~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/net/netfilter/nf_tables_api.c

Version: ~ [ linux-6.11-rc3 ] ~ [ linux-6.10.4 ] ~ [ linux-6.9.12 ] ~ [ linux-6.8.12 ] ~ [ linux-6.7.12 ] ~ [ linux-6.6.45 ] ~ [ linux-6.5.13 ] ~ [ linux-6.4.16 ] ~ [ linux-6.3.13 ] ~ [ linux-6.2.16 ] ~ [ linux-6.1.104 ] ~ [ linux-6.0.19 ] ~ [ linux-5.19.17 ] ~ [ linux-5.18.19 ] ~ [ linux-5.17.15 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.164 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.223 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.281 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.319 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.336 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.337 ] ~ [ linux-4.4.302 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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 #define NFT_SET_MAX_ANONLEN 16
 28 
 29 unsigned int nf_tables_net_id __read_mostly;
 30 
 31 static LIST_HEAD(nf_tables_expressions);
 32 static LIST_HEAD(nf_tables_objects);
 33 static LIST_HEAD(nf_tables_flowtables);
 34 static LIST_HEAD(nf_tables_destroy_list);
 35 static LIST_HEAD(nf_tables_gc_list);
 36 static DEFINE_SPINLOCK(nf_tables_destroy_list_lock);
 37 static DEFINE_SPINLOCK(nf_tables_gc_list_lock);
 38 
 39 enum {
 40         NFT_VALIDATE_SKIP       = 0,
 41         NFT_VALIDATE_NEED,
 42         NFT_VALIDATE_DO,
 43 };
 44 
 45 static struct rhltable nft_objname_ht;
 46 
 47 static u32 nft_chain_hash(const void *data, u32 len, u32 seed);
 48 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed);
 49 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *, const void *);
 50 
 51 static u32 nft_objname_hash(const void *data, u32 len, u32 seed);
 52 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed);
 53 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *, const void *);
 54 
 55 static const struct rhashtable_params nft_chain_ht_params = {
 56         .head_offset            = offsetof(struct nft_chain, rhlhead),
 57         .key_offset             = offsetof(struct nft_chain, name),
 58         .hashfn                 = nft_chain_hash,
 59         .obj_hashfn             = nft_chain_hash_obj,
 60         .obj_cmpfn              = nft_chain_hash_cmp,
 61         .automatic_shrinking    = true,
 62 };
 63 
 64 static const struct rhashtable_params nft_objname_ht_params = {
 65         .head_offset            = offsetof(struct nft_object, rhlhead),
 66         .key_offset             = offsetof(struct nft_object, key),
 67         .hashfn                 = nft_objname_hash,
 68         .obj_hashfn             = nft_objname_hash_obj,
 69         .obj_cmpfn              = nft_objname_hash_cmp,
 70         .automatic_shrinking    = true,
 71 };
 72 
 73 struct nft_audit_data {
 74         struct nft_table *table;
 75         int entries;
 76         int op;
 77         struct list_head list;
 78 };
 79 
 80 static const u8 nft2audit_op[NFT_MSG_MAX] = { // enum nf_tables_msg_types
 81         [NFT_MSG_NEWTABLE]      = AUDIT_NFT_OP_TABLE_REGISTER,
 82         [NFT_MSG_GETTABLE]      = AUDIT_NFT_OP_INVALID,
 83         [NFT_MSG_DELTABLE]      = AUDIT_NFT_OP_TABLE_UNREGISTER,
 84         [NFT_MSG_NEWCHAIN]      = AUDIT_NFT_OP_CHAIN_REGISTER,
 85         [NFT_MSG_GETCHAIN]      = AUDIT_NFT_OP_INVALID,
 86         [NFT_MSG_DELCHAIN]      = AUDIT_NFT_OP_CHAIN_UNREGISTER,
 87         [NFT_MSG_NEWRULE]       = AUDIT_NFT_OP_RULE_REGISTER,
 88         [NFT_MSG_GETRULE]       = AUDIT_NFT_OP_INVALID,
 89         [NFT_MSG_DELRULE]       = AUDIT_NFT_OP_RULE_UNREGISTER,
 90         [NFT_MSG_NEWSET]        = AUDIT_NFT_OP_SET_REGISTER,
 91         [NFT_MSG_GETSET]        = AUDIT_NFT_OP_INVALID,
 92         [NFT_MSG_DELSET]        = AUDIT_NFT_OP_SET_UNREGISTER,
 93         [NFT_MSG_NEWSETELEM]    = AUDIT_NFT_OP_SETELEM_REGISTER,
 94         [NFT_MSG_GETSETELEM]    = AUDIT_NFT_OP_INVALID,
 95         [NFT_MSG_DELSETELEM]    = AUDIT_NFT_OP_SETELEM_UNREGISTER,
 96         [NFT_MSG_NEWGEN]        = AUDIT_NFT_OP_GEN_REGISTER,
 97         [NFT_MSG_GETGEN]        = AUDIT_NFT_OP_INVALID,
 98         [NFT_MSG_TRACE]         = AUDIT_NFT_OP_INVALID,
 99         [NFT_MSG_NEWOBJ]        = AUDIT_NFT_OP_OBJ_REGISTER,
100         [NFT_MSG_GETOBJ]        = AUDIT_NFT_OP_INVALID,
101         [NFT_MSG_DELOBJ]        = AUDIT_NFT_OP_OBJ_UNREGISTER,
102         [NFT_MSG_GETOBJ_RESET]  = AUDIT_NFT_OP_OBJ_RESET,
103         [NFT_MSG_NEWFLOWTABLE]  = AUDIT_NFT_OP_FLOWTABLE_REGISTER,
104         [NFT_MSG_GETFLOWTABLE]  = AUDIT_NFT_OP_INVALID,
105         [NFT_MSG_DELFLOWTABLE]  = AUDIT_NFT_OP_FLOWTABLE_UNREGISTER,
106         [NFT_MSG_GETSETELEM_RESET] = AUDIT_NFT_OP_SETELEM_RESET,
107 };
108 
109 static void nft_validate_state_update(struct nft_table *table, u8 new_validate_state)
110 {
111         switch (table->validate_state) {
112         case NFT_VALIDATE_SKIP:
113                 WARN_ON_ONCE(new_validate_state == NFT_VALIDATE_DO);
114                 break;
115         case NFT_VALIDATE_NEED:
116                 break;
117         case NFT_VALIDATE_DO:
118                 if (new_validate_state == NFT_VALIDATE_NEED)
119                         return;
120         }
121 
122         table->validate_state = new_validate_state;
123 }
124 static void nf_tables_trans_destroy_work(struct work_struct *w);
125 static DECLARE_WORK(trans_destroy_work, nf_tables_trans_destroy_work);
126 
127 static void nft_trans_gc_work(struct work_struct *work);
128 static DECLARE_WORK(trans_gc_work, nft_trans_gc_work);
129 
130 static void nft_ctx_init(struct nft_ctx *ctx,
131                          struct net *net,
132                          const struct sk_buff *skb,
133                          const struct nlmsghdr *nlh,
134                          u8 family,
135                          struct nft_table *table,
136                          struct nft_chain *chain,
137                          const struct nlattr * const *nla)
138 {
139         ctx->net        = net;
140         ctx->family     = family;
141         ctx->level      = 0;
142         ctx->table      = table;
143         ctx->chain      = chain;
144         ctx->nla        = nla;
145         ctx->portid     = NETLINK_CB(skb).portid;
146         ctx->report     = nlmsg_report(nlh);
147         ctx->flags      = nlh->nlmsg_flags;
148         ctx->seq        = nlh->nlmsg_seq;
149 }
150 
151 static struct nft_trans *nft_trans_alloc_gfp(const struct nft_ctx *ctx,
152                                              int msg_type, u32 size, gfp_t gfp)
153 {
154         struct nft_trans *trans;
155 
156         trans = kzalloc(size, gfp);
157         if (trans == NULL)
158                 return NULL;
159 
160         INIT_LIST_HEAD(&trans->list);
161         trans->msg_type = msg_type;
162 
163         trans->net = ctx->net;
164         trans->table = ctx->table;
165         trans->seq = ctx->seq;
166         trans->flags = ctx->flags;
167         trans->report = ctx->report;
168 
169         return trans;
170 }
171 
172 static struct nft_trans *nft_trans_alloc(const struct nft_ctx *ctx,
173                                          int msg_type, u32 size)
174 {
175         return nft_trans_alloc_gfp(ctx, msg_type, size, GFP_KERNEL);
176 }
177 
178 static struct nft_trans_binding *nft_trans_get_binding(struct nft_trans *trans)
179 {
180         switch (trans->msg_type) {
181         case NFT_MSG_NEWCHAIN:
182         case NFT_MSG_NEWSET:
183                 return container_of(trans, struct nft_trans_binding, nft_trans);
184         }
185 
186         return NULL;
187 }
188 
189 static void nft_trans_list_del(struct nft_trans *trans)
190 {
191         struct nft_trans_binding *trans_binding;
192 
193         list_del(&trans->list);
194 
195         trans_binding = nft_trans_get_binding(trans);
196         if (trans_binding)
197                 list_del(&trans_binding->binding_list);
198 }
199 
200 static void nft_trans_destroy(struct nft_trans *trans)
201 {
202         nft_trans_list_del(trans);
203         kfree(trans);
204 }
205 
206 static void __nft_set_trans_bind(const struct nft_ctx *ctx, struct nft_set *set,
207                                  bool bind)
208 {
209         struct nftables_pernet *nft_net;
210         struct net *net = ctx->net;
211         struct nft_trans *trans;
212 
213         if (!nft_set_is_anonymous(set))
214                 return;
215 
216         nft_net = nft_pernet(net);
217         list_for_each_entry_reverse(trans, &nft_net->commit_list, list) {
218                 switch (trans->msg_type) {
219                 case NFT_MSG_NEWSET:
220                         if (nft_trans_set(trans) == set)
221                                 nft_trans_set_bound(trans) = bind;
222                         break;
223                 case NFT_MSG_NEWSETELEM:
224                         if (nft_trans_elem_set(trans) == set)
225                                 nft_trans_elem_set_bound(trans) = bind;
226                         break;
227                 }
228         }
229 }
230 
231 static void nft_set_trans_bind(const struct nft_ctx *ctx, struct nft_set *set)
232 {
233         return __nft_set_trans_bind(ctx, set, true);
234 }
235 
236 static void nft_set_trans_unbind(const struct nft_ctx *ctx, struct nft_set *set)
237 {
238         return __nft_set_trans_bind(ctx, set, false);
239 }
240 
241 static void __nft_chain_trans_bind(const struct nft_ctx *ctx,
242                                    struct nft_chain *chain, bool bind)
243 {
244         struct nftables_pernet *nft_net;
245         struct net *net = ctx->net;
246         struct nft_trans *trans;
247 
248         if (!nft_chain_binding(chain))
249                 return;
250 
251         nft_net = nft_pernet(net);
252         list_for_each_entry_reverse(trans, &nft_net->commit_list, list) {
253                 switch (trans->msg_type) {
254                 case NFT_MSG_NEWCHAIN:
255                         if (nft_trans_chain(trans) == chain)
256                                 nft_trans_chain_bound(trans) = bind;
257                         break;
258                 case NFT_MSG_NEWRULE:
259                         if (nft_trans_rule_chain(trans) == chain)
260                                 nft_trans_rule_bound(trans) = bind;
261                         break;
262                 }
263         }
264 }
265 
266 static void nft_chain_trans_bind(const struct nft_ctx *ctx,
267                                  struct nft_chain *chain)
268 {
269         __nft_chain_trans_bind(ctx, chain, true);
270 }
271 
272 int nf_tables_bind_chain(const struct nft_ctx *ctx, struct nft_chain *chain)
273 {
274         if (!nft_chain_binding(chain))
275                 return 0;
276 
277         if (nft_chain_binding(ctx->chain))
278                 return -EOPNOTSUPP;
279 
280         if (chain->bound)
281                 return -EBUSY;
282 
283         if (!nft_use_inc(&chain->use))
284                 return -EMFILE;
285 
286         chain->bound = true;
287         nft_chain_trans_bind(ctx, chain);
288 
289         return 0;
290 }
291 
292 void nf_tables_unbind_chain(const struct nft_ctx *ctx, struct nft_chain *chain)
293 {
294         __nft_chain_trans_bind(ctx, chain, false);
295 }
296 
297 static int nft_netdev_register_hooks(struct net *net,
298                                      struct list_head *hook_list)
299 {
300         struct nft_hook *hook;
301         int err, j;
302 
303         j = 0;
304         list_for_each_entry(hook, hook_list, list) {
305                 err = nf_register_net_hook(net, &hook->ops);
306                 if (err < 0)
307                         goto err_register;
308 
309                 j++;
310         }
311         return 0;
312 
313 err_register:
314         list_for_each_entry(hook, hook_list, list) {
315                 if (j-- <= 0)
316                         break;
317 
318                 nf_unregister_net_hook(net, &hook->ops);
319         }
320         return err;
321 }
322 
323 static void nft_netdev_unregister_hooks(struct net *net,
324                                         struct list_head *hook_list,
325                                         bool release_netdev)
326 {
327         struct nft_hook *hook, *next;
328 
329         list_for_each_entry_safe(hook, next, hook_list, list) {
330                 nf_unregister_net_hook(net, &hook->ops);
331                 if (release_netdev) {
332                         list_del(&hook->list);
333                         kfree_rcu(hook, rcu);
334                 }
335         }
336 }
337 
338 static int nf_tables_register_hook(struct net *net,
339                                    const struct nft_table *table,
340                                    struct nft_chain *chain)
341 {
342         struct nft_base_chain *basechain;
343         const struct nf_hook_ops *ops;
344 
345         if (table->flags & NFT_TABLE_F_DORMANT ||
346             !nft_is_base_chain(chain))
347                 return 0;
348 
349         basechain = nft_base_chain(chain);
350         ops = &basechain->ops;
351 
352         if (basechain->type->ops_register)
353                 return basechain->type->ops_register(net, ops);
354 
355         if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
356                 return nft_netdev_register_hooks(net, &basechain->hook_list);
357 
358         return nf_register_net_hook(net, &basechain->ops);
359 }
360 
361 static void __nf_tables_unregister_hook(struct net *net,
362                                         const struct nft_table *table,
363                                         struct nft_chain *chain,
364                                         bool release_netdev)
365 {
366         struct nft_base_chain *basechain;
367         const struct nf_hook_ops *ops;
368 
369         if (table->flags & NFT_TABLE_F_DORMANT ||
370             !nft_is_base_chain(chain))
371                 return;
372         basechain = nft_base_chain(chain);
373         ops = &basechain->ops;
374 
375         if (basechain->type->ops_unregister)
376                 return basechain->type->ops_unregister(net, ops);
377 
378         if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
379                 nft_netdev_unregister_hooks(net, &basechain->hook_list,
380                                             release_netdev);
381         else
382                 nf_unregister_net_hook(net, &basechain->ops);
383 }
384 
385 static void nf_tables_unregister_hook(struct net *net,
386                                       const struct nft_table *table,
387                                       struct nft_chain *chain)
388 {
389         return __nf_tables_unregister_hook(net, table, chain, false);
390 }
391 
392 static void nft_trans_commit_list_add_tail(struct net *net, struct nft_trans *trans)
393 {
394         struct nftables_pernet *nft_net = nft_pernet(net);
395         struct nft_trans_binding *binding;
396 
397         list_add_tail(&trans->list, &nft_net->commit_list);
398 
399         binding = nft_trans_get_binding(trans);
400         if (!binding)
401                 return;
402 
403         switch (trans->msg_type) {
404         case NFT_MSG_NEWSET:
405                 if (!nft_trans_set_update(trans) &&
406                     nft_set_is_anonymous(nft_trans_set(trans)))
407                         list_add_tail(&binding->binding_list, &nft_net->binding_list);
408                 break;
409         case NFT_MSG_NEWCHAIN:
410                 if (!nft_trans_chain_update(trans) &&
411                     nft_chain_binding(nft_trans_chain(trans)))
412                         list_add_tail(&binding->binding_list, &nft_net->binding_list);
413                 break;
414         }
415 }
416 
417 static int nft_trans_table_add(struct nft_ctx *ctx, int msg_type)
418 {
419         struct nft_trans *trans;
420 
421         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_table));
422         if (trans == NULL)
423                 return -ENOMEM;
424 
425         if (msg_type == NFT_MSG_NEWTABLE)
426                 nft_activate_next(ctx->net, ctx->table);
427 
428         nft_trans_commit_list_add_tail(ctx->net, trans);
429         return 0;
430 }
431 
432 static int nft_deltable(struct nft_ctx *ctx)
433 {
434         int err;
435 
436         err = nft_trans_table_add(ctx, NFT_MSG_DELTABLE);
437         if (err < 0)
438                 return err;
439 
440         nft_deactivate_next(ctx->net, ctx->table);
441         return err;
442 }
443 
444 static struct nft_trans *
445 nft_trans_alloc_chain(const struct nft_ctx *ctx, int msg_type)
446 {
447         struct nft_trans_chain *trans_chain;
448         struct nft_trans *trans;
449 
450         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_chain));
451         if (!trans)
452                 return NULL;
453 
454         trans_chain = nft_trans_container_chain(trans);
455         INIT_LIST_HEAD(&trans_chain->nft_trans_binding.binding_list);
456         trans_chain->chain = ctx->chain;
457 
458         return trans;
459 }
460 
461 static struct nft_trans *nft_trans_chain_add(struct nft_ctx *ctx, int msg_type)
462 {
463         struct nft_trans *trans;
464 
465         trans = nft_trans_alloc_chain(ctx, msg_type);
466         if (trans == NULL)
467                 return ERR_PTR(-ENOMEM);
468 
469         if (msg_type == NFT_MSG_NEWCHAIN) {
470                 nft_activate_next(ctx->net, ctx->chain);
471 
472                 if (ctx->nla[NFTA_CHAIN_ID]) {
473                         nft_trans_chain_id(trans) =
474                                 ntohl(nla_get_be32(ctx->nla[NFTA_CHAIN_ID]));
475                 }
476         }
477         nft_trans_commit_list_add_tail(ctx->net, trans);
478 
479         return trans;
480 }
481 
482 static int nft_delchain(struct nft_ctx *ctx)
483 {
484         struct nft_trans *trans;
485 
486         trans = nft_trans_chain_add(ctx, NFT_MSG_DELCHAIN);
487         if (IS_ERR(trans))
488                 return PTR_ERR(trans);
489 
490         nft_use_dec(&ctx->table->use);
491         nft_deactivate_next(ctx->net, ctx->chain);
492 
493         return 0;
494 }
495 
496 void nft_rule_expr_activate(const struct nft_ctx *ctx, struct nft_rule *rule)
497 {
498         struct nft_expr *expr;
499 
500         expr = nft_expr_first(rule);
501         while (nft_expr_more(rule, expr)) {
502                 if (expr->ops->activate)
503                         expr->ops->activate(ctx, expr);
504 
505                 expr = nft_expr_next(expr);
506         }
507 }
508 
509 void nft_rule_expr_deactivate(const struct nft_ctx *ctx, struct nft_rule *rule,
510                               enum nft_trans_phase phase)
511 {
512         struct nft_expr *expr;
513 
514         expr = nft_expr_first(rule);
515         while (nft_expr_more(rule, expr)) {
516                 if (expr->ops->deactivate)
517                         expr->ops->deactivate(ctx, expr, phase);
518 
519                 expr = nft_expr_next(expr);
520         }
521 }
522 
523 static int
524 nf_tables_delrule_deactivate(struct nft_ctx *ctx, struct nft_rule *rule)
525 {
526         /* You cannot delete the same rule twice */
527         if (nft_is_active_next(ctx->net, rule)) {
528                 nft_deactivate_next(ctx->net, rule);
529                 nft_use_dec(&ctx->chain->use);
530                 return 0;
531         }
532         return -ENOENT;
533 }
534 
535 static struct nft_trans *nft_trans_rule_add(struct nft_ctx *ctx, int msg_type,
536                                             struct nft_rule *rule)
537 {
538         struct nft_trans *trans;
539 
540         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_rule));
541         if (trans == NULL)
542                 return NULL;
543 
544         if (msg_type == NFT_MSG_NEWRULE && ctx->nla[NFTA_RULE_ID] != NULL) {
545                 nft_trans_rule_id(trans) =
546                         ntohl(nla_get_be32(ctx->nla[NFTA_RULE_ID]));
547         }
548         nft_trans_rule(trans) = rule;
549         nft_trans_rule_chain(trans) = ctx->chain;
550         nft_trans_commit_list_add_tail(ctx->net, trans);
551 
552         return trans;
553 }
554 
555 static int nft_delrule(struct nft_ctx *ctx, struct nft_rule *rule)
556 {
557         struct nft_flow_rule *flow;
558         struct nft_trans *trans;
559         int err;
560 
561         trans = nft_trans_rule_add(ctx, NFT_MSG_DELRULE, rule);
562         if (trans == NULL)
563                 return -ENOMEM;
564 
565         if (ctx->chain->flags & NFT_CHAIN_HW_OFFLOAD) {
566                 flow = nft_flow_rule_create(ctx->net, rule);
567                 if (IS_ERR(flow)) {
568                         nft_trans_destroy(trans);
569                         return PTR_ERR(flow);
570                 }
571 
572                 nft_trans_flow_rule(trans) = flow;
573         }
574 
575         err = nf_tables_delrule_deactivate(ctx, rule);
576         if (err < 0) {
577                 nft_trans_destroy(trans);
578                 return err;
579         }
580         nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_PREPARE);
581 
582         return 0;
583 }
584 
585 static int nft_delrule_by_chain(struct nft_ctx *ctx)
586 {
587         struct nft_rule *rule;
588         int err;
589 
590         list_for_each_entry(rule, &ctx->chain->rules, list) {
591                 if (!nft_is_active_next(ctx->net, rule))
592                         continue;
593 
594                 err = nft_delrule(ctx, rule);
595                 if (err < 0)
596                         return err;
597         }
598         return 0;
599 }
600 
601 static int __nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
602                                struct nft_set *set,
603                                const struct nft_set_desc *desc)
604 {
605         struct nft_trans_set *trans_set;
606         struct nft_trans *trans;
607 
608         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_set));
609         if (trans == NULL)
610                 return -ENOMEM;
611 
612         trans_set = nft_trans_container_set(trans);
613         INIT_LIST_HEAD(&trans_set->nft_trans_binding.binding_list);
614 
615         if (msg_type == NFT_MSG_NEWSET && ctx->nla[NFTA_SET_ID] && !desc) {
616                 nft_trans_set_id(trans) =
617                         ntohl(nla_get_be32(ctx->nla[NFTA_SET_ID]));
618                 nft_activate_next(ctx->net, set);
619         }
620         nft_trans_set(trans) = set;
621         if (desc) {
622                 nft_trans_set_update(trans) = true;
623                 nft_trans_set_gc_int(trans) = desc->gc_int;
624                 nft_trans_set_timeout(trans) = desc->timeout;
625                 nft_trans_set_size(trans) = desc->size;
626         }
627         nft_trans_commit_list_add_tail(ctx->net, trans);
628 
629         return 0;
630 }
631 
632 static int nft_trans_set_add(const struct nft_ctx *ctx, int msg_type,
633                              struct nft_set *set)
634 {
635         return __nft_trans_set_add(ctx, msg_type, set, NULL);
636 }
637 
638 static int nft_mapelem_deactivate(const struct nft_ctx *ctx,
639                                   struct nft_set *set,
640                                   const struct nft_set_iter *iter,
641                                   struct nft_elem_priv *elem_priv)
642 {
643         struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
644 
645         if (!nft_set_elem_active(ext, iter->genmask))
646                 return 0;
647 
648         nft_set_elem_change_active(ctx->net, set, ext);
649         nft_setelem_data_deactivate(ctx->net, set, elem_priv);
650 
651         return 0;
652 }
653 
654 struct nft_set_elem_catchall {
655         struct list_head        list;
656         struct rcu_head         rcu;
657         struct nft_elem_priv    *elem;
658 };
659 
660 static void nft_map_catchall_deactivate(const struct nft_ctx *ctx,
661                                         struct nft_set *set)
662 {
663         u8 genmask = nft_genmask_next(ctx->net);
664         struct nft_set_elem_catchall *catchall;
665         struct nft_set_ext *ext;
666 
667         list_for_each_entry(catchall, &set->catchall_list, list) {
668                 ext = nft_set_elem_ext(set, catchall->elem);
669                 if (!nft_set_elem_active(ext, genmask))
670                         continue;
671 
672                 nft_set_elem_change_active(ctx->net, set, ext);
673                 nft_setelem_data_deactivate(ctx->net, set, catchall->elem);
674                 break;
675         }
676 }
677 
678 static void nft_map_deactivate(const struct nft_ctx *ctx, struct nft_set *set)
679 {
680         struct nft_set_iter iter = {
681                 .genmask        = nft_genmask_next(ctx->net),
682                 .type           = NFT_ITER_UPDATE,
683                 .fn             = nft_mapelem_deactivate,
684         };
685 
686         set->ops->walk(ctx, set, &iter);
687         WARN_ON_ONCE(iter.err);
688 
689         nft_map_catchall_deactivate(ctx, set);
690 }
691 
692 static int nft_delset(const struct nft_ctx *ctx, struct nft_set *set)
693 {
694         int err;
695 
696         err = nft_trans_set_add(ctx, NFT_MSG_DELSET, set);
697         if (err < 0)
698                 return err;
699 
700         if (set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
701                 nft_map_deactivate(ctx, set);
702 
703         nft_deactivate_next(ctx->net, set);
704         nft_use_dec(&ctx->table->use);
705 
706         return err;
707 }
708 
709 static int nft_trans_obj_add(struct nft_ctx *ctx, int msg_type,
710                              struct nft_object *obj)
711 {
712         struct nft_trans *trans;
713 
714         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_obj));
715         if (trans == NULL)
716                 return -ENOMEM;
717 
718         if (msg_type == NFT_MSG_NEWOBJ)
719                 nft_activate_next(ctx->net, obj);
720 
721         nft_trans_obj(trans) = obj;
722         nft_trans_commit_list_add_tail(ctx->net, trans);
723 
724         return 0;
725 }
726 
727 static int nft_delobj(struct nft_ctx *ctx, struct nft_object *obj)
728 {
729         int err;
730 
731         err = nft_trans_obj_add(ctx, NFT_MSG_DELOBJ, obj);
732         if (err < 0)
733                 return err;
734 
735         nft_deactivate_next(ctx->net, obj);
736         nft_use_dec(&ctx->table->use);
737 
738         return err;
739 }
740 
741 static struct nft_trans *
742 nft_trans_flowtable_add(struct nft_ctx *ctx, int msg_type,
743                         struct nft_flowtable *flowtable)
744 {
745         struct nft_trans *trans;
746 
747         trans = nft_trans_alloc(ctx, msg_type,
748                                 sizeof(struct nft_trans_flowtable));
749         if (trans == NULL)
750                 return ERR_PTR(-ENOMEM);
751 
752         if (msg_type == NFT_MSG_NEWFLOWTABLE)
753                 nft_activate_next(ctx->net, flowtable);
754 
755         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
756         nft_trans_flowtable(trans) = flowtable;
757         nft_trans_commit_list_add_tail(ctx->net, trans);
758 
759         return trans;
760 }
761 
762 static int nft_delflowtable(struct nft_ctx *ctx,
763                             struct nft_flowtable *flowtable)
764 {
765         struct nft_trans *trans;
766 
767         trans = nft_trans_flowtable_add(ctx, NFT_MSG_DELFLOWTABLE, flowtable);
768         if (IS_ERR(trans))
769                 return PTR_ERR(trans);
770 
771         nft_deactivate_next(ctx->net, flowtable);
772         nft_use_dec(&ctx->table->use);
773 
774         return 0;
775 }
776 
777 static void __nft_reg_track_clobber(struct nft_regs_track *track, u8 dreg)
778 {
779         int i;
780 
781         for (i = track->regs[dreg].num_reg; i > 0; i--)
782                 __nft_reg_track_cancel(track, dreg - i);
783 }
784 
785 static void __nft_reg_track_update(struct nft_regs_track *track,
786                                    const struct nft_expr *expr,
787                                    u8 dreg, u8 num_reg)
788 {
789         track->regs[dreg].selector = expr;
790         track->regs[dreg].bitwise = NULL;
791         track->regs[dreg].num_reg = num_reg;
792 }
793 
794 void nft_reg_track_update(struct nft_regs_track *track,
795                           const struct nft_expr *expr, u8 dreg, u8 len)
796 {
797         unsigned int regcount;
798         int i;
799 
800         __nft_reg_track_clobber(track, dreg);
801 
802         regcount = DIV_ROUND_UP(len, NFT_REG32_SIZE);
803         for (i = 0; i < regcount; i++, dreg++)
804                 __nft_reg_track_update(track, expr, dreg, i);
805 }
806 EXPORT_SYMBOL_GPL(nft_reg_track_update);
807 
808 void nft_reg_track_cancel(struct nft_regs_track *track, u8 dreg, u8 len)
809 {
810         unsigned int regcount;
811         int i;
812 
813         __nft_reg_track_clobber(track, dreg);
814 
815         regcount = DIV_ROUND_UP(len, NFT_REG32_SIZE);
816         for (i = 0; i < regcount; i++, dreg++)
817                 __nft_reg_track_cancel(track, dreg);
818 }
819 EXPORT_SYMBOL_GPL(nft_reg_track_cancel);
820 
821 void __nft_reg_track_cancel(struct nft_regs_track *track, u8 dreg)
822 {
823         track->regs[dreg].selector = NULL;
824         track->regs[dreg].bitwise = NULL;
825         track->regs[dreg].num_reg = 0;
826 }
827 EXPORT_SYMBOL_GPL(__nft_reg_track_cancel);
828 
829 /*
830  * Tables
831  */
832 
833 static struct nft_table *nft_table_lookup(const struct net *net,
834                                           const struct nlattr *nla,
835                                           u8 family, u8 genmask, u32 nlpid)
836 {
837         struct nftables_pernet *nft_net;
838         struct nft_table *table;
839 
840         if (nla == NULL)
841                 return ERR_PTR(-EINVAL);
842 
843         nft_net = nft_pernet(net);
844         list_for_each_entry_rcu(table, &nft_net->tables, list,
845                                 lockdep_is_held(&nft_net->commit_mutex)) {
846                 if (!nla_strcmp(nla, table->name) &&
847                     table->family == family &&
848                     nft_active_genmask(table, genmask)) {
849                         if (nft_table_has_owner(table) &&
850                             nlpid && table->nlpid != nlpid)
851                                 return ERR_PTR(-EPERM);
852 
853                         return table;
854                 }
855         }
856 
857         return ERR_PTR(-ENOENT);
858 }
859 
860 static struct nft_table *nft_table_lookup_byhandle(const struct net *net,
861                                                    const struct nlattr *nla,
862                                                    int family, u8 genmask, u32 nlpid)
863 {
864         struct nftables_pernet *nft_net;
865         struct nft_table *table;
866 
867         nft_net = nft_pernet(net);
868         list_for_each_entry(table, &nft_net->tables, list) {
869                 if (be64_to_cpu(nla_get_be64(nla)) == table->handle &&
870                     table->family == family &&
871                     nft_active_genmask(table, genmask)) {
872                         if (nft_table_has_owner(table) &&
873                             nlpid && table->nlpid != nlpid)
874                                 return ERR_PTR(-EPERM);
875 
876                         return table;
877                 }
878         }
879 
880         return ERR_PTR(-ENOENT);
881 }
882 
883 static inline u64 nf_tables_alloc_handle(struct nft_table *table)
884 {
885         return ++table->hgenerator;
886 }
887 
888 static const struct nft_chain_type *chain_type[NFPROTO_NUMPROTO][NFT_CHAIN_T_MAX];
889 
890 static const struct nft_chain_type *
891 __nft_chain_type_get(u8 family, enum nft_chain_types type)
892 {
893         if (family >= NFPROTO_NUMPROTO ||
894             type >= NFT_CHAIN_T_MAX)
895                 return NULL;
896 
897         return chain_type[family][type];
898 }
899 
900 static const struct nft_chain_type *
901 __nf_tables_chain_type_lookup(const struct nlattr *nla, u8 family)
902 {
903         const struct nft_chain_type *type;
904         int i;
905 
906         for (i = 0; i < NFT_CHAIN_T_MAX; i++) {
907                 type = __nft_chain_type_get(family, i);
908                 if (!type)
909                         continue;
910                 if (!nla_strcmp(nla, type->name))
911                         return type;
912         }
913         return NULL;
914 }
915 
916 struct nft_module_request {
917         struct list_head        list;
918         char                    module[MODULE_NAME_LEN];
919         bool                    done;
920 };
921 
922 #ifdef CONFIG_MODULES
923 __printf(2, 3) int nft_request_module(struct net *net, const char *fmt,
924                                       ...)
