [NFTA_EXPR_DATA] = { .type = NLA_NESTED },
2640 };
2641
2642 static int nf_tables_fill_expr_info(struct sk_buff *skb,
2643 const struct nft_expr *expr)
2644 {
2645 if (nla_put_string(skb, NFTA_EXPR_NAME, expr->ops->type->name))
2646 goto nla_put_failure;
2647
2648 if (expr->ops->dump) {
2649 struct nlattr *data = nla_nest_start_noflag(skb,
2650 NFTA_EXPR_DATA);
2651 if (data == NULL)
2652 goto nla_put_failure;
2653 if (expr->ops->dump(skb, expr) < 0)
2654 goto nla_put_failure;
2655 nla_nest_end(skb, data);
2656 }
2657
2658 return skb->len;
2659
2660 nla_put_failure:
2661 return -1;
2662 };
2663
2664 int nft_expr_dump(struct sk_buff *skb, unsigned int attr,
2665 const struct nft_expr *expr)
2666 {
2667 struct nlattr *nest;
2668
2669 nest = nla_nest_start_noflag(skb, attr);
2670 if (!nest)
2671 goto nla_put_failure;
2672 if (nf_tables_fill_expr_info(skb, expr) < 0)
2673 goto nla_put_failure;
2674 nla_nest_end(skb, nest);
2675 return 0;
2676
2677 nla_put_failure:
2678 return -1;
2679 }
2680
2681 struct nft_expr_info {
2682 const struct nft_expr_ops *ops;
2683 const struct nlattr *attr;
2684 struct nlattr *tb[NFT_EXPR_MAXATTR + 1];
2685 };
2686
2687 static int nf_tables_expr_parse(const struct nft_ctx *ctx,
2688 const struct nlattr *nla,
2689 struct nft_expr_info *info)
2690 {
2691 const struct nft_expr_type *type;
2692 const struct nft_expr_ops *ops;
2693 struct nlattr *tb[NFTA_EXPR_MAX + 1];
2694 int err;
2695
2696 err = nla_parse_nested_deprecated(tb, NFTA_EXPR_MAX, nla,
2697 nft_expr_policy, NULL);
2698 if (err < 0)
2699 return err;
2700
2701 type = nft_expr_type_get(ctx->net, ctx->family, tb[NFTA_EXPR_NAME]);
2702 if (IS_ERR(type))
2703 return PTR_ERR(type);
2704
2705 if (tb[NFTA_EXPR_DATA]) {
2706 err = nla_parse_nested_deprecated(info->tb, type->maxattr,
2707 tb[NFTA_EXPR_DATA],
2708 type->policy, NULL);
2709 if (err < 0)
2710 goto err1;
2711 } else
2712 memset(info->tb, 0, sizeof(info->tb[0]) * (type->maxattr + 1));
2713
2714 if (type->select_ops != NULL) {
2715 ops = type->select_ops(ctx,
2716 (const struct nlattr * const *)info->tb);
2717 if (IS_ERR(ops)) {
2718 err = PTR_ERR(ops);
2719 #ifdef CONFIG_MODULES
2720 if (err == -EAGAIN)
2721 if (nft_expr_type_request_module(ctx->net,
2722 ctx->family,
2723 tb[NFTA_EXPR_NAME]) != -EAGAIN)
2724 err = -ENOENT;
2725 #endif
2726 goto err1;
2727 }
2728 } else
2729 ops = type->ops;
2730
2731 info->attr = nla;
2732 info->ops = ops;
2733
2734 return 0;
2735
2736 err1:
2737 module_put(type->owner);
2738 return err;
2739 }
2740
2741 static int nf_tables_newexpr(const struct nft_ctx *ctx,
2742 const struct nft_expr_info *expr_info,
2743 struct nft_expr *expr)
2744 {
2745 const struct nft_expr_ops *ops = expr_info->ops;
2746 int err;
2747
2748 expr->ops = ops;
2749 if (ops->init) {
2750 err = ops->init(ctx, expr, (const struct nlattr **)expr_info->tb);
2751 if (err < 0)
2752 goto err1;
2753 }
2754
2755 return 0;
2756 err1:
2757 expr->ops = NULL;
2758 return err;
2759 }
2760
2761 static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
2762 struct nft_expr *expr)
2763 {
2764 const struct nft_expr_type *type = expr->ops->type;
2765
2766 if (expr->ops->destroy)
2767 expr->ops->destroy(ctx, expr);
2768 module_put(type->owner);
2769 }
2770
2771 static struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
2772 const struct nlattr *nla)
2773 {
2774 struct nft_expr_info expr_info;
2775 struct nft_expr *expr;
2776 struct module *owner;
2777 int err;
2778
2779 err = nf_tables_expr_parse(ctx, nla, &expr_info);
2780 if (err < 0)
2781 goto err1;
2782
2783 err = -ENOMEM;
2784 expr = kzalloc(expr_info.ops->size, GFP_KERNEL);
2785 if (expr == NULL)
2786 goto err2;
2787
2788 err = nf_tables_newexpr(ctx, &expr_info, expr);
2789 if (err < 0)
2790 goto err3;
2791
2792 return expr;
2793 err3:
2794 kfree(expr);
2795 err2:
2796 owner = expr_info.ops->type->owner;
2797 if (expr_info.ops->type->release_ops)
2798 expr_info.ops->type->release_ops(expr_info.ops);
2799
2800 module_put(owner);
2801 err1:
2802 return ERR_PTR(err);
2803 }
2804
2805 int nft_expr_clone(struct nft_expr *dst, struct nft_expr *src)
2806 {
2807 int err;
2808
2809 if (src->ops->clone) {
2810 dst->ops = src->ops;
2811 err = src->ops->clone(dst, src);
2812 if (err < 0)
2813 return err;
2814 } else {
2815 memcpy(dst, src, src->ops->size);
2816 }
2817
2818 __module_get(src->ops->type->owner);
2819
2820 return 0;
2821 }
2822
2823 void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
2824 {
2825 nf_tables_expr_destroy(ctx, expr);
2826 kfree(expr);
2827 }
2828
2829 /*
2830 * Rules
2831 */
2832
2833 static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
2834 u64 handle)
2835 {
2836 struct nft_rule *rule;
2837
2838 // FIXME: this sucks
2839 list_for_each_entry_rcu(rule, &chain->rules, list) {
2840 if (handle == rule->handle)
2841 return rule;
2842 }
2843
2844 return ERR_PTR(-ENOENT);
2845 }
2846
2847 static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
2848 const struct nlattr *nla)
2849 {
2850 if (nla == NULL)
2851 return ERR_PTR(-EINVAL);
2852
2853 return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
2854 }
2855
2856 static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
2857 [NFTA_RULE_TABLE] = { .type = NLA_STRING,
2858 .len = NFT_TABLE_MAXNAMELEN - 1 },
2859 [NFTA_RULE_CHAIN] = { .type = NLA_STRING,
2860 .len = NFT_CHAIN_MAXNAMELEN - 1 },
2861 [NFTA_RULE_HANDLE] = { .type = NLA_U64 },
2862 [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
2863 [NFTA_RULE_COMPAT] = { .type = NLA_NESTED },
2864 [NFTA_RULE_POSITION] = { .type = NLA_U64 },
2865 [NFTA_RULE_USERDATA] = { .type = NLA_BINARY,
2866 .len = NFT_USERDATA_MAXLEN },
2867 [NFTA_RULE_ID] = { .type = NLA_U32 },
2868 [NFTA_RULE_POSITION_ID] = { .type = NLA_U32 },
2869 [NFTA_RULE_CHAIN_ID] = { .type = NLA_U32 },
2870 };
2871
2872 static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
2873 u32 portid, u32 seq, int event,
2874 u32 flags, int family,
2875 const struct nft_table *table,
2876 const struct nft_chain *chain,
2877 const struct nft_rule *rule, u64 handle)
2878 {
2879 struct nlmsghdr *nlh;
2880 const struct nft_expr *expr, *next;
2881 struct nlattr *list;
2882 u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
2883
2884 nlh = nfnl_msg_put(skb, portid, seq, type, flags, family, NFNETLINK_V0,
2885 nft_base_seq(net));
2886 if (!nlh)
2887 goto nla_put_failure;
2888
2889 if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
2890 goto nla_put_failure;
2891 if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
2892 goto nla_put_failure;
2893 if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
2894 NFTA_RULE_PAD))
2895 goto nla_put_failure;
2896
2897 if (event != NFT_MSG_DELRULE && handle) {
2898 if (nla_put_be64(skb, NFTA_RULE_POSITION, cpu_to_be64(handle),
2899 NFTA_RULE_PAD))
2900 goto nla_put_failure;
2901 }
2902
2903 if (chain->flags & NFT_CHAIN_HW_OFFLOAD)
2904 nft_flow_rule_stats(chain, rule);
2905
2906 list = nla_nest_start_noflag(skb, NFTA_RULE_EXPRESSIONS);
2907 if (list == NULL)
2908 goto nla_put_failure;
2909 nft_rule_for_each_expr(expr, next, rule) {
2910 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
2911 goto nla_put_failure;
2912 }
2913 nla_nest_end(skb, list);
2914
2915 if (rule->udata) {
2916 struct nft_userdata *udata = nft_userdata(rule);
2917 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
2918 udata->data) < 0)
2919 goto nla_put_failure;
2920 }
2921
2922 nlmsg_end(skb, nlh);
2923 return 0;
2924
2925 nla_put_failure:
2926 nlmsg_trim(skb, nlh);
2927 return -1;
2928 }
2929
2930 static void nf_tables_rule_notify(const struct nft_ctx *ctx,
2931 const struct nft_rule *rule, int event)
2932 {
2933 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
2934 const struct nft_rule *prule;
2935 struct sk_buff *skb;
2936 u64 handle = 0;
2937 u16 flags = 0;
2938 int err;
2939
2940 if (!ctx->report &&
2941 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
2942 return;
2943
2944 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
2945 if (skb == NULL)
2946 goto err;
2947
2948 if (event == NFT_MSG_NEWRULE &&
2949 !list_is_first(&rule->list, &ctx->chain->rules) &&
2950 !list_is_last(&rule->list, &ctx->chain->rules)) {
2951 prule = list_prev_entry(rule, list);
2952 handle = prule->handle;
2953 }
2954 if (ctx->flags & (NLM_F_APPEND | NLM_F_REPLACE))
2955 flags |= NLM_F_APPEND;
2956 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
2957 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
2958
2959 err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
2960 event, flags, ctx->family, ctx->table,
2961 ctx->chain, rule, handle);
2962 if (err < 0) {
2963 kfree_skb(skb);
2964 goto err;
2965 }
2966
2967 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
2968 return;
2969 err:
2970 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
2971 }
2972
2973 struct nft_rule_dump_ctx {
2974 char *table;
2975 char *chain;
2976 };
2977
2978 static int __nf_tables_dump_rules(struct sk_buff *skb,
2979 unsigned int *idx,
2980 struct netlink_callback *cb,
2981 const struct nft_table *table,
2982 const struct nft_chain *chain)
2983 {
2984 struct net *net = sock_net(skb->sk);
2985 const struct nft_rule *rule, *prule;
2986 unsigned int s_idx = cb->args[0];
2987 u64 handle;
2988
2989 prule = NULL;
2990 list_for_each_entry_rcu(rule, &chain->rules, list) {
2991 if (!nft_is_active(net, rule))
2992 goto cont_skip;
2993 if (*idx < s_idx)
2994 goto cont;
2995 if (*idx > s_idx) {
2996 memset(&cb->args[1], 0,
2997 sizeof(cb->args) - sizeof(cb->args[0]));
2998 }
2999 if (prule)
3000 handle = prule->handle;
3001 else
3002 handle = 0;
3003
3004 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
3005 cb->nlh->nlmsg_seq,
3006 NFT_MSG_NEWRULE,
3007 NLM_F_MULTI | NLM_F_APPEND,
3008 table->family,
3009 table, chain, rule, handle) < 0)
3010 return 1;
3011
3012 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
3013 cont:
3014 prule = rule;
3015 cont_skip:
3016 (*idx)++;
3017 }
3018 return 0;
3019 }
3020
3021 static int nf_tables_dump_rules(struct sk_buff *skb,
3022 struct netlink_callback *cb)
3023 {
3024 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3025 const struct nft_rule_dump_ctx *ctx = cb->data;
3026 struct nft_table *table;
3027 const struct nft_chain *chain;
3028 unsigned int idx = 0;
3029 struct net *net = sock_net(skb->sk);
3030 int family = nfmsg->nfgen_family;
3031 struct nftables_pernet *nft_net;
3032
3033 rcu_read_lock();
3034 nft_net = nft_pernet(net);
3035 cb->seq = nft_net->base_seq;
3036
3037 list_for_each_entry_rcu(table, &nft_net->tables, list) {
3038 if (family != NFPROTO_UNSPEC && family != table->family)
3039 continue;
3040
3041 if (ctx && ctx->table && strcmp(ctx->table, table->name) != 0)
3042 continue;
3043
3044 if (ctx && ctx->table && ctx->chain) {
3045 struct rhlist_head *list, *tmp;
3046
3047 list = rhltable_lookup(&table->chains_ht, ctx->chain,
3048 nft_chain_ht_params);
3049 if (!list)
3050 goto done;
3051
3052 rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
3053 if (!nft_is_active(net, chain))
3054 continue;
3055 __nf_tables_dump_rules(skb, &idx,
3056 cb, table, chain);
3057 break;
3058 }
3059 goto done;
3060 }
3061
3062 list_for_each_entry_rcu(chain, &table->chains, list) {
3063 if (__nf_tables_dump_rules(skb, &idx, cb, table, chain))
3064 goto done;
3065 }
3066
3067 if (ctx && ctx->table)
3068 break;
3069 }
3070 done:
3071 rcu_read_unlock();
3072
3073 cb->args[0] = idx;
3074 return skb->len;
3075 }
3076
3077 static int nf_tables_dump_rules_start(struct netlink_callback *cb)
3078 {
3079 const struct nlattr * const *nla = cb->data;
3080 struct nft_rule_dump_ctx *ctx = NULL;
3081
3082 if (nla[NFTA_RULE_TABLE] || nla[NFTA_RULE_CHAIN]) {
3083 ctx = kzalloc(sizeof(*ctx), GFP_ATOMIC);
3084 if (!ctx)
3085 return -ENOMEM;
3086
3087 if (nla[NFTA_RULE_TABLE]) {
3088 ctx->table = nla_strdup(nla[NFTA_RULE_TABLE],
3089 GFP_ATOMIC);
3090 if (!ctx->table) {
3091 kfree(ctx);
3092 return -ENOMEM;
3093 }
3094 }
3095 if (nla[NFTA_RULE_CHAIN]) {
3096 ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN],
3097 GFP_ATOMIC);
3098 if (!ctx->chain) {
3099 kfree(ctx->table);
3100 kfree(ctx);
3101 return -ENOMEM;
3102 }
3103 }
3104 }
3105
3106 cb->data = ctx;
3107 return 0;
3108 }
3109
3110 static int nf_tables_dump_rules_done(struct netlink_callback *cb)
3111 {
3112 struct nft_rule_dump_ctx *ctx = cb->data;
3113
3114 if (ctx) {
3115 kfree(ctx->table);
3116 kfree(ctx->chain);
3117 kfree(ctx);
3118 }
3119 return 0;
3120 }
3121
3122 /* called with rcu_read_lock held */
3123 static int nf_tables_getrule(struct sk_buff *skb, const struct nfnl_info *info,
3124 const struct nlattr * const nla[])
3125 {
3126 struct netlink_ext_ack *extack = info->extack;
3127 u8 genmask = nft_genmask_cur(info->net);
3128 u8 family = info->nfmsg->nfgen_family;
3129 const struct nft_chain *chain;
3130 const struct nft_rule *rule;
3131 struct net *net = info->net;
3132 struct nft_table *table;
3133 struct sk_buff *skb2;
3134 int err;
3135
3136 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3137 struct netlink_dump_control c = {
3138 .start= nf_tables_dump_rules_start,
3139 .dump = nf_tables_dump_rules,
3140 .done = nf_tables_dump_rules_done,
3141 .module = THIS_MODULE,
3142 .data = (void *)nla,
3143 };
3144
3145 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
3146 }
3147
3148 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask, 0);
3149 if (IS_ERR(table)) {
3150 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3151 return PTR_ERR(table);
3152 }
3153
3154 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
3155 if (IS_ERR(chain)) {
3156 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3157 return PTR_ERR(chain);
3158 }
3159
3160 rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3161 if (IS_ERR(rule)) {
3162 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3163 return PTR_ERR(rule);
3164 }
3165
3166 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3167 if (!skb2)
3168 return -ENOMEM;
3169
3170 err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
3171 info->nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
3172 family, table, chain, rule, 0);
3173 if (err < 0)
3174 goto err_fill_rule_info;
3175
3176 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
3177
3178 err_fill_rule_info:
3179 kfree_skb(skb2);
3180 return err;
3181 }
3182
3183 static void nf_tables_rule_destroy(const struct nft_ctx *ctx,
3184 struct nft_rule *rule)
3185 {
3186 struct nft_expr *expr, *next;
3187
3188 /*
3189 * Careful: some expressions might not be initialized in case this
3190 * is called on error from nf_tables_newrule().
3191 */
3192 expr = nft_expr_first(rule);
3193 while (nft_expr_more(rule, expr)) {
3194 next = nft_expr_next(expr);
3195 nf_tables_expr_destroy(ctx, expr);
3196 expr = next;
3197 }
3198 kfree(rule);
3199 }
3200
3201 void nf_tables_rule_release(const struct nft_ctx *ctx, struct nft_rule *rule)
3202 {
3203 nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_RELEASE);
3204 nf_tables_rule_destroy(ctx, rule);
3205 }
3206
3207 int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
3208 {
3209 struct nft_expr *expr, *last;
3210 const struct nft_data *data;
3211 struct nft_rule *rule;
3212 int err;
3213
3214 if (ctx->level == NFT_JUMP_STACK_SIZE)
3215 return -EMLINK;
3216
3217 list_for_each_entry(rule, &chain->rules, list) {
3218 if (!nft_is_active_next(ctx->net, rule))
3219 continue;
3220
3221 nft_rule_for_each_expr(expr, last, rule) {
3222 if (!expr->ops->validate)
3223 continue;
3224
3225 err = expr->ops->validate(ctx, expr, &data);
3226 if (err < 0)
3227 return err;
3228 }
3229 }
3230
3231 return 0;
3232 }
3233 EXPORT_SYMBOL_GPL(nft_chain_validate);
3234
3235 static int nft_table_validate(struct net *net, const struct nft_table *table)
3236 {
3237 struct nft_chain *chain;
3238 struct nft_ctx ctx = {
3239 .net = net,
3240 .family = table->family,
3241 };
3242 int err;
3243
3244 list_for_each_entry(chain, &table->chains, list) {
3245 if (!nft_is_base_chain(chain))
3246 continue;
3247
3248 ctx.chain = chain;
3249 err = nft_chain_validate(&ctx, chain);
3250 if (err < 0)
3251 return err;
3252 }
3253
3254 return 0;
3255 }
3256
3257 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3258 const struct nlattr *nla);
3259
3260 #define NFT_RULE_MAXEXPRS 128
3261
3262 static int nf_tables_newrule(struct sk_buff *skb, const struct nfnl_info *info,
3263 const struct nlattr * const nla[])
3264 {
3265 struct nftables_pernet *nft_net = nft_pernet(info->net);
3266 struct netlink_ext_ack *extack = info->extack;
3267 unsigned int size, i, n, ulen = 0, usize = 0;
3268 u8 genmask = nft_genmask_next(info->net);
3269 struct nft_rule *rule, *old_rule = NULL;
3270 struct nft_expr_info *expr_info = NULL;
3271 u8 family = info->nfmsg->nfgen_family;
3272 struct nft_flow_rule *flow = NULL;
3273 struct net *net = info->net;
3274 struct nft_userdata *udata;
3275 struct nft_table *table;
3276 struct nft_chain *chain;
3277 struct nft_trans *trans;
3278 u64 handle, pos_handle;
3279 struct nft_expr *expr;
3280 struct nft_ctx ctx;
3281 struct nlattr *tmp;
3282 int err, rem;
3283
3284 lockdep_assert_held(&nft_net->commit_mutex);
3285
3286 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3287 NETLINK_CB(skb).portid);
3288 if (IS_ERR(table)) {
3289 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3290 return PTR_ERR(table);
3291 }
3292
3293 if (nla[NFTA_RULE_CHAIN]) {
3294 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3295 genmask);
3296 if (IS_ERR(chain)) {
3297 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3298 return PTR_ERR(chain);
3299 }
3300 if (nft_chain_is_bound(chain))
3301 return -EOPNOTSUPP;
3302
3303 } else if (nla[NFTA_RULE_CHAIN_ID]) {
3304 chain = nft_chain_lookup_byid(net, nla[NFTA_RULE_CHAIN_ID]);
3305 if (IS_ERR(chain)) {
3306 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN_ID]);
3307 return PTR_ERR(chain);
3308 }
3309 } else {
3310 return -EINVAL;
3311 }
3312
3313 if (nla[NFTA_RULE_HANDLE]) {
3314 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
3315 rule = __nft_rule_lookup(chain, handle);
3316 if (IS_ERR(rule)) {
3317 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3318 return PTR_ERR(rule);
3319 }
3320
3321 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
3322 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3323 return -EEXIST;
3324 }
3325 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
3326 old_rule = rule;
3327 else
3328 return -EOPNOTSUPP;
3329 } else {
3330 if (!(info->nlh->nlmsg_flags & NLM_F_CREATE) ||
3331 info->nlh->nlmsg_flags & NLM_F_REPLACE)
3332 return -EINVAL;
3333 handle = nf_tables_alloc_handle(table);
3334
3335 if (chain->use == UINT_MAX)
3336 return -EOVERFLOW;
3337
3338 if (nla[NFTA_RULE_POSITION]) {
3339 pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
3340 old_rule = __nft_rule_lookup(chain, pos_handle);
3341 if (IS_ERR(old_rule)) {
3342 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
3343 return PTR_ERR(old_rule);
3344 }
3345 } else if (nla[NFTA_RULE_POSITION_ID]) {
3346 old_rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_POSITION_ID]);
3347 if (IS_ERR(old_rule)) {
3348 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION_ID]);
3349 return PTR_ERR(old_rule);
3350 }
3351 }
3352 }
3353
3354 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3355
3356 n = 0;
3357 size = 0;
3358 if (nla[NFTA_RULE_EXPRESSIONS]) {
3359 expr_info = kvmalloc_array(NFT_RULE_MAXEXPRS,
3360 sizeof(struct nft_expr_info),
3361 GFP_KERNEL);
3362 if (!expr_info)
3363 return -ENOMEM;
3364
3365 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
3366 err = -EINVAL;
3367 if (nla_type(tmp) != NFTA_LIST_ELEM)
3368 goto err_release_expr;
3369 if (n == NFT_RULE_MAXEXPRS)
3370 goto err_release_expr;
3371 err = nf_tables_expr_parse(&ctx, tmp, &expr_info[n]);
3372 if (err < 0) {
3373 NL_SET_BAD_ATTR(extack, tmp);
3374 goto err_release_expr;
3375 }
3376 size += expr_info[n].ops->size;
3377 n++;
3378 }
3379 }
3380 /* Check for overflow of dlen field */
3381 err = -EFBIG;
3382 if (size >= 1 << 12)
3383 goto err_release_expr;
3384
3385 if (nla[NFTA_RULE_USERDATA]) {
3386 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
3387 if (ulen > 0)
3388 usize = sizeof(struct nft_userdata) + ulen;
3389 }
3390
3391 err = -ENOMEM;
3392 rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL);
3393 if (rule == NULL)
3394 goto err_release_expr;
3395
3396 nft_activate_next(net, rule);
3397
3398 rule->handle = handle;
3399 rule->dlen = size;
3400 rule->udata = ulen ? 1 : 0;
3401
3402 if (ulen) {
3403 udata = nft_userdata(rule);
3404 udata->len = ulen - 1;
3405 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
3406 }
3407
3408 expr = nft_expr_first(rule);
3409 for (i = 0; i < n; i++) {
3410 err = nf_tables_newexpr(&ctx, &expr_info[i], expr);
3411 if (err < 0) {
3412 NL_SET_BAD_ATTR(extack, expr_info[i].attr);
3413 goto err_release_rule;
3414 }
3415
3416 if (expr_info[i].ops->validate)
3417 nft_validate_state_update(net, NFT_VALIDATE_NEED);
3418
3419 expr_info[i].ops = NULL;
3420 expr = nft_expr_next(expr);
3421 }
3422
3423 if (chain->flags & NFT_CHAIN_HW_OFFLOAD) {
3424 flow = nft_flow_rule_create(net, rule);
3425 if (IS_ERR(flow)) {
3426 err = PTR_ERR(flow);
3427 goto err_release_rule;
3428 }
3429 }
3430
3431 if (info->nlh->nlmsg_flags & NLM_F_REPLACE) {
3432 err = nft_delrule(&ctx, old_rule);
3433 if (err < 0)
3434 goto err_destroy_flow_rule;
3435
3436 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3437 if (trans == NULL) {
3438 err = -ENOMEM;
3439 goto err_destroy_flow_rule;
3440 }
3441 list_add_tail_rcu(&rule->list, &old_rule->list);
3442 } else {
3443 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3444 if (!trans) {
3445 err = -ENOMEM;
3446 goto err_destroy_flow_rule;
3447 }
3448
3449 if (info->nlh->nlmsg_flags & NLM_F_APPEND) {
3450 if (old_rule)
3451 list_add_rcu(&rule->list, &old_rule->list);
3452 else
3453 list_add_tail_rcu(&rule->list, &chain->rules);
3454 } else {
3455 if (old_rule)
3456 list_add_tail_rcu(&rule->list, &old_rule->list);
3457 else
3458 list_add_rcu(&rule->list, &chain->rules);
3459 }
3460 }
3461 kvfree(expr_info);
3462 chain->use++;
3463
3464 if (flow)
3465 nft_trans_flow_rule(trans) = flow;
3466
3467 if (nft_net->validate_state == NFT_VALIDATE_DO)
3468 return nft_table_validate(net, table);
3469
3470 return 0;
3471
3472 err_destroy_flow_rule:
3473 if (flow)
3474 nft_flow_rule_destroy(flow);
3475 err_release_rule:
3476 nf_tables_rule_release(&ctx, rule);
3477 err_release_expr:
3478 for (i = 0; i < n; i++) {
3479 if (expr_info[i].ops) {
3480 module_put(expr_info[i].ops->type->owner);
3481 if (expr_info[i].ops->type->release_ops)
3482 expr_info[i].ops->type->release_ops(expr_info[i].ops);
3483 }
3484 }
3485 kvfree(expr_info);
3486
3487 return err;
3488 }
3489
3490 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3491 const struct nlattr *nla)
3492 {
3493 struct nftables_pernet *nft_net = nft_pernet(net);
3494 u32 id = ntohl(nla_get_be32(nla));
3495 struct nft_trans *trans;
3496
3497 list_for_each_entry(trans, &nft_net->commit_list, list) {
3498 struct nft_rule *rule = nft_trans_rule(trans);
3499
3500 if (trans->msg_type == NFT_MSG_NEWRULE &&
3501 id == nft_trans_rule_id(trans))
3502 return rule;
3503 }
3504 return ERR_PTR(-ENOENT);
3505 }
3506
3507 static int nf_tables_delrule(struct sk_buff *skb, const struct nfnl_info *info,
3508 const struct nlattr * const nla[])
3509 {
3510 struct netlink_ext_ack *extack = info->extack;
3511 u8 genmask = nft_genmask_next(info->net);
3512 u8 family = info->nfmsg->nfgen_family;
3513 struct nft_chain *chain = NULL;
3514 struct net *net = info->net;
3515 struct nft_table *table;
3516 struct nft_rule *rule;
3517 struct nft_ctx ctx;
3518 int err = 0;
3519
3520 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3521 NETLINK_CB(skb).portid);
3522 if (IS_ERR(table)) {
3523 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3524 return PTR_ERR(table);
3525 }
3526
3527 if (nla[NFTA_RULE_CHAIN]) {
3528 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3529 genmask);
3530 if (IS_ERR(chain)) {
3531 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3532 return PTR_ERR(chain);
3533 }
3534 if (nft_chain_is_bound(chain))
3535 return -EOPNOTSUPP;
3536 }
3537
3538 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3539
3540 if (chain) {
3541 if (nla[NFTA_RULE_HANDLE]) {
3542 rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3543 if (IS_ERR(rule)) {
3544 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3545 return PTR_ERR(rule);
3546 }
3547
3548 err = nft_delrule(&ctx, rule);
3549 } else if (nla[NFTA_RULE_ID]) {
3550 rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_ID]);
3551 if (IS_ERR(rule)) {
3552 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
3553 return PTR_ERR(rule);
3554 }
3555
3556 err = nft_delrule(&ctx, rule);
3557 } else {
3558 err = nft_delrule_by_chain(&ctx);
3559 }
3560 } else {
3561 list_for_each_entry(chain, &table->chains, list) {
3562 if (!nft_is_active_next(net, chain))
3563 continue;
3564
3565 ctx.chain = chain;
3566 err = nft_delrule_by_chain(&ctx);
3567 if (err < 0)
3568 break;
3569 }
3570 }
3571
3572 return err;
3573 }
3574
3575 /*
3576 * Sets
3577 */
3578 static const struct nft_set_type *nft_set_types[] = {
3579 &nft_set_hash_fast_type,
3580 &nft_set_hash_type,
3581 &nft_set_rhash_type,
3582 &nft_set_bitmap_type,
3583 &nft_set_rbtree_type,
3584 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
3585 &nft_set_pipapo_avx2_type,
3586 #endif
3587 &nft_set_pipapo_type,
3588 };
3589
3590 #define NFT_SET_FEATURES (NFT_SET_INTERVAL | NFT_SET_MAP | \
3591 NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
3592 NFT_SET_EVAL)
3593
3594 static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
3595 {
3596 return (flags & type->features) == (flags & NFT_SET_FEATURES);
3597 }
3598
3599 /*
3600 * Select a set implementation based on the data characteristics and the
3601 * given policy. The total memory use might not be known if no size is
3602 * given, in that case the amount of memory per element is used.
