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