925 {
926         char module_name[MODULE_NAME_LEN];
927         struct nftables_pernet *nft_net;
928         struct nft_module_request *req;
929         va_list args;
930         int ret;
931 
932         va_start(args, fmt);
933         ret = vsnprintf(module_name, MODULE_NAME_LEN, fmt, args);
934         va_end(args);
935         if (ret >= MODULE_NAME_LEN)
936                 return 0;
937 
938         nft_net = nft_pernet(net);
939         list_for_each_entry(req, &nft_net->module_list, list) {
940                 if (!strcmp(req->module, module_name)) {
941                         if (req->done)
942                                 return 0;
943 
944                         /* A request to load this module already exists. */
945                         return -EAGAIN;
946                 }
947         }
948 
949         req = kmalloc(sizeof(*req), GFP_KERNEL);
950         if (!req)
951                 return -ENOMEM;
952 
953         req->done = false;
954         strscpy(req->module, module_name, MODULE_NAME_LEN);
955         list_add_tail(&req->list, &nft_net->module_list);
956 
957         return -EAGAIN;
958 }
959 EXPORT_SYMBOL_GPL(nft_request_module);
960 #endif
961 
962 static void lockdep_nfnl_nft_mutex_not_held(void)
963 {
964 #ifdef CONFIG_PROVE_LOCKING
965         if (debug_locks)
966                 WARN_ON_ONCE(lockdep_nfnl_is_held(NFNL_SUBSYS_NFTABLES));
967 #endif
968 }
969 
970 static const struct nft_chain_type *
971 nf_tables_chain_type_lookup(struct net *net, const struct nlattr *nla,
972                             u8 family, bool autoload)
973 {
974         const struct nft_chain_type *type;
975 
976         type = __nf_tables_chain_type_lookup(nla, family);
977         if (type != NULL)
978                 return type;
979 
980         lockdep_nfnl_nft_mutex_not_held();
981 #ifdef CONFIG_MODULES
982         if (autoload) {
983                 if (nft_request_module(net, "nft-chain-%u-%.*s", family,
984                                        nla_len(nla),
985                                        (const char *)nla_data(nla)) == -EAGAIN)
986                         return ERR_PTR(-EAGAIN);
987         }
988 #endif
989         return ERR_PTR(-ENOENT);
990 }
991 
992 static __be16 nft_base_seq(const struct net *net)
993 {
994         struct nftables_pernet *nft_net = nft_pernet(net);
995 
996         return htons(nft_net->base_seq & 0xffff);
997 }
998 
999 static const struct nla_policy nft_table_policy[NFTA_TABLE_MAX + 1] = {
1000         [NFTA_TABLE_NAME]       = { .type = NLA_STRING,
1001                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
1002         [NFTA_TABLE_FLAGS]      = { .type = NLA_U32 },
1003         [NFTA_TABLE_HANDLE]     = { .type = NLA_U64 },
1004         [NFTA_TABLE_USERDATA]   = { .type = NLA_BINARY,
1005                                     .len = NFT_USERDATA_MAXLEN }
1006 };
1007 
1008 static int nf_tables_fill_table_info(struct sk_buff *skb, struct net *net,
1009                                      u32 portid, u32 seq, int event, u32 flags,
1010                                      int family, const struct nft_table *table)
1011 {
1012         struct nlmsghdr *nlh;
1013 
1014         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
1015         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
1016                            NFNETLINK_V0, nft_base_seq(net));
1017         if (!nlh)
1018                 goto nla_put_failure;
1019 
1020         if (nla_put_string(skb, NFTA_TABLE_NAME, table->name) ||
1021             nla_put_be32(skb, NFTA_TABLE_USE, htonl(table->use)) ||
1022             nla_put_be64(skb, NFTA_TABLE_HANDLE, cpu_to_be64(table->handle),
1023                          NFTA_TABLE_PAD))
1024                 goto nla_put_failure;
1025 
1026         if (event == NFT_MSG_DELTABLE) {
1027                 nlmsg_end(skb, nlh);
1028                 return 0;
1029         }
1030 
1031         if (nla_put_be32(skb, NFTA_TABLE_FLAGS,
1032                          htonl(table->flags & NFT_TABLE_F_MASK)))
1033                 goto nla_put_failure;
1034 
1035         if (nft_table_has_owner(table) &&
1036             nla_put_be32(skb, NFTA_TABLE_OWNER, htonl(table->nlpid)))
1037                 goto nla_put_failure;
1038 
1039         if (table->udata) {
1040                 if (nla_put(skb, NFTA_TABLE_USERDATA, table->udlen, table->udata))
1041                         goto nla_put_failure;
1042         }
1043 
1044         nlmsg_end(skb, nlh);
1045         return 0;
1046 
1047 nla_put_failure:
1048         nlmsg_trim(skb, nlh);
1049         return -1;
1050 }
1051 
1052 struct nftnl_skb_parms {
1053         bool report;
1054 };
1055 #define NFT_CB(skb)     (*(struct nftnl_skb_parms*)&((skb)->cb))
1056 
1057 static void nft_notify_enqueue(struct sk_buff *skb, bool report,
1058                                struct list_head *notify_list)
1059 {
1060         NFT_CB(skb).report = report;
1061         list_add_tail(&skb->list, notify_list);
1062 }
1063 
1064 static void nf_tables_table_notify(const struct nft_ctx *ctx, int event)
1065 {
1066         struct nftables_pernet *nft_net;
1067         struct sk_buff *skb;
1068         u16 flags = 0;
1069         int err;
1070 
1071         if (!ctx->report &&
1072             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1073                 return;
1074 
1075         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1076         if (skb == NULL)
1077                 goto err;
1078 
1079         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
1080                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
1081 
1082         err = nf_tables_fill_table_info(skb, ctx->net, ctx->portid, ctx->seq,
1083                                         event, flags, ctx->family, ctx->table);
1084         if (err < 0) {
1085                 kfree_skb(skb);
1086                 goto err;
1087         }
1088 
1089         nft_net = nft_pernet(ctx->net);
1090         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
1091         return;
1092 err:
1093         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1094 }
1095 
1096 static int nf_tables_dump_tables(struct sk_buff *skb,
1097                                  struct netlink_callback *cb)
1098 {
1099         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1100         struct nftables_pernet *nft_net;
1101         const struct nft_table *table;
1102         unsigned int idx = 0, s_idx = cb->args[0];
1103         struct net *net = sock_net(skb->sk);
1104         int family = nfmsg->nfgen_family;
1105 
1106         rcu_read_lock();
1107         nft_net = nft_pernet(net);
1108         cb->seq = READ_ONCE(nft_net->base_seq);
1109 
1110         list_for_each_entry_rcu(table, &nft_net->tables, list) {
1111                 if (family != NFPROTO_UNSPEC && family != table->family)
1112                         continue;
1113 
1114                 if (idx < s_idx)
1115                         goto cont;
1116                 if (idx > s_idx)
1117                         memset(&cb->args[1], 0,
1118                                sizeof(cb->args) - sizeof(cb->args[0]));
1119                 if (!nft_is_active(net, table))
1120                         continue;
1121                 if (nf_tables_fill_table_info(skb, net,
1122                                               NETLINK_CB(cb->skb).portid,
1123                                               cb->nlh->nlmsg_seq,
1124                                               NFT_MSG_NEWTABLE, NLM_F_MULTI,
1125                                               table->family, table) < 0)
1126                         goto done;
1127 
1128                 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
1129 cont:
1130                 idx++;
1131         }
1132 done:
1133         rcu_read_unlock();
1134         cb->args[0] = idx;
1135         return skb->len;
1136 }
1137 
1138 static int nft_netlink_dump_start_rcu(struct sock *nlsk, struct sk_buff *skb,
1139                                       const struct nlmsghdr *nlh,
1140                                       struct netlink_dump_control *c)
1141 {
1142         int err;
1143 
1144         if (!try_module_get(THIS_MODULE))
1145                 return -EINVAL;
1146 
1147         rcu_read_unlock();
1148         err = netlink_dump_start(nlsk, skb, nlh, c);
1149         rcu_read_lock();
1150         module_put(THIS_MODULE);
1151 
1152         return err;
1153 }
1154 
1155 /* called with rcu_read_lock held */
1156 static int nf_tables_gettable(struct sk_buff *skb, const struct nfnl_info *info,
1157                               const struct nlattr * const nla[])
1158 {
1159         struct netlink_ext_ack *extack = info->extack;
1160         u8 genmask = nft_genmask_cur(info->net);
1161         u8 family = info->nfmsg->nfgen_family;
1162         const struct nft_table *table;
1163         struct net *net = info->net;
1164         struct sk_buff *skb2;
1165         int err;
1166 
1167         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
1168                 struct netlink_dump_control c = {
1169                         .dump = nf_tables_dump_tables,
1170                         .module = THIS_MODULE,
1171                 };
1172 
1173                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
1174         }
1175 
1176         table = nft_table_lookup(net, nla[NFTA_TABLE_NAME], family, genmask, 0);
1177         if (IS_ERR(table)) {
1178                 NL_SET_BAD_ATTR(extack, nla[NFTA_TABLE_NAME]);
1179                 return PTR_ERR(table);
1180         }
1181 
1182         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
1183         if (!skb2)
1184                 return -ENOMEM;
1185 
1186         err = nf_tables_fill_table_info(skb2, net, NETLINK_CB(skb).portid,
1187                                         info->nlh->nlmsg_seq, NFT_MSG_NEWTABLE,
1188                                         0, family, table);
1189         if (err < 0)
1190                 goto err_fill_table_info;
1191 
1192         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
1193 
1194 err_fill_table_info:
1195         kfree_skb(skb2);
1196         return err;
1197 }
1198 
1199 static void nft_table_disable(struct net *net, struct nft_table *table, u32 cnt)
1200 {
1201         struct nft_chain *chain;
1202         u32 i = 0;
1203 
1204         list_for_each_entry(chain, &table->chains, list) {
1205                 if (!nft_is_active_next(net, chain))
1206                         continue;
1207                 if (!nft_is_base_chain(chain))
1208                         continue;
1209 
1210                 if (cnt && i++ == cnt)
1211                         break;
1212 
1213                 nf_tables_unregister_hook(net, table, chain);
1214         }
1215 }
1216 
1217 static int nf_tables_table_enable(struct net *net, struct nft_table *table)
1218 {
1219         struct nft_chain *chain;
1220         int err, i = 0;
1221 
1222         list_for_each_entry(chain, &table->chains, list) {
1223                 if (!nft_is_active_next(net, chain))
1224                         continue;
1225                 if (!nft_is_base_chain(chain))
1226                         continue;
1227 
1228                 err = nf_tables_register_hook(net, table, chain);
1229                 if (err < 0)
1230                         goto err_register_hooks;
1231 
1232                 i++;
1233         }
1234         return 0;
1235 
1236 err_register_hooks:
1237         if (i)
1238                 nft_table_disable(net, table, i);
1239         return err;
1240 }
1241 
1242 static void nf_tables_table_disable(struct net *net, struct nft_table *table)
1243 {
1244         table->flags &= ~NFT_TABLE_F_DORMANT;
1245         nft_table_disable(net, table, 0);
1246         table->flags |= NFT_TABLE_F_DORMANT;
1247 }
1248 
1249 #define __NFT_TABLE_F_INTERNAL          (NFT_TABLE_F_MASK + 1)
1250 #define __NFT_TABLE_F_WAS_DORMANT       (__NFT_TABLE_F_INTERNAL << 0)
1251 #define __NFT_TABLE_F_WAS_AWAKEN        (__NFT_TABLE_F_INTERNAL << 1)
1252 #define __NFT_TABLE_F_WAS_ORPHAN        (__NFT_TABLE_F_INTERNAL << 2)
1253 #define __NFT_TABLE_F_UPDATE            (__NFT_TABLE_F_WAS_DORMANT | \
1254                                          __NFT_TABLE_F_WAS_AWAKEN | \
1255                                          __NFT_TABLE_F_WAS_ORPHAN)
1256 
1257 static bool nft_table_pending_update(const struct nft_ctx *ctx)
1258 {
1259         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
1260         struct nft_trans *trans;
1261 
1262         if (ctx->table->flags & __NFT_TABLE_F_UPDATE)
1263                 return true;
1264 
1265         list_for_each_entry(trans, &nft_net->commit_list, list) {
1266                 if (trans->table == ctx->table &&
1267                     ((trans->msg_type == NFT_MSG_NEWCHAIN &&
1268                       nft_trans_chain_update(trans)) ||
1269                      (trans->msg_type == NFT_MSG_DELCHAIN &&
1270                       nft_is_base_chain(nft_trans_chain(trans)))))
1271                         return true;
1272         }
1273 
1274         return false;
1275 }
1276 
1277 static int nf_tables_updtable(struct nft_ctx *ctx)
1278 {
1279         struct nft_trans *trans;
1280         u32 flags;
1281         int ret;
1282 
1283         if (!ctx->nla[NFTA_TABLE_FLAGS])
1284                 return 0;
1285 
1286         flags = ntohl(nla_get_be32(ctx->nla[NFTA_TABLE_FLAGS]));
1287         if (flags & ~NFT_TABLE_F_MASK)
1288                 return -EOPNOTSUPP;
1289 
1290         if (flags == (ctx->table->flags & NFT_TABLE_F_MASK))
1291                 return 0;
1292 
1293         if ((nft_table_has_owner(ctx->table) &&
1294              !(flags & NFT_TABLE_F_OWNER)) ||
1295             (flags & NFT_TABLE_F_OWNER &&
1296              !nft_table_is_orphan(ctx->table)))
1297                 return -EOPNOTSUPP;
1298 
1299         if ((flags ^ ctx->table->flags) & NFT_TABLE_F_PERSIST)
1300                 return -EOPNOTSUPP;
1301 
1302         /* No dormant off/on/off/on games in single transaction */
1303         if (nft_table_pending_update(ctx))
1304                 return -EINVAL;
1305 
1306         trans = nft_trans_alloc(ctx, NFT_MSG_NEWTABLE,
1307                                 sizeof(struct nft_trans_table));
1308         if (trans == NULL)
1309                 return -ENOMEM;
1310 
1311         if ((flags & NFT_TABLE_F_DORMANT) &&
1312             !(ctx->table->flags & NFT_TABLE_F_DORMANT)) {
1313                 ctx->table->flags |= NFT_TABLE_F_DORMANT;
1314                 if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE))
1315                         ctx->table->flags |= __NFT_TABLE_F_WAS_AWAKEN;
1316         } else if (!(flags & NFT_TABLE_F_DORMANT) &&
1317                    ctx->table->flags & NFT_TABLE_F_DORMANT) {
1318                 ctx->table->flags &= ~NFT_TABLE_F_DORMANT;
1319                 if (!(ctx->table->flags & __NFT_TABLE_F_UPDATE)) {
1320                         ret = nf_tables_table_enable(ctx->net, ctx->table);
1321                         if (ret < 0)
1322                                 goto err_register_hooks;
1323 
1324                         ctx->table->flags |= __NFT_TABLE_F_WAS_DORMANT;
1325                 }
1326         }
1327 
1328         if ((flags & NFT_TABLE_F_OWNER) &&
1329             !nft_table_has_owner(ctx->table)) {
1330                 ctx->table->nlpid = ctx->portid;
1331                 ctx->table->flags |= NFT_TABLE_F_OWNER |
1332                                      __NFT_TABLE_F_WAS_ORPHAN;
1333         }
1334 
1335         nft_trans_table_update(trans) = true;
1336         nft_trans_commit_list_add_tail(ctx->net, trans);
1337 
1338         return 0;
1339 
1340 err_register_hooks:
1341         ctx->table->flags |= NFT_TABLE_F_DORMANT;
1342         nft_trans_destroy(trans);
1343         return ret;
1344 }
1345 
1346 static u32 nft_chain_hash(const void *data, u32 len, u32 seed)
1347 {
1348         const char *name = data;
1349 
1350         return jhash(name, strlen(name), seed);
1351 }
1352 
1353 static u32 nft_chain_hash_obj(const void *data, u32 len, u32 seed)
1354 {
1355         const struct nft_chain *chain = data;
1356 
1357         return nft_chain_hash(chain->name, 0, seed);
1358 }
1359 
1360 static int nft_chain_hash_cmp(struct rhashtable_compare_arg *arg,
1361                               const void *ptr)
1362 {
1363         const struct nft_chain *chain = ptr;
1364         const char *name = arg->key;
1365 
1366         return strcmp(chain->name, name);
1367 }
1368 
1369 static u32 nft_objname_hash(const void *data, u32 len, u32 seed)
1370 {
1371         const struct nft_object_hash_key *k = data;
1372 
1373         seed ^= hash_ptr(k->table, 32);
1374 
1375         return jhash(k->name, strlen(k->name), seed);
1376 }
1377 
1378 static u32 nft_objname_hash_obj(const void *data, u32 len, u32 seed)
1379 {
1380         const struct nft_object *obj = data;
1381 
1382         return nft_objname_hash(&obj->key, 0, seed);
1383 }
1384 
1385 static int nft_objname_hash_cmp(struct rhashtable_compare_arg *arg,
1386                                 const void *ptr)
1387 {
1388         const struct nft_object_hash_key *k = arg->key;
1389         const struct nft_object *obj = ptr;
1390 
1391         if (obj->key.table != k->table)
1392                 return -1;
1393 
1394         return strcmp(obj->key.name, k->name);
1395 }
1396 
1397 static bool nft_supported_family(u8 family)
1398 {
1399         return false
1400 #ifdef CONFIG_NF_TABLES_INET
1401                 || family == NFPROTO_INET
1402 #endif
1403 #ifdef CONFIG_NF_TABLES_IPV4
1404                 || family == NFPROTO_IPV4
1405 #endif
1406 #ifdef CONFIG_NF_TABLES_ARP
1407                 || family == NFPROTO_ARP
1408 #endif
1409 #ifdef CONFIG_NF_TABLES_NETDEV
1410                 || family == NFPROTO_NETDEV
1411 #endif
1412 #if IS_ENABLED(CONFIG_NF_TABLES_BRIDGE)
1413                 || family == NFPROTO_BRIDGE
1414 #endif
1415 #ifdef CONFIG_NF_TABLES_IPV6
1416                 || family == NFPROTO_IPV6
1417 #endif
1418                 ;
1419 }
1420 
1421 static int nf_tables_newtable(struct sk_buff *skb, const struct nfnl_info *info,
1422                               const struct nlattr * const nla[])
1423 {
1424         struct nftables_pernet *nft_net = nft_pernet(info->net);
1425         struct netlink_ext_ack *extack = info->extack;
1426         u8 genmask = nft_genmask_next(info->net);
1427         u8 family = info->nfmsg->nfgen_family;
1428         struct net *net = info->net;
1429         const struct nlattr *attr;
1430         struct nft_table *table;
1431         struct nft_ctx ctx;
1432         u32 flags = 0;
1433         int err;
1434 
1435         if (!nft_supported_family(family))
1436                 return -EOPNOTSUPP;
1437 
1438         lockdep_assert_held(&nft_net->commit_mutex);
1439         attr = nla[NFTA_TABLE_NAME];
1440         table = nft_table_lookup(net, attr, family, genmask,
1441                                  NETLINK_CB(skb).portid);
1442         if (IS_ERR(table)) {
1443                 if (PTR_ERR(table) != -ENOENT)
1444                         return PTR_ERR(table);
1445         } else {
1446                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
1447                         NL_SET_BAD_ATTR(extack, attr);
1448                         return -EEXIST;
1449                 }
1450                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
1451                         return -EOPNOTSUPP;
1452 
1453                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1454 
1455                 return nf_tables_updtable(&ctx);
1456         }
1457 
1458         if (nla[NFTA_TABLE_FLAGS]) {
1459                 flags = ntohl(nla_get_be32(nla[NFTA_TABLE_FLAGS]));
1460                 if (flags & ~NFT_TABLE_F_MASK)
1461                         return -EOPNOTSUPP;
1462         }
1463 
1464         err = -ENOMEM;
1465         table = kzalloc(sizeof(*table), GFP_KERNEL_ACCOUNT);
1466         if (table == NULL)
1467                 goto err_kzalloc;
1468 
1469         table->validate_state = nft_net->validate_state;
1470         table->name = nla_strdup(attr, GFP_KERNEL_ACCOUNT);
1471         if (table->name == NULL)
1472                 goto err_strdup;
1473 
1474         if (nla[NFTA_TABLE_USERDATA]) {
1475                 table->udata = nla_memdup(nla[NFTA_TABLE_USERDATA], GFP_KERNEL_ACCOUNT);
1476                 if (table->udata == NULL)
1477                         goto err_table_udata;
1478 
1479                 table->udlen = nla_len(nla[NFTA_TABLE_USERDATA]);
1480         }
1481 
1482         err = rhltable_init(&table->chains_ht, &nft_chain_ht_params);
1483         if (err)
1484                 goto err_chain_ht;
1485 
1486         INIT_LIST_HEAD(&table->chains);
1487         INIT_LIST_HEAD(&table->sets);
1488         INIT_LIST_HEAD(&table->objects);
1489         INIT_LIST_HEAD(&table->flowtables);
1490         table->family = family;
1491         table->flags = flags;
1492         table->handle = ++nft_net->table_handle;
1493         if (table->flags & NFT_TABLE_F_OWNER)
1494                 table->nlpid = NETLINK_CB(skb).portid;
1495 
1496         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
1497         err = nft_trans_table_add(&ctx, NFT_MSG_NEWTABLE);
1498         if (err < 0)
1499                 goto err_trans;
1500 
1501         list_add_tail_rcu(&table->list, &nft_net->tables);
1502         return 0;
1503 err_trans:
1504         rhltable_destroy(&table->chains_ht);
1505 err_chain_ht:
1506         kfree(table->udata);
1507 err_table_udata:
1508         kfree(table->name);
1509 err_strdup:
1510         kfree(table);
1511 err_kzalloc:
1512         return err;
1513 }
1514 
1515 static int nft_flush_table(struct nft_ctx *ctx)
1516 {
1517         struct nft_flowtable *flowtable, *nft;
1518         struct nft_chain *chain, *nc;
1519         struct nft_object *obj, *ne;
1520         struct nft_set *set, *ns;
1521         int err;
1522 
1523         list_for_each_entry(chain, &ctx->table->chains, list) {
1524                 if (!nft_is_active_next(ctx->net, chain))
1525                         continue;
1526 
1527                 if (nft_chain_binding(chain))
1528                         continue;
1529 
1530                 ctx->chain = chain;
1531 
1532                 err = nft_delrule_by_chain(ctx);
1533                 if (err < 0)
1534                         goto out;
1535         }
1536 
1537         list_for_each_entry_safe(set, ns, &ctx->table->sets, list) {
1538                 if (!nft_is_active_next(ctx->net, set))
1539                         continue;
1540 
1541                 if (nft_set_is_anonymous(set))
1542                         continue;
1543 
1544                 err = nft_delset(ctx, set);
1545                 if (err < 0)
1546                         goto out;
1547         }
1548 
1549         list_for_each_entry_safe(flowtable, nft, &ctx->table->flowtables, list) {
1550                 if (!nft_is_active_next(ctx->net, flowtable))
1551                         continue;
1552 
1553                 err = nft_delflowtable(ctx, flowtable);
1554                 if (err < 0)
1555                         goto out;
1556         }
1557 
1558         list_for_each_entry_safe(obj, ne, &ctx->table->objects, list) {
1559                 if (!nft_is_active_next(ctx->net, obj))
1560                         continue;
1561 
1562                 err = nft_delobj(ctx, obj);
1563                 if (err < 0)
1564                         goto out;
1565         }
1566 
1567         list_for_each_entry_safe(chain, nc, &ctx->table->chains, list) {
1568                 if (!nft_is_active_next(ctx->net, chain))
1569                         continue;
1570 
1571                 if (nft_chain_binding(chain))
1572                         continue;
1573 
1574                 ctx->chain = chain;
1575 
1576                 err = nft_delchain(ctx);
1577                 if (err < 0)
1578                         goto out;
1579         }
1580 
1581         err = nft_deltable(ctx);
1582 out:
1583         return err;
1584 }
1585 
1586 static int nft_flush(struct nft_ctx *ctx, int family)
1587 {
1588         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
1589         const struct nlattr * const *nla = ctx->nla;
1590         struct nft_table *table, *nt;
1591         int err = 0;
1592 
1593         list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
1594                 if (family != AF_UNSPEC && table->family != family)
1595                         continue;
1596 
1597                 ctx->family = table->family;
1598 
1599                 if (!nft_is_active_next(ctx->net, table))
1600                         continue;
1601 
1602                 if (nft_table_has_owner(table) && table->nlpid != ctx->portid)
1603                         continue;
1604 
1605                 if (nla[NFTA_TABLE_NAME] &&
1606                     nla_strcmp(nla[NFTA_TABLE_NAME], table->name) != 0)
1607                         continue;
1608 
1609                 ctx->table = table;
1610 
1611                 err = nft_flush_table(ctx);
1612                 if (err < 0)
1613                         goto out;
1614         }
1615 out:
1616         return err;
1617 }
1618 
1619 static int nf_tables_deltable(struct sk_buff *skb, const struct nfnl_info *info,
1620                               const struct nlattr * const nla[])
1621 {
1622         struct netlink_ext_ack *extack = info->extack;
1623         u8 genmask = nft_genmask_next(info->net);
1624         u8 family = info->nfmsg->nfgen_family;
1625         struct net *net = info->net;
1626         const struct nlattr *attr;
1627         struct nft_table *table;
1628         struct nft_ctx ctx;
1629 
1630         nft_ctx_init(&ctx, net, skb, info->nlh, 0, NULL, NULL, nla);
1631         if (family == AF_UNSPEC ||
1632             (!nla[NFTA_TABLE_NAME] && !nla[NFTA_TABLE_HANDLE]))
1633                 return nft_flush(&ctx, family);
1634 
1635         if (nla[NFTA_TABLE_HANDLE]) {
1636                 attr = nla[NFTA_TABLE_HANDLE];
1637                 table = nft_table_lookup_byhandle(net, attr, family, genmask,
1638                                                   NETLINK_CB(skb).portid);
1639         } else {
1640                 attr = nla[NFTA_TABLE_NAME];
1641                 table = nft_table_lookup(net, attr, family, genmask,
1642                                          NETLINK_CB(skb).portid);
1643         }
1644 
1645         if (IS_ERR(table)) {
1646                 if (PTR_ERR(table) == -ENOENT &&
1647                     NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYTABLE)
1648                         return 0;
1649 
1650                 NL_SET_BAD_ATTR(extack, attr);
1651                 return PTR_ERR(table);
1652         }
1653 
1654         if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
1655             table->use > 0)
1656                 return -EBUSY;
1657 
1658         ctx.family = family;
1659         ctx.table = table;
1660 
1661         return nft_flush_table(&ctx);
1662 }
1663 
1664 static void nf_tables_table_destroy(struct nft_table *table)
1665 {
1666         if (WARN_ON(table->use > 0))
1667                 return;
1668 
1669         rhltable_destroy(&table->chains_ht);
1670         kfree(table->name);
1671         kfree(table->udata);
1672         kfree(table);
1673 }
1674 
1675 void nft_register_chain_type(const struct nft_chain_type *ctype)
1676 {
1677         nfnl_lock(NFNL_SUBSYS_NFTABLES);
1678         if (WARN_ON(__nft_chain_type_get(ctype->family, ctype->type))) {
1679                 nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1680                 return;
1681         }
1682         chain_type[ctype->family][ctype->type] = ctype;
1683         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1684 }
1685 EXPORT_SYMBOL_GPL(nft_register_chain_type);
1686 
1687 void nft_unregister_chain_type(const struct nft_chain_type *ctype)
1688 {
1689         nfnl_lock(NFNL_SUBSYS_NFTABLES);
1690         chain_type[ctype->family][ctype->type] = NULL;
1691         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
1692 }
1693 EXPORT_SYMBOL_GPL(nft_unregister_chain_type);
1694 
1695 /*
1696  * Chains
1697  */
1698 
1699 static struct nft_chain *
1700 nft_chain_lookup_byhandle(const struct nft_table *table, u64 handle, u8 genmask)
1701 {
1702         struct nft_chain *chain;
1703 
1704         list_for_each_entry(chain, &table->chains, list) {
1705                 if (chain->handle == handle &&
1706                     nft_active_genmask(chain, genmask))
1707                         return chain;
1708         }
1709 
1710         return ERR_PTR(-ENOENT);
1711 }
1712 
1713 static bool lockdep_commit_lock_is_held(const struct net *net)
1714 {
1715 #ifdef CONFIG_PROVE_LOCKING
1716         struct nftables_pernet *nft_net = nft_pernet(net);
1717 
1718         return lockdep_is_held(&nft_net->commit_mutex);
1719 #else
1720         return true;
1721 #endif
1722 }
1723 
1724 static struct nft_chain *nft_chain_lookup(struct net *net,
1725                                           struct nft_table *table,
1726                                           const struct nlattr *nla, u8 genmask)
1727 {
1728         char search[NFT_CHAIN_MAXNAMELEN + 1];
1729         struct rhlist_head *tmp, *list;
1730         struct nft_chain *chain;
1731 
1732         if (nla == NULL)
1733                 return ERR_PTR(-EINVAL);
1734 
1735         nla_strscpy(search, nla, sizeof(search));
1736 
1737         WARN_ON(!rcu_read_lock_held() &&
1738                 !lockdep_commit_lock_is_held(net));
1739 
1740         chain = ERR_PTR(-ENOENT);
1741         rcu_read_lock();
1742         list = rhltable_lookup(&table->chains_ht, search, nft_chain_ht_params);
1743         if (!list)
1744                 goto out_unlock;
1745 
1746         rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
1747                 if (nft_active_genmask(chain, genmask))
1748                         goto out_unlock;
1749         }
1750         chain = ERR_PTR(-ENOENT);
1751 out_unlock:
1752         rcu_read_unlock();
1753         return chain;
1754 }
1755 
1756 static const struct nla_policy nft_chain_policy[NFTA_CHAIN_MAX + 1] = {
1757         [NFTA_CHAIN_TABLE]      = { .type = NLA_STRING,
1758                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
1759         [NFTA_CHAIN_HANDLE]     = { .type = NLA_U64 },
1760         [NFTA_CHAIN_NAME]       = { .type = NLA_STRING,
1761                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
1762         [NFTA_CHAIN_HOOK]       = { .type = NLA_NESTED },
1763         [NFTA_CHAIN_POLICY]     = { .type = NLA_U32 },
1764         [NFTA_CHAIN_TYPE]       = { .type = NLA_STRING,
1765                                     .len = NFT_MODULE_AUTOLOAD_LIMIT },
1766         [NFTA_CHAIN_COUNTERS]   = { .type = NLA_NESTED },
1767         [NFTA_CHAIN_FLAGS]      = { .type = NLA_U32 },
1768         [NFTA_CHAIN_ID]         = { .type = NLA_U32 },
1769         [NFTA_CHAIN_USERDATA]   = { .type = NLA_BINARY,
1770                                     .len = NFT_USERDATA_MAXLEN },
1771 };
1772 
1773 static const struct nla_policy nft_hook_policy[NFTA_HOOK_MAX + 1] = {
1774         [NFTA_HOOK_HOOKNUM]     = { .type = NLA_U32 },
1775         [NFTA_HOOK_PRIORITY]    = { .type = NLA_U32 },
1776         [NFTA_HOOK_DEV]         = { .type = NLA_STRING,
1777                                     .len = IFNAMSIZ - 1 },
1778 };
1779 
1780 static int nft_dump_stats(struct sk_buff *skb, struct nft_stats __percpu *stats)
1781 {
1782         struct nft_stats *cpu_stats, total;
1783         struct nlattr *nest;
1784         unsigned int seq;
1785         u64 pkts, bytes;
1786         int cpu;
1787 
1788         if (!stats)
1789                 return 0;
1790 
1791         memset(&total, 0, sizeof(total));
1792         for_each_possible_cpu(cpu) {
1793                 cpu_stats = per_cpu_ptr(stats, cpu);
1794                 do {
1795                         seq = u64_stats_fetch_begin(&cpu_stats->syncp);
1796                         pkts = cpu_stats->pkts;
1797                         bytes = cpu_stats->bytes;
1798                 } while (u64_stats_fetch_retry(&cpu_stats->syncp, seq));
1799                 total.pkts += pkts;
1800                 total.bytes += bytes;
1801         }
1802         nest = nla_nest_start_noflag(skb, NFTA_CHAIN_COUNTERS);
1803         if (nest == NULL)
1804                 goto nla_put_failure;
1805 
1806         if (nla_put_be64(skb, NFTA_COUNTER_PACKETS, cpu_to_be64(total.pkts),
1807                          NFTA_COUNTER_PAD) ||
1808             nla_put_be64(skb, NFTA_COUNTER_BYTES, cpu_to_be64(total.bytes),
1809                          NFTA_COUNTER_PAD))
1810                 goto nla_put_failure;
1811 
1812         nla_nest_end(skb, nest);
1813         return 0;
1814 
1815 nla_put_failure:
1816         return -ENOSPC;
1817 }
1818 
1819 static int nft_dump_basechain_hook(struct sk_buff *skb, int family,
1820                                    const struct nft_base_chain *basechain,
1821                                    const struct list_head *hook_list)
1822 {
1823         const struct nf_hook_ops *ops = &basechain->ops;
1824         struct nft_hook *hook, *first = NULL;
1825         struct nlattr *nest, *nest_devs;
1826         int n = 0;
1827 
1828         nest = nla_nest_start_noflag(skb, NFTA_CHAIN_HOOK);
1829         if (nest == NULL)
1830                 goto nla_put_failure;
1831         if (nla_put_be32(skb, NFTA_HOOK_HOOKNUM, htonl(ops->hooknum)))
1832                 goto nla_put_failure;
1833         if (nla_put_be32(skb, NFTA_HOOK_PRIORITY, htonl(ops->priority)))
1834                 goto nla_put_failure;
1835 
1836         if (nft_base_chain_netdev(family, ops->hooknum)) {
1837                 nest_devs = nla_nest_start_noflag(skb, NFTA_HOOK_DEVS);
1838                 if (!nest_devs)
1839                         goto nla_put_failure;
1840 
1841                 if (!hook_list)
1842                         hook_list = &basechain->hook_list;
1843 
1844                 list_for_each_entry(hook, hook_list, list) {
1845                         if (!first)
1846                                 first = hook;
1847 
1848                         if (nla_put_string(skb, NFTA_DEVICE_NAME,
1849                                            hook->ops.dev->name))
1850                                 goto nla_put_failure;
1851                         n++;
1852                 }
1853                 nla_nest_end(skb, nest_devs);
1854 
1855                 if (n == 1 &&
1856                     nla_put_string(skb, NFTA_HOOK_DEV, first->ops.dev->name))
1857                         goto nla_put_failure;
1858         }
1859         nla_nest_end(skb, nest);
1860 
1861         return 0;
1862 nla_put_failure:
1863         return -1;
1864 }
1865 
1866 static int nf_tables_fill_chain_info(struct sk_buff *skb, struct net *net,
1867                                      u32 portid, u32 seq, int event, u32 flags,
1868                                      int family, const struct nft_table *table,
1869                                      const struct nft_chain *chain,
1870                                      const struct list_head *hook_list)
1871 {
1872         struct nlmsghdr *nlh;
1873 
1874         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
1875         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
1876                            NFNETLINK_V0, nft_base_seq(net));
1877         if (!nlh)
1878                 goto nla_put_failure;
1879 
1880         if (nla_put_string(skb, NFTA_CHAIN_TABLE, table->name) ||
1881             nla_put_string(skb, NFTA_CHAIN_NAME, chain->name) ||
1882             nla_put_be64(skb, NFTA_CHAIN_HANDLE, cpu_to_be64(chain->handle),
1883                          NFTA_CHAIN_PAD))
1884                 goto nla_put_failure;
1885 
1886         if (event == NFT_MSG_DELCHAIN && !hook_list) {
1887                 nlmsg_end(skb, nlh);
1888                 return 0;
1889         }
1890 
1891         if (nft_is_base_chain(chain)) {
1892                 const struct nft_base_chain *basechain = nft_base_chain(chain);
1893                 struct nft_stats __percpu *stats;
1894 
1895                 if (nft_dump_basechain_hook(skb, family, basechain, hook_list))
1896                         goto nla_put_failure;
1897 
1898                 if (nla_put_be32(skb, NFTA_CHAIN_POLICY,
1899                                  htonl(basechain->policy)))
1900                         goto nla_put_failure;
1901 
1902                 if (nla_put_string(skb, NFTA_CHAIN_TYPE, basechain->type->name))
1903                         goto nla_put_failure;
1904 
1905                 stats = rcu_dereference_check(basechain->stats,
1906                                               lockdep_commit_lock_is_held(net));
1907                 if (nft_dump_stats(skb, stats))
1908                         goto nla_put_failure;
1909         }
1910 
1911         if (chain->flags &&
1912             nla_put_be32(skb, NFTA_CHAIN_FLAGS, htonl(chain->flags)))
1913                 goto nla_put_failure;
1914 
1915         if (nla_put_be32(skb, NFTA_CHAIN_USE, htonl(chain->use)))
1916                 goto nla_put_failure;
1917 
1918         if (chain->udata &&
1919             nla_put(skb, NFTA_CHAIN_USERDATA, chain->udlen, chain->udata))
1920                 goto nla_put_failure;
1921 
1922         nlmsg_end(skb, nlh);
1923         return 0;
1924 
1925 nla_put_failure:
1926         nlmsg_trim(skb, nlh);
1927         return -1;
1928 }
1929 
1930 static void nf_tables_chain_notify(const struct nft_ctx *ctx, int event,
1931                                    const struct list_head *hook_list)
1932 {
1933         struct nftables_pernet *nft_net;
1934         struct sk_buff *skb;
1935         u16 flags = 0;
1936         int err;
1937 
1938         if (!ctx->report &&
1939             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
1940                 return;
1941 
1942         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
1943         if (skb == NULL)
1944                 goto err;
1945 
1946         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
1947                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
1948 
1949         err = nf_tables_fill_chain_info(skb, ctx->net, ctx->portid, ctx->seq,
1950                                         event, flags, ctx->family, ctx->table,
1951                                         ctx->chain, hook_list);
1952         if (err < 0) {
1953                 kfree_skb(skb);
1954                 goto err;
1955         }
1956 
1957         nft_net = nft_pernet(ctx->net);
1958         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
1959         return;
1960 err:
1961         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
1962 }
1963 
1964 static int nf_tables_dump_chains(struct sk_buff *skb,
1965                                  struct netlink_callback *cb)
1966 {
1967         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1968         unsigned int idx = 0, s_idx = cb->args[0];
1969         struct net *net = sock_net(skb->sk);
1970         int family = nfmsg->nfgen_family;
1971         struct nftables_pernet *nft_net;
1972         const struct nft_table *table;
1973         const struct nft_chain *chain;
1974 
1975         rcu_read_lock();
1976         nft_net = nft_pernet(net);
1977         cb->seq = READ_ONCE(nft_net->base_seq);
1978 
1979         list_for_each_entry_rcu(table, &nft_net->tables, list) {
1980                 if (family != NFPROTO_UNSPEC && family != table->family)
1981                         continue;
1982 
1983                 list_for_each_entry_rcu(chain, &table->chains, list) {
1984                         if (idx < s_idx)
1985                                 goto cont;
1986                         if (idx > s_idx)
1987                                 memset(&cb->args[1], 0,
1988                                        sizeof(cb->args) - sizeof(cb->args[0]));
1989                         if (!nft_is_active(net, chain))
1990                                 continue;
1991                         if (nf_tables_fill_chain_info(skb, net,
1992                                                       NETLINK_CB(cb->skb).portid,
1993                                                       cb->nlh->nlmsg_seq,
1994                                                       NFT_MSG_NEWCHAIN,
1995                                                       NLM_F_MULTI,
1996                                                       table->family, table,
1997                                                       chain, NULL) < 0)
1998                                 goto done;
1999 
2000                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
2001 cont:
2002                         idx++;
2003                 }
2004         }
2005 done:
2006         rcu_read_unlock();
2007         cb->args[0] = idx;
2008         return skb->len;
2009 }
2010 
2011 /* called with rcu_read_lock held */
2012 static int nf_tables_getchain(struct sk_buff *skb, const struct nfnl_info *info,
2013                               const struct nlattr * const nla[])
2014 {
2015         struct netlink_ext_ack *extack = info->extack;
2016         u8 genmask = nft_genmask_cur(info->net);
2017         u8 family = info->nfmsg->nfgen_family;
2018         const struct nft_chain *chain;
2019         struct net *net = info->net;
2020         struct nft_table *table;
2021         struct sk_buff *skb2;
2022         int err;
2023 
2024         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
2025                 struct netlink_dump_control c = {
2026                         .dump = nf_tables_dump_chains,
2027                         .module = THIS_MODULE,
2028                 };
2029 
2030                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
2031         }
2032 
2033         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask, 0);
2034         if (IS_ERR(table)) {
2035                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2036                 return PTR_ERR(table);
2037         }
2038 
2039         chain = nft_chain_lookup(net, table, nla[NFTA_CHAIN_NAME], genmask);
2040         if (IS_ERR(chain)) {
2041                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2042                 return PTR_ERR(chain);
2043         }
2044 
2045         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
2046         if (!skb2)
2047                 return -ENOMEM;
2048 
2049         err = nf_tables_fill_chain_info(skb2, net, NETLINK_CB(skb).portid,
2050                                         info->nlh->nlmsg_seq, NFT_MSG_NEWCHAIN,
2051                                         0, family, table, chain, NULL);
2052         if (err < 0)
2053                 goto err_fill_chain_info;
2054 
2055         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
2056 
2057 err_fill_chain_info:
2058         kfree_skb(skb2);
2059         return err;
2060 }
2061 
2062 static const struct nla_policy nft_counter_policy[NFTA_COUNTER_MAX + 1] = {
2063         [NFTA_COUNTER_PACKETS]  = { .type = NLA_U64 },
2064         [NFTA_COUNTER_BYTES]    = { .type = NLA_U64 },
2065 };
2066 
2067 static struct nft_stats __percpu *nft_stats_alloc(const struct nlattr *attr)
2068 {
2069         struct nlattr *tb[NFTA_COUNTER_MAX+1];
2070         struct nft_stats __percpu *newstats;
2071         struct nft_stats *stats;
2072         int err;
2073 
2074         err = nla_parse_nested_deprecated(tb, NFTA_COUNTER_MAX, attr,
2075                                           nft_counter_policy, NULL);
2076         if (err < 0)
2077                 return ERR_PTR(err);
2078 
2079         if (!tb[NFTA_COUNTER_BYTES] || !tb[NFTA_COUNTER_PACKETS])
2080                 return ERR_PTR(-EINVAL);
2081 
2082         newstats = netdev_alloc_pcpu_stats(struct nft_stats);
2083         if (newstats == NULL)
2084                 return ERR_PTR(-ENOMEM);
2085 
2086         /* Restore old counters on this cpu, no problem. Per-cpu statistics
2087          * are not exposed to userspace.
