]> git.proxmox.com Git - mirror_ubuntu-jammy-kernel.git/blob - net/netfilter/nf_tables_api.c
91865dc2bcbb89809000d4151e02a273b01be3ba
[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 err_expr_parse;
2782
2783 err = -EOPNOTSUPP;
2784 if (!(expr_info.ops->type->flags & NFT_EXPR_STATEFUL))
2785 goto err_expr_stateful;
2786
2787 err = -ENOMEM;
2788 expr = kzalloc(expr_info.ops->size, GFP_KERNEL);
2789 if (expr == NULL)
2790 goto err_expr_stateful;
2791
2792 err = nf_tables_newexpr(ctx, &expr_info, expr);
2793 if (err < 0)
2794 goto err_expr_new;
2795
2796 return expr;
2797 err_expr_new:
2798 kfree(expr);
2799 err_expr_stateful:
2800 owner = expr_info.ops->type->owner;
2801 if (expr_info.ops->type->release_ops)
2802 expr_info.ops->type->release_ops(expr_info.ops);
2803
2804 module_put(owner);
2805 err_expr_parse:
2806 return ERR_PTR(err);
2807 }
2808
2809 int nft_expr_clone(struct nft_expr *dst, struct nft_expr *src)
2810 {
2811 int err;
2812
2813 if (src->ops->clone) {
2814 dst->ops = src->ops;
2815 err = src->ops->clone(dst, src);
2816 if (err < 0)
2817 return err;
2818 } else {
2819 memcpy(dst, src, src->ops->size);
2820 }
2821
2822 __module_get(src->ops->type->owner);
2823
2824 return 0;
2825 }
2826
2827 void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
2828 {
2829 nf_tables_expr_destroy(ctx, expr);
2830 kfree(expr);
2831 }
2832
2833 /*
2834 * Rules
2835 */
2836
2837 static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
2838 u64 handle)
2839 {
2840 struct nft_rule *rule;
2841
2842 // FIXME: this sucks
2843 list_for_each_entry_rcu(rule, &chain->rules, list) {
2844 if (handle == rule->handle)
2845 return rule;
2846 }
2847
2848 return ERR_PTR(-ENOENT);
2849 }
2850
2851 static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
2852 const struct nlattr *nla)
2853 {
2854 if (nla == NULL)
2855 return ERR_PTR(-EINVAL);
2856
2857 return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
2858 }
2859
2860 static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
2861 [NFTA_RULE_TABLE] = { .type = NLA_STRING,
2862 .len = NFT_TABLE_MAXNAMELEN - 1 },
2863 [NFTA_RULE_CHAIN] = { .type = NLA_STRING,
2864 .len = NFT_CHAIN_MAXNAMELEN - 1 },
2865 [NFTA_RULE_HANDLE] = { .type = NLA_U64 },
2866 [NFTA_RULE_EXPRESSIONS] = { .type = NLA_NESTED },
2867 [NFTA_RULE_COMPAT] = { .type = NLA_NESTED },
2868 [NFTA_RULE_POSITION] = { .type = NLA_U64 },
2869 [NFTA_RULE_USERDATA] = { .type = NLA_BINARY,
2870 .len = NFT_USERDATA_MAXLEN },
2871 [NFTA_RULE_ID] = { .type = NLA_U32 },
2872 [NFTA_RULE_POSITION_ID] = { .type = NLA_U32 },
2873 [NFTA_RULE_CHAIN_ID] = { .type = NLA_U32 },
2874 };
2875
2876 static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
2877 u32 portid, u32 seq, int event,
2878 u32 flags, int family,
2879 const struct nft_table *table,
2880 const struct nft_chain *chain,
2881 const struct nft_rule *rule, u64 handle)
2882 {
2883 struct nlmsghdr *nlh;
2884 const struct nft_expr *expr, *next;
2885 struct nlattr *list;
2886 u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
2887
2888 nlh = nfnl_msg_put(skb, portid, seq, type, flags, family, NFNETLINK_V0,
2889 nft_base_seq(net));
2890 if (!nlh)
2891 goto nla_put_failure;
2892
2893 if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
2894 goto nla_put_failure;
2895 if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
2896 goto nla_put_failure;
2897 if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
2898 NFTA_RULE_PAD))
2899 goto nla_put_failure;
2900
2901 if (event != NFT_MSG_DELRULE && handle) {
2902 if (nla_put_be64(skb, NFTA_RULE_POSITION, cpu_to_be64(handle),
2903 NFTA_RULE_PAD))
2904 goto nla_put_failure;
2905 }
2906
2907 if (chain->flags & NFT_CHAIN_HW_OFFLOAD)
2908 nft_flow_rule_stats(chain, rule);
2909
2910 list = nla_nest_start_noflag(skb, NFTA_RULE_EXPRESSIONS);
2911 if (list == NULL)
2912 goto nla_put_failure;
2913 nft_rule_for_each_expr(expr, next, rule) {
2914 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
2915 goto nla_put_failure;
2916 }
2917 nla_nest_end(skb, list);
2918
2919 if (rule->udata) {
2920 struct nft_userdata *udata = nft_userdata(rule);
2921 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
2922 udata->data) < 0)
2923 goto nla_put_failure;
2924 }
2925
2926 nlmsg_end(skb, nlh);
2927 return 0;
2928
2929 nla_put_failure:
2930 nlmsg_trim(skb, nlh);
2931 return -1;
2932 }
2933
2934 static void nf_tables_rule_notify(const struct nft_ctx *ctx,
2935 const struct nft_rule *rule, int event)
2936 {
2937 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
2938 const struct nft_rule *prule;
2939 struct sk_buff *skb;
2940 u64 handle = 0;
2941 u16 flags = 0;
2942 int err;
2943
2944 if (!ctx->report &&
2945 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
2946 return;
2947
2948 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
2949 if (skb == NULL)
2950 goto err;
2951
2952 if (event == NFT_MSG_NEWRULE &&
2953 !list_is_first(&rule->list, &ctx->chain->rules) &&
2954 !list_is_last(&rule->list, &ctx->chain->rules)) {
2955 prule = list_prev_entry(rule, list);
2956 handle = prule->handle;
2957 }
2958 if (ctx->flags & (NLM_F_APPEND | NLM_F_REPLACE))
2959 flags |= NLM_F_APPEND;
2960 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
2961 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
2962
2963 err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
2964 event, flags, ctx->family, ctx->table,
2965 ctx->chain, rule, handle);
2966 if (err < 0) {
2967 kfree_skb(skb);
2968 goto err;
2969 }
2970
2971 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
2972 return;
2973 err:
2974 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
2975 }
2976
2977 struct nft_rule_dump_ctx {
2978 char *table;
2979 char *chain;
2980 };
2981
2982 static int __nf_tables_dump_rules(struct sk_buff *skb,
2983 unsigned int *idx,
2984 struct netlink_callback *cb,
2985 const struct nft_table *table,
2986 const struct nft_chain *chain)
2987 {
2988 struct net *net = sock_net(skb->sk);
2989 const struct nft_rule *rule, *prule;
2990 unsigned int s_idx = cb->args[0];
2991 u64 handle;
2992
2993 prule = NULL;
2994 list_for_each_entry_rcu(rule, &chain->rules, list) {
2995 if (!nft_is_active(net, rule))
2996 goto cont_skip;
2997 if (*idx < s_idx)
2998 goto cont;
2999 if (*idx > s_idx) {
3000 memset(&cb->args[1], 0,
3001 sizeof(cb->args) - sizeof(cb->args[0]));
3002 }
3003 if (prule)
3004 handle = prule->handle;
3005 else
3006 handle = 0;
3007
3008 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
3009 cb->nlh->nlmsg_seq,
3010 NFT_MSG_NEWRULE,
3011 NLM_F_MULTI | NLM_F_APPEND,
3012 table->family,
3013 table, chain, rule, handle) < 0)
3014 return 1;
3015
3016 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
3017 cont:
3018 prule = rule;
3019 cont_skip:
3020 (*idx)++;
3021 }
3022 return 0;
3023 }
3024
3025 static int nf_tables_dump_rules(struct sk_buff *skb,
3026 struct netlink_callback *cb)
3027 {
3028 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3029 const struct nft_rule_dump_ctx *ctx = cb->data;
3030 struct nft_table *table;
3031 const struct nft_chain *chain;
3032 unsigned int idx = 0;
3033 struct net *net = sock_net(skb->sk);
3034 int family = nfmsg->nfgen_family;
3035 struct nftables_pernet *nft_net;
3036
3037 rcu_read_lock();
3038 nft_net = nft_pernet(net);
3039 cb->seq = nft_net->base_seq;
3040
3041 list_for_each_entry_rcu(table, &nft_net->tables, list) {
3042 if (family != NFPROTO_UNSPEC && family != table->family)
3043 continue;
3044
3045 if (ctx && ctx->table && strcmp(ctx->table, table->name) != 0)
3046 continue;
3047
3048 if (ctx && ctx->table && ctx->chain) {
3049 struct rhlist_head *list, *tmp;
3050
3051 list = rhltable_lookup(&table->chains_ht, ctx->chain,
3052 nft_chain_ht_params);
3053 if (!list)
3054 goto done;
3055
3056 rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
3057 if (!nft_is_active(net, chain))
3058 continue;
3059 __nf_tables_dump_rules(skb, &idx,
3060 cb, table, chain);
3061 break;
3062 }
3063 goto done;
3064 }
3065
3066 list_for_each_entry_rcu(chain, &table->chains, list) {
3067 if (__nf_tables_dump_rules(skb, &idx, cb, table, chain))
3068 goto done;
3069 }
3070
3071 if (ctx && ctx->table)
3072 break;
3073 }
3074 done:
3075 rcu_read_unlock();
3076
3077 cb->args[0] = idx;
3078 return skb->len;
3079 }
3080
3081 static int nf_tables_dump_rules_start(struct netlink_callback *cb)
3082 {
3083 const struct nlattr * const *nla = cb->data;
3084 struct nft_rule_dump_ctx *ctx = NULL;
3085
3086 if (nla[NFTA_RULE_TABLE] || nla[NFTA_RULE_CHAIN]) {
3087 ctx = kzalloc(sizeof(*ctx), GFP_ATOMIC);
3088 if (!ctx)
3089 return -ENOMEM;
3090
3091 if (nla[NFTA_RULE_TABLE]) {
3092 ctx->table = nla_strdup(nla[NFTA_RULE_TABLE],
3093 GFP_ATOMIC);
3094 if (!ctx->table) {
3095 kfree(ctx);
3096 return -ENOMEM;
3097 }
3098 }
3099 if (nla[NFTA_RULE_CHAIN]) {
3100 ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN],
3101 GFP_ATOMIC);
3102 if (!ctx->chain) {
3103 kfree(ctx->table);
3104 kfree(ctx);
3105 return -ENOMEM;
3106 }
3107 }
3108 }
3109
3110 cb->data = ctx;
3111 return 0;
3112 }
3113
3114 static int nf_tables_dump_rules_done(struct netlink_callback *cb)
3115 {
3116 struct nft_rule_dump_ctx *ctx = cb->data;
3117
3118 if (ctx) {
3119 kfree(ctx->table);
3120 kfree(ctx->chain);
3121 kfree(ctx);
3122 }
3123 return 0;
3124 }
3125
3126 /* called with rcu_read_lock held */
3127 static int nf_tables_getrule(struct sk_buff *skb, const struct nfnl_info *info,
3128 const struct nlattr * const nla[])
3129 {
3130 struct netlink_ext_ack *extack = info->extack;
3131 u8 genmask = nft_genmask_cur(info->net);
3132 u8 family = info->nfmsg->nfgen_family;
3133 const struct nft_chain *chain;
3134 const struct nft_rule *rule;
3135 struct net *net = info->net;
3136 struct nft_table *table;
3137 struct sk_buff *skb2;
3138 int err;
3139
3140 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3141 struct netlink_dump_control c = {
3142 .start= nf_tables_dump_rules_start,
3143 .dump = nf_tables_dump_rules,
3144 .done = nf_tables_dump_rules_done,
3145 .module = THIS_MODULE,
3146 .data = (void *)nla,
3147 };
3148
3149 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
3150 }
3151
3152 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask, 0);
3153 if (IS_ERR(table)) {
3154 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3155 return PTR_ERR(table);
3156 }
3157
3158 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
3159 if (IS_ERR(chain)) {
3160 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3161 return PTR_ERR(chain);
3162 }
3163
3164 rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3165 if (IS_ERR(rule)) {
3166 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3167 return PTR_ERR(rule);
3168 }
3169
3170 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3171 if (!skb2)
3172 return -ENOMEM;
3173
3174 err = nf_tables_fill_rule_info(skb2, net, NETLINK_CB(skb).portid,
3175 info->nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
3176 family, table, chain, rule, 0);
3177 if (err < 0)
3178 goto err_fill_rule_info;
3179
3180 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
3181
3182 err_fill_rule_info:
3183 kfree_skb(skb2);
3184 return err;
3185 }
3186
3187 static void nf_tables_rule_destroy(const struct nft_ctx *ctx,
3188 struct nft_rule *rule)
3189 {
3190 struct nft_expr *expr, *next;
3191
3192 /*
3193 * Careful: some expressions might not be initialized in case this
3194 * is called on error from nf_tables_newrule().
3195 */
3196 expr = nft_expr_first(rule);
3197 while (nft_expr_more(rule, expr)) {
3198 next = nft_expr_next(expr);
3199 nf_tables_expr_destroy(ctx, expr);
3200 expr = next;
3201 }
3202 kfree(rule);
3203 }
3204
3205 void nf_tables_rule_release(const struct nft_ctx *ctx, struct nft_rule *rule)
3206 {
3207 nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_RELEASE);
3208 nf_tables_rule_destroy(ctx, rule);
3209 }
3210
3211 int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
3212 {
3213 struct nft_expr *expr, *last;
3214 const struct nft_data *data;
3215 struct nft_rule *rule;
3216 int err;
3217
3218 if (ctx->level == NFT_JUMP_STACK_SIZE)
3219 return -EMLINK;
3220
3221 list_for_each_entry(rule, &chain->rules, list) {
3222 if (!nft_is_active_next(ctx->net, rule))
3223 continue;
3224
3225 nft_rule_for_each_expr(expr, last, rule) {
3226 if (!expr->ops->validate)
3227 continue;
3228
3229 err = expr->ops->validate(ctx, expr, &data);
3230 if (err < 0)
3231 return err;
3232 }
3233 }
3234
3235 return 0;
3236 }
3237 EXPORT_SYMBOL_GPL(nft_chain_validate);
3238
3239 static int nft_table_validate(struct net *net, const struct nft_table *table)
3240 {
3241 struct nft_chain *chain;
3242 struct nft_ctx ctx = {
3243 .net = net,
3244 .family = table->family,
3245 };
3246 int err;
3247
3248 list_for_each_entry(chain, &table->chains, list) {
3249 if (!nft_is_base_chain(chain))
3250 continue;
3251
3252 ctx.chain = chain;
3253 err = nft_chain_validate(&ctx, chain);
3254 if (err < 0)
3255 return err;
3256 }
3257
3258 return 0;
3259 }
3260
3261 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3262 const struct nlattr *nla);
3263
3264 #define NFT_RULE_MAXEXPRS 128
3265
3266 static int nf_tables_newrule(struct sk_buff *skb, const struct nfnl_info *info,
3267 const struct nlattr * const nla[])
3268 {
3269 struct nftables_pernet *nft_net = nft_pernet(info->net);
3270 struct netlink_ext_ack *extack = info->extack;
3271 unsigned int size, i, n, ulen = 0, usize = 0;
3272 u8 genmask = nft_genmask_next(info->net);
3273 struct nft_rule *rule, *old_rule = NULL;
3274 struct nft_expr_info *expr_info = NULL;
3275 u8 family = info->nfmsg->nfgen_family;
3276 struct nft_flow_rule *flow = NULL;
3277 struct net *net = info->net;
3278 struct nft_userdata *udata;
3279 struct nft_table *table;
3280 struct nft_chain *chain;
3281 struct nft_trans *trans;
3282 u64 handle, pos_handle;
3283 struct nft_expr *expr;
3284 struct nft_ctx ctx;
3285 struct nlattr *tmp;
3286 int err, rem;
3287
3288 lockdep_assert_held(&nft_net->commit_mutex);
3289
3290 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3291 NETLINK_CB(skb).portid);
3292 if (IS_ERR(table)) {
3293 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3294 return PTR_ERR(table);
3295 }
3296
3297 if (nla[NFTA_RULE_CHAIN]) {
3298 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3299 genmask);
3300 if (IS_ERR(chain)) {
3301 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3302 return PTR_ERR(chain);
3303 }
3304 if (nft_chain_is_bound(chain))
3305 return -EOPNOTSUPP;
3306
3307 } else if (nla[NFTA_RULE_CHAIN_ID]) {
3308 chain = nft_chain_lookup_byid(net, nla[NFTA_RULE_CHAIN_ID]);
3309 if (IS_ERR(chain)) {
3310 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN_ID]);
3311 return PTR_ERR(chain);
3312 }
3313 } else {
3314 return -EINVAL;
3315 }
3316
3317 if (nla[NFTA_RULE_HANDLE]) {
3318 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
3319 rule = __nft_rule_lookup(chain, handle);
3320 if (IS_ERR(rule)) {
3321 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3322 return PTR_ERR(rule);
3323 }
3324
3325 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
3326 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3327 return -EEXIST;
3328 }
3329 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
3330 old_rule = rule;
3331 else
3332 return -EOPNOTSUPP;
3333 } else {
3334 if (!(info->nlh->nlmsg_flags & NLM_F_CREATE) ||
3335 info->nlh->nlmsg_flags & NLM_F_REPLACE)
3336 return -EINVAL;
3337 handle = nf_tables_alloc_handle(table);
3338
3339 if (chain->use == UINT_MAX)
3340 return -EOVERFLOW;
3341
3342 if (nla[NFTA_RULE_POSITION]) {
3343 pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
3344 old_rule = __nft_rule_lookup(chain, pos_handle);
3345 if (IS_ERR(old_rule)) {
3346 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
3347 return PTR_ERR(old_rule);
3348 }
3349 } else if (nla[NFTA_RULE_POSITION_ID]) {
3350 old_rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_POSITION_ID]);
3351 if (IS_ERR(old_rule)) {
3352 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION_ID]);
3353 return PTR_ERR(old_rule);
3354 }
3355 }
3356 }
3357
3358 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3359
3360 n = 0;
3361 size = 0;
3362 if (nla[NFTA_RULE_EXPRESSIONS]) {
3363 expr_info = kvmalloc_array(NFT_RULE_MAXEXPRS,
3364 sizeof(struct nft_expr_info),
3365 GFP_KERNEL);
3366 if (!expr_info)
3367 return -ENOMEM;
3368
3369 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
3370 err = -EINVAL;
3371 if (nla_type(tmp) != NFTA_LIST_ELEM)
3372 goto err_release_expr;
3373 if (n == NFT_RULE_MAXEXPRS)
3374 goto err_release_expr;
3375 err = nf_tables_expr_parse(&ctx, tmp, &expr_info[n]);
3376 if (err < 0) {
3377 NL_SET_BAD_ATTR(extack, tmp);
3378 goto err_release_expr;
3379 }
3380 size += expr_info[n].ops->size;
3381 n++;
3382 }
3383 }
3384 /* Check for overflow of dlen field */
3385 err = -EFBIG;
3386 if (size >= 1 << 12)
3387 goto err_release_expr;
3388
3389 if (nla[NFTA_RULE_USERDATA]) {
3390 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
3391 if (ulen > 0)
3392 usize = sizeof(struct nft_userdata) + ulen;
3393 }
3394
3395 err = -ENOMEM;
3396 rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL);
3397 if (rule == NULL)
3398 goto err_release_expr;
3399
3400 nft_activate_next(net, rule);
3401
3402 rule->handle = handle;
3403 rule->dlen = size;
3404 rule->udata = ulen ? 1 : 0;
3405
3406 if (ulen) {
3407 udata = nft_userdata(rule);
3408 udata->len = ulen - 1;
3409 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
3410 }
3411
3412 expr = nft_expr_first(rule);
3413 for (i = 0; i < n; i++) {
3414 err = nf_tables_newexpr(&ctx, &expr_info[i], expr);
3415 if (err < 0) {
3416 NL_SET_BAD_ATTR(extack, expr_info[i].attr);
3417 goto err_release_rule;
3418 }
3419
3420 if (expr_info[i].ops->validate)
3421 nft_validate_state_update(net, NFT_VALIDATE_NEED);
3422
3423 expr_info[i].ops = NULL;
3424 expr = nft_expr_next(expr);
3425 }
3426
3427 if (chain->flags & NFT_CHAIN_HW_OFFLOAD) {
3428 flow = nft_flow_rule_create(net, rule);
3429 if (IS_ERR(flow)) {
3430 err = PTR_ERR(flow);
3431 goto err_release_rule;
3432 }
3433 }
3434
3435 if (info->nlh->nlmsg_flags & NLM_F_REPLACE) {
3436 err = nft_delrule(&ctx, old_rule);
3437 if (err < 0)
3438 goto err_destroy_flow_rule;
3439
3440 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3441 if (trans == NULL) {
3442 err = -ENOMEM;
3443 goto err_destroy_flow_rule;
3444 }
3445 list_add_tail_rcu(&rule->list, &old_rule->list);
3446 } else {
3447 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
3448 if (!trans) {
3449 err = -ENOMEM;
3450 goto err_destroy_flow_rule;
3451 }
3452
3453 if (info->nlh->nlmsg_flags & NLM_F_APPEND) {
3454 if (old_rule)
3455 list_add_rcu(&rule->list, &old_rule->list);
3456 else
3457 list_add_tail_rcu(&rule->list, &chain->rules);
3458 } else {
3459 if (old_rule)
3460 list_add_tail_rcu(&rule->list, &old_rule->list);
3461 else
3462 list_add_rcu(&rule->list, &chain->rules);
3463 }
3464 }
3465 kvfree(expr_info);
3466 chain->use++;
3467
3468 if (flow)
3469 nft_trans_flow_rule(trans) = flow;
3470
3471 if (nft_net->validate_state == NFT_VALIDATE_DO)
3472 return nft_table_validate(net, table);
3473
3474 return 0;
3475
3476 err_destroy_flow_rule:
3477 if (flow)
3478 nft_flow_rule_destroy(flow);
3479 err_release_rule:
3480 nf_tables_rule_release(&ctx, rule);
3481 err_release_expr:
3482 for (i = 0; i < n; i++) {
3483 if (expr_info[i].ops) {
3484 module_put(expr_info[i].ops->type->owner);
3485 if (expr_info[i].ops->type->release_ops)
3486 expr_info[i].ops->type->release_ops(expr_info[i].ops);
3487 }
3488 }
3489 kvfree(expr_info);
3490
3491 return err;
3492 }
3493
3494 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3495 const struct nlattr *nla)
3496 {
3497 struct nftables_pernet *nft_net = nft_pernet(net);
3498 u32 id = ntohl(nla_get_be32(nla));
3499 struct nft_trans *trans;
3500
3501 list_for_each_entry(trans, &nft_net->commit_list, list) {
3502 struct nft_rule *rule = nft_trans_rule(trans);
3503
3504 if (trans->msg_type == NFT_MSG_NEWRULE &&
3505 id == nft_trans_rule_id(trans))
3506 return rule;
3507 }
3508 return ERR_PTR(-ENOENT);
3509 }
3510
3511 static int nf_tables_delrule(struct sk_buff *skb, const struct nfnl_info *info,
3512 const struct nlattr * const nla[])
3513 {
3514 struct netlink_ext_ack *extack = info->extack;
3515 u8 genmask = nft_genmask_next(info->net);
3516 u8 family = info->nfmsg->nfgen_family;
3517 struct nft_chain *chain = NULL;
3518 struct net *net = info->net;
3519 struct nft_table *table;
3520 struct nft_rule *rule;
3521 struct nft_ctx ctx;
3522 int err = 0;
3523
3524 table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
3525 NETLINK_CB(skb).portid);
3526 if (IS_ERR(table)) {
3527 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3528 return PTR_ERR(table);
3529 }
3530
3531 if (nla[NFTA_RULE_CHAIN]) {
3532 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
3533 genmask);
3534 if (IS_ERR(chain)) {
3535 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3536 return PTR_ERR(chain);
3537 }
3538 if (nft_chain_is_bound(chain))
3539 return -EOPNOTSUPP;
3540 }
3541
3542 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3543
3544 if (chain) {
3545 if (nla[NFTA_RULE_HANDLE]) {
3546 rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3547 if (IS_ERR(rule)) {
3548 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3549 return PTR_ERR(rule);
3550 }
3551
3552 err = nft_delrule(&ctx, rule);
3553 } else if (nla[NFTA_RULE_ID]) {
3554 rule = nft_rule_lookup_byid(net, nla[NFTA_RULE_ID]);
3555 if (IS_ERR(rule)) {
3556 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
3557 return PTR_ERR(rule);
3558 }
3559
3560 err = nft_delrule(&ctx, rule);
3561 } else {
3562 err = nft_delrule_by_chain(&ctx);
3563 }
3564 } else {
3565 list_for_each_entry(chain, &table->chains, list) {
3566 if (!nft_is_active_next(net, chain))
3567 continue;
3568
3569 ctx.chain = chain;
3570 err = nft_delrule_by_chain(&ctx);
3571 if (err < 0)
3572 break;
3573 }
3574 }
3575
3576 return err;
3577 }
3578
3579 /*
3580 * Sets
3581 */
3582 static const struct nft_set_type *nft_set_types[] = {
3583 &nft_set_hash_fast_type,
3584 &nft_set_hash_type,
3585 &nft_set_rhash_type,
3586 &nft_set_bitmap_type,
3587 &nft_set_rbtree_type,
3588 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
3589 &nft_set_pipapo_avx2_type,
3590 #endif
3591 &nft_set_pipapo_type,
3592 };
3593
3594 #define NFT_SET_FEATURES (NFT_SET_INTERVAL | NFT_SET_MAP | \
3595 NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
3596 NFT_SET_EVAL)
3597
3598 static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
3599 {
3600 return (flags & type->features) == (flags & NFT_SET_FEATURES);
3601 }
3602
3603 /*
3604 * Select a set implementation based on the data characteristics and the
3605 * given policy. The total memory use might not be known if no size is
3606 * given, in that case the amount of memory per element is used.