3603 */
3604 static const struct nft_set_ops *
3605 nft_select_set_ops(const struct nft_ctx *ctx,
3606 const struct nlattr * const nla[],
3607 const struct nft_set_desc *desc,
3608 enum nft_set_policies policy)
3609 {
3610 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3611 const struct nft_set_ops *ops, *bops;
3612 struct nft_set_estimate est, best;
3613 const struct nft_set_type *type;
3614 u32 flags = 0;
3615 int i;
3616
3617 lockdep_assert_held(&nft_net->commit_mutex);
3618 lockdep_nfnl_nft_mutex_not_held();
3619
3620 if (nla[NFTA_SET_FLAGS] != NULL)
3621 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
3622
3623 bops = NULL;
3624 best.size = ~0;
3625 best.lookup = ~0;
3626 best.space = ~0;
3627
3628 for (i = 0; i < ARRAY_SIZE(nft_set_types); i++) {
3629 type = nft_set_types[i];
3630 ops = &type->ops;
3631
3632 if (!nft_set_ops_candidate(type, flags))
3633 continue;
3634 if (!ops->estimate(desc, flags, &est))
3635 continue;
3636
3637 switch (policy) {
3638 case NFT_SET_POL_PERFORMANCE:
3639 if (est.lookup < best.lookup)
3640 break;
3641 if (est.lookup == best.lookup &&
3642 est.space < best.space)
3643 break;
3644 continue;
3645 case NFT_SET_POL_MEMORY:
3646 if (!desc->size) {
3647 if (est.space < best.space)
3648 break;
3649 if (est.space == best.space &&
3650 est.lookup < best.lookup)
3651 break;
3652 } else if (est.size < best.size || !bops) {
3653 break;
3654 }
3655 continue;
3656 default:
3657 break;
3658 }
3659
3660 bops = ops;
3661 best = est;
3662 }
3663
3664 if (bops != NULL)
3665 return bops;
3666
3667 return ERR_PTR(-EOPNOTSUPP);
3668 }
3669
3670 static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
3671 [NFTA_SET_TABLE] = { .type = NLA_STRING,
3672 .len = NFT_TABLE_MAXNAMELEN - 1 },
3673 [NFTA_SET_NAME] = { .type = NLA_STRING,
3674 .len = NFT_SET_MAXNAMELEN - 1 },
3675 [NFTA_SET_FLAGS] = { .type = NLA_U32 },
3676 [NFTA_SET_KEY_TYPE] = { .type = NLA_U32 },
3677 [NFTA_SET_KEY_LEN] = { .type = NLA_U32 },
3678 [NFTA_SET_DATA_TYPE] = { .type = NLA_U32 },
3679 [NFTA_SET_DATA_LEN] = { .type = NLA_U32 },
3680 [NFTA_SET_POLICY] = { .type = NLA_U32 },
3681 [NFTA_SET_DESC] = { .type = NLA_NESTED },
3682 [NFTA_SET_ID] = { .type = NLA_U32 },
3683 [NFTA_SET_TIMEOUT] = { .type = NLA_U64 },
3684 [NFTA_SET_GC_INTERVAL] = { .type = NLA_U32 },
3685 [NFTA_SET_USERDATA] = { .type = NLA_BINARY,
3686 .len = NFT_USERDATA_MAXLEN },
3687 [NFTA_SET_OBJ_TYPE] = { .type = NLA_U32 },
3688 [NFTA_SET_HANDLE] = { .type = NLA_U64 },
3689 [NFTA_SET_EXPR] = { .type = NLA_NESTED },
3690 [NFTA_SET_EXPRESSIONS] = { .type = NLA_NESTED },
3691 };
3692
3693 static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
3694 [NFTA_SET_DESC_SIZE] = { .type = NLA_U32 },
3695 [NFTA_SET_DESC_CONCAT] = { .type = NLA_NESTED },
3696 };
3697
3698 static struct nft_set *nft_set_lookup(const struct nft_table *table,
3699 const struct nlattr *nla, u8 genmask)
3700 {
3701 struct nft_set *set;
3702
3703 if (nla == NULL)
3704 return ERR_PTR(-EINVAL);
3705
3706 list_for_each_entry_rcu(set, &table->sets, list) {
3707 if (!nla_strcmp(nla, set->name) &&
3708 nft_active_genmask(set, genmask))
3709 return set;
3710 }
3711 return ERR_PTR(-ENOENT);
3712 }
3713
3714 static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
3715 const struct nlattr *nla,
3716 u8 genmask)
3717 {
3718 struct nft_set *set;
3719
3720 list_for_each_entry(set, &table->sets, list) {
3721 if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
3722 nft_active_genmask(set, genmask))
3723 return set;
3724 }
3725 return ERR_PTR(-ENOENT);
3726 }
3727
3728 static struct nft_set *nft_set_lookup_byid(const struct net *net,
3729 const struct nlattr *nla, u8 genmask)
3730 {
3731 struct nftables_pernet *nft_net = nft_pernet(net);
3732 u32 id = ntohl(nla_get_be32(nla));
3733 struct nft_trans *trans;
3734
3735 list_for_each_entry(trans, &nft_net->commit_list, list) {
3736 if (trans->msg_type == NFT_MSG_NEWSET) {
3737 struct nft_set *set = nft_trans_set(trans);
3738
3739 if (id == nft_trans_set_id(trans) &&
3740 nft_active_genmask(set, genmask))
3741 return set;
3742 }
3743 }
3744 return ERR_PTR(-ENOENT);
3745 }
3746
3747 struct nft_set *nft_set_lookup_global(const struct net *net,
3748 const struct nft_table *table,
3749 const struct nlattr *nla_set_name,
3750 const struct nlattr *nla_set_id,
3751 u8 genmask)
3752 {
3753 struct nft_set *set;
3754
3755 set = nft_set_lookup(table, nla_set_name, genmask);
3756 if (IS_ERR(set)) {
3757 if (!nla_set_id)
3758 return set;
3759
3760 set = nft_set_lookup_byid(net, nla_set_id, genmask);
3761 }
3762 return set;
3763 }
3764 EXPORT_SYMBOL_GPL(nft_set_lookup_global);
3765
3766 static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
3767 const char *name)
3768 {
3769 const struct nft_set *i;
3770 const char *p;
3771 unsigned long *inuse;
3772 unsigned int n = 0, min = 0;
3773
3774 p = strchr(name, '%');
3775 if (p != NULL) {
3776 if (p[1] != 'd' || strchr(p + 2, '%'))
3777 return -EINVAL;
3778
3779 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
3780 if (inuse == NULL)
3781 return -ENOMEM;
3782 cont:
3783 list_for_each_entry(i, &ctx->table->sets, list) {
3784 int tmp;
3785
3786 if (!nft_is_active_next(ctx->net, set))
3787 continue;
3788 if (!sscanf(i->name, name, &tmp))
3789 continue;
3790 if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
3791 continue;
3792
3793 set_bit(tmp - min, inuse);
3794 }
3795
3796 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
3797 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
3798 min += BITS_PER_BYTE * PAGE_SIZE;
3799 memset(inuse, 0, PAGE_SIZE);
3800 goto cont;
3801 }
3802 free_page((unsigned long)inuse);
3803 }
3804
3805 set->name = kasprintf(GFP_KERNEL, name, min + n);
3806 if (!set->name)
3807 return -ENOMEM;
3808
3809 list_for_each_entry(i, &ctx->table->sets, list) {
3810 if (!nft_is_active_next(ctx->net, i))
3811 continue;
3812 if (!strcmp(set->name, i->name)) {
3813 kfree(set->name);
3814 set->name = NULL;
3815 return -ENFILE;
3816 }
3817 }
3818 return 0;
3819 }
3820
3821 int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
3822 {
3823 u64 ms = be64_to_cpu(nla_get_be64(nla));
3824 u64 max = (u64)(~((u64)0));
3825
3826 max = div_u64(max, NSEC_PER_MSEC);
3827 if (ms >= max)
3828 return -ERANGE;
3829
3830 ms *= NSEC_PER_MSEC;
3831 *result = nsecs_to_jiffies64(ms);
3832 return 0;
3833 }
3834
3835 __be64 nf_jiffies64_to_msecs(u64 input)
3836 {
3837 return cpu_to_be64(jiffies64_to_msecs(input));
3838 }
3839
3840 static int nf_tables_fill_set_concat(struct sk_buff *skb,
3841 const struct nft_set *set)
3842 {
3843 struct nlattr *concat, *field;
3844 int i;
3845
3846 concat = nla_nest_start_noflag(skb, NFTA_SET_DESC_CONCAT);
3847 if (!concat)
3848 return -ENOMEM;
3849
3850 for (i = 0; i < set->field_count; i++) {
3851 field = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
3852 if (!field)
3853 return -ENOMEM;
3854
3855 if (nla_put_be32(skb, NFTA_SET_FIELD_LEN,
3856 htonl(set->field_len[i])))
3857 return -ENOMEM;
3858
3859 nla_nest_end(skb, field);
3860 }
3861
3862 nla_nest_end(skb, concat);
3863
3864 return 0;
3865 }
3866
3867 static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
3868 const struct nft_set *set, u16 event, u16 flags)
3869 {
3870 struct nlmsghdr *nlh;
3871 u32 portid = ctx->portid;
3872 struct nlattr *nest;
3873 u32 seq = ctx->seq;
3874 int i;
3875
3876 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3877 nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
3878 NFNETLINK_V0, nft_base_seq(ctx->net));
3879 if (!nlh)
3880 goto nla_put_failure;
3881
3882 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
3883 goto nla_put_failure;
3884 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
3885 goto nla_put_failure;
3886 if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
3887 NFTA_SET_PAD))
3888 goto nla_put_failure;
3889 if (set->flags != 0)
3890 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
3891 goto nla_put_failure;
3892
3893 if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
3894 goto nla_put_failure;
3895 if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
3896 goto nla_put_failure;
3897 if (set->flags & NFT_SET_MAP) {
3898 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
3899 goto nla_put_failure;
3900 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
3901 goto nla_put_failure;
3902 }
3903 if (set->flags & NFT_SET_OBJECT &&
3904 nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
3905 goto nla_put_failure;
3906
3907 if (set->timeout &&
3908 nla_put_be64(skb, NFTA_SET_TIMEOUT,
3909 nf_jiffies64_to_msecs(set->timeout),
3910 NFTA_SET_PAD))
3911 goto nla_put_failure;
3912 if (set->gc_int &&
3913 nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(set->gc_int)))
3914 goto nla_put_failure;
3915
3916 if (set->policy != NFT_SET_POL_PERFORMANCE) {
3917 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
3918 goto nla_put_failure;
3919 }
3920
3921 if (set->udata &&
3922 nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
3923 goto nla_put_failure;
3924
3925 nest = nla_nest_start_noflag(skb, NFTA_SET_DESC);
3926 if (!nest)
3927 goto nla_put_failure;
3928 if (set->size &&
3929 nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
3930 goto nla_put_failure;
3931
3932 if (set->field_count > 1 &&
3933 nf_tables_fill_set_concat(skb, set))
3934 goto nla_put_failure;
3935
3936 nla_nest_end(skb, nest);
3937
3938 if (set->num_exprs == 1) {
3939 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPR);
3940 if (nf_tables_fill_expr_info(skb, set->exprs[0]) < 0)
3941 goto nla_put_failure;
3942
3943 nla_nest_end(skb, nest);
3944 } else if (set->num_exprs > 1) {
3945 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPRESSIONS);
3946 if (nest == NULL)
3947 goto nla_put_failure;
3948
3949 for (i = 0; i < set->num_exprs; i++) {
3950 if (nft_expr_dump(skb, NFTA_LIST_ELEM,
3951 set->exprs[i]) < 0)
3952 goto nla_put_failure;
3953 }
3954 nla_nest_end(skb, nest);
3955 }
3956
3957 nlmsg_end(skb, nlh);
3958 return 0;
3959
3960 nla_put_failure:
3961 nlmsg_trim(skb, nlh);
3962 return -1;
3963 }
3964
3965 static void nf_tables_set_notify(const struct nft_ctx *ctx,
3966 const struct nft_set *set, int event,
3967 gfp_t gfp_flags)
3968 {
3969 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3970 u32 portid = ctx->portid;
3971 struct sk_buff *skb;
3972 u16 flags = 0;
3973 int err;
3974
3975 if (!ctx->report &&
3976 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3977 return;
3978
3979 skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
3980 if (skb == NULL)
3981 goto err;
3982
3983 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
3984 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
3985
3986 err = nf_tables_fill_set(skb, ctx, set, event, flags);
3987 if (err < 0) {
3988 kfree_skb(skb);
3989 goto err;
3990 }
3991
3992 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
3993 return;
3994 err:
3995 nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
3996 }
3997
3998 static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
3999 {
4000 const struct nft_set *set;
4001 unsigned int idx, s_idx = cb->args[0];
4002 struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
4003 struct net *net = sock_net(skb->sk);
4004 struct nft_ctx *ctx = cb->data, ctx_set;
4005 struct nftables_pernet *nft_net;
4006
4007 if (cb->args[1])
4008 return skb->len;
4009
4010 rcu_read_lock();
4011 nft_net = nft_pernet(net);
4012 cb->seq = nft_net->base_seq;
4013
4014 list_for_each_entry_rcu(table, &nft_net->tables, list) {
4015 if (ctx->family != NFPROTO_UNSPEC &&
4016 ctx->family != table->family)
4017 continue;
4018
4019 if (ctx->table && ctx->table != table)
4020 continue;
4021
4022 if (cur_table) {
4023 if (cur_table != table)
4024 continue;
4025
4026 cur_table = NULL;
4027 }
4028 idx = 0;
4029 list_for_each_entry_rcu(set, &table->sets, list) {
4030 if (idx < s_idx)
4031 goto cont;
4032 if (!nft_is_active(net, set))
4033 goto cont;
4034
4035 ctx_set = *ctx;
4036 ctx_set.table = table;
4037 ctx_set.family = table->family;
4038
4039 if (nf_tables_fill_set(skb, &ctx_set, set,
4040 NFT_MSG_NEWSET,
4041 NLM_F_MULTI) < 0) {
4042 cb->args[0] = idx;
4043 cb->args[2] = (unsigned long) table;
4044 goto done;
4045 }
4046 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
4047 cont:
4048 idx++;
4049 }
4050 if (s_idx)
4051 s_idx = 0;
4052 }
4053 cb->args[1] = 1;
4054 done:
4055 rcu_read_unlock();
4056 return skb->len;
4057 }
4058
4059 static int nf_tables_dump_sets_start(struct netlink_callback *cb)
4060 {
4061 struct nft_ctx *ctx_dump = NULL;
4062
4063 ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
4064 if (ctx_dump == NULL)
4065 return -ENOMEM;
4066
4067 cb->data = ctx_dump;
4068 return 0;
4069 }
4070
4071 static int nf_tables_dump_sets_done(struct netlink_callback *cb)
4072 {
4073 kfree(cb->data);
4074 return 0;
4075 }
4076
4077 /* called with rcu_read_lock held */
4078 static int nf_tables_getset(struct sk_buff *skb, const struct nfnl_info *info,
4079 const struct nlattr * const nla[])
4080 {
4081 struct netlink_ext_ack *extack = info->extack;
4082 u8 genmask = nft_genmask_cur(info->net);
4083 u8 family = info->nfmsg->nfgen_family;
4084 struct nft_table *table = NULL;
4085 struct net *net = info->net;
4086 const struct nft_set *set;
4087 struct sk_buff *skb2;
4088 struct nft_ctx ctx;
4089 int err;
4090
4091 if (nla[NFTA_SET_TABLE]) {
4092 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4093 genmask, 0);
4094 if (IS_ERR(table)) {
4095 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4096 return PTR_ERR(table);
4097 }
4098 }
4099
4100 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4101
4102 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
4103 struct netlink_dump_control c = {
4104 .start = nf_tables_dump_sets_start,
4105 .dump = nf_tables_dump_sets,
4106 .done = nf_tables_dump_sets_done,
4107 .data = &ctx,
4108 .module = THIS_MODULE,
4109 };
4110
4111 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
4112 }
4113
4114 /* Only accept unspec with dump */
4115 if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4116 return -EAFNOSUPPORT;
4117 if (!nla[NFTA_SET_TABLE])
4118 return -EINVAL;
4119
4120 set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4121 if (IS_ERR(set))
4122 return PTR_ERR(set);
4123
4124 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
4125 if (skb2 == NULL)
4126 return -ENOMEM;
4127
4128 err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
4129 if (err < 0)
4130 goto err_fill_set_info;
4131
4132 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
4133
4134 err_fill_set_info:
4135 kfree_skb(skb2);
4136 return err;
4137 }
4138
4139 static const struct nla_policy nft_concat_policy[NFTA_SET_FIELD_MAX + 1] = {
4140 [NFTA_SET_FIELD_LEN] = { .type = NLA_U32 },
4141 };
4142
4143 static int nft_set_desc_concat_parse(const struct nlattr *attr,
4144 struct nft_set_desc *desc)
4145 {
4146 struct nlattr *tb[NFTA_SET_FIELD_MAX + 1];
4147 u32 len;
4148 int err;
4149
4150 err = nla_parse_nested_deprecated(tb, NFTA_SET_FIELD_MAX, attr,
4151 nft_concat_policy, NULL);
4152 if (err < 0)
4153 return err;
4154
4155 if (!tb[NFTA_SET_FIELD_LEN])
4156 return -EINVAL;
4157
4158 len = ntohl(nla_get_be32(tb[NFTA_SET_FIELD_LEN]));
4159
4160 if (len * BITS_PER_BYTE / 32 > NFT_REG32_COUNT)
4161 return -E2BIG;
4162
4163 desc->field_len[desc->field_count++] = len;
4164
4165 return 0;
4166 }
4167
4168 static int nft_set_desc_concat(struct nft_set_desc *desc,
4169 const struct nlattr *nla)
4170 {
4171 struct nlattr *attr;
4172 int rem, err;
4173
4174 nla_for_each_nested(attr, nla, rem) {
4175 if (nla_type(attr) != NFTA_LIST_ELEM)
4176 return -EINVAL;
4177
4178 err = nft_set_desc_concat_parse(attr, desc);
4179 if (err < 0)
4180 return err;
4181 }
4182
4183 return 0;
4184 }
4185
4186 static int nf_tables_set_desc_parse(struct nft_set_desc *desc,
4187 const struct nlattr *nla)
4188 {
4189 struct nlattr *da[NFTA_SET_DESC_MAX + 1];
4190 int err;
4191
4192 err = nla_parse_nested_deprecated(da, NFTA_SET_DESC_MAX, nla,
4193 nft_set_desc_policy, NULL);
4194 if (err < 0)
4195 return err;
4196
4197 if (da[NFTA_SET_DESC_SIZE] != NULL)
4198 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
4199 if (da[NFTA_SET_DESC_CONCAT])
4200 err = nft_set_desc_concat(desc, da[NFTA_SET_DESC_CONCAT]);
4201
4202 return err;
4203 }
4204
4205 static int nf_tables_newset(struct sk_buff *skb, const struct nfnl_info *info,
4206 const struct nlattr * const nla[])
4207 {
4208 u32 ktype, dtype, flags, policy, gc_int, objtype;
4209 struct netlink_ext_ack *extack = info->extack;
4210 u8 genmask = nft_genmask_next(info->net);
4211 u8 family = info->nfmsg->nfgen_family;
4212 const struct nft_set_ops *ops;
4213 struct nft_expr *expr = NULL;
4214 struct net *net = info->net;
4215 struct nft_set_desc desc;
4216 struct nft_table *table;
4217 unsigned char *udata;
4218 struct nft_set *set;
4219 struct nft_ctx ctx;
4220 size_t alloc_size;
4221 u64 timeout;
4222 char *name;
4223 int err, i;
4224 u16 udlen;
4225 u64 size;
4226
4227 if (nla[NFTA_SET_TABLE] == NULL ||
4228 nla[NFTA_SET_NAME] == NULL ||
4229 nla[NFTA_SET_KEY_LEN] == NULL ||
4230 nla[NFTA_SET_ID] == NULL)
4231 return -EINVAL;
4232
4233 memset(&desc, 0, sizeof(desc));
4234
4235 ktype = NFT_DATA_VALUE;
4236 if (nla[NFTA_SET_KEY_TYPE] != NULL) {
4237 ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
4238 if ((ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
4239 return -EINVAL;
4240 }
4241
4242 desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
4243 if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
4244 return -EINVAL;
4245
4246 flags = 0;
4247 if (nla[NFTA_SET_FLAGS] != NULL) {
4248 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
4249 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
4250 NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
4251 NFT_SET_MAP | NFT_SET_EVAL |
4252 NFT_SET_OBJECT | NFT_SET_CONCAT | NFT_SET_EXPR))
4253 return -EOPNOTSUPP;
4254 /* Only one of these operations is supported */
4255 if ((flags & (NFT_SET_MAP | NFT_SET_OBJECT)) ==
4256 (NFT_SET_MAP | NFT_SET_OBJECT))
4257 return -EOPNOTSUPP;
4258 if ((flags & (NFT_SET_EVAL | NFT_SET_OBJECT)) ==
4259 (NFT_SET_EVAL | NFT_SET_OBJECT))
4260 return -EOPNOTSUPP;
4261 }
4262
4263 dtype = 0;
4264 if (nla[NFTA_SET_DATA_TYPE] != NULL) {
4265 if (!(flags & NFT_SET_MAP))
4266 return -EINVAL;
4267
4268 dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
4269 if ((dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
4270 dtype != NFT_DATA_VERDICT)
4271 return -EINVAL;
4272
4273 if (dtype != NFT_DATA_VERDICT) {
4274 if (nla[NFTA_SET_DATA_LEN] == NULL)
4275 return -EINVAL;
4276 desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
4277 if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
4278 return -EINVAL;
4279 } else
4280 desc.dlen = sizeof(struct nft_verdict);
4281 } else if (flags & NFT_SET_MAP)
4282 return -EINVAL;
4283
4284 if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
4285 if (!(flags & NFT_SET_OBJECT))
4286 return -EINVAL;
4287
4288 objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
4289 if (objtype == NFT_OBJECT_UNSPEC ||
4290 objtype > NFT_OBJECT_MAX)
4291 return -EOPNOTSUPP;
4292 } else if (flags & NFT_SET_OBJECT)
4293 return -EINVAL;
4294 else
4295 objtype = NFT_OBJECT_UNSPEC;
4296
4297 timeout = 0;
4298 if (nla[NFTA_SET_TIMEOUT] != NULL) {
4299 if (!(flags & NFT_SET_TIMEOUT))
4300 return -EINVAL;
4301
4302 err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &timeout);
4303 if (err)
4304 return err;
4305 }
4306 gc_int = 0;
4307 if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
4308 if (!(flags & NFT_SET_TIMEOUT))
4309 return -EINVAL;
4310 gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
4311 }
4312
4313 policy = NFT_SET_POL_PERFORMANCE;
4314 if (nla[NFTA_SET_POLICY] != NULL)
4315 policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
4316
4317 if (nla[NFTA_SET_DESC] != NULL) {
4318 err = nf_tables_set_desc_parse(&desc, nla[NFTA_SET_DESC]);
4319 if (err < 0)
4320 return err;
4321 }
4322
4323 if (nla[NFTA_SET_EXPR] || nla[NFTA_SET_EXPRESSIONS])
4324 desc.expr = true;
4325
4326 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask,
4327 NETLINK_CB(skb).portid);
4328 if (IS_ERR(table)) {
4329 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4330 return PTR_ERR(table);
4331 }
4332
4333 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4334
4335 set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4336 if (IS_ERR(set)) {
4337 if (PTR_ERR(set) != -ENOENT) {
4338 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4339 return PTR_ERR(set);
4340 }
4341 } else {
4342 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
4343 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4344 return -EEXIST;
4345 }
4346 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
4347 return -EOPNOTSUPP;
4348
4349 return 0;
4350 }
4351
4352 if (!(info->nlh->nlmsg_flags & NLM_F_CREATE))
4353 return -ENOENT;
4354
4355 ops = nft_select_set_ops(&ctx, nla, &desc, policy);
4356 if (IS_ERR(ops))
4357 return PTR_ERR(ops);
4358
4359 udlen = 0;
4360 if (nla[NFTA_SET_USERDATA])
4361 udlen = nla_len(nla[NFTA_SET_USERDATA]);
4362
4363 size = 0;
4364 if (ops->privsize != NULL)
4365 size = ops->privsize(nla, &desc);
4366 alloc_size = sizeof(*set) + size + udlen;
4367 if (alloc_size < size || alloc_size > INT_MAX)
4368 return -ENOMEM;
4369 set = kvzalloc(alloc_size, GFP_KERNEL);
4370 if (!set)
4371 return -ENOMEM;
4372
4373 name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL);
4374 if (!name) {
4375 err = -ENOMEM;
4376 goto err_set_name;
4377 }
4378
4379 err = nf_tables_set_alloc_name(&ctx, set, name);
4380 kfree(name);
4381 if (err < 0)
4382 goto err_set_name;
4383
4384 udata = NULL;
4385 if (udlen) {
4386 udata = set->data + size;
4387 nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
4388 }
4389
4390 INIT_LIST_HEAD(&set->bindings);
4391 INIT_LIST_HEAD(&set->catchall_list);
4392 set->table = table;
4393 write_pnet(&set->net, net);
4394 set->ops = ops;
4395 set->ktype = ktype;
4396 set->klen = desc.klen;
4397 set->dtype = dtype;
4398 set->objtype = objtype;
4399 set->dlen = desc.dlen;
4400 set->flags = flags;
4401 set->size = desc.size;
4402 set->policy = policy;
4403 set->udlen = udlen;
4404 set->udata = udata;
4405 set->timeout = timeout;
4406 set->gc_int = gc_int;
4407
4408 set->field_count = desc.field_count;
4409 for (i = 0; i < desc.field_count; i++)
4410 set->field_len[i] = desc.field_len[i];
4411
4412 err = ops->init(set, &desc, nla);
4413 if (err < 0)
4414 goto err_set_init;
4415
4416 if (nla[NFTA_SET_EXPR]) {
4417 expr = nft_set_elem_expr_alloc(&ctx, set, nla[NFTA_SET_EXPR]);
4418 if (IS_ERR(expr)) {
4419 err = PTR_ERR(expr);
4420 goto err_set_expr_alloc;
4421 }
4422 set->exprs[0] = expr;
4423 set->num_exprs++;
4424 } else if (nla[NFTA_SET_EXPRESSIONS]) {
4425 struct nft_expr *expr;
4426 struct nlattr *tmp;
4427 int left;
4428
4429 if (!(flags & NFT_SET_EXPR)) {
4430 err = -EINVAL;
4431 goto err_set_expr_alloc;
4432 }
4433 i = 0;
4434 nla_for_each_nested(tmp, nla[NFTA_SET_EXPRESSIONS], left) {
4435 if (i == NFT_SET_EXPR_MAX) {
4436 err = -E2BIG;
4437 goto err_set_expr_alloc;
4438 }
4439 if (nla_type(tmp) != NFTA_LIST_ELEM) {
4440 err = -EINVAL;
4441 goto err_set_expr_alloc;
4442 }
4443 expr = nft_set_elem_expr_alloc(&ctx, set, tmp);
4444 if (IS_ERR(expr)) {
4445 err = PTR_ERR(expr);
4446 goto err_set_expr_alloc;
4447 }
4448 set->exprs[i++] = expr;
4449 set->num_exprs++;
4450 }
4451 }
4452
4453 set->handle = nf_tables_alloc_handle(table);
4454
4455 err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
4456 if (err < 0)
4457 goto err_set_expr_alloc;
4458
4459 list_add_tail_rcu(&set->list, &table->sets);
4460 table->use++;
4461 return 0;
4462
4463 err_set_expr_alloc:
4464 for (i = 0; i < set->num_exprs; i++)
4465 nft_expr_destroy(&ctx, set->exprs[i]);
4466
4467 ops->destroy(set);
4468 err_set_init:
4469 kfree(set->name);
4470 err_set_name:
4471 kvfree(set);
4472 return err;
4473 }
4474
4475 struct nft_set_elem_catchall {
4476 struct list_head list;
4477 struct rcu_head rcu;
4478 void *elem;
4479 };
4480
4481 static void nft_set_catchall_destroy(const struct nft_ctx *ctx,
4482 struct nft_set *set)
4483 {
4484 struct nft_set_elem_catchall *catchall;
4485
4486 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4487 list_del_rcu(&catchall->list);
4488 nft_set_elem_destroy(set, catchall->elem, true);
4489 kfree_rcu(catchall);
4490 }
4491 }
4492
4493 static void nft_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
4494 {
4495 int i;
4496
4497 if (WARN_ON(set->use > 0))
4498 return;
4499
4500 for (i = 0; i < set->num_exprs; i++)
4501 nft_expr_destroy(ctx, set->exprs[i]);
4502
4503 set->ops->destroy(set);
4504 nft_set_catchall_destroy(ctx, set);
4505 kfree(set->name);
4506 kvfree(set);
4507 }
4508
4509 static int nf_tables_delset(struct sk_buff *skb, const struct nfnl_info *info,
4510 const struct nlattr * const nla[])
4511 {
4512 struct netlink_ext_ack *extack = info->extack;
4513 u8 genmask = nft_genmask_next(info->net);
4514 u8 family = info->nfmsg->nfgen_family;
4515 struct net *net = info->net;
4516 const struct nlattr *attr;
4517 struct nft_table *table;
4518 struct nft_set *set;
4519 struct nft_ctx ctx;
4520
4521 if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4522 return -EAFNOSUPPORT;
4523
4524 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4525 genmask, NETLINK_CB(skb).portid);
4526 if (IS_ERR(table)) {
4527 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4528 return PTR_ERR(table);
4529 }
4530
4531 if (nla[NFTA_SET_HANDLE]) {
4532 attr = nla[NFTA_SET_HANDLE];
4533 set = nft_set_lookup_byhandle(table, attr, genmask);
4534 } else {
4535 attr = nla[NFTA_SET_NAME];
4536 set = nft_set_lookup(table, attr, genmask);
4537 }
4538
4539 if (IS_ERR(set)) {
4540 NL_SET_BAD_ATTR(extack, attr);
4541 return PTR_ERR(set);
4542 }
4543 if (set->use ||
4544 (info->nlh->nlmsg_flags & NLM_F_NONREC &&
4545 atomic_read(&set->nelems) > 0)) {
4546 NL_SET_BAD_ATTR(extack, attr);
4547 return -EBUSY;
4548 }
4549
4550 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4551
4552 return nft_delset(&ctx, set);
4553 }
4554
4555 static int nft_validate_register_store(const struct nft_ctx *ctx,
4556 enum nft_registers reg,
4557 const struct nft_data *data,
4558 enum nft_data_types type,
4559 unsigned int len);
4560
4561 static int nft_setelem_data_validate(const struct nft_ctx *ctx,
4562 struct nft_set *set,
4563 struct nft_set_elem *elem)
4564 {
4565 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4566 enum nft_registers dreg;
4567
4568 dreg = nft_type_to_reg(set->dtype);
4569 return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
4570 set->dtype == NFT_DATA_VERDICT ?