2088          */
2089         preempt_disable();
2090         stats = this_cpu_ptr(newstats);
2091         stats->bytes = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_BYTES]));
2092         stats->pkts = be64_to_cpu(nla_get_be64(tb[NFTA_COUNTER_PACKETS]));
2093         preempt_enable();
2094 
2095         return newstats;
2096 }
2097 
2098 static void nft_chain_stats_replace(struct nft_trans_chain *trans)
2099 {
2100         const struct nft_trans *t = &trans->nft_trans_binding.nft_trans;
2101         struct nft_base_chain *chain = nft_base_chain(trans->chain);
2102 
2103         if (!trans->stats)
2104                 return;
2105 
2106         trans->stats =
2107                 rcu_replace_pointer(chain->stats, trans->stats,
2108                                     lockdep_commit_lock_is_held(t->net));
2109 
2110         if (!trans->stats)
2111                 static_branch_inc(&nft_counters_enabled);
2112 }
2113 
2114 static void nf_tables_chain_free_chain_rules(struct nft_chain *chain)
2115 {
2116         struct nft_rule_blob *g0 = rcu_dereference_raw(chain->blob_gen_0);
2117         struct nft_rule_blob *g1 = rcu_dereference_raw(chain->blob_gen_1);
2118 
2119         if (g0 != g1)
2120                 kvfree(g1);
2121         kvfree(g0);
2122 
2123         /* should be NULL either via abort or via successful commit */
2124         WARN_ON_ONCE(chain->blob_next);
2125         kvfree(chain->blob_next);
2126 }
2127 
2128 void nf_tables_chain_destroy(struct nft_chain *chain)
2129 {
2130         const struct nft_table *table = chain->table;
2131         struct nft_hook *hook, *next;
2132 
2133         if (WARN_ON(chain->use > 0))
2134                 return;
2135 
2136         /* no concurrent access possible anymore */
2137         nf_tables_chain_free_chain_rules(chain);
2138 
2139         if (nft_is_base_chain(chain)) {
2140                 struct nft_base_chain *basechain = nft_base_chain(chain);
2141 
2142                 if (nft_base_chain_netdev(table->family, basechain->ops.hooknum)) {
2143                         list_for_each_entry_safe(hook, next,
2144                                                  &basechain->hook_list, list) {
2145                                 list_del_rcu(&hook->list);
2146                                 kfree_rcu(hook, rcu);
2147                         }
2148                 }
2149                 module_put(basechain->type->owner);
2150                 if (rcu_access_pointer(basechain->stats)) {
2151                         static_branch_dec(&nft_counters_enabled);
2152                         free_percpu(rcu_dereference_raw(basechain->stats));
2153                 }
2154                 kfree(chain->name);
2155                 kfree(chain->udata);
2156                 kfree(basechain);
2157         } else {
2158                 kfree(chain->name);
2159                 kfree(chain->udata);
2160                 kfree(chain);
2161         }
2162 }
2163 
2164 static struct nft_hook *nft_netdev_hook_alloc(struct net *net,
2165                                               const struct nlattr *attr)
2166 {
2167         struct net_device *dev;
2168         char ifname[IFNAMSIZ];
2169         struct nft_hook *hook;
2170         int err;
2171 
2172         hook = kzalloc(sizeof(struct nft_hook), GFP_KERNEL_ACCOUNT);
2173         if (!hook) {
2174                 err = -ENOMEM;
2175                 goto err_hook_alloc;
2176         }
2177 
2178         nla_strscpy(ifname, attr, IFNAMSIZ);
2179         /* nf_tables_netdev_event() is called under rtnl_mutex, this is
2180          * indirectly serializing all the other holders of the commit_mutex with
2181          * the rtnl_mutex.
2182          */
2183         dev = __dev_get_by_name(net, ifname);
2184         if (!dev) {
2185                 err = -ENOENT;
2186                 goto err_hook_dev;
2187         }
2188         hook->ops.dev = dev;
2189 
2190         return hook;
2191 
2192 err_hook_dev:
2193         kfree(hook);
2194 err_hook_alloc:
2195         return ERR_PTR(err);
2196 }
2197 
2198 static struct nft_hook *nft_hook_list_find(struct list_head *hook_list,
2199                                            const struct nft_hook *this)
2200 {
2201         struct nft_hook *hook;
2202 
2203         list_for_each_entry(hook, hook_list, list) {
2204                 if (this->ops.dev == hook->ops.dev)
2205                         return hook;
2206         }
2207 
2208         return NULL;
2209 }
2210 
2211 static int nf_tables_parse_netdev_hooks(struct net *net,
2212                                         const struct nlattr *attr,
2213                                         struct list_head *hook_list,
2214                                         struct netlink_ext_ack *extack)
2215 {
2216         struct nft_hook *hook, *next;
2217         const struct nlattr *tmp;
2218         int rem, n = 0, err;
2219 
2220         nla_for_each_nested(tmp, attr, rem) {
2221                 if (nla_type(tmp) != NFTA_DEVICE_NAME) {
2222                         err = -EINVAL;
2223                         goto err_hook;
2224                 }
2225 
2226                 hook = nft_netdev_hook_alloc(net, tmp);
2227                 if (IS_ERR(hook)) {
2228                         NL_SET_BAD_ATTR(extack, tmp);
2229                         err = PTR_ERR(hook);
2230                         goto err_hook;
2231                 }
2232                 if (nft_hook_list_find(hook_list, hook)) {
2233                         NL_SET_BAD_ATTR(extack, tmp);
2234                         kfree(hook);
2235                         err = -EEXIST;
2236                         goto err_hook;
2237                 }
2238                 list_add_tail(&hook->list, hook_list);
2239                 n++;
2240 
2241                 if (n == NFT_NETDEVICE_MAX) {
2242                         err = -EFBIG;
2243                         goto err_hook;
2244                 }
2245         }
2246 
2247         return 0;
2248 
2249 err_hook:
2250         list_for_each_entry_safe(hook, next, hook_list, list) {
2251                 list_del(&hook->list);
2252                 kfree(hook);
2253         }
2254         return err;
2255 }
2256 
2257 struct nft_chain_hook {
2258         u32                             num;
2259         s32                             priority;
2260         const struct nft_chain_type     *type;
2261         struct list_head                list;
2262 };
2263 
2264 static int nft_chain_parse_netdev(struct net *net, struct nlattr *tb[],
2265                                   struct list_head *hook_list,
2266                                   struct netlink_ext_ack *extack, u32 flags)
2267 {
2268         struct nft_hook *hook;
2269         int err;
2270 
2271         if (tb[NFTA_HOOK_DEV]) {
2272                 hook = nft_netdev_hook_alloc(net, tb[NFTA_HOOK_DEV]);
2273                 if (IS_ERR(hook)) {
2274                         NL_SET_BAD_ATTR(extack, tb[NFTA_HOOK_DEV]);
2275                         return PTR_ERR(hook);
2276                 }
2277 
2278                 list_add_tail(&hook->list, hook_list);
2279         } else if (tb[NFTA_HOOK_DEVS]) {
2280                 err = nf_tables_parse_netdev_hooks(net, tb[NFTA_HOOK_DEVS],
2281                                                    hook_list, extack);
2282                 if (err < 0)
2283                         return err;
2284 
2285         }
2286 
2287         if (flags & NFT_CHAIN_HW_OFFLOAD &&
2288             list_empty(hook_list))
2289                 return -EINVAL;
2290 
2291         return 0;
2292 }
2293 
2294 static int nft_chain_parse_hook(struct net *net,
2295                                 struct nft_base_chain *basechain,
2296                                 const struct nlattr * const nla[],
2297                                 struct nft_chain_hook *hook, u8 family,
2298                                 u32 flags, struct netlink_ext_ack *extack)
2299 {
2300         struct nftables_pernet *nft_net = nft_pernet(net);
2301         struct nlattr *ha[NFTA_HOOK_MAX + 1];
2302         const struct nft_chain_type *type;
2303         int err;
2304 
2305         lockdep_assert_held(&nft_net->commit_mutex);
2306         lockdep_nfnl_nft_mutex_not_held();
2307 
2308         err = nla_parse_nested_deprecated(ha, NFTA_HOOK_MAX,
2309                                           nla[NFTA_CHAIN_HOOK],
2310                                           nft_hook_policy, NULL);
2311         if (err < 0)
2312                 return err;
2313 
2314         if (!basechain) {
2315                 if (!ha[NFTA_HOOK_HOOKNUM] ||
2316                     !ha[NFTA_HOOK_PRIORITY]) {
2317                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2318                         return -ENOENT;
2319                 }
2320 
2321                 hook->num = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
2322                 hook->priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
2323 
2324                 type = __nft_chain_type_get(family, NFT_CHAIN_T_DEFAULT);
2325                 if (!type)
2326                         return -EOPNOTSUPP;
2327 
2328                 if (nla[NFTA_CHAIN_TYPE]) {
2329                         type = nf_tables_chain_type_lookup(net, nla[NFTA_CHAIN_TYPE],
2330                                                            family, true);
2331                         if (IS_ERR(type)) {
2332                                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
2333                                 return PTR_ERR(type);
2334                         }
2335                 }
2336                 if (hook->num >= NFT_MAX_HOOKS || !(type->hook_mask & (1 << hook->num)))
2337                         return -EOPNOTSUPP;
2338 
2339                 if (type->type == NFT_CHAIN_T_NAT &&
2340                     hook->priority <= NF_IP_PRI_CONNTRACK)
2341                         return -EOPNOTSUPP;
2342         } else {
2343                 if (ha[NFTA_HOOK_HOOKNUM]) {
2344                         hook->num = ntohl(nla_get_be32(ha[NFTA_HOOK_HOOKNUM]));
2345                         if (hook->num != basechain->ops.hooknum)
2346                                 return -EOPNOTSUPP;
2347                 }
2348                 if (ha[NFTA_HOOK_PRIORITY]) {
2349                         hook->priority = ntohl(nla_get_be32(ha[NFTA_HOOK_PRIORITY]));
2350                         if (hook->priority != basechain->ops.priority)
2351                                 return -EOPNOTSUPP;
2352                 }
2353 
2354                 if (nla[NFTA_CHAIN_TYPE]) {
2355                         type = __nf_tables_chain_type_lookup(nla[NFTA_CHAIN_TYPE],
2356                                                              family);
2357                         if (!type) {
2358                                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
2359                                 return -ENOENT;
2360                         }
2361                 } else {
2362                         type = basechain->type;
2363                 }
2364         }
2365 
2366         if (!try_module_get(type->owner)) {
2367                 if (nla[NFTA_CHAIN_TYPE])
2368                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TYPE]);
2369                 return -ENOENT;
2370         }
2371 
2372         hook->type = type;
2373 
2374         INIT_LIST_HEAD(&hook->list);
2375         if (nft_base_chain_netdev(family, hook->num)) {
2376                 err = nft_chain_parse_netdev(net, ha, &hook->list, extack, flags);
2377                 if (err < 0) {
2378                         module_put(type->owner);
2379                         return err;
2380                 }
2381         } else if (ha[NFTA_HOOK_DEV] || ha[NFTA_HOOK_DEVS]) {
2382                 module_put(type->owner);
2383                 return -EOPNOTSUPP;
2384         }
2385 
2386         return 0;
2387 }
2388 
2389 static void nft_chain_release_hook(struct nft_chain_hook *hook)
2390 {
2391         struct nft_hook *h, *next;
2392 
2393         list_for_each_entry_safe(h, next, &hook->list, list) {
2394                 list_del(&h->list);
2395                 kfree(h);
2396         }
2397         module_put(hook->type->owner);
2398 }
2399 
2400 static void nft_last_rule(const struct nft_chain *chain, const void *ptr)
2401 {
2402         struct nft_rule_dp_last *lrule;
2403 
2404         BUILD_BUG_ON(offsetof(struct nft_rule_dp_last, end) != 0);
2405 
2406         lrule = (struct nft_rule_dp_last *)ptr;
2407         lrule->end.is_last = 1;
2408         lrule->chain = chain;
2409         /* blob size does not include the trailer rule */
2410 }
2411 
2412 static struct nft_rule_blob *nf_tables_chain_alloc_rules(const struct nft_chain *chain,
2413                                                          unsigned int size)
2414 {
2415         struct nft_rule_blob *blob;
2416 
2417         if (size > INT_MAX)
2418                 return NULL;
2419 
2420         size += sizeof(struct nft_rule_blob) + sizeof(struct nft_rule_dp_last);
2421 
2422         blob = kvmalloc(size, GFP_KERNEL_ACCOUNT);
2423         if (!blob)
2424                 return NULL;
2425 
2426         blob->size = 0;
2427         nft_last_rule(chain, blob->data);
2428 
2429         return blob;
2430 }
2431 
2432 static void nft_basechain_hook_init(struct nf_hook_ops *ops, u8 family,
2433                                     const struct nft_chain_hook *hook,
2434                                     struct nft_chain *chain)
2435 {
2436         ops->pf                 = family;
2437         ops->hooknum            = hook->num;
2438         ops->priority           = hook->priority;
2439         ops->priv               = chain;
2440         ops->hook               = hook->type->hooks[ops->hooknum];
2441         ops->hook_ops_type      = NF_HOOK_OP_NF_TABLES;
2442 }
2443 
2444 static int nft_basechain_init(struct nft_base_chain *basechain, u8 family,
2445                               struct nft_chain_hook *hook, u32 flags)
2446 {
2447         struct nft_chain *chain;
2448         struct nft_hook *h;
2449 
2450         basechain->type = hook->type;
2451         INIT_LIST_HEAD(&basechain->hook_list);
2452         chain = &basechain->chain;
2453 
2454         if (nft_base_chain_netdev(family, hook->num)) {
2455                 list_splice_init(&hook->list, &basechain->hook_list);
2456                 list_for_each_entry(h, &basechain->hook_list, list)
2457                         nft_basechain_hook_init(&h->ops, family, hook, chain);
2458         }
2459         nft_basechain_hook_init(&basechain->ops, family, hook, chain);
2460 
2461         chain->flags |= NFT_CHAIN_BASE | flags;
2462         basechain->policy = NF_ACCEPT;
2463         if (chain->flags & NFT_CHAIN_HW_OFFLOAD &&
2464             !nft_chain_offload_support(basechain)) {
2465                 list_splice_init(&basechain->hook_list, &hook->list);
2466                 return -EOPNOTSUPP;
2467         }
2468 
2469         flow_block_init(&basechain->flow_block);
2470 
2471         return 0;
2472 }
2473 
2474 int nft_chain_add(struct nft_table *table, struct nft_chain *chain)
2475 {
2476         int err;
2477 
2478         err = rhltable_insert_key(&table->chains_ht, chain->name,
2479                                   &chain->rhlhead, nft_chain_ht_params);
2480         if (err)
2481                 return err;
2482 
2483         list_add_tail_rcu(&chain->list, &table->chains);
2484 
2485         return 0;
2486 }
2487 
2488 static u64 chain_id;
2489 
2490 static int nf_tables_addchain(struct nft_ctx *ctx, u8 family, u8 genmask,
2491                               u8 policy, u32 flags,
2492                               struct netlink_ext_ack *extack)
2493 {
2494         const struct nlattr * const *nla = ctx->nla;
2495         struct nft_table *table = ctx->table;
2496         struct nft_base_chain *basechain;
2497         struct net *net = ctx->net;
2498         char name[NFT_NAME_MAXLEN];
2499         struct nft_rule_blob *blob;
2500         struct nft_trans *trans;
2501         struct nft_chain *chain;
2502         int err;
2503 
2504         if (nla[NFTA_CHAIN_HOOK]) {
2505                 struct nft_stats __percpu *stats = NULL;
2506                 struct nft_chain_hook hook = {};
2507 
2508                 if (table->flags & __NFT_TABLE_F_UPDATE)
2509                         return -EINVAL;
2510 
2511                 if (flags & NFT_CHAIN_BINDING)
2512                         return -EOPNOTSUPP;
2513 
2514                 err = nft_chain_parse_hook(net, NULL, nla, &hook, family, flags,
2515                                            extack);
2516                 if (err < 0)
2517                         return err;
2518 
2519                 basechain = kzalloc(sizeof(*basechain), GFP_KERNEL_ACCOUNT);
2520                 if (basechain == NULL) {
2521                         nft_chain_release_hook(&hook);
2522                         return -ENOMEM;
2523                 }
2524                 chain = &basechain->chain;
2525 
2526                 if (nla[NFTA_CHAIN_COUNTERS]) {
2527                         stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2528                         if (IS_ERR(stats)) {
2529                                 nft_chain_release_hook(&hook);
2530                                 kfree(basechain);
2531                                 return PTR_ERR(stats);
2532                         }
2533                         rcu_assign_pointer(basechain->stats, stats);
2534                 }
2535 
2536                 err = nft_basechain_init(basechain, family, &hook, flags);
2537                 if (err < 0) {
2538                         nft_chain_release_hook(&hook);
2539                         kfree(basechain);
2540                         free_percpu(stats);
2541                         return err;
2542                 }
2543                 if (stats)
2544                         static_branch_inc(&nft_counters_enabled);
2545         } else {
2546                 if (flags & NFT_CHAIN_BASE)
2547                         return -EINVAL;
2548                 if (flags & NFT_CHAIN_HW_OFFLOAD)
2549                         return -EOPNOTSUPP;
2550 
2551                 chain = kzalloc(sizeof(*chain), GFP_KERNEL_ACCOUNT);
2552                 if (chain == NULL)
2553                         return -ENOMEM;
2554 
2555                 chain->flags = flags;
2556         }
2557         ctx->chain = chain;
2558 
2559         INIT_LIST_HEAD(&chain->rules);
2560         chain->handle = nf_tables_alloc_handle(table);
2561         chain->table = table;
2562 
2563         if (nla[NFTA_CHAIN_NAME]) {
2564                 chain->name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL_ACCOUNT);
2565         } else {
2566                 if (!(flags & NFT_CHAIN_BINDING)) {
2567                         err = -EINVAL;
2568                         goto err_destroy_chain;
2569                 }
2570 
2571                 snprintf(name, sizeof(name), "__chain%llu", ++chain_id);
2572                 chain->name = kstrdup(name, GFP_KERNEL_ACCOUNT);
2573         }
2574 
2575         if (!chain->name) {
2576                 err = -ENOMEM;
2577                 goto err_destroy_chain;
2578         }
2579 
2580         if (nla[NFTA_CHAIN_USERDATA]) {
2581                 chain->udata = nla_memdup(nla[NFTA_CHAIN_USERDATA], GFP_KERNEL_ACCOUNT);
2582                 if (chain->udata == NULL) {
2583                         err = -ENOMEM;
2584                         goto err_destroy_chain;
2585                 }
2586                 chain->udlen = nla_len(nla[NFTA_CHAIN_USERDATA]);
2587         }
2588 
2589         blob = nf_tables_chain_alloc_rules(chain, 0);
2590         if (!blob) {
2591                 err = -ENOMEM;
2592                 goto err_destroy_chain;
2593         }
2594 
2595         RCU_INIT_POINTER(chain->blob_gen_0, blob);
2596         RCU_INIT_POINTER(chain->blob_gen_1, blob);
2597 
2598         if (!nft_use_inc(&table->use)) {
2599                 err = -EMFILE;
2600                 goto err_destroy_chain;
2601         }
2602 
2603         trans = nft_trans_chain_add(ctx, NFT_MSG_NEWCHAIN);
2604         if (IS_ERR(trans)) {
2605                 err = PTR_ERR(trans);
2606                 goto err_trans;
2607         }
2608 
2609         nft_trans_chain_policy(trans) = NFT_CHAIN_POLICY_UNSET;
2610         if (nft_is_base_chain(chain))
2611                 nft_trans_chain_policy(trans) = policy;
2612 
2613         err = nft_chain_add(table, chain);
2614         if (err < 0)
2615                 goto err_chain_add;
2616 
2617         /* This must be LAST to ensure no packets are walking over this chain. */
2618         err = nf_tables_register_hook(net, table, chain);
2619         if (err < 0)
2620                 goto err_register_hook;
2621 
2622         return 0;
2623 
2624 err_register_hook:
2625         nft_chain_del(chain);
2626 err_chain_add:
2627         nft_trans_destroy(trans);
2628 err_trans:
2629         nft_use_dec_restore(&table->use);
2630 err_destroy_chain:
2631         nf_tables_chain_destroy(chain);
2632 
2633         return err;
2634 }
2635 
2636 static int nf_tables_updchain(struct nft_ctx *ctx, u8 genmask, u8 policy,
2637                               u32 flags, const struct nlattr *attr,
2638                               struct netlink_ext_ack *extack)
2639 {
2640         const struct nlattr * const *nla = ctx->nla;
2641         struct nft_base_chain *basechain = NULL;
2642         struct nft_table *table = ctx->table;
2643         struct nft_chain *chain = ctx->chain;
2644         struct nft_chain_hook hook = {};
2645         struct nft_stats *stats = NULL;
2646         struct nft_hook *h, *next;
2647         struct nf_hook_ops *ops;
2648         struct nft_trans *trans;
2649         bool unregister = false;
2650         int err;
2651 
2652         if (chain->flags ^ flags)
2653                 return -EOPNOTSUPP;
2654 
2655         INIT_LIST_HEAD(&hook.list);
2656 
2657         if (nla[NFTA_CHAIN_HOOK]) {
2658                 if (!nft_is_base_chain(chain)) {
2659                         NL_SET_BAD_ATTR(extack, attr);
2660                         return -EEXIST;
2661                 }
2662 
2663                 basechain = nft_base_chain(chain);
2664                 err = nft_chain_parse_hook(ctx->net, basechain, nla, &hook,
2665                                            ctx->family, flags, extack);
2666                 if (err < 0)
2667                         return err;
2668 
2669                 if (basechain->type != hook.type) {
2670                         nft_chain_release_hook(&hook);
2671                         NL_SET_BAD_ATTR(extack, attr);
2672                         return -EEXIST;
2673                 }
2674 
2675                 if (nft_base_chain_netdev(ctx->family, basechain->ops.hooknum)) {
2676                         list_for_each_entry_safe(h, next, &hook.list, list) {
2677                                 h->ops.pf       = basechain->ops.pf;
2678                                 h->ops.hooknum  = basechain->ops.hooknum;
2679                                 h->ops.priority = basechain->ops.priority;
2680                                 h->ops.priv     = basechain->ops.priv;
2681                                 h->ops.hook     = basechain->ops.hook;
2682 
2683                                 if (nft_hook_list_find(&basechain->hook_list, h)) {
2684                                         list_del(&h->list);
2685                                         kfree(h);
2686                                 }
2687                         }
2688                 } else {
2689                         ops = &basechain->ops;
2690                         if (ops->hooknum != hook.num ||
2691                             ops->priority != hook.priority) {
2692                                 nft_chain_release_hook(&hook);
2693                                 NL_SET_BAD_ATTR(extack, attr);
2694                                 return -EEXIST;
2695                         }
2696                 }
2697         }
2698 
2699         if (nla[NFTA_CHAIN_HANDLE] &&
2700             nla[NFTA_CHAIN_NAME]) {
2701                 struct nft_chain *chain2;
2702 
2703                 chain2 = nft_chain_lookup(ctx->net, table,
2704                                           nla[NFTA_CHAIN_NAME], genmask);
2705                 if (!IS_ERR(chain2)) {
2706                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2707                         err = -EEXIST;
2708                         goto err_hooks;
2709                 }
2710         }
2711 
2712         if (table->flags & __NFT_TABLE_F_UPDATE &&
2713             !list_empty(&hook.list)) {
2714                 NL_SET_BAD_ATTR(extack, attr);
2715                 err = -EOPNOTSUPP;
2716                 goto err_hooks;
2717         }
2718 
2719         if (!(table->flags & NFT_TABLE_F_DORMANT) &&
2720             nft_is_base_chain(chain) &&
2721             !list_empty(&hook.list)) {
2722                 basechain = nft_base_chain(chain);
2723                 ops = &basechain->ops;
2724 
2725                 if (nft_base_chain_netdev(table->family, basechain->ops.hooknum)) {
2726                         err = nft_netdev_register_hooks(ctx->net, &hook.list);
2727                         if (err < 0)
2728                                 goto err_hooks;
2729                 }
2730         }
2731 
2732         unregister = true;
2733 
2734         if (nla[NFTA_CHAIN_COUNTERS]) {
2735                 if (!nft_is_base_chain(chain)) {
2736                         err = -EOPNOTSUPP;
2737                         goto err_hooks;
2738                 }
2739 
2740                 stats = nft_stats_alloc(nla[NFTA_CHAIN_COUNTERS]);
2741                 if (IS_ERR(stats)) {
2742                         err = PTR_ERR(stats);
2743                         goto err_hooks;
2744                 }
2745         }
2746 
2747         err = -ENOMEM;
2748         trans = nft_trans_alloc_chain(ctx, NFT_MSG_NEWCHAIN);
2749         if (trans == NULL)
2750                 goto err_trans;
2751 
2752         nft_trans_chain_stats(trans) = stats;
2753         nft_trans_chain_update(trans) = true;
2754 
2755         if (nla[NFTA_CHAIN_POLICY])
2756                 nft_trans_chain_policy(trans) = policy;
2757         else
2758                 nft_trans_chain_policy(trans) = -1;
2759 
2760         if (nla[NFTA_CHAIN_HANDLE] &&
2761             nla[NFTA_CHAIN_NAME]) {
2762                 struct nftables_pernet *nft_net = nft_pernet(ctx->net);
2763                 struct nft_trans *tmp;
2764                 char *name;
2765 
2766                 err = -ENOMEM;
2767                 name = nla_strdup(nla[NFTA_CHAIN_NAME], GFP_KERNEL_ACCOUNT);
2768                 if (!name)
2769                         goto err_trans;
2770 
2771                 err = -EEXIST;
2772                 list_for_each_entry(tmp, &nft_net->commit_list, list) {
2773                         if (tmp->msg_type == NFT_MSG_NEWCHAIN &&
2774                             tmp->table == table &&
2775                             nft_trans_chain_update(tmp) &&
2776                             nft_trans_chain_name(tmp) &&
2777                             strcmp(name, nft_trans_chain_name(tmp)) == 0) {
2778                                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_NAME]);
2779                                 kfree(name);
2780                                 goto err_trans;
2781                         }
2782                 }
2783 
2784                 nft_trans_chain_name(trans) = name;
2785         }
2786 
2787         nft_trans_basechain(trans) = basechain;
2788         INIT_LIST_HEAD(&nft_trans_chain_hooks(trans));
2789         list_splice(&hook.list, &nft_trans_chain_hooks(trans));
2790         if (nla[NFTA_CHAIN_HOOK])
2791                 module_put(hook.type->owner);
2792 
2793         nft_trans_commit_list_add_tail(ctx->net, trans);
2794 
2795         return 0;
2796 
2797 err_trans:
2798         free_percpu(stats);
2799         kfree(trans);
2800 err_hooks:
2801         if (nla[NFTA_CHAIN_HOOK]) {
2802                 list_for_each_entry_safe(h, next, &hook.list, list) {
2803                         if (unregister)
2804                                 nf_unregister_net_hook(ctx->net, &h->ops);
2805                         list_del(&h->list);
2806                         kfree_rcu(h, rcu);
2807                 }
2808                 module_put(hook.type->owner);
2809         }
2810 
2811         return err;
2812 }
2813 
2814 static struct nft_chain *nft_chain_lookup_byid(const struct net *net,
2815                                                const struct nft_table *table,
2816                                                const struct nlattr *nla, u8 genmask)
2817 {
2818         struct nftables_pernet *nft_net = nft_pernet(net);
2819         u32 id = ntohl(nla_get_be32(nla));
2820         struct nft_trans *trans;
2821 
2822         list_for_each_entry(trans, &nft_net->commit_list, list) {
2823                 if (trans->msg_type == NFT_MSG_NEWCHAIN &&
2824                     nft_trans_chain(trans)->table == table &&
2825                     id == nft_trans_chain_id(trans) &&
2826                     nft_active_genmask(nft_trans_chain(trans), genmask))
2827                         return nft_trans_chain(trans);
2828         }
2829         return ERR_PTR(-ENOENT);
2830 }
2831 
2832 static int nf_tables_newchain(struct sk_buff *skb, const struct nfnl_info *info,
2833                               const struct nlattr * const nla[])
2834 {
2835         struct nftables_pernet *nft_net = nft_pernet(info->net);
2836         struct netlink_ext_ack *extack = info->extack;
2837         u8 genmask = nft_genmask_next(info->net);
2838         u8 family = info->nfmsg->nfgen_family;
2839         struct nft_chain *chain = NULL;
2840         struct net *net = info->net;
2841         const struct nlattr *attr;
2842         struct nft_table *table;
2843         u8 policy = NF_ACCEPT;
2844         struct nft_ctx ctx;
2845         u64 handle = 0;
2846         u32 flags = 0;
2847 
2848         lockdep_assert_held(&nft_net->commit_mutex);
2849 
2850         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
2851                                  NETLINK_CB(skb).portid);
2852         if (IS_ERR(table)) {
2853                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
2854                 return PTR_ERR(table);
2855         }
2856 
2857         chain = NULL;
2858         attr = nla[NFTA_CHAIN_NAME];
2859 
2860         if (nla[NFTA_CHAIN_HANDLE]) {
2861                 handle = be64_to_cpu(nla_get_be64(nla[NFTA_CHAIN_HANDLE]));
2862                 chain = nft_chain_lookup_byhandle(table, handle, genmask);
2863                 if (IS_ERR(chain)) {
2864                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_HANDLE]);
2865                         return PTR_ERR(chain);
2866                 }
2867                 attr = nla[NFTA_CHAIN_HANDLE];
2868         } else if (nla[NFTA_CHAIN_NAME]) {
2869                 chain = nft_chain_lookup(net, table, attr, genmask);
2870                 if (IS_ERR(chain)) {
2871                         if (PTR_ERR(chain) != -ENOENT) {
2872                                 NL_SET_BAD_ATTR(extack, attr);
2873                                 return PTR_ERR(chain);
2874                         }
2875                         chain = NULL;
2876                 }
2877         } else if (!nla[NFTA_CHAIN_ID]) {
2878                 return -EINVAL;
2879         }
2880 
2881         if (nla[NFTA_CHAIN_POLICY]) {
2882                 if (chain != NULL &&
2883                     !nft_is_base_chain(chain)) {
2884                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2885                         return -EOPNOTSUPP;
2886                 }
2887 
2888                 if (chain == NULL &&
2889                     nla[NFTA_CHAIN_HOOK] == NULL) {
2890                         NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_POLICY]);
2891                         return -EOPNOTSUPP;
2892                 }
2893 
2894                 policy = ntohl(nla_get_be32(nla[NFTA_CHAIN_POLICY]));
2895                 switch (policy) {
2896                 case NF_DROP:
2897                 case NF_ACCEPT:
2898                         break;
2899                 default:
2900                         return -EINVAL;
2901                 }
2902         }
2903 
2904         if (nla[NFTA_CHAIN_FLAGS])
2905                 flags = ntohl(nla_get_be32(nla[NFTA_CHAIN_FLAGS]));
2906         else if (chain)
2907                 flags = chain->flags;
2908 
2909         if (flags & ~NFT_CHAIN_FLAGS)
2910                 return -EOPNOTSUPP;
2911 
2912         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
2913 
2914         if (chain != NULL) {
2915                 if (chain->flags & NFT_CHAIN_BINDING)
2916                         return -EINVAL;
2917 
2918                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
2919                         NL_SET_BAD_ATTR(extack, attr);
2920                         return -EEXIST;
2921                 }
2922                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
2923                         return -EOPNOTSUPP;
2924 
2925                 flags |= chain->flags & NFT_CHAIN_BASE;
2926                 return nf_tables_updchain(&ctx, genmask, policy, flags, attr,
2927                                           extack);
2928         }
2929 
2930         return nf_tables_addchain(&ctx, family, genmask, policy, flags, extack);
2931 }
2932 
2933 static int nft_delchain_hook(struct nft_ctx *ctx,
2934                              struct nft_base_chain *basechain,
2935                              struct netlink_ext_ack *extack)
2936 {
2937         const struct nft_chain *chain = &basechain->chain;
2938         const struct nlattr * const *nla = ctx->nla;
2939         struct nft_chain_hook chain_hook = {};
2940         struct nft_hook *this, *hook;
2941         LIST_HEAD(chain_del_list);
2942         struct nft_trans *trans;
2943         int err;
2944 
2945         if (ctx->table->flags & __NFT_TABLE_F_UPDATE)
2946                 return -EOPNOTSUPP;
2947 
2948         err = nft_chain_parse_hook(ctx->net, basechain, nla, &chain_hook,
2949                                    ctx->family, chain->flags, extack);
2950         if (err < 0)
2951                 return err;
2952 
2953         list_for_each_entry(this, &chain_hook.list, list) {
2954                 hook = nft_hook_list_find(&basechain->hook_list, this);
2955                 if (!hook) {
2956                         err = -ENOENT;
2957                         goto err_chain_del_hook;
2958                 }
2959                 list_move(&hook->list, &chain_del_list);
2960         }
2961 
2962         trans = nft_trans_alloc_chain(ctx, NFT_MSG_DELCHAIN);
2963         if (!trans) {
2964                 err = -ENOMEM;
2965                 goto err_chain_del_hook;
2966         }
2967 
2968         nft_trans_basechain(trans) = basechain;
2969         nft_trans_chain_update(trans) = true;
2970         INIT_LIST_HEAD(&nft_trans_chain_hooks(trans));
2971         list_splice(&chain_del_list, &nft_trans_chain_hooks(trans));
2972         nft_chain_release_hook(&chain_hook);
2973 
2974         nft_trans_commit_list_add_tail(ctx->net, trans);
2975 
2976         return 0;
2977 
2978 err_chain_del_hook:
2979         list_splice(&chain_del_list, &basechain->hook_list);
2980         nft_chain_release_hook(&chain_hook);
2981 
2982         return err;
2983 }
2984 
2985 static int nf_tables_delchain(struct sk_buff *skb, const struct nfnl_info *info,
2986                               const struct nlattr * const nla[])
2987 {
2988         struct netlink_ext_ack *extack = info->extack;
2989         u8 genmask = nft_genmask_next(info->net);
2990         u8 family = info->nfmsg->nfgen_family;
2991         struct net *net = info->net;
2992         const struct nlattr *attr;
2993         struct nft_table *table;
2994         struct nft_chain *chain;
2995         struct nft_rule *rule;
2996         struct nft_ctx ctx;
2997         u64 handle;
2998         u32 use;
2999         int err;
3000 
3001         table = nft_table_lookup(net, nla[NFTA_CHAIN_TABLE], family, genmask,
3002                                  NETLINK_CB(skb).portid);
3003         if (IS_ERR(table)) {
3004                 NL_SET_BAD_ATTR(extack, nla[NFTA_CHAIN_TABLE]);
3005                 return PTR_ERR(table);
3006         }
3007 
3008         if (nla[NFTA_CHAIN_HANDLE]) {
3009                 attr = nla[NFTA_CHAIN_HANDLE];
3010                 handle = be64_to_cpu(nla_get_be64(attr));
3011                 chain = nft_chain_lookup_byhandle(table, handle, genmask);
3012         } else {
3013                 attr = nla[NFTA_CHAIN_NAME];
3014                 chain = nft_chain_lookup(net, table, attr, genmask);
3015         }
3016         if (IS_ERR(chain)) {
3017                 if (PTR_ERR(chain) == -ENOENT &&
3018                     NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYCHAIN)
3019                         return 0;
3020 
3021                 NL_SET_BAD_ATTR(extack, attr);
3022                 return PTR_ERR(chain);
3023         }
3024 
3025         if (nft_chain_binding(chain))
3026                 return -EOPNOTSUPP;
3027 
3028         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
3029 
3030         if (nla[NFTA_CHAIN_HOOK]) {
3031                 if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYCHAIN ||
3032                     chain->flags & NFT_CHAIN_HW_OFFLOAD)
3033                         return -EOPNOTSUPP;
3034 
3035                 if (nft_is_base_chain(chain)) {
3036                         struct nft_base_chain *basechain = nft_base_chain(chain);
3037 
3038                         if (nft_base_chain_netdev(table->family, basechain->ops.hooknum))
3039                                 return nft_delchain_hook(&ctx, basechain, extack);
3040                 }
3041         }
3042 
3043         if (info->nlh->nlmsg_flags & NLM_F_NONREC &&
3044             chain->use > 0)
3045                 return -EBUSY;
3046 
3047         use = chain->use;
3048         list_for_each_entry(rule, &chain->rules, list) {
3049                 if (!nft_is_active_next(net, rule))
3050                         continue;
3051                 use--;
3052 
3053                 err = nft_delrule(&ctx, rule);
3054                 if (err < 0)
3055                         return err;
3056         }
3057 
3058         /* There are rules and elements that are still holding references to us,
3059          * we cannot do a recursive removal in this case.