3607 */
3608 static const struct nft_set_ops *
3609 nft_select_set_ops(const struct nft_ctx *ctx,
3610 const struct nlattr * const nla[],
3611 const struct nft_set_desc *desc,
3612 enum nft_set_policies policy)
3613 {
3614 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3615 const struct nft_set_ops *ops, *bops;
3616 struct nft_set_estimate est, best;
3617 const struct nft_set_type *type;
3618 u32 flags = 0;
3619 int i;
3620
3621 lockdep_assert_held(&nft_net->commit_mutex);
3622 lockdep_nfnl_nft_mutex_not_held();
3623
3624 if (nla[NFTA_SET_FLAGS] != NULL)
3625 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
3626
3627 bops = NULL;
3628 best.size = ~0;
3629 best.lookup = ~0;
3630 best.space = ~0;
3631
3632 for (i = 0; i < ARRAY_SIZE(nft_set_types); i++) {
3633 type = nft_set_types[i];
3634 ops = &type->ops;
3635
3636 if (!nft_set_ops_candidate(type, flags))
3637 continue;
3638 if (!ops->estimate(desc, flags, &est))
3639 continue;
3640
3641 switch (policy) {
3642 case NFT_SET_POL_PERFORMANCE:
3643 if (est.lookup < best.lookup)
3644 break;
3645 if (est.lookup == best.lookup &&
3646 est.space < best.space)
3647 break;
3648 continue;
3649 case NFT_SET_POL_MEMORY:
3650 if (!desc->size) {
3651 if (est.space < best.space)
3652 break;
3653 if (est.space == best.space &&
3654 est.lookup < best.lookup)
3655 break;
3656 } else if (est.size < best.size || !bops) {
3657 break;
3658 }
3659 continue;
3660 default:
3661 break;
3662 }
3663
3664 bops = ops;
3665 best = est;
3666 }
3667
3668 if (bops != NULL)
3669 return bops;
3670
3671 return ERR_PTR(-EOPNOTSUPP);
3672 }
3673
3674 static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
3675 [NFTA_SET_TABLE] = { .type = NLA_STRING,
3676 .len = NFT_TABLE_MAXNAMELEN - 1 },
3677 [NFTA_SET_NAME] = { .type = NLA_STRING,
3678 .len = NFT_SET_MAXNAMELEN - 1 },
3679 [NFTA_SET_FLAGS] = { .type = NLA_U32 },
3680 [NFTA_SET_KEY_TYPE] = { .type = NLA_U32 },
3681 [NFTA_SET_KEY_LEN] = { .type = NLA_U32 },
3682 [NFTA_SET_DATA_TYPE] = { .type = NLA_U32 },
3683 [NFTA_SET_DATA_LEN] = { .type = NLA_U32 },
3684 [NFTA_SET_POLICY] = { .type = NLA_U32 },
3685 [NFTA_SET_DESC] = { .type = NLA_NESTED },
3686 [NFTA_SET_ID] = { .type = NLA_U32 },
3687 [NFTA_SET_TIMEOUT] = { .type = NLA_U64 },
3688 [NFTA_SET_GC_INTERVAL] = { .type = NLA_U32 },
3689 [NFTA_SET_USERDATA] = { .type = NLA_BINARY,
3690 .len = NFT_USERDATA_MAXLEN },
3691 [NFTA_SET_OBJ_TYPE] = { .type = NLA_U32 },
3692 [NFTA_SET_HANDLE] = { .type = NLA_U64 },
3693 [NFTA_SET_EXPR] = { .type = NLA_NESTED },
3694 [NFTA_SET_EXPRESSIONS] = { .type = NLA_NESTED },
3695 };
3696
3697 static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
3698 [NFTA_SET_DESC_SIZE] = { .type = NLA_U32 },
3699 [NFTA_SET_DESC_CONCAT] = { .type = NLA_NESTED },
3700 };
3701
3702 static struct nft_set *nft_set_lookup(const struct nft_table *table,
3703 const struct nlattr *nla, u8 genmask)
3704 {
3705 struct nft_set *set;
3706
3707 if (nla == NULL)
3708 return ERR_PTR(-EINVAL);
3709
3710 list_for_each_entry_rcu(set, &table->sets, list) {
3711 if (!nla_strcmp(nla, set->name) &&
3712 nft_active_genmask(set, genmask))
3713 return set;
3714 }
3715 return ERR_PTR(-ENOENT);
3716 }
3717
3718 static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
3719 const struct nlattr *nla,
3720 u8 genmask)
3721 {
3722 struct nft_set *set;
3723
3724 list_for_each_entry(set, &table->sets, list) {
3725 if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
3726 nft_active_genmask(set, genmask))
3727 return set;
3728 }
3729 return ERR_PTR(-ENOENT);
3730 }
3731
3732 static struct nft_set *nft_set_lookup_byid(const struct net *net,
3733 const struct nlattr *nla, u8 genmask)
3734 {
3735 struct nftables_pernet *nft_net = nft_pernet(net);
3736 u32 id = ntohl(nla_get_be32(nla));
3737 struct nft_trans *trans;
3738
3739 list_for_each_entry(trans, &nft_net->commit_list, list) {
3740 if (trans->msg_type == NFT_MSG_NEWSET) {
3741 struct nft_set *set = nft_trans_set(trans);
3742
3743 if (id == nft_trans_set_id(trans) &&
3744 nft_active_genmask(set, genmask))
3745 return set;
3746 }
3747 }
3748 return ERR_PTR(-ENOENT);
3749 }
3750
3751 struct nft_set *nft_set_lookup_global(const struct net *net,
3752 const struct nft_table *table,
3753 const struct nlattr *nla_set_name,
3754 const struct nlattr *nla_set_id,
3755 u8 genmask)
3756 {
3757 struct nft_set *set;
3758
3759 set = nft_set_lookup(table, nla_set_name, genmask);
3760 if (IS_ERR(set)) {
3761 if (!nla_set_id)
3762 return set;
3763
3764 set = nft_set_lookup_byid(net, nla_set_id, genmask);
3765 }
3766 return set;
3767 }
3768 EXPORT_SYMBOL_GPL(nft_set_lookup_global);
3769
3770 static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
3771 const char *name)
3772 {
3773 const struct nft_set *i;
3774 const char *p;
3775 unsigned long *inuse;
3776 unsigned int n = 0, min = 0;
3777
3778 p = strchr(name, '%');
3779 if (p != NULL) {
3780 if (p[1] != 'd' || strchr(p + 2, '%'))
3781 return -EINVAL;
3782
3783 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
3784 if (inuse == NULL)
3785 return -ENOMEM;
3786 cont:
3787 list_for_each_entry(i, &ctx->table->sets, list) {
3788 int tmp;
3789
3790 if (!nft_is_active_next(ctx->net, set))
3791 continue;
3792 if (!sscanf(i->name, name, &tmp))
3793 continue;
3794 if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
3795 continue;
3796
3797 set_bit(tmp - min, inuse);
3798 }
3799
3800 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
3801 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
3802 min += BITS_PER_BYTE * PAGE_SIZE;
3803 memset(inuse, 0, PAGE_SIZE);
3804 goto cont;
3805 }
3806 free_page((unsigned long)inuse);
3807 }
3808
3809 set->name = kasprintf(GFP_KERNEL, name, min + n);
3810 if (!set->name)
3811 return -ENOMEM;
3812
3813 list_for_each_entry(i, &ctx->table->sets, list) {
3814 if (!nft_is_active_next(ctx->net, i))
3815 continue;
3816 if (!strcmp(set->name, i->name)) {
3817 kfree(set->name);
3818 set->name = NULL;
3819 return -ENFILE;
3820 }
3821 }
3822 return 0;
3823 }
3824
3825 int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
3826 {
3827 u64 ms = be64_to_cpu(nla_get_be64(nla));
3828 u64 max = (u64)(~((u64)0));
3829
3830 max = div_u64(max, NSEC_PER_MSEC);
3831 if (ms >= max)
3832 return -ERANGE;
3833
3834 ms *= NSEC_PER_MSEC;
3835 *result = nsecs_to_jiffies64(ms);
3836 return 0;
3837 }
3838
3839 __be64 nf_jiffies64_to_msecs(u64 input)
3840 {
3841 return cpu_to_be64(jiffies64_to_msecs(input));
3842 }
3843
3844 static int nf_tables_fill_set_concat(struct sk_buff *skb,
3845 const struct nft_set *set)
3846 {
3847 struct nlattr *concat, *field;
3848 int i;
3849
3850 concat = nla_nest_start_noflag(skb, NFTA_SET_DESC_CONCAT);
3851 if (!concat)
3852 return -ENOMEM;
3853
3854 for (i = 0; i < set->field_count; i++) {
3855 field = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
3856 if (!field)
3857 return -ENOMEM;
3858
3859 if (nla_put_be32(skb, NFTA_SET_FIELD_LEN,
3860 htonl(set->field_len[i])))
3861 return -ENOMEM;
3862
3863 nla_nest_end(skb, field);
3864 }
3865
3866 nla_nest_end(skb, concat);
3867
3868 return 0;
3869 }
3870
3871 static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
3872 const struct nft_set *set, u16 event, u16 flags)
3873 {
3874 struct nlmsghdr *nlh;
3875 u32 portid = ctx->portid;
3876 struct nlattr *nest;
3877 u32 seq = ctx->seq;
3878 int i;
3879
3880 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3881 nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
3882 NFNETLINK_V0, nft_base_seq(ctx->net));
3883 if (!nlh)
3884 goto nla_put_failure;
3885
3886 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
3887 goto nla_put_failure;
3888 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
3889 goto nla_put_failure;
3890 if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
3891 NFTA_SET_PAD))
3892 goto nla_put_failure;
3893 if (set->flags != 0)
3894 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
3895 goto nla_put_failure;
3896
3897 if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
3898 goto nla_put_failure;
3899 if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
3900 goto nla_put_failure;
3901 if (set->flags & NFT_SET_MAP) {
3902 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
3903 goto nla_put_failure;
3904 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
3905 goto nla_put_failure;
3906 }
3907 if (set->flags & NFT_SET_OBJECT &&
3908 nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
3909 goto nla_put_failure;
3910
3911 if (set->timeout &&
3912 nla_put_be64(skb, NFTA_SET_TIMEOUT,
3913 nf_jiffies64_to_msecs(set->timeout),
3914 NFTA_SET_PAD))
3915 goto nla_put_failure;
3916 if (set->gc_int &&
3917 nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(set->gc_int)))
3918 goto nla_put_failure;
3919
3920 if (set->policy != NFT_SET_POL_PERFORMANCE) {
3921 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
3922 goto nla_put_failure;
3923 }
3924
3925 if (set->udata &&
3926 nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
3927 goto nla_put_failure;
3928
3929 nest = nla_nest_start_noflag(skb, NFTA_SET_DESC);
3930 if (!nest)
3931 goto nla_put_failure;
3932 if (set->size &&
3933 nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
3934 goto nla_put_failure;
3935
3936 if (set->field_count > 1 &&
3937 nf_tables_fill_set_concat(skb, set))
3938 goto nla_put_failure;
3939
3940 nla_nest_end(skb, nest);
3941
3942 if (set->num_exprs == 1) {
3943 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPR);
3944 if (nf_tables_fill_expr_info(skb, set->exprs[0]) < 0)
3945 goto nla_put_failure;
3946
3947 nla_nest_end(skb, nest);
3948 } else if (set->num_exprs > 1) {
3949 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPRESSIONS);
3950 if (nest == NULL)
3951 goto nla_put_failure;
3952
3953 for (i = 0; i < set->num_exprs; i++) {
3954 if (nft_expr_dump(skb, NFTA_LIST_ELEM,
3955 set->exprs[i]) < 0)
3956 goto nla_put_failure;
3957 }
3958 nla_nest_end(skb, nest);
3959 }
3960
3961 nlmsg_end(skb, nlh);
3962 return 0;
3963
3964 nla_put_failure:
3965 nlmsg_trim(skb, nlh);
3966 return -1;
3967 }
3968
3969 static void nf_tables_set_notify(const struct nft_ctx *ctx,
3970 const struct nft_set *set, int event,
3971 gfp_t gfp_flags)
3972 {
3973 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3974 u32 portid = ctx->portid;
3975 struct sk_buff *skb;
3976 u16 flags = 0;
3977 int err;
3978
3979 if (!ctx->report &&
3980 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3981 return;
3982
3983 skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
3984 if (skb == NULL)
3985 goto err;
3986
3987 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
3988 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
3989
3990 err = nf_tables_fill_set(skb, ctx, set, event, flags);
3991 if (err < 0) {
3992 kfree_skb(skb);
3993 goto err;
3994 }
3995
3996 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
3997 return;
3998 err:
3999 nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
4000 }
4001
4002 static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
4003 {
4004 const struct nft_set *set;
4005 unsigned int idx, s_idx = cb->args[0];
4006 struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
4007 struct net *net = sock_net(skb->sk);
4008 struct nft_ctx *ctx = cb->data, ctx_set;
4009 struct nftables_pernet *nft_net;
4010
4011 if (cb->args[1])
4012 return skb->len;
4013
4014 rcu_read_lock();
4015 nft_net = nft_pernet(net);
4016 cb->seq = nft_net->base_seq;
4017
4018 list_for_each_entry_rcu(table, &nft_net->tables, list) {
4019 if (ctx->family != NFPROTO_UNSPEC &&
4020 ctx->family != table->family)
4021 continue;
4022
4023 if (ctx->table && ctx->table != table)
4024 continue;
4025
4026 if (cur_table) {
4027 if (cur_table != table)
4028 continue;
4029
4030 cur_table = NULL;
4031 }
4032 idx = 0;
4033 list_for_each_entry_rcu(set, &table->sets, list) {
4034 if (idx < s_idx)
4035 goto cont;
4036 if (!nft_is_active(net, set))
4037 goto cont;
4038
4039 ctx_set = *ctx;
4040 ctx_set.table = table;
4041 ctx_set.family = table->family;
4042
4043 if (nf_tables_fill_set(skb, &ctx_set, set,
4044 NFT_MSG_NEWSET,
4045 NLM_F_MULTI) < 0) {
4046 cb->args[0] = idx;
4047 cb->args[2] = (unsigned long) table;
4048 goto done;
4049 }
4050 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
4051 cont:
4052 idx++;
4053 }
4054 if (s_idx)
4055 s_idx = 0;
4056 }
4057 cb->args[1] = 1;
4058 done:
4059 rcu_read_unlock();
4060 return skb->len;
4061 }
4062
4063 static int nf_tables_dump_sets_start(struct netlink_callback *cb)
4064 {
4065 struct nft_ctx *ctx_dump = NULL;
4066
4067 ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
4068 if (ctx_dump == NULL)
4069 return -ENOMEM;
4070
4071 cb->data = ctx_dump;
4072 return 0;
4073 }
4074
4075 static int nf_tables_dump_sets_done(struct netlink_callback *cb)
4076 {
4077 kfree(cb->data);
4078 return 0;
4079 }
4080
4081 /* called with rcu_read_lock held */
4082 static int nf_tables_getset(struct sk_buff *skb, const struct nfnl_info *info,
4083 const struct nlattr * const nla[])
4084 {
4085 struct netlink_ext_ack *extack = info->extack;
4086 u8 genmask = nft_genmask_cur(info->net);
4087 u8 family = info->nfmsg->nfgen_family;
4088 struct nft_table *table = NULL;
4089 struct net *net = info->net;
4090 const struct nft_set *set;
4091 struct sk_buff *skb2;
4092 struct nft_ctx ctx;
4093 int err;
4094
4095 if (nla[NFTA_SET_TABLE]) {
4096 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4097 genmask, 0);
4098 if (IS_ERR(table)) {
4099 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4100 return PTR_ERR(table);
4101 }
4102 }
4103
4104 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4105
4106 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
4107 struct netlink_dump_control c = {
4108 .start = nf_tables_dump_sets_start,
4109 .dump = nf_tables_dump_sets,
4110 .done = nf_tables_dump_sets_done,
4111 .data = &ctx,
4112 .module = THIS_MODULE,
4113 };
4114
4115 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
4116 }
4117
4118 /* Only accept unspec with dump */
4119 if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4120 return -EAFNOSUPPORT;
4121 if (!nla[NFTA_SET_TABLE])
4122 return -EINVAL;
4123
4124 set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4125 if (IS_ERR(set))
4126 return PTR_ERR(set);
4127
4128 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
4129 if (skb2 == NULL)
4130 return -ENOMEM;
4131
4132 err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
4133 if (err < 0)
4134 goto err_fill_set_info;
4135
4136 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
4137
4138 err_fill_set_info:
4139 kfree_skb(skb2);
4140 return err;
4141 }
4142
4143 static const struct nla_policy nft_concat_policy[NFTA_SET_FIELD_MAX + 1] = {
4144 [NFTA_SET_FIELD_LEN] = { .type = NLA_U32 },
4145 };
4146
4147 static int nft_set_desc_concat_parse(const struct nlattr *attr,
4148 struct nft_set_desc *desc)
4149 {
4150 struct nlattr *tb[NFTA_SET_FIELD_MAX + 1];
4151 u32 len;
4152 int err;
4153
4154 err = nla_parse_nested_deprecated(tb, NFTA_SET_FIELD_MAX, attr,
4155 nft_concat_policy, NULL);
4156 if (err < 0)
4157 return err;
4158
4159 if (!tb[NFTA_SET_FIELD_LEN])
4160 return -EINVAL;
4161
4162 len = ntohl(nla_get_be32(tb[NFTA_SET_FIELD_LEN]));
4163
4164 if (len * BITS_PER_BYTE / 32 > NFT_REG32_COUNT)
4165 return -E2BIG;
4166
4167 desc->field_len[desc->field_count++] = len;
4168
4169 return 0;
4170 }
4171
4172 static int nft_set_desc_concat(struct nft_set_desc *desc,
4173 const struct nlattr *nla)
4174 {
4175 struct nlattr *attr;
4176 int rem, err;
4177
4178 nla_for_each_nested(attr, nla, rem) {
4179 if (nla_type(attr) != NFTA_LIST_ELEM)
4180 return -EINVAL;
4181
4182 err = nft_set_desc_concat_parse(attr, desc);
4183 if (err < 0)
4184 return err;
4185 }
4186
4187 return 0;
4188 }
4189
4190 static int nf_tables_set_desc_parse(struct nft_set_desc *desc,
4191 const struct nlattr *nla)
4192 {
4193 struct nlattr *da[NFTA_SET_DESC_MAX + 1];
4194 int err;
4195
4196 err = nla_parse_nested_deprecated(da, NFTA_SET_DESC_MAX, nla,
4197 nft_set_desc_policy, NULL);
4198 if (err < 0)
4199 return err;
4200
4201 if (da[NFTA_SET_DESC_SIZE] != NULL)
4202 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
4203 if (da[NFTA_SET_DESC_CONCAT])
4204 err = nft_set_desc_concat(desc, da[NFTA_SET_DESC_CONCAT]);
4205
4206 return err;
4207 }
4208
4209 static int nf_tables_newset(struct sk_buff *skb, const struct nfnl_info *info,
4210 const struct nlattr * const nla[])
4211 {
4212 u32 ktype, dtype, flags, policy, gc_int, objtype;
4213 struct netlink_ext_ack *extack = info->extack;
4214 u8 genmask = nft_genmask_next(info->net);
4215 u8 family = info->nfmsg->nfgen_family;
4216 const struct nft_set_ops *ops;
4217 struct nft_expr *expr = NULL;
4218 struct net *net = info->net;
4219 struct nft_set_desc desc;
4220 struct nft_table *table;
4221 unsigned char *udata;
4222 struct nft_set *set;
4223 struct nft_ctx ctx;
4224 size_t alloc_size;
4225 u64 timeout;
4226 char *name;
4227 int err, i;
4228 u16 udlen;
4229 u64 size;
4230
4231 if (nla[NFTA_SET_TABLE] == NULL ||
4232 nla[NFTA_SET_NAME] == NULL ||
4233 nla[NFTA_SET_KEY_LEN] == NULL ||
4234 nla[NFTA_SET_ID] == NULL)
4235 return -EINVAL;
4236
4237 memset(&desc, 0, sizeof(desc));
4238
4239 ktype = NFT_DATA_VALUE;
4240 if (nla[NFTA_SET_KEY_TYPE] != NULL) {
4241 ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
4242 if ((ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
4243 return -EINVAL;
4244 }
4245
4246 desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
4247 if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
4248 return -EINVAL;
4249
4250 flags = 0;
4251 if (nla[NFTA_SET_FLAGS] != NULL) {
4252 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
4253 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
4254 NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
4255 NFT_SET_MAP | NFT_SET_EVAL |
4256 NFT_SET_OBJECT | NFT_SET_CONCAT | NFT_SET_EXPR))
4257 return -EOPNOTSUPP;
4258 /* Only one of these operations is supported */
4259 if ((flags & (NFT_SET_MAP | NFT_SET_OBJECT)) ==
4260 (NFT_SET_MAP | NFT_SET_OBJECT))
4261 return -EOPNOTSUPP;
4262 if ((flags & (NFT_SET_EVAL | NFT_SET_OBJECT)) ==
4263 (NFT_SET_EVAL | NFT_SET_OBJECT))
4264 return -EOPNOTSUPP;
4265 }
4266
4267 dtype = 0;
4268 if (nla[NFTA_SET_DATA_TYPE] != NULL) {
4269 if (!(flags & NFT_SET_MAP))
4270 return -EINVAL;
4271
4272 dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
4273 if ((dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
4274 dtype != NFT_DATA_VERDICT)
4275 return -EINVAL;
4276
4277 if (dtype != NFT_DATA_VERDICT) {
4278 if (nla[NFTA_SET_DATA_LEN] == NULL)
4279 return -EINVAL;
4280 desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
4281 if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
4282 return -EINVAL;
4283 } else
4284 desc.dlen = sizeof(struct nft_verdict);
4285 } else if (flags & NFT_SET_MAP)
4286 return -EINVAL;
4287
4288 if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
4289 if (!(flags & NFT_SET_OBJECT))
4290 return -EINVAL;
4291
4292 objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
4293 if (objtype == NFT_OBJECT_UNSPEC ||
4294 objtype > NFT_OBJECT_MAX)
4295 return -EOPNOTSUPP;
4296 } else if (flags & NFT_SET_OBJECT)
4297 return -EINVAL;
4298 else
4299 objtype = NFT_OBJECT_UNSPEC;
4300
4301 timeout = 0;
4302 if (nla[NFTA_SET_TIMEOUT] != NULL) {
4303 if (!(flags & NFT_SET_TIMEOUT))
4304 return -EINVAL;
4305
4306 err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &timeout);
4307 if (err)
4308 return err;
4309 }
4310 gc_int = 0;
4311 if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
4312 if (!(flags & NFT_SET_TIMEOUT))
4313 return -EINVAL;
4314 gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
4315 }
4316
4317 policy = NFT_SET_POL_PERFORMANCE;
4318 if (nla[NFTA_SET_POLICY] != NULL)
4319 policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
4320
4321 if (nla[NFTA_SET_DESC] != NULL) {
4322 err = nf_tables_set_desc_parse(&desc, nla[NFTA_SET_DESC]);
4323 if (err < 0)
4324 return err;
4325 }
4326
4327 if (nla[NFTA_SET_EXPR] || nla[NFTA_SET_EXPRESSIONS])
4328 desc.expr = true;
4329
4330 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask,
4331 NETLINK_CB(skb).portid);
4332 if (IS_ERR(table)) {
4333 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4334 return PTR_ERR(table);
4335 }
4336
4337 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4338
4339 set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4340 if (IS_ERR(set)) {
4341 if (PTR_ERR(set) != -ENOENT) {
4342 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4343 return PTR_ERR(set);
4344 }
4345 } else {
4346 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
4347 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4348 return -EEXIST;
4349 }
4350 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
4351 return -EOPNOTSUPP;
4352
4353 return 0;
4354 }
4355
4356 if (!(info->nlh->nlmsg_flags & NLM_F_CREATE))
4357 return -ENOENT;
4358
4359 ops = nft_select_set_ops(&ctx, nla, &desc, policy);
4360 if (IS_ERR(ops))
4361 return PTR_ERR(ops);
4362
4363 udlen = 0;
4364 if (nla[NFTA_SET_USERDATA])
4365 udlen = nla_len(nla[NFTA_SET_USERDATA]);
4366
4367 size = 0;
4368 if (ops->privsize != NULL)
4369 size = ops->privsize(nla, &desc);
4370 alloc_size = sizeof(*set) + size + udlen;
4371 if (alloc_size < size || alloc_size > INT_MAX)
4372 return -ENOMEM;
4373 set = kvzalloc(alloc_size, GFP_KERNEL);
4374 if (!set)
4375 return -ENOMEM;
4376
4377 name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL);
4378 if (!name) {
4379 err = -ENOMEM;
4380 goto err_set_name;
4381 }
4382
4383 err = nf_tables_set_alloc_name(&ctx, set, name);
4384 kfree(name);
4385 if (err < 0)
4386 goto err_set_name;
4387
4388 udata = NULL;
4389 if (udlen) {
4390 udata = set->data + size;
4391 nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
4392 }
4393
4394 INIT_LIST_HEAD(&set->bindings);
4395 INIT_LIST_HEAD(&set->catchall_list);
4396 set->table = table;
4397 write_pnet(&set->net, net);
4398 set->ops = ops;
4399 set->ktype = ktype;
4400 set->klen = desc.klen;
4401 set->dtype = dtype;
4402 set->objtype = objtype;
4403 set->dlen = desc.dlen;
4404 set->flags = flags;
4405 set->size = desc.size;
4406 set->policy = policy;
4407 set->udlen = udlen;
4408 set->udata = udata;
4409 set->timeout = timeout;
4410 set->gc_int = gc_int;
4411
4412 set->field_count = desc.field_count;
4413 for (i = 0; i < desc.field_count; i++)
4414 set->field_len[i] = desc.field_len[i];
4415
4416 err = ops->init(set, &desc, nla);
4417 if (err < 0)
4418 goto err_set_init;
4419
4420 if (nla[NFTA_SET_EXPR]) {
4421 expr = nft_set_elem_expr_alloc(&ctx, set, nla[NFTA_SET_EXPR]);
4422 if (IS_ERR(expr)) {
4423 err = PTR_ERR(expr);
4424 goto err_set_expr_alloc;
4425 }
4426 set->exprs[0] = expr;
4427 set->num_exprs++;
4428 } else if (nla[NFTA_SET_EXPRESSIONS]) {
4429 struct nft_expr *expr;
4430 struct nlattr *tmp;
4431 int left;
4432
4433 if (!(flags & NFT_SET_EXPR)) {
4434 err = -EINVAL;
4435 goto err_set_expr_alloc;
4436 }
4437 i = 0;
4438 nla_for_each_nested(tmp, nla[NFTA_SET_EXPRESSIONS], left) {
4439 if (i == NFT_SET_EXPR_MAX) {
4440 err = -E2BIG;
4441 goto err_set_expr_alloc;
4442 }
4443 if (nla_type(tmp) != NFTA_LIST_ELEM) {
4444 err = -EINVAL;
4445 goto err_set_expr_alloc;
4446 }
4447 expr = nft_set_elem_expr_alloc(&ctx, set, tmp);
4448 if (IS_ERR(expr)) {
4449 err = PTR_ERR(expr);
4450 goto err_set_expr_alloc;
4451 }
4452 set->exprs[i++] = expr;
4453 set->num_exprs++;
4454 }
4455 }
4456
4457 set->handle = nf_tables_alloc_handle(table);
4458
4459 err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
4460 if (err < 0)
4461 goto err_set_expr_alloc;
4462
4463 list_add_tail_rcu(&set->list, &table->sets);
4464 table->use++;
4465 return 0;
4466
4467 err_set_expr_alloc:
4468 for (i = 0; i < set->num_exprs; i++)
4469 nft_expr_destroy(&ctx, set->exprs[i]);
4470
4471 ops->destroy(set);
4472 err_set_init:
4473 kfree(set->name);
4474 err_set_name:
4475 kvfree(set);
4476 return err;
4477 }
4478
4479 struct nft_set_elem_catchall {
4480 struct list_head list;
4481 struct rcu_head rcu;
4482 void *elem;
4483 };
4484
4485 static void nft_set_catchall_destroy(const struct nft_ctx *ctx,
4486 struct nft_set *set)
4487 {
4488 struct nft_set_elem_catchall *next, *catchall;
4489
4490 list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
4491 list_del_rcu(&catchall->list);
4492 nft_set_elem_destroy(set, catchall->elem, true);
4493 kfree_rcu(catchall, rcu);
4494 }
4495 }
4496
4497 static void nft_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
4498 {
4499 int i;
4500
4501 if (WARN_ON(set->use > 0))
4502 return;
4503
4504 for (i = 0; i < set->num_exprs; i++)
4505 nft_expr_destroy(ctx, set->exprs[i]);
4506
4507 set->ops->destroy(set);
4508 nft_set_catchall_destroy(ctx, set);
4509 kfree(set->name);
4510 kvfree(set);
4511 }
4512
4513 static int nf_tables_delset(struct sk_buff *skb, const struct nfnl_info *info,
4514 const struct nlattr * const nla[])
4515 {
4516 struct netlink_ext_ack *extack = info->extack;
4517 u8 genmask = nft_genmask_next(info->net);
4518 u8 family = info->nfmsg->nfgen_family;
4519 struct net *net = info->net;
4520 const struct nlattr *attr;
4521 struct nft_table *table;
4522 struct nft_set *set;
4523 struct nft_ctx ctx;
4524
4525 if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4526 return -EAFNOSUPPORT;
4527
4528 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4529 genmask, NETLINK_CB(skb).portid);
4530 if (IS_ERR(table)) {
4531 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4532 return PTR_ERR(table);
4533 }
4534
4535 if (nla[NFTA_SET_HANDLE]) {
4536 attr = nla[NFTA_SET_HANDLE];
4537 set = nft_set_lookup_byhandle(table, attr, genmask);
4538 } else {
4539 attr = nla[NFTA_SET_NAME];
4540 set = nft_set_lookup(table, attr, genmask);
4541 }
4542
4543 if (IS_ERR(set)) {
4544 NL_SET_BAD_ATTR(extack, attr);
4545 return PTR_ERR(set);
4546 }
4547 if (set->use ||
4548 (info->nlh->nlmsg_flags & NLM_F_NONREC &&
4549 atomic_read(&set->nelems) > 0)) {
4550 NL_SET_BAD_ATTR(extack, attr);
4551 return -EBUSY;
4552 }
4553
4554 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4555
4556 return nft_delset(&ctx, set);
4557 }
4558
4559 static int nft_validate_register_store(const struct nft_ctx *ctx,
4560 enum nft_registers reg,
4561 const struct nft_data *data,
4562 enum nft_data_types type,
4563 unsigned int len);
4564
4565 static int nft_setelem_data_validate(const struct nft_ctx *ctx,
4566 struct nft_set *set,
4567 struct nft_set_elem *elem)
4568 {
4569 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4570 enum nft_registers dreg;
4571
4572 dreg = nft_type_to_reg(set->dtype);
4573 return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
4574 set->dtype == NFT_DATA_VERDICT ?