4571 NFT_DATA_VERDICT : NFT_DATA_VALUE,
4572 set->dlen);
4573 }
4574
4575 static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
4576 struct nft_set *set,
4577 const struct nft_set_iter *iter,
4578 struct nft_set_elem *elem)
4579 {
4580 return nft_setelem_data_validate(ctx, set, elem);
4581 }
4582
4583 static int nft_set_catchall_bind_check(const struct nft_ctx *ctx,
4584 struct nft_set *set)
4585 {
4586 u8 genmask = nft_genmask_next(ctx->net);
4587 struct nft_set_elem_catchall *catchall;
4588 struct nft_set_elem elem;
4589 struct nft_set_ext *ext;
4590 int ret = 0;
4591
4592 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4593 ext = nft_set_elem_ext(set, catchall->elem);
4594 if (!nft_set_elem_active(ext, genmask))
4595 continue;
4596
4597 elem.priv = catchall->elem;
4598 ret = nft_setelem_data_validate(ctx, set, &elem);
4599 if (ret < 0)
4600 break;
4601 }
4602
4603 return ret;
4604 }
4605
4606 int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
4607 struct nft_set_binding *binding)
4608 {
4609 struct nft_set_binding *i;
4610 struct nft_set_iter iter;
4611
4612 if (set->use == UINT_MAX)
4613 return -EOVERFLOW;
4614
4615 if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
4616 return -EBUSY;
4617
4618 if (binding->flags & NFT_SET_MAP) {
4619 /* If the set is already bound to the same chain all
4620 * jumps are already validated for that chain.
4621 */
4622 list_for_each_entry(i, &set->bindings, list) {
4623 if (i->flags & NFT_SET_MAP &&
4624 i->chain == binding->chain)
4625 goto bind;
4626 }
4627
4628 iter.genmask = nft_genmask_next(ctx->net);
4629 iter.skip = 0;
4630 iter.count = 0;
4631 iter.err = 0;
4632 iter.fn = nf_tables_bind_check_setelem;
4633
4634 set->ops->walk(ctx, set, &iter);
4635 if (!iter.err)
4636 iter.err = nft_set_catchall_bind_check(ctx, set);
4637
4638 if (iter.err < 0)
4639 return iter.err;
4640 }
4641 bind:
4642 binding->chain = ctx->chain;
4643 list_add_tail_rcu(&binding->list, &set->bindings);
4644 nft_set_trans_bind(ctx, set);
4645 set->use++;
4646
4647 return 0;
4648 }
4649 EXPORT_SYMBOL_GPL(nf_tables_bind_set);
4650
4651 static void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
4652 struct nft_set_binding *binding, bool event)
4653 {
4654 list_del_rcu(&binding->list);
4655
4656 if (list_empty(&set->bindings) && nft_set_is_anonymous(set)) {
4657 list_del_rcu(&set->list);
4658 if (event)
4659 nf_tables_set_notify(ctx, set, NFT_MSG_DELSET,
4660 GFP_KERNEL);
4661 }
4662 }
4663
4664 void nf_tables_deactivate_set(const struct nft_ctx *ctx, struct nft_set *set,
4665 struct nft_set_binding *binding,
4666 enum nft_trans_phase phase)
4667 {
4668 switch (phase) {
4669 case NFT_TRANS_PREPARE:
4670 set->use--;
4671 return;
4672 case NFT_TRANS_ABORT:
4673 case NFT_TRANS_RELEASE:
4674 set->use--;
4675 fallthrough;
4676 default:
4677 nf_tables_unbind_set(ctx, set, binding,
4678 phase == NFT_TRANS_COMMIT);
4679 }
4680 }
4681 EXPORT_SYMBOL_GPL(nf_tables_deactivate_set);
4682
4683 void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
4684 {
4685 if (list_empty(&set->bindings) && nft_set_is_anonymous(set))
4686 nft_set_destroy(ctx, set);
4687 }
4688 EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
4689
4690 const struct nft_set_ext_type nft_set_ext_types[] = {
4691 [NFT_SET_EXT_KEY] = {
4692 .align = __alignof__(u32),
4693 },
4694 [NFT_SET_EXT_DATA] = {
4695 .align = __alignof__(u32),
4696 },
4697 [NFT_SET_EXT_EXPRESSIONS] = {
4698 .align = __alignof__(struct nft_set_elem_expr),
4699 },
4700 [NFT_SET_EXT_OBJREF] = {
4701 .len = sizeof(struct nft_object *),
4702 .align = __alignof__(struct nft_object *),
4703 },
4704 [NFT_SET_EXT_FLAGS] = {
4705 .len = sizeof(u8),
4706 .align = __alignof__(u8),
4707 },
4708 [NFT_SET_EXT_TIMEOUT] = {
4709 .len = sizeof(u64),
4710 .align = __alignof__(u64),
4711 },
4712 [NFT_SET_EXT_EXPIRATION] = {
4713 .len = sizeof(u64),
4714 .align = __alignof__(u64),
4715 },
4716 [NFT_SET_EXT_USERDATA] = {
4717 .len = sizeof(struct nft_userdata),
4718 .align = __alignof__(struct nft_userdata),
4719 },
4720 [NFT_SET_EXT_KEY_END] = {
4721 .align = __alignof__(u32),
4722 },
4723 };
4724
4725 /*
4726 * Set elements
4727 */
4728
4729 static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
4730 [NFTA_SET_ELEM_KEY] = { .type = NLA_NESTED },
4731 [NFTA_SET_ELEM_DATA] = { .type = NLA_NESTED },
4732 [NFTA_SET_ELEM_FLAGS] = { .type = NLA_U32 },
4733 [NFTA_SET_ELEM_TIMEOUT] = { .type = NLA_U64 },
4734 [NFTA_SET_ELEM_EXPIRATION] = { .type = NLA_U64 },
4735 [NFTA_SET_ELEM_USERDATA] = { .type = NLA_BINARY,
4736 .len = NFT_USERDATA_MAXLEN },
4737 [NFTA_SET_ELEM_EXPR] = { .type = NLA_NESTED },
4738 [NFTA_SET_ELEM_OBJREF] = { .type = NLA_STRING,
4739 .len = NFT_OBJ_MAXNAMELEN - 1 },
4740 [NFTA_SET_ELEM_KEY_END] = { .type = NLA_NESTED },
4741 [NFTA_SET_ELEM_EXPRESSIONS] = { .type = NLA_NESTED },
4742 };
4743
4744 static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
4745 [NFTA_SET_ELEM_LIST_TABLE] = { .type = NLA_STRING,
4746 .len = NFT_TABLE_MAXNAMELEN - 1 },
4747 [NFTA_SET_ELEM_LIST_SET] = { .type = NLA_STRING,
4748 .len = NFT_SET_MAXNAMELEN - 1 },
4749 [NFTA_SET_ELEM_LIST_ELEMENTS] = { .type = NLA_NESTED },
4750 [NFTA_SET_ELEM_LIST_SET_ID] = { .type = NLA_U32 },
4751 };
4752
4753 static int nft_set_elem_expr_dump(struct sk_buff *skb,
4754 const struct nft_set *set,
4755 const struct nft_set_ext *ext)
4756 {
4757 struct nft_set_elem_expr *elem_expr;
4758 u32 size, num_exprs = 0;
4759 struct nft_expr *expr;
4760 struct nlattr *nest;
4761
4762 elem_expr = nft_set_ext_expr(ext);
4763 nft_setelem_expr_foreach(expr, elem_expr, size)
4764 num_exprs++;
4765
4766 if (num_exprs == 1) {
4767 expr = nft_setelem_expr_at(elem_expr, 0);
4768 if (nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, expr) < 0)
4769 return -1;
4770
4771 return 0;
4772 } else if (num_exprs > 1) {
4773 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_EXPRESSIONS);
4774 if (nest == NULL)
4775 goto nla_put_failure;
4776
4777 nft_setelem_expr_foreach(expr, elem_expr, size) {
4778 expr = nft_setelem_expr_at(elem_expr, size);
4779 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
4780 goto nla_put_failure;
4781 }
4782 nla_nest_end(skb, nest);
4783 }
4784 return 0;
4785
4786 nla_put_failure:
4787 return -1;
4788 }
4789
4790 static int nf_tables_fill_setelem(struct sk_buff *skb,
4791 const struct nft_set *set,
4792 const struct nft_set_elem *elem)
4793 {
4794 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4795 unsigned char *b = skb_tail_pointer(skb);
4796 struct nlattr *nest;
4797
4798 nest = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4799 if (nest == NULL)
4800 goto nla_put_failure;
4801
4802 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
4803 nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
4804 NFT_DATA_VALUE, set->klen) < 0)
4805 goto nla_put_failure;
4806
4807 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
4808 nft_data_dump(skb, NFTA_SET_ELEM_KEY_END, nft_set_ext_key_end(ext),
4809 NFT_DATA_VALUE, set->klen) < 0)
4810 goto nla_put_failure;
4811
4812 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
4813 nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
4814 set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
4815 set->dlen) < 0)
4816 goto nla_put_failure;
4817
4818 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS) &&
4819 nft_set_elem_expr_dump(skb, set, ext))
4820 goto nla_put_failure;
4821
4822 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
4823 nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
4824 (*nft_set_ext_obj(ext))->key.name) < 0)
4825 goto nla_put_failure;
4826
4827 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
4828 nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
4829 htonl(*nft_set_ext_flags(ext))))
4830 goto nla_put_failure;
4831
4832 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
4833 nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
4834 nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
4835 NFTA_SET_ELEM_PAD))
4836 goto nla_put_failure;
4837
4838 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
4839 u64 expires, now = get_jiffies_64();
4840
4841 expires = *nft_set_ext_expiration(ext);
4842 if (time_before64(now, expires))
4843 expires -= now;
4844 else
4845 expires = 0;
4846
4847 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
4848 nf_jiffies64_to_msecs(expires),
4849 NFTA_SET_ELEM_PAD))
4850 goto nla_put_failure;
4851 }
4852
4853 if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
4854 struct nft_userdata *udata;
4855
4856 udata = nft_set_ext_userdata(ext);
4857 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
4858 udata->len + 1, udata->data))
4859 goto nla_put_failure;
4860 }
4861
4862 nla_nest_end(skb, nest);
4863 return 0;
4864
4865 nla_put_failure:
4866 nlmsg_trim(skb, b);
4867 return -EMSGSIZE;
4868 }
4869
4870 struct nft_set_dump_args {
4871 const struct netlink_callback *cb;
4872 struct nft_set_iter iter;
4873 struct sk_buff *skb;
4874 };
4875
4876 static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
4877 struct nft_set *set,
4878 const struct nft_set_iter *iter,
4879 struct nft_set_elem *elem)
4880 {
4881 struct nft_set_dump_args *args;
4882
4883 args = container_of(iter, struct nft_set_dump_args, iter);
4884 return nf_tables_fill_setelem(args->skb, set, elem);
4885 }
4886
4887 struct nft_set_dump_ctx {
4888 const struct nft_set *set;
4889 struct nft_ctx ctx;
4890 };
4891
4892 static int nft_set_catchall_dump(struct net *net, struct sk_buff *skb,
4893 const struct nft_set *set)
4894 {
4895 struct nft_set_elem_catchall *catchall;
4896 u8 genmask = nft_genmask_cur(net);
4897 struct nft_set_elem elem;
4898 struct nft_set_ext *ext;
4899 int ret = 0;
4900
4901 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4902 ext = nft_set_elem_ext(set, catchall->elem);
4903 if (!nft_set_elem_active(ext, genmask) ||
4904 nft_set_elem_expired(ext))
4905 continue;
4906
4907 elem.priv = catchall->elem;
4908 ret = nf_tables_fill_setelem(skb, set, &elem);
4909 break;
4910 }
4911
4912 return ret;
4913 }
4914
4915 static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
4916 {
4917 struct nft_set_dump_ctx *dump_ctx = cb->data;
4918 struct net *net = sock_net(skb->sk);
4919 struct nftables_pernet *nft_net;
4920 struct nft_table *table;
4921 struct nft_set *set;
4922 struct nft_set_dump_args args;
4923 bool set_found = false;
4924 struct nlmsghdr *nlh;
4925 struct nlattr *nest;
4926 u32 portid, seq;
4927 int event;
4928
4929 rcu_read_lock();
4930 nft_net = nft_pernet(net);
4931 list_for_each_entry_rcu(table, &nft_net->tables, list) {
4932 if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
4933 dump_ctx->ctx.family != table->family)
4934 continue;
4935
4936 if (table != dump_ctx->ctx.table)
4937 continue;
4938
4939 list_for_each_entry_rcu(set, &table->sets, list) {
4940 if (set == dump_ctx->set) {
4941 set_found = true;
4942 break;
4943 }
4944 }
4945 break;
4946 }
4947
4948 if (!set_found) {
4949 rcu_read_unlock();
4950 return -ENOENT;
4951 }
4952
4953 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
4954 portid = NETLINK_CB(cb->skb).portid;
4955 seq = cb->nlh->nlmsg_seq;
4956
4957 nlh = nfnl_msg_put(skb, portid, seq, event, NLM_F_MULTI,
4958 table->family, NFNETLINK_V0, nft_base_seq(net));
4959 if (!nlh)
4960 goto nla_put_failure;
4961
4962 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
4963 goto nla_put_failure;
4964 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
4965 goto nla_put_failure;
4966
4967 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
4968 if (nest == NULL)
4969 goto nla_put_failure;
4970
4971 args.cb = cb;
4972 args.skb = skb;
4973 args.iter.genmask = nft_genmask_cur(net);
4974 args.iter.skip = cb->args[0];
4975 args.iter.count = 0;
4976 args.iter.err = 0;
4977 args.iter.fn = nf_tables_dump_setelem;
4978 set->ops->walk(&dump_ctx->ctx, set, &args.iter);
4979
4980 if (!args.iter.err && args.iter.count == cb->args[0])
4981 args.iter.err = nft_set_catchall_dump(net, skb, set);
4982 rcu_read_unlock();
4983
4984 nla_nest_end(skb, nest);
4985 nlmsg_end(skb, nlh);
4986
4987 if (args.iter.err && args.iter.err != -EMSGSIZE)
4988 return args.iter.err;
4989 if (args.iter.count == cb->args[0])
4990 return 0;
4991
4992 cb->args[0] = args.iter.count;
4993 return skb->len;
4994
4995 nla_put_failure:
4996 rcu_read_unlock();
4997 return -ENOSPC;
4998 }
4999
5000 static int nf_tables_dump_set_start(struct netlink_callback *cb)
5001 {
5002 struct nft_set_dump_ctx *dump_ctx = cb->data;
5003
5004 cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
5005
5006 return cb->data ? 0 : -ENOMEM;
5007 }
5008
5009 static int nf_tables_dump_set_done(struct netlink_callback *cb)
5010 {
5011 kfree(cb->data);
5012 return 0;
5013 }
5014
5015 static int nf_tables_fill_setelem_info(struct sk_buff *skb,
5016 const struct nft_ctx *ctx, u32 seq,
5017 u32 portid, int event, u16 flags,
5018 const struct nft_set *set,
5019 const struct nft_set_elem *elem)
5020 {
5021 struct nlmsghdr *nlh;
5022 struct nlattr *nest;
5023 int err;
5024
5025 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
5026 nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
5027 NFNETLINK_V0, nft_base_seq(ctx->net));
5028 if (!nlh)
5029 goto nla_put_failure;
5030
5031 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
5032 goto nla_put_failure;
5033 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
5034 goto nla_put_failure;
5035
5036 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5037 if (nest == NULL)
5038 goto nla_put_failure;
5039
5040 err = nf_tables_fill_setelem(skb, set, elem);
5041 if (err < 0)
5042 goto nla_put_failure;
5043
5044 nla_nest_end(skb, nest);
5045
5046 nlmsg_end(skb, nlh);
5047 return 0;
5048
5049 nla_put_failure:
5050 nlmsg_trim(skb, nlh);
5051 return -1;
5052 }
5053
5054 static int nft_setelem_parse_flags(const struct nft_set *set,
5055 const struct nlattr *attr, u32 *flags)
5056 {
5057 if (attr == NULL)
5058 return 0;
5059
5060 *flags = ntohl(nla_get_be32(attr));
5061 if (*flags & ~(NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5062 return -EOPNOTSUPP;
5063 if (!(set->flags & NFT_SET_INTERVAL) &&
5064 *flags & NFT_SET_ELEM_INTERVAL_END)
5065 return -EINVAL;
5066
5067 return 0;
5068 }
5069
5070 static int nft_setelem_parse_key(struct nft_ctx *ctx, struct nft_set *set,
5071 struct nft_data *key, struct nlattr *attr)
5072 {
5073 struct nft_data_desc desc;
5074 int err;
5075
5076 err = nft_data_init(ctx, key, NFT_DATA_VALUE_MAXLEN, &desc, attr);
5077 if (err < 0)
5078 return err;
5079
5080 if (desc.type != NFT_DATA_VALUE || desc.len != set->klen) {
5081 nft_data_release(key, desc.type);
5082 return -EINVAL;
5083 }
5084
5085 return 0;
5086 }
5087
5088 static int nft_setelem_parse_data(struct nft_ctx *ctx, struct nft_set *set,
5089 struct nft_data_desc *desc,
5090 struct nft_data *data,
5091 struct nlattr *attr)
5092 {
5093 int err;
5094
5095 err = nft_data_init(ctx, data, NFT_DATA_VALUE_MAXLEN, desc, attr);
5096 if (err < 0)
5097 return err;
5098
5099 if (desc->type != NFT_DATA_VERDICT && desc->len != set->dlen) {
5100 nft_data_release(data, desc->type);
5101 return -EINVAL;
5102 }
5103
5104 return 0;
5105 }
5106
5107 static void *nft_setelem_catchall_get(const struct net *net,
5108 const struct nft_set *set)
5109 {
5110 struct nft_set_elem_catchall *catchall;
5111 u8 genmask = nft_genmask_cur(net);
5112 struct nft_set_ext *ext;
5113 void *priv = NULL;
5114
5115 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5116 ext = nft_set_elem_ext(set, catchall->elem);
5117 if (!nft_set_elem_active(ext, genmask) ||
5118 nft_set_elem_expired(ext))
5119 continue;
5120
5121 priv = catchall->elem;
5122 break;
5123 }
5124
5125 return priv;
5126 }
5127
5128 static int nft_setelem_get(struct nft_ctx *ctx, struct nft_set *set,
5129 struct nft_set_elem *elem, u32 flags)
5130 {
5131 void *priv;
5132
5133 if (!(flags & NFT_SET_ELEM_CATCHALL)) {
5134 priv = set->ops->get(ctx->net, set, elem, flags);
5135 if (IS_ERR(priv))
5136 return PTR_ERR(priv);
5137 } else {
5138 priv = nft_setelem_catchall_get(ctx->net, set);
5139 if (!priv)
5140 return -ENOENT;
5141 }
5142 elem->priv = priv;
5143
5144 return 0;
5145 }
5146
5147 static int nft_get_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5148 const struct nlattr *attr)
5149 {
5150 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5151 struct nft_set_elem elem;
5152 struct sk_buff *skb;
5153 uint32_t flags = 0;
5154 int err;
5155
5156 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5157 nft_set_elem_policy, NULL);
5158 if (err < 0)
5159 return err;
5160
5161 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5162 if (err < 0)
5163 return err;
5164
5165 if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5166 return -EINVAL;
5167
5168 if (nla[NFTA_SET_ELEM_KEY]) {
5169 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5170 nla[NFTA_SET_ELEM_KEY]);
5171 if (err < 0)
5172 return err;
5173 }
5174
5175 if (nla[NFTA_SET_ELEM_KEY_END]) {
5176 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5177 nla[NFTA_SET_ELEM_KEY_END]);
5178 if (err < 0)
5179 return err;
5180 }
5181
5182 err = nft_setelem_get(ctx, set, &elem, flags);
5183 if (err < 0)
5184 return err;
5185
5186 err = -ENOMEM;
5187 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
5188 if (skb == NULL)
5189 return err;
5190
5191 err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
5192 NFT_MSG_NEWSETELEM, 0, set, &elem);
5193 if (err < 0)
5194 goto err_fill_setelem;
5195
5196 return nfnetlink_unicast(skb, ctx->net, ctx->portid);
5197
5198 err_fill_setelem:
5199 kfree_skb(skb);
5200 return err;
5201 }
5202
5203 /* called with rcu_read_lock held */
5204 static int nf_tables_getsetelem(struct sk_buff *skb,
5205 const struct nfnl_info *info,
5206 const struct nlattr * const nla[])
5207 {
5208 struct netlink_ext_ack *extack = info->extack;
5209 u8 genmask = nft_genmask_cur(info->net);
5210 u8 family = info->nfmsg->nfgen_family;
5211 struct net *net = info->net;
5212 struct nft_table *table;
5213 struct nft_set *set;
5214 struct nlattr *attr;
5215 struct nft_ctx ctx;
5216 int rem, err = 0;
5217
5218 table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
5219 genmask, NETLINK_CB(skb).portid);
5220 if (IS_ERR(table)) {
5221 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
5222 return PTR_ERR(table);
5223 }
5224
5225 set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
5226 if (IS_ERR(set))
5227 return PTR_ERR(set);
5228
5229 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5230
5231 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
5232 struct netlink_dump_control c = {
5233 .start = nf_tables_dump_set_start,
5234 .dump = nf_tables_dump_set,
5235 .done = nf_tables_dump_set_done,
5236 .module = THIS_MODULE,
5237 };
5238 struct nft_set_dump_ctx dump_ctx = {
5239 .set = set,
5240 .ctx = ctx,
5241 };
5242
5243 c.data = &dump_ctx;
5244 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
5245 }
5246
5247 if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
5248 return -EINVAL;
5249
5250 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
5251 err = nft_get_set_elem(&ctx, set, attr);
5252 if (err < 0)
5253 break;
5254 }
5255
5256 return err;
5257 }
5258
5259 static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
5260 const struct nft_set *set,
5261 const struct nft_set_elem *elem,
5262 int event)
5263 {
5264 struct nftables_pernet *nft_net;
5265 struct net *net = ctx->net;
5266 u32 portid = ctx->portid;
5267 struct sk_buff *skb;
5268 u16 flags = 0;
5269 int err;
5270
5271 if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5272 return;
5273
5274 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5275 if (skb == NULL)
5276 goto err;
5277
5278 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
5279 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
5280
5281 err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
5282 set, elem);
5283 if (err < 0) {
5284 kfree_skb(skb);
5285 goto err;
5286 }
5287
5288 nft_net = nft_pernet(net);
5289 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
5290 return;
5291 err:
5292 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
5293 }
5294
5295 static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
5296 int msg_type,
5297 struct nft_set *set)
5298 {
5299 struct nft_trans *trans;
5300
5301 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
5302 if (trans == NULL)
5303 return NULL;
5304
5305 nft_trans_elem_set(trans) = set;
5306 return trans;
5307 }
5308
5309 struct nft_expr *nft_set_elem_expr_alloc(const struct nft_ctx *ctx,
5310 const struct nft_set *set,
5311 const struct nlattr *attr)
5312 {
5313 struct nft_expr *expr;
5314 int err;
5315
5316 expr = nft_expr_init(ctx, attr);
5317 if (IS_ERR(expr))
5318 return expr;
5319
5320 err = -EOPNOTSUPP;
5321 if (!(expr->ops->type->flags & NFT_EXPR_STATEFUL))
5322 goto err_set_elem_expr;
5323
5324 if (expr->ops->type->flags & NFT_EXPR_GC) {
5325 if (set->flags & NFT_SET_TIMEOUT)
5326 goto err_set_elem_expr;
5327 if (!set->ops->gc_init)
5328 goto err_set_elem_expr;
5329 set->ops->gc_init(set);
5330 }
5331
5332 return expr;
5333
5334 err_set_elem_expr:
5335 nft_expr_destroy(ctx, expr);
5336 return ERR_PTR(err);
5337 }
5338
5339 void *nft_set_elem_init(const struct nft_set *set,
5340 const struct nft_set_ext_tmpl *tmpl,
5341 const u32 *key, const u32 *key_end,
5342 const u32 *data, u64 timeout, u64 expiration, gfp_t gfp)
5343 {
5344 struct nft_set_ext *ext;
5345 void *elem;
5346
5347 elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
5348 if (elem == NULL)
5349 return NULL;
5350
5351 ext = nft_set_elem_ext(set, elem);
5352 nft_set_ext_init(ext, tmpl);
5353
5354 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY))
5355 memcpy(nft_set_ext_key(ext), key, set->klen);
5356 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END))
5357 memcpy(nft_set_ext_key_end(ext), key_end, set->klen);
5358 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5359 memcpy(nft_set_ext_data(ext), data, set->dlen);
5360 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5361 *nft_set_ext_expiration(ext) = get_jiffies_64() + expiration;
5362 if (expiration == 0)
5363 *nft_set_ext_expiration(ext) += timeout;
5364 }
5365 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
5366 *nft_set_ext_timeout(ext) = timeout;
5367
5368 return elem;
5369 }
5370
5371 static void __nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5372 struct nft_expr *expr)
5373 {
5374 if (expr->ops->destroy_clone) {
5375 expr->ops->destroy_clone(ctx, expr);
5376 module_put(expr->ops->type->owner);
5377 } else {
5378 nf_tables_expr_destroy(ctx, expr);
5379 }
5380 }
5381
5382 static void nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5383 struct nft_set_elem_expr *elem_expr)
5384 {
5385 struct nft_expr *expr;
5386 u32 size;
5387
5388 nft_setelem_expr_foreach(expr, elem_expr, size)
5389 __nft_set_elem_expr_destroy(ctx, expr);
5390 }
5391
5392 void nft_set_elem_destroy(const struct nft_set *set, void *elem,
5393 bool destroy_expr)
5394 {
5395 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5396 struct nft_ctx ctx = {
5397 .net = read_pnet(&set->net),
5398 .family = set->table->family,
5399 };
5400
5401 nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
5402 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5403 nft_data_release(nft_set_ext_data(ext), set->dtype);
5404 if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5405 nft_set_elem_expr_destroy(&ctx, nft_set_ext_expr(ext));
5406
5407 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
5408 (*nft_set_ext_obj(ext))->use--;
5409 kfree(elem);
5410 }
5411 EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
5412
5413 /* Only called from commit path, nft_setelem_data_deactivate() already deals
5414 * with the refcounting from the preparation phase.