3060          */
3061         if (use > 0) {
3062                 NL_SET_BAD_ATTR(extack, attr);
3063                 return -EBUSY;
3064         }
3065 
3066         return nft_delchain(&ctx);
3067 }
3068 
3069 /*
3070  * Expressions
3071  */
3072 
3073 /**
3074  *      nft_register_expr - register nf_tables expr type
3075  *      @type: expr type
3076  *
3077  *      Registers the expr type for use with nf_tables. Returns zero on
3078  *      success or a negative errno code otherwise.
3079  */
3080 int nft_register_expr(struct nft_expr_type *type)
3081 {
3082         if (WARN_ON_ONCE(type->maxattr > NFT_EXPR_MAXATTR))
3083                 return -ENOMEM;
3084 
3085         nfnl_lock(NFNL_SUBSYS_NFTABLES);
3086         if (type->family == NFPROTO_UNSPEC)
3087                 list_add_tail_rcu(&type->list, &nf_tables_expressions);
3088         else
3089                 list_add_rcu(&type->list, &nf_tables_expressions);
3090         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
3091         return 0;
3092 }
3093 EXPORT_SYMBOL_GPL(nft_register_expr);
3094 
3095 /**
3096  *      nft_unregister_expr - unregister nf_tables expr type
3097  *      @type: expr type
3098  *
3099  *      Unregisters the expr typefor use with nf_tables.
3100  */
3101 void nft_unregister_expr(struct nft_expr_type *type)
3102 {
3103         nfnl_lock(NFNL_SUBSYS_NFTABLES);
3104         list_del_rcu(&type->list);
3105         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
3106 }
3107 EXPORT_SYMBOL_GPL(nft_unregister_expr);
3108 
3109 static const struct nft_expr_type *__nft_expr_type_get(u8 family,
3110                                                        struct nlattr *nla)
3111 {
3112         const struct nft_expr_type *type, *candidate = NULL;
3113 
3114         list_for_each_entry_rcu(type, &nf_tables_expressions, list) {
3115                 if (!nla_strcmp(nla, type->name)) {
3116                         if (!type->family && !candidate)
3117                                 candidate = type;
3118                         else if (type->family == family)
3119                                 candidate = type;
3120                 }
3121         }
3122         return candidate;
3123 }
3124 
3125 #ifdef CONFIG_MODULES
3126 static int nft_expr_type_request_module(struct net *net, u8 family,
3127                                         struct nlattr *nla)
3128 {
3129         if (nft_request_module(net, "nft-expr-%u-%.*s", family,
3130                                nla_len(nla), (char *)nla_data(nla)) == -EAGAIN)
3131                 return -EAGAIN;
3132 
3133         return 0;
3134 }
3135 #endif
3136 
3137 static const struct nft_expr_type *nft_expr_type_get(struct net *net,
3138                                                      u8 family,
3139                                                      struct nlattr *nla)
3140 {
3141         const struct nft_expr_type *type;
3142 
3143         if (nla == NULL)
3144                 return ERR_PTR(-EINVAL);
3145 
3146         rcu_read_lock();
3147         type = __nft_expr_type_get(family, nla);
3148         if (type != NULL && try_module_get(type->owner)) {
3149                 rcu_read_unlock();
3150                 return type;
3151         }
3152         rcu_read_unlock();
3153 
3154         lockdep_nfnl_nft_mutex_not_held();
3155 #ifdef CONFIG_MODULES
3156         if (type == NULL) {
3157                 if (nft_expr_type_request_module(net, family, nla) == -EAGAIN)
3158                         return ERR_PTR(-EAGAIN);
3159 
3160                 if (nft_request_module(net, "nft-expr-%.*s",
3161                                        nla_len(nla),
3162                                        (char *)nla_data(nla)) == -EAGAIN)
3163                         return ERR_PTR(-EAGAIN);
3164         }
3165 #endif
3166         return ERR_PTR(-ENOENT);
3167 }
3168 
3169 static const struct nla_policy nft_expr_policy[NFTA_EXPR_MAX + 1] = {
3170         [NFTA_EXPR_NAME]        = { .type = NLA_STRING,
3171                                     .len = NFT_MODULE_AUTOLOAD_LIMIT },
3172         [NFTA_EXPR_DATA]        = { .type = NLA_NESTED },
3173 };
3174 
3175 static int nf_tables_fill_expr_info(struct sk_buff *skb,
3176                                     const struct nft_expr *expr, bool reset)
3177 {
3178         if (nla_put_string(skb, NFTA_EXPR_NAME, expr->ops->type->name))
3179                 goto nla_put_failure;
3180 
3181         if (expr->ops->dump) {
3182                 struct nlattr *data = nla_nest_start_noflag(skb,
3183                                                             NFTA_EXPR_DATA);
3184                 if (data == NULL)
3185                         goto nla_put_failure;
3186                 if (expr->ops->dump(skb, expr, reset) < 0)
3187                         goto nla_put_failure;
3188                 nla_nest_end(skb, data);
3189         }
3190 
3191         return skb->len;
3192 
3193 nla_put_failure:
3194         return -1;
3195 };
3196 
3197 int nft_expr_dump(struct sk_buff *skb, unsigned int attr,
3198                   const struct nft_expr *expr, bool reset)
3199 {
3200         struct nlattr *nest;
3201 
3202         nest = nla_nest_start_noflag(skb, attr);
3203         if (!nest)
3204                 goto nla_put_failure;
3205         if (nf_tables_fill_expr_info(skb, expr, reset) < 0)
3206                 goto nla_put_failure;
3207         nla_nest_end(skb, nest);
3208         return 0;
3209 
3210 nla_put_failure:
3211         return -1;
3212 }
3213 
3214 struct nft_expr_info {
3215         const struct nft_expr_ops       *ops;
3216         const struct nlattr             *attr;
3217         struct nlattr                   *tb[NFT_EXPR_MAXATTR + 1];
3218 };
3219 
3220 static int nf_tables_expr_parse(const struct nft_ctx *ctx,
3221                                 const struct nlattr *nla,
3222                                 struct nft_expr_info *info)
3223 {
3224         const struct nft_expr_type *type;
3225         const struct nft_expr_ops *ops;
3226         struct nlattr *tb[NFTA_EXPR_MAX + 1];
3227         int err;
3228 
3229         err = nla_parse_nested_deprecated(tb, NFTA_EXPR_MAX, nla,
3230                                           nft_expr_policy, NULL);
3231         if (err < 0)
3232                 return err;
3233 
3234         type = nft_expr_type_get(ctx->net, ctx->family, tb[NFTA_EXPR_NAME]);
3235         if (IS_ERR(type))
3236                 return PTR_ERR(type);
3237 
3238         if (tb[NFTA_EXPR_DATA]) {
3239                 err = nla_parse_nested_deprecated(info->tb, type->maxattr,
3240                                                   tb[NFTA_EXPR_DATA],
3241                                                   type->policy, NULL);
3242                 if (err < 0)
3243                         goto err1;
3244         } else
3245                 memset(info->tb, 0, sizeof(info->tb[0]) * (type->maxattr + 1));
3246 
3247         if (type->select_ops != NULL) {
3248                 ops = type->select_ops(ctx,
3249                                        (const struct nlattr * const *)info->tb);
3250                 if (IS_ERR(ops)) {
3251                         err = PTR_ERR(ops);
3252 #ifdef CONFIG_MODULES
3253                         if (err == -EAGAIN)
3254                                 if (nft_expr_type_request_module(ctx->net,
3255                                                                  ctx->family,
3256                                                                  tb[NFTA_EXPR_NAME]) != -EAGAIN)
3257                                         err = -ENOENT;
3258 #endif
3259                         goto err1;
3260                 }
3261         } else
3262                 ops = type->ops;
3263 
3264         info->attr = nla;
3265         info->ops = ops;
3266 
3267         return 0;
3268 
3269 err1:
3270         module_put(type->owner);
3271         return err;
3272 }
3273 
3274 int nft_expr_inner_parse(const struct nft_ctx *ctx, const struct nlattr *nla,
3275                          struct nft_expr_info *info)
3276 {
3277         struct nlattr *tb[NFTA_EXPR_MAX + 1];
3278         const struct nft_expr_type *type;
3279         int err;
3280 
3281         err = nla_parse_nested_deprecated(tb, NFTA_EXPR_MAX, nla,
3282                                           nft_expr_policy, NULL);
3283         if (err < 0)
3284                 return err;
3285 
3286         if (!tb[NFTA_EXPR_DATA] || !tb[NFTA_EXPR_NAME])
3287                 return -EINVAL;
3288 
3289         type = __nft_expr_type_get(ctx->family, tb[NFTA_EXPR_NAME]);
3290         if (!type)
3291                 return -ENOENT;
3292 
3293         if (!type->inner_ops)
3294                 return -EOPNOTSUPP;
3295 
3296         err = nla_parse_nested_deprecated(info->tb, type->maxattr,
3297                                           tb[NFTA_EXPR_DATA],
3298                                           type->policy, NULL);
3299         if (err < 0)
3300                 goto err_nla_parse;
3301 
3302         info->attr = nla;
3303         info->ops = type->inner_ops;
3304 
3305         return 0;
3306 
3307 err_nla_parse:
3308         return err;
3309 }
3310 
3311 static int nf_tables_newexpr(const struct nft_ctx *ctx,
3312                              const struct nft_expr_info *expr_info,
3313                              struct nft_expr *expr)
3314 {
3315         const struct nft_expr_ops *ops = expr_info->ops;
3316         int err;
3317 
3318         expr->ops = ops;
3319         if (ops->init) {
3320                 err = ops->init(ctx, expr, (const struct nlattr **)expr_info->tb);
3321                 if (err < 0)
3322                         goto err1;
3323         }
3324 
3325         return 0;
3326 err1:
3327         expr->ops = NULL;
3328         return err;
3329 }
3330 
3331 static void nf_tables_expr_destroy(const struct nft_ctx *ctx,
3332                                    struct nft_expr *expr)
3333 {
3334         const struct nft_expr_type *type = expr->ops->type;
3335 
3336         if (expr->ops->destroy)
3337                 expr->ops->destroy(ctx, expr);
3338         module_put(type->owner);
3339 }
3340 
3341 static struct nft_expr *nft_expr_init(const struct nft_ctx *ctx,
3342                                       const struct nlattr *nla)
3343 {
3344         struct nft_expr_info expr_info;
3345         struct nft_expr *expr;
3346         struct module *owner;
3347         int err;
3348 
3349         err = nf_tables_expr_parse(ctx, nla, &expr_info);
3350         if (err < 0)
3351                 goto err_expr_parse;
3352 
3353         err = -EOPNOTSUPP;
3354         if (!(expr_info.ops->type->flags & NFT_EXPR_STATEFUL))
3355                 goto err_expr_stateful;
3356 
3357         err = -ENOMEM;
3358         expr = kzalloc(expr_info.ops->size, GFP_KERNEL_ACCOUNT);
3359         if (expr == NULL)
3360                 goto err_expr_stateful;
3361 
3362         err = nf_tables_newexpr(ctx, &expr_info, expr);
3363         if (err < 0)
3364                 goto err_expr_new;
3365 
3366         return expr;
3367 err_expr_new:
3368         kfree(expr);
3369 err_expr_stateful:
3370         owner = expr_info.ops->type->owner;
3371         if (expr_info.ops->type->release_ops)
3372                 expr_info.ops->type->release_ops(expr_info.ops);
3373 
3374         module_put(owner);
3375 err_expr_parse:
3376         return ERR_PTR(err);
3377 }
3378 
3379 int nft_expr_clone(struct nft_expr *dst, struct nft_expr *src, gfp_t gfp)
3380 {
3381         int err;
3382 
3383         if (WARN_ON_ONCE(!src->ops->clone))
3384                 return -EINVAL;
3385 
3386         dst->ops = src->ops;
3387         err = src->ops->clone(dst, src, gfp);
3388         if (err < 0)
3389                 return err;
3390 
3391         __module_get(src->ops->type->owner);
3392 
3393         return 0;
3394 }
3395 
3396 void nft_expr_destroy(const struct nft_ctx *ctx, struct nft_expr *expr)
3397 {
3398         nf_tables_expr_destroy(ctx, expr);
3399         kfree(expr);
3400 }
3401 
3402 /*
3403  * Rules
3404  */
3405 
3406 static struct nft_rule *__nft_rule_lookup(const struct nft_chain *chain,
3407                                           u64 handle)
3408 {
3409         struct nft_rule *rule;
3410 
3411         // FIXME: this sucks
3412         list_for_each_entry_rcu(rule, &chain->rules, list) {
3413                 if (handle == rule->handle)
3414                         return rule;
3415         }
3416 
3417         return ERR_PTR(-ENOENT);
3418 }
3419 
3420 static struct nft_rule *nft_rule_lookup(const struct nft_chain *chain,
3421                                         const struct nlattr *nla)
3422 {
3423         if (nla == NULL)
3424                 return ERR_PTR(-EINVAL);
3425 
3426         return __nft_rule_lookup(chain, be64_to_cpu(nla_get_be64(nla)));
3427 }
3428 
3429 static const struct nla_policy nft_rule_policy[NFTA_RULE_MAX + 1] = {
3430         [NFTA_RULE_TABLE]       = { .type = NLA_STRING,
3431                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
3432         [NFTA_RULE_CHAIN]       = { .type = NLA_STRING,
3433                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
3434         [NFTA_RULE_HANDLE]      = { .type = NLA_U64 },
3435         [NFTA_RULE_EXPRESSIONS] = NLA_POLICY_NESTED_ARRAY(nft_expr_policy),
3436         [NFTA_RULE_COMPAT]      = { .type = NLA_NESTED },
3437         [NFTA_RULE_POSITION]    = { .type = NLA_U64 },
3438         [NFTA_RULE_USERDATA]    = { .type = NLA_BINARY,
3439                                     .len = NFT_USERDATA_MAXLEN },
3440         [NFTA_RULE_ID]          = { .type = NLA_U32 },
3441         [NFTA_RULE_POSITION_ID] = { .type = NLA_U32 },
3442         [NFTA_RULE_CHAIN_ID]    = { .type = NLA_U32 },
3443 };
3444 
3445 static int nf_tables_fill_rule_info(struct sk_buff *skb, struct net *net,
3446                                     u32 portid, u32 seq, int event,
3447                                     u32 flags, int family,
3448                                     const struct nft_table *table,
3449                                     const struct nft_chain *chain,
3450                                     const struct nft_rule *rule, u64 handle,
3451                                     bool reset)
3452 {
3453         struct nlmsghdr *nlh;
3454         const struct nft_expr *expr, *next;
3455         struct nlattr *list;
3456         u16 type = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
3457 
3458         nlh = nfnl_msg_put(skb, portid, seq, type, flags, family, NFNETLINK_V0,
3459                            nft_base_seq(net));
3460         if (!nlh)
3461                 goto nla_put_failure;
3462 
3463         if (nla_put_string(skb, NFTA_RULE_TABLE, table->name))
3464                 goto nla_put_failure;
3465         if (nla_put_string(skb, NFTA_RULE_CHAIN, chain->name))
3466                 goto nla_put_failure;
3467         if (nla_put_be64(skb, NFTA_RULE_HANDLE, cpu_to_be64(rule->handle),
3468                          NFTA_RULE_PAD))
3469                 goto nla_put_failure;
3470 
3471         if (event != NFT_MSG_DELRULE && handle) {
3472                 if (nla_put_be64(skb, NFTA_RULE_POSITION, cpu_to_be64(handle),
3473                                  NFTA_RULE_PAD))
3474                         goto nla_put_failure;
3475         }
3476 
3477         if (chain->flags & NFT_CHAIN_HW_OFFLOAD)
3478                 nft_flow_rule_stats(chain, rule);
3479 
3480         list = nla_nest_start_noflag(skb, NFTA_RULE_EXPRESSIONS);
3481         if (list == NULL)
3482                 goto nla_put_failure;
3483         nft_rule_for_each_expr(expr, next, rule) {
3484                 if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr, reset) < 0)
3485                         goto nla_put_failure;
3486         }
3487         nla_nest_end(skb, list);
3488 
3489         if (rule->udata) {
3490                 struct nft_userdata *udata = nft_userdata(rule);
3491                 if (nla_put(skb, NFTA_RULE_USERDATA, udata->len + 1,
3492                             udata->data) < 0)
3493                         goto nla_put_failure;
3494         }
3495 
3496         nlmsg_end(skb, nlh);
3497         return 0;
3498 
3499 nla_put_failure:
3500         nlmsg_trim(skb, nlh);
3501         return -1;
3502 }
3503 
3504 static void nf_tables_rule_notify(const struct nft_ctx *ctx,
3505                                   const struct nft_rule *rule, int event)
3506 {
3507         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
3508         const struct nft_rule *prule;
3509         struct sk_buff *skb;
3510         u64 handle = 0;
3511         u16 flags = 0;
3512         int err;
3513 
3514         if (!ctx->report &&
3515             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
3516                 return;
3517 
3518         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
3519         if (skb == NULL)
3520                 goto err;
3521 
3522         if (event == NFT_MSG_NEWRULE &&
3523             !list_is_first(&rule->list, &ctx->chain->rules) &&
3524             !list_is_last(&rule->list, &ctx->chain->rules)) {
3525                 prule = list_prev_entry(rule, list);
3526                 handle = prule->handle;
3527         }
3528         if (ctx->flags & (NLM_F_APPEND | NLM_F_REPLACE))
3529                 flags |= NLM_F_APPEND;
3530         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
3531                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
3532 
3533         err = nf_tables_fill_rule_info(skb, ctx->net, ctx->portid, ctx->seq,
3534                                        event, flags, ctx->family, ctx->table,
3535                                        ctx->chain, rule, handle, false);
3536         if (err < 0) {
3537                 kfree_skb(skb);
3538                 goto err;
3539         }
3540 
3541         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
3542         return;
3543 err:
3544         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
3545 }
3546 
3547 static void audit_log_rule_reset(const struct nft_table *table,
3548                                  unsigned int base_seq,
3549                                  unsigned int nentries)
3550 {
3551         char *buf = kasprintf(GFP_ATOMIC, "%s:%u",
3552                               table->name, base_seq);
3553 
3554         audit_log_nfcfg(buf, table->family, nentries,
3555                         AUDIT_NFT_OP_RULE_RESET, GFP_ATOMIC);
3556         kfree(buf);
3557 }
3558 
3559 struct nft_rule_dump_ctx {
3560         unsigned int s_idx;
3561         char *table;
3562         char *chain;
3563         bool reset;
3564 };
3565 
3566 static int __nf_tables_dump_rules(struct sk_buff *skb,
3567                                   unsigned int *idx,
3568                                   struct netlink_callback *cb,
3569                                   const struct nft_table *table,
3570                                   const struct nft_chain *chain)
3571 {
3572         struct nft_rule_dump_ctx *ctx = (void *)cb->ctx;
3573         struct net *net = sock_net(skb->sk);
3574         const struct nft_rule *rule, *prule;
3575         unsigned int entries = 0;
3576         int ret = 0;
3577         u64 handle;
3578 
3579         prule = NULL;
3580         list_for_each_entry_rcu(rule, &chain->rules, list) {
3581                 if (!nft_is_active(net, rule))
3582                         goto cont_skip;
3583                 if (*idx < ctx->s_idx)
3584                         goto cont;
3585                 if (prule)
3586                         handle = prule->handle;
3587                 else
3588                         handle = 0;
3589 
3590                 if (nf_tables_fill_rule_info(skb, net, NETLINK_CB(cb->skb).portid,
3591                                         cb->nlh->nlmsg_seq,
3592                                         NFT_MSG_NEWRULE,
3593                                         NLM_F_MULTI | NLM_F_APPEND,
3594                                         table->family,
3595                                         table, chain, rule, handle, ctx->reset) < 0) {
3596                         ret = 1;
3597                         break;
3598                 }
3599                 entries++;
3600                 nl_dump_check_consistent(cb, nlmsg_hdr(skb));
3601 cont:
3602                 prule = rule;
3603 cont_skip:
3604                 (*idx)++;
3605         }
3606 
3607         if (ctx->reset && entries)
3608                 audit_log_rule_reset(table, cb->seq, entries);
3609 
3610         return ret;
3611 }
3612 
3613 static int nf_tables_dump_rules(struct sk_buff *skb,
3614                                 struct netlink_callback *cb)
3615 {
3616         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3617         struct nft_rule_dump_ctx *ctx = (void *)cb->ctx;
3618         struct nft_table *table;
3619         const struct nft_chain *chain;
3620         unsigned int idx = 0;
3621         struct net *net = sock_net(skb->sk);
3622         int family = nfmsg->nfgen_family;
3623         struct nftables_pernet *nft_net;
3624 
3625         rcu_read_lock();
3626         nft_net = nft_pernet(net);
3627         cb->seq = READ_ONCE(nft_net->base_seq);
3628 
3629         list_for_each_entry_rcu(table, &nft_net->tables, list) {
3630                 if (family != NFPROTO_UNSPEC && family != table->family)
3631                         continue;
3632 
3633                 if (ctx->table && strcmp(ctx->table, table->name) != 0)
3634                         continue;
3635 
3636                 if (ctx->table && ctx->chain) {
3637                         struct rhlist_head *list, *tmp;
3638 
3639                         list = rhltable_lookup(&table->chains_ht, ctx->chain,
3640                                                nft_chain_ht_params);
3641                         if (!list)
3642                                 goto done;
3643 
3644                         rhl_for_each_entry_rcu(chain, tmp, list, rhlhead) {
3645                                 if (!nft_is_active(net, chain))
3646                                         continue;
3647                                 __nf_tables_dump_rules(skb, &idx,
3648                                                        cb, table, chain);
3649                                 break;
3650                         }
3651                         goto done;
3652                 }
3653 
3654                 list_for_each_entry_rcu(chain, &table->chains, list) {
3655                         if (__nf_tables_dump_rules(skb, &idx,
3656                                                    cb, table, chain))
3657                                 goto done;
3658                 }
3659 
3660                 if (ctx->table)
3661                         break;
3662         }
3663 done:
3664         rcu_read_unlock();
3665 
3666         ctx->s_idx = idx;
3667         return skb->len;
3668 }
3669 
3670 static int nf_tables_dumpreset_rules(struct sk_buff *skb,
3671                                      struct netlink_callback *cb)
3672 {
3673         struct nftables_pernet *nft_net = nft_pernet(sock_net(skb->sk));
3674         int ret;
3675 
3676         /* Mutex is held is to prevent that two concurrent dump-and-reset calls
3677          * do not underrun counters and quotas. The commit_mutex is used for
3678          * the lack a better lock, this is not transaction path.
3679          */
3680         mutex_lock(&nft_net->commit_mutex);
3681         ret = nf_tables_dump_rules(skb, cb);
3682         mutex_unlock(&nft_net->commit_mutex);
3683 
3684         return ret;
3685 }
3686 
3687 static int nf_tables_dump_rules_start(struct netlink_callback *cb)
3688 {
3689         struct nft_rule_dump_ctx *ctx = (void *)cb->ctx;
3690         const struct nlattr * const *nla = cb->data;
3691 
3692         BUILD_BUG_ON(sizeof(*ctx) > sizeof(cb->ctx));
3693 
3694         if (nla[NFTA_RULE_TABLE]) {
3695                 ctx->table = nla_strdup(nla[NFTA_RULE_TABLE], GFP_ATOMIC);
3696                 if (!ctx->table)
3697                         return -ENOMEM;
3698         }
3699         if (nla[NFTA_RULE_CHAIN]) {
3700                 ctx->chain = nla_strdup(nla[NFTA_RULE_CHAIN], GFP_ATOMIC);
3701                 if (!ctx->chain) {
3702                         kfree(ctx->table);
3703                         return -ENOMEM;
3704                 }
3705         }
3706         return 0;
3707 }
3708 
3709 static int nf_tables_dumpreset_rules_start(struct netlink_callback *cb)
3710 {
3711         struct nft_rule_dump_ctx *ctx = (void *)cb->ctx;
3712 
3713         ctx->reset = true;
3714 
3715         return nf_tables_dump_rules_start(cb);
3716 }
3717 
3718 static int nf_tables_dump_rules_done(struct netlink_callback *cb)
3719 {
3720         struct nft_rule_dump_ctx *ctx = (void *)cb->ctx;
3721 
3722         kfree(ctx->table);
3723         kfree(ctx->chain);
3724         return 0;
3725 }
3726 
3727 /* called with rcu_read_lock held */
3728 static struct sk_buff *
3729 nf_tables_getrule_single(u32 portid, const struct nfnl_info *info,
3730                          const struct nlattr * const nla[], bool reset)
3731 {
3732         struct netlink_ext_ack *extack = info->extack;
3733         u8 genmask = nft_genmask_cur(info->net);
3734         u8 family = info->nfmsg->nfgen_family;
3735         const struct nft_chain *chain;
3736         const struct nft_rule *rule;
3737         struct net *net = info->net;
3738         struct nft_table *table;
3739         struct sk_buff *skb2;
3740         int err;
3741 
3742         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask, 0);
3743         if (IS_ERR(table)) {
3744                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
3745                 return ERR_CAST(table);
3746         }
3747 
3748         chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN], genmask);
3749         if (IS_ERR(chain)) {
3750                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
3751                 return ERR_CAST(chain);
3752         }
3753 
3754         rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
3755         if (IS_ERR(rule)) {
3756                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
3757                 return ERR_CAST(rule);
3758         }
3759 
3760         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
3761         if (!skb2)
3762                 return ERR_PTR(-ENOMEM);
3763 
3764         err = nf_tables_fill_rule_info(skb2, net, portid,
3765                                        info->nlh->nlmsg_seq, NFT_MSG_NEWRULE, 0,
3766                                        family, table, chain, rule, 0, reset);
3767         if (err < 0) {
3768                 kfree_skb(skb2);
3769                 return ERR_PTR(err);
3770         }
3771 
3772         return skb2;
3773 }
3774 
3775 static int nf_tables_getrule(struct sk_buff *skb, const struct nfnl_info *info,
3776                              const struct nlattr * const nla[])
3777 {
3778         u32 portid = NETLINK_CB(skb).portid;
3779         struct net *net = info->net;
3780         struct sk_buff *skb2;
3781 
3782         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3783                 struct netlink_dump_control c = {
3784                         .start= nf_tables_dump_rules_start,
3785                         .dump = nf_tables_dump_rules,
3786                         .done = nf_tables_dump_rules_done,
3787                         .module = THIS_MODULE,
3788                         .data = (void *)nla,
3789                 };
3790 
3791                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
3792         }
3793 
3794         skb2 = nf_tables_getrule_single(portid, info, nla, false);
3795         if (IS_ERR(skb2))
3796                 return PTR_ERR(skb2);
3797 
3798         return nfnetlink_unicast(skb2, net, portid);
3799 }
3800 
3801 static int nf_tables_getrule_reset(struct sk_buff *skb,
3802                                    const struct nfnl_info *info,
3803                                    const struct nlattr * const nla[])
3804 {
3805         struct nftables_pernet *nft_net = nft_pernet(info->net);
3806         u32 portid = NETLINK_CB(skb).portid;
3807         struct net *net = info->net;
3808         struct sk_buff *skb2;
3809         char *buf;
3810 
3811         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
3812                 struct netlink_dump_control c = {
3813                         .start= nf_tables_dumpreset_rules_start,
3814                         .dump = nf_tables_dumpreset_rules,
3815                         .done = nf_tables_dump_rules_done,
3816                         .module = THIS_MODULE,
3817                         .data = (void *)nla,
3818                 };
3819 
3820                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
3821         }
3822 
3823         if (!try_module_get(THIS_MODULE))
3824                 return -EINVAL;
3825         rcu_read_unlock();
3826         mutex_lock(&nft_net->commit_mutex);
3827         skb2 = nf_tables_getrule_single(portid, info, nla, true);
3828         mutex_unlock(&nft_net->commit_mutex);
3829         rcu_read_lock();
3830         module_put(THIS_MODULE);
3831 
3832         if (IS_ERR(skb2))
3833                 return PTR_ERR(skb2);
3834 
3835         buf = kasprintf(GFP_ATOMIC, "%.*s:%u",
3836                         nla_len(nla[NFTA_RULE_TABLE]),
3837                         (char *)nla_data(nla[NFTA_RULE_TABLE]),
3838                         nft_net->base_seq);
3839         audit_log_nfcfg(buf, info->nfmsg->nfgen_family, 1,
3840                         AUDIT_NFT_OP_RULE_RESET, GFP_ATOMIC);
3841         kfree(buf);
3842 
3843         return nfnetlink_unicast(skb2, net, portid);
3844 }
3845 
3846 void nf_tables_rule_destroy(const struct nft_ctx *ctx, struct nft_rule *rule)
3847 {
3848         struct nft_expr *expr, *next;
3849 
3850         /*
3851          * Careful: some expressions might not be initialized in case this
3852          * is called on error from nf_tables_newrule().
3853          */
3854         expr = nft_expr_first(rule);
3855         while (nft_expr_more(rule, expr)) {
3856                 next = nft_expr_next(expr);
3857                 nf_tables_expr_destroy(ctx, expr);
3858                 expr = next;
3859         }
3860         kfree(rule);
3861 }
3862 
3863 static void nf_tables_rule_release(const struct nft_ctx *ctx, struct nft_rule *rule)
3864 {
3865         nft_rule_expr_deactivate(ctx, rule, NFT_TRANS_RELEASE);
3866         nf_tables_rule_destroy(ctx, rule);
3867 }
3868 
3869 /** nft_chain_validate - loop detection and hook validation
3870  *
3871  * @ctx: context containing call depth and base chain
3872  * @chain: chain to validate
3873  *
3874  * Walk through the rules of the given chain and chase all jumps/gotos
3875  * and set lookups until either the jump limit is hit or all reachable
3876  * chains have been validated.
3877  */
3878 int nft_chain_validate(const struct nft_ctx *ctx, const struct nft_chain *chain)
3879 {
3880         struct nft_expr *expr, *last;
3881         const struct nft_data *data;
3882         struct nft_rule *rule;
3883         int err;
3884 
3885         if (ctx->level == NFT_JUMP_STACK_SIZE)
3886                 return -EMLINK;
3887 
3888         list_for_each_entry(rule, &chain->rules, list) {
3889                 if (fatal_signal_pending(current))
3890                         return -EINTR;
3891 
3892                 if (!nft_is_active_next(ctx->net, rule))
3893                         continue;
3894 
3895                 nft_rule_for_each_expr(expr, last, rule) {
3896                         if (!expr->ops->validate)
3897                                 continue;
3898 
3899                         /* This may call nft_chain_validate() recursively,
3900                          * callers that do so must increment ctx->level.