4575 NFT_DATA_VERDICT : NFT_DATA_VALUE,
4576 set->dlen);
4577 }
4578
4579 static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
4580 struct nft_set *set,
4581 const struct nft_set_iter *iter,
4582 struct nft_set_elem *elem)
4583 {
4584 return nft_setelem_data_validate(ctx, set, elem);
4585 }
4586
4587 static int nft_set_catchall_bind_check(const struct nft_ctx *ctx,
4588 struct nft_set *set)
4589 {
4590 u8 genmask = nft_genmask_next(ctx->net);
4591 struct nft_set_elem_catchall *catchall;
4592 struct nft_set_elem elem;
4593 struct nft_set_ext *ext;
4594 int ret = 0;
4595
4596 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4597 ext = nft_set_elem_ext(set, catchall->elem);
4598 if (!nft_set_elem_active(ext, genmask))
4599 continue;
4600
4601 elem.priv = catchall->elem;
4602 ret = nft_setelem_data_validate(ctx, set, &elem);
4603 if (ret < 0)
4604 break;
4605 }
4606
4607 return ret;
4608 }
4609
4610 int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
4611 struct nft_set_binding *binding)
4612 {
4613 struct nft_set_binding *i;
4614 struct nft_set_iter iter;
4615
4616 if (set->use == UINT_MAX)
4617 return -EOVERFLOW;
4618
4619 if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
4620 return -EBUSY;
4621
4622 if (binding->flags & NFT_SET_MAP) {
4623 /* If the set is already bound to the same chain all
4624 * jumps are already validated for that chain.
4625 */
4626 list_for_each_entry(i, &set->bindings, list) {
4627 if (i->flags & NFT_SET_MAP &&
4628 i->chain == binding->chain)
4629 goto bind;
4630 }
4631
4632 iter.genmask = nft_genmask_next(ctx->net);
4633 iter.skip = 0;
4634 iter.count = 0;
4635 iter.err = 0;
4636 iter.fn = nf_tables_bind_check_setelem;
4637
4638 set->ops->walk(ctx, set, &iter);
4639 if (!iter.err)
4640 iter.err = nft_set_catchall_bind_check(ctx, set);
4641
4642 if (iter.err < 0)
4643 return iter.err;
4644 }
4645 bind:
4646 binding->chain = ctx->chain;
4647 list_add_tail_rcu(&binding->list, &set->bindings);
4648 nft_set_trans_bind(ctx, set);
4649 set->use++;
4650
4651 return 0;
4652 }
4653 EXPORT_SYMBOL_GPL(nf_tables_bind_set);
4654
4655 static void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
4656 struct nft_set_binding *binding, bool event)
4657 {
4658 list_del_rcu(&binding->list);
4659
4660 if (list_empty(&set->bindings) && nft_set_is_anonymous(set)) {
4661 list_del_rcu(&set->list);
4662 if (event)
4663 nf_tables_set_notify(ctx, set, NFT_MSG_DELSET,
4664 GFP_KERNEL);
4665 }
4666 }
4667
4668 void nf_tables_deactivate_set(const struct nft_ctx *ctx, struct nft_set *set,
4669 struct nft_set_binding *binding,
4670 enum nft_trans_phase phase)
4671 {
4672 switch (phase) {
4673 case NFT_TRANS_PREPARE:
4674 set->use--;
4675 return;
4676 case NFT_TRANS_ABORT:
4677 case NFT_TRANS_RELEASE:
4678 set->use--;
4679 fallthrough;
4680 default:
4681 nf_tables_unbind_set(ctx, set, binding,
4682 phase == NFT_TRANS_COMMIT);
4683 }
4684 }
4685 EXPORT_SYMBOL_GPL(nf_tables_deactivate_set);
4686
4687 void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
4688 {
4689 if (list_empty(&set->bindings) && nft_set_is_anonymous(set))
4690 nft_set_destroy(ctx, set);
4691 }
4692 EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
4693
4694 const struct nft_set_ext_type nft_set_ext_types[] = {
4695 [NFT_SET_EXT_KEY] = {
4696 .align = __alignof__(u32),
4697 },
4698 [NFT_SET_EXT_DATA] = {
4699 .align = __alignof__(u32),
4700 },
4701 [NFT_SET_EXT_EXPRESSIONS] = {
4702 .align = __alignof__(struct nft_set_elem_expr),
4703 },
4704 [NFT_SET_EXT_OBJREF] = {
4705 .len = sizeof(struct nft_object *),
4706 .align = __alignof__(struct nft_object *),
4707 },
4708 [NFT_SET_EXT_FLAGS] = {
4709 .len = sizeof(u8),
4710 .align = __alignof__(u8),
4711 },
4712 [NFT_SET_EXT_TIMEOUT] = {
4713 .len = sizeof(u64),
4714 .align = __alignof__(u64),
4715 },
4716 [NFT_SET_EXT_EXPIRATION] = {
4717 .len = sizeof(u64),
4718 .align = __alignof__(u64),
4719 },
4720 [NFT_SET_EXT_USERDATA] = {
4721 .len = sizeof(struct nft_userdata),
4722 .align = __alignof__(struct nft_userdata),
4723 },
4724 [NFT_SET_EXT_KEY_END] = {
4725 .align = __alignof__(u32),
4726 },
4727 };
4728
4729 /*
4730 * Set elements
4731 */
4732
4733 static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
4734 [NFTA_SET_ELEM_KEY] = { .type = NLA_NESTED },
4735 [NFTA_SET_ELEM_DATA] = { .type = NLA_NESTED },
4736 [NFTA_SET_ELEM_FLAGS] = { .type = NLA_U32 },
4737 [NFTA_SET_ELEM_TIMEOUT] = { .type = NLA_U64 },
4738 [NFTA_SET_ELEM_EXPIRATION] = { .type = NLA_U64 },
4739 [NFTA_SET_ELEM_USERDATA] = { .type = NLA_BINARY,
4740 .len = NFT_USERDATA_MAXLEN },
4741 [NFTA_SET_ELEM_EXPR] = { .type = NLA_NESTED },
4742 [NFTA_SET_ELEM_OBJREF] = { .type = NLA_STRING,
4743 .len = NFT_OBJ_MAXNAMELEN - 1 },
4744 [NFTA_SET_ELEM_KEY_END] = { .type = NLA_NESTED },
4745 [NFTA_SET_ELEM_EXPRESSIONS] = { .type = NLA_NESTED },
4746 };
4747
4748 static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
4749 [NFTA_SET_ELEM_LIST_TABLE] = { .type = NLA_STRING,
4750 .len = NFT_TABLE_MAXNAMELEN - 1 },
4751 [NFTA_SET_ELEM_LIST_SET] = { .type = NLA_STRING,
4752 .len = NFT_SET_MAXNAMELEN - 1 },
4753 [NFTA_SET_ELEM_LIST_ELEMENTS] = { .type = NLA_NESTED },
4754 [NFTA_SET_ELEM_LIST_SET_ID] = { .type = NLA_U32 },
4755 };
4756
4757 static int nft_set_elem_expr_dump(struct sk_buff *skb,
4758 const struct nft_set *set,
4759 const struct nft_set_ext *ext)
4760 {
4761 struct nft_set_elem_expr *elem_expr;
4762 u32 size, num_exprs = 0;
4763 struct nft_expr *expr;
4764 struct nlattr *nest;
4765
4766 elem_expr = nft_set_ext_expr(ext);
4767 nft_setelem_expr_foreach(expr, elem_expr, size)
4768 num_exprs++;
4769
4770 if (num_exprs == 1) {
4771 expr = nft_setelem_expr_at(elem_expr, 0);
4772 if (nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, expr) < 0)
4773 return -1;
4774
4775 return 0;
4776 } else if (num_exprs > 1) {
4777 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_EXPRESSIONS);
4778 if (nest == NULL)
4779 goto nla_put_failure;
4780
4781 nft_setelem_expr_foreach(expr, elem_expr, size) {
4782 expr = nft_setelem_expr_at(elem_expr, size);
4783 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr) < 0)
4784 goto nla_put_failure;
4785 }
4786 nla_nest_end(skb, nest);
4787 }
4788 return 0;
4789
4790 nla_put_failure:
4791 return -1;
4792 }
4793
4794 static int nf_tables_fill_setelem(struct sk_buff *skb,
4795 const struct nft_set *set,
4796 const struct nft_set_elem *elem)
4797 {
4798 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
4799 unsigned char *b = skb_tail_pointer(skb);
4800 struct nlattr *nest;
4801
4802 nest = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4803 if (nest == NULL)
4804 goto nla_put_failure;
4805
4806 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
4807 nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
4808 NFT_DATA_VALUE, set->klen) < 0)
4809 goto nla_put_failure;
4810
4811 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
4812 nft_data_dump(skb, NFTA_SET_ELEM_KEY_END, nft_set_ext_key_end(ext),
4813 NFT_DATA_VALUE, set->klen) < 0)
4814 goto nla_put_failure;
4815
4816 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
4817 nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
4818 set->dtype == NFT_DATA_VERDICT ? NFT_DATA_VERDICT : NFT_DATA_VALUE,
4819 set->dlen) < 0)
4820 goto nla_put_failure;
4821
4822 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS) &&
4823 nft_set_elem_expr_dump(skb, set, ext))
4824 goto nla_put_failure;
4825
4826 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
4827 nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
4828 (*nft_set_ext_obj(ext))->key.name) < 0)
4829 goto nla_put_failure;
4830
4831 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
4832 nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
4833 htonl(*nft_set_ext_flags(ext))))
4834 goto nla_put_failure;
4835
4836 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
4837 nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
4838 nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
4839 NFTA_SET_ELEM_PAD))
4840 goto nla_put_failure;
4841
4842 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
4843 u64 expires, now = get_jiffies_64();
4844
4845 expires = *nft_set_ext_expiration(ext);
4846 if (time_before64(now, expires))
4847 expires -= now;
4848 else
4849 expires = 0;
4850
4851 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
4852 nf_jiffies64_to_msecs(expires),
4853 NFTA_SET_ELEM_PAD))
4854 goto nla_put_failure;
4855 }
4856
4857 if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
4858 struct nft_userdata *udata;
4859
4860 udata = nft_set_ext_userdata(ext);
4861 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
4862 udata->len + 1, udata->data))
4863 goto nla_put_failure;
4864 }
4865
4866 nla_nest_end(skb, nest);
4867 return 0;
4868
4869 nla_put_failure:
4870 nlmsg_trim(skb, b);
4871 return -EMSGSIZE;
4872 }
4873
4874 struct nft_set_dump_args {
4875 const struct netlink_callback *cb;
4876 struct nft_set_iter iter;
4877 struct sk_buff *skb;
4878 };
4879
4880 static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
4881 struct nft_set *set,
4882 const struct nft_set_iter *iter,
4883 struct nft_set_elem *elem)
4884 {
4885 struct nft_set_dump_args *args;
4886
4887 args = container_of(iter, struct nft_set_dump_args, iter);
4888 return nf_tables_fill_setelem(args->skb, set, elem);
4889 }
4890
4891 struct nft_set_dump_ctx {
4892 const struct nft_set *set;
4893 struct nft_ctx ctx;
4894 };
4895
4896 static int nft_set_catchall_dump(struct net *net, struct sk_buff *skb,
4897 const struct nft_set *set)
4898 {
4899 struct nft_set_elem_catchall *catchall;
4900 u8 genmask = nft_genmask_cur(net);
4901 struct nft_set_elem elem;
4902 struct nft_set_ext *ext;
4903 int ret = 0;
4904
4905 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
4906 ext = nft_set_elem_ext(set, catchall->elem);
4907 if (!nft_set_elem_active(ext, genmask) ||
4908 nft_set_elem_expired(ext))
4909 continue;
4910
4911 elem.priv = catchall->elem;
4912 ret = nf_tables_fill_setelem(skb, set, &elem);
4913 break;
4914 }
4915
4916 return ret;
4917 }
4918
4919 static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
4920 {
4921 struct nft_set_dump_ctx *dump_ctx = cb->data;
4922 struct net *net = sock_net(skb->sk);
4923 struct nftables_pernet *nft_net;
4924 struct nft_table *table;
4925 struct nft_set *set;
4926 struct nft_set_dump_args args;
4927 bool set_found = false;
4928 struct nlmsghdr *nlh;
4929 struct nlattr *nest;
4930 u32 portid, seq;
4931 int event;
4932
4933 rcu_read_lock();
4934 nft_net = nft_pernet(net);
4935 list_for_each_entry_rcu(table, &nft_net->tables, list) {
4936 if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
4937 dump_ctx->ctx.family != table->family)
4938 continue;
4939
4940 if (table != dump_ctx->ctx.table)
4941 continue;
4942
4943 list_for_each_entry_rcu(set, &table->sets, list) {
4944 if (set == dump_ctx->set) {
4945 set_found = true;
4946 break;
4947 }
4948 }
4949 break;
4950 }
4951
4952 if (!set_found) {
4953 rcu_read_unlock();
4954 return -ENOENT;
4955 }
4956
4957 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
4958 portid = NETLINK_CB(cb->skb).portid;
4959 seq = cb->nlh->nlmsg_seq;
4960
4961 nlh = nfnl_msg_put(skb, portid, seq, event, NLM_F_MULTI,
4962 table->family, NFNETLINK_V0, nft_base_seq(net));
4963 if (!nlh)
4964 goto nla_put_failure;
4965
4966 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
4967 goto nla_put_failure;
4968 if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
4969 goto nla_put_failure;
4970
4971 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
4972 if (nest == NULL)
4973 goto nla_put_failure;
4974
4975 args.cb = cb;
4976 args.skb = skb;
4977 args.iter.genmask = nft_genmask_cur(net);
4978 args.iter.skip = cb->args[0];
4979 args.iter.count = 0;
4980 args.iter.err = 0;
4981 args.iter.fn = nf_tables_dump_setelem;
4982 set->ops->walk(&dump_ctx->ctx, set, &args.iter);
4983
4984 if (!args.iter.err && args.iter.count == cb->args[0])
4985 args.iter.err = nft_set_catchall_dump(net, skb, set);
4986 rcu_read_unlock();
4987
4988 nla_nest_end(skb, nest);
4989 nlmsg_end(skb, nlh);
4990
4991 if (args.iter.err && args.iter.err != -EMSGSIZE)
4992 return args.iter.err;
4993 if (args.iter.count == cb->args[0])
4994 return 0;
4995
4996 cb->args[0] = args.iter.count;
4997 return skb->len;
4998
4999 nla_put_failure:
5000 rcu_read_unlock();
5001 return -ENOSPC;
5002 }
5003
5004 static int nf_tables_dump_set_start(struct netlink_callback *cb)
5005 {
5006 struct nft_set_dump_ctx *dump_ctx = cb->data;
5007
5008 cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
5009
5010 return cb->data ? 0 : -ENOMEM;
5011 }
5012
5013 static int nf_tables_dump_set_done(struct netlink_callback *cb)
5014 {
5015 kfree(cb->data);
5016 return 0;
5017 }
5018
5019 static int nf_tables_fill_setelem_info(struct sk_buff *skb,
5020 const struct nft_ctx *ctx, u32 seq,
5021 u32 portid, int event, u16 flags,
5022 const struct nft_set *set,
5023 const struct nft_set_elem *elem)
5024 {
5025 struct nlmsghdr *nlh;
5026 struct nlattr *nest;
5027 int err;
5028
5029 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
5030 nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
5031 NFNETLINK_V0, nft_base_seq(ctx->net));
5032 if (!nlh)
5033 goto nla_put_failure;
5034
5035 if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
5036 goto nla_put_failure;
5037 if (nla_put_string(skb, NFTA_SET_NAME, set->name))
5038 goto nla_put_failure;
5039
5040 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5041 if (nest == NULL)
5042 goto nla_put_failure;
5043
5044 err = nf_tables_fill_setelem(skb, set, elem);
5045 if (err < 0)
5046 goto nla_put_failure;
5047
5048 nla_nest_end(skb, nest);
5049
5050 nlmsg_end(skb, nlh);
5051 return 0;
5052
5053 nla_put_failure:
5054 nlmsg_trim(skb, nlh);
5055 return -1;
5056 }
5057
5058 static int nft_setelem_parse_flags(const struct nft_set *set,
5059 const struct nlattr *attr, u32 *flags)
5060 {
5061 if (attr == NULL)
5062 return 0;
5063
5064 *flags = ntohl(nla_get_be32(attr));
5065 if (*flags & ~(NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
5066 return -EOPNOTSUPP;
5067 if (!(set->flags & NFT_SET_INTERVAL) &&
5068 *flags & NFT_SET_ELEM_INTERVAL_END)
5069 return -EINVAL;
5070
5071 return 0;
5072 }
5073
5074 static int nft_setelem_parse_key(struct nft_ctx *ctx, struct nft_set *set,
5075 struct nft_data *key, struct nlattr *attr)
5076 {
5077 struct nft_data_desc desc;
5078 int err;
5079
5080 err = nft_data_init(ctx, key, NFT_DATA_VALUE_MAXLEN, &desc, attr);
5081 if (err < 0)
5082 return err;
5083
5084 if (desc.type != NFT_DATA_VALUE || desc.len != set->klen) {
5085 nft_data_release(key, desc.type);
5086 return -EINVAL;
5087 }
5088
5089 return 0;
5090 }
5091
5092 static int nft_setelem_parse_data(struct nft_ctx *ctx, struct nft_set *set,
5093 struct nft_data_desc *desc,
5094 struct nft_data *data,
5095 struct nlattr *attr)
5096 {
5097 int err;
5098
5099 err = nft_data_init(ctx, data, NFT_DATA_VALUE_MAXLEN, desc, attr);
5100 if (err < 0)
5101 return err;
5102
5103 if (desc->type != NFT_DATA_VERDICT && desc->len != set->dlen) {
5104 nft_data_release(data, desc->type);
5105 return -EINVAL;
5106 }
5107
5108 return 0;
5109 }
5110
5111 static void *nft_setelem_catchall_get(const struct net *net,
5112 const struct nft_set *set)
5113 {
5114 struct nft_set_elem_catchall *catchall;
5115 u8 genmask = nft_genmask_cur(net);
5116 struct nft_set_ext *ext;
5117 void *priv = NULL;
5118
5119 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5120 ext = nft_set_elem_ext(set, catchall->elem);
5121 if (!nft_set_elem_active(ext, genmask) ||
5122 nft_set_elem_expired(ext))
5123 continue;
5124
5125 priv = catchall->elem;
5126 break;
5127 }
5128
5129 return priv;
5130 }
5131
5132 static int nft_setelem_get(struct nft_ctx *ctx, struct nft_set *set,
5133 struct nft_set_elem *elem, u32 flags)
5134 {
5135 void *priv;
5136
5137 if (!(flags & NFT_SET_ELEM_CATCHALL)) {
5138 priv = set->ops->get(ctx->net, set, elem, flags);
5139 if (IS_ERR(priv))
5140 return PTR_ERR(priv);
5141 } else {
5142 priv = nft_setelem_catchall_get(ctx->net, set);
5143 if (!priv)
5144 return -ENOENT;
5145 }
5146 elem->priv = priv;
5147
5148 return 0;
5149 }
5150
5151 static int nft_get_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5152 const struct nlattr *attr)
5153 {
5154 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5155 struct nft_set_elem elem;
5156 struct sk_buff *skb;
5157 uint32_t flags = 0;
5158 int err;
5159
5160 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5161 nft_set_elem_policy, NULL);
5162 if (err < 0)
5163 return err;
5164
5165 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5166 if (err < 0)
5167 return err;
5168
5169 if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5170 return -EINVAL;
5171
5172 if (nla[NFTA_SET_ELEM_KEY]) {
5173 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5174 nla[NFTA_SET_ELEM_KEY]);
5175 if (err < 0)
5176 return err;
5177 }
5178
5179 if (nla[NFTA_SET_ELEM_KEY_END]) {
5180 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5181 nla[NFTA_SET_ELEM_KEY_END]);
5182 if (err < 0)
5183 return err;
5184 }
5185
5186 err = nft_setelem_get(ctx, set, &elem, flags);
5187 if (err < 0)
5188 return err;
5189
5190 err = -ENOMEM;
5191 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
5192 if (skb == NULL)
5193 return err;
5194
5195 err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
5196 NFT_MSG_NEWSETELEM, 0, set, &elem);
5197 if (err < 0)
5198 goto err_fill_setelem;
5199
5200 return nfnetlink_unicast(skb, ctx->net, ctx->portid);
5201
5202 err_fill_setelem:
5203 kfree_skb(skb);
5204 return err;
5205 }
5206
5207 /* called with rcu_read_lock held */
5208 static int nf_tables_getsetelem(struct sk_buff *skb,
5209 const struct nfnl_info *info,
5210 const struct nlattr * const nla[])
5211 {
5212 struct netlink_ext_ack *extack = info->extack;
5213 u8 genmask = nft_genmask_cur(info->net);
5214 u8 family = info->nfmsg->nfgen_family;
5215 struct net *net = info->net;
5216 struct nft_table *table;
5217 struct nft_set *set;
5218 struct nlattr *attr;
5219 struct nft_ctx ctx;
5220 int rem, err = 0;
5221
5222 table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
5223 genmask, NETLINK_CB(skb).portid);
5224 if (IS_ERR(table)) {
5225 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
5226 return PTR_ERR(table);
5227 }
5228
5229 set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
5230 if (IS_ERR(set))
5231 return PTR_ERR(set);
5232
5233 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5234
5235 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
5236 struct netlink_dump_control c = {
5237 .start = nf_tables_dump_set_start,
5238 .dump = nf_tables_dump_set,
5239 .done = nf_tables_dump_set_done,
5240 .module = THIS_MODULE,
5241 };
5242 struct nft_set_dump_ctx dump_ctx = {
5243 .set = set,
5244 .ctx = ctx,
5245 };
5246
5247 c.data = &dump_ctx;
5248 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
5249 }
5250
5251 if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
5252 return -EINVAL;
5253
5254 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
5255 err = nft_get_set_elem(&ctx, set, attr);
5256 if (err < 0)
5257 break;
5258 }
5259
5260 return err;
5261 }
5262
5263 static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
5264 const struct nft_set *set,
5265 const struct nft_set_elem *elem,
5266 int event)
5267 {
5268 struct nftables_pernet *nft_net;
5269 struct net *net = ctx->net;
5270 u32 portid = ctx->portid;
5271 struct sk_buff *skb;
5272 u16 flags = 0;
5273 int err;
5274
5275 if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
5276 return;
5277
5278 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
5279 if (skb == NULL)
5280 goto err;
5281
5282 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
5283 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
5284
5285 err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
5286 set, elem);
5287 if (err < 0) {
5288 kfree_skb(skb);
5289 goto err;
5290 }
5291
5292 nft_net = nft_pernet(net);
5293 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
5294 return;
5295 err:
5296 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
5297 }
5298
5299 static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
5300 int msg_type,
5301 struct nft_set *set)
5302 {
5303 struct nft_trans *trans;
5304
5305 trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
5306 if (trans == NULL)
5307 return NULL;
5308
5309 nft_trans_elem_set(trans) = set;
5310 return trans;
5311 }
5312
5313 struct nft_expr *nft_set_elem_expr_alloc(const struct nft_ctx *ctx,
5314 const struct nft_set *set,
5315 const struct nlattr *attr)
5316 {
5317 struct nft_expr *expr;
5318 int err;
5319
5320 expr = nft_expr_init(ctx, attr);
5321 if (IS_ERR(expr))
5322 return expr;
5323
5324 err = -EOPNOTSUPP;
5325 if (expr->ops->type->flags & NFT_EXPR_GC) {
5326 if (set->flags & NFT_SET_TIMEOUT)
5327 goto err_set_elem_expr;
5328 if (!set->ops->gc_init)
5329 goto err_set_elem_expr;
5330 set->ops->gc_init(set);
5331 }
5332
5333 return expr;
5334
5335 err_set_elem_expr:
5336 nft_expr_destroy(ctx, expr);
5337 return ERR_PTR(err);
5338 }
5339
5340 void *nft_set_elem_init(const struct nft_set *set,
5341 const struct nft_set_ext_tmpl *tmpl,
5342 const u32 *key, const u32 *key_end,
5343 const u32 *data, u64 timeout, u64 expiration, gfp_t gfp)
5344 {
5345 struct nft_set_ext *ext;
5346 void *elem;
5347
5348 elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
5349 if (elem == NULL)
5350 return NULL;
5351
5352 ext = nft_set_elem_ext(set, elem);
5353 nft_set_ext_init(ext, tmpl);
5354
5355 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY))
5356 memcpy(nft_set_ext_key(ext), key, set->klen);
5357 if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END))
5358 memcpy(nft_set_ext_key_end(ext), key_end, set->klen);
5359 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5360 memcpy(nft_set_ext_data(ext), data, set->dlen);
5361 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5362 *nft_set_ext_expiration(ext) = get_jiffies_64() + expiration;
5363 if (expiration == 0)
5364 *nft_set_ext_expiration(ext) += timeout;
5365 }
5366 if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
5367 *nft_set_ext_timeout(ext) = timeout;
5368
5369 return elem;
5370 }
5371
5372 static void __nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5373 struct nft_expr *expr)
5374 {
5375 if (expr->ops->destroy_clone) {
5376 expr->ops->destroy_clone(ctx, expr);
5377 module_put(expr->ops->type->owner);
5378 } else {
5379 nf_tables_expr_destroy(ctx, expr);
5380 }
5381 }
5382
5383 static void nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
5384 struct nft_set_elem_expr *elem_expr)
5385 {
5386 struct nft_expr *expr;
5387 u32 size;
5388
5389 nft_setelem_expr_foreach(expr, elem_expr, size)
5390 __nft_set_elem_expr_destroy(ctx, expr);
5391 }
5392
5393 void nft_set_elem_destroy(const struct nft_set *set, void *elem,
5394 bool destroy_expr)
5395 {
5396 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5397 struct nft_ctx ctx = {
5398 .net = read_pnet(&set->net),
5399 .family = set->table->family,
5400 };
5401
5402 nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
5403 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
5404 nft_data_release(nft_set_ext_data(ext), set->dtype);
5405 if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5406 nft_set_elem_expr_destroy(&ctx, nft_set_ext_expr(ext));
5407
5408 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
5409 (*nft_set_ext_obj(ext))->use--;
5410 kfree(elem);
5411 }
5412 EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
5413
5414 /* Only called from commit path, nft_setelem_data_deactivate() already deals
5415 * with the refcounting from the preparation phase.