5415 */
5416 static void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
5417 const struct nft_set *set, void *elem)
5418 {
5419 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5420
5421 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5422 nft_set_elem_expr_destroy(ctx, nft_set_ext_expr(ext));
5423
5424 kfree(elem);
5425 }
5426
5427 int nft_set_elem_expr_clone(const struct nft_ctx *ctx, struct nft_set *set,
5428 struct nft_expr *expr_array[])
5429 {
5430 struct nft_expr *expr;
5431 int err, i, k;
5432
5433 for (i = 0; i < set->num_exprs; i++) {
5434 expr = kzalloc(set->exprs[i]->ops->size, GFP_KERNEL);
5435 if (!expr)
5436 goto err_expr;
5437
5438 err = nft_expr_clone(expr, set->exprs[i]);
5439 if (err < 0) {
5440 nft_expr_destroy(ctx, expr);
5441 goto err_expr;
5442 }
5443 expr_array[i] = expr;
5444 }
5445
5446 return 0;
5447
5448 err_expr:
5449 for (k = i - 1; k >= 0; k--)
5450 nft_expr_destroy(ctx, expr_array[k]);
5451
5452 return -ENOMEM;
5453 }
5454
5455 static int nft_set_elem_expr_setup(struct nft_ctx *ctx,
5456 const struct nft_set_ext *ext,
5457 struct nft_expr *expr_array[],
5458 u32 num_exprs)
5459 {
5460 struct nft_set_elem_expr *elem_expr = nft_set_ext_expr(ext);
5461 struct nft_expr *expr;
5462 int i, err;
5463
5464 for (i = 0; i < num_exprs; i++) {
5465 expr = nft_setelem_expr_at(elem_expr, elem_expr->size);
5466 err = nft_expr_clone(expr, expr_array[i]);
5467 if (err < 0)
5468 goto err_elem_expr_setup;
5469
5470 elem_expr->size += expr_array[i]->ops->size;
5471 nft_expr_destroy(ctx, expr_array[i]);
5472 expr_array[i] = NULL;
5473 }
5474
5475 return 0;
5476
5477 err_elem_expr_setup:
5478 for (; i < num_exprs; i++) {
5479 nft_expr_destroy(ctx, expr_array[i]);
5480 expr_array[i] = NULL;
5481 }
5482
5483 return -ENOMEM;
5484 }
5485
5486 struct nft_set_ext *nft_set_catchall_lookup(const struct net *net,
5487 const struct nft_set *set)
5488 {
5489 struct nft_set_elem_catchall *catchall;
5490 u8 genmask = nft_genmask_cur(net);
5491 struct nft_set_ext *ext;
5492
5493 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5494 ext = nft_set_elem_ext(set, catchall->elem);
5495 if (nft_set_elem_active(ext, genmask) &&
5496 !nft_set_elem_expired(ext))
5497 return ext;
5498 }
5499
5500 return NULL;
5501 }
5502 EXPORT_SYMBOL_GPL(nft_set_catchall_lookup);
5503
5504 void *nft_set_catchall_gc(const struct nft_set *set)
5505 {
5506 struct nft_set_elem_catchall *catchall, *next;
5507 struct nft_set_ext *ext;
5508 void *elem = NULL;
5509
5510 list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5511 ext = nft_set_elem_ext(set, catchall->elem);
5512
5513 if (!nft_set_elem_expired(ext) ||
5514 nft_set_elem_mark_busy(ext))
5515 continue;
5516
5517 elem = catchall->elem;
5518 list_del_rcu(&catchall->list);
5519 kfree_rcu(catchall, rcu);
5520 break;
5521 }
5522
5523 return elem;
5524 }
5525 EXPORT_SYMBOL_GPL(nft_set_catchall_gc);
5526
5527 static int nft_setelem_catchall_insert(const struct net *net,
5528 struct nft_set *set,
5529 const struct nft_set_elem *elem,
5530 struct nft_set_ext **pext)
5531 {
5532 struct nft_set_elem_catchall *catchall;
5533 u8 genmask = nft_genmask_next(net);
5534 struct nft_set_ext *ext;
5535
5536 list_for_each_entry(catchall, &set->catchall_list, list) {
5537 ext = nft_set_elem_ext(set, catchall->elem);
5538 if (nft_set_elem_active(ext, genmask)) {
5539 *pext = ext;
5540 return -EEXIST;
5541 }
5542 }
5543
5544 catchall = kmalloc(sizeof(*catchall), GFP_KERNEL);
5545 if (!catchall)
5546 return -ENOMEM;
5547
5548 catchall->elem = elem->priv;
5549 list_add_tail_rcu(&catchall->list, &set->catchall_list);
5550
5551 return 0;
5552 }
5553
5554 static int nft_setelem_insert(const struct net *net,
5555 struct nft_set *set,
5556 const struct nft_set_elem *elem,
5557 struct nft_set_ext **ext, unsigned int flags)
5558 {
5559 int ret;
5560
5561 if (flags & NFT_SET_ELEM_CATCHALL)
5562 ret = nft_setelem_catchall_insert(net, set, elem, ext);
5563 else
5564 ret = set->ops->insert(net, set, elem, ext);
5565
5566 return ret;
5567 }
5568
5569 static bool nft_setelem_is_catchall(const struct nft_set *set,
5570 const struct nft_set_elem *elem)
5571 {
5572 struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5573
5574 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
5575 *nft_set_ext_flags(ext) & NFT_SET_ELEM_CATCHALL)
5576 return true;
5577
5578 return false;
5579 }
5580
5581 static void nft_setelem_activate(struct net *net, struct nft_set *set,
5582 struct nft_set_elem *elem)
5583 {
5584 struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5585
5586 if (nft_setelem_is_catchall(set, elem)) {
5587 nft_set_elem_change_active(net, set, ext);
5588 nft_set_elem_clear_busy(ext);
5589 } else {
5590 set->ops->activate(net, set, elem);
5591 }
5592 }
5593
5594 static int nft_setelem_catchall_deactivate(const struct net *net,
5595 struct nft_set *set,
5596 struct nft_set_elem *elem)
5597 {
5598 struct nft_set_elem_catchall *catchall;
5599 struct nft_set_ext *ext;
5600
5601 list_for_each_entry(catchall, &set->catchall_list, list) {
5602 ext = nft_set_elem_ext(set, catchall->elem);
5603 if (!nft_is_active(net, ext) ||
5604 nft_set_elem_mark_busy(ext))
5605 continue;
5606
5607 kfree(elem->priv);
5608 elem->priv = catchall->elem;
5609 nft_set_elem_change_active(net, set, ext);
5610 return 0;
5611 }
5612
5613 return -ENOENT;
5614 }
5615
5616 static int __nft_setelem_deactivate(const struct net *net,
5617 struct nft_set *set,
5618 struct nft_set_elem *elem)
5619 {
5620 void *priv;
5621
5622 priv = set->ops->deactivate(net, set, elem);
5623 if (!priv)
5624 return -ENOENT;
5625
5626 kfree(elem->priv);
5627 elem->priv = priv;
5628 set->ndeact++;
5629
5630 return 0;
5631 }
5632
5633 static int nft_setelem_deactivate(const struct net *net,
5634 struct nft_set *set,
5635 struct nft_set_elem *elem, u32 flags)
5636 {
5637 int ret;
5638
5639 if (flags & NFT_SET_ELEM_CATCHALL)
5640 ret = nft_setelem_catchall_deactivate(net, set, elem);
5641 else
5642 ret = __nft_setelem_deactivate(net, set, elem);
5643
5644 return ret;
5645 }
5646
5647 static void nft_setelem_catchall_remove(const struct net *net,
5648 const struct nft_set *set,
5649 const struct nft_set_elem *elem)
5650 {
5651 struct nft_set_elem_catchall *catchall, *next;
5652
5653 list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5654 if (catchall->elem == elem->priv) {
5655 list_del_rcu(&catchall->list);
5656 kfree_rcu(catchall);
5657 break;
5658 }
5659 }
5660 }
5661
5662 static void nft_setelem_remove(const struct net *net,
5663 const struct nft_set *set,
5664 const struct nft_set_elem *elem)
5665 {
5666 if (nft_setelem_is_catchall(set, elem))
5667 nft_setelem_catchall_remove(net, set, elem);
5668 else
5669 set->ops->remove(net, set, elem);
5670 }
5671
5672 static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5673 const struct nlattr *attr, u32 nlmsg_flags)
5674 {
5675 struct nft_expr *expr_array[NFT_SET_EXPR_MAX] = {};
5676 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5677 u8 genmask = nft_genmask_next(ctx->net);
5678 u32 flags = 0, size = 0, num_exprs = 0;
5679 struct nft_set_ext_tmpl tmpl;
5680 struct nft_set_ext *ext, *ext2;
5681 struct nft_set_elem elem;
5682 struct nft_set_binding *binding;
5683 struct nft_object *obj = NULL;
5684 struct nft_userdata *udata;
5685 struct nft_data_desc desc;
5686 enum nft_registers dreg;
5687 struct nft_trans *trans;
5688 u64 timeout;
5689 u64 expiration;
5690 int err, i;
5691 u8 ulen;
5692
5693 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5694 nft_set_elem_policy, NULL);
5695 if (err < 0)
5696 return err;
5697
5698 nft_set_ext_prepare(&tmpl);
5699
5700 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5701 if (err < 0)
5702 return err;
5703
5704 if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5705 return -EINVAL;
5706
5707 if (flags != 0)
5708 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
5709
5710 if (set->flags & NFT_SET_MAP) {
5711 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
5712 !(flags & NFT_SET_ELEM_INTERVAL_END))
5713 return -EINVAL;
5714 } else {
5715 if (nla[NFTA_SET_ELEM_DATA] != NULL)
5716 return -EINVAL;
5717 }
5718
5719 if ((flags & NFT_SET_ELEM_INTERVAL_END) &&
5720 (nla[NFTA_SET_ELEM_DATA] ||
5721 nla[NFTA_SET_ELEM_OBJREF] ||
5722 nla[NFTA_SET_ELEM_TIMEOUT] ||
5723 nla[NFTA_SET_ELEM_EXPIRATION] ||
5724 nla[NFTA_SET_ELEM_USERDATA] ||
5725 nla[NFTA_SET_ELEM_EXPR] ||
5726 nla[NFTA_SET_ELEM_EXPRESSIONS]))
5727 return -EINVAL;
5728
5729 timeout = 0;
5730 if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
5731 if (!(set->flags & NFT_SET_TIMEOUT))
5732 return -EINVAL;
5733 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
5734 &timeout);
5735 if (err)
5736 return err;
5737 } else if (set->flags & NFT_SET_TIMEOUT) {
5738 timeout = set->timeout;
5739 }
5740
5741 expiration = 0;
5742 if (nla[NFTA_SET_ELEM_EXPIRATION] != NULL) {
5743 if (!(set->flags & NFT_SET_TIMEOUT))
5744 return -EINVAL;
5745 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_EXPIRATION],
5746 &expiration);
5747 if (err)
5748 return err;
5749 }
5750
5751 if (nla[NFTA_SET_ELEM_EXPR]) {
5752 struct nft_expr *expr;
5753
5754 if (set->num_exprs && set->num_exprs != 1)
5755 return -EOPNOTSUPP;
5756
5757 expr = nft_set_elem_expr_alloc(ctx, set,
5758 nla[NFTA_SET_ELEM_EXPR]);
5759 if (IS_ERR(expr))
5760 return PTR_ERR(expr);
5761
5762 expr_array[0] = expr;
5763 num_exprs = 1;
5764
5765 if (set->num_exprs && set->exprs[0]->ops != expr->ops) {
5766 err = -EOPNOTSUPP;
5767 goto err_set_elem_expr;
5768 }
5769 } else if (nla[NFTA_SET_ELEM_EXPRESSIONS]) {
5770 struct nft_expr *expr;
5771 struct nlattr *tmp;
5772 int left;
5773
5774 i = 0;
5775 nla_for_each_nested(tmp, nla[NFTA_SET_ELEM_EXPRESSIONS], left) {
5776 if (i == NFT_SET_EXPR_MAX ||
5777 (set->num_exprs && set->num_exprs == i)) {
5778 err = -E2BIG;
5779 goto err_set_elem_expr;
5780 }
5781 if (nla_type(tmp) != NFTA_LIST_ELEM) {
5782 err = -EINVAL;
5783 goto err_set_elem_expr;
5784 }
5785 expr = nft_set_elem_expr_alloc(ctx, set, tmp);
5786 if (IS_ERR(expr)) {
5787 err = PTR_ERR(expr);
5788 goto err_set_elem_expr;
5789 }
5790 expr_array[i] = expr;
5791 num_exprs++;
5792
5793 if (set->num_exprs && expr->ops != set->exprs[i]->ops) {
5794 err = -EOPNOTSUPP;
5795 goto err_set_elem_expr;
5796 }
5797 i++;
5798 }
5799 if (set->num_exprs && set->num_exprs != i) {
5800 err = -EOPNOTSUPP;
5801 goto err_set_elem_expr;
5802 }
5803 } else if (set->num_exprs > 0) {
5804 err = nft_set_elem_expr_clone(ctx, set, expr_array);
5805 if (err < 0)
5806 goto err_set_elem_expr_clone;
5807
5808 num_exprs = set->num_exprs;
5809 }
5810
5811 if (nla[NFTA_SET_ELEM_KEY]) {
5812 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5813 nla[NFTA_SET_ELEM_KEY]);
5814 if (err < 0)
5815 goto err_set_elem_expr;
5816
5817 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
5818 }
5819
5820 if (nla[NFTA_SET_ELEM_KEY_END]) {
5821 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5822 nla[NFTA_SET_ELEM_KEY_END]);
5823 if (err < 0)
5824 goto err_parse_key;
5825
5826 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
5827 }
5828
5829 if (timeout > 0) {
5830 nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
5831 if (timeout != set->timeout)
5832 nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
5833 }
5834
5835 if (num_exprs) {
5836 for (i = 0; i < num_exprs; i++)
5837 size += expr_array[i]->ops->size;
5838
5839 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_EXPRESSIONS,
5840 sizeof(struct nft_set_elem_expr) +
5841 size);
5842 }
5843
5844 if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
5845 if (!(set->flags & NFT_SET_OBJECT)) {
5846 err = -EINVAL;
5847 goto err_parse_key_end;
5848 }
5849 obj = nft_obj_lookup(ctx->net, ctx->table,
5850 nla[NFTA_SET_ELEM_OBJREF],
5851 set->objtype, genmask);
5852 if (IS_ERR(obj)) {
5853 err = PTR_ERR(obj);
5854 goto err_parse_key_end;
5855 }
5856 nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
5857 }
5858
5859 if (nla[NFTA_SET_ELEM_DATA] != NULL) {
5860 err = nft_setelem_parse_data(ctx, set, &desc, &elem.data.val,
5861 nla[NFTA_SET_ELEM_DATA]);
5862 if (err < 0)
5863 goto err_parse_key_end;
5864
5865 dreg = nft_type_to_reg(set->dtype);
5866 list_for_each_entry(binding, &set->bindings, list) {
5867 struct nft_ctx bind_ctx = {
5868 .net = ctx->net,
5869 .family = ctx->family,
5870 .table = ctx->table,
5871 .chain = (struct nft_chain *)binding->chain,
5872 };
5873
5874 if (!(binding->flags & NFT_SET_MAP))
5875 continue;
5876
5877 err = nft_validate_register_store(&bind_ctx, dreg,
5878 &elem.data.val,
5879 desc.type, desc.len);
5880 if (err < 0)
5881 goto err_parse_data;
5882
5883 if (desc.type == NFT_DATA_VERDICT &&
5884 (elem.data.val.verdict.code == NFT_GOTO ||
5885 elem.data.val.verdict.code == NFT_JUMP))
5886 nft_validate_state_update(ctx->net,
5887 NFT_VALIDATE_NEED);
5888 }
5889
5890 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, desc.len);
5891 }
5892
5893 /* The full maximum length of userdata can exceed the maximum
5894 * offset value (U8_MAX) for following extensions, therefor it
5895 * must be the last extension added.
5896 */
5897 ulen = 0;
5898 if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
5899 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
5900 if (ulen > 0)
5901 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
5902 ulen);
5903 }
5904
5905 err = -ENOMEM;
5906 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
5907 elem.key_end.val.data, elem.data.val.data,
5908 timeout, expiration, GFP_KERNEL);
5909 if (elem.priv == NULL)
5910 goto err_parse_data;
5911
5912 ext = nft_set_elem_ext(set, elem.priv);
5913 if (flags)
5914 *nft_set_ext_flags(ext) = flags;
5915 if (ulen > 0) {
5916 udata = nft_set_ext_userdata(ext);
5917 udata->len = ulen - 1;
5918 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
5919 }
5920 if (obj) {
5921 *nft_set_ext_obj(ext) = obj;
5922 obj->use++;
5923 }
5924 err = nft_set_elem_expr_setup(ctx, ext, expr_array, num_exprs);
5925 if (err < 0)
5926 goto err_elem_expr;
5927
5928 trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
5929 if (trans == NULL) {
5930 err = -ENOMEM;
5931 goto err_elem_expr;
5932 }
5933
5934 ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
5935
5936 err = nft_setelem_insert(ctx->net, set, &elem, &ext2, flags);
5937 if (err) {
5938 if (err == -EEXIST) {
5939 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
5940 nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
5941 nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
5942 nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF))
5943 goto err_element_clash;
5944 if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5945 nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
5946 memcmp(nft_set_ext_data(ext),
5947 nft_set_ext_data(ext2), set->dlen) != 0) ||
5948 (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
5949 nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
5950 *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
5951 goto err_element_clash;
5952 else if (!(nlmsg_flags & NLM_F_EXCL))
5953 err = 0;
5954 } else if (err == -ENOTEMPTY) {
5955 /* ENOTEMPTY reports overlapping between this element
5956 * and an existing one.
5957 */
5958 err = -EEXIST;
5959 }
5960 goto err_element_clash;
5961 }
5962
5963 if (!(flags & NFT_SET_ELEM_CATCHALL) && set->size &&
5964 !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact)) {
5965 err = -ENFILE;
5966 goto err_set_full;
5967 }
5968
5969 nft_trans_elem(trans) = elem;
5970 nft_trans_commit_list_add_tail(ctx->net, trans);
5971 return 0;
5972
5973 err_set_full:
5974 nft_setelem_remove(ctx->net, set, &elem);
5975 err_element_clash:
5976 kfree(trans);
5977 err_elem_expr:
5978 if (obj)
5979 obj->use--;
5980
5981 nf_tables_set_elem_destroy(ctx, set, elem.priv);
5982 err_parse_data:
5983 if (nla[NFTA_SET_ELEM_DATA] != NULL)
5984 nft_data_release(&elem.data.val, desc.type);
5985 err_parse_key_end:
5986 nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
5987 err_parse_key:
5988 nft_data_release(&elem.key.val, NFT_DATA_VALUE);
5989 err_set_elem_expr:
5990 for (i = 0; i < num_exprs && expr_array[i]; i++)
5991 nft_expr_destroy(ctx, expr_array[i]);
5992 err_set_elem_expr_clone:
5993 return err;
5994 }
5995
5996 static int nf_tables_newsetelem(struct sk_buff *skb,
5997 const struct nfnl_info *info,
5998 const struct nlattr * const nla[])
5999 {
6000 struct nftables_pernet *nft_net = nft_pernet(info->net);
6001 struct netlink_ext_ack *extack = info->extack;
6002 u8 genmask = nft_genmask_next(info->net);
6003 u8 family = info->nfmsg->nfgen_family;
6004 struct net *net = info->net;
6005 const struct nlattr *attr;
6006 struct nft_table *table;
6007 struct nft_set *set;
6008 struct nft_ctx ctx;
6009 int rem, err;
6010
6011 if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
6012 return -EINVAL;
6013
6014 table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6015 genmask, NETLINK_CB(skb).portid);
6016 if (IS_ERR(table)) {
6017 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6018 return PTR_ERR(table);
6019 }
6020
6021 set = nft_set_lookup_global(net, table, nla[NFTA_SET_ELEM_LIST_SET],
6022 nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
6023 if (IS_ERR(set))
6024 return PTR_ERR(set);
6025
6026 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6027 return -EBUSY;
6028
6029 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6030
6031 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6032 err = nft_add_set_elem(&ctx, set, attr, info->nlh->nlmsg_flags);
6033 if (err < 0)
6034 return err;
6035 }
6036
6037 if (nft_net->validate_state == NFT_VALIDATE_DO)
6038 return nft_table_validate(net, table);
6039
6040 return 0;
6041 }
6042
6043 /**
6044 * nft_data_hold - hold a nft_data item
6045 *
6046 * @data: struct nft_data to release
6047 * @type: type of data
6048 *
6049 * Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
6050 * NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
6051 * NFT_GOTO verdicts. This function must be called on active data objects
6052 * from the second phase of the commit protocol.
6053 */
6054 void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
6055 {
6056 struct nft_chain *chain;
6057 struct nft_rule *rule;
6058
6059 if (type == NFT_DATA_VERDICT) {
6060 switch (data->verdict.code) {
6061 case NFT_JUMP:
6062 case NFT_GOTO:
6063 chain = data->verdict.chain;
6064 chain->use++;
6065
6066 if (!nft_chain_is_bound(chain))
6067 break;
6068
6069 chain->table->use++;
6070 list_for_each_entry(rule, &chain->rules, list)
6071 chain->use++;
6072
6073 nft_chain_add(chain->table, chain);
6074 break;
6075 }
6076 }
6077 }
6078
6079 static void nft_setelem_data_activate(const struct net *net,
6080 const struct nft_set *set,
6081 struct nft_set_elem *elem)
6082 {
6083 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6084
6085 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6086 nft_data_hold(nft_set_ext_data(ext), set->dtype);
6087 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6088 (*nft_set_ext_obj(ext))->use++;
6089 }
6090
6091 static void nft_setelem_data_deactivate(const struct net *net,
6092 const struct nft_set *set,
6093 struct nft_set_elem *elem)
6094 {
6095 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6096
6097 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6098 nft_data_release(nft_set_ext_data(ext), set->dtype);
6099 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6100 (*nft_set_ext_obj(ext))->use--;
6101 }
6102
6103 static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
6104 const struct nlattr *attr)
6105 {
6106 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6107 struct nft_set_ext_tmpl tmpl;
6108 struct nft_set_elem elem;
6109 struct nft_set_ext *ext;
6110 struct nft_trans *trans;
6111 u32 flags = 0;
6112 int err;
6113
6114 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6115 nft_set_elem_policy, NULL);
6116 if (err < 0)
6117 return err;
6118
6119 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6120 if (err < 0)
6121 return err;
6122
6123 if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
6124 return -EINVAL;
6125
6126 nft_set_ext_prepare(&tmpl);
6127
6128 if (flags != 0)
6129 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
6130
6131 if (nla[NFTA_SET_ELEM_KEY]) {
6132 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6133 nla[NFTA_SET_ELEM_KEY]);
6134 if (err < 0)
6135 return err;
6136
6137 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6138 }
6139
6140 if (nla[NFTA_SET_ELEM_KEY_END]) {
6141 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6142 nla[NFTA_SET_ELEM_KEY_END]);
6143 if (err < 0)
6144 return err;
6145
6146 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6147 }
6148
6149 err = -ENOMEM;
6150 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6151 elem.key_end.val.data, NULL, 0, 0,
6152 GFP_KERNEL);
6153 if (elem.priv == NULL)
6154 goto fail_elem;
6155
6156 ext = nft_set_elem_ext(set, elem.priv);
6157 if (flags)
6158 *nft_set_ext_flags(ext) = flags;
6159
6160 trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
6161 if (trans == NULL)
6162 goto fail_trans;
6163
6164 err = nft_setelem_deactivate(ctx->net, set, &elem, flags);
6165 if (err < 0)
6166 goto fail_ops;
6167
6168 nft_setelem_data_deactivate(ctx->net, set, &elem);
6169
6170 nft_trans_elem(trans) = elem;
6171 nft_trans_commit_list_add_tail(ctx->net, trans);
6172 return 0;
6173
6174 fail_ops:
6175 kfree(trans);
6176 fail_trans:
6177 kfree(elem.priv);
6178 fail_elem:
6179 nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6180 return err;
6181 }
6182
6183 static int nft_setelem_flush(const struct nft_ctx *ctx,
6184 struct nft_set *set,
6185 const struct nft_set_iter *iter,
6186 struct nft_set_elem *elem)
6187 {
6188 struct nft_trans *trans;
6189 int err;
6190
6191 trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6192 sizeof(struct nft_trans_elem), GFP_ATOMIC);
6193 if (!trans)
6194 return -ENOMEM;
6195
6196 if (!set->ops->flush(ctx->net, set, elem->priv)) {
6197 err = -ENOENT;
6198 goto err1;
6199 }
6200 set->ndeact++;
6201
6202 nft_setelem_data_deactivate(ctx->net, set, elem);
6203 nft_trans_elem_set(trans) = set;
6204 nft_trans_elem(trans) = *elem;
6205 nft_trans_commit_list_add_tail(ctx->net, trans);
6206
6207 return 0;
6208 err1:
6209 kfree(trans);
6210 return err;
6211 }
6212
6213 static int __nft_set_catchall_flush(const struct nft_ctx *ctx,
6214 struct nft_set *set,
6215 struct nft_set_elem *elem)
6216 {
6217 struct nft_trans *trans;
6218
6219 trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6220 sizeof(struct nft_trans_elem), GFP_KERNEL);
6221 if (!trans)
6222 return -ENOMEM;
6223
6224 nft_setelem_data_deactivate(ctx->net, set, elem);
6225 nft_trans_elem_set(trans) = set;
6226 nft_trans_elem(trans) = *elem;
6227 nft_trans_commit_list_add_tail(ctx->net, trans);
6228
6229 return 0;
6230 }
6231
6232 static int nft_set_catchall_flush(const struct nft_ctx *ctx,
6233 struct nft_set *set)
6234 {
6235 u8 genmask = nft_genmask_next(ctx->net);
6236 struct nft_set_elem_catchall *catchall;
6237 struct nft_set_elem elem;
6238 struct nft_set_ext *ext;
6239 int ret = 0;
6240
6241 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6242 ext = nft_set_elem_ext(set, catchall->elem);
6243 if (!nft_set_elem_active(ext, genmask) ||
6244 nft_set_elem_mark_busy(ext))
6245 continue;
6246
6247 elem.priv = catchall->elem;
6248 ret = __nft_set_catchall_flush(ctx, set, &elem);
6249 if (ret < 0)
6250 break;
6251 }
6252
6253 return ret;
6254 }
6255
6256 static int nft_set_flush(struct nft_ctx *ctx, struct nft_set *set, u8 genmask)
6257 {
6258 struct nft_set_iter iter = {
6259 .genmask = genmask,
6260 .fn = nft_setelem_flush,
6261 };
6262
6263 set->ops->walk(ctx, set, &iter);
6264 if (!iter.err)
6265 iter.err = nft_set_catchall_flush(ctx, set);
6266
6267 return iter.err;
6268 }
6269
6270 static int nf_tables_delsetelem(struct sk_buff *skb,
6271 const struct nfnl_info *info,
6272 const struct nlattr * const nla[])
6273 {
6274 struct netlink_ext_ack *extack = info->extack;
6275 u8 genmask = nft_genmask_next(info->net);
6276 u8 family = info->nfmsg->nfgen_family;
6277 struct net *net = info->net;
6278 const struct nlattr *attr;
6279 struct nft_table *table;
6280 struct nft_set *set;
6281 struct nft_ctx ctx;
6282 int rem, err = 0;
6283
6284 table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6285 genmask, NETLINK_CB(skb).portid);
6286 if (IS_ERR(table)) {
6287 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6288 return PTR_ERR(table);
6289 }
6290
6291 set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
6292 if (IS_ERR(set))
6293 return PTR_ERR(set);
6294 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6295 return -EBUSY;
6296
6297 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6298
6299 if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
6300 return nft_set_flush(&ctx, set, genmask);
6301
6302 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6303 err = nft_del_setelem(&ctx, set, attr);
6304 if (err < 0)
6305 break;
6306 }
6307 return err;
6308 }
6309
6310 void nft_set_gc_batch_release(struct rcu_head *rcu)
6311 {
6312 struct nft_set_gc_batch *gcb;
6313 unsigned int i;
6314
6315 gcb = container_of(rcu, struct nft_set_gc_batch, head.rcu);
6316 for (i = 0; i < gcb->head.cnt; i++)
6317 nft_set_elem_destroy(gcb->head.set, gcb->elems[i], true);
6318 kfree(gcb);
6319 }
6320
6321 struct nft_set_gc_batch *nft_set_gc_batch_alloc(const struct nft_set *set,
6322 gfp_t gfp)
6323 {
6324 struct nft_set_gc_batch *gcb;
6325
6326 gcb = kzalloc(sizeof(*gcb), gfp);
6327 if (gcb == NULL)
6328 return gcb;
6329 gcb->head.set = set;
6330 return gcb;
6331 }
6332
6333 /*
6334 * Stateful objects
6335 */
6336
6337 /**
6338 * nft_register_obj- register nf_tables stateful object type
6339 * @obj_type: object type
6340 *
6341 * Registers the object type for use with nf_tables. Returns zero on
6342 * success or a negative errno code otherwise.