3901                          */
3902                         err = expr->ops->validate(ctx, expr, &data);
3903                         if (err < 0)
3904                                 return err;
3905                 }
3906         }
3907 
3908         return 0;
3909 }
3910 EXPORT_SYMBOL_GPL(nft_chain_validate);
3911 
3912 static int nft_table_validate(struct net *net, const struct nft_table *table)
3913 {
3914         struct nft_chain *chain;
3915         struct nft_ctx ctx = {
3916                 .net    = net,
3917                 .family = table->family,
3918         };
3919         int err;
3920 
3921         list_for_each_entry(chain, &table->chains, list) {
3922                 if (!nft_is_base_chain(chain))
3923                         continue;
3924 
3925                 ctx.chain = chain;
3926                 err = nft_chain_validate(&ctx, chain);
3927                 if (err < 0)
3928                         return err;
3929 
3930                 cond_resched();
3931         }
3932 
3933         return 0;
3934 }
3935 
3936 int nft_setelem_validate(const struct nft_ctx *ctx, struct nft_set *set,
3937                          const struct nft_set_iter *iter,
3938                          struct nft_elem_priv *elem_priv)
3939 {
3940         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
3941         struct nft_ctx *pctx = (struct nft_ctx *)ctx;
3942         const struct nft_data *data;
3943         int err;
3944 
3945         if (!nft_set_elem_active(ext, iter->genmask))
3946                 return 0;
3947 
3948         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
3949             *nft_set_ext_flags(ext) & NFT_SET_ELEM_INTERVAL_END)
3950                 return 0;
3951 
3952         data = nft_set_ext_data(ext);
3953         switch (data->verdict.code) {
3954         case NFT_JUMP:
3955         case NFT_GOTO:
3956                 pctx->level++;
3957                 err = nft_chain_validate(ctx, data->verdict.chain);
3958                 if (err < 0)
3959                         return err;
3960                 pctx->level--;
3961                 break;
3962         default:
3963                 break;
3964         }
3965 
3966         return 0;
3967 }
3968 
3969 int nft_set_catchall_validate(const struct nft_ctx *ctx, struct nft_set *set)
3970 {
3971         struct nft_set_iter dummy_iter = {
3972                 .genmask        = nft_genmask_next(ctx->net),
3973         };
3974         struct nft_set_elem_catchall *catchall;
3975 
3976         struct nft_set_ext *ext;
3977         int ret = 0;
3978 
3979         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
3980                 ext = nft_set_elem_ext(set, catchall->elem);
3981                 if (!nft_set_elem_active(ext, dummy_iter.genmask))
3982                         continue;
3983 
3984                 ret = nft_setelem_validate(ctx, set, &dummy_iter, catchall->elem);
3985                 if (ret < 0)
3986                         return ret;
3987         }
3988 
3989         return ret;
3990 }
3991 
3992 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
3993                                              const struct nft_chain *chain,
3994                                              const struct nlattr *nla);
3995 
3996 #define NFT_RULE_MAXEXPRS       128
3997 
3998 static int nf_tables_newrule(struct sk_buff *skb, const struct nfnl_info *info,
3999                              const struct nlattr * const nla[])
4000 {
4001         struct nftables_pernet *nft_net = nft_pernet(info->net);
4002         struct netlink_ext_ack *extack = info->extack;
4003         unsigned int size, i, n, ulen = 0, usize = 0;
4004         u8 genmask = nft_genmask_next(info->net);
4005         struct nft_rule *rule, *old_rule = NULL;
4006         struct nft_expr_info *expr_info = NULL;
4007         u8 family = info->nfmsg->nfgen_family;
4008         struct nft_flow_rule *flow = NULL;
4009         struct net *net = info->net;
4010         struct nft_userdata *udata;
4011         struct nft_table *table;
4012         struct nft_chain *chain;
4013         struct nft_trans *trans;
4014         u64 handle, pos_handle;
4015         struct nft_expr *expr;
4016         struct nft_ctx ctx;
4017         struct nlattr *tmp;
4018         int err, rem;
4019 
4020         lockdep_assert_held(&nft_net->commit_mutex);
4021 
4022         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
4023                                  NETLINK_CB(skb).portid);
4024         if (IS_ERR(table)) {
4025                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
4026                 return PTR_ERR(table);
4027         }
4028 
4029         if (nla[NFTA_RULE_CHAIN]) {
4030                 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
4031                                          genmask);
4032                 if (IS_ERR(chain)) {
4033                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
4034                         return PTR_ERR(chain);
4035                 }
4036 
4037         } else if (nla[NFTA_RULE_CHAIN_ID]) {
4038                 chain = nft_chain_lookup_byid(net, table, nla[NFTA_RULE_CHAIN_ID],
4039                                               genmask);
4040                 if (IS_ERR(chain)) {
4041                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN_ID]);
4042                         return PTR_ERR(chain);
4043                 }
4044         } else {
4045                 return -EINVAL;
4046         }
4047 
4048         if (nft_chain_is_bound(chain))
4049                 return -EOPNOTSUPP;
4050 
4051         if (nla[NFTA_RULE_HANDLE]) {
4052                 handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_HANDLE]));
4053                 rule = __nft_rule_lookup(chain, handle);
4054                 if (IS_ERR(rule)) {
4055                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
4056                         return PTR_ERR(rule);
4057                 }
4058 
4059                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
4060                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
4061                         return -EEXIST;
4062                 }
4063                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
4064                         old_rule = rule;
4065                 else
4066                         return -EOPNOTSUPP;
4067         } else {
4068                 if (!(info->nlh->nlmsg_flags & NLM_F_CREATE) ||
4069                     info->nlh->nlmsg_flags & NLM_F_REPLACE)
4070                         return -EINVAL;
4071                 handle = nf_tables_alloc_handle(table);
4072 
4073                 if (nla[NFTA_RULE_POSITION]) {
4074                         pos_handle = be64_to_cpu(nla_get_be64(nla[NFTA_RULE_POSITION]));
4075                         old_rule = __nft_rule_lookup(chain, pos_handle);
4076                         if (IS_ERR(old_rule)) {
4077                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION]);
4078                                 return PTR_ERR(old_rule);
4079                         }
4080                 } else if (nla[NFTA_RULE_POSITION_ID]) {
4081                         old_rule = nft_rule_lookup_byid(net, chain, nla[NFTA_RULE_POSITION_ID]);
4082                         if (IS_ERR(old_rule)) {
4083                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_POSITION_ID]);
4084                                 return PTR_ERR(old_rule);
4085                         }
4086                 }
4087         }
4088 
4089         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
4090 
4091         n = 0;
4092         size = 0;
4093         if (nla[NFTA_RULE_EXPRESSIONS]) {
4094                 expr_info = kvmalloc_array(NFT_RULE_MAXEXPRS,
4095                                            sizeof(struct nft_expr_info),
4096                                            GFP_KERNEL);
4097                 if (!expr_info)
4098                         return -ENOMEM;
4099 
4100                 nla_for_each_nested(tmp, nla[NFTA_RULE_EXPRESSIONS], rem) {
4101                         err = -EINVAL;
4102                         if (nla_type(tmp) != NFTA_LIST_ELEM)
4103                                 goto err_release_expr;
4104                         if (n == NFT_RULE_MAXEXPRS)
4105                                 goto err_release_expr;
4106                         err = nf_tables_expr_parse(&ctx, tmp, &expr_info[n]);
4107                         if (err < 0) {
4108                                 NL_SET_BAD_ATTR(extack, tmp);
4109                                 goto err_release_expr;
4110                         }
4111                         size += expr_info[n].ops->size;
4112                         n++;
4113                 }
4114         }
4115         /* Check for overflow of dlen field */
4116         err = -EFBIG;
4117         if (size >= 1 << 12)
4118                 goto err_release_expr;
4119 
4120         if (nla[NFTA_RULE_USERDATA]) {
4121                 ulen = nla_len(nla[NFTA_RULE_USERDATA]);
4122                 if (ulen > 0)
4123                         usize = sizeof(struct nft_userdata) + ulen;
4124         }
4125 
4126         err = -ENOMEM;
4127         rule = kzalloc(sizeof(*rule) + size + usize, GFP_KERNEL_ACCOUNT);
4128         if (rule == NULL)
4129                 goto err_release_expr;
4130 
4131         nft_activate_next(net, rule);
4132 
4133         rule->handle = handle;
4134         rule->dlen   = size;
4135         rule->udata  = ulen ? 1 : 0;
4136 
4137         if (ulen) {
4138                 udata = nft_userdata(rule);
4139                 udata->len = ulen - 1;
4140                 nla_memcpy(udata->data, nla[NFTA_RULE_USERDATA], ulen);
4141         }
4142 
4143         expr = nft_expr_first(rule);
4144         for (i = 0; i < n; i++) {
4145                 err = nf_tables_newexpr(&ctx, &expr_info[i], expr);
4146                 if (err < 0) {
4147                         NL_SET_BAD_ATTR(extack, expr_info[i].attr);
4148                         goto err_release_rule;
4149                 }
4150 
4151                 if (expr_info[i].ops->validate)
4152                         nft_validate_state_update(table, NFT_VALIDATE_NEED);
4153 
4154                 expr_info[i].ops = NULL;
4155                 expr = nft_expr_next(expr);
4156         }
4157 
4158         if (chain->flags & NFT_CHAIN_HW_OFFLOAD) {
4159                 flow = nft_flow_rule_create(net, rule);
4160                 if (IS_ERR(flow)) {
4161                         err = PTR_ERR(flow);
4162                         goto err_release_rule;
4163                 }
4164         }
4165 
4166         if (!nft_use_inc(&chain->use)) {
4167                 err = -EMFILE;
4168                 goto err_release_rule;
4169         }
4170 
4171         if (info->nlh->nlmsg_flags & NLM_F_REPLACE) {
4172                 if (nft_chain_binding(chain)) {
4173                         err = -EOPNOTSUPP;
4174                         goto err_destroy_flow_rule;
4175                 }
4176 
4177                 err = nft_delrule(&ctx, old_rule);
4178                 if (err < 0)
4179                         goto err_destroy_flow_rule;
4180 
4181                 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
4182                 if (trans == NULL) {
4183                         err = -ENOMEM;
4184                         goto err_destroy_flow_rule;
4185                 }
4186                 list_add_tail_rcu(&rule->list, &old_rule->list);
4187         } else {
4188                 trans = nft_trans_rule_add(&ctx, NFT_MSG_NEWRULE, rule);
4189                 if (!trans) {
4190                         err = -ENOMEM;
4191                         goto err_destroy_flow_rule;
4192                 }
4193 
4194                 if (info->nlh->nlmsg_flags & NLM_F_APPEND) {
4195                         if (old_rule)
4196                                 list_add_rcu(&rule->list, &old_rule->list);
4197                         else
4198                                 list_add_tail_rcu(&rule->list, &chain->rules);
4199                  } else {
4200                         if (old_rule)
4201                                 list_add_tail_rcu(&rule->list, &old_rule->list);
4202                         else
4203                                 list_add_rcu(&rule->list, &chain->rules);
4204                 }
4205         }
4206         kvfree(expr_info);
4207 
4208         if (flow)
4209                 nft_trans_flow_rule(trans) = flow;
4210 
4211         if (table->validate_state == NFT_VALIDATE_DO)
4212                 return nft_table_validate(net, table);
4213 
4214         return 0;
4215 
4216 err_destroy_flow_rule:
4217         nft_use_dec_restore(&chain->use);
4218         if (flow)
4219                 nft_flow_rule_destroy(flow);
4220 err_release_rule:
4221         nft_rule_expr_deactivate(&ctx, rule, NFT_TRANS_PREPARE_ERROR);
4222         nf_tables_rule_destroy(&ctx, rule);
4223 err_release_expr:
4224         for (i = 0; i < n; i++) {
4225                 if (expr_info[i].ops) {
4226                         module_put(expr_info[i].ops->type->owner);
4227                         if (expr_info[i].ops->type->release_ops)
4228                                 expr_info[i].ops->type->release_ops(expr_info[i].ops);
4229                 }
4230         }
4231         kvfree(expr_info);
4232 
4233         return err;
4234 }
4235 
4236 static struct nft_rule *nft_rule_lookup_byid(const struct net *net,
4237                                              const struct nft_chain *chain,
4238                                              const struct nlattr *nla)
4239 {
4240         struct nftables_pernet *nft_net = nft_pernet(net);
4241         u32 id = ntohl(nla_get_be32(nla));
4242         struct nft_trans *trans;
4243 
4244         list_for_each_entry(trans, &nft_net->commit_list, list) {
4245                 if (trans->msg_type == NFT_MSG_NEWRULE &&
4246                     nft_trans_rule_chain(trans) == chain &&
4247                     id == nft_trans_rule_id(trans))
4248                         return nft_trans_rule(trans);
4249         }
4250         return ERR_PTR(-ENOENT);
4251 }
4252 
4253 static int nf_tables_delrule(struct sk_buff *skb, const struct nfnl_info *info,
4254                              const struct nlattr * const nla[])
4255 {
4256         struct netlink_ext_ack *extack = info->extack;
4257         u8 genmask = nft_genmask_next(info->net);
4258         u8 family = info->nfmsg->nfgen_family;
4259         struct nft_chain *chain = NULL;
4260         struct net *net = info->net;
4261         struct nft_table *table;
4262         struct nft_rule *rule;
4263         struct nft_ctx ctx;
4264         int err = 0;
4265 
4266         table = nft_table_lookup(net, nla[NFTA_RULE_TABLE], family, genmask,
4267                                  NETLINK_CB(skb).portid);
4268         if (IS_ERR(table)) {
4269                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_TABLE]);
4270                 return PTR_ERR(table);
4271         }
4272 
4273         if (nla[NFTA_RULE_CHAIN]) {
4274                 chain = nft_chain_lookup(net, table, nla[NFTA_RULE_CHAIN],
4275                                          genmask);
4276                 if (IS_ERR(chain)) {
4277                         if (PTR_ERR(chain) == -ENOENT &&
4278                             NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYRULE)
4279                                 return 0;
4280 
4281                         NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_CHAIN]);
4282                         return PTR_ERR(chain);
4283                 }
4284                 if (nft_chain_binding(chain))
4285                         return -EOPNOTSUPP;
4286         }
4287 
4288         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, chain, nla);
4289 
4290         if (chain) {
4291                 if (nla[NFTA_RULE_HANDLE]) {
4292                         rule = nft_rule_lookup(chain, nla[NFTA_RULE_HANDLE]);
4293                         if (IS_ERR(rule)) {
4294                                 if (PTR_ERR(rule) == -ENOENT &&
4295                                     NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYRULE)
4296                                         return 0;
4297 
4298                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_HANDLE]);
4299                                 return PTR_ERR(rule);
4300                         }
4301 
4302                         err = nft_delrule(&ctx, rule);
4303                 } else if (nla[NFTA_RULE_ID]) {
4304                         rule = nft_rule_lookup_byid(net, chain, nla[NFTA_RULE_ID]);
4305                         if (IS_ERR(rule)) {
4306                                 NL_SET_BAD_ATTR(extack, nla[NFTA_RULE_ID]);
4307                                 return PTR_ERR(rule);
4308                         }
4309 
4310                         err = nft_delrule(&ctx, rule);
4311                 } else {
4312                         err = nft_delrule_by_chain(&ctx);
4313                 }
4314         } else {
4315                 list_for_each_entry(chain, &table->chains, list) {
4316                         if (!nft_is_active_next(net, chain))
4317                                 continue;
4318                         if (nft_chain_binding(chain))
4319                                 continue;
4320 
4321                         ctx.chain = chain;
4322                         err = nft_delrule_by_chain(&ctx);
4323                         if (err < 0)
4324                                 break;
4325                 }
4326         }
4327 
4328         return err;
4329 }
4330 
4331 /*
4332  * Sets
4333  */
4334 static const struct nft_set_type *nft_set_types[] = {
4335         &nft_set_hash_fast_type,
4336         &nft_set_hash_type,
4337         &nft_set_rhash_type,
4338         &nft_set_bitmap_type,
4339         &nft_set_rbtree_type,
4340 #if defined(CONFIG_X86_64) && !defined(CONFIG_UML)
4341         &nft_set_pipapo_avx2_type,
4342 #endif
4343         &nft_set_pipapo_type,
4344 };
4345 
4346 #define NFT_SET_FEATURES        (NFT_SET_INTERVAL | NFT_SET_MAP | \
4347                                  NFT_SET_TIMEOUT | NFT_SET_OBJECT | \
4348                                  NFT_SET_EVAL)
4349 
4350 static bool nft_set_ops_candidate(const struct nft_set_type *type, u32 flags)
4351 {
4352         return (flags & type->features) == (flags & NFT_SET_FEATURES);
4353 }
4354 
4355 /*
4356  * Select a set implementation based on the data characteristics and the
4357  * given policy. The total memory use might not be known if no size is
4358  * given, in that case the amount of memory per element is used.
4359  */
4360 static const struct nft_set_ops *
4361 nft_select_set_ops(const struct nft_ctx *ctx, u32 flags,
4362                    const struct nft_set_desc *desc)
4363 {
4364         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
4365         const struct nft_set_ops *ops, *bops;
4366         struct nft_set_estimate est, best;
4367         const struct nft_set_type *type;
4368         int i;
4369 
4370         lockdep_assert_held(&nft_net->commit_mutex);
4371         lockdep_nfnl_nft_mutex_not_held();
4372 
4373         bops        = NULL;
4374         best.size   = ~0;
4375         best.lookup = ~0;
4376         best.space  = ~0;
4377 
4378         for (i = 0; i < ARRAY_SIZE(nft_set_types); i++) {
4379                 type = nft_set_types[i];
4380                 ops = &type->ops;
4381 
4382                 if (!nft_set_ops_candidate(type, flags))
4383                         continue;
4384                 if (!ops->estimate(desc, flags, &est))
4385                         continue;
4386 
4387                 switch (desc->policy) {
4388                 case NFT_SET_POL_PERFORMANCE:
4389                         if (est.lookup < best.lookup)
4390                                 break;
4391                         if (est.lookup == best.lookup &&
4392                             est.space < best.space)
4393                                 break;
4394                         continue;
4395                 case NFT_SET_POL_MEMORY:
4396                         if (!desc->size) {
4397                                 if (est.space < best.space)
4398                                         break;
4399                                 if (est.space == best.space &&
4400                                     est.lookup < best.lookup)
4401                                         break;
4402                         } else if (est.size < best.size || !bops) {
4403                                 break;
4404                         }
4405                         continue;
4406                 default:
4407                         break;
4408                 }
4409 
4410                 bops = ops;
4411                 best = est;
4412         }
4413 
4414         if (bops != NULL)
4415                 return bops;
4416 
4417         return ERR_PTR(-EOPNOTSUPP);
4418 }
4419 
4420 static const struct nla_policy nft_set_policy[NFTA_SET_MAX + 1] = {
4421         [NFTA_SET_TABLE]                = { .type = NLA_STRING,
4422                                             .len = NFT_TABLE_MAXNAMELEN - 1 },
4423         [NFTA_SET_NAME]                 = { .type = NLA_STRING,
4424                                             .len = NFT_SET_MAXNAMELEN - 1 },
4425         [NFTA_SET_FLAGS]                = { .type = NLA_U32 },
4426         [NFTA_SET_KEY_TYPE]             = { .type = NLA_U32 },
4427         [NFTA_SET_KEY_LEN]              = { .type = NLA_U32 },
4428         [NFTA_SET_DATA_TYPE]            = { .type = NLA_U32 },
4429         [NFTA_SET_DATA_LEN]             = { .type = NLA_U32 },
4430         [NFTA_SET_POLICY]               = { .type = NLA_U32 },
4431         [NFTA_SET_DESC]                 = { .type = NLA_NESTED },
4432         [NFTA_SET_ID]                   = { .type = NLA_U32 },
4433         [NFTA_SET_TIMEOUT]              = { .type = NLA_U64 },
4434         [NFTA_SET_GC_INTERVAL]          = { .type = NLA_U32 },
4435         [NFTA_SET_USERDATA]             = { .type = NLA_BINARY,
4436                                             .len  = NFT_USERDATA_MAXLEN },
4437         [NFTA_SET_OBJ_TYPE]             = { .type = NLA_U32 },
4438         [NFTA_SET_HANDLE]               = { .type = NLA_U64 },
4439         [NFTA_SET_EXPR]                 = { .type = NLA_NESTED },
4440         [NFTA_SET_EXPRESSIONS]          = NLA_POLICY_NESTED_ARRAY(nft_expr_policy),
4441 };
4442 
4443 static const struct nla_policy nft_concat_policy[NFTA_SET_FIELD_MAX + 1] = {
4444         [NFTA_SET_FIELD_LEN]    = { .type = NLA_U32 },
4445 };
4446 
4447 static const struct nla_policy nft_set_desc_policy[NFTA_SET_DESC_MAX + 1] = {
4448         [NFTA_SET_DESC_SIZE]            = { .type = NLA_U32 },
4449         [NFTA_SET_DESC_CONCAT]          = NLA_POLICY_NESTED_ARRAY(nft_concat_policy),
4450 };
4451 
4452 static struct nft_set *nft_set_lookup(const struct nft_table *table,
4453                                       const struct nlattr *nla, u8 genmask)
4454 {
4455         struct nft_set *set;
4456 
4457         if (nla == NULL)
4458                 return ERR_PTR(-EINVAL);
4459 
4460         list_for_each_entry_rcu(set, &table->sets, list) {
4461                 if (!nla_strcmp(nla, set->name) &&
4462                     nft_active_genmask(set, genmask))
4463                         return set;
4464         }
4465         return ERR_PTR(-ENOENT);
4466 }
4467 
4468 static struct nft_set *nft_set_lookup_byhandle(const struct nft_table *table,
4469                                                const struct nlattr *nla,
4470                                                u8 genmask)
4471 {
4472         struct nft_set *set;
4473 
4474         list_for_each_entry(set, &table->sets, list) {
4475                 if (be64_to_cpu(nla_get_be64(nla)) == set->handle &&
4476                     nft_active_genmask(set, genmask))
4477                         return set;
4478         }
4479         return ERR_PTR(-ENOENT);
4480 }
4481 
4482 static struct nft_set *nft_set_lookup_byid(const struct net *net,
4483                                            const struct nft_table *table,
4484                                            const struct nlattr *nla, u8 genmask)
4485 {
4486         struct nftables_pernet *nft_net = nft_pernet(net);
4487         u32 id = ntohl(nla_get_be32(nla));
4488         struct nft_trans *trans;
4489 
4490         list_for_each_entry(trans, &nft_net->commit_list, list) {
4491                 if (trans->msg_type == NFT_MSG_NEWSET) {
4492                         struct nft_set *set = nft_trans_set(trans);
4493 
4494                         if (id == nft_trans_set_id(trans) &&
4495                             set->table == table &&
4496                             nft_active_genmask(set, genmask))
4497                                 return set;
4498                 }
4499         }
4500         return ERR_PTR(-ENOENT);
4501 }
4502 
4503 struct nft_set *nft_set_lookup_global(const struct net *net,
4504                                       const struct nft_table *table,
4505                                       const struct nlattr *nla_set_name,
4506                                       const struct nlattr *nla_set_id,
4507                                       u8 genmask)
4508 {
4509         struct nft_set *set;
4510 
4511         set = nft_set_lookup(table, nla_set_name, genmask);
4512         if (IS_ERR(set)) {
4513                 if (!nla_set_id)
4514                         return set;
4515 
4516                 set = nft_set_lookup_byid(net, table, nla_set_id, genmask);
4517         }
4518         return set;
4519 }
4520 EXPORT_SYMBOL_GPL(nft_set_lookup_global);
4521 
4522 static int nf_tables_set_alloc_name(struct nft_ctx *ctx, struct nft_set *set,
4523                                     const char *name)
4524 {
4525         const struct nft_set *i;
4526         const char *p;
4527         unsigned long *inuse;
4528         unsigned int n = 0, min = 0;
4529 
4530         p = strchr(name, '%');
4531         if (p != NULL) {
4532                 if (p[1] != 'd' || strchr(p + 2, '%'))
4533                         return -EINVAL;
4534 
4535                 if (strnlen(name, NFT_SET_MAX_ANONLEN) >= NFT_SET_MAX_ANONLEN)
4536                         return -EINVAL;
4537 
4538                 inuse = (unsigned long *)get_zeroed_page(GFP_KERNEL);
4539                 if (inuse == NULL)
4540                         return -ENOMEM;
4541 cont:
4542                 list_for_each_entry(i, &ctx->table->sets, list) {
4543                         int tmp;
4544 
4545                         if (!nft_is_active_next(ctx->net, i))
4546                                 continue;
4547                         if (!sscanf(i->name, name, &tmp))
4548                                 continue;
4549                         if (tmp < min || tmp >= min + BITS_PER_BYTE * PAGE_SIZE)
4550                                 continue;
4551 
4552                         set_bit(tmp - min, inuse);
4553                 }
4554 
4555                 n = find_first_zero_bit(inuse, BITS_PER_BYTE * PAGE_SIZE);
4556                 if (n >= BITS_PER_BYTE * PAGE_SIZE) {
4557                         min += BITS_PER_BYTE * PAGE_SIZE;
4558                         memset(inuse, 0, PAGE_SIZE);
4559                         goto cont;
4560                 }
4561                 free_page((unsigned long)inuse);
4562         }
4563 
4564         set->name = kasprintf(GFP_KERNEL_ACCOUNT, name, min + n);
4565         if (!set->name)
4566                 return -ENOMEM;
4567 
4568         list_for_each_entry(i, &ctx->table->sets, list) {
4569                 if (!nft_is_active_next(ctx->net, i))
4570                         continue;
4571                 if (!strcmp(set->name, i->name)) {
4572                         kfree(set->name);
4573                         set->name = NULL;
4574                         return -ENFILE;
4575                 }
4576         }
4577         return 0;
4578 }
4579 
4580 int nf_msecs_to_jiffies64(const struct nlattr *nla, u64 *result)
4581 {
4582         u64 ms = be64_to_cpu(nla_get_be64(nla));
4583         u64 max = (u64)(~((u64)0));
4584 
4585         max = div_u64(max, NSEC_PER_MSEC);
4586         if (ms >= max)
4587                 return -ERANGE;
4588 
4589         ms *= NSEC_PER_MSEC;
4590         *result = nsecs_to_jiffies64(ms);
4591         return 0;
4592 }
4593 
4594 __be64 nf_jiffies64_to_msecs(u64 input)
4595 {
4596         return cpu_to_be64(jiffies64_to_msecs(input));
4597 }
4598 
4599 static int nf_tables_fill_set_concat(struct sk_buff *skb,
4600                                      const struct nft_set *set)
4601 {
4602         struct nlattr *concat, *field;
4603         int i;
4604 
4605         concat = nla_nest_start_noflag(skb, NFTA_SET_DESC_CONCAT);
4606         if (!concat)
4607                 return -ENOMEM;
4608 
4609         for (i = 0; i < set->field_count; i++) {
4610                 field = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
4611                 if (!field)
4612                         return -ENOMEM;
4613 
4614                 if (nla_put_be32(skb, NFTA_SET_FIELD_LEN,
4615                                  htonl(set->field_len[i])))
4616                         return -ENOMEM;
4617 
4618                 nla_nest_end(skb, field);
4619         }
4620 
4621         nla_nest_end(skb, concat);
4622 
4623         return 0;
4624 }
4625 
4626 static int nf_tables_fill_set(struct sk_buff *skb, const struct nft_ctx *ctx,
4627                               const struct nft_set *set, u16 event, u16 flags)
4628 {
4629         u64 timeout = READ_ONCE(set->timeout);
4630         u32 gc_int = READ_ONCE(set->gc_int);
4631         u32 portid = ctx->portid;
4632         struct nlmsghdr *nlh;
4633         struct nlattr *nest;
4634         u32 seq = ctx->seq;
4635         int i;
4636 
4637         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
4638         nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
4639                            NFNETLINK_V0, nft_base_seq(ctx->net));
4640         if (!nlh)
4641                 goto nla_put_failure;
4642 
4643         if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
4644                 goto nla_put_failure;
4645         if (nla_put_string(skb, NFTA_SET_NAME, set->name))
4646                 goto nla_put_failure;
4647         if (nla_put_be64(skb, NFTA_SET_HANDLE, cpu_to_be64(set->handle),
4648                          NFTA_SET_PAD))
4649                 goto nla_put_failure;
4650 
4651         if (event == NFT_MSG_DELSET) {
4652                 nlmsg_end(skb, nlh);
4653                 return 0;
4654         }
4655 
4656         if (set->flags != 0)
4657                 if (nla_put_be32(skb, NFTA_SET_FLAGS, htonl(set->flags)))
4658                         goto nla_put_failure;
4659 
4660         if (nla_put_be32(skb, NFTA_SET_KEY_TYPE, htonl(set->ktype)))
4661                 goto nla_put_failure;
4662         if (nla_put_be32(skb, NFTA_SET_KEY_LEN, htonl(set->klen)))
4663                 goto nla_put_failure;
4664         if (set->flags & NFT_SET_MAP) {
4665                 if (nla_put_be32(skb, NFTA_SET_DATA_TYPE, htonl(set->dtype)))
4666                         goto nla_put_failure;
4667                 if (nla_put_be32(skb, NFTA_SET_DATA_LEN, htonl(set->dlen)))
4668                         goto nla_put_failure;
4669         }
4670         if (set->flags & NFT_SET_OBJECT &&
4671             nla_put_be32(skb, NFTA_SET_OBJ_TYPE, htonl(set->objtype)))
4672                 goto nla_put_failure;
4673 
4674         if (timeout &&
4675             nla_put_be64(skb, NFTA_SET_TIMEOUT,
4676                          nf_jiffies64_to_msecs(timeout),
4677                          NFTA_SET_PAD))
4678                 goto nla_put_failure;
4679         if (gc_int &&
4680             nla_put_be32(skb, NFTA_SET_GC_INTERVAL, htonl(gc_int)))
4681                 goto nla_put_failure;
4682 
4683         if (set->policy != NFT_SET_POL_PERFORMANCE) {
4684                 if (nla_put_be32(skb, NFTA_SET_POLICY, htonl(set->policy)))
4685                         goto nla_put_failure;
4686         }
4687 
4688         if (set->udata &&
4689             nla_put(skb, NFTA_SET_USERDATA, set->udlen, set->udata))
4690                 goto nla_put_failure;
4691 
4692         nest = nla_nest_start_noflag(skb, NFTA_SET_DESC);
4693         if (!nest)
4694                 goto nla_put_failure;
4695         if (set->size &&
4696             nla_put_be32(skb, NFTA_SET_DESC_SIZE, htonl(set->size)))
4697                 goto nla_put_failure;
4698 
4699         if (set->field_count > 1 &&
4700             nf_tables_fill_set_concat(skb, set))
4701                 goto nla_put_failure;
4702 
4703         nla_nest_end(skb, nest);
4704 
4705         if (set->num_exprs == 1) {
4706                 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPR);
4707                 if (nf_tables_fill_expr_info(skb, set->exprs[0], false) < 0)
4708                         goto nla_put_failure;
4709 
4710                 nla_nest_end(skb, nest);
4711         } else if (set->num_exprs > 1) {
4712                 nest = nla_nest_start_noflag(skb, NFTA_SET_EXPRESSIONS);
4713                 if (nest == NULL)
4714                         goto nla_put_failure;
4715 
4716                 for (i = 0; i < set->num_exprs; i++) {
4717                         if (nft_expr_dump(skb, NFTA_LIST_ELEM,
4718                                           set->exprs[i], false) < 0)
4719                                 goto nla_put_failure;
4720                 }
4721                 nla_nest_end(skb, nest);
4722         }
4723 
4724         nlmsg_end(skb, nlh);
4725         return 0;
4726 
4727 nla_put_failure:
4728         nlmsg_trim(skb, nlh);
4729         return -1;
4730 }
4731 
4732 static void nf_tables_set_notify(const struct nft_ctx *ctx,
4733                                  const struct nft_set *set, int event,
4734                                  gfp_t gfp_flags)
4735 {
4736         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
4737         u32 portid = ctx->portid;
4738         struct sk_buff *skb;
4739         u16 flags = 0;
4740         int err;
4741 
4742         if (!ctx->report &&
4743             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
4744                 return;
4745 
4746         skb = nlmsg_new(NLMSG_GOODSIZE, gfp_flags);
4747         if (skb == NULL)
4748                 goto err;
4749 
4750         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
4751                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
4752 
4753         err = nf_tables_fill_set(skb, ctx, set, event, flags);
4754         if (err < 0) {
4755                 kfree_skb(skb);
4756                 goto err;
4757         }
4758 
4759         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
4760         return;
4761 err:
4762         nfnetlink_set_err(ctx->net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
4763 }
4764 
4765 static int nf_tables_dump_sets(struct sk_buff *skb, struct netlink_callback *cb)
4766 {
4767         const struct nft_set *set;
4768         unsigned int idx, s_idx = cb->args[0];
4769         struct nft_table *table, *cur_table = (struct nft_table *)cb->args[2];
4770         struct net *net = sock_net(skb->sk);
4771         struct nft_ctx *ctx = cb->data, ctx_set;
4772         struct nftables_pernet *nft_net;
4773 
4774         if (cb->args[1])
4775                 return skb->len;
4776 
4777         rcu_read_lock();
4778         nft_net = nft_pernet(net);
4779         cb->seq = READ_ONCE(nft_net->base_seq);
4780 
4781         list_for_each_entry_rcu(table, &nft_net->tables, list) {
4782                 if (ctx->family != NFPROTO_UNSPEC &&
4783                     ctx->family != table->family)
4784                         continue;
4785 
4786                 if (ctx->table && ctx->table != table)
4787                         continue;
4788 
4789                 if (cur_table) {
4790                         if (cur_table != table)
4791                                 continue;
4792 
4793                         cur_table = NULL;
4794                 }
4795                 idx = 0;
4796                 list_for_each_entry_rcu(set, &table->sets, list) {
4797                         if (idx < s_idx)
4798                                 goto cont;
4799                         if (!nft_is_active(net, set))
4800                                 goto cont;
4801 
4802                         ctx_set = *ctx;
4803                         ctx_set.table = table;
4804                         ctx_set.family = table->family;
4805 
4806                         if (nf_tables_fill_set(skb, &ctx_set, set,
4807                                                NFT_MSG_NEWSET,
4808                                                NLM_F_MULTI) < 0) {
4809                                 cb->args[0] = idx;
4810                                 cb->args[2] = (unsigned long) table;
4811                                 goto done;
4812                         }
4813                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
4814 cont:
4815                         idx++;
4816                 }
4817                 if (s_idx)
4818                         s_idx = 0;
4819         }
4820         cb->args[1] = 1;
4821 done:
4822         rcu_read_unlock();
4823         return skb->len;
4824 }
4825 
4826 static int nf_tables_dump_sets_start(struct netlink_callback *cb)
4827 {
4828         struct nft_ctx *ctx_dump = NULL;
4829 
4830         ctx_dump = kmemdup(cb->data, sizeof(*ctx_dump), GFP_ATOMIC);
4831         if (ctx_dump == NULL)
4832                 return -ENOMEM;
4833 
4834         cb->data = ctx_dump;
4835         return 0;
4836 }
4837 
4838 static int nf_tables_dump_sets_done(struct netlink_callback *cb)
4839 {
4840         kfree(cb->data);
4841         return 0;
4842 }
4843 
4844 /* called with rcu_read_lock held */
4845 static int nf_tables_getset(struct sk_buff *skb, const struct nfnl_info *info,
4846                             const struct nlattr * const nla[])
4847 {
4848         struct netlink_ext_ack *extack = info->extack;
4849         u8 genmask = nft_genmask_cur(info->net);
4850         u8 family = info->nfmsg->nfgen_family;
4851         struct nft_table *table = NULL;
4852         struct net *net = info->net;
4853         const struct nft_set *set;
4854         struct sk_buff *skb2;
4855         struct nft_ctx ctx;
4856         int err;
4857 
4858         if (nla[NFTA_SET_TABLE]) {
4859                 table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
4860                                          genmask, 0);
4861                 if (IS_ERR(table)) {
4862                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
4863                         return PTR_ERR(table);
4864                 }
4865         }
4866 
4867         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
4868 
4869         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
4870                 struct netlink_dump_control c = {
4871                         .start = nf_tables_dump_sets_start,
4872                         .dump = nf_tables_dump_sets,
4873                         .done = nf_tables_dump_sets_done,
4874                         .data = &ctx,
4875                         .module = THIS_MODULE,
4876                 };
4877 
4878                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
4879         }
4880 
4881         /* Only accept unspec with dump */
4882         if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
4883                 return -EAFNOSUPPORT;
4884         if (!nla[NFTA_SET_TABLE])
4885                 return -EINVAL;
4886 
4887         set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
4888         if (IS_ERR(set)) {
4889                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
4890                 return PTR_ERR(set);
4891         }
4892 
4893         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
4894         if (skb2 == NULL)
4895                 return -ENOMEM;
4896 
4897         err = nf_tables_fill_set(skb2, &ctx, set, NFT_MSG_NEWSET, 0);
4898         if (err < 0)
4899                 goto err_fill_set_info;
4900 
4901         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
4902 
4903 err_fill_set_info:
4904         kfree_skb(skb2);
4905         return err;
4906 }
4907 
4908 static int nft_set_desc_concat_parse(const struct nlattr *attr,
4909                                      struct nft_set_desc *desc)
4910 {
4911         struct nlattr *tb[NFTA_SET_FIELD_MAX + 1];
4912         u32 len;
4913         int err;
4914 
4915         if (desc->field_count >= ARRAY_SIZE(desc->field_len))
4916                 return -E2BIG;
4917 
4918         err = nla_parse_nested_deprecated(tb, NFTA_SET_FIELD_MAX, attr,
4919                                           nft_concat_policy, NULL);
4920         if (err < 0)
4921                 return err;
4922 
4923         if (!tb[NFTA_SET_FIELD_LEN])
4924                 return -EINVAL;
4925 
4926         len = ntohl(nla_get_be32(tb[NFTA_SET_FIELD_LEN]));
4927         if (!len || len > U8_MAX)
4928                 return -EINVAL;
4929 
4930         desc->field_len[desc->field_count++] = len;
4931 
4932         return 0;
4933 }
4934 
4935 static int nft_set_desc_concat(struct nft_set_desc *desc,
4936                                const struct nlattr *nla)
4937 {
4938         u32 num_regs = 0, key_num_regs = 0;
4939         struct nlattr *attr;
4940         int rem, err, i;
4941 
4942         nla_for_each_nested(attr, nla, rem) {
4943                 if (nla_type(attr) != NFTA_LIST_ELEM)
4944                         return -EINVAL;
4945 
4946                 err = nft_set_desc_concat_parse(attr, desc);
4947                 if (err < 0)
4948                         return err;
4949         }
4950 
4951         for (i = 0; i < desc->field_count; i++)
4952                 num_regs += DIV_ROUND_UP(desc->field_len[i], sizeof(u32));
4953 
4954         key_num_regs = DIV_ROUND_UP(desc->klen, sizeof(u32));
4955         if (key_num_regs != num_regs)
4956                 return -EINVAL;
4957 
4958         if (num_regs > NFT_REG32_COUNT)
4959                 return -E2BIG;