5416 */
5417 static void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
5418 const struct nft_set *set, void *elem)
5419 {
5420 struct nft_set_ext *ext = nft_set_elem_ext(set, elem);
5421
5422 if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
5423 nft_set_elem_expr_destroy(ctx, nft_set_ext_expr(ext));
5424
5425 kfree(elem);
5426 }
5427
5428 int nft_set_elem_expr_clone(const struct nft_ctx *ctx, struct nft_set *set,
5429 struct nft_expr *expr_array[])
5430 {
5431 struct nft_expr *expr;
5432 int err, i, k;
5433
5434 for (i = 0; i < set->num_exprs; i++) {
5435 expr = kzalloc(set->exprs[i]->ops->size, GFP_KERNEL);
5436 if (!expr)
5437 goto err_expr;
5438
5439 err = nft_expr_clone(expr, set->exprs[i]);
5440 if (err < 0) {
5441 nft_expr_destroy(ctx, expr);
5442 goto err_expr;
5443 }
5444 expr_array[i] = expr;
5445 }
5446
5447 return 0;
5448
5449 err_expr:
5450 for (k = i - 1; k >= 0; k--)
5451 nft_expr_destroy(ctx, expr_array[k]);
5452
5453 return -ENOMEM;
5454 }
5455
5456 static int nft_set_elem_expr_setup(struct nft_ctx *ctx,
5457 const struct nft_set_ext *ext,
5458 struct nft_expr *expr_array[],
5459 u32 num_exprs)
5460 {
5461 struct nft_set_elem_expr *elem_expr = nft_set_ext_expr(ext);
5462 struct nft_expr *expr;
5463 int i, err;
5464
5465 for (i = 0; i < num_exprs; i++) {
5466 expr = nft_setelem_expr_at(elem_expr, elem_expr->size);
5467 err = nft_expr_clone(expr, expr_array[i]);
5468 if (err < 0)
5469 goto err_elem_expr_setup;
5470
5471 elem_expr->size += expr_array[i]->ops->size;
5472 nft_expr_destroy(ctx, expr_array[i]);
5473 expr_array[i] = NULL;
5474 }
5475
5476 return 0;
5477
5478 err_elem_expr_setup:
5479 for (; i < num_exprs; i++) {
5480 nft_expr_destroy(ctx, expr_array[i]);
5481 expr_array[i] = NULL;
5482 }
5483
5484 return -ENOMEM;
5485 }
5486
5487 struct nft_set_ext *nft_set_catchall_lookup(const struct net *net,
5488 const struct nft_set *set)
5489 {
5490 struct nft_set_elem_catchall *catchall;
5491 u8 genmask = nft_genmask_cur(net);
5492 struct nft_set_ext *ext;
5493
5494 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5495 ext = nft_set_elem_ext(set, catchall->elem);
5496 if (nft_set_elem_active(ext, genmask) &&
5497 !nft_set_elem_expired(ext))
5498 return ext;
5499 }
5500
5501 return NULL;
5502 }
5503 EXPORT_SYMBOL_GPL(nft_set_catchall_lookup);
5504
5505 void *nft_set_catchall_gc(const struct nft_set *set)
5506 {
5507 struct nft_set_elem_catchall *catchall, *next;
5508 struct nft_set_ext *ext;
5509 void *elem = NULL;
5510
5511 list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5512 ext = nft_set_elem_ext(set, catchall->elem);
5513
5514 if (!nft_set_elem_expired(ext) ||
5515 nft_set_elem_mark_busy(ext))
5516 continue;
5517
5518 elem = catchall->elem;
5519 list_del_rcu(&catchall->list);
5520 kfree_rcu(catchall, rcu);
5521 break;
5522 }
5523
5524 return elem;
5525 }
5526 EXPORT_SYMBOL_GPL(nft_set_catchall_gc);
5527
5528 static int nft_setelem_catchall_insert(const struct net *net,
5529 struct nft_set *set,
5530 const struct nft_set_elem *elem,
5531 struct nft_set_ext **pext)
5532 {
5533 struct nft_set_elem_catchall *catchall;
5534 u8 genmask = nft_genmask_next(net);
5535 struct nft_set_ext *ext;
5536
5537 list_for_each_entry(catchall, &set->catchall_list, list) {
5538 ext = nft_set_elem_ext(set, catchall->elem);
5539 if (nft_set_elem_active(ext, genmask)) {
5540 *pext = ext;
5541 return -EEXIST;
5542 }
5543 }
5544
5545 catchall = kmalloc(sizeof(*catchall), GFP_KERNEL);
5546 if (!catchall)
5547 return -ENOMEM;
5548
5549 catchall->elem = elem->priv;
5550 list_add_tail_rcu(&catchall->list, &set->catchall_list);
5551
5552 return 0;
5553 }
5554
5555 static int nft_setelem_insert(const struct net *net,
5556 struct nft_set *set,
5557 const struct nft_set_elem *elem,
5558 struct nft_set_ext **ext, unsigned int flags)
5559 {
5560 int ret;
5561
5562 if (flags & NFT_SET_ELEM_CATCHALL)
5563 ret = nft_setelem_catchall_insert(net, set, elem, ext);
5564 else
5565 ret = set->ops->insert(net, set, elem, ext);
5566
5567 return ret;
5568 }
5569
5570 static bool nft_setelem_is_catchall(const struct nft_set *set,
5571 const struct nft_set_elem *elem)
5572 {
5573 struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5574
5575 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
5576 *nft_set_ext_flags(ext) & NFT_SET_ELEM_CATCHALL)
5577 return true;
5578
5579 return false;
5580 }
5581
5582 static void nft_setelem_activate(struct net *net, struct nft_set *set,
5583 struct nft_set_elem *elem)
5584 {
5585 struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
5586
5587 if (nft_setelem_is_catchall(set, elem)) {
5588 nft_set_elem_change_active(net, set, ext);
5589 nft_set_elem_clear_busy(ext);
5590 } else {
5591 set->ops->activate(net, set, elem);
5592 }
5593 }
5594
5595 static int nft_setelem_catchall_deactivate(const struct net *net,
5596 struct nft_set *set,
5597 struct nft_set_elem *elem)
5598 {
5599 struct nft_set_elem_catchall *catchall;
5600 struct nft_set_ext *ext;
5601
5602 list_for_each_entry(catchall, &set->catchall_list, list) {
5603 ext = nft_set_elem_ext(set, catchall->elem);
5604 if (!nft_is_active(net, ext) ||
5605 nft_set_elem_mark_busy(ext))
5606 continue;
5607
5608 kfree(elem->priv);
5609 elem->priv = catchall->elem;
5610 nft_set_elem_change_active(net, set, ext);
5611 return 0;
5612 }
5613
5614 return -ENOENT;
5615 }
5616
5617 static int __nft_setelem_deactivate(const struct net *net,
5618 struct nft_set *set,
5619 struct nft_set_elem *elem)
5620 {
5621 void *priv;
5622
5623 priv = set->ops->deactivate(net, set, elem);
5624 if (!priv)
5625 return -ENOENT;
5626
5627 kfree(elem->priv);
5628 elem->priv = priv;
5629 set->ndeact++;
5630
5631 return 0;
5632 }
5633
5634 static int nft_setelem_deactivate(const struct net *net,
5635 struct nft_set *set,
5636 struct nft_set_elem *elem, u32 flags)
5637 {
5638 int ret;
5639
5640 if (flags & NFT_SET_ELEM_CATCHALL)
5641 ret = nft_setelem_catchall_deactivate(net, set, elem);
5642 else
5643 ret = __nft_setelem_deactivate(net, set, elem);
5644
5645 return ret;
5646 }
5647
5648 static void nft_setelem_catchall_remove(const struct net *net,
5649 const struct nft_set *set,
5650 const struct nft_set_elem *elem)
5651 {
5652 struct nft_set_elem_catchall *catchall, *next;
5653
5654 list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5655 if (catchall->elem == elem->priv) {
5656 list_del_rcu(&catchall->list);
5657 kfree_rcu(catchall, rcu);
5658 break;
5659 }
5660 }
5661 }
5662
5663 static void nft_setelem_remove(const struct net *net,
5664 const struct nft_set *set,
5665 const struct nft_set_elem *elem)
5666 {
5667 if (nft_setelem_is_catchall(set, elem))
5668 nft_setelem_catchall_remove(net, set, elem);
5669 else
5670 set->ops->remove(net, set, elem);
5671 }
5672
5673 static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
5674 const struct nlattr *attr, u32 nlmsg_flags)
5675 {
5676 struct nft_expr *expr_array[NFT_SET_EXPR_MAX] = {};
5677 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
5678 u8 genmask = nft_genmask_next(ctx->net);
5679 u32 flags = 0, size = 0, num_exprs = 0;
5680 struct nft_set_ext_tmpl tmpl;
5681 struct nft_set_ext *ext, *ext2;
5682 struct nft_set_elem elem;
5683 struct nft_set_binding *binding;
5684 struct nft_object *obj = NULL;
5685 struct nft_userdata *udata;
5686 struct nft_data_desc desc;
5687 enum nft_registers dreg;
5688 struct nft_trans *trans;
5689 u64 timeout;
5690 u64 expiration;
5691 int err, i;
5692 u8 ulen;
5693
5694 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
5695 nft_set_elem_policy, NULL);
5696 if (err < 0)
5697 return err;
5698
5699 nft_set_ext_prepare(&tmpl);
5700
5701 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
5702 if (err < 0)
5703 return err;
5704
5705 if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
5706 return -EINVAL;
5707
5708 if (flags != 0)
5709 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
5710
5711 if (set->flags & NFT_SET_MAP) {
5712 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
5713 !(flags & NFT_SET_ELEM_INTERVAL_END))
5714 return -EINVAL;
5715 } else {
5716 if (nla[NFTA_SET_ELEM_DATA] != NULL)
5717 return -EINVAL;
5718 }
5719
5720 if ((flags & NFT_SET_ELEM_INTERVAL_END) &&
5721 (nla[NFTA_SET_ELEM_DATA] ||
5722 nla[NFTA_SET_ELEM_OBJREF] ||
5723 nla[NFTA_SET_ELEM_TIMEOUT] ||
5724 nla[NFTA_SET_ELEM_EXPIRATION] ||
5725 nla[NFTA_SET_ELEM_USERDATA] ||
5726 nla[NFTA_SET_ELEM_EXPR] ||
5727 nla[NFTA_SET_ELEM_EXPRESSIONS]))
5728 return -EINVAL;
5729
5730 timeout = 0;
5731 if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
5732 if (!(set->flags & NFT_SET_TIMEOUT))
5733 return -EINVAL;
5734 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
5735 &timeout);
5736 if (err)
5737 return err;
5738 } else if (set->flags & NFT_SET_TIMEOUT) {
5739 timeout = set->timeout;
5740 }
5741
5742 expiration = 0;
5743 if (nla[NFTA_SET_ELEM_EXPIRATION] != NULL) {
5744 if (!(set->flags & NFT_SET_TIMEOUT))
5745 return -EINVAL;
5746 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_EXPIRATION],
5747 &expiration);
5748 if (err)
5749 return err;
5750 }
5751
5752 if (nla[NFTA_SET_ELEM_EXPR]) {
5753 struct nft_expr *expr;
5754
5755 if (set->num_exprs && set->num_exprs != 1)
5756 return -EOPNOTSUPP;
5757
5758 expr = nft_set_elem_expr_alloc(ctx, set,
5759 nla[NFTA_SET_ELEM_EXPR]);
5760 if (IS_ERR(expr))
5761 return PTR_ERR(expr);
5762
5763 expr_array[0] = expr;
5764 num_exprs = 1;
5765
5766 if (set->num_exprs && set->exprs[0]->ops != expr->ops) {
5767 err = -EOPNOTSUPP;
5768 goto err_set_elem_expr;
5769 }
5770 } else if (nla[NFTA_SET_ELEM_EXPRESSIONS]) {
5771 struct nft_expr *expr;
5772 struct nlattr *tmp;
5773 int left;
5774
5775 i = 0;
5776 nla_for_each_nested(tmp, nla[NFTA_SET_ELEM_EXPRESSIONS], left) {
5777 if (i == NFT_SET_EXPR_MAX ||
5778 (set->num_exprs && set->num_exprs == i)) {
5779 err = -E2BIG;
5780 goto err_set_elem_expr;
5781 }
5782 if (nla_type(tmp) != NFTA_LIST_ELEM) {
5783 err = -EINVAL;
5784 goto err_set_elem_expr;
5785 }
5786 expr = nft_set_elem_expr_alloc(ctx, set, tmp);
5787 if (IS_ERR(expr)) {
5788 err = PTR_ERR(expr);
5789 goto err_set_elem_expr;
5790 }
5791 expr_array[i] = expr;
5792 num_exprs++;
5793
5794 if (set->num_exprs && expr->ops != set->exprs[i]->ops) {
5795 err = -EOPNOTSUPP;
5796 goto err_set_elem_expr;
5797 }
5798 i++;
5799 }
5800 if (set->num_exprs && set->num_exprs != i) {
5801 err = -EOPNOTSUPP;
5802 goto err_set_elem_expr;
5803 }
5804 } else if (set->num_exprs > 0) {
5805 err = nft_set_elem_expr_clone(ctx, set, expr_array);
5806 if (err < 0)
5807 goto err_set_elem_expr_clone;
5808
5809 num_exprs = set->num_exprs;
5810 }
5811
5812 if (nla[NFTA_SET_ELEM_KEY]) {
5813 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
5814 nla[NFTA_SET_ELEM_KEY]);
5815 if (err < 0)
5816 goto err_set_elem_expr;
5817
5818 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
5819 }
5820
5821 if (nla[NFTA_SET_ELEM_KEY_END]) {
5822 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
5823 nla[NFTA_SET_ELEM_KEY_END]);
5824 if (err < 0)
5825 goto err_parse_key;
5826
5827 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
5828 }
5829
5830 if (timeout > 0) {
5831 nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
5832 if (timeout != set->timeout)
5833 nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
5834 }
5835
5836 if (num_exprs) {
5837 for (i = 0; i < num_exprs; i++)
5838 size += expr_array[i]->ops->size;
5839
5840 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_EXPRESSIONS,
5841 sizeof(struct nft_set_elem_expr) +
5842 size);
5843 }
5844
5845 if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
5846 if (!(set->flags & NFT_SET_OBJECT)) {
5847 err = -EINVAL;
5848 goto err_parse_key_end;
5849 }
5850 obj = nft_obj_lookup(ctx->net, ctx->table,
5851 nla[NFTA_SET_ELEM_OBJREF],
5852 set->objtype, genmask);
5853 if (IS_ERR(obj)) {
5854 err = PTR_ERR(obj);
5855 goto err_parse_key_end;
5856 }
5857 nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
5858 }
5859
5860 if (nla[NFTA_SET_ELEM_DATA] != NULL) {
5861 err = nft_setelem_parse_data(ctx, set, &desc, &elem.data.val,
5862 nla[NFTA_SET_ELEM_DATA]);
5863 if (err < 0)
5864 goto err_parse_key_end;
5865
5866 dreg = nft_type_to_reg(set->dtype);
5867 list_for_each_entry(binding, &set->bindings, list) {
5868 struct nft_ctx bind_ctx = {
5869 .net = ctx->net,
5870 .family = ctx->family,
5871 .table = ctx->table,
5872 .chain = (struct nft_chain *)binding->chain,
5873 };
5874
5875 if (!(binding->flags & NFT_SET_MAP))
5876 continue;
5877
5878 err = nft_validate_register_store(&bind_ctx, dreg,
5879 &elem.data.val,
5880 desc.type, desc.len);
5881 if (err < 0)
5882 goto err_parse_data;
5883
5884 if (desc.type == NFT_DATA_VERDICT &&
5885 (elem.data.val.verdict.code == NFT_GOTO ||
5886 elem.data.val.verdict.code == NFT_JUMP))
5887 nft_validate_state_update(ctx->net,
5888 NFT_VALIDATE_NEED);
5889 }
5890
5891 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, desc.len);
5892 }
5893
5894 /* The full maximum length of userdata can exceed the maximum
5895 * offset value (U8_MAX) for following extensions, therefor it
5896 * must be the last extension added.
5897 */
5898 ulen = 0;
5899 if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
5900 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
5901 if (ulen > 0)
5902 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
5903 ulen);
5904 }
5905
5906 err = -ENOMEM;
5907 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
5908 elem.key_end.val.data, elem.data.val.data,
5909 timeout, expiration, GFP_KERNEL);
5910 if (elem.priv == NULL)
5911 goto err_parse_data;
5912
5913 ext = nft_set_elem_ext(set, elem.priv);
5914 if (flags)
5915 *nft_set_ext_flags(ext) = flags;
5916 if (ulen > 0) {
5917 udata = nft_set_ext_userdata(ext);
5918 udata->len = ulen - 1;
5919 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
5920 }
5921 if (obj) {
5922 *nft_set_ext_obj(ext) = obj;
5923 obj->use++;
5924 }
5925 err = nft_set_elem_expr_setup(ctx, ext, expr_array, num_exprs);
5926 if (err < 0)
5927 goto err_elem_expr;
5928
5929 trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
5930 if (trans == NULL) {
5931 err = -ENOMEM;
5932 goto err_elem_expr;
5933 }
5934
5935 ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
5936
5937 err = nft_setelem_insert(ctx->net, set, &elem, &ext2, flags);
5938 if (err) {
5939 if (err == -EEXIST) {
5940 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
5941 nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
5942 nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
5943 nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF))
5944 goto err_element_clash;
5945 if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5946 nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
5947 memcmp(nft_set_ext_data(ext),
5948 nft_set_ext_data(ext2), set->dlen) != 0) ||
5949 (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
5950 nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
5951 *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
5952 goto err_element_clash;
5953 else if (!(nlmsg_flags & NLM_F_EXCL))
5954 err = 0;
5955 } else if (err == -ENOTEMPTY) {
5956 /* ENOTEMPTY reports overlapping between this element
5957 * and an existing one.
5958 */
5959 err = -EEXIST;
5960 }
5961 goto err_element_clash;
5962 }
5963
5964 if (!(flags & NFT_SET_ELEM_CATCHALL) && set->size &&
5965 !atomic_add_unless(&set->nelems, 1, set->size + set->ndeact)) {
5966 err = -ENFILE;
5967 goto err_set_full;
5968 }
5969
5970 nft_trans_elem(trans) = elem;
5971 nft_trans_commit_list_add_tail(ctx->net, trans);
5972 return 0;
5973
5974 err_set_full:
5975 nft_setelem_remove(ctx->net, set, &elem);
5976 err_element_clash:
5977 kfree(trans);
5978 err_elem_expr:
5979 if (obj)
5980 obj->use--;
5981
5982 nf_tables_set_elem_destroy(ctx, set, elem.priv);
5983 err_parse_data:
5984 if (nla[NFTA_SET_ELEM_DATA] != NULL)
5985 nft_data_release(&elem.data.val, desc.type);
5986 err_parse_key_end:
5987 nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
5988 err_parse_key:
5989 nft_data_release(&elem.key.val, NFT_DATA_VALUE);
5990 err_set_elem_expr:
5991 for (i = 0; i < num_exprs && expr_array[i]; i++)
5992 nft_expr_destroy(ctx, expr_array[i]);
5993 err_set_elem_expr_clone:
5994 return err;
5995 }
5996
5997 static int nf_tables_newsetelem(struct sk_buff *skb,
5998 const struct nfnl_info *info,
5999 const struct nlattr * const nla[])
6000 {
6001 struct nftables_pernet *nft_net = nft_pernet(info->net);
6002 struct netlink_ext_ack *extack = info->extack;
6003 u8 genmask = nft_genmask_next(info->net);
6004 u8 family = info->nfmsg->nfgen_family;
6005 struct net *net = info->net;
6006 const struct nlattr *attr;
6007 struct nft_table *table;
6008 struct nft_set *set;
6009 struct nft_ctx ctx;
6010 int rem, err;
6011
6012 if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
6013 return -EINVAL;
6014
6015 table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6016 genmask, NETLINK_CB(skb).portid);
6017 if (IS_ERR(table)) {
6018 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6019 return PTR_ERR(table);
6020 }
6021
6022 set = nft_set_lookup_global(net, table, nla[NFTA_SET_ELEM_LIST_SET],
6023 nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
6024 if (IS_ERR(set))
6025 return PTR_ERR(set);
6026
6027 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6028 return -EBUSY;
6029
6030 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6031
6032 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6033 err = nft_add_set_elem(&ctx, set, attr, info->nlh->nlmsg_flags);
6034 if (err < 0)
6035 return err;
6036 }
6037
6038 if (nft_net->validate_state == NFT_VALIDATE_DO)
6039 return nft_table_validate(net, table);
6040
6041 return 0;
6042 }
6043
6044 /**
6045 * nft_data_hold - hold a nft_data item
6046 *
6047 * @data: struct nft_data to release
6048 * @type: type of data
6049 *
6050 * Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
6051 * NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
6052 * NFT_GOTO verdicts. This function must be called on active data objects
6053 * from the second phase of the commit protocol.
6054 */
6055 void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
6056 {
6057 struct nft_chain *chain;
6058 struct nft_rule *rule;
6059
6060 if (type == NFT_DATA_VERDICT) {
6061 switch (data->verdict.code) {
6062 case NFT_JUMP:
6063 case NFT_GOTO:
6064 chain = data->verdict.chain;
6065 chain->use++;
6066
6067 if (!nft_chain_is_bound(chain))
6068 break;
6069
6070 chain->table->use++;
6071 list_for_each_entry(rule, &chain->rules, list)
6072 chain->use++;
6073
6074 nft_chain_add(chain->table, chain);
6075 break;
6076 }
6077 }
6078 }
6079
6080 static void nft_setelem_data_activate(const struct net *net,
6081 const struct nft_set *set,
6082 struct nft_set_elem *elem)
6083 {
6084 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6085
6086 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6087 nft_data_hold(nft_set_ext_data(ext), set->dtype);
6088 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6089 (*nft_set_ext_obj(ext))->use++;
6090 }
6091
6092 static void nft_setelem_data_deactivate(const struct net *net,
6093 const struct nft_set *set,
6094 struct nft_set_elem *elem)
6095 {
6096 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
6097
6098 if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6099 nft_data_release(nft_set_ext_data(ext), set->dtype);
6100 if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6101 (*nft_set_ext_obj(ext))->use--;
6102 }
6103
6104 static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
6105 const struct nlattr *attr)
6106 {
6107 struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6108 struct nft_set_ext_tmpl tmpl;
6109 struct nft_set_elem elem;
6110 struct nft_set_ext *ext;
6111 struct nft_trans *trans;
6112 u32 flags = 0;
6113 int err;
6114
6115 err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6116 nft_set_elem_policy, NULL);
6117 if (err < 0)
6118 return err;
6119
6120 err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6121 if (err < 0)
6122 return err;
6123
6124 if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
6125 return -EINVAL;
6126
6127 nft_set_ext_prepare(&tmpl);
6128
6129 if (flags != 0)
6130 nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
6131
6132 if (nla[NFTA_SET_ELEM_KEY]) {
6133 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6134 nla[NFTA_SET_ELEM_KEY]);
6135 if (err < 0)
6136 return err;
6137
6138 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6139 }
6140
6141 if (nla[NFTA_SET_ELEM_KEY_END]) {
6142 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6143 nla[NFTA_SET_ELEM_KEY_END]);
6144 if (err < 0)
6145 return err;
6146
6147 nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
6148 }
6149
6150 err = -ENOMEM;
6151 elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
6152 elem.key_end.val.data, NULL, 0, 0,
6153 GFP_KERNEL);
6154 if (elem.priv == NULL)
6155 goto fail_elem;
6156
6157 ext = nft_set_elem_ext(set, elem.priv);
6158 if (flags)
6159 *nft_set_ext_flags(ext) = flags;
6160
6161 trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
6162 if (trans == NULL)
6163 goto fail_trans;
6164
6165 err = nft_setelem_deactivate(ctx->net, set, &elem, flags);
6166 if (err < 0)
6167 goto fail_ops;
6168
6169 nft_setelem_data_deactivate(ctx->net, set, &elem);
6170
6171 nft_trans_elem(trans) = elem;
6172 nft_trans_commit_list_add_tail(ctx->net, trans);
6173 return 0;
6174
6175 fail_ops:
6176 kfree(trans);
6177 fail_trans:
6178 kfree(elem.priv);
6179 fail_elem:
6180 nft_data_release(&elem.key.val, NFT_DATA_VALUE);
6181 return err;
6182 }
6183
6184 static int nft_setelem_flush(const struct nft_ctx *ctx,
6185 struct nft_set *set,
6186 const struct nft_set_iter *iter,
6187 struct nft_set_elem *elem)
6188 {
6189 struct nft_trans *trans;
6190 int err;
6191
6192 trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6193 sizeof(struct nft_trans_elem), GFP_ATOMIC);
6194 if (!trans)
6195 return -ENOMEM;
6196
6197 if (!set->ops->flush(ctx->net, set, elem->priv)) {
6198 err = -ENOENT;
6199 goto err1;
6200 }
6201 set->ndeact++;
6202
6203 nft_setelem_data_deactivate(ctx->net, set, elem);
6204 nft_trans_elem_set(trans) = set;
6205 nft_trans_elem(trans) = *elem;
6206 nft_trans_commit_list_add_tail(ctx->net, trans);
6207
6208 return 0;
6209 err1:
6210 kfree(trans);
6211 return err;
6212 }
6213
6214 static int __nft_set_catchall_flush(const struct nft_ctx *ctx,
6215 struct nft_set *set,
6216 struct nft_set_elem *elem)
6217 {
6218 struct nft_trans *trans;
6219
6220 trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
6221 sizeof(struct nft_trans_elem), GFP_KERNEL);
6222 if (!trans)
6223 return -ENOMEM;
6224
6225 nft_setelem_data_deactivate(ctx->net, set, elem);
6226 nft_trans_elem_set(trans) = set;
6227 nft_trans_elem(trans) = *elem;
6228 nft_trans_commit_list_add_tail(ctx->net, trans);
6229
6230 return 0;
6231 }
6232
6233 static int nft_set_catchall_flush(const struct nft_ctx *ctx,
6234 struct nft_set *set)
6235 {
6236 u8 genmask = nft_genmask_next(ctx->net);
6237 struct nft_set_elem_catchall *catchall;
6238 struct nft_set_elem elem;
6239 struct nft_set_ext *ext;
6240 int ret = 0;
6241
6242 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6243 ext = nft_set_elem_ext(set, catchall->elem);
6244 if (!nft_set_elem_active(ext, genmask) ||
6245 nft_set_elem_mark_busy(ext))
6246 continue;
6247
6248 elem.priv = catchall->elem;
6249 ret = __nft_set_catchall_flush(ctx, set, &elem);
6250 if (ret < 0)
6251 break;
6252 }
6253
6254 return ret;
6255 }
6256
6257 static int nft_set_flush(struct nft_ctx *ctx, struct nft_set *set, u8 genmask)
6258 {
6259 struct nft_set_iter iter = {
6260 .genmask = genmask,
6261 .fn = nft_setelem_flush,
6262 };
6263
6264 set->ops->walk(ctx, set, &iter);
6265 if (!iter.err)
6266 iter.err = nft_set_catchall_flush(ctx, set);
6267
6268 return iter.err;
6269 }
6270
6271 static int nf_tables_delsetelem(struct sk_buff *skb,
6272 const struct nfnl_info *info,
6273 const struct nlattr * const nla[])
6274 {
6275 struct netlink_ext_ack *extack = info->extack;
6276 u8 genmask = nft_genmask_next(info->net);
6277 u8 family = info->nfmsg->nfgen_family;
6278 struct net *net = info->net;
6279 const struct nlattr *attr;
6280 struct nft_table *table;
6281 struct nft_set *set;
6282 struct nft_ctx ctx;
6283 int rem, err = 0;
6284
6285 table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6286 genmask, NETLINK_CB(skb).portid);
6287 if (IS_ERR(table)) {
6288 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6289 return PTR_ERR(table);
6290 }
6291
6292 set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
6293 if (IS_ERR(set))
6294 return PTR_ERR(set);
6295 if (!list_empty(&set->bindings) && set->flags & NFT_SET_CONSTANT)
6296 return -EBUSY;
6297
6298 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6299
6300 if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
6301 return nft_set_flush(&ctx, set, genmask);
6302
6303 nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6304 err = nft_del_setelem(&ctx, set, attr);
6305 if (err < 0)
6306 break;
6307 }
6308 return err;
6309 }
6310
6311 void nft_set_gc_batch_release(struct rcu_head *rcu)
6312 {
6313 struct nft_set_gc_batch *gcb;
6314 unsigned int i;
6315
6316 gcb = container_of(rcu, struct nft_set_gc_batch, head.rcu);
6317 for (i = 0; i < gcb->head.cnt; i++)
6318 nft_set_elem_destroy(gcb->head.set, gcb->elems[i], true);
6319 kfree(gcb);
6320 }
6321
6322 struct nft_set_gc_batch *nft_set_gc_batch_alloc(const struct nft_set *set,
6323 gfp_t gfp)
6324 {
6325 struct nft_set_gc_batch *gcb;
6326
6327 gcb = kzalloc(sizeof(*gcb), gfp);
6328 if (gcb == NULL)
6329 return gcb;
6330 gcb->head.set = set;
6331 return gcb;
6332 }
6333
6334 /*
6335 * Stateful objects
6336 */
6337
6338 /**
6339 * nft_register_obj- register nf_tables stateful object type
6340 * @obj_type: object type
6341 *
6342 * Registers the object type for use with nf_tables. Returns zero on
6343 * success or a negative errno code otherwise.
6344 */
6345 int nft_register_obj(struct nft_object_type *obj_type)
6346 {
6347 if (obj_type->type == NFT_OBJECT_UNSPEC)
6348 return -EINVAL;
6349
6350 nfnl_lock(NFNL_SUBSYS_NFTABLES);
6351 list_add_rcu(&obj_type->list, &nf_tables_objects);
6352 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6353 return 0;
6354 }
6355 EXPORT_SYMBOL_GPL(nft_register_obj);
6356
6357 /**
6358 * nft_unregister_obj - unregister nf_tables object type
6359 * @obj_type: object type
6360 *
6361 * Unregisters the object type for use with nf_tables.