6343 */
6344 int nft_register_obj(struct nft_object_type *obj_type)
6345 {
6346 if (obj_type->type == NFT_OBJECT_UNSPEC)
6347 return -EINVAL;
6348
6349 nfnl_lock(NFNL_SUBSYS_NFTABLES);
6350 list_add_rcu(&obj_type->list, &nf_tables_objects);
6351 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6352 return 0;
6353 }
6354 EXPORT_SYMBOL_GPL(nft_register_obj);
6355
6356 /**
6357 * nft_unregister_obj - unregister nf_tables object type
6358 * @obj_type: object type
6359 *
6360 * Unregisters the object type for use with nf_tables.
6361 */
6362 void nft_unregister_obj(struct nft_object_type *obj_type)
6363 {
6364 nfnl_lock(NFNL_SUBSYS_NFTABLES);
6365 list_del_rcu(&obj_type->list);
6366 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6367 }
6368 EXPORT_SYMBOL_GPL(nft_unregister_obj);
6369
6370 struct nft_object *nft_obj_lookup(const struct net *net,
6371 const struct nft_table *table,
6372 const struct nlattr *nla, u32 objtype,
6373 u8 genmask)
6374 {
6375 struct nft_object_hash_key k = { .table = table };
6376 char search[NFT_OBJ_MAXNAMELEN];
6377 struct rhlist_head *tmp, *list;
6378 struct nft_object *obj;
6379
6380 nla_strscpy(search, nla, sizeof(search));
6381 k.name = search;
6382
6383 WARN_ON_ONCE(!rcu_read_lock_held() &&
6384 !lockdep_commit_lock_is_held(net));
6385
6386 rcu_read_lock();
6387 list = rhltable_lookup(&nft_objname_ht, &k, nft_objname_ht_params);
6388 if (!list)
6389 goto out;
6390
6391 rhl_for_each_entry_rcu(obj, tmp, list, rhlhead) {
6392 if (objtype == obj->ops->type->type &&
6393 nft_active_genmask(obj, genmask)) {
6394 rcu_read_unlock();
6395 return obj;
6396 }
6397 }
6398 out:
6399 rcu_read_unlock();
6400 return ERR_PTR(-ENOENT);
6401 }
6402 EXPORT_SYMBOL_GPL(nft_obj_lookup);
6403
6404 static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
6405 const struct nlattr *nla,
6406 u32 objtype, u8 genmask)
6407 {
6408 struct nft_object *obj;
6409
6410 list_for_each_entry(obj, &table->objects, list) {
6411 if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
6412 objtype == obj->ops->type->type &&
6413 nft_active_genmask(obj, genmask))
6414 return obj;
6415 }
6416 return ERR_PTR(-ENOENT);
6417 }
6418
6419 static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
6420 [NFTA_OBJ_TABLE] = { .type = NLA_STRING,
6421 .len = NFT_TABLE_MAXNAMELEN - 1 },
6422 [NFTA_OBJ_NAME] = { .type = NLA_STRING,
6423 .len = NFT_OBJ_MAXNAMELEN - 1 },
6424 [NFTA_OBJ_TYPE] = { .type = NLA_U32 },
6425 [NFTA_OBJ_DATA] = { .type = NLA_NESTED },
6426 [NFTA_OBJ_HANDLE] = { .type = NLA_U64},
6427 [NFTA_OBJ_USERDATA] = { .type = NLA_BINARY,
6428 .len = NFT_USERDATA_MAXLEN },
6429 };
6430
6431 static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
6432 const struct nft_object_type *type,
6433 const struct nlattr *attr)
6434 {
6435 struct nlattr **tb;
6436 const struct nft_object_ops *ops;
6437 struct nft_object *obj;
6438 int err = -ENOMEM;
6439
6440 tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
6441 if (!tb)
6442 goto err1;
6443
6444 if (attr) {
6445 err = nla_parse_nested_deprecated(tb, type->maxattr, attr,
6446 type->policy, NULL);
6447 if (err < 0)
6448 goto err2;
6449 } else {
6450 memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
6451 }
6452
6453 if (type->select_ops) {
6454 ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
6455 if (IS_ERR(ops)) {
6456 err = PTR_ERR(ops);
6457 goto err2;
6458 }
6459 } else {
6460 ops = type->ops;
6461 }
6462
6463 err = -ENOMEM;
6464 obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL);
6465 if (!obj)
6466 goto err2;
6467
6468 err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
6469 if (err < 0)
6470 goto err3;
6471
6472 obj->ops = ops;
6473
6474 kfree(tb);
6475 return obj;
6476 err3:
6477 kfree(obj);
6478 err2:
6479 kfree(tb);
6480 err1:
6481 return ERR_PTR(err);
6482 }
6483
6484 static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
6485 struct nft_object *obj, bool reset)
6486 {
6487 struct nlattr *nest;
6488
6489 nest = nla_nest_start_noflag(skb, attr);
6490 if (!nest)
6491 goto nla_put_failure;
6492 if (obj->ops->dump(skb, obj, reset) < 0)
6493 goto nla_put_failure;
6494 nla_nest_end(skb, nest);
6495 return 0;
6496
6497 nla_put_failure:
6498 return -1;
6499 }
6500
6501 static const struct nft_object_type *__nft_obj_type_get(u32 objtype)
6502 {
6503 const struct nft_object_type *type;
6504
6505 list_for_each_entry(type, &nf_tables_objects, list) {
6506 if (objtype == type->type)
6507 return type;
6508 }
6509 return NULL;
6510 }
6511
6512 static const struct nft_object_type *
6513 nft_obj_type_get(struct net *net, u32 objtype)
6514 {
6515 const struct nft_object_type *type;
6516
6517 type = __nft_obj_type_get(objtype);
6518 if (type != NULL && try_module_get(type->owner))
6519 return type;
6520
6521 lockdep_nfnl_nft_mutex_not_held();
6522 #ifdef CONFIG_MODULES
6523 if (type == NULL) {
6524 if (nft_request_module(net, "nft-obj-%u", objtype) == -EAGAIN)
6525 return ERR_PTR(-EAGAIN);
6526 }
6527 #endif
6528 return ERR_PTR(-ENOENT);
6529 }
6530
6531 static int nf_tables_updobj(const struct nft_ctx *ctx,
6532 const struct nft_object_type *type,
6533 const struct nlattr *attr,
6534 struct nft_object *obj)
6535 {
6536 struct nft_object *newobj;
6537 struct nft_trans *trans;
6538 int err;
6539
6540 trans = nft_trans_alloc(ctx, NFT_MSG_NEWOBJ,
6541 sizeof(struct nft_trans_obj));
6542 if (!trans)
6543 return -ENOMEM;
6544
6545 newobj = nft_obj_init(ctx, type, attr);
6546 if (IS_ERR(newobj)) {
6547 err = PTR_ERR(newobj);
6548 goto err_free_trans;
6549 }
6550
6551 nft_trans_obj(trans) = obj;
6552 nft_trans_obj_update(trans) = true;
6553 nft_trans_obj_newobj(trans) = newobj;
6554 nft_trans_commit_list_add_tail(ctx->net, trans);
6555
6556 return 0;
6557
6558 err_free_trans:
6559 kfree(trans);
6560 return err;
6561 }
6562
6563 static int nf_tables_newobj(struct sk_buff *skb, const struct nfnl_info *info,
6564 const struct nlattr * const nla[])
6565 {
6566 struct netlink_ext_ack *extack = info->extack;
6567 u8 genmask = nft_genmask_next(info->net);
6568 u8 family = info->nfmsg->nfgen_family;
6569 const struct nft_object_type *type;
6570 struct net *net = info->net;
6571 struct nft_table *table;
6572 struct nft_object *obj;
6573 struct nft_ctx ctx;
6574 u32 objtype;
6575 int err;
6576
6577 if (!nla[NFTA_OBJ_TYPE] ||
6578 !nla[NFTA_OBJ_NAME] ||
6579 !nla[NFTA_OBJ_DATA])
6580 return -EINVAL;
6581
6582 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
6583 NETLINK_CB(skb).portid);
6584 if (IS_ERR(table)) {
6585 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6586 return PTR_ERR(table);
6587 }
6588
6589 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6590 obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6591 if (IS_ERR(obj)) {
6592 err = PTR_ERR(obj);
6593 if (err != -ENOENT) {
6594 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6595 return err;
6596 }
6597 } else {
6598 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
6599 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6600 return -EEXIST;
6601 }
6602 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
6603 return -EOPNOTSUPP;
6604
6605 type = __nft_obj_type_get(objtype);
6606 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6607
6608 return nf_tables_updobj(&ctx, type, nla[NFTA_OBJ_DATA], obj);
6609 }
6610
6611 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6612
6613 type = nft_obj_type_get(net, objtype);
6614 if (IS_ERR(type))
6615 return PTR_ERR(type);
6616
6617 obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
6618 if (IS_ERR(obj)) {
6619 err = PTR_ERR(obj);
6620 goto err_init;
6621 }
6622 obj->key.table = table;
6623 obj->handle = nf_tables_alloc_handle(table);
6624
6625 obj->key.name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL);
6626 if (!obj->key.name) {
6627 err = -ENOMEM;
6628 goto err_strdup;
6629 }
6630
6631 if (nla[NFTA_OBJ_USERDATA]) {
6632 obj->udata = nla_memdup(nla[NFTA_OBJ_USERDATA], GFP_KERNEL);
6633 if (obj->udata == NULL)
6634 goto err_userdata;
6635
6636 obj->udlen = nla_len(nla[NFTA_OBJ_USERDATA]);
6637 }
6638
6639 err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
6640 if (err < 0)
6641 goto err_trans;
6642
6643 err = rhltable_insert(&nft_objname_ht, &obj->rhlhead,
6644 nft_objname_ht_params);
6645 if (err < 0)
6646 goto err_obj_ht;
6647
6648 list_add_tail_rcu(&obj->list, &table->objects);
6649 table->use++;
6650 return 0;
6651 err_obj_ht:
6652 /* queued in transaction log */
6653 INIT_LIST_HEAD(&obj->list);
6654 return err;
6655 err_trans:
6656 kfree(obj->udata);
6657 err_userdata:
6658 kfree(obj->key.name);
6659 err_strdup:
6660 if (obj->ops->destroy)
6661 obj->ops->destroy(&ctx, obj);
6662 kfree(obj);
6663 err_init:
6664 module_put(type->owner);
6665 return err;
6666 }
6667
6668 static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
6669 u32 portid, u32 seq, int event, u32 flags,
6670 int family, const struct nft_table *table,
6671 struct nft_object *obj, bool reset)
6672 {
6673 struct nlmsghdr *nlh;
6674
6675 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
6676 nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
6677 NFNETLINK_V0, nft_base_seq(net));
6678 if (!nlh)
6679 goto nla_put_failure;
6680
6681 if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
6682 nla_put_string(skb, NFTA_OBJ_NAME, obj->key.name) ||
6683 nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
6684 nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
6685 nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset) ||
6686 nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
6687 NFTA_OBJ_PAD))
6688 goto nla_put_failure;
6689
6690 if (obj->udata &&
6691 nla_put(skb, NFTA_OBJ_USERDATA, obj->udlen, obj->udata))
6692 goto nla_put_failure;
6693
6694 nlmsg_end(skb, nlh);
6695 return 0;
6696
6697 nla_put_failure:
6698 nlmsg_trim(skb, nlh);
6699 return -1;
6700 }
6701
6702 struct nft_obj_filter {
6703 char *table;
6704 u32 type;
6705 };
6706
6707 static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
6708 {
6709 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
6710 const struct nft_table *table;
6711 unsigned int idx = 0, s_idx = cb->args[0];
6712 struct nft_obj_filter *filter = cb->data;
6713 struct net *net = sock_net(skb->sk);
6714 int family = nfmsg->nfgen_family;
6715 struct nftables_pernet *nft_net;
6716 struct nft_object *obj;
6717 bool reset = false;
6718
6719 if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6720 reset = true;
6721
6722 rcu_read_lock();
6723 nft_net = nft_pernet(net);
6724 cb->seq = nft_net->base_seq;
6725
6726 list_for_each_entry_rcu(table, &nft_net->tables, list) {
6727 if (family != NFPROTO_UNSPEC && family != table->family)
6728 continue;
6729
6730 list_for_each_entry_rcu(obj, &table->objects, list) {
6731 if (!nft_is_active(net, obj))
6732 goto cont;
6733 if (idx < s_idx)
6734 goto cont;
6735 if (idx > s_idx)
6736 memset(&cb->args[1], 0,
6737 sizeof(cb->args) - sizeof(cb->args[0]));
6738 if (filter && filter->table &&
6739 strcmp(filter->table, table->name))
6740 goto cont;
6741 if (filter &&
6742 filter->type != NFT_OBJECT_UNSPEC &&
6743 obj->ops->type->type != filter->type)
6744 goto cont;
6745 if (reset) {
6746 char *buf = kasprintf(GFP_ATOMIC,
6747 "%s:%u",
6748 table->name,
6749 nft_net->base_seq);
6750
6751 audit_log_nfcfg(buf,
6752 family,
6753 obj->handle,
6754 AUDIT_NFT_OP_OBJ_RESET,
6755 GFP_ATOMIC);
6756 kfree(buf);
6757 }
6758
6759 if (nf_tables_fill_obj_info(skb, net, NETLINK_CB(cb->skb).portid,
6760 cb->nlh->nlmsg_seq,
6761 NFT_MSG_NEWOBJ,
6762 NLM_F_MULTI | NLM_F_APPEND,
6763 table->family, table,
6764 obj, reset) < 0)
6765 goto done;
6766
6767 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
6768 cont:
6769 idx++;
6770 }
6771 }
6772 done:
6773 rcu_read_unlock();
6774
6775 cb->args[0] = idx;
6776 return skb->len;
6777 }
6778
6779 static int nf_tables_dump_obj_start(struct netlink_callback *cb)
6780 {
6781 const struct nlattr * const *nla = cb->data;
6782 struct nft_obj_filter *filter = NULL;
6783
6784 if (nla[NFTA_OBJ_TABLE] || nla[NFTA_OBJ_TYPE]) {
6785 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
6786 if (!filter)
6787 return -ENOMEM;
6788
6789 if (nla[NFTA_OBJ_TABLE]) {
6790 filter->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
6791 if (!filter->table) {
6792 kfree(filter);
6793 return -ENOMEM;
6794 }
6795 }
6796
6797 if (nla[NFTA_OBJ_TYPE])
6798 filter->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6799 }
6800
6801 cb->data = filter;
6802 return 0;
6803 }
6804
6805 static int nf_tables_dump_obj_done(struct netlink_callback *cb)
6806 {
6807 struct nft_obj_filter *filter = cb->data;
6808
6809 if (filter) {
6810 kfree(filter->table);
6811 kfree(filter);
6812 }
6813
6814 return 0;
6815 }
6816
6817 /* called with rcu_read_lock held */
6818 static int nf_tables_getobj(struct sk_buff *skb, const struct nfnl_info *info,
6819 const struct nlattr * const nla[])
6820 {
6821 struct netlink_ext_ack *extack = info->extack;
6822 u8 genmask = nft_genmask_cur(info->net);
6823 u8 family = info->nfmsg->nfgen_family;
6824 const struct nft_table *table;
6825 struct net *net = info->net;
6826 struct nft_object *obj;
6827 struct sk_buff *skb2;
6828 bool reset = false;
6829 u32 objtype;
6830 int err;
6831
6832 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
6833 struct netlink_dump_control c = {
6834 .start = nf_tables_dump_obj_start,
6835 .dump = nf_tables_dump_obj,
6836 .done = nf_tables_dump_obj_done,
6837 .module = THIS_MODULE,
6838 .data = (void *)nla,
6839 };
6840
6841 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
6842 }
6843
6844 if (!nla[NFTA_OBJ_NAME] ||
6845 !nla[NFTA_OBJ_TYPE])
6846 return -EINVAL;
6847
6848 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask, 0);
6849 if (IS_ERR(table)) {
6850 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6851 return PTR_ERR(table);
6852 }
6853
6854 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6855 obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6856 if (IS_ERR(obj)) {
6857 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6858 return PTR_ERR(obj);
6859 }
6860
6861 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
6862 if (!skb2)
6863 return -ENOMEM;
6864
6865 if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6866 reset = true;
6867
6868 if (reset) {
6869 const struct nftables_pernet *nft_net;
6870 char *buf;
6871
6872 nft_net = nft_pernet(net);
6873 buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, nft_net->base_seq);
6874
6875 audit_log_nfcfg(buf,
6876 family,
6877 obj->handle,
6878 AUDIT_NFT_OP_OBJ_RESET,
6879 GFP_ATOMIC);
6880 kfree(buf);
6881 }
6882
6883 err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
6884 info->nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
6885 family, table, obj, reset);
6886 if (err < 0)
6887 goto err_fill_obj_info;
6888
6889 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
6890
6891 err_fill_obj_info:
6892 kfree_skb(skb2);
6893 return err;
6894 }
6895
6896 static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
6897 {
6898 if (obj->ops->destroy)
6899 obj->ops->destroy(ctx, obj);
6900
6901 module_put(obj->ops->type->owner);
6902 kfree(obj->key.name);
6903 kfree(obj->udata);
6904 kfree(obj);
6905 }
6906
6907 static int nf_tables_delobj(struct sk_buff *skb, const struct nfnl_info *info,
6908 const struct nlattr * const nla[])
6909 {
6910 struct netlink_ext_ack *extack = info->extack;
6911 u8 genmask = nft_genmask_next(info->net);
6912 u8 family = info->nfmsg->nfgen_family;
6913 struct net *net = info->net;
6914 const struct nlattr *attr;
6915 struct nft_table *table;
6916 struct nft_object *obj;
6917 struct nft_ctx ctx;
6918 u32 objtype;
6919
6920 if (!nla[NFTA_OBJ_TYPE] ||
6921 (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
6922 return -EINVAL;
6923
6924 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
6925 NETLINK_CB(skb).portid);
6926 if (IS_ERR(table)) {
6927 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6928 return PTR_ERR(table);
6929 }
6930
6931 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6932 if (nla[NFTA_OBJ_HANDLE]) {
6933 attr = nla[NFTA_OBJ_HANDLE];
6934 obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
6935 } else {
6936 attr = nla[NFTA_OBJ_NAME];
6937 obj = nft_obj_lookup(net, table, attr, objtype, genmask);
6938 }
6939
6940 if (IS_ERR(obj)) {
6941 NL_SET_BAD_ATTR(extack, attr);
6942 return PTR_ERR(obj);
6943 }
6944 if (obj->use > 0) {
6945 NL_SET_BAD_ATTR(extack, attr);
6946 return -EBUSY;
6947 }
6948
6949 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6950
6951 return nft_delobj(&ctx, obj);
6952 }
6953
6954 void nft_obj_notify(struct net *net, const struct nft_table *table,
6955 struct nft_object *obj, u32 portid, u32 seq, int event,
6956 u16 flags, int family, int report, gfp_t gfp)
6957 {
6958 struct nftables_pernet *nft_net = nft_pernet(net);
6959 struct sk_buff *skb;
6960 int err;
6961 char *buf = kasprintf(gfp, "%s:%u",
6962 table->name, nft_net->base_seq);
6963
6964 audit_log_nfcfg(buf,
6965 family,
6966 obj->handle,
6967 event == NFT_MSG_NEWOBJ ?
6968 AUDIT_NFT_OP_OBJ_REGISTER :
6969 AUDIT_NFT_OP_OBJ_UNREGISTER,
6970 gfp);
6971 kfree(buf);
6972
6973 if (!report &&
6974 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
6975 return;
6976
6977 skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
6978 if (skb == NULL)
6979 goto err;
6980
6981 err = nf_tables_fill_obj_info(skb, net, portid, seq, event,
6982 flags & (NLM_F_CREATE | NLM_F_EXCL),
6983 family, table, obj, false);
6984 if (err < 0) {
6985 kfree_skb(skb);
6986 goto err;
6987 }
6988
6989 nft_notify_enqueue(skb, report, &nft_net->notify_list);
6990 return;
6991 err:
6992 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
6993 }
6994 EXPORT_SYMBOL_GPL(nft_obj_notify);
6995
6996 static void nf_tables_obj_notify(const struct nft_ctx *ctx,
6997 struct nft_object *obj, int event)
6998 {
6999 nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid, ctx->seq, event,
7000 ctx->flags, ctx->family, ctx->report, GFP_KERNEL);
7001 }
7002
7003 /*
7004 * Flow tables
7005 */
7006 void nft_register_flowtable_type(struct nf_flowtable_type *type)
7007 {
7008 nfnl_lock(NFNL_SUBSYS_NFTABLES);
7009 list_add_tail_rcu(&type->list, &nf_tables_flowtables);
7010 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7011 }
7012 EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
7013
7014 void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
7015 {
7016 nfnl_lock(NFNL_SUBSYS_NFTABLES);
7017 list_del_rcu(&type->list);
7018 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7019 }
7020 EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
7021
7022 static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
7023 [NFTA_FLOWTABLE_TABLE] = { .type = NLA_STRING,
7024 .len = NFT_NAME_MAXLEN - 1 },
7025 [NFTA_FLOWTABLE_NAME] = { .type = NLA_STRING,
7026 .len = NFT_NAME_MAXLEN - 1 },
7027 [NFTA_FLOWTABLE_HOOK] = { .type = NLA_NESTED },
7028 [NFTA_FLOWTABLE_HANDLE] = { .type = NLA_U64 },
7029 [NFTA_FLOWTABLE_FLAGS] = { .type = NLA_U32 },
7030 };
7031
7032 struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
7033 const struct nlattr *nla, u8 genmask)
7034 {
7035 struct nft_flowtable *flowtable;
7036
7037 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7038 if (!nla_strcmp(nla, flowtable->name) &&
7039 nft_active_genmask(flowtable, genmask))
7040 return flowtable;
7041 }
7042 return ERR_PTR(-ENOENT);
7043 }
7044 EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
7045
7046 void nf_tables_deactivate_flowtable(const struct nft_ctx *ctx,
7047 struct nft_flowtable *flowtable,
7048 enum nft_trans_phase phase)
7049 {
7050 switch (phase) {
7051 case NFT_TRANS_PREPARE:
7052 case NFT_TRANS_ABORT:
7053 case NFT_TRANS_RELEASE:
7054 flowtable->use--;
7055 fallthrough;
7056 default:
7057 return;
7058 }
7059 }
7060 EXPORT_SYMBOL_GPL(nf_tables_deactivate_flowtable);
7061
7062 static struct nft_flowtable *
7063 nft_flowtable_lookup_byhandle(const struct nft_table *table,
7064 const struct nlattr *nla, u8 genmask)
7065 {
7066 struct nft_flowtable *flowtable;
7067
7068 list_for_each_entry(flowtable, &table->flowtables, list) {
7069 if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
7070 nft_active_genmask(flowtable, genmask))
7071 return flowtable;
7072 }
7073 return ERR_PTR(-ENOENT);
7074 }
7075
7076 struct nft_flowtable_hook {
7077 u32 num;
7078 int priority;
7079 struct list_head list;
7080 };
7081
7082 static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
7083 [NFTA_FLOWTABLE_HOOK_NUM] = { .type = NLA_U32 },
7084 [NFTA_FLOWTABLE_HOOK_PRIORITY] = { .type = NLA_U32 },
7085 [NFTA_FLOWTABLE_HOOK_DEVS] = { .type = NLA_NESTED },
7086 };
7087
7088 static int nft_flowtable_parse_hook(const struct nft_ctx *ctx,
7089 const struct nlattr *attr,
7090 struct nft_flowtable_hook *flowtable_hook,
7091 struct nft_flowtable *flowtable, bool add)
7092 {
7093 struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
7094 struct nft_hook *hook;
7095 int hooknum, priority;
7096 int err;
7097
7098 INIT_LIST_HEAD(&flowtable_hook->list);
7099
7100 err = nla_parse_nested_deprecated(tb, NFTA_FLOWTABLE_HOOK_MAX, attr,
7101 nft_flowtable_hook_policy, NULL);
7102 if (err < 0)
7103 return err;
7104
7105 if (add) {
7106 if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
7107 !tb[NFTA_FLOWTABLE_HOOK_PRIORITY])
7108 return -EINVAL;
7109
7110 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7111 if (hooknum != NF_NETDEV_INGRESS)
7112 return -EOPNOTSUPP;
7113
7114 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7115
7116 flowtable_hook->priority = priority;
7117 flowtable_hook->num = hooknum;
7118 } else {
7119 if (tb[NFTA_FLOWTABLE_HOOK_NUM]) {
7120 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7121 if (hooknum != flowtable->hooknum)
7122 return -EOPNOTSUPP;
7123 }
7124
7125 if (tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
7126 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7127 if (priority != flowtable->data.priority)
7128 return -EOPNOTSUPP;
7129 }
7130
7131 flowtable_hook->priority = flowtable->data.priority;
7132 flowtable_hook->num = flowtable->hooknum;
7133 }
7134
7135 if (tb[NFTA_FLOWTABLE_HOOK_DEVS]) {
7136 err = nf_tables_parse_netdev_hooks(ctx->net,
7137 tb[NFTA_FLOWTABLE_HOOK_DEVS],
7138 &flowtable_hook->list);
7139 if (err < 0)
7140 return err;
7141 }
7142
7143 list_for_each_entry(hook, &flowtable_hook->list, list) {
7144 hook->ops.pf = NFPROTO_NETDEV;
7145 hook->ops.hooknum = flowtable_hook->num;
7146 hook->ops.priority = flowtable_hook->priority;
7147 hook->ops.priv = &flowtable->data;
7148 hook->ops.hook = flowtable->data.type->hook;
7149 }
7150
7151 return err;
7152 }
7153
7154 static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
7155 {
7156 const struct nf_flowtable_type *type;
7157
7158 list_for_each_entry(type, &nf_tables_flowtables, list) {
7159 if (family == type->family)
7160 return type;
7161 }
7162 return NULL;
7163 }
7164
7165 static const struct nf_flowtable_type *
7166 nft_flowtable_type_get(struct net *net, u8 family)
7167 {
7168 const struct nf_flowtable_type *type;
7169
7170 type = __nft_flowtable_type_get(family);
7171 if (type != NULL && try_module_get(type->owner))
7172 return type;
7173
7174 lockdep_nfnl_nft_mutex_not_held();
7175 #ifdef CONFIG_MODULES
7176 if (type == NULL) {
7177 if (nft_request_module(net, "nf-flowtable-%u", family) == -EAGAIN)
7178 return ERR_PTR(-EAGAIN);
7179 }
7180 #endif
7181 return ERR_PTR(-ENOENT);
7182 }
7183
7184 /* Only called from error and netdev event paths. */
7185 static void nft_unregister_flowtable_hook(struct net *net,
7186 struct nft_flowtable *flowtable,
7187 struct nft_hook *hook)
7188 {
7189 nf_unregister_net_hook(net, &hook->ops);
7190 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7191 FLOW_BLOCK_UNBIND);
7192 }
7193
7194 static void nft_unregister_flowtable_net_hooks(struct net *net,
7195 struct list_head *hook_list)
7196 {
7197 struct nft_hook *hook;
7198
7199 list_for_each_entry(hook, hook_list, list)
7200 nf_unregister_net_hook(net, &hook->ops);
7201 }
7202
7203 static int nft_register_flowtable_net_hooks(struct net *net,
7204 struct nft_table *table,
7205 struct list_head *hook_list,
7206 struct nft_flowtable *flowtable)
7207 {
7208 struct nft_hook *hook, *hook2, *next;
7209 struct nft_flowtable *ft;
7210 int err, i = 0;
7211
7212 list_for_each_entry(hook, hook_list, list) {
7213 list_for_each_entry(ft, &table->flowtables, list) {
7214 if (!nft_is_active_next(net, ft))
7215 continue;
7216
7217 list_for_each_entry(hook2, &ft->hook_list, list) {
7218 if (hook->ops.dev == hook2->ops.dev &&
7219 hook->ops.pf == hook2->ops.pf) {
7220 err = -EEXIST;
7221 goto err_unregister_net_hooks;
7222 }
7223 }
7224 }
7225
7226 err = flowtable->data.type->setup(&flowtable->data,
7227 hook->ops.dev,
7228 FLOW_BLOCK_BIND);