4960 
4961         return 0;
4962 }
4963 
4964 static int nf_tables_set_desc_parse(struct nft_set_desc *desc,
4965                                     const struct nlattr *nla)
4966 {
4967         struct nlattr *da[NFTA_SET_DESC_MAX + 1];
4968         int err;
4969 
4970         err = nla_parse_nested_deprecated(da, NFTA_SET_DESC_MAX, nla,
4971                                           nft_set_desc_policy, NULL);
4972         if (err < 0)
4973                 return err;
4974 
4975         if (da[NFTA_SET_DESC_SIZE] != NULL)
4976                 desc->size = ntohl(nla_get_be32(da[NFTA_SET_DESC_SIZE]));
4977         if (da[NFTA_SET_DESC_CONCAT])
4978                 err = nft_set_desc_concat(desc, da[NFTA_SET_DESC_CONCAT]);
4979 
4980         return err;
4981 }
4982 
4983 static int nft_set_expr_alloc(struct nft_ctx *ctx, struct nft_set *set,
4984                               const struct nlattr * const *nla,
4985                               struct nft_expr **exprs, int *num_exprs,
4986                               u32 flags)
4987 {
4988         struct nft_expr *expr;
4989         int err, i;
4990 
4991         if (nla[NFTA_SET_EXPR]) {
4992                 expr = nft_set_elem_expr_alloc(ctx, set, nla[NFTA_SET_EXPR]);
4993                 if (IS_ERR(expr)) {
4994                         err = PTR_ERR(expr);
4995                         goto err_set_expr_alloc;
4996                 }
4997                 exprs[0] = expr;
4998                 (*num_exprs)++;
4999         } else if (nla[NFTA_SET_EXPRESSIONS]) {
5000                 struct nlattr *tmp;
5001                 int left;
5002 
5003                 if (!(flags & NFT_SET_EXPR)) {
5004                         err = -EINVAL;
5005                         goto err_set_expr_alloc;
5006                 }
5007                 i = 0;
5008                 nla_for_each_nested(tmp, nla[NFTA_SET_EXPRESSIONS], left) {
5009                         if (i == NFT_SET_EXPR_MAX) {
5010                                 err = -E2BIG;
5011                                 goto err_set_expr_alloc;
5012                         }
5013                         if (nla_type(tmp) != NFTA_LIST_ELEM) {
5014                                 err = -EINVAL;
5015                                 goto err_set_expr_alloc;
5016                         }
5017                         expr = nft_set_elem_expr_alloc(ctx, set, tmp);
5018                         if (IS_ERR(expr)) {
5019                                 err = PTR_ERR(expr);
5020                                 goto err_set_expr_alloc;
5021                         }
5022                         exprs[i++] = expr;
5023                         (*num_exprs)++;
5024                 }
5025         }
5026 
5027         return 0;
5028 
5029 err_set_expr_alloc:
5030         for (i = 0; i < *num_exprs; i++)
5031                 nft_expr_destroy(ctx, exprs[i]);
5032 
5033         return err;
5034 }
5035 
5036 static bool nft_set_is_same(const struct nft_set *set,
5037                             const struct nft_set_desc *desc,
5038                             struct nft_expr *exprs[], u32 num_exprs, u32 flags)
5039 {
5040         int i;
5041 
5042         if (set->ktype != desc->ktype ||
5043             set->dtype != desc->dtype ||
5044             set->flags != flags ||
5045             set->klen != desc->klen ||
5046             set->dlen != desc->dlen ||
5047             set->field_count != desc->field_count ||
5048             set->num_exprs != num_exprs)
5049                 return false;
5050 
5051         for (i = 0; i < desc->field_count; i++) {
5052                 if (set->field_len[i] != desc->field_len[i])
5053                         return false;
5054         }
5055 
5056         for (i = 0; i < num_exprs; i++) {
5057                 if (set->exprs[i]->ops != exprs[i]->ops)
5058                         return false;
5059         }
5060 
5061         return true;
5062 }
5063 
5064 static int nf_tables_newset(struct sk_buff *skb, const struct nfnl_info *info,
5065                             const struct nlattr * const nla[])
5066 {
5067         struct netlink_ext_ack *extack = info->extack;
5068         u8 genmask = nft_genmask_next(info->net);
5069         u8 family = info->nfmsg->nfgen_family;
5070         const struct nft_set_ops *ops;
5071         struct net *net = info->net;
5072         struct nft_set_desc desc;
5073         struct nft_table *table;
5074         unsigned char *udata;
5075         struct nft_set *set;
5076         struct nft_ctx ctx;
5077         size_t alloc_size;
5078         int num_exprs = 0;
5079         char *name;
5080         int err, i;
5081         u16 udlen;
5082         u32 flags;
5083         u64 size;
5084 
5085         if (nla[NFTA_SET_TABLE] == NULL ||
5086             nla[NFTA_SET_NAME] == NULL ||
5087             nla[NFTA_SET_KEY_LEN] == NULL ||
5088             nla[NFTA_SET_ID] == NULL)
5089                 return -EINVAL;
5090 
5091         memset(&desc, 0, sizeof(desc));
5092 
5093         desc.ktype = NFT_DATA_VALUE;
5094         if (nla[NFTA_SET_KEY_TYPE] != NULL) {
5095                 desc.ktype = ntohl(nla_get_be32(nla[NFTA_SET_KEY_TYPE]));
5096                 if ((desc.ktype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK)
5097                         return -EINVAL;
5098         }
5099 
5100         desc.klen = ntohl(nla_get_be32(nla[NFTA_SET_KEY_LEN]));
5101         if (desc.klen == 0 || desc.klen > NFT_DATA_VALUE_MAXLEN)
5102                 return -EINVAL;
5103 
5104         flags = 0;
5105         if (nla[NFTA_SET_FLAGS] != NULL) {
5106                 flags = ntohl(nla_get_be32(nla[NFTA_SET_FLAGS]));
5107                 if (flags & ~(NFT_SET_ANONYMOUS | NFT_SET_CONSTANT |
5108                               NFT_SET_INTERVAL | NFT_SET_TIMEOUT |
5109                               NFT_SET_MAP | NFT_SET_EVAL |
5110                               NFT_SET_OBJECT | NFT_SET_CONCAT | NFT_SET_EXPR))
5111                         return -EOPNOTSUPP;
5112                 /* Only one of these operations is supported */
5113                 if ((flags & (NFT_SET_MAP | NFT_SET_OBJECT)) ==
5114                              (NFT_SET_MAP | NFT_SET_OBJECT))
5115                         return -EOPNOTSUPP;
5116                 if ((flags & (NFT_SET_EVAL | NFT_SET_OBJECT)) ==
5117                              (NFT_SET_EVAL | NFT_SET_OBJECT))
5118                         return -EOPNOTSUPP;
5119                 if ((flags & (NFT_SET_ANONYMOUS | NFT_SET_TIMEOUT | NFT_SET_EVAL)) ==
5120                              (NFT_SET_ANONYMOUS | NFT_SET_TIMEOUT))
5121                         return -EOPNOTSUPP;
5122                 if ((flags & (NFT_SET_CONSTANT | NFT_SET_TIMEOUT)) ==
5123                              (NFT_SET_CONSTANT | NFT_SET_TIMEOUT))
5124                         return -EOPNOTSUPP;
5125         }
5126 
5127         desc.dtype = 0;
5128         if (nla[NFTA_SET_DATA_TYPE] != NULL) {
5129                 if (!(flags & NFT_SET_MAP))
5130                         return -EINVAL;
5131 
5132                 desc.dtype = ntohl(nla_get_be32(nla[NFTA_SET_DATA_TYPE]));
5133                 if ((desc.dtype & NFT_DATA_RESERVED_MASK) == NFT_DATA_RESERVED_MASK &&
5134                     desc.dtype != NFT_DATA_VERDICT)
5135                         return -EINVAL;
5136 
5137                 if (desc.dtype != NFT_DATA_VERDICT) {
5138                         if (nla[NFTA_SET_DATA_LEN] == NULL)
5139                                 return -EINVAL;
5140                         desc.dlen = ntohl(nla_get_be32(nla[NFTA_SET_DATA_LEN]));
5141                         if (desc.dlen == 0 || desc.dlen > NFT_DATA_VALUE_MAXLEN)
5142                                 return -EINVAL;
5143                 } else
5144                         desc.dlen = sizeof(struct nft_verdict);
5145         } else if (flags & NFT_SET_MAP)
5146                 return -EINVAL;
5147 
5148         if (nla[NFTA_SET_OBJ_TYPE] != NULL) {
5149                 if (!(flags & NFT_SET_OBJECT))
5150                         return -EINVAL;
5151 
5152                 desc.objtype = ntohl(nla_get_be32(nla[NFTA_SET_OBJ_TYPE]));
5153                 if (desc.objtype == NFT_OBJECT_UNSPEC ||
5154                     desc.objtype > NFT_OBJECT_MAX)
5155                         return -EOPNOTSUPP;
5156         } else if (flags & NFT_SET_OBJECT)
5157                 return -EINVAL;
5158         else
5159                 desc.objtype = NFT_OBJECT_UNSPEC;
5160 
5161         desc.timeout = 0;
5162         if (nla[NFTA_SET_TIMEOUT] != NULL) {
5163                 if (!(flags & NFT_SET_TIMEOUT))
5164                         return -EINVAL;
5165 
5166                 if (flags & NFT_SET_ANONYMOUS)
5167                         return -EOPNOTSUPP;
5168 
5169                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_TIMEOUT], &desc.timeout);
5170                 if (err)
5171                         return err;
5172         }
5173         desc.gc_int = 0;
5174         if (nla[NFTA_SET_GC_INTERVAL] != NULL) {
5175                 if (!(flags & NFT_SET_TIMEOUT))
5176                         return -EINVAL;
5177 
5178                 if (flags & NFT_SET_ANONYMOUS)
5179                         return -EOPNOTSUPP;
5180 
5181                 desc.gc_int = ntohl(nla_get_be32(nla[NFTA_SET_GC_INTERVAL]));
5182         }
5183 
5184         desc.policy = NFT_SET_POL_PERFORMANCE;
5185         if (nla[NFTA_SET_POLICY] != NULL) {
5186                 desc.policy = ntohl(nla_get_be32(nla[NFTA_SET_POLICY]));
5187                 switch (desc.policy) {
5188                 case NFT_SET_POL_PERFORMANCE:
5189                 case NFT_SET_POL_MEMORY:
5190                         break;
5191                 default:
5192                         return -EOPNOTSUPP;
5193                 }
5194         }
5195 
5196         if (nla[NFTA_SET_DESC] != NULL) {
5197                 err = nf_tables_set_desc_parse(&desc, nla[NFTA_SET_DESC]);
5198                 if (err < 0)
5199                         return err;
5200 
5201                 if (desc.field_count > 1) {
5202                         if (!(flags & NFT_SET_CONCAT))
5203                                 return -EINVAL;
5204                 } else if (flags & NFT_SET_CONCAT) {
5205                         return -EINVAL;
5206                 }
5207         } else if (flags & NFT_SET_CONCAT) {
5208                 return -EINVAL;
5209         }
5210 
5211         if (nla[NFTA_SET_EXPR] || nla[NFTA_SET_EXPRESSIONS])
5212                 desc.expr = true;
5213 
5214         table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family, genmask,
5215                                  NETLINK_CB(skb).portid);
5216         if (IS_ERR(table)) {
5217                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
5218                 return PTR_ERR(table);
5219         }
5220 
5221         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5222 
5223         set = nft_set_lookup(table, nla[NFTA_SET_NAME], genmask);
5224         if (IS_ERR(set)) {
5225                 if (PTR_ERR(set) != -ENOENT) {
5226                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
5227                         return PTR_ERR(set);
5228                 }
5229         } else {
5230                 struct nft_expr *exprs[NFT_SET_EXPR_MAX] = {};
5231 
5232                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
5233                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
5234                         return -EEXIST;
5235                 }
5236                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
5237                         return -EOPNOTSUPP;
5238 
5239                 if (nft_set_is_anonymous(set))
5240                         return -EOPNOTSUPP;
5241 
5242                 err = nft_set_expr_alloc(&ctx, set, nla, exprs, &num_exprs, flags);
5243                 if (err < 0)
5244                         return err;
5245 
5246                 err = 0;
5247                 if (!nft_set_is_same(set, &desc, exprs, num_exprs, flags)) {
5248                         NL_SET_BAD_ATTR(extack, nla[NFTA_SET_NAME]);
5249                         err = -EEXIST;
5250                 }
5251 
5252                 for (i = 0; i < num_exprs; i++)
5253                         nft_expr_destroy(&ctx, exprs[i]);
5254 
5255                 if (err < 0)
5256                         return err;
5257 
5258                 return __nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set, &desc);
5259         }
5260 
5261         if (!(info->nlh->nlmsg_flags & NLM_F_CREATE))
5262                 return -ENOENT;
5263 
5264         ops = nft_select_set_ops(&ctx, flags, &desc);
5265         if (IS_ERR(ops))
5266                 return PTR_ERR(ops);
5267 
5268         udlen = 0;
5269         if (nla[NFTA_SET_USERDATA])
5270                 udlen = nla_len(nla[NFTA_SET_USERDATA]);
5271 
5272         size = 0;
5273         if (ops->privsize != NULL)
5274                 size = ops->privsize(nla, &desc);
5275         alloc_size = sizeof(*set) + size + udlen;
5276         if (alloc_size < size || alloc_size > INT_MAX)
5277                 return -ENOMEM;
5278 
5279         if (!nft_use_inc(&table->use))
5280                 return -EMFILE;
5281 
5282         set = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT);
5283         if (!set) {
5284                 err = -ENOMEM;
5285                 goto err_alloc;
5286         }
5287 
5288         name = nla_strdup(nla[NFTA_SET_NAME], GFP_KERNEL_ACCOUNT);
5289         if (!name) {
5290                 err = -ENOMEM;
5291                 goto err_set_name;
5292         }
5293 
5294         err = nf_tables_set_alloc_name(&ctx, set, name);
5295         kfree(name);
5296         if (err < 0)
5297                 goto err_set_name;
5298 
5299         udata = NULL;
5300         if (udlen) {
5301                 udata = set->data + size;
5302                 nla_memcpy(udata, nla[NFTA_SET_USERDATA], udlen);
5303         }
5304 
5305         INIT_LIST_HEAD(&set->bindings);
5306         INIT_LIST_HEAD(&set->catchall_list);
5307         refcount_set(&set->refs, 1);
5308         set->table = table;
5309         write_pnet(&set->net, net);
5310         set->ops = ops;
5311         set->ktype = desc.ktype;
5312         set->klen = desc.klen;
5313         set->dtype = desc.dtype;
5314         set->objtype = desc.objtype;
5315         set->dlen = desc.dlen;
5316         set->flags = flags;
5317         set->size = desc.size;
5318         set->policy = desc.policy;
5319         set->udlen = udlen;
5320         set->udata = udata;
5321         set->timeout = desc.timeout;
5322         set->gc_int = desc.gc_int;
5323 
5324         set->field_count = desc.field_count;
5325         for (i = 0; i < desc.field_count; i++)
5326                 set->field_len[i] = desc.field_len[i];
5327 
5328         err = ops->init(set, &desc, nla);
5329         if (err < 0)
5330                 goto err_set_init;
5331 
5332         err = nft_set_expr_alloc(&ctx, set, nla, set->exprs, &num_exprs, flags);
5333         if (err < 0)
5334                 goto err_set_destroy;
5335 
5336         set->num_exprs = num_exprs;
5337         set->handle = nf_tables_alloc_handle(table);
5338         INIT_LIST_HEAD(&set->pending_update);
5339 
5340         err = nft_trans_set_add(&ctx, NFT_MSG_NEWSET, set);
5341         if (err < 0)
5342                 goto err_set_expr_alloc;
5343 
5344         list_add_tail_rcu(&set->list, &table->sets);
5345 
5346         return 0;
5347 
5348 err_set_expr_alloc:
5349         for (i = 0; i < set->num_exprs; i++)
5350                 nft_expr_destroy(&ctx, set->exprs[i]);
5351 err_set_destroy:
5352         ops->destroy(&ctx, set);
5353 err_set_init:
5354         kfree(set->name);
5355 err_set_name:
5356         kvfree(set);
5357 err_alloc:
5358         nft_use_dec_restore(&table->use);
5359 
5360         return err;
5361 }
5362 
5363 static void nft_set_catchall_destroy(const struct nft_ctx *ctx,
5364                                      struct nft_set *set)
5365 {
5366         struct nft_set_elem_catchall *next, *catchall;
5367 
5368         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
5369                 list_del_rcu(&catchall->list);
5370                 nf_tables_set_elem_destroy(ctx, set, catchall->elem);
5371                 kfree_rcu(catchall, rcu);
5372         }
5373 }
5374 
5375 static void nft_set_put(struct nft_set *set)
5376 {
5377         if (refcount_dec_and_test(&set->refs)) {
5378                 kfree(set->name);
5379                 kvfree(set);
5380         }
5381 }
5382 
5383 static void nft_set_destroy(const struct nft_ctx *ctx, struct nft_set *set)
5384 {
5385         int i;
5386 
5387         if (WARN_ON(set->use > 0))
5388                 return;
5389 
5390         for (i = 0; i < set->num_exprs; i++)
5391                 nft_expr_destroy(ctx, set->exprs[i]);
5392 
5393         set->ops->destroy(ctx, set);
5394         nft_set_catchall_destroy(ctx, set);
5395         nft_set_put(set);
5396 }
5397 
5398 static int nf_tables_delset(struct sk_buff *skb, const struct nfnl_info *info,
5399                             const struct nlattr * const nla[])
5400 {
5401         struct netlink_ext_ack *extack = info->extack;
5402         u8 genmask = nft_genmask_next(info->net);
5403         u8 family = info->nfmsg->nfgen_family;
5404         struct net *net = info->net;
5405         const struct nlattr *attr;
5406         struct nft_table *table;
5407         struct nft_set *set;
5408         struct nft_ctx ctx;
5409 
5410         if (info->nfmsg->nfgen_family == NFPROTO_UNSPEC)
5411                 return -EAFNOSUPPORT;
5412 
5413         table = nft_table_lookup(net, nla[NFTA_SET_TABLE], family,
5414                                  genmask, NETLINK_CB(skb).portid);
5415         if (IS_ERR(table)) {
5416                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_TABLE]);
5417                 return PTR_ERR(table);
5418         }
5419 
5420         if (nla[NFTA_SET_HANDLE]) {
5421                 attr = nla[NFTA_SET_HANDLE];
5422                 set = nft_set_lookup_byhandle(table, attr, genmask);
5423         } else {
5424                 attr = nla[NFTA_SET_NAME];
5425                 set = nft_set_lookup(table, attr, genmask);
5426         }
5427 
5428         if (IS_ERR(set)) {
5429                 if (PTR_ERR(set) == -ENOENT &&
5430                     NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYSET)
5431                         return 0;
5432 
5433                 NL_SET_BAD_ATTR(extack, attr);
5434                 return PTR_ERR(set);
5435         }
5436         if (set->use ||
5437             (info->nlh->nlmsg_flags & NLM_F_NONREC &&
5438              atomic_read(&set->nelems) > 0)) {
5439                 NL_SET_BAD_ATTR(extack, attr);
5440                 return -EBUSY;
5441         }
5442 
5443         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
5444 
5445         return nft_delset(&ctx, set);
5446 }
5447 
5448 static int nft_validate_register_store(const struct nft_ctx *ctx,
5449                                        enum nft_registers reg,
5450                                        const struct nft_data *data,
5451                                        enum nft_data_types type,
5452                                        unsigned int len);
5453 
5454 static int nft_setelem_data_validate(const struct nft_ctx *ctx,
5455                                      struct nft_set *set,
5456                                      struct nft_elem_priv *elem_priv)
5457 {
5458         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
5459         enum nft_registers dreg;
5460 
5461         dreg = nft_type_to_reg(set->dtype);
5462         return nft_validate_register_store(ctx, dreg, nft_set_ext_data(ext),
5463                                            set->dtype == NFT_DATA_VERDICT ?
5464                                            NFT_DATA_VERDICT : NFT_DATA_VALUE,
5465                                            set->dlen);
5466 }
5467 
5468 static int nf_tables_bind_check_setelem(const struct nft_ctx *ctx,
5469                                         struct nft_set *set,
5470                                         const struct nft_set_iter *iter,
5471                                         struct nft_elem_priv *elem_priv)
5472 {
5473         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
5474 
5475         if (!nft_set_elem_active(ext, iter->genmask))
5476                 return 0;
5477 
5478         return nft_setelem_data_validate(ctx, set, elem_priv);
5479 }
5480 
5481 static int nft_set_catchall_bind_check(const struct nft_ctx *ctx,
5482                                        struct nft_set *set)
5483 {
5484         u8 genmask = nft_genmask_next(ctx->net);
5485         struct nft_set_elem_catchall *catchall;
5486         struct nft_set_ext *ext;
5487         int ret = 0;
5488 
5489         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5490                 ext = nft_set_elem_ext(set, catchall->elem);
5491                 if (!nft_set_elem_active(ext, genmask))
5492                         continue;
5493 
5494                 ret = nft_setelem_data_validate(ctx, set, catchall->elem);
5495                 if (ret < 0)
5496                         break;
5497         }
5498 
5499         return ret;
5500 }
5501 
5502 int nf_tables_bind_set(const struct nft_ctx *ctx, struct nft_set *set,
5503                        struct nft_set_binding *binding)
5504 {
5505         struct nft_set_binding *i;
5506         struct nft_set_iter iter;
5507 
5508         if (!list_empty(&set->bindings) && nft_set_is_anonymous(set))
5509                 return -EBUSY;
5510 
5511         if (binding->flags & NFT_SET_MAP) {
5512                 /* If the set is already bound to the same chain all
5513                  * jumps are already validated for that chain.
5514                  */
5515                 list_for_each_entry(i, &set->bindings, list) {
5516                         if (i->flags & NFT_SET_MAP &&
5517                             i->chain == binding->chain)
5518                                 goto bind;
5519                 }
5520 
5521                 iter.genmask    = nft_genmask_next(ctx->net);
5522                 iter.type       = NFT_ITER_UPDATE;
5523                 iter.skip       = 0;
5524                 iter.count      = 0;
5525                 iter.err        = 0;
5526                 iter.fn         = nf_tables_bind_check_setelem;
5527 
5528                 set->ops->walk(ctx, set, &iter);
5529                 if (!iter.err)
5530                         iter.err = nft_set_catchall_bind_check(ctx, set);
5531 
5532                 if (iter.err < 0)
5533                         return iter.err;
5534         }
5535 bind:
5536         if (!nft_use_inc(&set->use))
5537                 return -EMFILE;
5538 
5539         binding->chain = ctx->chain;
5540         list_add_tail_rcu(&binding->list, &set->bindings);
5541         nft_set_trans_bind(ctx, set);
5542 
5543         return 0;
5544 }
5545 EXPORT_SYMBOL_GPL(nf_tables_bind_set);
5546 
5547 static void nf_tables_unbind_set(const struct nft_ctx *ctx, struct nft_set *set,
5548                                  struct nft_set_binding *binding, bool event)
5549 {
5550         list_del_rcu(&binding->list);
5551 
5552         if (list_empty(&set->bindings) && nft_set_is_anonymous(set)) {
5553                 list_del_rcu(&set->list);
5554                 set->dead = 1;
5555                 if (event)
5556                         nf_tables_set_notify(ctx, set, NFT_MSG_DELSET,
5557                                              GFP_KERNEL);
5558         }
5559 }
5560 
5561 static void nft_setelem_data_activate(const struct net *net,
5562                                       const struct nft_set *set,
5563                                       struct nft_elem_priv *elem_priv);
5564 
5565 static int nft_mapelem_activate(const struct nft_ctx *ctx,
5566                                 struct nft_set *set,
5567                                 const struct nft_set_iter *iter,
5568                                 struct nft_elem_priv *elem_priv)
5569 {
5570         struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
5571 
5572         /* called from abort path, reverse check to undo changes. */
5573         if (nft_set_elem_active(ext, iter->genmask))
5574                 return 0;
5575 
5576         nft_clear(ctx->net, ext);
5577         nft_setelem_data_activate(ctx->net, set, elem_priv);
5578 
5579         return 0;
5580 }
5581 
5582 static void nft_map_catchall_activate(const struct nft_ctx *ctx,
5583                                       struct nft_set *set)
5584 {
5585         u8 genmask = nft_genmask_next(ctx->net);
5586         struct nft_set_elem_catchall *catchall;
5587         struct nft_set_ext *ext;
5588 
5589         list_for_each_entry(catchall, &set->catchall_list, list) {
5590                 ext = nft_set_elem_ext(set, catchall->elem);
5591                 if (!nft_set_elem_active(ext, genmask))
5592                         continue;
5593 
5594                 nft_clear(ctx->net, ext);
5595                 nft_setelem_data_activate(ctx->net, set, catchall->elem);
5596                 break;
5597         }
5598 }
5599 
5600 static void nft_map_activate(const struct nft_ctx *ctx, struct nft_set *set)
5601 {
5602         struct nft_set_iter iter = {
5603                 .genmask        = nft_genmask_next(ctx->net),
5604                 .type           = NFT_ITER_UPDATE,
5605                 .fn             = nft_mapelem_activate,
5606         };
5607 
5608         set->ops->walk(ctx, set, &iter);
5609         WARN_ON_ONCE(iter.err);
5610 
5611         nft_map_catchall_activate(ctx, set);
5612 }
5613 
5614 void nf_tables_activate_set(const struct nft_ctx *ctx, struct nft_set *set)
5615 {
5616         if (nft_set_is_anonymous(set)) {
5617                 if (set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
5618                         nft_map_activate(ctx, set);
5619 
5620                 nft_clear(ctx->net, set);
5621         }
5622 
5623         nft_use_inc_restore(&set->use);
5624 }
5625 EXPORT_SYMBOL_GPL(nf_tables_activate_set);
5626 
5627 void nf_tables_deactivate_set(const struct nft_ctx *ctx, struct nft_set *set,
5628                               struct nft_set_binding *binding,
5629                               enum nft_trans_phase phase)
5630 {
5631         switch (phase) {
5632         case NFT_TRANS_PREPARE_ERROR:
5633                 nft_set_trans_unbind(ctx, set);
5634                 if (nft_set_is_anonymous(set))
5635                         nft_deactivate_next(ctx->net, set);
5636                 else
5637                         list_del_rcu(&binding->list);
5638 
5639                 nft_use_dec(&set->use);
5640                 break;
5641         case NFT_TRANS_PREPARE:
5642                 if (nft_set_is_anonymous(set)) {
5643                         if (set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
5644                                 nft_map_deactivate(ctx, set);
5645 
5646                         nft_deactivate_next(ctx->net, set);
5647                 }
5648                 nft_use_dec(&set->use);
5649                 return;
5650         case NFT_TRANS_ABORT:
5651         case NFT_TRANS_RELEASE:
5652                 if (nft_set_is_anonymous(set) &&
5653                     set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
5654                         nft_map_deactivate(ctx, set);
5655 
5656                 nft_use_dec(&set->use);
5657                 fallthrough;
5658         default:
5659                 nf_tables_unbind_set(ctx, set, binding,
5660                                      phase == NFT_TRANS_COMMIT);
5661         }
5662 }
5663 EXPORT_SYMBOL_GPL(nf_tables_deactivate_set);
5664 
5665 void nf_tables_destroy_set(const struct nft_ctx *ctx, struct nft_set *set)
5666 {
5667         if (list_empty(&set->bindings) && nft_set_is_anonymous(set))
5668                 nft_set_destroy(ctx, set);
5669 }
5670 EXPORT_SYMBOL_GPL(nf_tables_destroy_set);
5671 
5672 const struct nft_set_ext_type nft_set_ext_types[] = {
5673         [NFT_SET_EXT_KEY]               = {
5674                 .align  = __alignof__(u32),
5675         },
5676         [NFT_SET_EXT_DATA]              = {
5677                 .align  = __alignof__(u32),
5678         },
5679         [NFT_SET_EXT_EXPRESSIONS]       = {
5680                 .align  = __alignof__(struct nft_set_elem_expr),
5681         },
5682         [NFT_SET_EXT_OBJREF]            = {
5683                 .len    = sizeof(struct nft_object *),
5684                 .align  = __alignof__(struct nft_object *),
5685         },
5686         [NFT_SET_EXT_FLAGS]             = {
5687                 .len    = sizeof(u8),
5688                 .align  = __alignof__(u8),
5689         },
5690         [NFT_SET_EXT_TIMEOUT]           = {
5691                 .len    = sizeof(u64),
5692                 .align  = __alignof__(u64),
5693         },
5694         [NFT_SET_EXT_EXPIRATION]        = {
5695                 .len    = sizeof(u64),
5696                 .align  = __alignof__(u64),
5697         },
5698         [NFT_SET_EXT_USERDATA]          = {
5699                 .len    = sizeof(struct nft_userdata),
5700                 .align  = __alignof__(struct nft_userdata),
5701         },
5702         [NFT_SET_EXT_KEY_END]           = {
5703                 .align  = __alignof__(u32),
5704         },
5705 };
5706 
5707 /*
5708  * Set elements
5709  */
5710 
5711 static const struct nla_policy nft_set_elem_policy[NFTA_SET_ELEM_MAX + 1] = {
5712         [NFTA_SET_ELEM_KEY]             = { .type = NLA_NESTED },
5713         [NFTA_SET_ELEM_DATA]            = { .type = NLA_NESTED },
5714         [NFTA_SET_ELEM_FLAGS]           = { .type = NLA_U32 },
5715         [NFTA_SET_ELEM_TIMEOUT]         = { .type = NLA_U64 },
5716         [NFTA_SET_ELEM_EXPIRATION]      = { .type = NLA_U64 },
5717         [NFTA_SET_ELEM_USERDATA]        = { .type = NLA_BINARY,
5718                                             .len = NFT_USERDATA_MAXLEN },
5719         [NFTA_SET_ELEM_EXPR]            = { .type = NLA_NESTED },
5720         [NFTA_SET_ELEM_OBJREF]          = { .type = NLA_STRING,
5721                                             .len = NFT_OBJ_MAXNAMELEN - 1 },
5722         [NFTA_SET_ELEM_KEY_END]         = { .type = NLA_NESTED },
5723         [NFTA_SET_ELEM_EXPRESSIONS]     = NLA_POLICY_NESTED_ARRAY(nft_expr_policy),
5724 };
5725 
5726 static const struct nla_policy nft_set_elem_list_policy[NFTA_SET_ELEM_LIST_MAX + 1] = {
5727         [NFTA_SET_ELEM_LIST_TABLE]      = { .type = NLA_STRING,
5728                                             .len = NFT_TABLE_MAXNAMELEN - 1 },
5729         [NFTA_SET_ELEM_LIST_SET]        = { .type = NLA_STRING,
5730                                             .len = NFT_SET_MAXNAMELEN - 1 },
5731         [NFTA_SET_ELEM_LIST_ELEMENTS]   = NLA_POLICY_NESTED_ARRAY(nft_set_elem_policy),
5732         [NFTA_SET_ELEM_LIST_SET_ID]     = { .type = NLA_U32 },
5733 };
5734 
5735 static int nft_set_elem_expr_dump(struct sk_buff *skb,
5736                                   const struct nft_set *set,
5737                                   const struct nft_set_ext *ext,
5738                                   bool reset)
5739 {
5740         struct nft_set_elem_expr *elem_expr;
5741         u32 size, num_exprs = 0;
5742         struct nft_expr *expr;
5743         struct nlattr *nest;
5744 
5745         elem_expr = nft_set_ext_expr(ext);
5746         nft_setelem_expr_foreach(expr, elem_expr, size)
5747                 num_exprs++;
5748 
5749         if (num_exprs == 1) {
5750                 expr = nft_setelem_expr_at(elem_expr, 0);
5751                 if (nft_expr_dump(skb, NFTA_SET_ELEM_EXPR, expr, reset) < 0)
5752                         return -1;
5753 
5754                 return 0;
5755         } else if (num_exprs > 1) {
5756                 nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_EXPRESSIONS);
5757                 if (nest == NULL)
5758                         goto nla_put_failure;
5759 
5760                 nft_setelem_expr_foreach(expr, elem_expr, size) {
5761                         expr = nft_setelem_expr_at(elem_expr, size);
5762                         if (nft_expr_dump(skb, NFTA_LIST_ELEM, expr, reset) < 0)
5763                                 goto nla_put_failure;
5764                 }
5765                 nla_nest_end(skb, nest);
5766         }
5767         return 0;
5768 
5769 nla_put_failure:
5770         return -1;
5771 }
5772 
5773 static int nf_tables_fill_setelem(struct sk_buff *skb,
5774                                   const struct nft_set *set,
5775                                   const struct nft_elem_priv *elem_priv,
5776                                   bool reset)
5777 {
5778         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
5779         unsigned char *b = skb_tail_pointer(skb);
5780         struct nlattr *nest;
5781 
5782         nest = nla_nest_start_noflag(skb, NFTA_LIST_ELEM);
5783         if (nest == NULL)
5784                 goto nla_put_failure;
5785 
5786         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
5787             nft_data_dump(skb, NFTA_SET_ELEM_KEY, nft_set_ext_key(ext),
5788                           NFT_DATA_VALUE, set->klen) < 0)
5789                 goto nla_put_failure;
5790 
5791         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
5792             nft_data_dump(skb, NFTA_SET_ELEM_KEY_END, nft_set_ext_key_end(ext),
5793                           NFT_DATA_VALUE, set->klen) < 0)
5794                 goto nla_put_failure;
5795 
5796         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
5797             nft_data_dump(skb, NFTA_SET_ELEM_DATA, nft_set_ext_data(ext),
5798                           nft_set_datatype(set), set->dlen) < 0)
5799                 goto nla_put_failure;
5800 
5801         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS) &&
5802             nft_set_elem_expr_dump(skb, set, ext, reset))
5803                 goto nla_put_failure;
5804 
5805         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
5806             nla_put_string(skb, NFTA_SET_ELEM_OBJREF,
5807                            (*nft_set_ext_obj(ext))->key.name) < 0)
5808                 goto nla_put_failure;
5809 
5810         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
5811             nla_put_be32(skb, NFTA_SET_ELEM_FLAGS,
5812                          htonl(*nft_set_ext_flags(ext))))
5813                 goto nla_put_failure;
5814 
5815         if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT) &&
5816             nla_put_be64(skb, NFTA_SET_ELEM_TIMEOUT,
5817                          nf_jiffies64_to_msecs(*nft_set_ext_timeout(ext)),
5818                          NFTA_SET_ELEM_PAD))
5819                 goto nla_put_failure;
5820 
5821         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
5822                 u64 expires, now = get_jiffies_64();
5823 
5824                 expires = *nft_set_ext_expiration(ext);
5825                 if (time_before64(now, expires))
5826                         expires -= now;
5827                 else
5828                         expires = 0;
5829 
5830                 if (nla_put_be64(skb, NFTA_SET_ELEM_EXPIRATION,
5831                                  nf_jiffies64_to_msecs(expires),
5832                                  NFTA_SET_ELEM_PAD))
5833                         goto nla_put_failure;
5834         }
5835 
5836         if (nft_set_ext_exists(ext, NFT_SET_EXT_USERDATA)) {
5837                 struct nft_userdata *udata;
5838 
5839                 udata = nft_set_ext_userdata(ext);
5840                 if (nla_put(skb, NFTA_SET_ELEM_USERDATA,
5841                             udata->len + 1, udata->data))
5842                         goto nla_put_failure;
5843         }
5844 
5845         nla_nest_end(skb, nest);
5846         return 0;
5847 
5848 nla_put_failure:
5849         nlmsg_trim(skb, b);
5850         return -EMSGSIZE;
5851 }
5852 
5853 struct nft_set_dump_args {
5854         const struct netlink_callback   *cb;
5855         struct nft_set_iter             iter;
5856         struct sk_buff                  *skb;
5857         bool                            reset;
5858 };
5859 
5860 static int nf_tables_dump_setelem(const struct nft_ctx *ctx,
5861                                   struct nft_set *set,
5862                                   const struct nft_set_iter *iter,
5863                                   struct nft_elem_priv *elem_priv)
5864 {
5865         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
5866         struct nft_set_dump_args *args;
5867 
5868         if (!nft_set_elem_active(ext, iter->genmask))
5869                 return 0;
5870 
5871         if (nft_set_elem_expired(ext) || nft_set_elem_is_dead(ext))
5872                 return 0;
5873 
5874         args = container_of(iter, struct nft_set_dump_args, iter);
5875         return nf_tables_fill_setelem(args->skb, set, elem_priv, args->reset);
5876 }
5877 
5878 static void audit_log_nft_set_reset(const struct nft_table *table,
5879                                     unsigned int base_seq,
5880                                     unsigned int nentries)
5881 {
5882         char *buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, base_seq);
5883 
5884         audit_log_nfcfg(buf, table->family, nentries,
5885                         AUDIT_NFT_OP_SETELEM_RESET, GFP_ATOMIC);
5886         kfree(buf);
5887 }
5888 
5889 struct nft_set_dump_ctx {
5890         const struct nft_set    *set;
5891         struct nft_ctx          ctx;
5892         bool                    reset;
5893 };
5894 
5895 static int nft_set_catchall_dump(struct net *net, struct sk_buff *skb,
5896                                  const struct nft_set *set, bool reset,
5897                                  unsigned int base_seq)
5898 {
5899         struct nft_set_elem_catchall *catchall;
5900         u8 genmask = nft_genmask_cur(net);
5901         struct nft_set_ext *ext;
5902         int ret = 0;
5903 
5904         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
5905                 ext = nft_set_elem_ext(set, catchall->elem);
5906                 if (!nft_set_elem_active(ext, genmask) ||
5907                     nft_set_elem_expired(ext))
5908                         continue;
5909 
5910                 ret = nf_tables_fill_setelem(skb, set, catchall->elem, reset);
5911                 if (reset && !ret)
5912                         audit_log_nft_set_reset(set->table, base_seq, 1);
5913                 break;
5914         }
5915 
5916         return ret;
5917 }
5918 
5919 static int nf_tables_dump_set(struct sk_buff *skb, struct netlink_callback *cb)
5920 {
5921         struct nft_set_dump_ctx *dump_ctx = cb->data;
5922         struct net *net = sock_net(skb->sk);
5923         struct nftables_pernet *nft_net;
5924         struct nft_table *table;
5925         struct nft_set *set;
5926         struct nft_set_dump_args args;
5927         bool set_found = false;
5928         struct nlmsghdr *nlh;
5929         struct nlattr *nest;
5930         u32 portid, seq;
5931         int event;
5932 
5933         rcu_read_lock();
5934         nft_net = nft_pernet(net);
5935         cb->seq = READ_ONCE(nft_net->base_seq);
5936 
5937         list_for_each_entry_rcu(table, &nft_net->tables, list) {
5938                 if (dump_ctx->ctx.family != NFPROTO_UNSPEC &&
5939                     dump_ctx->ctx.family != table->family)
5940                         continue;
5941 
5942                 if (table != dump_ctx->ctx.table)
5943                         continue;
5944 
5945                 list_for_each_entry_rcu(set, &table->sets, list) {
5946                         if (set == dump_ctx->set) {
5947                                 set_found = true;
5948                                 break;
5949                         }
5950                 }
5951                 break;
5952         }
5953 
5954         if (!set_found) {
5955                 rcu_read_unlock();
5956                 return -ENOENT;
5957         }
5958 
5959         event  = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWSETELEM);
5960         portid = NETLINK_CB(cb->skb).portid;
5961         seq    = cb->nlh->nlmsg_seq;
5962 
5963         nlh = nfnl_msg_put(skb, portid, seq, event, NLM_F_MULTI,
5964                            table->family, NFNETLINK_V0, nft_base_seq(net));
5965         if (!nlh)
5966                 goto nla_put_failure;
5967 
5968         if (nla_put_string(skb, NFTA_SET_ELEM_LIST_TABLE, table->name))
5969                 goto nla_put_failure;
5970         if (nla_put_string(skb, NFTA_SET_ELEM_LIST_SET, set->name))
5971                 goto nla_put_failure;
5972 
5973         nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
5974         if (nest == NULL)
5975                 goto nla_put_failure;
5976 
5977         args.cb                 = cb;
5978         args.skb                = skb;
5979         args.reset              = dump_ctx->reset;
5980         args.iter.genmask       = nft_genmask_cur(net);
5981         args.iter.type          = NFT_ITER_READ;
5982         args.iter.skip          = cb->args[0];
5983         args.iter.count         = 0;
5984         args.iter.err           = 0;
5985         args.iter.fn            = nf_tables_dump_setelem;
5986         set->ops->walk(&dump_ctx->ctx, set, &args.iter);
5987 
5988         if (!args.iter.err && args.iter.count == cb->args[0])
5989                 args.iter.err = nft_set_catchall_dump(net, skb, set,
5990                                                       dump_ctx->reset, cb->seq);
5991         nla_nest_end(skb, nest);
5992         nlmsg_end(skb, nlh);
5993 
5994         rcu_read_unlock();
5995 
5996         if (args.iter.err && args.iter.err != -EMSGSIZE)
5997                 return args.iter.err;
5998         if (args.iter.count == cb->args[0])
5999                 return 0;
6000 
6001         cb->args[0] = args.iter.count;
6002         return skb->len;
6003 
6004 nla_put_failure:
6005         rcu_read_unlock();
6006         return -ENOSPC;
6007 }
6008 
6009 static int nf_tables_dumpreset_set(struct sk_buff *skb,
6010                                    struct netlink_callback *cb)
6011 {