6362 */
6363 void nft_unregister_obj(struct nft_object_type *obj_type)
6364 {
6365 nfnl_lock(NFNL_SUBSYS_NFTABLES);
6366 list_del_rcu(&obj_type->list);
6367 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
6368 }
6369 EXPORT_SYMBOL_GPL(nft_unregister_obj);
6370
6371 struct nft_object *nft_obj_lookup(const struct net *net,
6372 const struct nft_table *table,
6373 const struct nlattr *nla, u32 objtype,
6374 u8 genmask)
6375 {
6376 struct nft_object_hash_key k = { .table = table };
6377 char search[NFT_OBJ_MAXNAMELEN];
6378 struct rhlist_head *tmp, *list;
6379 struct nft_object *obj;
6380
6381 nla_strscpy(search, nla, sizeof(search));
6382 k.name = search;
6383
6384 WARN_ON_ONCE(!rcu_read_lock_held() &&
6385 !lockdep_commit_lock_is_held(net));
6386
6387 rcu_read_lock();
6388 list = rhltable_lookup(&nft_objname_ht, &k, nft_objname_ht_params);
6389 if (!list)
6390 goto out;
6391
6392 rhl_for_each_entry_rcu(obj, tmp, list, rhlhead) {
6393 if (objtype == obj->ops->type->type &&
6394 nft_active_genmask(obj, genmask)) {
6395 rcu_read_unlock();
6396 return obj;
6397 }
6398 }
6399 out:
6400 rcu_read_unlock();
6401 return ERR_PTR(-ENOENT);
6402 }
6403 EXPORT_SYMBOL_GPL(nft_obj_lookup);
6404
6405 static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
6406 const struct nlattr *nla,
6407 u32 objtype, u8 genmask)
6408 {
6409 struct nft_object *obj;
6410
6411 list_for_each_entry(obj, &table->objects, list) {
6412 if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
6413 objtype == obj->ops->type->type &&
6414 nft_active_genmask(obj, genmask))
6415 return obj;
6416 }
6417 return ERR_PTR(-ENOENT);
6418 }
6419
6420 static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
6421 [NFTA_OBJ_TABLE] = { .type = NLA_STRING,
6422 .len = NFT_TABLE_MAXNAMELEN - 1 },
6423 [NFTA_OBJ_NAME] = { .type = NLA_STRING,
6424 .len = NFT_OBJ_MAXNAMELEN - 1 },
6425 [NFTA_OBJ_TYPE] = { .type = NLA_U32 },
6426 [NFTA_OBJ_DATA] = { .type = NLA_NESTED },
6427 [NFTA_OBJ_HANDLE] = { .type = NLA_U64},
6428 [NFTA_OBJ_USERDATA] = { .type = NLA_BINARY,
6429 .len = NFT_USERDATA_MAXLEN },
6430 };
6431
6432 static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
6433 const struct nft_object_type *type,
6434 const struct nlattr *attr)
6435 {
6436 struct nlattr **tb;
6437 const struct nft_object_ops *ops;
6438 struct nft_object *obj;
6439 int err = -ENOMEM;
6440
6441 tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
6442 if (!tb)
6443 goto err1;
6444
6445 if (attr) {
6446 err = nla_parse_nested_deprecated(tb, type->maxattr, attr,
6447 type->policy, NULL);
6448 if (err < 0)
6449 goto err2;
6450 } else {
6451 memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
6452 }
6453
6454 if (type->select_ops) {
6455 ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
6456 if (IS_ERR(ops)) {
6457 err = PTR_ERR(ops);
6458 goto err2;
6459 }
6460 } else {
6461 ops = type->ops;
6462 }
6463
6464 err = -ENOMEM;
6465 obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL);
6466 if (!obj)
6467 goto err2;
6468
6469 err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
6470 if (err < 0)
6471 goto err3;
6472
6473 obj->ops = ops;
6474
6475 kfree(tb);
6476 return obj;
6477 err3:
6478 kfree(obj);
6479 err2:
6480 kfree(tb);
6481 err1:
6482 return ERR_PTR(err);
6483 }
6484
6485 static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
6486 struct nft_object *obj, bool reset)
6487 {
6488 struct nlattr *nest;
6489
6490 nest = nla_nest_start_noflag(skb, attr);
6491 if (!nest)
6492 goto nla_put_failure;
6493 if (obj->ops->dump(skb, obj, reset) < 0)
6494 goto nla_put_failure;
6495 nla_nest_end(skb, nest);
6496 return 0;
6497
6498 nla_put_failure:
6499 return -1;
6500 }
6501
6502 static const struct nft_object_type *__nft_obj_type_get(u32 objtype)
6503 {
6504 const struct nft_object_type *type;
6505
6506 list_for_each_entry(type, &nf_tables_objects, list) {
6507 if (objtype == type->type)
6508 return type;
6509 }
6510 return NULL;
6511 }
6512
6513 static const struct nft_object_type *
6514 nft_obj_type_get(struct net *net, u32 objtype)
6515 {
6516 const struct nft_object_type *type;
6517
6518 type = __nft_obj_type_get(objtype);
6519 if (type != NULL && try_module_get(type->owner))
6520 return type;
6521
6522 lockdep_nfnl_nft_mutex_not_held();
6523 #ifdef CONFIG_MODULES
6524 if (type == NULL) {
6525 if (nft_request_module(net, "nft-obj-%u", objtype) == -EAGAIN)
6526 return ERR_PTR(-EAGAIN);
6527 }
6528 #endif
6529 return ERR_PTR(-ENOENT);
6530 }
6531
6532 static int nf_tables_updobj(const struct nft_ctx *ctx,
6533 const struct nft_object_type *type,
6534 const struct nlattr *attr,
6535 struct nft_object *obj)
6536 {
6537 struct nft_object *newobj;
6538 struct nft_trans *trans;
6539 int err = -ENOMEM;
6540
6541 if (!try_module_get(type->owner))
6542 return -ENOENT;
6543
6544 trans = nft_trans_alloc(ctx, NFT_MSG_NEWOBJ,
6545 sizeof(struct nft_trans_obj));
6546 if (!trans)
6547 goto err_trans;
6548
6549 newobj = nft_obj_init(ctx, type, attr);
6550 if (IS_ERR(newobj)) {
6551 err = PTR_ERR(newobj);
6552 goto err_free_trans;
6553 }
6554
6555 nft_trans_obj(trans) = obj;
6556 nft_trans_obj_update(trans) = true;
6557 nft_trans_obj_newobj(trans) = newobj;
6558 nft_trans_commit_list_add_tail(ctx->net, trans);
6559
6560 return 0;
6561
6562 err_free_trans:
6563 kfree(trans);
6564 err_trans:
6565 module_put(type->owner);
6566 return err;
6567 }
6568
6569 static int nf_tables_newobj(struct sk_buff *skb, const struct nfnl_info *info,
6570 const struct nlattr * const nla[])
6571 {
6572 struct netlink_ext_ack *extack = info->extack;
6573 u8 genmask = nft_genmask_next(info->net);
6574 u8 family = info->nfmsg->nfgen_family;
6575 const struct nft_object_type *type;
6576 struct net *net = info->net;
6577 struct nft_table *table;
6578 struct nft_object *obj;
6579 struct nft_ctx ctx;
6580 u32 objtype;
6581 int err;
6582
6583 if (!nla[NFTA_OBJ_TYPE] ||
6584 !nla[NFTA_OBJ_NAME] ||
6585 !nla[NFTA_OBJ_DATA])
6586 return -EINVAL;
6587
6588 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
6589 NETLINK_CB(skb).portid);
6590 if (IS_ERR(table)) {
6591 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6592 return PTR_ERR(table);
6593 }
6594
6595 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6596 obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6597 if (IS_ERR(obj)) {
6598 err = PTR_ERR(obj);
6599 if (err != -ENOENT) {
6600 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6601 return err;
6602 }
6603 } else {
6604 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
6605 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6606 return -EEXIST;
6607 }
6608 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
6609 return -EOPNOTSUPP;
6610
6611 type = __nft_obj_type_get(objtype);
6612 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6613
6614 return nf_tables_updobj(&ctx, type, nla[NFTA_OBJ_DATA], obj);
6615 }
6616
6617 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6618
6619 type = nft_obj_type_get(net, objtype);
6620 if (IS_ERR(type))
6621 return PTR_ERR(type);
6622
6623 obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
6624 if (IS_ERR(obj)) {
6625 err = PTR_ERR(obj);
6626 goto err_init;
6627 }
6628 obj->key.table = table;
6629 obj->handle = nf_tables_alloc_handle(table);
6630
6631 obj->key.name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL);
6632 if (!obj->key.name) {
6633 err = -ENOMEM;
6634 goto err_strdup;
6635 }
6636
6637 if (nla[NFTA_OBJ_USERDATA]) {
6638 obj->udata = nla_memdup(nla[NFTA_OBJ_USERDATA], GFP_KERNEL);
6639 if (obj->udata == NULL)
6640 goto err_userdata;
6641
6642 obj->udlen = nla_len(nla[NFTA_OBJ_USERDATA]);
6643 }
6644
6645 err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
6646 if (err < 0)
6647 goto err_trans;
6648
6649 err = rhltable_insert(&nft_objname_ht, &obj->rhlhead,
6650 nft_objname_ht_params);
6651 if (err < 0)
6652 goto err_obj_ht;
6653
6654 list_add_tail_rcu(&obj->list, &table->objects);
6655 table->use++;
6656 return 0;
6657 err_obj_ht:
6658 /* queued in transaction log */
6659 INIT_LIST_HEAD(&obj->list);
6660 return err;
6661 err_trans:
6662 kfree(obj->udata);
6663 err_userdata:
6664 kfree(obj->key.name);
6665 err_strdup:
6666 if (obj->ops->destroy)
6667 obj->ops->destroy(&ctx, obj);
6668 kfree(obj);
6669 err_init:
6670 module_put(type->owner);
6671 return err;
6672 }
6673
6674 static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
6675 u32 portid, u32 seq, int event, u32 flags,
6676 int family, const struct nft_table *table,
6677 struct nft_object *obj, bool reset)
6678 {
6679 struct nlmsghdr *nlh;
6680
6681 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
6682 nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
6683 NFNETLINK_V0, nft_base_seq(net));
6684 if (!nlh)
6685 goto nla_put_failure;
6686
6687 if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
6688 nla_put_string(skb, NFTA_OBJ_NAME, obj->key.name) ||
6689 nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
6690 nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
6691 nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset) ||
6692 nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
6693 NFTA_OBJ_PAD))
6694 goto nla_put_failure;
6695
6696 if (obj->udata &&
6697 nla_put(skb, NFTA_OBJ_USERDATA, obj->udlen, obj->udata))
6698 goto nla_put_failure;
6699
6700 nlmsg_end(skb, nlh);
6701 return 0;
6702
6703 nla_put_failure:
6704 nlmsg_trim(skb, nlh);
6705 return -1;
6706 }
6707
6708 struct nft_obj_filter {
6709 char *table;
6710 u32 type;
6711 };
6712
6713 static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
6714 {
6715 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
6716 const struct nft_table *table;
6717 unsigned int idx = 0, s_idx = cb->args[0];
6718 struct nft_obj_filter *filter = cb->data;
6719 struct net *net = sock_net(skb->sk);
6720 int family = nfmsg->nfgen_family;
6721 struct nftables_pernet *nft_net;
6722 struct nft_object *obj;
6723 bool reset = false;
6724
6725 if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6726 reset = true;
6727
6728 rcu_read_lock();
6729 nft_net = nft_pernet(net);
6730 cb->seq = nft_net->base_seq;
6731
6732 list_for_each_entry_rcu(table, &nft_net->tables, list) {
6733 if (family != NFPROTO_UNSPEC && family != table->family)
6734 continue;
6735
6736 list_for_each_entry_rcu(obj, &table->objects, list) {
6737 if (!nft_is_active(net, obj))
6738 goto cont;
6739 if (idx < s_idx)
6740 goto cont;
6741 if (idx > s_idx)
6742 memset(&cb->args[1], 0,
6743 sizeof(cb->args) - sizeof(cb->args[0]));
6744 if (filter && filter->table &&
6745 strcmp(filter->table, table->name))
6746 goto cont;
6747 if (filter &&
6748 filter->type != NFT_OBJECT_UNSPEC &&
6749 obj->ops->type->type != filter->type)
6750 goto cont;
6751 if (reset) {
6752 char *buf = kasprintf(GFP_ATOMIC,
6753 "%s:%u",
6754 table->name,
6755 nft_net->base_seq);
6756
6757 audit_log_nfcfg(buf,
6758 family,
6759 obj->handle,
6760 AUDIT_NFT_OP_OBJ_RESET,
6761 GFP_ATOMIC);
6762 kfree(buf);
6763 }
6764
6765 if (nf_tables_fill_obj_info(skb, net, NETLINK_CB(cb->skb).portid,
6766 cb->nlh->nlmsg_seq,
6767 NFT_MSG_NEWOBJ,
6768 NLM_F_MULTI | NLM_F_APPEND,
6769 table->family, table,
6770 obj, reset) < 0)
6771 goto done;
6772
6773 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
6774 cont:
6775 idx++;
6776 }
6777 }
6778 done:
6779 rcu_read_unlock();
6780
6781 cb->args[0] = idx;
6782 return skb->len;
6783 }
6784
6785 static int nf_tables_dump_obj_start(struct netlink_callback *cb)
6786 {
6787 const struct nlattr * const *nla = cb->data;
6788 struct nft_obj_filter *filter = NULL;
6789
6790 if (nla[NFTA_OBJ_TABLE] || nla[NFTA_OBJ_TYPE]) {
6791 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
6792 if (!filter)
6793 return -ENOMEM;
6794
6795 if (nla[NFTA_OBJ_TABLE]) {
6796 filter->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
6797 if (!filter->table) {
6798 kfree(filter);
6799 return -ENOMEM;
6800 }
6801 }
6802
6803 if (nla[NFTA_OBJ_TYPE])
6804 filter->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6805 }
6806
6807 cb->data = filter;
6808 return 0;
6809 }
6810
6811 static int nf_tables_dump_obj_done(struct netlink_callback *cb)
6812 {
6813 struct nft_obj_filter *filter = cb->data;
6814
6815 if (filter) {
6816 kfree(filter->table);
6817 kfree(filter);
6818 }
6819
6820 return 0;
6821 }
6822
6823 /* called with rcu_read_lock held */
6824 static int nf_tables_getobj(struct sk_buff *skb, const struct nfnl_info *info,
6825 const struct nlattr * const nla[])
6826 {
6827 struct netlink_ext_ack *extack = info->extack;
6828 u8 genmask = nft_genmask_cur(info->net);
6829 u8 family = info->nfmsg->nfgen_family;
6830 const struct nft_table *table;
6831 struct net *net = info->net;
6832 struct nft_object *obj;
6833 struct sk_buff *skb2;
6834 bool reset = false;
6835 u32 objtype;
6836 int err;
6837
6838 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
6839 struct netlink_dump_control c = {
6840 .start = nf_tables_dump_obj_start,
6841 .dump = nf_tables_dump_obj,
6842 .done = nf_tables_dump_obj_done,
6843 .module = THIS_MODULE,
6844 .data = (void *)nla,
6845 };
6846
6847 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
6848 }
6849
6850 if (!nla[NFTA_OBJ_NAME] ||
6851 !nla[NFTA_OBJ_TYPE])
6852 return -EINVAL;
6853
6854 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask, 0);
6855 if (IS_ERR(table)) {
6856 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6857 return PTR_ERR(table);
6858 }
6859
6860 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6861 obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
6862 if (IS_ERR(obj)) {
6863 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
6864 return PTR_ERR(obj);
6865 }
6866
6867 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
6868 if (!skb2)
6869 return -ENOMEM;
6870
6871 if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
6872 reset = true;
6873
6874 if (reset) {
6875 const struct nftables_pernet *nft_net;
6876 char *buf;
6877
6878 nft_net = nft_pernet(net);
6879 buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, nft_net->base_seq);
6880
6881 audit_log_nfcfg(buf,
6882 family,
6883 obj->handle,
6884 AUDIT_NFT_OP_OBJ_RESET,
6885 GFP_ATOMIC);
6886 kfree(buf);
6887 }
6888
6889 err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
6890 info->nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
6891 family, table, obj, reset);
6892 if (err < 0)
6893 goto err_fill_obj_info;
6894
6895 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
6896
6897 err_fill_obj_info:
6898 kfree_skb(skb2);
6899 return err;
6900 }
6901
6902 static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
6903 {
6904 if (obj->ops->destroy)
6905 obj->ops->destroy(ctx, obj);
6906
6907 module_put(obj->ops->type->owner);
6908 kfree(obj->key.name);
6909 kfree(obj->udata);
6910 kfree(obj);
6911 }
6912
6913 static int nf_tables_delobj(struct sk_buff *skb, const struct nfnl_info *info,
6914 const struct nlattr * const nla[])
6915 {
6916 struct netlink_ext_ack *extack = info->extack;
6917 u8 genmask = nft_genmask_next(info->net);
6918 u8 family = info->nfmsg->nfgen_family;
6919 struct net *net = info->net;
6920 const struct nlattr *attr;
6921 struct nft_table *table;
6922 struct nft_object *obj;
6923 struct nft_ctx ctx;
6924 u32 objtype;
6925
6926 if (!nla[NFTA_OBJ_TYPE] ||
6927 (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
6928 return -EINVAL;
6929
6930 table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
6931 NETLINK_CB(skb).portid);
6932 if (IS_ERR(table)) {
6933 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
6934 return PTR_ERR(table);
6935 }
6936
6937 objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
6938 if (nla[NFTA_OBJ_HANDLE]) {
6939 attr = nla[NFTA_OBJ_HANDLE];
6940 obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
6941 } else {
6942 attr = nla[NFTA_OBJ_NAME];
6943 obj = nft_obj_lookup(net, table, attr, objtype, genmask);
6944 }
6945
6946 if (IS_ERR(obj)) {
6947 NL_SET_BAD_ATTR(extack, attr);
6948 return PTR_ERR(obj);
6949 }
6950 if (obj->use > 0) {
6951 NL_SET_BAD_ATTR(extack, attr);
6952 return -EBUSY;
6953 }
6954
6955 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
6956
6957 return nft_delobj(&ctx, obj);
6958 }
6959
6960 void nft_obj_notify(struct net *net, const struct nft_table *table,
6961 struct nft_object *obj, u32 portid, u32 seq, int event,
6962 u16 flags, int family, int report, gfp_t gfp)
6963 {
6964 struct nftables_pernet *nft_net = nft_pernet(net);
6965 struct sk_buff *skb;
6966 int err;
6967 char *buf = kasprintf(gfp, "%s:%u",
6968 table->name, nft_net->base_seq);
6969
6970 audit_log_nfcfg(buf,
6971 family,
6972 obj->handle,
6973 event == NFT_MSG_NEWOBJ ?
6974 AUDIT_NFT_OP_OBJ_REGISTER :
6975 AUDIT_NFT_OP_OBJ_UNREGISTER,
6976 gfp);
6977 kfree(buf);
6978
6979 if (!report &&
6980 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
6981 return;
6982
6983 skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
6984 if (skb == NULL)
6985 goto err;
6986
6987 err = nf_tables_fill_obj_info(skb, net, portid, seq, event,
6988 flags & (NLM_F_CREATE | NLM_F_EXCL),
6989 family, table, obj, false);
6990 if (err < 0) {
6991 kfree_skb(skb);
6992 goto err;
6993 }
6994
6995 nft_notify_enqueue(skb, report, &nft_net->notify_list);
6996 return;
6997 err:
6998 nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
6999 }
7000 EXPORT_SYMBOL_GPL(nft_obj_notify);
7001
7002 static void nf_tables_obj_notify(const struct nft_ctx *ctx,
7003 struct nft_object *obj, int event)
7004 {
7005 nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid, ctx->seq, event,
7006 ctx->flags, ctx->family, ctx->report, GFP_KERNEL);
7007 }
7008
7009 /*
7010 * Flow tables
7011 */
7012 void nft_register_flowtable_type(struct nf_flowtable_type *type)
7013 {
7014 nfnl_lock(NFNL_SUBSYS_NFTABLES);
7015 list_add_tail_rcu(&type->list, &nf_tables_flowtables);
7016 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7017 }
7018 EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
7019
7020 void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
7021 {
7022 nfnl_lock(NFNL_SUBSYS_NFTABLES);
7023 list_del_rcu(&type->list);
7024 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7025 }
7026 EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
7027
7028 static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
7029 [NFTA_FLOWTABLE_TABLE] = { .type = NLA_STRING,
7030 .len = NFT_NAME_MAXLEN - 1 },
7031 [NFTA_FLOWTABLE_NAME] = { .type = NLA_STRING,
7032 .len = NFT_NAME_MAXLEN - 1 },
7033 [NFTA_FLOWTABLE_HOOK] = { .type = NLA_NESTED },
7034 [NFTA_FLOWTABLE_HANDLE] = { .type = NLA_U64 },
7035 [NFTA_FLOWTABLE_FLAGS] = { .type = NLA_U32 },
7036 };
7037
7038 struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
7039 const struct nlattr *nla, u8 genmask)
7040 {
7041 struct nft_flowtable *flowtable;
7042
7043 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7044 if (!nla_strcmp(nla, flowtable->name) &&
7045 nft_active_genmask(flowtable, genmask))
7046 return flowtable;
7047 }
7048 return ERR_PTR(-ENOENT);
7049 }
7050 EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
7051
7052 void nf_tables_deactivate_flowtable(const struct nft_ctx *ctx,
7053 struct nft_flowtable *flowtable,
7054 enum nft_trans_phase phase)
7055 {
7056 switch (phase) {
7057 case NFT_TRANS_PREPARE:
7058 case NFT_TRANS_ABORT:
7059 case NFT_TRANS_RELEASE:
7060 flowtable->use--;
7061 fallthrough;
7062 default:
7063 return;
7064 }
7065 }
7066 EXPORT_SYMBOL_GPL(nf_tables_deactivate_flowtable);
7067
7068 static struct nft_flowtable *
7069 nft_flowtable_lookup_byhandle(const struct nft_table *table,
7070 const struct nlattr *nla, u8 genmask)
7071 {
7072 struct nft_flowtable *flowtable;
7073
7074 list_for_each_entry(flowtable, &table->flowtables, list) {
7075 if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
7076 nft_active_genmask(flowtable, genmask))
7077 return flowtable;
7078 }
7079 return ERR_PTR(-ENOENT);
7080 }
7081
7082 struct nft_flowtable_hook {
7083 u32 num;
7084 int priority;
7085 struct list_head list;
7086 };
7087
7088 static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
7089 [NFTA_FLOWTABLE_HOOK_NUM] = { .type = NLA_U32 },
7090 [NFTA_FLOWTABLE_HOOK_PRIORITY] = { .type = NLA_U32 },
7091 [NFTA_FLOWTABLE_HOOK_DEVS] = { .type = NLA_NESTED },
7092 };
7093
7094 static int nft_flowtable_parse_hook(const struct nft_ctx *ctx,
7095 const struct nlattr *attr,
7096 struct nft_flowtable_hook *flowtable_hook,
7097 struct nft_flowtable *flowtable, bool add)
7098 {
7099 struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
7100 struct nft_hook *hook;
7101 int hooknum, priority;
7102 int err;
7103
7104 INIT_LIST_HEAD(&flowtable_hook->list);
7105
7106 err = nla_parse_nested_deprecated(tb, NFTA_FLOWTABLE_HOOK_MAX, attr,
7107 nft_flowtable_hook_policy, NULL);
7108 if (err < 0)
7109 return err;
7110
7111 if (add) {
7112 if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
7113 !tb[NFTA_FLOWTABLE_HOOK_PRIORITY])
7114 return -EINVAL;
7115
7116 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7117 if (hooknum != NF_NETDEV_INGRESS)
7118 return -EOPNOTSUPP;
7119
7120 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7121
7122 flowtable_hook->priority = priority;
7123 flowtable_hook->num = hooknum;
7124 } else {
7125 if (tb[NFTA_FLOWTABLE_HOOK_NUM]) {
7126 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
7127 if (hooknum != flowtable->hooknum)
7128 return -EOPNOTSUPP;
7129 }
7130
7131 if (tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
7132 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
7133 if (priority != flowtable->data.priority)
7134 return -EOPNOTSUPP;
7135 }
7136
7137 flowtable_hook->priority = flowtable->data.priority;
7138 flowtable_hook->num = flowtable->hooknum;
7139 }
7140
7141 if (tb[NFTA_FLOWTABLE_HOOK_DEVS]) {
7142 err = nf_tables_parse_netdev_hooks(ctx->net,
7143 tb[NFTA_FLOWTABLE_HOOK_DEVS],
7144 &flowtable_hook->list);
7145 if (err < 0)
7146 return err;
7147 }
7148
7149 list_for_each_entry(hook, &flowtable_hook->list, list) {
7150 hook->ops.pf = NFPROTO_NETDEV;
7151 hook->ops.hooknum = flowtable_hook->num;
7152 hook->ops.priority = flowtable_hook->priority;
7153 hook->ops.priv = &flowtable->data;
7154 hook->ops.hook = flowtable->data.type->hook;
7155 }
7156
7157 return err;
7158 }
7159
7160 static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
7161 {
7162 const struct nf_flowtable_type *type;
7163
7164 list_for_each_entry(type, &nf_tables_flowtables, list) {
7165 if (family == type->family)
7166 return type;
7167 }
7168 return NULL;
7169 }
7170
7171 static const struct nf_flowtable_type *
7172 nft_flowtable_type_get(struct net *net, u8 family)
7173 {
7174 const struct nf_flowtable_type *type;
7175
7176 type = __nft_flowtable_type_get(family);
7177 if (type != NULL && try_module_get(type->owner))
7178 return type;
7179
7180 lockdep_nfnl_nft_mutex_not_held();
7181 #ifdef CONFIG_MODULES
7182 if (type == NULL) {
7183 if (nft_request_module(net, "nf-flowtable-%u", family) == -EAGAIN)
7184 return ERR_PTR(-EAGAIN);
7185 }
7186 #endif
7187 return ERR_PTR(-ENOENT);
7188 }
7189
7190 /* Only called from error and netdev event paths. */
7191 static void nft_unregister_flowtable_hook(struct net *net,
7192 struct nft_flowtable *flowtable,
7193 struct nft_hook *hook)
7194 {
7195 nf_unregister_net_hook(net, &hook->ops);
7196 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7197 FLOW_BLOCK_UNBIND);
7198 }
7199
7200 static void nft_unregister_flowtable_net_hooks(struct net *net,
7201 struct list_head *hook_list)
7202 {
7203 struct nft_hook *hook;
7204
7205 list_for_each_entry(hook, hook_list, list)
7206 nf_unregister_net_hook(net, &hook->ops);
7207 }
7208
7209 static int nft_register_flowtable_net_hooks(struct net *net,
7210 struct nft_table *table,
7211 struct list_head *hook_list,
7212 struct nft_flowtable *flowtable)
7213 {
7214 struct nft_hook *hook, *hook2, *next;
7215 struct nft_flowtable *ft;
7216 int err, i = 0;
7217
7218 list_for_each_entry(hook, hook_list, list) {
7219 list_for_each_entry(ft, &table->flowtables, list) {
7220 if (!nft_is_active_next(net, ft))
7221 continue;
7222
7223 list_for_each_entry(hook2, &ft->hook_list, list) {
7224 if (hook->ops.dev == hook2->ops.dev &&
7225 hook->ops.pf == hook2->ops.pf) {
7226 err = -EEXIST;
7227 goto err_unregister_net_hooks;
7228 }
7229 }
7230 }
7231
7232 err = flowtable->data.type->setup(&flowtable->data,
7233 hook->ops.dev,
7234 FLOW_BLOCK_BIND);
7235 if (err < 0)
7236 goto err_unregister_net_hooks;
7237
7238 err = nf_register_net_hook(net, &hook->ops);
7239 if (err < 0) {
7240 flowtable->data.type->setup(&flowtable->data,
7241 hook->ops.dev,
7242 FLOW_BLOCK_UNBIND);
7243 goto err_unregister_net_hooks;