7229 if (err < 0)
7230 goto err_unregister_net_hooks;
7231
7232 err = nf_register_net_hook(net, &hook->ops);
7233 if (err < 0) {
7234 flowtable->data.type->setup(&flowtable->data,
7235 hook->ops.dev,
7236 FLOW_BLOCK_UNBIND);
7237 goto err_unregister_net_hooks;
7238 }
7239
7240 i++;
7241 }
7242
7243 return 0;
7244
7245 err_unregister_net_hooks:
7246 list_for_each_entry_safe(hook, next, hook_list, list) {
7247 if (i-- <= 0)
7248 break;
7249
7250 nft_unregister_flowtable_hook(net, flowtable, hook);
7251 list_del_rcu(&hook->list);
7252 kfree_rcu(hook, rcu);
7253 }
7254
7255 return err;
7256 }
7257
7258 static void nft_flowtable_hooks_destroy(struct list_head *hook_list)
7259 {
7260 struct nft_hook *hook, *next;
7261
7262 list_for_each_entry_safe(hook, next, hook_list, list) {
7263 list_del_rcu(&hook->list);
7264 kfree_rcu(hook, rcu);
7265 }
7266 }
7267
7268 static int nft_flowtable_update(struct nft_ctx *ctx, const struct nlmsghdr *nlh,
7269 struct nft_flowtable *flowtable)
7270 {
7271 const struct nlattr * const *nla = ctx->nla;
7272 struct nft_flowtable_hook flowtable_hook;
7273 struct nft_hook *hook, *next;
7274 struct nft_trans *trans;
7275 bool unregister = false;
7276 u32 flags;
7277 int err;
7278
7279 err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7280 &flowtable_hook, flowtable, false);
7281 if (err < 0)
7282 return err;
7283
7284 list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7285 if (nft_hook_list_find(&flowtable->hook_list, hook)) {
7286 list_del(&hook->list);
7287 kfree(hook);
7288 }
7289 }
7290
7291 if (nla[NFTA_FLOWTABLE_FLAGS]) {
7292 flags = ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7293 if (flags & ~NFT_FLOWTABLE_MASK)
7294 return -EOPNOTSUPP;
7295 if ((flowtable->data.flags & NFT_FLOWTABLE_HW_OFFLOAD) ^
7296 (flags & NFT_FLOWTABLE_HW_OFFLOAD))
7297 return -EOPNOTSUPP;
7298 } else {
7299 flags = flowtable->data.flags;
7300 }
7301
7302 err = nft_register_flowtable_net_hooks(ctx->net, ctx->table,
7303 &flowtable_hook.list, flowtable);
7304 if (err < 0)
7305 goto err_flowtable_update_hook;
7306
7307 trans = nft_trans_alloc(ctx, NFT_MSG_NEWFLOWTABLE,
7308 sizeof(struct nft_trans_flowtable));
7309 if (!trans) {
7310 unregister = true;
7311 err = -ENOMEM;
7312 goto err_flowtable_update_hook;
7313 }
7314
7315 nft_trans_flowtable_flags(trans) = flags;
7316 nft_trans_flowtable(trans) = flowtable;
7317 nft_trans_flowtable_update(trans) = true;
7318 INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7319 list_splice(&flowtable_hook.list, &nft_trans_flowtable_hooks(trans));
7320
7321 nft_trans_commit_list_add_tail(ctx->net, trans);
7322
7323 return 0;
7324
7325 err_flowtable_update_hook:
7326 list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7327 if (unregister)
7328 nft_unregister_flowtable_hook(ctx->net, flowtable, hook);
7329 list_del_rcu(&hook->list);
7330 kfree_rcu(hook, rcu);
7331 }
7332
7333 return err;
7334
7335 }
7336
7337 static int nf_tables_newflowtable(struct sk_buff *skb,
7338 const struct nfnl_info *info,
7339 const struct nlattr * const nla[])
7340 {
7341 struct netlink_ext_ack *extack = info->extack;
7342 struct nft_flowtable_hook flowtable_hook;
7343 u8 genmask = nft_genmask_next(info->net);
7344 u8 family = info->nfmsg->nfgen_family;
7345 const struct nf_flowtable_type *type;
7346 struct nft_flowtable *flowtable;
7347 struct nft_hook *hook, *next;
7348 struct net *net = info->net;
7349 struct nft_table *table;
7350 struct nft_ctx ctx;
7351 int err;
7352
7353 if (!nla[NFTA_FLOWTABLE_TABLE] ||
7354 !nla[NFTA_FLOWTABLE_NAME] ||
7355 !nla[NFTA_FLOWTABLE_HOOK])
7356 return -EINVAL;
7357
7358 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7359 genmask, NETLINK_CB(skb).portid);
7360 if (IS_ERR(table)) {
7361 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7362 return PTR_ERR(table);
7363 }
7364
7365 flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7366 genmask);
7367 if (IS_ERR(flowtable)) {
7368 err = PTR_ERR(flowtable);
7369 if (err != -ENOENT) {
7370 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7371 return err;
7372 }
7373 } else {
7374 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
7375 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7376 return -EEXIST;
7377 }
7378
7379 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7380
7381 return nft_flowtable_update(&ctx, info->nlh, flowtable);
7382 }
7383
7384 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7385
7386 flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL);
7387 if (!flowtable)
7388 return -ENOMEM;
7389
7390 flowtable->table = table;
7391 flowtable->handle = nf_tables_alloc_handle(table);
7392 INIT_LIST_HEAD(&flowtable->hook_list);
7393
7394 flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL);
7395 if (!flowtable->name) {
7396 err = -ENOMEM;
7397 goto err1;
7398 }
7399
7400 type = nft_flowtable_type_get(net, family);
7401 if (IS_ERR(type)) {
7402 err = PTR_ERR(type);
7403 goto err2;
7404 }
7405
7406 if (nla[NFTA_FLOWTABLE_FLAGS]) {
7407 flowtable->data.flags =
7408 ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7409 if (flowtable->data.flags & ~NFT_FLOWTABLE_MASK) {
7410 err = -EOPNOTSUPP;
7411 goto err3;
7412 }
7413 }
7414
7415 write_pnet(&flowtable->data.net, net);
7416 flowtable->data.type = type;
7417 err = type->init(&flowtable->data);
7418 if (err < 0)
7419 goto err3;
7420
7421 err = nft_flowtable_parse_hook(&ctx, nla[NFTA_FLOWTABLE_HOOK],
7422 &flowtable_hook, flowtable, true);
7423 if (err < 0)
7424 goto err4;
7425
7426 list_splice(&flowtable_hook.list, &flowtable->hook_list);
7427 flowtable->data.priority = flowtable_hook.priority;
7428 flowtable->hooknum = flowtable_hook.num;
7429
7430 err = nft_register_flowtable_net_hooks(ctx.net, table,
7431 &flowtable->hook_list,
7432 flowtable);
7433 if (err < 0) {
7434 nft_flowtable_hooks_destroy(&flowtable->hook_list);
7435 goto err4;
7436 }
7437
7438 err = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
7439 if (err < 0)
7440 goto err5;
7441
7442 list_add_tail_rcu(&flowtable->list, &table->flowtables);
7443 table->use++;
7444
7445 return 0;
7446 err5:
7447 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7448 nft_unregister_flowtable_hook(net, flowtable, hook);
7449 list_del_rcu(&hook->list);
7450 kfree_rcu(hook, rcu);
7451 }
7452 err4:
7453 flowtable->data.type->free(&flowtable->data);
7454 err3:
7455 module_put(type->owner);
7456 err2:
7457 kfree(flowtable->name);
7458 err1:
7459 kfree(flowtable);
7460 return err;
7461 }
7462
7463 static void nft_flowtable_hook_release(struct nft_flowtable_hook *flowtable_hook)
7464 {
7465 struct nft_hook *this, *next;
7466
7467 list_for_each_entry_safe(this, next, &flowtable_hook->list, list) {
7468 list_del(&this->list);
7469 kfree(this);
7470 }
7471 }
7472
7473 static int nft_delflowtable_hook(struct nft_ctx *ctx,
7474 struct nft_flowtable *flowtable)
7475 {
7476 const struct nlattr * const *nla = ctx->nla;
7477 struct nft_flowtable_hook flowtable_hook;
7478 struct nft_hook *this, *hook;
7479 struct nft_trans *trans;
7480 int err;
7481
7482 err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7483 &flowtable_hook, flowtable, false);
7484 if (err < 0)
7485 return err;
7486
7487 list_for_each_entry(this, &flowtable_hook.list, list) {
7488 hook = nft_hook_list_find(&flowtable->hook_list, this);
7489 if (!hook) {
7490 err = -ENOENT;
7491 goto err_flowtable_del_hook;
7492 }
7493 hook->inactive = true;
7494 }
7495
7496 trans = nft_trans_alloc(ctx, NFT_MSG_DELFLOWTABLE,
7497 sizeof(struct nft_trans_flowtable));
7498 if (!trans) {
7499 err = -ENOMEM;
7500 goto err_flowtable_del_hook;
7501 }
7502
7503 nft_trans_flowtable(trans) = flowtable;
7504 nft_trans_flowtable_update(trans) = true;
7505 INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7506 nft_flowtable_hook_release(&flowtable_hook);
7507
7508 nft_trans_commit_list_add_tail(ctx->net, trans);
7509
7510 return 0;
7511
7512 err_flowtable_del_hook:
7513 list_for_each_entry(this, &flowtable_hook.list, list) {
7514 hook = nft_hook_list_find(&flowtable->hook_list, this);
7515 if (!hook)
7516 break;
7517
7518 hook->inactive = false;
7519 }
7520 nft_flowtable_hook_release(&flowtable_hook);
7521
7522 return err;
7523 }
7524
7525 static int nf_tables_delflowtable(struct sk_buff *skb,
7526 const struct nfnl_info *info,
7527 const struct nlattr * const nla[])
7528 {
7529 struct netlink_ext_ack *extack = info->extack;
7530 u8 genmask = nft_genmask_next(info->net);
7531 u8 family = info->nfmsg->nfgen_family;
7532 struct nft_flowtable *flowtable;
7533 struct net *net = info->net;
7534 const struct nlattr *attr;
7535 struct nft_table *table;
7536 struct nft_ctx ctx;
7537
7538 if (!nla[NFTA_FLOWTABLE_TABLE] ||
7539 (!nla[NFTA_FLOWTABLE_NAME] &&
7540 !nla[NFTA_FLOWTABLE_HANDLE]))
7541 return -EINVAL;
7542
7543 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7544 genmask, NETLINK_CB(skb).portid);
7545 if (IS_ERR(table)) {
7546 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7547 return PTR_ERR(table);
7548 }
7549
7550 if (nla[NFTA_FLOWTABLE_HANDLE]) {
7551 attr = nla[NFTA_FLOWTABLE_HANDLE];
7552 flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
7553 } else {
7554 attr = nla[NFTA_FLOWTABLE_NAME];
7555 flowtable = nft_flowtable_lookup(table, attr, genmask);
7556 }
7557
7558 if (IS_ERR(flowtable)) {
7559 NL_SET_BAD_ATTR(extack, attr);
7560 return PTR_ERR(flowtable);
7561 }
7562
7563 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7564
7565 if (nla[NFTA_FLOWTABLE_HOOK])
7566 return nft_delflowtable_hook(&ctx, flowtable);
7567
7568 if (flowtable->use > 0) {
7569 NL_SET_BAD_ATTR(extack, attr);
7570 return -EBUSY;
7571 }
7572
7573 return nft_delflowtable(&ctx, flowtable);
7574 }
7575
7576 static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
7577 u32 portid, u32 seq, int event,
7578 u32 flags, int family,
7579 struct nft_flowtable *flowtable,
7580 struct list_head *hook_list)
7581 {
7582 struct nlattr *nest, *nest_devs;
7583 struct nft_hook *hook;
7584 struct nlmsghdr *nlh;
7585
7586 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
7587 nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
7588 NFNETLINK_V0, nft_base_seq(net));
7589 if (!nlh)
7590 goto nla_put_failure;
7591
7592 if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
7593 nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
7594 nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
7595 nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
7596 NFTA_FLOWTABLE_PAD) ||
7597 nla_put_be32(skb, NFTA_FLOWTABLE_FLAGS, htonl(flowtable->data.flags)))
7598 goto nla_put_failure;
7599
7600 nest = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK);
7601 if (!nest)
7602 goto nla_put_failure;
7603 if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
7604 nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->data.priority)))
7605 goto nla_put_failure;
7606
7607 nest_devs = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK_DEVS);
7608 if (!nest_devs)
7609 goto nla_put_failure;
7610
7611 list_for_each_entry_rcu(hook, hook_list, list) {
7612 if (nla_put_string(skb, NFTA_DEVICE_NAME, hook->ops.dev->name))
7613 goto nla_put_failure;
7614 }
7615 nla_nest_end(skb, nest_devs);
7616 nla_nest_end(skb, nest);
7617
7618 nlmsg_end(skb, nlh);
7619 return 0;
7620
7621 nla_put_failure:
7622 nlmsg_trim(skb, nlh);
7623 return -1;
7624 }
7625
7626 struct nft_flowtable_filter {
7627 char *table;
7628 };
7629
7630 static int nf_tables_dump_flowtable(struct sk_buff *skb,
7631 struct netlink_callback *cb)
7632 {
7633 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
7634 struct nft_flowtable_filter *filter = cb->data;
7635 unsigned int idx = 0, s_idx = cb->args[0];
7636 struct net *net = sock_net(skb->sk);
7637 int family = nfmsg->nfgen_family;
7638 struct nft_flowtable *flowtable;
7639 struct nftables_pernet *nft_net;
7640 const struct nft_table *table;
7641
7642 rcu_read_lock();
7643 nft_net = nft_pernet(net);
7644 cb->seq = nft_net->base_seq;
7645
7646 list_for_each_entry_rcu(table, &nft_net->tables, list) {
7647 if (family != NFPROTO_UNSPEC && family != table->family)
7648 continue;
7649
7650 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7651 if (!nft_is_active(net, flowtable))
7652 goto cont;
7653 if (idx < s_idx)
7654 goto cont;
7655 if (idx > s_idx)
7656 memset(&cb->args[1], 0,
7657 sizeof(cb->args) - sizeof(cb->args[0]));
7658 if (filter && filter->table &&
7659 strcmp(filter->table, table->name))
7660 goto cont;
7661
7662 if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
7663 cb->nlh->nlmsg_seq,
7664 NFT_MSG_NEWFLOWTABLE,
7665 NLM_F_MULTI | NLM_F_APPEND,
7666 table->family,
7667 flowtable,
7668 &flowtable->hook_list) < 0)
7669 goto done;
7670
7671 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
7672 cont:
7673 idx++;
7674 }
7675 }
7676 done:
7677 rcu_read_unlock();
7678
7679 cb->args[0] = idx;
7680 return skb->len;
7681 }
7682
7683 static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
7684 {
7685 const struct nlattr * const *nla = cb->data;
7686 struct nft_flowtable_filter *filter = NULL;
7687
7688 if (nla[NFTA_FLOWTABLE_TABLE]) {
7689 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
7690 if (!filter)
7691 return -ENOMEM;
7692
7693 filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
7694 GFP_ATOMIC);
7695 if (!filter->table) {
7696 kfree(filter);
7697 return -ENOMEM;
7698 }
7699 }
7700
7701 cb->data = filter;
7702 return 0;
7703 }
7704
7705 static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
7706 {
7707 struct nft_flowtable_filter *filter = cb->data;
7708
7709 if (!filter)
7710 return 0;
7711
7712 kfree(filter->table);
7713 kfree(filter);
7714
7715 return 0;
7716 }
7717
7718 /* called with rcu_read_lock held */
7719 static int nf_tables_getflowtable(struct sk_buff *skb,
7720 const struct nfnl_info *info,
7721 const struct nlattr * const nla[])
7722 {
7723 u8 genmask = nft_genmask_cur(info->net);
7724 u8 family = info->nfmsg->nfgen_family;
7725 struct nft_flowtable *flowtable;
7726 const struct nft_table *table;
7727 struct net *net = info->net;
7728 struct sk_buff *skb2;
7729 int err;
7730
7731 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
7732 struct netlink_dump_control c = {
7733 .start = nf_tables_dump_flowtable_start,
7734 .dump = nf_tables_dump_flowtable,
7735 .done = nf_tables_dump_flowtable_done,
7736 .module = THIS_MODULE,
7737 .data = (void *)nla,
7738 };
7739
7740 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
7741 }
7742
7743 if (!nla[NFTA_FLOWTABLE_NAME])
7744 return -EINVAL;
7745
7746 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7747 genmask, 0);
7748 if (IS_ERR(table))
7749 return PTR_ERR(table);
7750
7751 flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7752 genmask);
7753 if (IS_ERR(flowtable))
7754 return PTR_ERR(flowtable);
7755
7756 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7757 if (!skb2)
7758 return -ENOMEM;
7759
7760 err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
7761 info->nlh->nlmsg_seq,
7762 NFT_MSG_NEWFLOWTABLE, 0, family,
7763 flowtable, &flowtable->hook_list);
7764 if (err < 0)
7765 goto err_fill_flowtable_info;
7766
7767 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7768
7769 err_fill_flowtable_info:
7770 kfree_skb(skb2);
7771 return err;
7772 }
7773
7774 static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
7775 struct nft_flowtable *flowtable,
7776 struct list_head *hook_list,
7777 int event)
7778 {
7779 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
7780 struct sk_buff *skb;
7781 u16 flags = 0;
7782 int err;
7783
7784 if (!ctx->report &&
7785 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
7786 return;
7787
7788 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
7789 if (skb == NULL)
7790 goto err;
7791
7792 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
7793 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
7794
7795 err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
7796 ctx->seq, event, flags,
7797 ctx->family, flowtable, hook_list);
7798 if (err < 0) {
7799 kfree_skb(skb);
7800 goto err;
7801 }
7802
7803 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
7804 return;
7805 err:
7806 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
7807 }
7808
7809 static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
7810 {
7811 struct nft_hook *hook, *next;
7812
7813 flowtable->data.type->free(&flowtable->data);
7814 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7815 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7816 FLOW_BLOCK_UNBIND);
7817 list_del_rcu(&hook->list);
7818 kfree(hook);
7819 }
7820 kfree(flowtable->name);
7821 module_put(flowtable->data.type->owner);
7822 kfree(flowtable);
7823 }
7824
7825 static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
7826 u32 portid, u32 seq)
7827 {
7828 struct nftables_pernet *nft_net = nft_pernet(net);
7829 struct nlmsghdr *nlh;
7830 char buf[TASK_COMM_LEN];
7831 int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
7832
7833 nlh = nfnl_msg_put(skb, portid, seq, event, 0, AF_UNSPEC,
7834 NFNETLINK_V0, nft_base_seq(net));
7835 if (!nlh)
7836 goto nla_put_failure;
7837
7838 if (nla_put_be32(skb, NFTA_GEN_ID, htonl(nft_net->base_seq)) ||
7839 nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
7840 nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
7841 goto nla_put_failure;
7842
7843 nlmsg_end(skb, nlh);
7844 return 0;
7845
7846 nla_put_failure:
7847 nlmsg_trim(skb, nlh);
7848 return -EMSGSIZE;
7849 }
7850
7851 static void nft_flowtable_event(unsigned long event, struct net_device *dev,
7852 struct nft_flowtable *flowtable)
7853 {
7854 struct nft_hook *hook;
7855
7856 list_for_each_entry(hook, &flowtable->hook_list, list) {
7857 if (hook->ops.dev != dev)
7858 continue;
7859
7860 /* flow_offload_netdev_event() cleans up entries for us. */
7861 nft_unregister_flowtable_hook(dev_net(dev), flowtable, hook);
7862 list_del_rcu(&hook->list);
7863 kfree_rcu(hook, rcu);
7864 break;
7865 }
7866 }
7867
7868 static int nf_tables_flowtable_event(struct notifier_block *this,
7869 unsigned long event, void *ptr)
7870 {
7871 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
7872 struct nft_flowtable *flowtable;
7873 struct nftables_pernet *nft_net;
7874 struct nft_table *table;
7875 struct net *net;
7876
7877 if (event != NETDEV_UNREGISTER)
7878 return 0;
7879
7880 net = dev_net(dev);
7881 nft_net = nft_pernet(net);
7882 mutex_lock(&nft_net->commit_mutex);
7883 list_for_each_entry(table, &nft_net->tables, list) {
7884 list_for_each_entry(flowtable, &table->flowtables, list) {
7885 nft_flowtable_event(event, dev, flowtable);
7886 }
7887 }
7888 mutex_unlock(&nft_net->commit_mutex);
7889
7890 return NOTIFY_DONE;
7891 }
7892
7893 static struct notifier_block nf_tables_flowtable_notifier = {
7894 .notifier_call = nf_tables_flowtable_event,
7895 };
7896
7897 static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
7898 int event)
7899 {
7900 struct nlmsghdr *nlh = nlmsg_hdr(skb);
7901 struct sk_buff *skb2;
7902 int err;
7903
7904 if (!nlmsg_report(nlh) &&
7905 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
7906 return;
7907
7908 skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
7909 if (skb2 == NULL)
7910 goto err;
7911
7912 err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
7913 nlh->nlmsg_seq);
7914 if (err < 0) {
7915 kfree_skb(skb2);
7916 goto err;
7917 }
7918
7919 nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
7920 nlmsg_report(nlh), GFP_KERNEL);
7921 return;
7922 err:
7923 nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
7924 -ENOBUFS);
7925 }
7926
7927 static int nf_tables_getgen(struct sk_buff *skb, const struct nfnl_info *info,
7928 const struct nlattr * const nla[])
7929 {
7930 struct sk_buff *skb2;
7931 int err;
7932
7933 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7934 if (skb2 == NULL)
7935 return -ENOMEM;
7936
7937 err = nf_tables_fill_gen_info(skb2, info->net, NETLINK_CB(skb).portid,
7938 info->nlh->nlmsg_seq);
7939 if (err < 0)
7940 goto err_fill_gen_info;
7941
7942 return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
7943
7944 err_fill_gen_info:
7945 kfree_skb(skb2);
7946 return err;
7947 }
7948
7949 static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
7950 [NFT_MSG_NEWTABLE] = {
7951 .call = nf_tables_newtable,
7952 .type = NFNL_CB_BATCH,
7953 .attr_count = NFTA_TABLE_MAX,
7954 .policy = nft_table_policy,
7955 },
7956 [NFT_MSG_GETTABLE] = {
7957 .call = nf_tables_gettable,
7958 .type = NFNL_CB_RCU,
7959 .attr_count = NFTA_TABLE_MAX,
7960 .policy = nft_table_policy,
7961 },
7962 [NFT_MSG_DELTABLE] = {
7963 .call = nf_tables_deltable,
7964 .type = NFNL_CB_BATCH,
7965 .attr_count = NFTA_TABLE_MAX,
7966 .policy = nft_table_policy,
7967 },
7968 [NFT_MSG_NEWCHAIN] = {
7969 .call = nf_tables_newchain,
7970 .type = NFNL_CB_BATCH,
7971 .attr_count = NFTA_CHAIN_MAX,
7972 .policy = nft_chain_policy,
7973 },
7974 [NFT_MSG_GETCHAIN] = {
7975 .call = nf_tables_getchain,
7976 .type = NFNL_CB_RCU,
7977 .attr_count = NFTA_CHAIN_MAX,
7978 .policy = nft_chain_policy,
7979 },
7980 [NFT_MSG_DELCHAIN] = {
7981 .call = nf_tables_delchain,
7982 .type = NFNL_CB_BATCH,
7983 .attr_count = NFTA_CHAIN_MAX,
7984 .policy = nft_chain_policy,
7985 },
7986 [NFT_MSG_NEWRULE] = {
7987 .call = nf_tables_newrule,
7988 .type = NFNL_CB_BATCH,
7989 .attr_count = NFTA_RULE_MAX,
7990 .policy = nft_rule_policy,
7991 },
7992 [NFT_MSG_GETRULE] = {
7993 .call = nf_tables_getrule,
7994 .type = NFNL_CB_RCU,
7995 .attr_count = NFTA_RULE_MAX,
7996 .policy = nft_rule_policy,
7997 },
7998 [NFT_MSG_DELRULE] = {
7999 .call = nf_tables_delrule,
8000 .type = NFNL_CB_BATCH,
8001 .attr_count = NFTA_RULE_MAX,
8002 .policy = nft_rule_policy,
8003 },
8004 [NFT_MSG_NEWSET] = {
8005 .call = nf_tables_newset,
8006 .type = NFNL_CB_BATCH,
8007 .attr_count = NFTA_SET_MAX,
8008 .policy = nft_set_policy,
8009 },
8010 [NFT_MSG_GETSET] = {
8011 .call = nf_tables_getset,
8012 .type = NFNL_CB_RCU,
8013 .attr_count = NFTA_SET_MAX,
8014 .policy = nft_set_policy,
8015 },
8016 [NFT_MSG_DELSET] = {
8017 .call = nf_tables_delset,
8018 .type = NFNL_CB_BATCH,
8019 .attr_count = NFTA_SET_MAX,
8020 .policy = nft_set_policy,
8021 },
8022 [NFT_MSG_NEWSETELEM] = {
8023 .call = nf_tables_newsetelem,
8024 .type = NFNL_CB_BATCH,
8025 .attr_count = NFTA_SET_ELEM_LIST_MAX,
8026 .policy = nft_set_elem_list_policy,
8027 },
8028 [NFT_MSG_GETSETELEM] = {
8029 .call = nf_tables_getsetelem,
8030 .type = NFNL_CB_RCU,
8031 .attr_count = NFTA_SET_ELEM_LIST_MAX,
8032 .policy = nft_set_elem_list_policy,
8033 },
8034 [NFT_MSG_DELSETELEM] = {
8035 .call = nf_tables_delsetelem,
8036 .type = NFNL_CB_BATCH,
8037 .attr_count = NFTA_SET_ELEM_LIST_MAX,
8038 .policy = nft_set_elem_list_policy,
8039 },
8040 [NFT_MSG_GETGEN] = {
8041 .call = nf_tables_getgen,
8042 .type = NFNL_CB_RCU,
8043 },
8044 [NFT_MSG_NEWOBJ] = {
8045 .call = nf_tables_newobj,
8046 .type = NFNL_CB_BATCH,
8047 .attr_count = NFTA_OBJ_MAX,
8048 .policy = nft_obj_policy,
8049 },
8050 [NFT_MSG_GETOBJ] = {
8051 .call = nf_tables_getobj,
8052 .type = NFNL_CB_RCU,
8053 .attr_count = NFTA_OBJ_MAX,
8054 .policy = nft_obj_policy,
8055 },
8056 [NFT_MSG_DELOBJ] = {
8057 .call = nf_tables_delobj,
8058 .type = NFNL_CB_BATCH,
8059 .attr_count = NFTA_OBJ_MAX,
8060 .policy = nft_obj_policy,
8061 },
8062 [NFT_MSG_GETOBJ_RESET] = {
8063 .call = nf_tables_getobj,
8064 .type = NFNL_CB_RCU,
8065 .attr_count = NFTA_OBJ_MAX,
8066 .policy = nft_obj_policy,
8067 },
8068 [NFT_MSG_NEWFLOWTABLE] = {
8069 .call = nf_tables_newflowtable,
8070 .type = NFNL_CB_BATCH,
8071 .attr_count = NFTA_FLOWTABLE_MAX,
8072 .policy = nft_flowtable_policy,
8073 },
8074 [NFT_MSG_GETFLOWTABLE] = {
8075 .call = nf_tables_getflowtable,
8076 .type = NFNL_CB_RCU,
8077 .attr_count = NFTA_FLOWTABLE_MAX,
8078 .policy = nft_flowtable_policy,
8079 },
8080 [NFT_MSG_DELFLOWTABLE] = {
8081 .call = nf_tables_delflowtable,
8082 .type = NFNL_CB_BATCH,
8083 .attr_count = NFTA_FLOWTABLE_MAX,
8084 .policy = nft_flowtable_policy,
8085 },
8086 };
8087
8088 static int nf_tables_validate(struct net *net)
8089 {
8090 struct nftables_pernet *nft_net = nft_pernet(net);
8091 struct nft_table *table;
8092
8093 switch (nft_net->validate_state) {
8094 case NFT_VALIDATE_SKIP:
8095 break;
8096 case NFT_VALIDATE_NEED:
8097 nft_validate_state_update(net, NFT_VALIDATE_DO);
8098 fallthrough;
8099 case NFT_VALIDATE_DO:
8100 list_for_each_entry(table, &nft_net->tables, list) {
8101 if (nft_table_validate(net, table) < 0)
8102 return -EAGAIN;
8103 }
8104 break;
8105 }
8106
8107 return 0;
8108 }
8109
8110 /* a drop policy has to be deferred until all rules have been activated,
8111 * otherwise a large ruleset that contains a drop-policy base chain will
8112 * cause all packets to get dropped until the full transaction has been
8113 * processed.