6012         struct nftables_pernet *nft_net = nft_pernet(sock_net(skb->sk));
6013         struct nft_set_dump_ctx *dump_ctx = cb->data;
6014         int ret, skip = cb->args[0];
6015 
6016         mutex_lock(&nft_net->commit_mutex);
6017 
6018         ret = nf_tables_dump_set(skb, cb);
6019 
6020         if (cb->args[0] > skip)
6021                 audit_log_nft_set_reset(dump_ctx->ctx.table, cb->seq,
6022                                         cb->args[0] - skip);
6023 
6024         mutex_unlock(&nft_net->commit_mutex);
6025 
6026         return ret;
6027 }
6028 
6029 static int nf_tables_dump_set_start(struct netlink_callback *cb)
6030 {
6031         struct nft_set_dump_ctx *dump_ctx = cb->data;
6032 
6033         cb->data = kmemdup(dump_ctx, sizeof(*dump_ctx), GFP_ATOMIC);
6034 
6035         return cb->data ? 0 : -ENOMEM;
6036 }
6037 
6038 static int nf_tables_dump_set_done(struct netlink_callback *cb)
6039 {
6040         kfree(cb->data);
6041         return 0;
6042 }
6043 
6044 static int nf_tables_fill_setelem_info(struct sk_buff *skb,
6045                                        const struct nft_ctx *ctx, u32 seq,
6046                                        u32 portid, int event, u16 flags,
6047                                        const struct nft_set *set,
6048                                        const struct nft_elem_priv *elem_priv,
6049                                        bool reset)
6050 {
6051         struct nlmsghdr *nlh;
6052         struct nlattr *nest;
6053         int err;
6054 
6055         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
6056         nlh = nfnl_msg_put(skb, portid, seq, event, flags, ctx->family,
6057                            NFNETLINK_V0, nft_base_seq(ctx->net));
6058         if (!nlh)
6059                 goto nla_put_failure;
6060 
6061         if (nla_put_string(skb, NFTA_SET_TABLE, ctx->table->name))
6062                 goto nla_put_failure;
6063         if (nla_put_string(skb, NFTA_SET_NAME, set->name))
6064                 goto nla_put_failure;
6065 
6066         nest = nla_nest_start_noflag(skb, NFTA_SET_ELEM_LIST_ELEMENTS);
6067         if (nest == NULL)
6068                 goto nla_put_failure;
6069 
6070         err = nf_tables_fill_setelem(skb, set, elem_priv, reset);
6071         if (err < 0)
6072                 goto nla_put_failure;
6073 
6074         nla_nest_end(skb, nest);
6075 
6076         nlmsg_end(skb, nlh);
6077         return 0;
6078 
6079 nla_put_failure:
6080         nlmsg_trim(skb, nlh);
6081         return -1;
6082 }
6083 
6084 static int nft_setelem_parse_flags(const struct nft_set *set,
6085                                    const struct nlattr *attr, u32 *flags)
6086 {
6087         if (attr == NULL)
6088                 return 0;
6089 
6090         *flags = ntohl(nla_get_be32(attr));
6091         if (*flags & ~(NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
6092                 return -EOPNOTSUPP;
6093         if (!(set->flags & NFT_SET_INTERVAL) &&
6094             *flags & NFT_SET_ELEM_INTERVAL_END)
6095                 return -EINVAL;
6096         if ((*flags & (NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL)) ==
6097             (NFT_SET_ELEM_INTERVAL_END | NFT_SET_ELEM_CATCHALL))
6098                 return -EINVAL;
6099 
6100         return 0;
6101 }
6102 
6103 static int nft_setelem_parse_key(struct nft_ctx *ctx, const struct nft_set *set,
6104                                  struct nft_data *key, struct nlattr *attr)
6105 {
6106         struct nft_data_desc desc = {
6107                 .type   = NFT_DATA_VALUE,
6108                 .size   = NFT_DATA_VALUE_MAXLEN,
6109                 .len    = set->klen,
6110         };
6111 
6112         return nft_data_init(ctx, key, &desc, attr);
6113 }
6114 
6115 static int nft_setelem_parse_data(struct nft_ctx *ctx, struct nft_set *set,
6116                                   struct nft_data_desc *desc,
6117                                   struct nft_data *data,
6118                                   struct nlattr *attr)
6119 {
6120         u32 dtype;
6121 
6122         if (set->dtype == NFT_DATA_VERDICT)
6123                 dtype = NFT_DATA_VERDICT;
6124         else
6125                 dtype = NFT_DATA_VALUE;
6126 
6127         desc->type = dtype;
6128         desc->size = NFT_DATA_VALUE_MAXLEN;
6129         desc->len = set->dlen;
6130         desc->flags = NFT_DATA_DESC_SETELEM;
6131 
6132         return nft_data_init(ctx, data, desc, attr);
6133 }
6134 
6135 static void *nft_setelem_catchall_get(const struct net *net,
6136                                       const struct nft_set *set)
6137 {
6138         struct nft_set_elem_catchall *catchall;
6139         u8 genmask = nft_genmask_cur(net);
6140         struct nft_set_ext *ext;
6141         void *priv = NULL;
6142 
6143         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6144                 ext = nft_set_elem_ext(set, catchall->elem);
6145                 if (!nft_set_elem_active(ext, genmask) ||
6146                     nft_set_elem_expired(ext))
6147                         continue;
6148 
6149                 priv = catchall->elem;
6150                 break;
6151         }
6152 
6153         return priv;
6154 }
6155 
6156 static int nft_setelem_get(struct nft_ctx *ctx, const struct nft_set *set,
6157                            struct nft_set_elem *elem, u32 flags)
6158 {
6159         void *priv;
6160 
6161         if (!(flags & NFT_SET_ELEM_CATCHALL)) {
6162                 priv = set->ops->get(ctx->net, set, elem, flags);
6163                 if (IS_ERR(priv))
6164                         return PTR_ERR(priv);
6165         } else {
6166                 priv = nft_setelem_catchall_get(ctx->net, set);
6167                 if (!priv)
6168                         return -ENOENT;
6169         }
6170         elem->priv = priv;
6171 
6172         return 0;
6173 }
6174 
6175 static int nft_get_set_elem(struct nft_ctx *ctx, const struct nft_set *set,
6176                             const struct nlattr *attr, bool reset)
6177 {
6178         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6179         struct nft_set_elem elem;
6180         struct sk_buff *skb;
6181         uint32_t flags = 0;
6182         int err;
6183 
6184         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6185                                           nft_set_elem_policy, NULL);
6186         if (err < 0)
6187                 return err;
6188 
6189         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6190         if (err < 0)
6191                 return err;
6192 
6193         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
6194                 return -EINVAL;
6195 
6196         if (nla[NFTA_SET_ELEM_KEY]) {
6197                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6198                                             nla[NFTA_SET_ELEM_KEY]);
6199                 if (err < 0)
6200                         return err;
6201         }
6202 
6203         if (nla[NFTA_SET_ELEM_KEY_END]) {
6204                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
6205                                             nla[NFTA_SET_ELEM_KEY_END]);
6206                 if (err < 0)
6207                         return err;
6208         }
6209 
6210         err = nft_setelem_get(ctx, set, &elem, flags);
6211         if (err < 0)
6212                 return err;
6213 
6214         err = -ENOMEM;
6215         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_ATOMIC);
6216         if (skb == NULL)
6217                 return err;
6218 
6219         err = nf_tables_fill_setelem_info(skb, ctx, ctx->seq, ctx->portid,
6220                                           NFT_MSG_NEWSETELEM, 0, set, elem.priv,
6221                                           reset);
6222         if (err < 0)
6223                 goto err_fill_setelem;
6224 
6225         return nfnetlink_unicast(skb, ctx->net, ctx->portid);
6226 
6227 err_fill_setelem:
6228         kfree_skb(skb);
6229         return err;
6230 }
6231 
6232 static int nft_set_dump_ctx_init(struct nft_set_dump_ctx *dump_ctx,
6233                                  const struct sk_buff *skb,
6234                                  const struct nfnl_info *info,
6235                                  const struct nlattr * const nla[],
6236                                  bool reset)
6237 {
6238         struct netlink_ext_ack *extack = info->extack;
6239         u8 genmask = nft_genmask_cur(info->net);
6240         u8 family = info->nfmsg->nfgen_family;
6241         struct net *net = info->net;
6242         struct nft_table *table;
6243         struct nft_set *set;
6244 
6245         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
6246                                  genmask, 0);
6247         if (IS_ERR(table)) {
6248                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
6249                 return PTR_ERR(table);
6250         }
6251 
6252         set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
6253         if (IS_ERR(set)) {
6254                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_SET]);
6255                 return PTR_ERR(set);
6256         }
6257 
6258         nft_ctx_init(&dump_ctx->ctx, net, skb,
6259                      info->nlh, family, table, NULL, nla);
6260         dump_ctx->set = set;
6261         dump_ctx->reset = reset;
6262         return 0;
6263 }
6264 
6265 /* called with rcu_read_lock held */
6266 static int nf_tables_getsetelem(struct sk_buff *skb,
6267                                 const struct nfnl_info *info,
6268                                 const struct nlattr * const nla[])
6269 {
6270         struct netlink_ext_ack *extack = info->extack;
6271         struct nft_set_dump_ctx dump_ctx;
6272         struct nlattr *attr;
6273         int rem, err = 0;
6274 
6275         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
6276                 struct netlink_dump_control c = {
6277                         .start = nf_tables_dump_set_start,
6278                         .dump = nf_tables_dump_set,
6279                         .done = nf_tables_dump_set_done,
6280                         .module = THIS_MODULE,
6281                 };
6282 
6283                 err = nft_set_dump_ctx_init(&dump_ctx, skb, info, nla, false);
6284                 if (err)
6285                         return err;
6286 
6287                 c.data = &dump_ctx;
6288                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
6289         }
6290 
6291         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
6292                 return -EINVAL;
6293 
6294         err = nft_set_dump_ctx_init(&dump_ctx, skb, info, nla, false);
6295         if (err)
6296                 return err;
6297 
6298         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6299                 err = nft_get_set_elem(&dump_ctx.ctx, dump_ctx.set, attr, false);
6300                 if (err < 0) {
6301                         NL_SET_BAD_ATTR(extack, attr);
6302                         break;
6303                 }
6304         }
6305 
6306         return err;
6307 }
6308 
6309 static int nf_tables_getsetelem_reset(struct sk_buff *skb,
6310                                       const struct nfnl_info *info,
6311                                       const struct nlattr * const nla[])
6312 {
6313         struct nftables_pernet *nft_net = nft_pernet(info->net);
6314         struct netlink_ext_ack *extack = info->extack;
6315         struct nft_set_dump_ctx dump_ctx;
6316         int rem, err = 0, nelems = 0;
6317         struct nlattr *attr;
6318 
6319         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
6320                 struct netlink_dump_control c = {
6321                         .start = nf_tables_dump_set_start,
6322                         .dump = nf_tables_dumpreset_set,
6323                         .done = nf_tables_dump_set_done,
6324                         .module = THIS_MODULE,
6325                 };
6326 
6327                 err = nft_set_dump_ctx_init(&dump_ctx, skb, info, nla, true);
6328                 if (err)
6329                         return err;
6330 
6331                 c.data = &dump_ctx;
6332                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
6333         }
6334 
6335         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
6336                 return -EINVAL;
6337 
6338         if (!try_module_get(THIS_MODULE))
6339                 return -EINVAL;
6340         rcu_read_unlock();
6341         mutex_lock(&nft_net->commit_mutex);
6342         rcu_read_lock();
6343 
6344         err = nft_set_dump_ctx_init(&dump_ctx, skb, info, nla, true);
6345         if (err)
6346                 goto out_unlock;
6347 
6348         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
6349                 err = nft_get_set_elem(&dump_ctx.ctx, dump_ctx.set, attr, true);
6350                 if (err < 0) {
6351                         NL_SET_BAD_ATTR(extack, attr);
6352                         break;
6353                 }
6354                 nelems++;
6355         }
6356         audit_log_nft_set_reset(dump_ctx.ctx.table, nft_net->base_seq, nelems);
6357 
6358 out_unlock:
6359         rcu_read_unlock();
6360         mutex_unlock(&nft_net->commit_mutex);
6361         rcu_read_lock();
6362         module_put(THIS_MODULE);
6363 
6364         return err;
6365 }
6366 
6367 static void nf_tables_setelem_notify(const struct nft_ctx *ctx,
6368                                      const struct nft_set *set,
6369                                      const struct nft_elem_priv *elem_priv,
6370                                      int event)
6371 {
6372         struct nftables_pernet *nft_net;
6373         struct net *net = ctx->net;
6374         u32 portid = ctx->portid;
6375         struct sk_buff *skb;
6376         u16 flags = 0;
6377         int err;
6378 
6379         if (!ctx->report && !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
6380                 return;
6381 
6382         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
6383         if (skb == NULL)
6384                 goto err;
6385 
6386         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
6387                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
6388 
6389         err = nf_tables_fill_setelem_info(skb, ctx, 0, portid, event, flags,
6390                                           set, elem_priv, false);
6391         if (err < 0) {
6392                 kfree_skb(skb);
6393                 goto err;
6394         }
6395 
6396         nft_net = nft_pernet(net);
6397         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
6398         return;
6399 err:
6400         nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
6401 }
6402 
6403 static struct nft_trans *nft_trans_elem_alloc(struct nft_ctx *ctx,
6404                                               int msg_type,
6405                                               struct nft_set *set)
6406 {
6407         struct nft_trans *trans;
6408 
6409         trans = nft_trans_alloc(ctx, msg_type, sizeof(struct nft_trans_elem));
6410         if (trans == NULL)
6411                 return NULL;
6412 
6413         nft_trans_elem_set(trans) = set;
6414         return trans;
6415 }
6416 
6417 struct nft_expr *nft_set_elem_expr_alloc(const struct nft_ctx *ctx,
6418                                          const struct nft_set *set,
6419                                          const struct nlattr *attr)
6420 {
6421         struct nft_expr *expr;
6422         int err;
6423 
6424         expr = nft_expr_init(ctx, attr);
6425         if (IS_ERR(expr))
6426                 return expr;
6427 
6428         err = -EOPNOTSUPP;
6429         if (expr->ops->type->flags & NFT_EXPR_GC) {
6430                 if (set->flags & NFT_SET_TIMEOUT)
6431                         goto err_set_elem_expr;
6432                 if (!set->ops->gc_init)
6433                         goto err_set_elem_expr;
6434                 set->ops->gc_init(set);
6435         }
6436 
6437         return expr;
6438 
6439 err_set_elem_expr:
6440         nft_expr_destroy(ctx, expr);
6441         return ERR_PTR(err);
6442 }
6443 
6444 static int nft_set_ext_check(const struct nft_set_ext_tmpl *tmpl, u8 id, u32 len)
6445 {
6446         len += nft_set_ext_types[id].len;
6447         if (len > tmpl->ext_len[id] ||
6448             len > U8_MAX)
6449                 return -1;
6450 
6451         return 0;
6452 }
6453 
6454 static int nft_set_ext_memcpy(const struct nft_set_ext_tmpl *tmpl, u8 id,
6455                               void *to, const void *from, u32 len)
6456 {
6457         if (nft_set_ext_check(tmpl, id, len) < 0)
6458                 return -1;
6459 
6460         memcpy(to, from, len);
6461 
6462         return 0;
6463 }
6464 
6465 struct nft_elem_priv *nft_set_elem_init(const struct nft_set *set,
6466                                         const struct nft_set_ext_tmpl *tmpl,
6467                                         const u32 *key, const u32 *key_end,
6468                                         const u32 *data,
6469                                         u64 timeout, u64 expiration, gfp_t gfp)
6470 {
6471         struct nft_set_ext *ext;
6472         void *elem;
6473 
6474         elem = kzalloc(set->ops->elemsize + tmpl->len, gfp);
6475         if (elem == NULL)
6476                 return ERR_PTR(-ENOMEM);
6477 
6478         ext = nft_set_elem_ext(set, elem);
6479         nft_set_ext_init(ext, tmpl);
6480 
6481         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY) &&
6482             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_KEY,
6483                                nft_set_ext_key(ext), key, set->klen) < 0)
6484                 goto err_ext_check;
6485 
6486         if (nft_set_ext_exists(ext, NFT_SET_EXT_KEY_END) &&
6487             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_KEY_END,
6488                                nft_set_ext_key_end(ext), key_end, set->klen) < 0)
6489                 goto err_ext_check;
6490 
6491         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
6492             nft_set_ext_memcpy(tmpl, NFT_SET_EXT_DATA,
6493                                nft_set_ext_data(ext), data, set->dlen) < 0)
6494                 goto err_ext_check;
6495 
6496         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPIRATION)) {
6497                 *nft_set_ext_expiration(ext) = get_jiffies_64() + expiration;
6498                 if (expiration == 0)
6499                         *nft_set_ext_expiration(ext) += timeout;
6500         }
6501         if (nft_set_ext_exists(ext, NFT_SET_EXT_TIMEOUT))
6502                 *nft_set_ext_timeout(ext) = timeout;
6503 
6504         return elem;
6505 
6506 err_ext_check:
6507         kfree(elem);
6508 
6509         return ERR_PTR(-EINVAL);
6510 }
6511 
6512 static void __nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
6513                                         struct nft_expr *expr)
6514 {
6515         if (expr->ops->destroy_clone) {
6516                 expr->ops->destroy_clone(ctx, expr);
6517                 module_put(expr->ops->type->owner);
6518         } else {
6519                 nf_tables_expr_destroy(ctx, expr);
6520         }
6521 }
6522 
6523 static void nft_set_elem_expr_destroy(const struct nft_ctx *ctx,
6524                                       struct nft_set_elem_expr *elem_expr)
6525 {
6526         struct nft_expr *expr;
6527         u32 size;
6528 
6529         nft_setelem_expr_foreach(expr, elem_expr, size)
6530                 __nft_set_elem_expr_destroy(ctx, expr);
6531 }
6532 
6533 /* Drop references and destroy. Called from gc, dynset and abort path. */
6534 void nft_set_elem_destroy(const struct nft_set *set,
6535                           const struct nft_elem_priv *elem_priv,
6536                           bool destroy_expr)
6537 {
6538         struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
6539         struct nft_ctx ctx = {
6540                 .net    = read_pnet(&set->net),
6541                 .family = set->table->family,
6542         };
6543 
6544         nft_data_release(nft_set_ext_key(ext), NFT_DATA_VALUE);
6545         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
6546                 nft_data_release(nft_set_ext_data(ext), set->dtype);
6547         if (destroy_expr && nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
6548                 nft_set_elem_expr_destroy(&ctx, nft_set_ext_expr(ext));
6549         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
6550                 nft_use_dec(&(*nft_set_ext_obj(ext))->use);
6551 
6552         kfree(elem_priv);
6553 }
6554 EXPORT_SYMBOL_GPL(nft_set_elem_destroy);
6555 
6556 /* Destroy element. References have been already dropped in the preparation
6557  * path via nft_setelem_data_deactivate().
6558  */
6559 void nf_tables_set_elem_destroy(const struct nft_ctx *ctx,
6560                                 const struct nft_set *set,
6561                                 const struct nft_elem_priv *elem_priv)
6562 {
6563         struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
6564 
6565         if (nft_set_ext_exists(ext, NFT_SET_EXT_EXPRESSIONS))
6566                 nft_set_elem_expr_destroy(ctx, nft_set_ext_expr(ext));
6567 
6568         kfree(elem_priv);
6569 }
6570 
6571 int nft_set_elem_expr_clone(const struct nft_ctx *ctx, struct nft_set *set,
6572                             struct nft_expr *expr_array[])
6573 {
6574         struct nft_expr *expr;
6575         int err, i, k;
6576 
6577         for (i = 0; i < set->num_exprs; i++) {
6578                 expr = kzalloc(set->exprs[i]->ops->size, GFP_KERNEL_ACCOUNT);
6579                 if (!expr)
6580                         goto err_expr;
6581 
6582                 err = nft_expr_clone(expr, set->exprs[i], GFP_KERNEL_ACCOUNT);
6583                 if (err < 0) {
6584                         kfree(expr);
6585                         goto err_expr;
6586                 }
6587                 expr_array[i] = expr;
6588         }
6589 
6590         return 0;
6591 
6592 err_expr:
6593         for (k = i - 1; k >= 0; k--)
6594                 nft_expr_destroy(ctx, expr_array[k]);
6595 
6596         return -ENOMEM;
6597 }
6598 
6599 static int nft_set_elem_expr_setup(struct nft_ctx *ctx,
6600                                    const struct nft_set_ext_tmpl *tmpl,
6601                                    const struct nft_set_ext *ext,
6602                                    struct nft_expr *expr_array[],
6603                                    u32 num_exprs)
6604 {
6605         struct nft_set_elem_expr *elem_expr = nft_set_ext_expr(ext);
6606         u32 len = sizeof(struct nft_set_elem_expr);
6607         struct nft_expr *expr;
6608         int i, err;
6609 
6610         if (num_exprs == 0)
6611                 return 0;
6612 
6613         for (i = 0; i < num_exprs; i++)
6614                 len += expr_array[i]->ops->size;
6615 
6616         if (nft_set_ext_check(tmpl, NFT_SET_EXT_EXPRESSIONS, len) < 0)
6617                 return -EINVAL;
6618 
6619         for (i = 0; i < num_exprs; i++) {
6620                 expr = nft_setelem_expr_at(elem_expr, elem_expr->size);
6621                 err = nft_expr_clone(expr, expr_array[i], GFP_KERNEL_ACCOUNT);
6622                 if (err < 0)
6623                         goto err_elem_expr_setup;
6624 
6625                 elem_expr->size += expr_array[i]->ops->size;
6626                 nft_expr_destroy(ctx, expr_array[i]);
6627                 expr_array[i] = NULL;
6628         }
6629 
6630         return 0;
6631 
6632 err_elem_expr_setup:
6633         for (; i < num_exprs; i++) {
6634                 nft_expr_destroy(ctx, expr_array[i]);
6635                 expr_array[i] = NULL;
6636         }
6637 
6638         return -ENOMEM;
6639 }
6640 
6641 struct nft_set_ext *nft_set_catchall_lookup(const struct net *net,
6642                                             const struct nft_set *set)
6643 {
6644         struct nft_set_elem_catchall *catchall;
6645         u8 genmask = nft_genmask_cur(net);
6646         struct nft_set_ext *ext;
6647 
6648         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
6649                 ext = nft_set_elem_ext(set, catchall->elem);
6650                 if (nft_set_elem_active(ext, genmask) &&
6651                     !nft_set_elem_expired(ext) &&
6652                     !nft_set_elem_is_dead(ext))
6653                         return ext;
6654         }
6655 
6656         return NULL;
6657 }
6658 EXPORT_SYMBOL_GPL(nft_set_catchall_lookup);
6659 
6660 static int nft_setelem_catchall_insert(const struct net *net,
6661                                        struct nft_set *set,
6662                                        const struct nft_set_elem *elem,
6663                                        struct nft_elem_priv **priv)
6664 {
6665         struct nft_set_elem_catchall *catchall;
6666         u8 genmask = nft_genmask_next(net);
6667         struct nft_set_ext *ext;
6668 
6669         list_for_each_entry(catchall, &set->catchall_list, list) {
6670                 ext = nft_set_elem_ext(set, catchall->elem);
6671                 if (nft_set_elem_active(ext, genmask)) {
6672                         *priv = catchall->elem;
6673                         return -EEXIST;
6674                 }
6675         }
6676 
6677         catchall = kmalloc(sizeof(*catchall), GFP_KERNEL);
6678         if (!catchall)
6679                 return -ENOMEM;
6680 
6681         catchall->elem = elem->priv;
6682         list_add_tail_rcu(&catchall->list, &set->catchall_list);
6683 
6684         return 0;
6685 }
6686 
6687 static int nft_setelem_insert(const struct net *net,
6688                               struct nft_set *set,
6689                               const struct nft_set_elem *elem,
6690                               struct nft_elem_priv **elem_priv,
6691                               unsigned int flags)
6692 {
6693         int ret;
6694 
6695         if (flags & NFT_SET_ELEM_CATCHALL)
6696                 ret = nft_setelem_catchall_insert(net, set, elem, elem_priv);
6697         else
6698                 ret = set->ops->insert(net, set, elem, elem_priv);
6699 
6700         return ret;
6701 }
6702 
6703 static bool nft_setelem_is_catchall(const struct nft_set *set,
6704                                     const struct nft_elem_priv *elem_priv)
6705 {
6706         struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
6707 
6708         if (nft_set_ext_exists(ext, NFT_SET_EXT_FLAGS) &&
6709             *nft_set_ext_flags(ext) & NFT_SET_ELEM_CATCHALL)
6710                 return true;
6711 
6712         return false;
6713 }
6714 
6715 static void nft_setelem_activate(struct net *net, struct nft_set *set,
6716                                  struct nft_elem_priv *elem_priv)
6717 {
6718         struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
6719 
6720         if (nft_setelem_is_catchall(set, elem_priv)) {
6721                 nft_clear(net, ext);
6722         } else {
6723                 set->ops->activate(net, set, elem_priv);
6724         }
6725 }
6726 
6727 static int nft_setelem_catchall_deactivate(const struct net *net,
6728                                            struct nft_set *set,
6729                                            struct nft_set_elem *elem)
6730 {
6731         struct nft_set_elem_catchall *catchall;
6732         struct nft_set_ext *ext;
6733 
6734         list_for_each_entry(catchall, &set->catchall_list, list) {
6735                 ext = nft_set_elem_ext(set, catchall->elem);
6736                 if (!nft_is_active_next(net, ext))
6737                         continue;
6738 
6739                 kfree(elem->priv);
6740                 elem->priv = catchall->elem;
6741                 nft_set_elem_change_active(net, set, ext);
6742                 return 0;
6743         }
6744 
6745         return -ENOENT;
6746 }
6747 
6748 static int __nft_setelem_deactivate(const struct net *net,
6749                                     struct nft_set *set,
6750                                     struct nft_set_elem *elem)
6751 {
6752         void *priv;
6753 
6754         priv = set->ops->deactivate(net, set, elem);
6755         if (!priv)
6756                 return -ENOENT;
6757 
6758         kfree(elem->priv);
6759         elem->priv = priv;
6760         set->ndeact++;
6761 
6762         return 0;
6763 }
6764 
6765 static int nft_setelem_deactivate(const struct net *net,
6766                                   struct nft_set *set,
6767                                   struct nft_set_elem *elem, u32 flags)
6768 {
6769         int ret;
6770 
6771         if (flags & NFT_SET_ELEM_CATCHALL)
6772                 ret = nft_setelem_catchall_deactivate(net, set, elem);
6773         else
6774                 ret = __nft_setelem_deactivate(net, set, elem);
6775 
6776         return ret;
6777 }
6778 
6779 static void nft_setelem_catchall_destroy(struct nft_set_elem_catchall *catchall)
6780 {
6781         list_del_rcu(&catchall->list);
6782         kfree_rcu(catchall, rcu);
6783 }
6784 
6785 static void nft_setelem_catchall_remove(const struct net *net,
6786                                         const struct nft_set *set,
6787                                         struct nft_elem_priv *elem_priv)
6788 {
6789         struct nft_set_elem_catchall *catchall, *next;
6790 
6791         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
6792                 if (catchall->elem == elem_priv) {
6793                         nft_setelem_catchall_destroy(catchall);
6794                         break;
6795                 }
6796         }
6797 }
6798 
6799 static void nft_setelem_remove(const struct net *net,
6800                                const struct nft_set *set,
6801                                struct nft_elem_priv *elem_priv)
6802 {
6803         if (nft_setelem_is_catchall(set, elem_priv))
6804                 nft_setelem_catchall_remove(net, set, elem_priv);
6805         else
6806                 set->ops->remove(net, set, elem_priv);
6807 }
6808 
6809 static bool nft_setelem_valid_key_end(const struct nft_set *set,
6810                                       struct nlattr **nla, u32 flags)
6811 {
6812         if ((set->flags & (NFT_SET_CONCAT | NFT_SET_INTERVAL)) ==
6813                           (NFT_SET_CONCAT | NFT_SET_INTERVAL)) {
6814                 if (flags & NFT_SET_ELEM_INTERVAL_END)
6815                         return false;
6816 
6817                 if (nla[NFTA_SET_ELEM_KEY_END] &&
6818                     flags & NFT_SET_ELEM_CATCHALL)
6819                         return false;
6820         } else {
6821                 if (nla[NFTA_SET_ELEM_KEY_END])
6822                         return false;
6823         }
6824 
6825         return true;
6826 }
6827 
6828 static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
6829                             const struct nlattr *attr, u32 nlmsg_flags)
6830 {
6831         struct nft_expr *expr_array[NFT_SET_EXPR_MAX] = {};
6832         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
6833         u8 genmask = nft_genmask_next(ctx->net);
6834         u32 flags = 0, size = 0, num_exprs = 0;
6835         struct nft_set_ext_tmpl tmpl;
6836         struct nft_set_ext *ext, *ext2;
6837         struct nft_set_elem elem;
6838         struct nft_set_binding *binding;
6839         struct nft_elem_priv *elem_priv;
6840         struct nft_object *obj = NULL;
6841         struct nft_userdata *udata;
6842         struct nft_data_desc desc;
6843         enum nft_registers dreg;
6844         struct nft_trans *trans;
6845         u64 expiration;
6846         u64 timeout;
6847         int err, i;
6848         u8 ulen;
6849 
6850         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
6851                                           nft_set_elem_policy, NULL);
6852         if (err < 0)
6853                 return err;
6854 
6855         nft_set_ext_prepare(&tmpl);
6856 
6857         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
6858         if (err < 0)
6859                 return err;
6860 
6861         if (((flags & NFT_SET_ELEM_CATCHALL) && nla[NFTA_SET_ELEM_KEY]) ||
6862             (!(flags & NFT_SET_ELEM_CATCHALL) && !nla[NFTA_SET_ELEM_KEY]))
6863                 return -EINVAL;
6864 
6865         if (flags != 0) {
6866                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
6867                 if (err < 0)
6868                         return err;
6869         }
6870 
6871         if (set->flags & NFT_SET_MAP) {
6872                 if (nla[NFTA_SET_ELEM_DATA] == NULL &&
6873                     !(flags & NFT_SET_ELEM_INTERVAL_END))
6874                         return -EINVAL;
6875         } else {
6876                 if (nla[NFTA_SET_ELEM_DATA] != NULL)
6877                         return -EINVAL;
6878         }
6879 
6880         if (set->flags & NFT_SET_OBJECT) {
6881                 if (!nla[NFTA_SET_ELEM_OBJREF] &&
6882                     !(flags & NFT_SET_ELEM_INTERVAL_END))
6883                         return -EINVAL;
6884         } else {
6885                 if (nla[NFTA_SET_ELEM_OBJREF])
6886                         return -EINVAL;
6887         }
6888 
6889         if (!nft_setelem_valid_key_end(set, nla, flags))
6890                 return -EINVAL;
6891 
6892         if ((flags & NFT_SET_ELEM_INTERVAL_END) &&
6893              (nla[NFTA_SET_ELEM_DATA] ||
6894               nla[NFTA_SET_ELEM_OBJREF] ||
6895               nla[NFTA_SET_ELEM_TIMEOUT] ||
6896               nla[NFTA_SET_ELEM_EXPIRATION] ||
6897               nla[NFTA_SET_ELEM_USERDATA] ||
6898               nla[NFTA_SET_ELEM_EXPR] ||
6899               nla[NFTA_SET_ELEM_KEY_END] ||
6900               nla[NFTA_SET_ELEM_EXPRESSIONS]))
6901                 return -EINVAL;
6902 
6903         timeout = 0;
6904         if (nla[NFTA_SET_ELEM_TIMEOUT] != NULL) {
6905                 if (!(set->flags & NFT_SET_TIMEOUT))
6906                         return -EINVAL;
6907                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_TIMEOUT],
6908                                             &timeout);
6909                 if (err)
6910                         return err;
6911         } else if (set->flags & NFT_SET_TIMEOUT &&
6912                    !(flags & NFT_SET_ELEM_INTERVAL_END)) {
6913                 timeout = READ_ONCE(set->timeout);
6914         }
6915 
6916         expiration = 0;
6917         if (nla[NFTA_SET_ELEM_EXPIRATION] != NULL) {
6918                 if (!(set->flags & NFT_SET_TIMEOUT))
6919                         return -EINVAL;
6920                 err = nf_msecs_to_jiffies64(nla[NFTA_SET_ELEM_EXPIRATION],
6921                                             &expiration);
6922                 if (err)
6923                         return err;
6924         }
6925 
6926         if (nla[NFTA_SET_ELEM_EXPR]) {
6927                 struct nft_expr *expr;
6928 
6929                 if (set->num_exprs && set->num_exprs != 1)
6930                         return -EOPNOTSUPP;
6931 
6932                 expr = nft_set_elem_expr_alloc(ctx, set,
6933                                                nla[NFTA_SET_ELEM_EXPR]);
6934                 if (IS_ERR(expr))
6935                         return PTR_ERR(expr);
6936 
6937                 expr_array[0] = expr;
6938                 num_exprs = 1;
6939 
6940                 if (set->num_exprs && set->exprs[0]->ops != expr->ops) {
6941                         err = -EOPNOTSUPP;
6942                         goto err_set_elem_expr;
6943                 }
6944         } else if (nla[NFTA_SET_ELEM_EXPRESSIONS]) {
6945                 struct nft_expr *expr;
6946                 struct nlattr *tmp;
6947                 int left;
6948 
6949                 i = 0;
6950                 nla_for_each_nested(tmp, nla[NFTA_SET_ELEM_EXPRESSIONS], left) {
6951                         if (i == NFT_SET_EXPR_MAX ||
6952                             (set->num_exprs && set->num_exprs == i)) {
6953                                 err = -E2BIG;
6954                                 goto err_set_elem_expr;
6955                         }
6956                         if (nla_type(tmp) != NFTA_LIST_ELEM) {
6957                                 err = -EINVAL;
6958                                 goto err_set_elem_expr;
6959                         }
6960                         expr = nft_set_elem_expr_alloc(ctx, set, tmp);
6961                         if (IS_ERR(expr)) {
6962                                 err = PTR_ERR(expr);
6963                                 goto err_set_elem_expr;
6964                         }
6965                         expr_array[i] = expr;
6966                         num_exprs++;
6967 
6968                         if (set->num_exprs && expr->ops != set->exprs[i]->ops) {
6969                                 err = -EOPNOTSUPP;
6970                                 goto err_set_elem_expr;
6971                         }
6972                         i++;
6973                 }
6974                 if (set->num_exprs && set->num_exprs != i) {
6975                         err = -EOPNOTSUPP;
6976                         goto err_set_elem_expr;
6977                 }
6978         } else if (set->num_exprs > 0 &&
6979                    !(flags & NFT_SET_ELEM_INTERVAL_END)) {
6980                 err = nft_set_elem_expr_clone(ctx, set, expr_array);
6981                 if (err < 0)
6982                         goto err_set_elem_expr_clone;
6983 
6984                 num_exprs = set->num_exprs;
6985         }
6986 
6987         if (nla[NFTA_SET_ELEM_KEY]) {
6988                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
6989                                             nla[NFTA_SET_ELEM_KEY]);
6990                 if (err < 0)
6991                         goto err_set_elem_expr;
6992 
6993                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
6994                 if (err < 0)
6995                         goto err_parse_key;
6996         }
6997 
6998         if (nla[NFTA_SET_ELEM_KEY_END]) {
6999                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
7000                                             nla[NFTA_SET_ELEM_KEY_END]);
7001                 if (err < 0)
7002                         goto err_parse_key;
7003 
7004                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
7005                 if (err < 0)
7006                         goto err_parse_key_end;
7007         }
7008 
7009         if (timeout > 0) {
7010                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_EXPIRATION);
7011                 if (err < 0)
7012                         goto err_parse_key_end;
7013 
7014                 if (timeout != READ_ONCE(set->timeout)) {
7015                         err = nft_set_ext_add(&tmpl, NFT_SET_EXT_TIMEOUT);
7016                         if (err < 0)
7017                                 goto err_parse_key_end;
7018                 }
7019         }
7020 
7021         if (num_exprs) {
7022                 for (i = 0; i < num_exprs; i++)
7023                         size += expr_array[i]->ops->size;
7024 
7025                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_EXPRESSIONS,
7026                                              sizeof(struct nft_set_elem_expr) + size);
7027                 if (err < 0)
7028                         goto err_parse_key_end;
7029         }
7030 
7031         if (nla[NFTA_SET_ELEM_OBJREF] != NULL) {
7032                 obj = nft_obj_lookup(ctx->net, ctx->table,
7033                                      nla[NFTA_SET_ELEM_OBJREF],
7034                                      set->objtype, genmask);
7035                 if (IS_ERR(obj)) {
7036                         err = PTR_ERR(obj);
7037                         obj = NULL;
7038                         goto err_parse_key_end;
7039                 }
7040 
7041                 if (!nft_use_inc(&obj->use)) {
7042                         err = -EMFILE;
7043                         obj = NULL;
7044                         goto err_parse_key_end;
7045                 }
7046 
7047                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_OBJREF);
7048                 if (err < 0)
7049                         goto err_parse_key_end;
7050         }
7051 
7052         if (nla[NFTA_SET_ELEM_DATA] != NULL) {
7053                 err = nft_setelem_parse_data(ctx, set, &desc, &elem.data.val,
7054                                              nla[NFTA_SET_ELEM_DATA]);
7055                 if (err < 0)
7056                         goto err_parse_key_end;
7057 
7058                 dreg = nft_type_to_reg(set->dtype);
7059                 list_for_each_entry(binding, &set->bindings, list) {
7060                         struct nft_ctx bind_ctx = {
7061                                 .net    = ctx->net,
7062                                 .family = ctx->family,
7063                                 .table  = ctx->table,
7064                                 .chain  = (struct nft_chain *)binding->chain,
7065                         };
7066 
7067                         if (!(binding->flags & NFT_SET_MAP))
7068                                 continue;
7069 
7070                         err = nft_validate_register_store(&bind_ctx, dreg,
7071                                                           &elem.data.val,
7072                                                           desc.type, desc.len);
7073                         if (err < 0)
7074                                 goto err_parse_data;
7075 
7076                         if (desc.type == NFT_DATA_VERDICT &&
7077                             (elem.data.val.verdict.code == NFT_GOTO ||
7078                              elem.data.val.verdict.code == NFT_JUMP))
7079                                 nft_validate_state_update(ctx->table,
7080                                                           NFT_VALIDATE_NEED);
7081                 }
7082 
7083                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_DATA, desc.len);
7084                 if (err < 0)
7085                         goto err_parse_data;
7086         }
7087 
7088         /* The full maximum length of userdata can exceed the maximum
7089          * offset value (U8_MAX) for following extensions, therefor it
7090          * must be the last extension added.