7244 }
7245
7246 i++;
7247 }
7248
7249 return 0;
7250
7251 err_unregister_net_hooks:
7252 list_for_each_entry_safe(hook, next, hook_list, list) {
7253 if (i-- <= 0)
7254 break;
7255
7256 nft_unregister_flowtable_hook(net, flowtable, hook);
7257 list_del_rcu(&hook->list);
7258 kfree_rcu(hook, rcu);
7259 }
7260
7261 return err;
7262 }
7263
7264 static void nft_flowtable_hooks_destroy(struct list_head *hook_list)
7265 {
7266 struct nft_hook *hook, *next;
7267
7268 list_for_each_entry_safe(hook, next, hook_list, list) {
7269 list_del_rcu(&hook->list);
7270 kfree_rcu(hook, rcu);
7271 }
7272 }
7273
7274 static int nft_flowtable_update(struct nft_ctx *ctx, const struct nlmsghdr *nlh,
7275 struct nft_flowtable *flowtable)
7276 {
7277 const struct nlattr * const *nla = ctx->nla;
7278 struct nft_flowtable_hook flowtable_hook;
7279 struct nft_hook *hook, *next;
7280 struct nft_trans *trans;
7281 bool unregister = false;
7282 u32 flags;
7283 int err;
7284
7285 err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7286 &flowtable_hook, flowtable, false);
7287 if (err < 0)
7288 return err;
7289
7290 list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7291 if (nft_hook_list_find(&flowtable->hook_list, hook)) {
7292 list_del(&hook->list);
7293 kfree(hook);
7294 }
7295 }
7296
7297 if (nla[NFTA_FLOWTABLE_FLAGS]) {
7298 flags = ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7299 if (flags & ~NFT_FLOWTABLE_MASK)
7300 return -EOPNOTSUPP;
7301 if ((flowtable->data.flags & NFT_FLOWTABLE_HW_OFFLOAD) ^
7302 (flags & NFT_FLOWTABLE_HW_OFFLOAD))
7303 return -EOPNOTSUPP;
7304 } else {
7305 flags = flowtable->data.flags;
7306 }
7307
7308 err = nft_register_flowtable_net_hooks(ctx->net, ctx->table,
7309 &flowtable_hook.list, flowtable);
7310 if (err < 0)
7311 goto err_flowtable_update_hook;
7312
7313 trans = nft_trans_alloc(ctx, NFT_MSG_NEWFLOWTABLE,
7314 sizeof(struct nft_trans_flowtable));
7315 if (!trans) {
7316 unregister = true;
7317 err = -ENOMEM;
7318 goto err_flowtable_update_hook;
7319 }
7320
7321 nft_trans_flowtable_flags(trans) = flags;
7322 nft_trans_flowtable(trans) = flowtable;
7323 nft_trans_flowtable_update(trans) = true;
7324 INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7325 list_splice(&flowtable_hook.list, &nft_trans_flowtable_hooks(trans));
7326
7327 nft_trans_commit_list_add_tail(ctx->net, trans);
7328
7329 return 0;
7330
7331 err_flowtable_update_hook:
7332 list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
7333 if (unregister)
7334 nft_unregister_flowtable_hook(ctx->net, flowtable, hook);
7335 list_del_rcu(&hook->list);
7336 kfree_rcu(hook, rcu);
7337 }
7338
7339 return err;
7340
7341 }
7342
7343 static int nf_tables_newflowtable(struct sk_buff *skb,
7344 const struct nfnl_info *info,
7345 const struct nlattr * const nla[])
7346 {
7347 struct netlink_ext_ack *extack = info->extack;
7348 struct nft_flowtable_hook flowtable_hook;
7349 u8 genmask = nft_genmask_next(info->net);
7350 u8 family = info->nfmsg->nfgen_family;
7351 const struct nf_flowtable_type *type;
7352 struct nft_flowtable *flowtable;
7353 struct nft_hook *hook, *next;
7354 struct net *net = info->net;
7355 struct nft_table *table;
7356 struct nft_ctx ctx;
7357 int err;
7358
7359 if (!nla[NFTA_FLOWTABLE_TABLE] ||
7360 !nla[NFTA_FLOWTABLE_NAME] ||
7361 !nla[NFTA_FLOWTABLE_HOOK])
7362 return -EINVAL;
7363
7364 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7365 genmask, NETLINK_CB(skb).portid);
7366 if (IS_ERR(table)) {
7367 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7368 return PTR_ERR(table);
7369 }
7370
7371 flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7372 genmask);
7373 if (IS_ERR(flowtable)) {
7374 err = PTR_ERR(flowtable);
7375 if (err != -ENOENT) {
7376 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7377 return err;
7378 }
7379 } else {
7380 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
7381 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
7382 return -EEXIST;
7383 }
7384
7385 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7386
7387 return nft_flowtable_update(&ctx, info->nlh, flowtable);
7388 }
7389
7390 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7391
7392 flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL);
7393 if (!flowtable)
7394 return -ENOMEM;
7395
7396 flowtable->table = table;
7397 flowtable->handle = nf_tables_alloc_handle(table);
7398 INIT_LIST_HEAD(&flowtable->hook_list);
7399
7400 flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL);
7401 if (!flowtable->name) {
7402 err = -ENOMEM;
7403 goto err1;
7404 }
7405
7406 type = nft_flowtable_type_get(net, family);
7407 if (IS_ERR(type)) {
7408 err = PTR_ERR(type);
7409 goto err2;
7410 }
7411
7412 if (nla[NFTA_FLOWTABLE_FLAGS]) {
7413 flowtable->data.flags =
7414 ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
7415 if (flowtable->data.flags & ~NFT_FLOWTABLE_MASK) {
7416 err = -EOPNOTSUPP;
7417 goto err3;
7418 }
7419 }
7420
7421 write_pnet(&flowtable->data.net, net);
7422 flowtable->data.type = type;
7423 err = type->init(&flowtable->data);
7424 if (err < 0)
7425 goto err3;
7426
7427 err = nft_flowtable_parse_hook(&ctx, nla[NFTA_FLOWTABLE_HOOK],
7428 &flowtable_hook, flowtable, true);
7429 if (err < 0)
7430 goto err4;
7431
7432 list_splice(&flowtable_hook.list, &flowtable->hook_list);
7433 flowtable->data.priority = flowtable_hook.priority;
7434 flowtable->hooknum = flowtable_hook.num;
7435
7436 err = nft_register_flowtable_net_hooks(ctx.net, table,
7437 &flowtable->hook_list,
7438 flowtable);
7439 if (err < 0) {
7440 nft_flowtable_hooks_destroy(&flowtable->hook_list);
7441 goto err4;
7442 }
7443
7444 err = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
7445 if (err < 0)
7446 goto err5;
7447
7448 list_add_tail_rcu(&flowtable->list, &table->flowtables);
7449 table->use++;
7450
7451 return 0;
7452 err5:
7453 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7454 nft_unregister_flowtable_hook(net, flowtable, hook);
7455 list_del_rcu(&hook->list);
7456 kfree_rcu(hook, rcu);
7457 }
7458 err4:
7459 flowtable->data.type->free(&flowtable->data);
7460 err3:
7461 module_put(type->owner);
7462 err2:
7463 kfree(flowtable->name);
7464 err1:
7465 kfree(flowtable);
7466 return err;
7467 }
7468
7469 static void nft_flowtable_hook_release(struct nft_flowtable_hook *flowtable_hook)
7470 {
7471 struct nft_hook *this, *next;
7472
7473 list_for_each_entry_safe(this, next, &flowtable_hook->list, list) {
7474 list_del(&this->list);
7475 kfree(this);
7476 }
7477 }
7478
7479 static int nft_delflowtable_hook(struct nft_ctx *ctx,
7480 struct nft_flowtable *flowtable)
7481 {
7482 const struct nlattr * const *nla = ctx->nla;
7483 struct nft_flowtable_hook flowtable_hook;
7484 struct nft_hook *this, *hook;
7485 struct nft_trans *trans;
7486 int err;
7487
7488 err = nft_flowtable_parse_hook(ctx, nla[NFTA_FLOWTABLE_HOOK],
7489 &flowtable_hook, flowtable, false);
7490 if (err < 0)
7491 return err;
7492
7493 list_for_each_entry(this, &flowtable_hook.list, list) {
7494 hook = nft_hook_list_find(&flowtable->hook_list, this);
7495 if (!hook) {
7496 err = -ENOENT;
7497 goto err_flowtable_del_hook;
7498 }
7499 hook->inactive = true;
7500 }
7501
7502 trans = nft_trans_alloc(ctx, NFT_MSG_DELFLOWTABLE,
7503 sizeof(struct nft_trans_flowtable));
7504 if (!trans) {
7505 err = -ENOMEM;
7506 goto err_flowtable_del_hook;
7507 }
7508
7509 nft_trans_flowtable(trans) = flowtable;
7510 nft_trans_flowtable_update(trans) = true;
7511 INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
7512 nft_flowtable_hook_release(&flowtable_hook);
7513
7514 nft_trans_commit_list_add_tail(ctx->net, trans);
7515
7516 return 0;
7517
7518 err_flowtable_del_hook:
7519 list_for_each_entry(this, &flowtable_hook.list, list) {
7520 hook = nft_hook_list_find(&flowtable->hook_list, this);
7521 if (!hook)
7522 break;
7523
7524 hook->inactive = false;
7525 }
7526 nft_flowtable_hook_release(&flowtable_hook);
7527
7528 return err;
7529 }
7530
7531 static int nf_tables_delflowtable(struct sk_buff *skb,
7532 const struct nfnl_info *info,
7533 const struct nlattr * const nla[])
7534 {
7535 struct netlink_ext_ack *extack = info->extack;
7536 u8 genmask = nft_genmask_next(info->net);
7537 u8 family = info->nfmsg->nfgen_family;
7538 struct nft_flowtable *flowtable;
7539 struct net *net = info->net;
7540 const struct nlattr *attr;
7541 struct nft_table *table;
7542 struct nft_ctx ctx;
7543
7544 if (!nla[NFTA_FLOWTABLE_TABLE] ||
7545 (!nla[NFTA_FLOWTABLE_NAME] &&
7546 !nla[NFTA_FLOWTABLE_HANDLE]))
7547 return -EINVAL;
7548
7549 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7550 genmask, NETLINK_CB(skb).portid);
7551 if (IS_ERR(table)) {
7552 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
7553 return PTR_ERR(table);
7554 }
7555
7556 if (nla[NFTA_FLOWTABLE_HANDLE]) {
7557 attr = nla[NFTA_FLOWTABLE_HANDLE];
7558 flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
7559 } else {
7560 attr = nla[NFTA_FLOWTABLE_NAME];
7561 flowtable = nft_flowtable_lookup(table, attr, genmask);
7562 }
7563
7564 if (IS_ERR(flowtable)) {
7565 NL_SET_BAD_ATTR(extack, attr);
7566 return PTR_ERR(flowtable);
7567 }
7568
7569 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7570
7571 if (nla[NFTA_FLOWTABLE_HOOK])
7572 return nft_delflowtable_hook(&ctx, flowtable);
7573
7574 if (flowtable->use > 0) {
7575 NL_SET_BAD_ATTR(extack, attr);
7576 return -EBUSY;
7577 }
7578
7579 return nft_delflowtable(&ctx, flowtable);
7580 }
7581
7582 static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
7583 u32 portid, u32 seq, int event,
7584 u32 flags, int family,
7585 struct nft_flowtable *flowtable,
7586 struct list_head *hook_list)
7587 {
7588 struct nlattr *nest, *nest_devs;
7589 struct nft_hook *hook;
7590 struct nlmsghdr *nlh;
7591
7592 event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
7593 nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
7594 NFNETLINK_V0, nft_base_seq(net));
7595 if (!nlh)
7596 goto nla_put_failure;
7597
7598 if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
7599 nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
7600 nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
7601 nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
7602 NFTA_FLOWTABLE_PAD) ||
7603 nla_put_be32(skb, NFTA_FLOWTABLE_FLAGS, htonl(flowtable->data.flags)))
7604 goto nla_put_failure;
7605
7606 nest = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK);
7607 if (!nest)
7608 goto nla_put_failure;
7609 if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
7610 nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->data.priority)))
7611 goto nla_put_failure;
7612
7613 nest_devs = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK_DEVS);
7614 if (!nest_devs)
7615 goto nla_put_failure;
7616
7617 list_for_each_entry_rcu(hook, hook_list, list) {
7618 if (nla_put_string(skb, NFTA_DEVICE_NAME, hook->ops.dev->name))
7619 goto nla_put_failure;
7620 }
7621 nla_nest_end(skb, nest_devs);
7622 nla_nest_end(skb, nest);
7623
7624 nlmsg_end(skb, nlh);
7625 return 0;
7626
7627 nla_put_failure:
7628 nlmsg_trim(skb, nlh);
7629 return -1;
7630 }
7631
7632 struct nft_flowtable_filter {
7633 char *table;
7634 };
7635
7636 static int nf_tables_dump_flowtable(struct sk_buff *skb,
7637 struct netlink_callback *cb)
7638 {
7639 const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
7640 struct nft_flowtable_filter *filter = cb->data;
7641 unsigned int idx = 0, s_idx = cb->args[0];
7642 struct net *net = sock_net(skb->sk);
7643 int family = nfmsg->nfgen_family;
7644 struct nft_flowtable *flowtable;
7645 struct nftables_pernet *nft_net;
7646 const struct nft_table *table;
7647
7648 rcu_read_lock();
7649 nft_net = nft_pernet(net);
7650 cb->seq = nft_net->base_seq;
7651
7652 list_for_each_entry_rcu(table, &nft_net->tables, list) {
7653 if (family != NFPROTO_UNSPEC && family != table->family)
7654 continue;
7655
7656 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
7657 if (!nft_is_active(net, flowtable))
7658 goto cont;
7659 if (idx < s_idx)
7660 goto cont;
7661 if (idx > s_idx)
7662 memset(&cb->args[1], 0,
7663 sizeof(cb->args) - sizeof(cb->args[0]));
7664 if (filter && filter->table &&
7665 strcmp(filter->table, table->name))
7666 goto cont;
7667
7668 if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
7669 cb->nlh->nlmsg_seq,
7670 NFT_MSG_NEWFLOWTABLE,
7671 NLM_F_MULTI | NLM_F_APPEND,
7672 table->family,
7673 flowtable,
7674 &flowtable->hook_list) < 0)
7675 goto done;
7676
7677 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
7678 cont:
7679 idx++;
7680 }
7681 }
7682 done:
7683 rcu_read_unlock();
7684
7685 cb->args[0] = idx;
7686 return skb->len;
7687 }
7688
7689 static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
7690 {
7691 const struct nlattr * const *nla = cb->data;
7692 struct nft_flowtable_filter *filter = NULL;
7693
7694 if (nla[NFTA_FLOWTABLE_TABLE]) {
7695 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
7696 if (!filter)
7697 return -ENOMEM;
7698
7699 filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
7700 GFP_ATOMIC);
7701 if (!filter->table) {
7702 kfree(filter);
7703 return -ENOMEM;
7704 }
7705 }
7706
7707 cb->data = filter;
7708 return 0;
7709 }
7710
7711 static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
7712 {
7713 struct nft_flowtable_filter *filter = cb->data;
7714
7715 if (!filter)
7716 return 0;
7717
7718 kfree(filter->table);
7719 kfree(filter);
7720
7721 return 0;
7722 }
7723
7724 /* called with rcu_read_lock held */
7725 static int nf_tables_getflowtable(struct sk_buff *skb,
7726 const struct nfnl_info *info,
7727 const struct nlattr * const nla[])
7728 {
7729 u8 genmask = nft_genmask_cur(info->net);
7730 u8 family = info->nfmsg->nfgen_family;
7731 struct nft_flowtable *flowtable;
7732 const struct nft_table *table;
7733 struct net *net = info->net;
7734 struct sk_buff *skb2;
7735 int err;
7736
7737 if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
7738 struct netlink_dump_control c = {
7739 .start = nf_tables_dump_flowtable_start,
7740 .dump = nf_tables_dump_flowtable,
7741 .done = nf_tables_dump_flowtable_done,
7742 .module = THIS_MODULE,
7743 .data = (void *)nla,
7744 };
7745
7746 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
7747 }
7748
7749 if (!nla[NFTA_FLOWTABLE_NAME])
7750 return -EINVAL;
7751
7752 table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
7753 genmask, 0);
7754 if (IS_ERR(table))
7755 return PTR_ERR(table);
7756
7757 flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
7758 genmask);
7759 if (IS_ERR(flowtable))
7760 return PTR_ERR(flowtable);
7761
7762 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7763 if (!skb2)
7764 return -ENOMEM;
7765
7766 err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
7767 info->nlh->nlmsg_seq,
7768 NFT_MSG_NEWFLOWTABLE, 0, family,
7769 flowtable, &flowtable->hook_list);
7770 if (err < 0)
7771 goto err_fill_flowtable_info;
7772
7773 return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
7774
7775 err_fill_flowtable_info:
7776 kfree_skb(skb2);
7777 return err;
7778 }
7779
7780 static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
7781 struct nft_flowtable *flowtable,
7782 struct list_head *hook_list,
7783 int event)
7784 {
7785 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
7786 struct sk_buff *skb;
7787 u16 flags = 0;
7788 int err;
7789
7790 if (!ctx->report &&
7791 !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
7792 return;
7793
7794 skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
7795 if (skb == NULL)
7796 goto err;
7797
7798 if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
7799 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
7800
7801 err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
7802 ctx->seq, event, flags,
7803 ctx->family, flowtable, hook_list);
7804 if (err < 0) {
7805 kfree_skb(skb);
7806 goto err;
7807 }
7808
7809 nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
7810 return;
7811 err:
7812 nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
7813 }
7814
7815 static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
7816 {
7817 struct nft_hook *hook, *next;
7818
7819 flowtable->data.type->free(&flowtable->data);
7820 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
7821 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
7822 FLOW_BLOCK_UNBIND);
7823 list_del_rcu(&hook->list);
7824 kfree(hook);
7825 }
7826 kfree(flowtable->name);
7827 module_put(flowtable->data.type->owner);
7828 kfree(flowtable);
7829 }
7830
7831 static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
7832 u32 portid, u32 seq)
7833 {
7834 struct nftables_pernet *nft_net = nft_pernet(net);
7835 struct nlmsghdr *nlh;
7836 char buf[TASK_COMM_LEN];
7837 int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
7838
7839 nlh = nfnl_msg_put(skb, portid, seq, event, 0, AF_UNSPEC,
7840 NFNETLINK_V0, nft_base_seq(net));
7841 if (!nlh)
7842 goto nla_put_failure;
7843
7844 if (nla_put_be32(skb, NFTA_GEN_ID, htonl(nft_net->base_seq)) ||
7845 nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
7846 nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
7847 goto nla_put_failure;
7848
7849 nlmsg_end(skb, nlh);
7850 return 0;
7851
7852 nla_put_failure:
7853 nlmsg_trim(skb, nlh);
7854 return -EMSGSIZE;
7855 }
7856
7857 static void nft_flowtable_event(unsigned long event, struct net_device *dev,
7858 struct nft_flowtable *flowtable)
7859 {
7860 struct nft_hook *hook;
7861
7862 list_for_each_entry(hook, &flowtable->hook_list, list) {
7863 if (hook->ops.dev != dev)
7864 continue;
7865
7866 /* flow_offload_netdev_event() cleans up entries for us. */
7867 nft_unregister_flowtable_hook(dev_net(dev), flowtable, hook);
7868 list_del_rcu(&hook->list);
7869 kfree_rcu(hook, rcu);
7870 break;
7871 }
7872 }
7873
7874 static int nf_tables_flowtable_event(struct notifier_block *this,
7875 unsigned long event, void *ptr)
7876 {
7877 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
7878 struct nft_flowtable *flowtable;
7879 struct nftables_pernet *nft_net;
7880 struct nft_table *table;
7881 struct net *net;
7882
7883 if (event != NETDEV_UNREGISTER)
7884 return 0;
7885
7886 net = dev_net(dev);
7887 nft_net = nft_pernet(net);
7888 mutex_lock(&nft_net->commit_mutex);
7889 list_for_each_entry(table, &nft_net->tables, list) {
7890 list_for_each_entry(flowtable, &table->flowtables, list) {
7891 nft_flowtable_event(event, dev, flowtable);
7892 }
7893 }
7894 mutex_unlock(&nft_net->commit_mutex);
7895
7896 return NOTIFY_DONE;
7897 }
7898
7899 static struct notifier_block nf_tables_flowtable_notifier = {
7900 .notifier_call = nf_tables_flowtable_event,
7901 };
7902
7903 static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
7904 int event)
7905 {
7906 struct nlmsghdr *nlh = nlmsg_hdr(skb);
7907 struct sk_buff *skb2;
7908 int err;
7909
7910 if (!nlmsg_report(nlh) &&
7911 !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
7912 return;
7913
7914 skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
7915 if (skb2 == NULL)
7916 goto err;
7917
7918 err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
7919 nlh->nlmsg_seq);
7920 if (err < 0) {
7921 kfree_skb(skb2);
7922 goto err;
7923 }
7924
7925 nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
7926 nlmsg_report(nlh), GFP_KERNEL);
7927 return;
7928 err:
7929 nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
7930 -ENOBUFS);
7931 }
7932
7933 static int nf_tables_getgen(struct sk_buff *skb, const struct nfnl_info *info,
7934 const struct nlattr * const nla[])
7935 {
7936 struct sk_buff *skb2;
7937 int err;
7938
7939 skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
7940 if (skb2 == NULL)
7941 return -ENOMEM;
7942
7943 err = nf_tables_fill_gen_info(skb2, info->net, NETLINK_CB(skb).portid,
7944 info->nlh->nlmsg_seq);
7945 if (err < 0)
7946 goto err_fill_gen_info;
7947
7948 return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
7949
7950 err_fill_gen_info:
7951 kfree_skb(skb2);
7952 return err;
7953 }
7954
7955 static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
7956 [NFT_MSG_NEWTABLE] = {
7957 .call = nf_tables_newtable,
7958 .type = NFNL_CB_BATCH,
7959 .attr_count = NFTA_TABLE_MAX,
7960 .policy = nft_table_policy,
7961 },
7962 [NFT_MSG_GETTABLE] = {
7963 .call = nf_tables_gettable,
7964 .type = NFNL_CB_RCU,
7965 .attr_count = NFTA_TABLE_MAX,
7966 .policy = nft_table_policy,
7967 },
7968 [NFT_MSG_DELTABLE] = {
7969 .call = nf_tables_deltable,
7970 .type = NFNL_CB_BATCH,
7971 .attr_count = NFTA_TABLE_MAX,
7972 .policy = nft_table_policy,
7973 },
7974 [NFT_MSG_NEWCHAIN] = {
7975 .call = nf_tables_newchain,
7976 .type = NFNL_CB_BATCH,
7977 .attr_count = NFTA_CHAIN_MAX,
7978 .policy = nft_chain_policy,
7979 },
7980 [NFT_MSG_GETCHAIN] = {
7981 .call = nf_tables_getchain,
7982 .type = NFNL_CB_RCU,
7983 .attr_count = NFTA_CHAIN_MAX,
7984 .policy = nft_chain_policy,
7985 },
7986 [NFT_MSG_DELCHAIN] = {
7987 .call = nf_tables_delchain,
7988 .type = NFNL_CB_BATCH,
7989 .attr_count = NFTA_CHAIN_MAX,
7990 .policy = nft_chain_policy,
7991 },
7992 [NFT_MSG_NEWRULE] = {
7993 .call = nf_tables_newrule,
7994 .type = NFNL_CB_BATCH,
7995 .attr_count = NFTA_RULE_MAX,
7996 .policy = nft_rule_policy,
7997 },
7998 [NFT_MSG_GETRULE] = {
7999 .call = nf_tables_getrule,
8000 .type = NFNL_CB_RCU,
8001 .attr_count = NFTA_RULE_MAX,
8002 .policy = nft_rule_policy,
8003 },
8004 [NFT_MSG_DELRULE] = {
8005 .call = nf_tables_delrule,
8006 .type = NFNL_CB_BATCH,
8007 .attr_count = NFTA_RULE_MAX,
8008 .policy = nft_rule_policy,
8009 },
8010 [NFT_MSG_NEWSET] = {
8011 .call = nf_tables_newset,
8012 .type = NFNL_CB_BATCH,
8013 .attr_count = NFTA_SET_MAX,
8014 .policy = nft_set_policy,
8015 },
8016 [NFT_MSG_GETSET] = {
8017 .call = nf_tables_getset,
8018 .type = NFNL_CB_RCU,
8019 .attr_count = NFTA_SET_MAX,
8020 .policy = nft_set_policy,
8021 },
8022 [NFT_MSG_DELSET] = {
8023 .call = nf_tables_delset,
8024 .type = NFNL_CB_BATCH,
8025 .attr_count = NFTA_SET_MAX,
8026 .policy = nft_set_policy,
8027 },
8028 [NFT_MSG_NEWSETELEM] = {
8029 .call = nf_tables_newsetelem,
8030 .type = NFNL_CB_BATCH,
8031 .attr_count = NFTA_SET_ELEM_LIST_MAX,
8032 .policy = nft_set_elem_list_policy,
8033 },
8034 [NFT_MSG_GETSETELEM] = {
8035 .call = nf_tables_getsetelem,
8036 .type = NFNL_CB_RCU,
8037 .attr_count = NFTA_SET_ELEM_LIST_MAX,
8038 .policy = nft_set_elem_list_policy,
8039 },
8040 [NFT_MSG_DELSETELEM] = {
8041 .call = nf_tables_delsetelem,
8042 .type = NFNL_CB_BATCH,
8043 .attr_count = NFTA_SET_ELEM_LIST_MAX,
8044 .policy = nft_set_elem_list_policy,
8045 },
8046 [NFT_MSG_GETGEN] = {
8047 .call = nf_tables_getgen,
8048 .type = NFNL_CB_RCU,
8049 },
8050 [NFT_MSG_NEWOBJ] = {
8051 .call = nf_tables_newobj,
8052 .type = NFNL_CB_BATCH,
8053 .attr_count = NFTA_OBJ_MAX,
8054 .policy = nft_obj_policy,
8055 },
8056 [NFT_MSG_GETOBJ] = {
8057 .call = nf_tables_getobj,
8058 .type = NFNL_CB_RCU,
8059 .attr_count = NFTA_OBJ_MAX,
8060 .policy = nft_obj_policy,
8061 },
8062 [NFT_MSG_DELOBJ] = {
8063 .call = nf_tables_delobj,
8064 .type = NFNL_CB_BATCH,
8065 .attr_count = NFTA_OBJ_MAX,
8066 .policy = nft_obj_policy,
8067 },
8068 [NFT_MSG_GETOBJ_RESET] = {
8069 .call = nf_tables_getobj,
8070 .type = NFNL_CB_RCU,
8071 .attr_count = NFTA_OBJ_MAX,
8072 .policy = nft_obj_policy,
8073 },
8074 [NFT_MSG_NEWFLOWTABLE] = {
8075 .call = nf_tables_newflowtable,
8076 .type = NFNL_CB_BATCH,
8077 .attr_count = NFTA_FLOWTABLE_MAX,
8078 .policy = nft_flowtable_policy,
8079 },
8080 [NFT_MSG_GETFLOWTABLE] = {
8081 .call = nf_tables_getflowtable,
8082 .type = NFNL_CB_RCU,
8083 .attr_count = NFTA_FLOWTABLE_MAX,
8084 .policy = nft_flowtable_policy,
8085 },
8086 [NFT_MSG_DELFLOWTABLE] = {
8087 .call = nf_tables_delflowtable,
8088 .type = NFNL_CB_BATCH,
8089 .attr_count = NFTA_FLOWTABLE_MAX,
8090 .policy = nft_flowtable_policy,
8091 },
8092 };
8093
8094 static int nf_tables_validate(struct net *net)
8095 {
8096 struct nftables_pernet *nft_net = nft_pernet(net);
8097 struct nft_table *table;
8098
8099 switch (nft_net->validate_state) {
8100 case NFT_VALIDATE_SKIP:
8101 break;
8102 case NFT_VALIDATE_NEED:
8103 nft_validate_state_update(net, NFT_VALIDATE_DO);
8104 fallthrough;
8105 case NFT_VALIDATE_DO:
8106 list_for_each_entry(table, &nft_net->tables, list) {
8107 if (nft_table_validate(net, table) < 0)
8108 return -EAGAIN;
8109 }
8110 break;
8111 }
8112
8113 return 0;
8114 }
8115
8116 /* a drop policy has to be deferred until all rules have been activated,
8117 * otherwise a large ruleset that contains a drop-policy base chain will
8118 * cause all packets to get dropped until the full transaction has been
8119 * processed.
8120 *
8121 * We defer the drop policy until the transaction has been finalized.