8114 *
8115 * We defer the drop policy until the transaction has been finalized.
8116 */
8117 static void nft_chain_commit_drop_policy(struct nft_trans *trans)
8118 {
8119 struct nft_base_chain *basechain;
8120
8121 if (nft_trans_chain_policy(trans) != NF_DROP)
8122 return;
8123
8124 if (!nft_is_base_chain(trans->ctx.chain))
8125 return;
8126
8127 basechain = nft_base_chain(trans->ctx.chain);
8128 basechain->policy = NF_DROP;
8129 }
8130
8131 static void nft_chain_commit_update(struct nft_trans *trans)
8132 {
8133 struct nft_base_chain *basechain;
8134
8135 if (nft_trans_chain_name(trans)) {
8136 rhltable_remove(&trans->ctx.table->chains_ht,
8137 &trans->ctx.chain->rhlhead,
8138 nft_chain_ht_params);
8139 swap(trans->ctx.chain->name, nft_trans_chain_name(trans));
8140 rhltable_insert_key(&trans->ctx.table->chains_ht,
8141 trans->ctx.chain->name,
8142 &trans->ctx.chain->rhlhead,
8143 nft_chain_ht_params);
8144 }
8145
8146 if (!nft_is_base_chain(trans->ctx.chain))
8147 return;
8148
8149 nft_chain_stats_replace(trans);
8150
8151 basechain = nft_base_chain(trans->ctx.chain);
8152
8153 switch (nft_trans_chain_policy(trans)) {
8154 case NF_DROP:
8155 case NF_ACCEPT:
8156 basechain->policy = nft_trans_chain_policy(trans);
8157 break;
8158 }
8159 }
8160
8161 static void nft_obj_commit_update(struct nft_trans *trans)
8162 {
8163 struct nft_object *newobj;
8164 struct nft_object *obj;
8165
8166 obj = nft_trans_obj(trans);
8167 newobj = nft_trans_obj_newobj(trans);
8168
8169 if (obj->ops->update)
8170 obj->ops->update(obj, newobj);
8171
8172 kfree(newobj);
8173 }
8174
8175 static void nft_commit_release(struct nft_trans *trans)
8176 {
8177 switch (trans->msg_type) {
8178 case NFT_MSG_DELTABLE:
8179 nf_tables_table_destroy(&trans->ctx);
8180 break;
8181 case NFT_MSG_NEWCHAIN:
8182 free_percpu(nft_trans_chain_stats(trans));
8183 kfree(nft_trans_chain_name(trans));
8184 break;
8185 case NFT_MSG_DELCHAIN:
8186 nf_tables_chain_destroy(&trans->ctx);
8187 break;
8188 case NFT_MSG_DELRULE:
8189 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8190 break;
8191 case NFT_MSG_DELSET:
8192 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8193 break;
8194 case NFT_MSG_DELSETELEM:
8195 nf_tables_set_elem_destroy(&trans->ctx,
8196 nft_trans_elem_set(trans),
8197 nft_trans_elem(trans).priv);
8198 break;
8199 case NFT_MSG_DELOBJ:
8200 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8201 break;
8202 case NFT_MSG_DELFLOWTABLE:
8203 if (nft_trans_flowtable_update(trans))
8204 nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8205 else
8206 nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8207 break;
8208 }
8209
8210 if (trans->put_net)
8211 put_net(trans->ctx.net);
8212
8213 kfree(trans);
8214 }
8215
8216 static void nf_tables_trans_destroy_work(struct work_struct *w)
8217 {
8218 struct nft_trans *trans, *next;
8219 LIST_HEAD(head);
8220
8221 spin_lock(&nf_tables_destroy_list_lock);
8222 list_splice_init(&nf_tables_destroy_list, &head);
8223 spin_unlock(&nf_tables_destroy_list_lock);
8224
8225 if (list_empty(&head))
8226 return;
8227
8228 synchronize_rcu();
8229
8230 list_for_each_entry_safe(trans, next, &head, list) {
8231 list_del(&trans->list);
8232 nft_commit_release(trans);
8233 }
8234 }
8235
8236 void nf_tables_trans_destroy_flush_work(void)
8237 {
8238 flush_work(&trans_destroy_work);
8239 }
8240 EXPORT_SYMBOL_GPL(nf_tables_trans_destroy_flush_work);
8241
8242 static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
8243 {
8244 struct nft_rule *rule;
8245 unsigned int alloc = 0;
8246 int i;
8247
8248 /* already handled or inactive chain? */
8249 if (chain->rules_next || !nft_is_active_next(net, chain))
8250 return 0;
8251
8252 rule = list_entry(&chain->rules, struct nft_rule, list);
8253 i = 0;
8254
8255 list_for_each_entry_continue(rule, &chain->rules, list) {
8256 if (nft_is_active_next(net, rule))
8257 alloc++;
8258 }
8259
8260 chain->rules_next = nf_tables_chain_alloc_rules(chain, alloc);
8261 if (!chain->rules_next)
8262 return -ENOMEM;
8263
8264 list_for_each_entry_continue(rule, &chain->rules, list) {
8265 if (nft_is_active_next(net, rule))
8266 chain->rules_next[i++] = rule;
8267 }
8268
8269 chain->rules_next[i] = NULL;
8270 return 0;
8271 }
8272
8273 static void nf_tables_commit_chain_prepare_cancel(struct net *net)
8274 {
8275 struct nftables_pernet *nft_net = nft_pernet(net);
8276 struct nft_trans *trans, *next;
8277
8278 list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8279 struct nft_chain *chain = trans->ctx.chain;
8280
8281 if (trans->msg_type == NFT_MSG_NEWRULE ||
8282 trans->msg_type == NFT_MSG_DELRULE) {
8283 kvfree(chain->rules_next);
8284 chain->rules_next = NULL;
8285 }
8286 }
8287 }
8288
8289 static void __nf_tables_commit_chain_free_rules_old(struct rcu_head *h)
8290 {
8291 struct nft_rules_old *o = container_of(h, struct nft_rules_old, h);
8292
8293 kvfree(o->start);
8294 }
8295
8296 static void nf_tables_commit_chain_free_rules_old(struct nft_rule **rules)
8297 {
8298 struct nft_rule **r = rules;
8299 struct nft_rules_old *old;
8300
8301 while (*r)
8302 r++;
8303
8304 r++; /* rcu_head is after end marker */
8305 old = (void *) r;
8306 old->start = rules;
8307
8308 call_rcu(&old->h, __nf_tables_commit_chain_free_rules_old);
8309 }
8310
8311 static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
8312 {
8313 struct nft_rule **g0, **g1;
8314 bool next_genbit;
8315
8316 next_genbit = nft_gencursor_next(net);
8317
8318 g0 = rcu_dereference_protected(chain->rules_gen_0,
8319 lockdep_commit_lock_is_held(net));
8320 g1 = rcu_dereference_protected(chain->rules_gen_1,
8321 lockdep_commit_lock_is_held(net));
8322
8323 /* No changes to this chain? */
8324 if (chain->rules_next == NULL) {
8325 /* chain had no change in last or next generation */
8326 if (g0 == g1)
8327 return;
8328 /*
8329 * chain had no change in this generation; make sure next
8330 * one uses same rules as current generation.
8331 */
8332 if (next_genbit) {
8333 rcu_assign_pointer(chain->rules_gen_1, g0);
8334 nf_tables_commit_chain_free_rules_old(g1);
8335 } else {
8336 rcu_assign_pointer(chain->rules_gen_0, g1);
8337 nf_tables_commit_chain_free_rules_old(g0);
8338 }
8339
8340 return;
8341 }
8342
8343 if (next_genbit)
8344 rcu_assign_pointer(chain->rules_gen_1, chain->rules_next);
8345 else
8346 rcu_assign_pointer(chain->rules_gen_0, chain->rules_next);
8347
8348 chain->rules_next = NULL;
8349
8350 if (g0 == g1)
8351 return;
8352
8353 if (next_genbit)
8354 nf_tables_commit_chain_free_rules_old(g1);
8355 else
8356 nf_tables_commit_chain_free_rules_old(g0);
8357 }
8358
8359 static void nft_obj_del(struct nft_object *obj)
8360 {
8361 rhltable_remove(&nft_objname_ht, &obj->rhlhead, nft_objname_ht_params);
8362 list_del_rcu(&obj->list);
8363 }
8364
8365 void nft_chain_del(struct nft_chain *chain)
8366 {
8367 struct nft_table *table = chain->table;
8368
8369 WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
8370 nft_chain_ht_params));
8371 list_del_rcu(&chain->list);
8372 }
8373
8374 static void nft_flowtable_hooks_del(struct nft_flowtable *flowtable,
8375 struct list_head *hook_list)
8376 {
8377 struct nft_hook *hook, *next;
8378
8379 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
8380 if (hook->inactive)
8381 list_move(&hook->list, hook_list);
8382 }
8383 }
8384
8385 static void nf_tables_module_autoload_cleanup(struct net *net)
8386 {
8387 struct nftables_pernet *nft_net = nft_pernet(net);
8388 struct nft_module_request *req, *next;
8389
8390 WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
8391 list_for_each_entry_safe(req, next, &nft_net->module_list, list) {
8392 WARN_ON_ONCE(!req->done);
8393 list_del(&req->list);
8394 kfree(req);
8395 }
8396 }
8397
8398 static void nf_tables_commit_release(struct net *net)
8399 {
8400 struct nftables_pernet *nft_net = nft_pernet(net);
8401 struct nft_trans *trans;
8402
8403 /* all side effects have to be made visible.
8404 * For example, if a chain named 'foo' has been deleted, a
8405 * new transaction must not find it anymore.
8406 *
8407 * Memory reclaim happens asynchronously from work queue
8408 * to prevent expensive synchronize_rcu() in commit phase.
8409 */
8410 if (list_empty(&nft_net->commit_list)) {
8411 nf_tables_module_autoload_cleanup(net);
8412 mutex_unlock(&nft_net->commit_mutex);
8413 return;
8414 }
8415
8416 trans = list_last_entry(&nft_net->commit_list,
8417 struct nft_trans, list);
8418 get_net(trans->ctx.net);
8419 WARN_ON_ONCE(trans->put_net);
8420
8421 trans->put_net = true;
8422 spin_lock(&nf_tables_destroy_list_lock);
8423 list_splice_tail_init(&nft_net->commit_list, &nf_tables_destroy_list);
8424 spin_unlock(&nf_tables_destroy_list_lock);
8425
8426 nf_tables_module_autoload_cleanup(net);
8427 schedule_work(&trans_destroy_work);
8428
8429 mutex_unlock(&nft_net->commit_mutex);
8430 }
8431
8432 static void nft_commit_notify(struct net *net, u32 portid)
8433 {
8434 struct nftables_pernet *nft_net = nft_pernet(net);
8435 struct sk_buff *batch_skb = NULL, *nskb, *skb;
8436 unsigned char *data;
8437 int len;
8438
8439 list_for_each_entry_safe(skb, nskb, &nft_net->notify_list, list) {
8440 if (!batch_skb) {
8441 new_batch:
8442 batch_skb = skb;
8443 len = NLMSG_GOODSIZE - skb->len;
8444 list_del(&skb->list);
8445 continue;
8446 }
8447 len -= skb->len;
8448 if (len > 0 && NFT_CB(skb).report == NFT_CB(batch_skb).report) {
8449 data = skb_put(batch_skb, skb->len);
8450 memcpy(data, skb->data, skb->len);
8451 list_del(&skb->list);
8452 kfree_skb(skb);
8453 continue;
8454 }
8455 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8456 NFT_CB(batch_skb).report, GFP_KERNEL);
8457 goto new_batch;
8458 }
8459
8460 if (batch_skb) {
8461 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8462 NFT_CB(batch_skb).report, GFP_KERNEL);
8463 }
8464
8465 WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
8466 }
8467
8468 static int nf_tables_commit_audit_alloc(struct list_head *adl,
8469 struct nft_table *table)
8470 {
8471 struct nft_audit_data *adp;
8472
8473 list_for_each_entry(adp, adl, list) {
8474 if (adp->table == table)
8475 return 0;
8476 }
8477 adp = kzalloc(sizeof(*adp), GFP_KERNEL);
8478 if (!adp)
8479 return -ENOMEM;
8480 adp->table = table;
8481 list_add(&adp->list, adl);
8482 return 0;
8483 }
8484
8485 static void nf_tables_commit_audit_free(struct list_head *adl)
8486 {
8487 struct nft_audit_data *adp, *adn;
8488
8489 list_for_each_entry_safe(adp, adn, adl, list) {
8490 list_del(&adp->list);
8491 kfree(adp);
8492 }
8493 }
8494
8495 static void nf_tables_commit_audit_collect(struct list_head *adl,
8496 struct nft_table *table, u32 op)
8497 {
8498 struct nft_audit_data *adp;
8499
8500 list_for_each_entry(adp, adl, list) {
8501 if (adp->table == table)
8502 goto found;
8503 }
8504 WARN_ONCE(1, "table=%s not expected in commit list", table->name);
8505 return;
8506 found:
8507 adp->entries++;
8508 if (!adp->op || adp->op > op)
8509 adp->op = op;
8510 }
8511
8512 #define AUNFTABLENAMELEN (NFT_TABLE_MAXNAMELEN + 22)
8513
8514 static void nf_tables_commit_audit_log(struct list_head *adl, u32 generation)
8515 {
8516 struct nft_audit_data *adp, *adn;
8517 char aubuf[AUNFTABLENAMELEN];
8518
8519 list_for_each_entry_safe(adp, adn, adl, list) {
8520 snprintf(aubuf, AUNFTABLENAMELEN, "%s:%u", adp->table->name,
8521 generation);
8522 audit_log_nfcfg(aubuf, adp->table->family, adp->entries,
8523 nft2audit_op[adp->op], GFP_KERNEL);
8524 list_del(&adp->list);
8525 kfree(adp);
8526 }
8527 }
8528
8529 static int nf_tables_commit(struct net *net, struct sk_buff *skb)
8530 {
8531 struct nftables_pernet *nft_net = nft_pernet(net);
8532 struct nft_trans *trans, *next;
8533 struct nft_trans_elem *te;
8534 struct nft_chain *chain;
8535 struct nft_table *table;
8536 LIST_HEAD(adl);
8537 int err;
8538
8539 if (list_empty(&nft_net->commit_list)) {
8540 mutex_unlock(&nft_net->commit_mutex);
8541 return 0;
8542 }
8543
8544 /* 0. Validate ruleset, otherwise roll back for error reporting. */
8545 if (nf_tables_validate(net) < 0)
8546 return -EAGAIN;
8547
8548 err = nft_flow_rule_offload_commit(net);
8549 if (err < 0)
8550 return err;
8551
8552 /* 1. Allocate space for next generation rules_gen_X[] */
8553 list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8554 int ret;
8555
8556 ret = nf_tables_commit_audit_alloc(&adl, trans->ctx.table);
8557 if (ret) {
8558 nf_tables_commit_chain_prepare_cancel(net);
8559 nf_tables_commit_audit_free(&adl);
8560 return ret;
8561 }
8562 if (trans->msg_type == NFT_MSG_NEWRULE ||
8563 trans->msg_type == NFT_MSG_DELRULE) {
8564 chain = trans->ctx.chain;
8565
8566 ret = nf_tables_commit_chain_prepare(net, chain);
8567 if (ret < 0) {
8568 nf_tables_commit_chain_prepare_cancel(net);
8569 nf_tables_commit_audit_free(&adl);
8570 return ret;
8571 }
8572 }
8573 }
8574
8575 /* step 2. Make rules_gen_X visible to packet path */
8576 list_for_each_entry(table, &nft_net->tables, list) {
8577 list_for_each_entry(chain, &table->chains, list)
8578 nf_tables_commit_chain(net, chain);
8579 }
8580
8581 /*
8582 * Bump generation counter, invalidate any dump in progress.
8583 * Cannot fail after this point.
8584 */
8585 while (++nft_net->base_seq == 0)
8586 ;
8587
8588 /* step 3. Start new generation, rules_gen_X now in use. */
8589 net->nft.gencursor = nft_gencursor_next(net);
8590
8591 list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8592 nf_tables_commit_audit_collect(&adl, trans->ctx.table,
8593 trans->msg_type);
8594 switch (trans->msg_type) {
8595 case NFT_MSG_NEWTABLE:
8596 if (nft_trans_table_update(trans)) {
8597 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
8598 nft_trans_destroy(trans);
8599 break;
8600 }
8601 if (trans->ctx.table->flags & NFT_TABLE_F_DORMANT)
8602 nf_tables_table_disable(net, trans->ctx.table);
8603
8604 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
8605 } else {
8606 nft_clear(net, trans->ctx.table);
8607 }
8608 nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
8609 nft_trans_destroy(trans);
8610 break;
8611 case NFT_MSG_DELTABLE:
8612 list_del_rcu(&trans->ctx.table->list);
8613 nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
8614 break;
8615 case NFT_MSG_NEWCHAIN:
8616 if (nft_trans_chain_update(trans)) {
8617 nft_chain_commit_update(trans);
8618 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8619 /* trans destroyed after rcu grace period */
8620 } else {
8621 nft_chain_commit_drop_policy(trans);
8622 nft_clear(net, trans->ctx.chain);
8623 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8624 nft_trans_destroy(trans);
8625 }
8626 break;
8627 case NFT_MSG_DELCHAIN:
8628 nft_chain_del(trans->ctx.chain);
8629 nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
8630 nf_tables_unregister_hook(trans->ctx.net,
8631 trans->ctx.table,
8632 trans->ctx.chain);
8633 break;
8634 case NFT_MSG_NEWRULE:
8635 nft_clear(trans->ctx.net, nft_trans_rule(trans));
8636 nf_tables_rule_notify(&trans->ctx,
8637 nft_trans_rule(trans),
8638 NFT_MSG_NEWRULE);
8639 nft_trans_destroy(trans);
8640 break;
8641 case NFT_MSG_DELRULE:
8642 list_del_rcu(&nft_trans_rule(trans)->list);
8643 nf_tables_rule_notify(&trans->ctx,
8644 nft_trans_rule(trans),
8645 NFT_MSG_DELRULE);
8646 nft_rule_expr_deactivate(&trans->ctx,
8647 nft_trans_rule(trans),
8648 NFT_TRANS_COMMIT);
8649 break;
8650 case NFT_MSG_NEWSET:
8651 nft_clear(net, nft_trans_set(trans));
8652 /* This avoids hitting -EBUSY when deleting the table
8653 * from the transaction.
8654 */
8655 if (nft_set_is_anonymous(nft_trans_set(trans)) &&
8656 !list_empty(&nft_trans_set(trans)->bindings))
8657 trans->ctx.table->use--;
8658
8659 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
8660 NFT_MSG_NEWSET, GFP_KERNEL);
8661 nft_trans_destroy(trans);
8662 break;
8663 case NFT_MSG_DELSET:
8664 list_del_rcu(&nft_trans_set(trans)->list);
8665 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
8666 NFT_MSG_DELSET, GFP_KERNEL);
8667 break;
8668 case NFT_MSG_NEWSETELEM:
8669 te = (struct nft_trans_elem *)trans->data;
8670
8671 nft_setelem_activate(net, te->set, &te->elem);
8672 nf_tables_setelem_notify(&trans->ctx, te->set,
8673 &te->elem,
8674 NFT_MSG_NEWSETELEM);
8675 nft_trans_destroy(trans);
8676 break;
8677 case NFT_MSG_DELSETELEM:
8678 te = (struct nft_trans_elem *)trans->data;
8679
8680 nf_tables_setelem_notify(&trans->ctx, te->set,
8681 &te->elem,
8682 NFT_MSG_DELSETELEM);
8683 nft_setelem_remove(net, te->set, &te->elem);
8684 if (!nft_setelem_is_catchall(te->set, &te->elem)) {
8685 atomic_dec(&te->set->nelems);
8686 te->set->ndeact--;
8687 }
8688 break;
8689 case NFT_MSG_NEWOBJ:
8690 if (nft_trans_obj_update(trans)) {
8691 nft_obj_commit_update(trans);
8692 nf_tables_obj_notify(&trans->ctx,
8693 nft_trans_obj(trans),
8694 NFT_MSG_NEWOBJ);
8695 } else {
8696 nft_clear(net, nft_trans_obj(trans));
8697 nf_tables_obj_notify(&trans->ctx,
8698 nft_trans_obj(trans),
8699 NFT_MSG_NEWOBJ);
8700 nft_trans_destroy(trans);
8701 }
8702 break;
8703 case NFT_MSG_DELOBJ:
8704 nft_obj_del(nft_trans_obj(trans));
8705 nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
8706 NFT_MSG_DELOBJ);
8707 break;
8708 case NFT_MSG_NEWFLOWTABLE:
8709 if (nft_trans_flowtable_update(trans)) {
8710 nft_trans_flowtable(trans)->data.flags =
8711 nft_trans_flowtable_flags(trans);
8712 nf_tables_flowtable_notify(&trans->ctx,
8713 nft_trans_flowtable(trans),
8714 &nft_trans_flowtable_hooks(trans),
8715 NFT_MSG_NEWFLOWTABLE);
8716 list_splice(&nft_trans_flowtable_hooks(trans),
8717 &nft_trans_flowtable(trans)->hook_list);
8718 } else {
8719 nft_clear(net, nft_trans_flowtable(trans));
8720 nf_tables_flowtable_notify(&trans->ctx,
8721 nft_trans_flowtable(trans),
8722 &nft_trans_flowtable(trans)->hook_list,
8723 NFT_MSG_NEWFLOWTABLE);
8724 }
8725 nft_trans_destroy(trans);
8726 break;
8727 case NFT_MSG_DELFLOWTABLE:
8728 if (nft_trans_flowtable_update(trans)) {
8729 nft_flowtable_hooks_del(nft_trans_flowtable(trans),
8730 &nft_trans_flowtable_hooks(trans));
8731 nf_tables_flowtable_notify(&trans->ctx,
8732 nft_trans_flowtable(trans),
8733 &nft_trans_flowtable_hooks(trans),
8734 NFT_MSG_DELFLOWTABLE);
8735 nft_unregister_flowtable_net_hooks(net,
8736 &nft_trans_flowtable_hooks(trans));
8737 } else {
8738 list_del_rcu(&nft_trans_flowtable(trans)->list);
8739 nf_tables_flowtable_notify(&trans->ctx,
8740 nft_trans_flowtable(trans),
8741 &nft_trans_flowtable(trans)->hook_list,
8742 NFT_MSG_DELFLOWTABLE);
8743 nft_unregister_flowtable_net_hooks(net,
8744 &nft_trans_flowtable(trans)->hook_list);
8745 }
8746 break;
8747 }
8748 }
8749
8750 nft_commit_notify(net, NETLINK_CB(skb).portid);
8751 nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
8752 nf_tables_commit_audit_log(&adl, nft_net->base_seq);
8753 nf_tables_commit_release(net);
8754
8755 return 0;
8756 }
8757
8758 static void nf_tables_module_autoload(struct net *net)
8759 {
8760 struct nftables_pernet *nft_net = nft_pernet(net);
8761 struct nft_module_request *req, *next;
8762 LIST_HEAD(module_list);
8763
8764 list_splice_init(&nft_net->module_list, &module_list);
8765 mutex_unlock(&nft_net->commit_mutex);
8766 list_for_each_entry_safe(req, next, &module_list, list) {
8767 request_module("%s", req->module);
8768 req->done = true;
8769 }
8770 mutex_lock(&nft_net->commit_mutex);
8771 list_splice(&module_list, &nft_net->module_list);
8772 }
8773
8774 static void nf_tables_abort_release(struct nft_trans *trans)
8775 {
8776 switch (trans->msg_type) {
8777 case NFT_MSG_NEWTABLE:
8778 nf_tables_table_destroy(&trans->ctx);
8779 break;
8780 case NFT_MSG_NEWCHAIN:
8781 nf_tables_chain_destroy(&trans->ctx);
8782 break;
8783 case NFT_MSG_NEWRULE:
8784 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8785 break;
8786 case NFT_MSG_NEWSET:
8787 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8788 break;
8789 case NFT_MSG_NEWSETELEM:
8790 nft_set_elem_destroy(nft_trans_elem_set(trans),
8791 nft_trans_elem(trans).priv, true);
8792 break;
8793 case NFT_MSG_NEWOBJ:
8794 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8795 break;
8796 case NFT_MSG_NEWFLOWTABLE:
8797 if (nft_trans_flowtable_update(trans))
8798 nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8799 else
8800 nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8801 break;
8802 }
8803 kfree(trans);
8804 }
8805
8806 static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)
8807 {
8808 struct nftables_pernet *nft_net = nft_pernet(net);
8809 struct nft_trans *trans, *next;
8810 struct nft_trans_elem *te;
8811 struct nft_hook *hook;
8812
8813 if (action == NFNL_ABORT_VALIDATE &&
8814 nf_tables_validate(net) < 0)
8815 return -EAGAIN;
8816
8817 list_for_each_entry_safe_reverse(trans, next, &nft_net->commit_list,
8818 list) {
8819 switch (trans->msg_type) {
8820 case NFT_MSG_NEWTABLE:
8821 if (nft_trans_table_update(trans)) {
8822 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
8823 nft_trans_destroy(trans);
8824 break;
8825 }
8826 if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_DORMANT) {
8827 nf_tables_table_disable(net, trans->ctx.table);
8828 trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
8829 } else if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_AWAKEN) {
8830 trans->ctx.table->flags &= ~NFT_TABLE_F_DORMANT;
8831 }
8832 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
8833 nft_trans_destroy(trans);
8834 } else {
8835 list_del_rcu(&trans->ctx.table->list);
8836 }
8837 break;
8838 case NFT_MSG_DELTABLE:
8839 nft_clear(trans->ctx.net, trans->ctx.table);
8840 nft_trans_destroy(trans);
8841 break;
8842 case NFT_MSG_NEWCHAIN:
8843 if (nft_trans_chain_update(trans)) {
8844 free_percpu(nft_trans_chain_stats(trans));
8845 kfree(nft_trans_chain_name(trans));
8846 nft_trans_destroy(trans);
8847 } else {
8848 if (nft_chain_is_bound(trans->ctx.chain)) {
8849 nft_trans_destroy(trans);
8850 break;
8851 }
8852 trans->ctx.table->use--;
8853 nft_chain_del(trans->ctx.chain);
8854 nf_tables_unregister_hook(trans->ctx.net,
8855 trans->ctx.table,
8856 trans->ctx.chain);
8857 }
8858 break;
8859 case NFT_MSG_DELCHAIN:
8860 trans->ctx.table->use++;
8861 nft_clear(trans->ctx.net, trans->ctx.chain);
8862 nft_trans_destroy(trans);
8863 break;
8864 case NFT_MSG_NEWRULE:
8865 trans->ctx.chain->use--;
8866 list_del_rcu(&nft_trans_rule(trans)->list);
8867 nft_rule_expr_deactivate(&trans->ctx,
8868 nft_trans_rule(trans),
8869 NFT_TRANS_ABORT);
8870 if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
8871 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
8872 break;
8873 case NFT_MSG_DELRULE:
8874 trans->ctx.chain->use++;
8875 nft_clear(trans->ctx.net, nft_trans_rule(trans));
8876 nft_rule_expr_activate(&trans->ctx, nft_trans_rule(trans));
8877 if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
8878 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
8879
8880 nft_trans_destroy(trans);
8881 break;
8882 case NFT_MSG_NEWSET:
8883 trans->ctx.table->use--;
8884 if (nft_trans_set_bound(trans)) {
8885 nft_trans_destroy(trans);
8886 break;
8887 }
8888 list_del_rcu(&nft_trans_set(trans)->list);
8889 break;
8890 case NFT_MSG_DELSET:
8891 trans->ctx.table->use++;
8892 nft_clear(trans->ctx.net, nft_trans_set(trans));
8893 nft_trans_destroy(trans);
8894 break;
8895 case NFT_MSG_NEWSETELEM:
8896 if (nft_trans_elem_set_bound(trans)) {
8897 nft_trans_destroy(trans);
8898 break;
8899 }
8900 te = (struct nft_trans_elem *)trans->data;
8901 nft_setelem_remove(net, te->set, &te->elem);
8902 if (!nft_setelem_is_catchall(te->set, &te->elem))
8903 atomic_dec(&te->set->nelems);
8904 break;
8905 case NFT_MSG_DELSETELEM:
8906 te = (struct nft_trans_elem *)trans->data;
8907
8908 nft_setelem_data_activate(net, te->set, &te->elem);
8909 nft_setelem_activate(net, te->set, &te->elem);
8910 if (!nft_setelem_is_catchall(te->set, &te->elem))
8911 te->set->ndeact--;
8912
8913 nft_trans_destroy(trans);
8914 break;
8915 case NFT_MSG_NEWOBJ:
8916 if (nft_trans_obj_update(trans)) {
8917 kfree(nft_trans_obj_newobj(trans));
8918 nft_trans_destroy(trans);
8919 } else {
8920 trans->ctx.table->use--;
8921 nft_obj_del(nft_trans_obj(trans));
8922 }
8923 break;
8924 case NFT_MSG_DELOBJ:
8925 trans->ctx.table->use++;
8926 nft_clear(trans->ctx.net, nft_trans_obj(trans));
8927 nft_trans_destroy(trans);
8928 break;
8929 case NFT_MSG_NEWFLOWTABLE:
8930 if (nft_trans_flowtable_update(trans)) {
8931 nft_unregister_flowtable_net_hooks(net,
8932 &nft_trans_flowtable_hooks(trans));
8933 } else {
8934 trans->ctx.table->use--;
8935 list_del_rcu(&nft_trans_flowtable(trans)->list);
8936 nft_unregister_flowtable_net_hooks(net,
8937 &nft_trans_flowtable(trans)->hook_list);
8938 }
8939 break;
8940 case NFT_MSG_DELFLOWTABLE:
8941 if (nft_trans_flowtable_update(trans)) {
8942 list_for_each_entry(hook, &nft_trans_flowtable(trans)->hook_list, list)
8943 hook->inactive = false;
8944 } else {
8945 trans->ctx.table->use++;
8946 nft_clear(trans->ctx.net, nft_trans_flowtable(trans));
8947 }
8948 nft_trans_destroy(trans);
8949 break;
8950 }
8951 }
8952
8953 synchronize_rcu();
8954
8955 list_for_each_entry_safe_reverse(trans, next,
8956 &nft_net->commit_list, list) {
8957 list_del(&trans->list);
8958 nf_tables_abort_release(trans);
8959 }
8960
8961 if (action == NFNL_ABORT_AUTOLOAD)
8962 nf_tables_module_autoload(net);
8963 else
8964 nf_tables_module_autoload_cleanup(net);
8965
8966 return 0;
8967 }
8968
8969 static void nf_tables_cleanup(struct net *net)
8970 {
8971 nft_validate_state_update(net, NFT_VALIDATE_SKIP);
8972 }
8973
8974 static int nf_tables_abort(struct net *net, struct sk_buff *skb,
8975 enum nfnl_abort_action action)
8976 {
8977 struct nftables_pernet *nft_net = nft_pernet(net);
8978 int ret = __nf_tables_abort(net, action);
8979
8980 mutex_unlock(&nft_net->commit_mutex);
8981
8982 return ret;
8983 }
8984
8985 static bool nf_tables_valid_genid(struct net *net, u32 genid)
8986 {
8987 struct nftables_pernet *nft_net = nft_pernet(net);
8988 bool genid_ok;
8989
8990 mutex_lock(&nft_net->commit_mutex);
8991
8992 genid_ok = genid == 0 || nft_net->base_seq == genid;
8993 if (!genid_ok)
8994 mutex_unlock(&nft_net->commit_mutex);
8995
8996 /* else, commit mutex has to be released by commit or abort function */
8997 return genid_ok;
8998 }
8999
9000 static const struct nfnetlink_subsystem nf_tables_subsys = {
9001 .name = "nf_tables",
9002 .subsys_id = NFNL_SUBSYS_NFTABLES,
9003 .cb_count = NFT_MSG_MAX,
9004 .cb = nf_tables_cb,
9005 .commit = nf_tables_commit,
9006 .abort = nf_tables_abort,
9007 .cleanup = nf_tables_cleanup,
9008 .valid_genid = nf_tables_valid_genid,
9009 .owner = THIS_MODULE,
9010 };
9011
9012 int nft_chain_validate_dependency(const struct nft_chain *chain,
9013 enum nft_chain_types type)
9014 {
9015 const struct nft_base_chain *basechain;
9016
9017 if (nft_is_base_chain(chain)) {
9018 basechain = nft_base_chain(chain);
9019 if (basechain->type->type != type)
9020 return -EOPNOTSUPP;
9021 }
9022 return 0;
9023 }
9024 EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
9025
9026 int nft_chain_validate_hooks(const struct nft_chain *chain,
9027 unsigned int hook_flags)
9028 {
9029 struct nft_base_chain *basechain;
9030
9031 if (nft_is_base_chain(chain)) {
9032 basechain = nft_base_chain(chain);
9033
9034 if ((1 << basechain->ops.hooknum) & hook_flags)
9035 return 0;
9036
9037 return -EOPNOTSUPP;
9038 }
9039
9040 return 0;
9041 }
9042 EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
9043
9044 /*
9045 * Loop detection - walk through the ruleset beginning at the destination chain
9046 * of a new jump until either the source chain is reached (loop) or all
9047 * reachable chains have been traversed.