8122 */
8123 static void nft_chain_commit_drop_policy(struct nft_trans *trans)
8124 {
8125 struct nft_base_chain *basechain;
8126
8127 if (nft_trans_chain_policy(trans) != NF_DROP)
8128 return;
8129
8130 if (!nft_is_base_chain(trans->ctx.chain))
8131 return;
8132
8133 basechain = nft_base_chain(trans->ctx.chain);
8134 basechain->policy = NF_DROP;
8135 }
8136
8137 static void nft_chain_commit_update(struct nft_trans *trans)
8138 {
8139 struct nft_base_chain *basechain;
8140
8141 if (nft_trans_chain_name(trans)) {
8142 rhltable_remove(&trans->ctx.table->chains_ht,
8143 &trans->ctx.chain->rhlhead,
8144 nft_chain_ht_params);
8145 swap(trans->ctx.chain->name, nft_trans_chain_name(trans));
8146 rhltable_insert_key(&trans->ctx.table->chains_ht,
8147 trans->ctx.chain->name,
8148 &trans->ctx.chain->rhlhead,
8149 nft_chain_ht_params);
8150 }
8151
8152 if (!nft_is_base_chain(trans->ctx.chain))
8153 return;
8154
8155 nft_chain_stats_replace(trans);
8156
8157 basechain = nft_base_chain(trans->ctx.chain);
8158
8159 switch (nft_trans_chain_policy(trans)) {
8160 case NF_DROP:
8161 case NF_ACCEPT:
8162 basechain->policy = nft_trans_chain_policy(trans);
8163 break;
8164 }
8165 }
8166
8167 static void nft_obj_commit_update(struct nft_trans *trans)
8168 {
8169 struct nft_object *newobj;
8170 struct nft_object *obj;
8171
8172 obj = nft_trans_obj(trans);
8173 newobj = nft_trans_obj_newobj(trans);
8174
8175 if (obj->ops->update)
8176 obj->ops->update(obj, newobj);
8177
8178 nft_obj_destroy(&trans->ctx, newobj);
8179 }
8180
8181 static void nft_commit_release(struct nft_trans *trans)
8182 {
8183 switch (trans->msg_type) {
8184 case NFT_MSG_DELTABLE:
8185 nf_tables_table_destroy(&trans->ctx);
8186 break;
8187 case NFT_MSG_NEWCHAIN:
8188 free_percpu(nft_trans_chain_stats(trans));
8189 kfree(nft_trans_chain_name(trans));
8190 break;
8191 case NFT_MSG_DELCHAIN:
8192 nf_tables_chain_destroy(&trans->ctx);
8193 break;
8194 case NFT_MSG_DELRULE:
8195 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8196 break;
8197 case NFT_MSG_DELSET:
8198 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8199 break;
8200 case NFT_MSG_DELSETELEM:
8201 nf_tables_set_elem_destroy(&trans->ctx,
8202 nft_trans_elem_set(trans),
8203 nft_trans_elem(trans).priv);
8204 break;
8205 case NFT_MSG_DELOBJ:
8206 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8207 break;
8208 case NFT_MSG_DELFLOWTABLE:
8209 if (nft_trans_flowtable_update(trans))
8210 nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8211 else
8212 nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8213 break;
8214 }
8215
8216 if (trans->put_net)
8217 put_net(trans->ctx.net);
8218
8219 kfree(trans);
8220 }
8221
8222 static void nf_tables_trans_destroy_work(struct work_struct *w)
8223 {
8224 struct nft_trans *trans, *next;
8225 LIST_HEAD(head);
8226
8227 spin_lock(&nf_tables_destroy_list_lock);
8228 list_splice_init(&nf_tables_destroy_list, &head);
8229 spin_unlock(&nf_tables_destroy_list_lock);
8230
8231 if (list_empty(&head))
8232 return;
8233
8234 synchronize_rcu();
8235
8236 list_for_each_entry_safe(trans, next, &head, list) {
8237 list_del(&trans->list);
8238 nft_commit_release(trans);
8239 }
8240 }
8241
8242 void nf_tables_trans_destroy_flush_work(void)
8243 {
8244 flush_work(&trans_destroy_work);
8245 }
8246 EXPORT_SYMBOL_GPL(nf_tables_trans_destroy_flush_work);
8247
8248 static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
8249 {
8250 struct nft_rule *rule;
8251 unsigned int alloc = 0;
8252 int i;
8253
8254 /* already handled or inactive chain? */
8255 if (chain->rules_next || !nft_is_active_next(net, chain))
8256 return 0;
8257
8258 rule = list_entry(&chain->rules, struct nft_rule, list);
8259 i = 0;
8260
8261 list_for_each_entry_continue(rule, &chain->rules, list) {
8262 if (nft_is_active_next(net, rule))
8263 alloc++;
8264 }
8265
8266 chain->rules_next = nf_tables_chain_alloc_rules(chain, alloc);
8267 if (!chain->rules_next)
8268 return -ENOMEM;
8269
8270 list_for_each_entry_continue(rule, &chain->rules, list) {
8271 if (nft_is_active_next(net, rule))
8272 chain->rules_next[i++] = rule;
8273 }
8274
8275 chain->rules_next[i] = NULL;
8276 return 0;
8277 }
8278
8279 static void nf_tables_commit_chain_prepare_cancel(struct net *net)
8280 {
8281 struct nftables_pernet *nft_net = nft_pernet(net);
8282 struct nft_trans *trans, *next;
8283
8284 list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8285 struct nft_chain *chain = trans->ctx.chain;
8286
8287 if (trans->msg_type == NFT_MSG_NEWRULE ||
8288 trans->msg_type == NFT_MSG_DELRULE) {
8289 kvfree(chain->rules_next);
8290 chain->rules_next = NULL;
8291 }
8292 }
8293 }
8294
8295 static void __nf_tables_commit_chain_free_rules_old(struct rcu_head *h)
8296 {
8297 struct nft_rules_old *o = container_of(h, struct nft_rules_old, h);
8298
8299 kvfree(o->start);
8300 }
8301
8302 static void nf_tables_commit_chain_free_rules_old(struct nft_rule **rules)
8303 {
8304 struct nft_rule **r = rules;
8305 struct nft_rules_old *old;
8306
8307 while (*r)
8308 r++;
8309
8310 r++; /* rcu_head is after end marker */
8311 old = (void *) r;
8312 old->start = rules;
8313
8314 call_rcu(&old->h, __nf_tables_commit_chain_free_rules_old);
8315 }
8316
8317 static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
8318 {
8319 struct nft_rule **g0, **g1;
8320 bool next_genbit;
8321
8322 next_genbit = nft_gencursor_next(net);
8323
8324 g0 = rcu_dereference_protected(chain->rules_gen_0,
8325 lockdep_commit_lock_is_held(net));
8326 g1 = rcu_dereference_protected(chain->rules_gen_1,
8327 lockdep_commit_lock_is_held(net));
8328
8329 /* No changes to this chain? */
8330 if (chain->rules_next == NULL) {
8331 /* chain had no change in last or next generation */
8332 if (g0 == g1)
8333 return;
8334 /*
8335 * chain had no change in this generation; make sure next
8336 * one uses same rules as current generation.
8337 */
8338 if (next_genbit) {
8339 rcu_assign_pointer(chain->rules_gen_1, g0);
8340 nf_tables_commit_chain_free_rules_old(g1);
8341 } else {
8342 rcu_assign_pointer(chain->rules_gen_0, g1);
8343 nf_tables_commit_chain_free_rules_old(g0);
8344 }
8345
8346 return;
8347 }
8348
8349 if (next_genbit)
8350 rcu_assign_pointer(chain->rules_gen_1, chain->rules_next);
8351 else
8352 rcu_assign_pointer(chain->rules_gen_0, chain->rules_next);
8353
8354 chain->rules_next = NULL;
8355
8356 if (g0 == g1)
8357 return;
8358
8359 if (next_genbit)
8360 nf_tables_commit_chain_free_rules_old(g1);
8361 else
8362 nf_tables_commit_chain_free_rules_old(g0);
8363 }
8364
8365 static void nft_obj_del(struct nft_object *obj)
8366 {
8367 rhltable_remove(&nft_objname_ht, &obj->rhlhead, nft_objname_ht_params);
8368 list_del_rcu(&obj->list);
8369 }
8370
8371 void nft_chain_del(struct nft_chain *chain)
8372 {
8373 struct nft_table *table = chain->table;
8374
8375 WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
8376 nft_chain_ht_params));
8377 list_del_rcu(&chain->list);
8378 }
8379
8380 static void nft_flowtable_hooks_del(struct nft_flowtable *flowtable,
8381 struct list_head *hook_list)
8382 {
8383 struct nft_hook *hook, *next;
8384
8385 list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
8386 if (hook->inactive)
8387 list_move(&hook->list, hook_list);
8388 }
8389 }
8390
8391 static void nf_tables_module_autoload_cleanup(struct net *net)
8392 {
8393 struct nftables_pernet *nft_net = nft_pernet(net);
8394 struct nft_module_request *req, *next;
8395
8396 WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
8397 list_for_each_entry_safe(req, next, &nft_net->module_list, list) {
8398 WARN_ON_ONCE(!req->done);
8399 list_del(&req->list);
8400 kfree(req);
8401 }
8402 }
8403
8404 static void nf_tables_commit_release(struct net *net)
8405 {
8406 struct nftables_pernet *nft_net = nft_pernet(net);
8407 struct nft_trans *trans;
8408
8409 /* all side effects have to be made visible.
8410 * For example, if a chain named 'foo' has been deleted, a
8411 * new transaction must not find it anymore.
8412 *
8413 * Memory reclaim happens asynchronously from work queue
8414 * to prevent expensive synchronize_rcu() in commit phase.
8415 */
8416 if (list_empty(&nft_net->commit_list)) {
8417 nf_tables_module_autoload_cleanup(net);
8418 mutex_unlock(&nft_net->commit_mutex);
8419 return;
8420 }
8421
8422 trans = list_last_entry(&nft_net->commit_list,
8423 struct nft_trans, list);
8424 get_net(trans->ctx.net);
8425 WARN_ON_ONCE(trans->put_net);
8426
8427 trans->put_net = true;
8428 spin_lock(&nf_tables_destroy_list_lock);
8429 list_splice_tail_init(&nft_net->commit_list, &nf_tables_destroy_list);
8430 spin_unlock(&nf_tables_destroy_list_lock);
8431
8432 nf_tables_module_autoload_cleanup(net);
8433 schedule_work(&trans_destroy_work);
8434
8435 mutex_unlock(&nft_net->commit_mutex);
8436 }
8437
8438 static void nft_commit_notify(struct net *net, u32 portid)
8439 {
8440 struct nftables_pernet *nft_net = nft_pernet(net);
8441 struct sk_buff *batch_skb = NULL, *nskb, *skb;
8442 unsigned char *data;
8443 int len;
8444
8445 list_for_each_entry_safe(skb, nskb, &nft_net->notify_list, list) {
8446 if (!batch_skb) {
8447 new_batch:
8448 batch_skb = skb;
8449 len = NLMSG_GOODSIZE - skb->len;
8450 list_del(&skb->list);
8451 continue;
8452 }
8453 len -= skb->len;
8454 if (len > 0 && NFT_CB(skb).report == NFT_CB(batch_skb).report) {
8455 data = skb_put(batch_skb, skb->len);
8456 memcpy(data, skb->data, skb->len);
8457 list_del(&skb->list);
8458 kfree_skb(skb);
8459 continue;
8460 }
8461 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8462 NFT_CB(batch_skb).report, GFP_KERNEL);
8463 goto new_batch;
8464 }
8465
8466 if (batch_skb) {
8467 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
8468 NFT_CB(batch_skb).report, GFP_KERNEL);
8469 }
8470
8471 WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
8472 }
8473
8474 static int nf_tables_commit_audit_alloc(struct list_head *adl,
8475 struct nft_table *table)
8476 {
8477 struct nft_audit_data *adp;
8478
8479 list_for_each_entry(adp, adl, list) {
8480 if (adp->table == table)
8481 return 0;
8482 }
8483 adp = kzalloc(sizeof(*adp), GFP_KERNEL);
8484 if (!adp)
8485 return -ENOMEM;
8486 adp->table = table;
8487 list_add(&adp->list, adl);
8488 return 0;
8489 }
8490
8491 static void nf_tables_commit_audit_free(struct list_head *adl)
8492 {
8493 struct nft_audit_data *adp, *adn;
8494
8495 list_for_each_entry_safe(adp, adn, adl, list) {
8496 list_del(&adp->list);
8497 kfree(adp);
8498 }
8499 }
8500
8501 static void nf_tables_commit_audit_collect(struct list_head *adl,
8502 struct nft_table *table, u32 op)
8503 {
8504 struct nft_audit_data *adp;
8505
8506 list_for_each_entry(adp, adl, list) {
8507 if (adp->table == table)
8508 goto found;
8509 }
8510 WARN_ONCE(1, "table=%s not expected in commit list", table->name);
8511 return;
8512 found:
8513 adp->entries++;
8514 if (!adp->op || adp->op > op)
8515 adp->op = op;
8516 }
8517
8518 #define AUNFTABLENAMELEN (NFT_TABLE_MAXNAMELEN + 22)
8519
8520 static void nf_tables_commit_audit_log(struct list_head *adl, u32 generation)
8521 {
8522 struct nft_audit_data *adp, *adn;
8523 char aubuf[AUNFTABLENAMELEN];
8524
8525 list_for_each_entry_safe(adp, adn, adl, list) {
8526 snprintf(aubuf, AUNFTABLENAMELEN, "%s:%u", adp->table->name,
8527 generation);
8528 audit_log_nfcfg(aubuf, adp->table->family, adp->entries,
8529 nft2audit_op[adp->op], GFP_KERNEL);
8530 list_del(&adp->list);
8531 kfree(adp);
8532 }
8533 }
8534
8535 static int nf_tables_commit(struct net *net, struct sk_buff *skb)
8536 {
8537 struct nftables_pernet *nft_net = nft_pernet(net);
8538 struct nft_trans *trans, *next;
8539 struct nft_trans_elem *te;
8540 struct nft_chain *chain;
8541 struct nft_table *table;
8542 LIST_HEAD(adl);
8543 int err;
8544
8545 if (list_empty(&nft_net->commit_list)) {
8546 mutex_unlock(&nft_net->commit_mutex);
8547 return 0;
8548 }
8549
8550 /* 0. Validate ruleset, otherwise roll back for error reporting. */
8551 if (nf_tables_validate(net) < 0)
8552 return -EAGAIN;
8553
8554 err = nft_flow_rule_offload_commit(net);
8555 if (err < 0)
8556 return err;
8557
8558 /* 1. Allocate space for next generation rules_gen_X[] */
8559 list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8560 int ret;
8561
8562 ret = nf_tables_commit_audit_alloc(&adl, trans->ctx.table);
8563 if (ret) {
8564 nf_tables_commit_chain_prepare_cancel(net);
8565 nf_tables_commit_audit_free(&adl);
8566 return ret;
8567 }
8568 if (trans->msg_type == NFT_MSG_NEWRULE ||
8569 trans->msg_type == NFT_MSG_DELRULE) {
8570 chain = trans->ctx.chain;
8571
8572 ret = nf_tables_commit_chain_prepare(net, chain);
8573 if (ret < 0) {
8574 nf_tables_commit_chain_prepare_cancel(net);
8575 nf_tables_commit_audit_free(&adl);
8576 return ret;
8577 }
8578 }
8579 }
8580
8581 /* step 2. Make rules_gen_X visible to packet path */
8582 list_for_each_entry(table, &nft_net->tables, list) {
8583 list_for_each_entry(chain, &table->chains, list)
8584 nf_tables_commit_chain(net, chain);
8585 }
8586
8587 /*
8588 * Bump generation counter, invalidate any dump in progress.
8589 * Cannot fail after this point.
8590 */
8591 while (++nft_net->base_seq == 0)
8592 ;
8593
8594 /* step 3. Start new generation, rules_gen_X now in use. */
8595 net->nft.gencursor = nft_gencursor_next(net);
8596
8597 list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
8598 nf_tables_commit_audit_collect(&adl, trans->ctx.table,
8599 trans->msg_type);
8600 switch (trans->msg_type) {
8601 case NFT_MSG_NEWTABLE:
8602 if (nft_trans_table_update(trans)) {
8603 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
8604 nft_trans_destroy(trans);
8605 break;
8606 }
8607 if (trans->ctx.table->flags & NFT_TABLE_F_DORMANT)
8608 nf_tables_table_disable(net, trans->ctx.table);
8609
8610 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
8611 } else {
8612 nft_clear(net, trans->ctx.table);
8613 }
8614 nf_tables_table_notify(&trans->ctx, NFT_MSG_NEWTABLE);
8615 nft_trans_destroy(trans);
8616 break;
8617 case NFT_MSG_DELTABLE:
8618 list_del_rcu(&trans->ctx.table->list);
8619 nf_tables_table_notify(&trans->ctx, NFT_MSG_DELTABLE);
8620 break;
8621 case NFT_MSG_NEWCHAIN:
8622 if (nft_trans_chain_update(trans)) {
8623 nft_chain_commit_update(trans);
8624 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8625 /* trans destroyed after rcu grace period */
8626 } else {
8627 nft_chain_commit_drop_policy(trans);
8628 nft_clear(net, trans->ctx.chain);
8629 nf_tables_chain_notify(&trans->ctx, NFT_MSG_NEWCHAIN);
8630 nft_trans_destroy(trans);
8631 }
8632 break;
8633 case NFT_MSG_DELCHAIN:
8634 nft_chain_del(trans->ctx.chain);
8635 nf_tables_chain_notify(&trans->ctx, NFT_MSG_DELCHAIN);
8636 nf_tables_unregister_hook(trans->ctx.net,
8637 trans->ctx.table,
8638 trans->ctx.chain);
8639 break;
8640 case NFT_MSG_NEWRULE:
8641 nft_clear(trans->ctx.net, nft_trans_rule(trans));
8642 nf_tables_rule_notify(&trans->ctx,
8643 nft_trans_rule(trans),
8644 NFT_MSG_NEWRULE);
8645 nft_trans_destroy(trans);
8646 break;
8647 case NFT_MSG_DELRULE:
8648 list_del_rcu(&nft_trans_rule(trans)->list);
8649 nf_tables_rule_notify(&trans->ctx,
8650 nft_trans_rule(trans),
8651 NFT_MSG_DELRULE);
8652 nft_rule_expr_deactivate(&trans->ctx,
8653 nft_trans_rule(trans),
8654 NFT_TRANS_COMMIT);
8655 break;
8656 case NFT_MSG_NEWSET:
8657 nft_clear(net, nft_trans_set(trans));
8658 /* This avoids hitting -EBUSY when deleting the table
8659 * from the transaction.
8660 */
8661 if (nft_set_is_anonymous(nft_trans_set(trans)) &&
8662 !list_empty(&nft_trans_set(trans)->bindings))
8663 trans->ctx.table->use--;
8664
8665 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
8666 NFT_MSG_NEWSET, GFP_KERNEL);
8667 nft_trans_destroy(trans);
8668 break;
8669 case NFT_MSG_DELSET:
8670 list_del_rcu(&nft_trans_set(trans)->list);
8671 nf_tables_set_notify(&trans->ctx, nft_trans_set(trans),
8672 NFT_MSG_DELSET, GFP_KERNEL);
8673 break;
8674 case NFT_MSG_NEWSETELEM:
8675 te = (struct nft_trans_elem *)trans->data;
8676
8677 nft_setelem_activate(net, te->set, &te->elem);
8678 nf_tables_setelem_notify(&trans->ctx, te->set,
8679 &te->elem,
8680 NFT_MSG_NEWSETELEM);
8681 nft_trans_destroy(trans);
8682 break;
8683 case NFT_MSG_DELSETELEM:
8684 te = (struct nft_trans_elem *)trans->data;
8685
8686 nf_tables_setelem_notify(&trans->ctx, te->set,
8687 &te->elem,
8688 NFT_MSG_DELSETELEM);
8689 nft_setelem_remove(net, te->set, &te->elem);
8690 if (!nft_setelem_is_catchall(te->set, &te->elem)) {
8691 atomic_dec(&te->set->nelems);
8692 te->set->ndeact--;
8693 }
8694 break;
8695 case NFT_MSG_NEWOBJ:
8696 if (nft_trans_obj_update(trans)) {
8697 nft_obj_commit_update(trans);
8698 nf_tables_obj_notify(&trans->ctx,
8699 nft_trans_obj(trans),
8700 NFT_MSG_NEWOBJ);
8701 } else {
8702 nft_clear(net, nft_trans_obj(trans));
8703 nf_tables_obj_notify(&trans->ctx,
8704 nft_trans_obj(trans),
8705 NFT_MSG_NEWOBJ);
8706 nft_trans_destroy(trans);
8707 }
8708 break;
8709 case NFT_MSG_DELOBJ:
8710 nft_obj_del(nft_trans_obj(trans));
8711 nf_tables_obj_notify(&trans->ctx, nft_trans_obj(trans),
8712 NFT_MSG_DELOBJ);
8713 break;
8714 case NFT_MSG_NEWFLOWTABLE:
8715 if (nft_trans_flowtable_update(trans)) {
8716 nft_trans_flowtable(trans)->data.flags =
8717 nft_trans_flowtable_flags(trans);
8718 nf_tables_flowtable_notify(&trans->ctx,
8719 nft_trans_flowtable(trans),
8720 &nft_trans_flowtable_hooks(trans),
8721 NFT_MSG_NEWFLOWTABLE);
8722 list_splice(&nft_trans_flowtable_hooks(trans),
8723 &nft_trans_flowtable(trans)->hook_list);
8724 } else {
8725 nft_clear(net, nft_trans_flowtable(trans));
8726 nf_tables_flowtable_notify(&trans->ctx,
8727 nft_trans_flowtable(trans),
8728 &nft_trans_flowtable(trans)->hook_list,
8729 NFT_MSG_NEWFLOWTABLE);
8730 }
8731 nft_trans_destroy(trans);
8732 break;
8733 case NFT_MSG_DELFLOWTABLE:
8734 if (nft_trans_flowtable_update(trans)) {
8735 nft_flowtable_hooks_del(nft_trans_flowtable(trans),
8736 &nft_trans_flowtable_hooks(trans));
8737 nf_tables_flowtable_notify(&trans->ctx,
8738 nft_trans_flowtable(trans),
8739 &nft_trans_flowtable_hooks(trans),
8740 NFT_MSG_DELFLOWTABLE);
8741 nft_unregister_flowtable_net_hooks(net,
8742 &nft_trans_flowtable_hooks(trans));
8743 } else {
8744 list_del_rcu(&nft_trans_flowtable(trans)->list);
8745 nf_tables_flowtable_notify(&trans->ctx,
8746 nft_trans_flowtable(trans),
8747 &nft_trans_flowtable(trans)->hook_list,
8748 NFT_MSG_DELFLOWTABLE);
8749 nft_unregister_flowtable_net_hooks(net,
8750 &nft_trans_flowtable(trans)->hook_list);
8751 }
8752 break;
8753 }
8754 }
8755
8756 nft_commit_notify(net, NETLINK_CB(skb).portid);
8757 nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
8758 nf_tables_commit_audit_log(&adl, nft_net->base_seq);
8759 nf_tables_commit_release(net);
8760
8761 return 0;
8762 }
8763
8764 static void nf_tables_module_autoload(struct net *net)
8765 {
8766 struct nftables_pernet *nft_net = nft_pernet(net);
8767 struct nft_module_request *req, *next;
8768 LIST_HEAD(module_list);
8769
8770 list_splice_init(&nft_net->module_list, &module_list);
8771 mutex_unlock(&nft_net->commit_mutex);
8772 list_for_each_entry_safe(req, next, &module_list, list) {
8773 request_module("%s", req->module);
8774 req->done = true;
8775 }
8776 mutex_lock(&nft_net->commit_mutex);
8777 list_splice(&module_list, &nft_net->module_list);
8778 }
8779
8780 static void nf_tables_abort_release(struct nft_trans *trans)
8781 {
8782 switch (trans->msg_type) {
8783 case NFT_MSG_NEWTABLE:
8784 nf_tables_table_destroy(&trans->ctx);
8785 break;
8786 case NFT_MSG_NEWCHAIN:
8787 nf_tables_chain_destroy(&trans->ctx);
8788 break;
8789 case NFT_MSG_NEWRULE:
8790 nf_tables_rule_destroy(&trans->ctx, nft_trans_rule(trans));
8791 break;
8792 case NFT_MSG_NEWSET:
8793 nft_set_destroy(&trans->ctx, nft_trans_set(trans));
8794 break;
8795 case NFT_MSG_NEWSETELEM:
8796 nft_set_elem_destroy(nft_trans_elem_set(trans),
8797 nft_trans_elem(trans).priv, true);
8798 break;
8799 case NFT_MSG_NEWOBJ:
8800 nft_obj_destroy(&trans->ctx, nft_trans_obj(trans));
8801 break;
8802 case NFT_MSG_NEWFLOWTABLE:
8803 if (nft_trans_flowtable_update(trans))
8804 nft_flowtable_hooks_destroy(&nft_trans_flowtable_hooks(trans));
8805 else
8806 nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
8807 break;
8808 }
8809 kfree(trans);
8810 }
8811
8812 static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)
8813 {
8814 struct nftables_pernet *nft_net = nft_pernet(net);
8815 struct nft_trans *trans, *next;
8816 struct nft_trans_elem *te;
8817 struct nft_hook *hook;
8818
8819 if (action == NFNL_ABORT_VALIDATE &&
8820 nf_tables_validate(net) < 0)
8821 return -EAGAIN;
8822
8823 list_for_each_entry_safe_reverse(trans, next, &nft_net->commit_list,
8824 list) {
8825 switch (trans->msg_type) {
8826 case NFT_MSG_NEWTABLE:
8827 if (nft_trans_table_update(trans)) {
8828 if (!(trans->ctx.table->flags & __NFT_TABLE_F_UPDATE)) {
8829 nft_trans_destroy(trans);
8830 break;
8831 }
8832 if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_DORMANT) {
8833 nf_tables_table_disable(net, trans->ctx.table);
8834 trans->ctx.table->flags |= NFT_TABLE_F_DORMANT;
8835 } else if (trans->ctx.table->flags & __NFT_TABLE_F_WAS_AWAKEN) {
8836 trans->ctx.table->flags &= ~NFT_TABLE_F_DORMANT;
8837 }
8838 trans->ctx.table->flags &= ~__NFT_TABLE_F_UPDATE;
8839 nft_trans_destroy(trans);
8840 } else {
8841 list_del_rcu(&trans->ctx.table->list);
8842 }
8843 break;
8844 case NFT_MSG_DELTABLE:
8845 nft_clear(trans->ctx.net, trans->ctx.table);
8846 nft_trans_destroy(trans);
8847 break;
8848 case NFT_MSG_NEWCHAIN:
8849 if (nft_trans_chain_update(trans)) {
8850 free_percpu(nft_trans_chain_stats(trans));
8851 kfree(nft_trans_chain_name(trans));
8852 nft_trans_destroy(trans);
8853 } else {
8854 if (nft_chain_is_bound(trans->ctx.chain)) {
8855 nft_trans_destroy(trans);
8856 break;
8857 }
8858 trans->ctx.table->use--;
8859 nft_chain_del(trans->ctx.chain);
8860 nf_tables_unregister_hook(trans->ctx.net,
8861 trans->ctx.table,
8862 trans->ctx.chain);
8863 }
8864 break;
8865 case NFT_MSG_DELCHAIN:
8866 trans->ctx.table->use++;
8867 nft_clear(trans->ctx.net, trans->ctx.chain);
8868 nft_trans_destroy(trans);
8869 break;
8870 case NFT_MSG_NEWRULE:
8871 trans->ctx.chain->use--;
8872 list_del_rcu(&nft_trans_rule(trans)->list);
8873 nft_rule_expr_deactivate(&trans->ctx,
8874 nft_trans_rule(trans),
8875 NFT_TRANS_ABORT);
8876 if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
8877 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
8878 break;
8879 case NFT_MSG_DELRULE:
8880 trans->ctx.chain->use++;
8881 nft_clear(trans->ctx.net, nft_trans_rule(trans));
8882 nft_rule_expr_activate(&trans->ctx, nft_trans_rule(trans));
8883 if (trans->ctx.chain->flags & NFT_CHAIN_HW_OFFLOAD)
8884 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
8885
8886 nft_trans_destroy(trans);
8887 break;
8888 case NFT_MSG_NEWSET:
8889 trans->ctx.table->use--;
8890 if (nft_trans_set_bound(trans)) {
8891 nft_trans_destroy(trans);
8892 break;
8893 }
8894 list_del_rcu(&nft_trans_set(trans)->list);
8895 break;
8896 case NFT_MSG_DELSET:
8897 trans->ctx.table->use++;
8898 nft_clear(trans->ctx.net, nft_trans_set(trans));
8899 nft_trans_destroy(trans);
8900 break;
8901 case NFT_MSG_NEWSETELEM:
8902 if (nft_trans_elem_set_bound(trans)) {
8903 nft_trans_destroy(trans);
8904 break;
8905 }
8906 te = (struct nft_trans_elem *)trans->data;
8907 nft_setelem_remove(net, te->set, &te->elem);
8908 if (!nft_setelem_is_catchall(te->set, &te->elem))
8909 atomic_dec(&te->set->nelems);
8910 break;
8911 case NFT_MSG_DELSETELEM:
8912 te = (struct nft_trans_elem *)trans->data;
8913
8914 nft_setelem_data_activate(net, te->set, &te->elem);
8915 nft_setelem_activate(net, te->set, &te->elem);
8916 if (!nft_setelem_is_catchall(te->set, &te->elem))
8917 te->set->ndeact--;
8918
8919 nft_trans_destroy(trans);
8920 break;
8921 case NFT_MSG_NEWOBJ:
8922 if (nft_trans_obj_update(trans)) {
8923 nft_obj_destroy(&trans->ctx, nft_trans_obj_newobj(trans));
8924 nft_trans_destroy(trans);
8925 } else {
8926 trans->ctx.table->use--;
8927 nft_obj_del(nft_trans_obj(trans));
8928 }
8929 break;
8930 case NFT_MSG_DELOBJ:
8931 trans->ctx.table->use++;
8932 nft_clear(trans->ctx.net, nft_trans_obj(trans));
8933 nft_trans_destroy(trans);
8934 break;
8935 case NFT_MSG_NEWFLOWTABLE:
8936 if (nft_trans_flowtable_update(trans)) {
8937 nft_unregister_flowtable_net_hooks(net,
8938 &nft_trans_flowtable_hooks(trans));
8939 } else {
8940 trans->ctx.table->use--;
8941 list_del_rcu(&nft_trans_flowtable(trans)->list);
8942 nft_unregister_flowtable_net_hooks(net,
8943 &nft_trans_flowtable(trans)->hook_list);
8944 }
8945 break;
8946 case NFT_MSG_DELFLOWTABLE:
8947 if (nft_trans_flowtable_update(trans)) {
8948 list_for_each_entry(hook, &nft_trans_flowtable(trans)->hook_list, list)
8949 hook->inactive = false;
8950 } else {
8951 trans->ctx.table->use++;
8952 nft_clear(trans->ctx.net, nft_trans_flowtable(trans));
8953 }
8954 nft_trans_destroy(trans);
8955 break;
8956 }
8957 }
8958
8959 synchronize_rcu();
8960
8961 list_for_each_entry_safe_reverse(trans, next,
8962 &nft_net->commit_list, list) {
8963 list_del(&trans->list);
8964 nf_tables_abort_release(trans);
8965 }
8966
8967 if (action == NFNL_ABORT_AUTOLOAD)
8968 nf_tables_module_autoload(net);
8969 else
8970 nf_tables_module_autoload_cleanup(net);
8971
8972 return 0;
8973 }
8974
8975 static void nf_tables_cleanup(struct net *net)
8976 {
8977 nft_validate_state_update(net, NFT_VALIDATE_SKIP);
8978 }
8979
8980 static int nf_tables_abort(struct net *net, struct sk_buff *skb,
8981 enum nfnl_abort_action action)
8982 {
8983 struct nftables_pernet *nft_net = nft_pernet(net);
8984 int ret = __nf_tables_abort(net, action);
8985
8986 mutex_unlock(&nft_net->commit_mutex);
8987
8988 return ret;
8989 }
8990
8991 static bool nf_tables_valid_genid(struct net *net, u32 genid)
8992 {
8993 struct nftables_pernet *nft_net = nft_pernet(net);
8994 bool genid_ok;
8995
8996 mutex_lock(&nft_net->commit_mutex);
8997
8998 genid_ok = genid == 0 || nft_net->base_seq == genid;
8999 if (!genid_ok)
9000 mutex_unlock(&nft_net->commit_mutex);
9001
9002 /* else, commit mutex has to be released by commit or abort function */
9003 return genid_ok;
9004 }
9005
9006 static const struct nfnetlink_subsystem nf_tables_subsys = {
9007 .name = "nf_tables",
9008 .subsys_id = NFNL_SUBSYS_NFTABLES,
9009 .cb_count = NFT_MSG_MAX,
9010 .cb = nf_tables_cb,
9011 .commit = nf_tables_commit,
9012 .abort = nf_tables_abort,
9013 .cleanup = nf_tables_cleanup,
9014 .valid_genid = nf_tables_valid_genid,
9015 .owner = THIS_MODULE,
9016 };
9017
9018 int nft_chain_validate_dependency(const struct nft_chain *chain,
9019 enum nft_chain_types type)
9020 {
9021 const struct nft_base_chain *basechain;
9022
9023 if (nft_is_base_chain(chain)) {
9024 basechain = nft_base_chain(chain);
9025 if (basechain->type->type != type)
9026 return -EOPNOTSUPP;
9027 }
9028 return 0;
9029 }
9030 EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
9031
9032 int nft_chain_validate_hooks(const struct nft_chain *chain,
9033 unsigned int hook_flags)
9034 {
9035 struct nft_base_chain *basechain;
9036
9037 if (nft_is_base_chain(chain)) {
9038 basechain = nft_base_chain(chain);
9039
9040 if ((1 << basechain->ops.hooknum) & hook_flags)
9041 return 0;
9042
9043 return -EOPNOTSUPP;
9044 }
9045
9046 return 0;
9047 }
9048 EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
9049
9050 /*
9051 * Loop detection - walk through the ruleset beginning at the destination chain
9052 * of a new jump until either the source chain is reached (loop) or all
9053 * reachable chains have been traversed.
9054 *
9055 * The loop check is performed whenever a new jump verdict is added to an
9056 * expression or verdict map or a verdict map is bound to a new chain.
9057 */
9058
9059 static int nf_tables_check_loops(const struct nft_ctx *ctx,
9060 const struct nft_chain *chain);
9061
9062 static int nft_check_loops(const struct nft_ctx *ctx,
9063 const struct nft_set_ext *ext)
9064 {
9065 const struct nft_data *data;
9066 int ret;
9067
9068 data = nft_set_ext_data(ext);
9069 switch (data->verdict.code) {
9070 case NFT_JUMP:
9071 case NFT_GOTO:
9072 ret = nf_tables_check_loops(ctx, data->verdict.chain);
9073 break;
9074 default:
9075 ret = 0;
9076 break;
9077 }
9078
9079 return ret;
9080 }
9081
9082 static int nf_tables_loop_check_setelem(const struct nft_ctx *ctx,
9083 struct nft_set *set,
9084 const struct nft_set_iter *iter,
9085 struct nft_set_elem *elem)
9086 {
9087 const struct nft_set_ext *ext = nft_set_elem_ext(set, elem->priv);
9088
9089 if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
9090 *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
9091 return 0;
9092
9093 return nft_check_loops(ctx, ext);
9094 }
9095
9096 static int nft_set_catchall_loops(const struct nft_ctx *ctx,
9097 struct nft_set *set)
9098 {
9099 u8 genmask = nft_genmask_next(ctx->net);
9100 struct nft_set_elem_catchall *catchall;
9101 struct nft_set_ext *ext;
9102 int ret = 0;
9103
9104 list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
9105 ext = nft_set_elem_ext(set, catchall->elem);
9106 if (!nft_set_elem_active(ext, genmask))
9107 continue;
9108
9109 ret = nft_check_loops(ctx, ext);
9110 if (ret < 0)
9111 return ret;
9112 }
9113
9114 return ret;
9115 }
9116
9117 static int nf_tables_check_loops(const struct nft_ctx *ctx,
9118 const struct nft_chain *chain)
9119 {
9120 const struct nft_rule *rule;
9121 const struct nft_expr *expr, *last;
9122 struct nft_set *set;
9123 struct nft_set_binding *binding;
9124 struct nft_set_iter iter;
9125
9126 if (ctx->chain == chain)
9127 return -ELOOP;
9128
9129 list_for_each_entry(rule, &chain->rules, list) {
9130 nft_rule_for_each_expr(expr, last, rule) {
9131 struct nft_immediate_expr *priv;
9132 const struct nft_data *data;
9133 int err;
9134
9135 if (strcmp(expr->ops->type->name, "immediate"))
9136 continue;
9137
9138 priv = nft_expr_priv(expr);
9139 if (priv->dreg != NFT_REG_VERDICT)
9140 continue;
9141
9142 data = &priv->data;
9143 switch (data->verdict.code) {
9144 case NFT_JUMP:
9145 case NFT_GOTO:
9146 err = nf_tables_check_loops(ctx,
9147 data->verdict.chain);
9148 if (err < 0)
9149 return err;
9150 break;
9151 default:
9152 break;
9153 }
9154 }
9155 }
9156
9157 list_for_each_entry(set, &ctx->table->sets, list) {
9158 if (!nft_is_active_next(ctx->net, set))
9159 continue;
9160 if (!(set->flags & NFT_SET_MAP) ||
9161 set->dtype != NFT_DATA_VERDICT)
9162 continue;
9163
9164 list_for_each_entry(binding, &set->bindings, list) {
9165 if (!(binding->flags & NFT_SET_MAP) ||
9166 binding->chain != chain)
9167 continue;
9168
9169 iter.genmask = nft_genmask_next(ctx->net);
9170 iter.skip = 0;
9171 iter.count = 0;
9172 iter.err = 0;
9173 iter.fn = nf_tables_loop_check_setelem;
9174
9175 set->ops->walk(ctx, set, &iter);
9176 if (!iter.err)
9177 iter.err = nft_set_catchall_loops(ctx, set);
9178
9179 if (iter.err < 0)
9180 return iter.err;
9181 }
9182 }
9183
9184 return 0;
9185 }
9186
9187 /**
9188 * nft_parse_u32_check - fetch u32 attribute and check for maximum value
9189 *
9190 * @attr: netlink attribute to fetch value from
9191 * @max: maximum value to be stored in dest
9192 * @dest: pointer to the variable
9193 *
9194 * Parse, check and store a given u32 netlink attribute into variable.
9195 * This function returns -ERANGE if the value goes over maximum value.
9196 * Otherwise a 0 is returned and the attribute value is stored in the
9197 * destination variable.
9198 */
9199 int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
9200 {
9201 u32 val;
9202
9203 val = ntohl(nla_get_be32(attr));
9204 if (val > max)
9205 return -ERANGE;
9206
9207 *dest = val;
9208 return 0;
9209 }
9210 EXPORT_SYMBOL_GPL(nft_parse_u32_check);
9211
9212 static int nft_parse_register(const struct nlattr *attr, u32 *preg)
9213 {
9214 unsigned int reg;
9215
9216 reg = ntohl(nla_get_be32(attr));
9217 switch (reg) {
9218 case NFT_REG_VERDICT...NFT_REG_4:
9219 *preg = reg * NFT_REG_SIZE / NFT_REG32_SIZE;
9220 break;
9221 case NFT_REG32_00...NFT_REG32_15:
9222 *preg = reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
9223 break;
9224 default:
9225 return -ERANGE;
9226 }
9227
9228 return 0;
9229 }
9230
9231 /**
9232 * nft_dump_register - dump a register value to a netlink attribute
9233 *
9234 * @skb: socket buffer
9235 * @attr: attribute number
9236 * @reg: register number
9237 *
9238 * Construct a netlink attribute containing the register number. For
9239 * compatibility reasons, register numbers being a multiple of 4 are
9240 * translated to the corresponding 128 bit register numbers.
9241 */
9242 int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
9243 {
9244 if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
9245 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
9246 else
9247 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
9248
9249 return nla_put_be32(skb, attr, htonl(reg));
9250 }
9251 EXPORT_SYMBOL_GPL(nft_dump_register);
9252
9253 static int nft_validate_register_load(enum nft_registers reg, unsigned int len)
9254 {
9255 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9256 return -EINVAL;
9257 if (len == 0)
9258 return -EINVAL;
9259 if (reg * NFT_REG32_SIZE + len > sizeof_field(struct nft_regs, data))
9260 return -ERANGE;
9261
9262 return 0;
9263 }
9264
9265 int nft_parse_register_load(const struct nlattr *attr, u8 *sreg, u32 len)
9266 {
9267 u32 reg;
9268 int err;
9269
9270 err = nft_parse_register(attr, &reg);
9271 if (err < 0)
9272 return err;
9273
9274 err = nft_validate_register_load(reg, len);
9275 if (err < 0)
9276 return err;
9277
9278 *sreg = reg;
9279 return 0;
9280 }
9281 EXPORT_SYMBOL_GPL(nft_parse_register_load);
9282
9283 static int nft_validate_register_store(const struct nft_ctx *ctx,
9284 enum nft_registers reg,
9285 const struct nft_data *data,
9286 enum nft_data_types type,
9287 unsigned int len)
9288 {
9289 int err;
9290
9291 switch (reg) {
9292 case NFT_REG_VERDICT:
9293 if (type != NFT_DATA_VERDICT)
9294 return -EINVAL;
9295
9296 if (data != NULL &&
9297 (data->verdict.code == NFT_GOTO ||
9298 data->verdict.code == NFT_JUMP)) {
9299 err = nf_tables_check_loops(ctx, data->verdict.chain);
9300 if (err < 0)
9301 return err;
9302 }
9303
9304 return 0;
9305 default:
9306 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
9307 return -EINVAL;
9308 if (len == 0)
9309 return -EINVAL;
9310 if (reg * NFT_REG32_SIZE + len >
9311 sizeof_field(struct nft_regs, data))
9312 return -ERANGE;
9313
9314 if (data != NULL && type != NFT_DATA_VALUE)
9315 return -EINVAL;
9316 return 0;
9317 }
9318 }
9319
9320 int nft_parse_register_store(const struct nft_ctx *ctx,
9321 const struct nlattr *attr, u8 *dreg,
9322 const struct nft_data *data,
9323 enum nft_data_types type, unsigned int len)
9324 {
9325 int err;
9326 u32 reg;
9327
9328 err = nft_parse_register(attr, &reg);
9329 if (err < 0)
9330 return err;
9331
9332 err = nft_validate_register_store(ctx, reg, data, type, len);
9333 if (err < 0)
9334 return err;
9335
9336 *dreg = reg;
9337 return 0;
9338 }
9339 EXPORT_SYMBOL_GPL(nft_parse_register_store);
9340
9341 static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
9342 [NFTA_VERDICT_CODE] = { .type = NLA_U32 },
9343 [NFTA_VERDICT_CHAIN] = { .type = NLA_STRING,
9344 .len = NFT_CHAIN_MAXNAMELEN - 1 },
9345 [NFTA_VERDICT_CHAIN_ID] = { .type = NLA_U32 },
9346 };
9347
9348 static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
9349 struct nft_data_desc *desc, const struct nlattr *nla)
9350 {
9351 u8 genmask = nft_genmask_next(ctx->net);
9352 struct nlattr *tb[NFTA_VERDICT_MAX + 1];
9353 struct nft_chain *chain;
9354 int err;
9355
9356 err = nla_parse_nested_deprecated(tb, NFTA_VERDICT_MAX, nla,
9357 nft_verdict_policy, NULL);
9358 if (err < 0)
9359 return err;
9360
9361 if (!tb[NFTA_VERDICT_CODE])
9362 return -EINVAL;
9363 data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
9364
9365 switch (data->verdict.code) {
9366 default:
9367 switch (data->verdict.code & NF_VERDICT_MASK) {
9368 case NF_ACCEPT:
9369 case NF_DROP:
9370 case NF_QUEUE:
9371 break;
9372 default:
9373 return -EINVAL;
9374 }
9375 fallthrough;
9376 case NFT_CONTINUE:
9377 case NFT_BREAK:
9378 case NFT_RETURN:
9379 break;
9380 case NFT_JUMP:
9381 case NFT_GOTO:
9382 if (tb[NFTA_VERDICT_CHAIN]) {
9383 chain = nft_chain_lookup(ctx->net, ctx->table,
9384 tb[NFTA_VERDICT_CHAIN],
9385 genmask);
9386 } else if (tb[NFTA_VERDICT_CHAIN_ID]) {
9387 chain = nft_chain_lookup_byid(ctx->net,
9388 tb[NFTA_VERDICT_CHAIN_ID]);
9389 if (IS_ERR(chain))
9390 return PTR_ERR(chain);
9391 } else {
9392 return -EINVAL;
9393 }
9394
9395 if (IS_ERR(chain))
9396 return PTR_ERR(chain);
9397 if (nft_is_base_chain(chain))
9398 return -EOPNOTSUPP;
9399
9400 chain->use++;
9401 data->verdict.chain = chain;
9402 break;
9403 }
9404
9405 desc->len = sizeof(data->verdict);
9406 desc->type = NFT_DATA_VERDICT;
9407 return 0;
9408 }
9409
9410 static void nft_verdict_uninit(const struct nft_data *data)
9411 {
9412 struct nft_chain *chain;
9413 struct nft_rule *rule;
9414
9415 switch (data->verdict.code) {
9416 case NFT_JUMP:
9417 case NFT_GOTO:
9418 chain = data->verdict.chain;
9419 chain->use--;
9420
9421 if (!nft_chain_is_bound(chain))
9422 break;
9423
9424 chain->table->use--;
9425 list_for_each_entry(rule, &chain->rules, list)
9426 chain->use--;
9427
9428 nft_chain_del(chain);
9429 break;
9430 }
9431 }
9432
9433 int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
9434 {
9435 struct nlattr *nest;
9436
9437 nest = nla_nest_start_noflag(skb, type);
9438 if (!nest)
9439 goto nla_put_failure;
9440
9441 if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
9442 goto nla_put_failure;
9443
9444 switch (v->code) {
9445 case NFT_JUMP:
9446 case NFT_GOTO:
9447 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
9448 v->chain->name))
9449 goto nla_put_failure;
9450 }
9451 nla_nest_end(skb, nest);
9452 return 0;
9453
9454 nla_put_failure:
9455 return -1;
9456 }
9457
9458 static int nft_value_init(const struct nft_ctx *ctx,
9459 struct nft_data *data, unsigned int size,
9460 struct nft_data_desc *desc, const struct nlattr *nla)
9461 {
9462 unsigned int len;
9463
9464 len = nla_len(nla);
9465 if (len == 0)
9466 return -EINVAL;
9467 if (len > size)
9468 return -EOVERFLOW;
9469
9470 nla_memcpy(data->data, nla, len);
9471 desc->type = NFT_DATA_VALUE;
9472 desc->len = len;
9473 return 0;
9474 }
9475
9476 static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
9477 unsigned int len)
9478 {
9479 return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
9480 }
9481
9482 static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
9483 [NFTA_DATA_VALUE] = { .type = NLA_BINARY },
9484 [NFTA_DATA_VERDICT] = { .type = NLA_NESTED },
9485 };
9486
9487 /**
9488 * nft_data_init - parse nf_tables data netlink attributes
9489 *
9490 * @ctx: context of the expression using the data
9491 * @data: destination struct nft_data
9492 * @size: maximum data length
9493 * @desc: data description
9494 * @nla: netlink attribute containing data
9495 *
9496 * Parse the netlink data attributes and initialize a struct nft_data.
9497 * The type and length of data are returned in the data description.
9498 *
9499 * The caller can indicate that it only wants to accept data of type
9500 * NFT_DATA_VALUE by passing NULL for the ctx argument.
9501 */
9502 int nft_data_init(const struct nft_ctx *ctx,
9503 struct nft_data *data, unsigned int size,
9504 struct nft_data_desc *desc, const struct nlattr *nla)
9505 {
9506 struct nlattr *tb[NFTA_DATA_MAX + 1];
9507 int err;
9508
9509 err = nla_parse_nested_deprecated(tb, NFTA_DATA_MAX, nla,
9510 nft_data_policy, NULL);
9511 if (err < 0)
9512 return err;
9513
9514 if (tb[NFTA_DATA_VALUE])
9515 return nft_value_init(ctx, data, size, desc,
9516 tb[NFTA_DATA_VALUE]);
9517 if (tb[NFTA_DATA_VERDICT] && ctx != NULL)
9518 return nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
9519 return -EINVAL;
9520 }
9521 EXPORT_SYMBOL_GPL(nft_data_init);
9522
9523 /**
9524 * nft_data_release - release a nft_data item
9525 *
9526 * @data: struct nft_data to release
9527 * @type: type of data
9528 *
9529 * Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
9530 * all others need to be released by calling this function.
9531 */
9532 void nft_data_release(const struct nft_data *data, enum nft_data_types type)
9533 {
9534 if (type < NFT_DATA_VERDICT)
9535 return;
9536 switch (type) {
9537 case NFT_DATA_VERDICT:
9538 return nft_verdict_uninit(data);
9539 default:
9540 WARN_ON(1);
9541 }
9542 }
9543 EXPORT_SYMBOL_GPL(nft_data_release);
9544
9545 int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
9546 enum nft_data_types type, unsigned int len)
9547 {
9548 struct nlattr *nest;
9549 int err;
9550
9551 nest = nla_nest_start_noflag(skb, attr);
9552 if (nest == NULL)
9553 return -1;
9554
9555 switch (type) {
9556 case NFT_DATA_VALUE:
9557 err = nft_value_dump(skb, data, len);
9558 break;
9559 case NFT_DATA_VERDICT:
9560 err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
9561 break;
9562 default:
9563 err = -EINVAL;
9564 WARN_ON(1);
9565 }
9566
9567 nla_nest_end(skb, nest);
9568 return err;
9569 }
9570 EXPORT_SYMBOL_GPL(nft_data_dump);
9571
9572 int __nft_release_basechain(struct nft_ctx *ctx)
9573 {
9574 struct nft_rule *rule, *nr;
9575
9576 if (WARN_ON(!nft_is_base_chain(ctx->chain)))
9577 return 0;
9578
9579 nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
9580 list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
9581 list_del(&rule->list);
9582 ctx->chain->use--;
9583 nf_tables_rule_release(ctx, rule);
9584 }
9585 nft_chain_del(ctx->chain);
9586 ctx->table->use--;
9587 nf_tables_chain_destroy(ctx);
9588
9589 return 0;
9590 }
9591 EXPORT_SYMBOL_GPL(__nft_release_basechain);
9592
9593 static void __nft_release_hook(struct net *net, struct nft_table *table)
9594 {
9595 struct nft_flowtable *flowtable;
9596 struct nft_chain *chain;
9597
9598 list_for_each_entry(chain, &table->chains, list)
9599 nf_tables_unregister_hook(net, table, chain);
9600 list_for_each_entry(flowtable, &table->flowtables, list)
9601 nft_unregister_flowtable_net_hooks(net, &flowtable->hook_list);
9602 }
9603
9604 static void __nft_release_hooks(struct net *net)
9605 {
9606 struct nftables_pernet *nft_net = nft_pernet(net);
9607 struct nft_table *table;
9608
9609 list_for_each_entry(table, &nft_net->tables, list) {
9610 if (nft_table_has_owner(table))
9611 continue;
9612
9613 __nft_release_hook(net, table);
9614 }
9615 }
9616
9617 static void __nft_release_table(struct net *net, struct nft_table *table)
9618 {
9619 struct nft_flowtable *flowtable, *nf;
9620 struct nft_chain *chain, *nc;
9621 struct nft_object *obj, *ne;
9622 struct nft_rule *rule, *nr;
9623 struct nft_set *set, *ns;
9624 struct nft_ctx ctx = {
9625 .net = net,
9626 .family = NFPROTO_NETDEV,
9627 };
9628
9629 ctx.family = table->family;
9630 ctx.table = table;
9631 list_for_each_entry(chain, &table->chains, list) {
9632 ctx.chain = chain;
9633 list_for_each_entry_safe(rule, nr, &chain->rules, list) {
9634 list_del(&rule->list);
9635 chain->use--;
9636 nf_tables_rule_release(&ctx, rule);
9637 }
9638 }
9639 list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
9640 list_del(&flowtable->list);
9641 table->use--;
9642 nf_tables_flowtable_destroy(flowtable);
9643 }
9644 list_for_each_entry_safe(set, ns, &table->sets, list) {
9645 list_del(&set->list);
9646 table->use--;
9647 nft_set_destroy(&ctx, set);
9648 }
9649 list_for_each_entry_safe(obj, ne, &table->objects, list) {
9650 nft_obj_del(obj);
9651 table->use--;
9652 nft_obj_destroy(&ctx, obj);
9653 }
9654 list_for_each_entry_safe(chain, nc, &table->chains, list) {
9655 ctx.chain = chain;
9656 nft_chain_del(chain);
9657 table->use--;
9658 nf_tables_chain_destroy(&ctx);
9659 }
9660 nf_tables_table_destroy(&ctx);
9661 }
9662
9663 static void __nft_release_tables(struct net *net)
9664 {
9665 struct nftables_pernet *nft_net = nft_pernet(net);
9666 struct nft_table *table, *nt;
9667
9668 list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
9669 if (nft_table_has_owner(table))
9670 continue;
9671
9672 list_del(&table->list);
9673
9674 __nft_release_table(net, table);
9675 }
9676 }
9677
9678 static int nft_rcv_nl_event(struct notifier_block *this, unsigned long event,
9679 void *ptr)
9680 {
9681 struct nft_table *table, *to_delete[8];
9682 struct nftables_pernet *nft_net;
9683 struct netlink_notify *n = ptr;
9684 struct net *net = n->net;
9685 unsigned int deleted;
9686 bool restart = false;
9687
9688 if (event != NETLINK_URELEASE || n->protocol != NETLINK_NETFILTER)
9689 return NOTIFY_DONE;
9690
9691 nft_net = nft_pernet(net);
9692 deleted = 0;
9693 mutex_lock(&nft_net->commit_mutex);
9694 again:
9695 list_for_each_entry(table, &nft_net->tables, list) {
9696 if (nft_table_has_owner(table) &&
9697 n->portid == table->nlpid) {
9698 __nft_release_hook(net, table);
9699 list_del_rcu(&table->list);
9700 to_delete[deleted++] = table;
9701 if (deleted >= ARRAY_SIZE(to_delete))
9702 break;
9703 }
9704 }
9705 if (deleted) {
9706 restart = deleted >= ARRAY_SIZE(to_delete);
9707 synchronize_rcu();
9708 while (deleted)
9709 __nft_release_table(net, to_delete[--deleted]);
9710
9711 if (restart)
9712 goto again;
9713 }
9714 mutex_unlock(&nft_net->commit_mutex);
9715
9716 return NOTIFY_DONE;
9717 }
9718
9719 static struct notifier_block nft_nl_notifier = {
9720 .notifier_call = nft_rcv_nl_event,
9721 };
9722
9723 static int __net_init nf_tables_init_net(struct net *net)
9724 {
9725 struct nftables_pernet *nft_net = nft_pernet(net);
9726
9727 INIT_LIST_HEAD(&nft_net->tables);
9728 INIT_LIST_HEAD(&nft_net->commit_list);
9729 INIT_LIST_HEAD(&nft_net->module_list);
9730 INIT_LIST_HEAD(&nft_net->notify_list);
9731 mutex_init(&nft_net->commit_mutex);
9732 nft_net->base_seq = 1;
9733 nft_net->validate_state = NFT_VALIDATE_SKIP;
9734
9735 return 0;
9736 }
9737
9738 static void __net_exit nf_tables_pre_exit_net(struct net *net)
9739 {
9740 __nft_release_hooks(net);
9741 }
9742
9743 static void __net_exit nf_tables_exit_net(struct net *net)
9744 {
9745 struct nftables_pernet *nft_net = nft_pernet(net);
9746
9747 mutex_lock(&nft_net->commit_mutex);
9748 if (!list_empty(&nft_net->commit_list))
9749 __nf_tables_abort(net, NFNL_ABORT_NONE);
9750 __nft_release_tables(net);
9751 mutex_unlock(&nft_net->commit_mutex);
9752 WARN_ON_ONCE(!list_empty(&nft_net->tables));
9753 WARN_ON_ONCE(!list_empty(&nft_net->module_list));
9754 WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
9755 }
9756
9757 static struct pernet_operations nf_tables_net_ops = {
9758 .init = nf_tables_init_net,
9759 .pre_exit = nf_tables_pre_exit_net,
9760 .exit = nf_tables_exit_net,
9761 .id = &nf_tables_net_id,
9762 .size = sizeof(struct nftables_pernet),
9763 };
9764
9765 static int __init nf_tables_module_init(void)
9766 {
9767 int err;
9768
9769 err = register_pernet_subsys(&nf_tables_net_ops);
9770 if (err < 0)
9771 return err;
9772
9773 err = nft_chain_filter_init();
9774 if (err < 0)
9775 goto err_chain_filter;
9776
9777 err = nf_tables_core_module_init();
9778 if (err < 0)
9779 goto err_core_module;
9780
9781 err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
9782 if (err < 0)
9783 goto err_netdev_notifier;
9784
9785 err = rhltable_init(&nft_objname_ht, &nft_objname_ht_params);
9786 if (err < 0)
9787 goto err_rht_objname;
9788
9789 err = nft_offload_init();
9790 if (err < 0)
9791 goto err_offload;
9792
9793 err = netlink_register_notifier(&nft_nl_notifier);
9794 if (err < 0)
9795 goto err_netlink_notifier;
9796
9797 /* must be last */
9798 err = nfnetlink_subsys_register(&nf_tables_subsys);
9799 if (err < 0)
9800 goto err_nfnl_subsys;
9801
9802 nft_chain_route_init();
9803
9804 return err;
9805
9806 err_nfnl_subsys:
9807 netlink_unregister_notifier(&nft_nl_notifier);
9808 err_netlink_notifier:
9809 nft_offload_exit();
9810 err_offload:
9811 rhltable_destroy(&nft_objname_ht);
9812 err_rht_objname:
9813 unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
9814 err_netdev_notifier:
9815 nf_tables_core_module_exit();
9816 err_core_module:
9817 nft_chain_filter_fini();
9818 err_chain_filter:
9819 unregister_pernet_subsys(&nf_tables_net_ops);
9820 return err;
9821 }
9822
9823 static void __exit nf_tables_module_exit(void)
9824 {
9825 nfnetlink_subsys_unregister(&nf_tables_subsys);
9826 netlink_unregister_notifier(&nft_nl_notifier);
9827 nft_offload_exit();
9828 unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
9829 nft_chain_filter_fini();
9830 nft_chain_route_fini();
9831 unregister_pernet_subsys(&nf_tables_net_ops);
9832 cancel_work_sync(&trans_destroy_work);
9833 rcu_barrier();
9834 rhltable_destroy(&nft_objname_ht);
9835 nf_tables_core_module_exit();
9836 }
9837
9838 module_init(nf_tables_module_init);
9839 module_exit(nf_tables_module_exit);
9840
9841 MODULE_LICENSE("GPL");
9842 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
9843 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);