9048 *
9049 * The loop check is performed whenever a new jump verdict is added to an
9050 * expression or verdict map or a verdict map is bound to a new chain.
9051 */
9052
9053 static int nf_tables_check_loops(const struct nft_ctx *ctx,
9054 const struct nft_chain *chain);
9055
9056 static int nft_check_loops(const struct nft_ctx *ctx,
9057 const struct nft_set_ext *ext)
9058 {
9059 const struct nft_data *data;
9060 int ret;
9061
9062 data = nft_set_ext_data(ext);
9063 switch (data->verdict.code) {
9064 case NFT_JUMP:
9065 case NFT_GOTO:
9066 ret = nf_tables_check_loops(ctx, data->verdict.chain);
9067 break;
9068 default:
9069 ret = 0;
9070 break;
9071 }
9072
9073 return ret;
9074 }
9075
9076 static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
9077 struct nft_set *set,
9078 const struct nft_set_iter *iter,
9079 struct nft_set_elem *elem)
9080 {
9081 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
9082
9083 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
9084 *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
9085 return 0;
9086
9087 return nft_check_loops(ctx, ext);
9088 }
9089
9090 static int nft_set_catchall_loops(const struct nft_ctx *ctx,
9091 struct nft_set *set)
9092 {
9093 u8 genmask = nft_genmask_next(ctx->net);
9094 struct nft_set_elem_catchall *catchall;
9095 struct nft_set_ext *ext;
9096 int ret = 0;
9097
9098 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
9099 ext = nft_set_elem_ext(set, catchall->elem);
9100 if (!nft_set_elem_active(ext, genmask))
9101 continue;
9102
9103 ret = nft_check_loops(ctx, ext);
9104 if (ret < 0)
9105 return ret;
9106 }
9107
9108 return ret;
9109 }
9110
9111 static int nf_tables_check_loops(const struct nft_ctx *ctx,
9112 const struct nft_chain *chain)
9113 {
9114 const struct nft_rule *rule;
9115 const struct nft_expr *expr, *last;
9116 struct nft_set *set;
9117 struct nft_set_binding *binding;
9118 struct nft_set_iter iter;
9119
9120 if (ctx->chain == chain)
9121 return -ELOOP;
9122
9123 list_for_each_entry(rule, &chain->rules, list) {
9124 nft_rule_for_each_expr(expr, last, rule) {
9125 struct nft_immediate_expr *priv;
9126 const struct nft_data *data;
9127 int err;
9128
9129 if (strcmp(expr->ops->type->name, "immediate"))
9130 continue;
9131
9132 priv = nft_expr_priv(expr);
9133 if (priv->dreg != NFT_REG_VERDICT)
9134 continue;
9135
9136 data = &priv->data;
9137 switch (data->verdict.code) {
9138 case NFT_JUMP:
9139 case NFT_GOTO:
9140 err = nf_tables_check_loops(ctx,
9141 data->verdict.chain);
9142 if (err < 0)
9143 return err;
9144 break;
9145 default:
9146 break;
9147 }
9148 }
9149 }
9150
9151 list_for_each_entry(set, &ctx->table->sets, list) {
9152 if (!nft_is_active_next(ctx->net, set))
9153 continue;
9154 if (!(set->flags & NFT_SET_MAP) ||
9155 set->dtype != NFT_DATA_VERDICT)
9156 continue;
9157
9158 list_for_each_entry(binding, &set->bindings, list) {
9159 if (!(binding->flags & NFT_SET_MAP) ||
9160 binding->chain != chain)
9161 continue;
9162
9163 iter.genmask = nft_genmask_next(ctx->net);
9164 iter.skip = 0;
9165 iter.count = 0;
9166 iter.err = 0;
9167 iter.fn = nf_tables_loop_check_setelem;
9168
9169 set->ops->walk(ctx, set, &iter);
9170 if (!iter.err)
9171 iter.err = nft_set_catchall_loops(ctx, set);
9172
9173 if (iter.err < 0)
9174 return iter.err;
9175 }
9176 }
9177
9178 return 0;
9179 }
9180
9181 /**
9182 * nft_parse_u32_check - fetch u32 attribute and check for maximum value
9183 *
9184 * @attr: netlink attribute to fetch value from
9185 * @max: maximum value to be stored in dest
9186 * @dest: pointer to the variable
9187 *
9188 * Parse, check and store a given u32 netlink attribute into variable.
9189 * This function returns -ERANGE if the value goes over maximum value.
9190 * Otherwise a 0 is returned and the attribute value is stored in the
9191 * destination variable.
9192 */
9193 int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
9194 {
9195 u32 val;
9196
9197 val = ntohl(nla_get_be32(attr));
9198 if (val > max)
9199 return -ERANGE;
9200
9201 *dest = val;
9202 return 0;
9203 }
9204 EXPORT_SYMBOL_GPL(nft_parse_u32_check);
9205
9206 static unsigned int nft_parse_register(const struct nlattr *attr)
9207 {
9208 unsigned int reg;
9209
9210 reg = ntohl(nla_get_be32(attr));
9211 switch (reg) {
9212 case NFT_REG_VERDICT...NFT_REG_4:
9213 return reg * NFT_REG_SIZE / NFT_REG32_SIZE;
9214 default:
9215 return reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
9216 }
9217 }
9218
9219 /**
9220 * nft_dump_register - dump a register value to a netlink attribute
9221 *
9222 * @skb: socket buffer
9223 * @attr: attribute number
9224 * @reg: register number
9225 *
9226 * Construct a netlink attribute containing the register number. For
9227 * compatibility reasons, register numbers being a multiple of 4 are
9228 * translated to the corresponding 128 bit register numbers.
9229 */
9230 int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
9231 {
9232 if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
9233 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
9234 else
9235 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
9236
9237 return nla_put_be32(skb, attr, htonl(reg));
9238 }
9239 EXPORT_SYMBOL_GPL(nft_dump_register);
9240
9241 static int nft_validate_register_load(enum nft_registers reg, unsigned int len)
9242 {
9243 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9244 return -EINVAL;
9245 if (len == 0)
9246 return -EINVAL;
9247 if (reg * NFT_REG32_SIZE + len > sizeof_field(struct nft_regs, data))
9248 return -ERANGE;
9249
9250 return 0;
9251 }
9252
9253 int nft_parse_register_load(const struct nlattr *attr, u8 *sreg, u32 len)
9254 {
9255 u32 reg;
9256 int err;
9257
9258 reg = nft_parse_register(attr);
9259 err = nft_validate_register_load(reg, len);
9260 if (err < 0)
9261 return err;
9262
9263 *sreg = reg;
9264 return 0;
9265 }
9266 EXPORT_SYMBOL_GPL(nft_parse_register_load);
9267
9268 static int nft_validate_register_store(const struct nft_ctx *ctx,
9269 enum nft_registers reg,
9270 const struct nft_data *data,
9271 enum nft_data_types type,
9272 unsigned int len)
9273 {
9274 int err;
9275
9276 switch (reg) {
9277 case NFT_REG_VERDICT:
9278 if (type != NFT_DATA_VERDICT)
9279 return -EINVAL;
9280
9281 if (data != NULL &&
9282 (data->verdict.code == NFT_GOTO ||
9283 data->verdict.code == NFT_JUMP)) {
9284 err = nf_tables_check_loops(ctx, data->verdict.chain);
9285 if (err < 0)
9286 return err;
9287 }
9288
9289 return 0;
9290 default:
9291 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9292 return -EINVAL;
9293 if (len == 0)
9294 return -EINVAL;
9295 if (reg * NFT_REG32_SIZE + len >
9296 sizeof_field(struct nft_regs, data))
9297 return -ERANGE;
9298
9299 if (data != NULL && type != NFT_DATA_VALUE)
9300 return -EINVAL;
9301 return 0;
9302 }
9303 }
9304
9305 int nft_parse_register_store(const struct nft_ctx *ctx,
9306 const struct nlattr *attr, u8 *dreg,
9307 const struct nft_data *data,
9308 enum nft_data_types type, unsigned int len)
9309 {
9310 int err;
9311 u32 reg;
9312
9313 reg = nft_parse_register(attr);
9314 err = nft_validate_register_store(ctx, reg, data, type, len);
9315 if (err < 0)
9316 return err;
9317
9318 *dreg = reg;
9319 return 0;
9320 }
9321 EXPORT_SYMBOL_GPL(nft_parse_register_store);
9322
9323 static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
9324 [NFTA_VERDICT_CODE] = { .type = NLA_U32 },
9325 [NFTA_VERDICT_CHAIN] = { .type = NLA_STRING,
9326 .len = NFT_CHAIN_MAXNAMELEN - 1 },
9327 [NFTA_VERDICT_CHAIN_ID] = { .type = NLA_U32 },
9328 };
9329
9330 static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
9331 struct nft_data_desc *desc, const struct nlattr *nla)
9332 {
9333 u8 genmask = nft_genmask_next(ctx->net);
9334 struct nlattr *tb[NFTA_VERDICT_MAX + 1];
9335 struct nft_chain *chain;
9336 int err;
9337
9338 err = nla_parse_nested_deprecated(tb, NFTA_VERDICT_MAX, nla,
9339 nft_verdict_policy, NULL);
9340 if (err < 0)
9341 return err;
9342
9343 if (!tb[NFTA_VERDICT_CODE])
9344 return -EINVAL;
9345 data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
9346
9347 switch (data->verdict.code) {
9348 default:
9349 switch (data->verdict.code & NF_VERDICT_MASK) {
9350 case NF_ACCEPT:
9351 case NF_DROP:
9352 case NF_QUEUE:
9353 break;
9354 default:
9355 return -EINVAL;
9356 }
9357 fallthrough;
9358 case NFT_CONTINUE:
9359 case NFT_BREAK:
9360 case NFT_RETURN:
9361 break;
9362 case NFT_JUMP:
9363 case NFT_GOTO:
9364 if (tb[NFTA_VERDICT_CHAIN]) {
9365 chain = nft_chain_lookup(ctx->net, ctx->table,
9366 tb[NFTA_VERDICT_CHAIN],
9367 genmask);
9368 } else if (tb[NFTA_VERDICT_CHAIN_ID]) {
9369 chain = nft_chain_lookup_byid(ctx->net,
9370 tb[NFTA_VERDICT_CHAIN_ID]);
9371 if (IS_ERR(chain))
9372 return PTR_ERR(chain);
9373 } else {
9374 return -EINVAL;
9375 }
9376
9377 if (IS_ERR(chain))
9378 return PTR_ERR(chain);
9379 if (nft_is_base_chain(chain))
9380 return -EOPNOTSUPP;
9381
9382 chain->use++;
9383 data->verdict.chain = chain;
9384 break;
9385 }
9386
9387 desc->len = sizeof(data->verdict);
9388 desc->type = NFT_DATA_VERDICT;
9389 return 0;
9390 }
9391
9392 static void nft_verdict_uninit(const struct nft_data *data)
9393 {
9394 struct nft_chain *chain;
9395 struct nft_rule *rule;
9396
9397 switch (data->verdict.code) {
9398 case NFT_JUMP:
9399 case NFT_GOTO:
9400 chain = data->verdict.chain;
9401 chain->use--;
9402
9403 if (!nft_chain_is_bound(chain))
9404 break;
9405
9406 chain->table->use--;
9407 list_for_each_entry(rule, &chain->rules, list)
9408 chain->use--;
9409
9410 nft_chain_del(chain);
9411 break;
9412 }
9413 }
9414
9415 int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
9416 {
9417 struct nlattr *nest;
9418
9419 nest = nla_nest_start_noflag(skb, type);
9420 if (!nest)
9421 goto nla_put_failure;
9422
9423 if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
9424 goto nla_put_failure;
9425
9426 switch (v->code) {
9427 case NFT_JUMP:
9428 case NFT_GOTO:
9429 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
9430 v->chain->name))
9431 goto nla_put_failure;
9432 }
9433 nla_nest_end(skb, nest);
9434 return 0;
9435
9436 nla_put_failure:
9437 return -1;
9438 }
9439
9440 static int nft_value_init(const struct nft_ctx *ctx,
9441 struct nft_data *data, unsigned int size,
9442 struct nft_data_desc *desc, const struct nlattr *nla)
9443 {
9444 unsigned int len;
9445
9446 len = nla_len(nla);
9447 if (len == 0)
9448 return -EINVAL;
9449 if (len > size)
9450 return -EOVERFLOW;
9451
9452 nla_memcpy(data->data, nla, len);
9453 desc->type = NFT_DATA_VALUE;
9454 desc->len = len;
9455 return 0;
9456 }
9457
9458 static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
9459 unsigned int len)
9460 {
9461 return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
9462 }
9463
9464 static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
9465 [NFTA_DATA_VALUE] = { .type = NLA_BINARY },
9466 [NFTA_DATA_VERDICT] = { .type = NLA_NESTED },
9467 };
9468
9469 /**
9470 * nft_data_init - parse nf_tables data netlink attributes
9471 *
9472 * @ctx: context of the expression using the data
9473 * @data: destination struct nft_data
9474 * @size: maximum data length
9475 * @desc: data description
9476 * @nla: netlink attribute containing data
9477 *
9478 * Parse the netlink data attributes and initialize a struct nft_data.
9479 * The type and length of data are returned in the data description.
9480 *
9481 * The caller can indicate that it only wants to accept data of type
9482 * NFT_DATA_VALUE by passing NULL for the ctx argument.
9483 */
9484 int nft_data_init(const struct nft_ctx *ctx,
9485 struct nft_data *data, unsigned int size,
9486 struct nft_data_desc *desc, const struct nlattr *nla)
9487 {
9488 struct nlattr *tb[NFTA_DATA_MAX + 1];
9489 int err;
9490
9491 err = nla_parse_nested_deprecated(tb, NFTA_DATA_MAX, nla,
9492 nft_data_policy, NULL);
9493 if (err < 0)
9494 return err;
9495
9496 if (tb[NFTA_DATA_VALUE])
9497 return nft_value_init(ctx, data, size, desc,
9498 tb[NFTA_DATA_VALUE]);
9499 if (tb[NFTA_DATA_VERDICT] && ctx != NULL)
9500 return nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
9501 return -EINVAL;
9502 }
9503 EXPORT_SYMBOL_GPL(nft_data_init);
9504
9505 /**
9506 * nft_data_release - release a nft_data item
9507 *
9508 * @data: struct nft_data to release
9509 * @type: type of data
9510 *
9511 * Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
9512 * all others need to be released by calling this function.
9513 */
9514 void nft_data_release(const struct nft_data *data, enum nft_data_types type)
9515 {
9516 if (type < NFT_DATA_VERDICT)
9517 return;
9518 switch (type) {
9519 case NFT_DATA_VERDICT:
9520 return nft_verdict_uninit(data);
9521 default:
9522 WARN_ON(1);
9523 }
9524 }
9525 EXPORT_SYMBOL_GPL(nft_data_release);
9526
9527 int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
9528 enum nft_data_types type, unsigned int len)
9529 {
9530 struct nlattr *nest;
9531 int err;
9532
9533 nest = nla_nest_start_noflag(skb, attr);
9534 if (nest == NULL)
9535 return -1;
9536
9537 switch (type) {
9538 case NFT_DATA_VALUE:
9539 err = nft_value_dump(skb, data, len);
9540 break;
9541 case NFT_DATA_VERDICT:
9542 err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
9543 break;
9544 default:
9545 err = -EINVAL;
9546 WARN_ON(1);
9547 }
9548
9549 nla_nest_end(skb, nest);
9550 return err;
9551 }
9552 EXPORT_SYMBOL_GPL(nft_data_dump);
9553
9554 int __nft_release_basechain(struct nft_ctx *ctx)
9555 {
9556 struct nft_rule *rule, *nr;
9557
9558 if (WARN_ON(!nft_is_base_chain(ctx->chain)))
9559 return 0;
9560
9561 nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
9562 list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
9563 list_del(&rule->list);
9564 ctx->chain->use--;
9565 nf_tables_rule_release(ctx, rule);
9566 }
9567 nft_chain_del(ctx->chain);
9568 ctx->table->use--;
9569 nf_tables_chain_destroy(ctx);
9570
9571 return 0;
9572 }
9573 EXPORT_SYMBOL_GPL(__nft_release_basechain);
9574
9575 static void __nft_release_hook(struct net *net, struct nft_table *table)
9576 {
9577 struct nft_chain *chain;
9578
9579 list_for_each_entry(chain, &table->chains, list)
9580 nf_tables_unregister_hook(net, table, chain);
9581 }
9582
9583 static void __nft_release_hooks(struct net *net)
9584 {
9585 struct nftables_pernet *nft_net = nft_pernet(net);
9586 struct nft_table *table;
9587
9588 list_for_each_entry(table, &nft_net->tables, list) {
9589 if (nft_table_has_owner(table))
9590 continue;
9591
9592 __nft_release_hook(net, table);
9593 }
9594 }
9595
9596 static void __nft_release_table(struct net *net, struct nft_table *table)
9597 {
9598 struct nft_flowtable *flowtable, *nf;
9599 struct nft_chain *chain, *nc;
9600 struct nft_object *obj, *ne;
9601 struct nft_rule *rule, *nr;
9602 struct nft_set *set, *ns;
9603 struct nft_ctx ctx = {
9604 .net = net,
9605 .family = NFPROTO_NETDEV,
9606 };
9607
9608 ctx.family = table->family;
9609 ctx.table = table;
9610 list_for_each_entry(chain, &table->chains, list) {
9611 ctx.chain = chain;
9612 list_for_each_entry_safe(rule, nr, &chain->rules, list) {
9613 list_del(&rule->list);
9614 chain->use--;
9615 nf_tables_rule_release(&ctx, rule);
9616 }
9617 }
9618 list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
9619 list_del(&flowtable->list);
9620 table->use--;
9621 nf_tables_flowtable_destroy(flowtable);
9622 }
9623 list_for_each_entry_safe(set, ns, &table->sets, list) {
9624 list_del(&set->list);
9625 table->use--;
9626 nft_set_destroy(&ctx, set);
9627 }
9628 list_for_each_entry_safe(obj, ne, &table->objects, list) {
9629 nft_obj_del(obj);
9630 table->use--;
9631 nft_obj_destroy(&ctx, obj);
9632 }
9633 list_for_each_entry_safe(chain, nc, &table->chains, list) {
9634 ctx.chain = chain;
9635 nft_chain_del(chain);
9636 table->use--;
9637 nf_tables_chain_destroy(&ctx);
9638 }
9639 nf_tables_table_destroy(&ctx);
9640 }
9641
9642 static void __nft_release_tables(struct net *net)
9643 {
9644 struct nftables_pernet *nft_net = nft_pernet(net);
9645 struct nft_table *table, *nt;
9646
9647 list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
9648 if (nft_table_has_owner(table))
9649 continue;
9650
9651 list_del(&table->list);
9652
9653 __nft_release_table(net, table);
9654 }
9655 }
9656
9657 static int nft_rcv_nl_event(struct notifier_block *this, unsigned long event,
9658 void *ptr)
9659 {
9660 struct nft_table *table, *to_delete[8];
9661 struct nftables_pernet *nft_net;
9662 struct netlink_notify *n = ptr;
9663 struct net *net = n->net;
9664 unsigned int deleted;
9665 bool restart = false;
9666
9667 if (event != NETLINK_URELEASE || n->protocol != NETLINK_NETFILTER)
9668 return NOTIFY_DONE;
9669
9670 nft_net = nft_pernet(net);
9671 deleted = 0;
9672 mutex_lock(&nft_net->commit_mutex);
9673 again:
9674 list_for_each_entry(table, &nft_net->tables, list) {
9675 if (nft_table_has_owner(table) &&
9676 n->portid == table->nlpid) {
9677 __nft_release_hook(net, table);
9678 list_del_rcu(&table->list);
9679 to_delete[deleted++] = table;
9680 if (deleted >= ARRAY_SIZE(to_delete))
9681 break;
9682 }
9683 }
9684 if (deleted) {
9685 restart = deleted >= ARRAY_SIZE(to_delete);
9686 synchronize_rcu();
9687 while (deleted)
9688 __nft_release_table(net, to_delete[--deleted]);
9689
9690 if (restart)
9691 goto again;
9692 }
9693 mutex_unlock(&nft_net->commit_mutex);
9694
9695 return NOTIFY_DONE;
9696 }
9697
9698 static struct notifier_block nft_nl_notifier = {
9699 .notifier_call = nft_rcv_nl_event,
9700 };
9701
9702 static int __net_init nf_tables_init_net(struct net *net)
9703 {
9704 struct nftables_pernet *nft_net = nft_pernet(net);
9705
9706 INIT_LIST_HEAD(&nft_net->tables);
9707 INIT_LIST_HEAD(&nft_net->commit_list);
9708 INIT_LIST_HEAD(&nft_net->module_list);
9709 INIT_LIST_HEAD(&nft_net->notify_list);
9710 mutex_init(&nft_net->commit_mutex);
9711 nft_net->base_seq = 1;
9712 nft_net->validate_state = NFT_VALIDATE_SKIP;
9713
9714 return 0;
9715 }
9716
9717 static void __net_exit nf_tables_pre_exit_net(struct net *net)
9718 {
9719 __nft_release_hooks(net);
9720 }
9721
9722 static void __net_exit nf_tables_exit_net(struct net *net)
9723 {
9724 struct nftables_pernet *nft_net = nft_pernet(net);
9725
9726 mutex_lock(&nft_net->commit_mutex);
9727 if (!list_empty(&nft_net->commit_list))
9728 __nf_tables_abort(net, NFNL_ABORT_NONE);
9729 __nft_release_tables(net);
9730 mutex_unlock(&nft_net->commit_mutex);
9731 WARN_ON_ONCE(!list_empty(&nft_net->tables));
9732 WARN_ON_ONCE(!list_empty(&nft_net->module_list));
9733 WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
9734 }
9735
9736 static struct pernet_operations nf_tables_net_ops = {
9737 .init = nf_tables_init_net,
9738 .pre_exit = nf_tables_pre_exit_net,
9739 .exit = nf_tables_exit_net,
9740 .id = &nf_tables_net_id,
9741 .size = sizeof(struct nftables_pernet),
9742 };
9743
9744 static int __init nf_tables_module_init(void)
9745 {
9746 int err;
9747
9748 err = register_pernet_subsys(&nf_tables_net_ops);
9749 if (err < 0)
9750 return err;
9751
9752 err = nft_chain_filter_init();
9753 if (err < 0)
9754 goto err_chain_filter;
9755
9756 err = nf_tables_core_module_init();
9757 if (err < 0)
9758 goto err_core_module;
9759
9760 err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
9761 if (err < 0)
9762 goto err_netdev_notifier;
9763
9764 err = rhltable_init(&nft_objname_ht, &nft_objname_ht_params);
9765 if (err < 0)
9766 goto err_rht_objname;
9767
9768 err = nft_offload_init();
9769 if (err < 0)
9770 goto err_offload;
9771
9772 err = netlink_register_notifier(&nft_nl_notifier);
9773 if (err < 0)
9774 goto err_netlink_notifier;
9775
9776 /* must be last */
9777 err = nfnetlink_subsys_register(&nf_tables_subsys);
9778 if (err < 0)
9779 goto err_nfnl_subsys;
9780
9781 nft_chain_route_init();
9782
9783 return err;
9784
9785 err_nfnl_subsys:
9786 netlink_unregister_notifier(&nft_nl_notifier);
9787 err_netlink_notifier:
9788 nft_offload_exit();
9789 err_offload:
9790 rhltable_destroy(&nft_objname_ht);
9791 err_rht_objname:
9792 unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
9793 err_netdev_notifier:
9794 nf_tables_core_module_exit();
9795 err_core_module:
9796 nft_chain_filter_fini();
9797 err_chain_filter:
9798 unregister_pernet_subsys(&nf_tables_net_ops);
9799 return err;
9800 }
9801
9802 static void __exit nf_tables_module_exit(void)
9803 {
9804 nfnetlink_subsys_unregister(&nf_tables_subsys);
9805 netlink_unregister_notifier(&nft_nl_notifier);
9806 nft_offload_exit();
9807 unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
9808 nft_chain_filter_fini();
9809 nft_chain_route_fini();
9810 unregister_pernet_subsys(&nf_tables_net_ops);
9811 cancel_work_sync(&trans_destroy_work);
9812 rcu_barrier();
9813 rhltable_destroy(&nft_objname_ht);
9814 nf_tables_core_module_exit();
9815 }
9816
9817 module_init(nf_tables_module_init);
9818 module_exit(nf_tables_module_exit);
9819
9820 MODULE_LICENSE("GPL");
9821 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
9822 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);