7091          */
7092         ulen = 0;
7093         if (nla[NFTA_SET_ELEM_USERDATA] != NULL) {
7094                 ulen = nla_len(nla[NFTA_SET_ELEM_USERDATA]);
7095                 if (ulen > 0) {
7096                         err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_USERDATA,
7097                                                      ulen);
7098                         if (err < 0)
7099                                 goto err_parse_data;
7100                 }
7101         }
7102 
7103         elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
7104                                       elem.key_end.val.data, elem.data.val.data,
7105                                       timeout, expiration, GFP_KERNEL_ACCOUNT);
7106         if (IS_ERR(elem.priv)) {
7107                 err = PTR_ERR(elem.priv);
7108                 goto err_parse_data;
7109         }
7110 
7111         ext = nft_set_elem_ext(set, elem.priv);
7112         if (flags)
7113                 *nft_set_ext_flags(ext) = flags;
7114 
7115         if (obj)
7116                 *nft_set_ext_obj(ext) = obj;
7117 
7118         if (ulen > 0) {
7119                 if (nft_set_ext_check(&tmpl, NFT_SET_EXT_USERDATA, ulen) < 0) {
7120                         err = -EINVAL;
7121                         goto err_elem_free;
7122                 }
7123                 udata = nft_set_ext_userdata(ext);
7124                 udata->len = ulen - 1;
7125                 nla_memcpy(&udata->data, nla[NFTA_SET_ELEM_USERDATA], ulen);
7126         }
7127         err = nft_set_elem_expr_setup(ctx, &tmpl, ext, expr_array, num_exprs);
7128         if (err < 0)
7129                 goto err_elem_free;
7130 
7131         trans = nft_trans_elem_alloc(ctx, NFT_MSG_NEWSETELEM, set);
7132         if (trans == NULL) {
7133                 err = -ENOMEM;
7134                 goto err_elem_free;
7135         }
7136 
7137         ext->genmask = nft_genmask_cur(ctx->net);
7138 
7139         err = nft_setelem_insert(ctx->net, set, &elem, &elem_priv, flags);
7140         if (err) {
7141                 if (err == -EEXIST) {
7142                         ext2 = nft_set_elem_ext(set, elem_priv);
7143                         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
7144                             nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
7145                             nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
7146                             nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF))
7147                                 goto err_element_clash;
7148                         if ((nft_set_ext_exists(ext, NFT_SET_EXT_DATA) &&
7149                              nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) &&
7150                              memcmp(nft_set_ext_data(ext),
7151                                     nft_set_ext_data(ext2), set->dlen) != 0) ||
7152                             (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) &&
7153                              nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
7154                              *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
7155                                 goto err_element_clash;
7156                         else if (!(nlmsg_flags & NLM_F_EXCL))
7157                                 err = 0;
7158                 } else if (err == -ENOTEMPTY) {
7159                         /* ENOTEMPTY reports overlapping between this element
7160                          * and an existing one.
7161                          */
7162                         err = -EEXIST;
7163                 }
7164                 goto err_element_clash;
7165         }
7166 
7167         if (!(flags & NFT_SET_ELEM_CATCHALL)) {
7168                 unsigned int max = set->size ? set->size + set->ndeact : UINT_MAX;
7169 
7170                 if (!atomic_add_unless(&set->nelems, 1, max)) {
7171                         err = -ENFILE;
7172                         goto err_set_full;
7173                 }
7174         }
7175 
7176         nft_trans_elem_priv(trans) = elem.priv;
7177         nft_trans_commit_list_add_tail(ctx->net, trans);
7178         return 0;
7179 
7180 err_set_full:
7181         nft_setelem_remove(ctx->net, set, elem.priv);
7182 err_element_clash:
7183         kfree(trans);
7184 err_elem_free:
7185         nf_tables_set_elem_destroy(ctx, set, elem.priv);
7186 err_parse_data:
7187         if (nla[NFTA_SET_ELEM_DATA] != NULL)
7188                 nft_data_release(&elem.data.val, desc.type);
7189 err_parse_key_end:
7190         if (obj)
7191                 nft_use_dec_restore(&obj->use);
7192 
7193         nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
7194 err_parse_key:
7195         nft_data_release(&elem.key.val, NFT_DATA_VALUE);
7196 err_set_elem_expr:
7197         for (i = 0; i < num_exprs && expr_array[i]; i++)
7198                 nft_expr_destroy(ctx, expr_array[i]);
7199 err_set_elem_expr_clone:
7200         return err;
7201 }
7202 
7203 static int nf_tables_newsetelem(struct sk_buff *skb,
7204                                 const struct nfnl_info *info,
7205                                 const struct nlattr * const nla[])
7206 {
7207         struct netlink_ext_ack *extack = info->extack;
7208         u8 genmask = nft_genmask_next(info->net);
7209         u8 family = info->nfmsg->nfgen_family;
7210         struct net *net = info->net;
7211         const struct nlattr *attr;
7212         struct nft_table *table;
7213         struct nft_set *set;
7214         struct nft_ctx ctx;
7215         int rem, err;
7216 
7217         if (nla[NFTA_SET_ELEM_LIST_ELEMENTS] == NULL)
7218                 return -EINVAL;
7219 
7220         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
7221                                  genmask, NETLINK_CB(skb).portid);
7222         if (IS_ERR(table)) {
7223                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
7224                 return PTR_ERR(table);
7225         }
7226 
7227         set = nft_set_lookup_global(net, table, nla[NFTA_SET_ELEM_LIST_SET],
7228                                     nla[NFTA_SET_ELEM_LIST_SET_ID], genmask);
7229         if (IS_ERR(set)) {
7230                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_SET]);
7231                 return PTR_ERR(set);
7232         }
7233 
7234         if (!list_empty(&set->bindings) &&
7235             (set->flags & (NFT_SET_CONSTANT | NFT_SET_ANONYMOUS)))
7236                 return -EBUSY;
7237 
7238         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7239 
7240         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
7241                 err = nft_add_set_elem(&ctx, set, attr, info->nlh->nlmsg_flags);
7242                 if (err < 0) {
7243                         NL_SET_BAD_ATTR(extack, attr);
7244                         return err;
7245                 }
7246         }
7247 
7248         if (table->validate_state == NFT_VALIDATE_DO)
7249                 return nft_table_validate(net, table);
7250 
7251         return 0;
7252 }
7253 
7254 /**
7255  *      nft_data_hold - hold a nft_data item
7256  *
7257  *      @data: struct nft_data to release
7258  *      @type: type of data
7259  *
7260  *      Hold a nft_data item. NFT_DATA_VALUE types can be silently discarded,
7261  *      NFT_DATA_VERDICT bumps the reference to chains in case of NFT_JUMP and
7262  *      NFT_GOTO verdicts. This function must be called on active data objects
7263  *      from the second phase of the commit protocol.
7264  */
7265 void nft_data_hold(const struct nft_data *data, enum nft_data_types type)
7266 {
7267         struct nft_chain *chain;
7268 
7269         if (type == NFT_DATA_VERDICT) {
7270                 switch (data->verdict.code) {
7271                 case NFT_JUMP:
7272                 case NFT_GOTO:
7273                         chain = data->verdict.chain;
7274                         nft_use_inc_restore(&chain->use);
7275                         break;
7276                 }
7277         }
7278 }
7279 
7280 static int nft_setelem_active_next(const struct net *net,
7281                                    const struct nft_set *set,
7282                                    struct nft_elem_priv *elem_priv)
7283 {
7284         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
7285         u8 genmask = nft_genmask_next(net);
7286 
7287         return nft_set_elem_active(ext, genmask);
7288 }
7289 
7290 static void nft_setelem_data_activate(const struct net *net,
7291                                       const struct nft_set *set,
7292                                       struct nft_elem_priv *elem_priv)
7293 {
7294         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
7295 
7296         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
7297                 nft_data_hold(nft_set_ext_data(ext), set->dtype);
7298         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
7299                 nft_use_inc_restore(&(*nft_set_ext_obj(ext))->use);
7300 }
7301 
7302 void nft_setelem_data_deactivate(const struct net *net,
7303                                  const struct nft_set *set,
7304                                  struct nft_elem_priv *elem_priv)
7305 {
7306         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
7307 
7308         if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA))
7309                 nft_data_release(nft_set_ext_data(ext), set->dtype);
7310         if (nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF))
7311                 nft_use_dec(&(*nft_set_ext_obj(ext))->use);
7312 }
7313 
7314 static int nft_del_setelem(struct nft_ctx *ctx, struct nft_set *set,
7315                            const struct nlattr *attr)
7316 {
7317         struct nlattr *nla[NFTA_SET_ELEM_MAX + 1];
7318         struct nft_set_ext_tmpl tmpl;
7319         struct nft_set_elem elem;
7320         struct nft_set_ext *ext;
7321         struct nft_trans *trans;
7322         u32 flags = 0;
7323         int err;
7324 
7325         err = nla_parse_nested_deprecated(nla, NFTA_SET_ELEM_MAX, attr,
7326                                           nft_set_elem_policy, NULL);
7327         if (err < 0)
7328                 return err;
7329 
7330         err = nft_setelem_parse_flags(set, nla[NFTA_SET_ELEM_FLAGS], &flags);
7331         if (err < 0)
7332                 return err;
7333 
7334         if (!nla[NFTA_SET_ELEM_KEY] && !(flags & NFT_SET_ELEM_CATCHALL))
7335                 return -EINVAL;
7336 
7337         if (!nft_setelem_valid_key_end(set, nla, flags))
7338                 return -EINVAL;
7339 
7340         nft_set_ext_prepare(&tmpl);
7341 
7342         if (flags != 0) {
7343                 err = nft_set_ext_add(&tmpl, NFT_SET_EXT_FLAGS);
7344                 if (err < 0)
7345                         return err;
7346         }
7347 
7348         if (nla[NFTA_SET_ELEM_KEY]) {
7349                 err = nft_setelem_parse_key(ctx, set, &elem.key.val,
7350                                             nla[NFTA_SET_ELEM_KEY]);
7351                 if (err < 0)
7352                         return err;
7353 
7354                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY, set->klen);
7355                 if (err < 0)
7356                         goto fail_elem;
7357         }
7358 
7359         if (nla[NFTA_SET_ELEM_KEY_END]) {
7360                 err = nft_setelem_parse_key(ctx, set, &elem.key_end.val,
7361                                             nla[NFTA_SET_ELEM_KEY_END]);
7362                 if (err < 0)
7363                         goto fail_elem;
7364 
7365                 err = nft_set_ext_add_length(&tmpl, NFT_SET_EXT_KEY_END, set->klen);
7366                 if (err < 0)
7367                         goto fail_elem_key_end;
7368         }
7369 
7370         err = -ENOMEM;
7371         elem.priv = nft_set_elem_init(set, &tmpl, elem.key.val.data,
7372                                       elem.key_end.val.data, NULL, 0, 0,
7373                                       GFP_KERNEL_ACCOUNT);
7374         if (IS_ERR(elem.priv)) {
7375                 err = PTR_ERR(elem.priv);
7376                 goto fail_elem_key_end;
7377         }
7378 
7379         ext = nft_set_elem_ext(set, elem.priv);
7380         if (flags)
7381                 *nft_set_ext_flags(ext) = flags;
7382 
7383         trans = nft_trans_elem_alloc(ctx, NFT_MSG_DELSETELEM, set);
7384         if (trans == NULL)
7385                 goto fail_trans;
7386 
7387         err = nft_setelem_deactivate(ctx->net, set, &elem, flags);
7388         if (err < 0)
7389                 goto fail_ops;
7390 
7391         nft_setelem_data_deactivate(ctx->net, set, elem.priv);
7392 
7393         nft_trans_elem_priv(trans) = elem.priv;
7394         nft_trans_commit_list_add_tail(ctx->net, trans);
7395         return 0;
7396 
7397 fail_ops:
7398         kfree(trans);
7399 fail_trans:
7400         kfree(elem.priv);
7401 fail_elem_key_end:
7402         nft_data_release(&elem.key_end.val, NFT_DATA_VALUE);
7403 fail_elem:
7404         nft_data_release(&elem.key.val, NFT_DATA_VALUE);
7405         return err;
7406 }
7407 
7408 static int nft_setelem_flush(const struct nft_ctx *ctx,
7409                              struct nft_set *set,
7410                              const struct nft_set_iter *iter,
7411                              struct nft_elem_priv *elem_priv)
7412 {
7413         const struct nft_set_ext *ext = nft_set_elem_ext(set, elem_priv);
7414         struct nft_trans *trans;
7415 
7416         if (!nft_set_elem_active(ext, iter->genmask))
7417                 return 0;
7418 
7419         trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
7420                                     sizeof(struct nft_trans_elem), GFP_ATOMIC);
7421         if (!trans)
7422                 return -ENOMEM;
7423 
7424         set->ops->flush(ctx->net, set, elem_priv);
7425         set->ndeact++;
7426 
7427         nft_setelem_data_deactivate(ctx->net, set, elem_priv);
7428         nft_trans_elem_set(trans) = set;
7429         nft_trans_elem_priv(trans) = elem_priv;
7430         nft_trans_commit_list_add_tail(ctx->net, trans);
7431 
7432         return 0;
7433 }
7434 
7435 static int __nft_set_catchall_flush(const struct nft_ctx *ctx,
7436                                     struct nft_set *set,
7437                                     struct nft_elem_priv *elem_priv)
7438 {
7439         struct nft_trans *trans;
7440 
7441         trans = nft_trans_alloc_gfp(ctx, NFT_MSG_DELSETELEM,
7442                                     sizeof(struct nft_trans_elem), GFP_KERNEL);
7443         if (!trans)
7444                 return -ENOMEM;
7445 
7446         nft_setelem_data_deactivate(ctx->net, set, elem_priv);
7447         nft_trans_elem_set(trans) = set;
7448         nft_trans_elem_priv(trans) = elem_priv;
7449         nft_trans_commit_list_add_tail(ctx->net, trans);
7450 
7451         return 0;
7452 }
7453 
7454 static int nft_set_catchall_flush(const struct nft_ctx *ctx,
7455                                   struct nft_set *set)
7456 {
7457         u8 genmask = nft_genmask_next(ctx->net);
7458         struct nft_set_elem_catchall *catchall;
7459         struct nft_set_ext *ext;
7460         int ret = 0;
7461 
7462         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
7463                 ext = nft_set_elem_ext(set, catchall->elem);
7464                 if (!nft_set_elem_active(ext, genmask))
7465                         continue;
7466 
7467                 ret = __nft_set_catchall_flush(ctx, set, catchall->elem);
7468                 if (ret < 0)
7469                         break;
7470                 nft_set_elem_change_active(ctx->net, set, ext);
7471         }
7472 
7473         return ret;
7474 }
7475 
7476 static int nft_set_flush(struct nft_ctx *ctx, struct nft_set *set, u8 genmask)
7477 {
7478         struct nft_set_iter iter = {
7479                 .genmask        = genmask,
7480                 .type           = NFT_ITER_UPDATE,
7481                 .fn             = nft_setelem_flush,
7482         };
7483 
7484         set->ops->walk(ctx, set, &iter);
7485         if (!iter.err)
7486                 iter.err = nft_set_catchall_flush(ctx, set);
7487 
7488         return iter.err;
7489 }
7490 
7491 static int nf_tables_delsetelem(struct sk_buff *skb,
7492                                 const struct nfnl_info *info,
7493                                 const struct nlattr * const nla[])
7494 {
7495         struct netlink_ext_ack *extack = info->extack;
7496         u8 genmask = nft_genmask_next(info->net);
7497         u8 family = info->nfmsg->nfgen_family;
7498         struct net *net = info->net;
7499         const struct nlattr *attr;
7500         struct nft_table *table;
7501         struct nft_set *set;
7502         struct nft_ctx ctx;
7503         int rem, err = 0;
7504 
7505         table = nft_table_lookup(net, nla[NFTA_SET_ELEM_LIST_TABLE], family,
7506                                  genmask, NETLINK_CB(skb).portid);
7507         if (IS_ERR(table)) {
7508                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_TABLE]);
7509                 return PTR_ERR(table);
7510         }
7511 
7512         set = nft_set_lookup(table, nla[NFTA_SET_ELEM_LIST_SET], genmask);
7513         if (IS_ERR(set)) {
7514                 NL_SET_BAD_ATTR(extack, nla[NFTA_SET_ELEM_LIST_SET]);
7515                 return PTR_ERR(set);
7516         }
7517 
7518         if (nft_set_is_anonymous(set))
7519                 return -EOPNOTSUPP;
7520 
7521         if (!list_empty(&set->bindings) && (set->flags & NFT_SET_CONSTANT))
7522                 return -EBUSY;
7523 
7524         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7525 
7526         if (!nla[NFTA_SET_ELEM_LIST_ELEMENTS])
7527                 return nft_set_flush(&ctx, set, genmask);
7528 
7529         nla_for_each_nested(attr, nla[NFTA_SET_ELEM_LIST_ELEMENTS], rem) {
7530                 err = nft_del_setelem(&ctx, set, attr);
7531                 if (err == -ENOENT &&
7532                     NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYSETELEM)
7533                         continue;
7534 
7535                 if (err < 0) {
7536                         NL_SET_BAD_ATTR(extack, attr);
7537                         return err;
7538                 }
7539         }
7540 
7541         return 0;
7542 }
7543 
7544 /*
7545  * Stateful objects
7546  */
7547 
7548 /**
7549  *      nft_register_obj- register nf_tables stateful object type
7550  *      @obj_type: object type
7551  *
7552  *      Registers the object type for use with nf_tables. Returns zero on
7553  *      success or a negative errno code otherwise.
7554  */
7555 int nft_register_obj(struct nft_object_type *obj_type)
7556 {
7557         if (obj_type->type == NFT_OBJECT_UNSPEC)
7558                 return -EINVAL;
7559 
7560         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7561         list_add_rcu(&obj_type->list, &nf_tables_objects);
7562         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7563         return 0;
7564 }
7565 EXPORT_SYMBOL_GPL(nft_register_obj);
7566 
7567 /**
7568  *      nft_unregister_obj - unregister nf_tables object type
7569  *      @obj_type: object type
7570  *
7571  *      Unregisters the object type for use with nf_tables.
7572  */
7573 void nft_unregister_obj(struct nft_object_type *obj_type)
7574 {
7575         nfnl_lock(NFNL_SUBSYS_NFTABLES);
7576         list_del_rcu(&obj_type->list);
7577         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
7578 }
7579 EXPORT_SYMBOL_GPL(nft_unregister_obj);
7580 
7581 struct nft_object *nft_obj_lookup(const struct net *net,
7582                                   const struct nft_table *table,
7583                                   const struct nlattr *nla, u32 objtype,
7584                                   u8 genmask)
7585 {
7586         struct nft_object_hash_key k = { .table = table };
7587         char search[NFT_OBJ_MAXNAMELEN];
7588         struct rhlist_head *tmp, *list;
7589         struct nft_object *obj;
7590 
7591         nla_strscpy(search, nla, sizeof(search));
7592         k.name = search;
7593 
7594         WARN_ON_ONCE(!rcu_read_lock_held() &&
7595                      !lockdep_commit_lock_is_held(net));
7596 
7597         rcu_read_lock();
7598         list = rhltable_lookup(&nft_objname_ht, &k, nft_objname_ht_params);
7599         if (!list)
7600                 goto out;
7601 
7602         rhl_for_each_entry_rcu(obj, tmp, list, rhlhead) {
7603                 if (objtype == obj->ops->type->type &&
7604                     nft_active_genmask(obj, genmask)) {
7605                         rcu_read_unlock();
7606                         return obj;
7607                 }
7608         }
7609 out:
7610         rcu_read_unlock();
7611         return ERR_PTR(-ENOENT);
7612 }
7613 EXPORT_SYMBOL_GPL(nft_obj_lookup);
7614 
7615 static struct nft_object *nft_obj_lookup_byhandle(const struct nft_table *table,
7616                                                   const struct nlattr *nla,
7617                                                   u32 objtype, u8 genmask)
7618 {
7619         struct nft_object *obj;
7620 
7621         list_for_each_entry(obj, &table->objects, list) {
7622                 if (be64_to_cpu(nla_get_be64(nla)) == obj->handle &&
7623                     objtype == obj->ops->type->type &&
7624                     nft_active_genmask(obj, genmask))
7625                         return obj;
7626         }
7627         return ERR_PTR(-ENOENT);
7628 }
7629 
7630 static const struct nla_policy nft_obj_policy[NFTA_OBJ_MAX + 1] = {
7631         [NFTA_OBJ_TABLE]        = { .type = NLA_STRING,
7632                                     .len = NFT_TABLE_MAXNAMELEN - 1 },
7633         [NFTA_OBJ_NAME]         = { .type = NLA_STRING,
7634                                     .len = NFT_OBJ_MAXNAMELEN - 1 },
7635         [NFTA_OBJ_TYPE]         = { .type = NLA_U32 },
7636         [NFTA_OBJ_DATA]         = { .type = NLA_NESTED },
7637         [NFTA_OBJ_HANDLE]       = { .type = NLA_U64},
7638         [NFTA_OBJ_USERDATA]     = { .type = NLA_BINARY,
7639                                     .len = NFT_USERDATA_MAXLEN },
7640 };
7641 
7642 static struct nft_object *nft_obj_init(const struct nft_ctx *ctx,
7643                                        const struct nft_object_type *type,
7644                                        const struct nlattr *attr)
7645 {
7646         struct nlattr **tb;
7647         const struct nft_object_ops *ops;
7648         struct nft_object *obj;
7649         int err = -ENOMEM;
7650 
7651         tb = kmalloc_array(type->maxattr + 1, sizeof(*tb), GFP_KERNEL);
7652         if (!tb)
7653                 goto err1;
7654 
7655         if (attr) {
7656                 err = nla_parse_nested_deprecated(tb, type->maxattr, attr,
7657                                                   type->policy, NULL);
7658                 if (err < 0)
7659                         goto err2;
7660         } else {
7661                 memset(tb, 0, sizeof(tb[0]) * (type->maxattr + 1));
7662         }
7663 
7664         if (type->select_ops) {
7665                 ops = type->select_ops(ctx, (const struct nlattr * const *)tb);
7666                 if (IS_ERR(ops)) {
7667                         err = PTR_ERR(ops);
7668                         goto err2;
7669                 }
7670         } else {
7671                 ops = type->ops;
7672         }
7673 
7674         err = -ENOMEM;
7675         obj = kzalloc(sizeof(*obj) + ops->size, GFP_KERNEL_ACCOUNT);
7676         if (!obj)
7677                 goto err2;
7678 
7679         err = ops->init(ctx, (const struct nlattr * const *)tb, obj);
7680         if (err < 0)
7681                 goto err3;
7682 
7683         obj->ops = ops;
7684 
7685         kfree(tb);
7686         return obj;
7687 err3:
7688         kfree(obj);
7689 err2:
7690         kfree(tb);
7691 err1:
7692         return ERR_PTR(err);
7693 }
7694 
7695 static int nft_object_dump(struct sk_buff *skb, unsigned int attr,
7696                            struct nft_object *obj, bool reset)
7697 {
7698         struct nlattr *nest;
7699 
7700         nest = nla_nest_start_noflag(skb, attr);
7701         if (!nest)
7702                 goto nla_put_failure;
7703         if (obj->ops->dump(skb, obj, reset) < 0)
7704                 goto nla_put_failure;
7705         nla_nest_end(skb, nest);
7706         return 0;
7707 
7708 nla_put_failure:
7709         return -1;
7710 }
7711 
7712 static const struct nft_object_type *__nft_obj_type_get(u32 objtype, u8 family)
7713 {
7714         const struct nft_object_type *type;
7715 
7716         list_for_each_entry_rcu(type, &nf_tables_objects, list) {
7717                 if (type->family != NFPROTO_UNSPEC &&
7718                     type->family != family)
7719                         continue;
7720 
7721                 if (objtype == type->type)
7722                         return type;
7723         }
7724         return NULL;
7725 }
7726 
7727 static const struct nft_object_type *
7728 nft_obj_type_get(struct net *net, u32 objtype, u8 family)
7729 {
7730         const struct nft_object_type *type;
7731 
7732         rcu_read_lock();
7733         type = __nft_obj_type_get(objtype, family);
7734         if (type != NULL && try_module_get(type->owner)) {
7735                 rcu_read_unlock();
7736                 return type;
7737         }
7738         rcu_read_unlock();
7739 
7740         lockdep_nfnl_nft_mutex_not_held();
7741 #ifdef CONFIG_MODULES
7742         if (type == NULL) {
7743                 if (nft_request_module(net, "nft-obj-%u", objtype) == -EAGAIN)
7744                         return ERR_PTR(-EAGAIN);
7745         }
7746 #endif
7747         return ERR_PTR(-ENOENT);
7748 }
7749 
7750 static int nf_tables_updobj(const struct nft_ctx *ctx,
7751                             const struct nft_object_type *type,
7752                             const struct nlattr *attr,
7753                             struct nft_object *obj)
7754 {
7755         struct nft_object *newobj;
7756         struct nft_trans *trans;
7757         int err = -ENOMEM;
7758 
7759         if (!try_module_get(type->owner))
7760                 return -ENOENT;
7761 
7762         trans = nft_trans_alloc(ctx, NFT_MSG_NEWOBJ,
7763                                 sizeof(struct nft_trans_obj));
7764         if (!trans)
7765                 goto err_trans;
7766 
7767         newobj = nft_obj_init(ctx, type, attr);
7768         if (IS_ERR(newobj)) {
7769                 err = PTR_ERR(newobj);
7770                 goto err_free_trans;
7771         }
7772 
7773         nft_trans_obj(trans) = obj;
7774         nft_trans_obj_update(trans) = true;
7775         nft_trans_obj_newobj(trans) = newobj;
7776         nft_trans_commit_list_add_tail(ctx->net, trans);
7777 
7778         return 0;
7779 
7780 err_free_trans:
7781         kfree(trans);
7782 err_trans:
7783         module_put(type->owner);
7784         return err;
7785 }
7786 
7787 static int nf_tables_newobj(struct sk_buff *skb, const struct nfnl_info *info,
7788                             const struct nlattr * const nla[])
7789 {
7790         struct netlink_ext_ack *extack = info->extack;
7791         u8 genmask = nft_genmask_next(info->net);
7792         u8 family = info->nfmsg->nfgen_family;
7793         const struct nft_object_type *type;
7794         struct net *net = info->net;
7795         struct nft_table *table;
7796         struct nft_object *obj;
7797         struct nft_ctx ctx;
7798         u32 objtype;
7799         int err;
7800 
7801         if (!nla[NFTA_OBJ_TYPE] ||
7802             !nla[NFTA_OBJ_NAME] ||
7803             !nla[NFTA_OBJ_DATA])
7804                 return -EINVAL;
7805 
7806         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
7807                                  NETLINK_CB(skb).portid);
7808         if (IS_ERR(table)) {
7809                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
7810                 return PTR_ERR(table);
7811         }
7812 
7813         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
7814         obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
7815         if (IS_ERR(obj)) {
7816                 err = PTR_ERR(obj);
7817                 if (err != -ENOENT) {
7818                         NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
7819                         return err;
7820                 }
7821         } else {
7822                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
7823                         NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
7824                         return -EEXIST;
7825                 }
7826                 if (info->nlh->nlmsg_flags & NLM_F_REPLACE)
7827                         return -EOPNOTSUPP;
7828 
7829                 type = __nft_obj_type_get(objtype, family);
7830                 if (WARN_ON_ONCE(!type))
7831                         return -ENOENT;
7832 
7833                 if (!obj->ops->update)
7834                         return 0;
7835 
7836                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7837 
7838                 return nf_tables_updobj(&ctx, type, nla[NFTA_OBJ_DATA], obj);
7839         }
7840 
7841         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
7842 
7843         if (!nft_use_inc(&table->use))
7844                 return -EMFILE;
7845 
7846         type = nft_obj_type_get(net, objtype, family);
7847         if (IS_ERR(type)) {
7848                 err = PTR_ERR(type);
7849                 goto err_type;
7850         }
7851 
7852         obj = nft_obj_init(&ctx, type, nla[NFTA_OBJ_DATA]);
7853         if (IS_ERR(obj)) {
7854                 err = PTR_ERR(obj);
7855                 goto err_init;
7856         }
7857         obj->key.table = table;
7858         obj->handle = nf_tables_alloc_handle(table);
7859 
7860         obj->key.name = nla_strdup(nla[NFTA_OBJ_NAME], GFP_KERNEL_ACCOUNT);
7861         if (!obj->key.name) {
7862                 err = -ENOMEM;
7863                 goto err_strdup;
7864         }
7865 
7866         if (nla[NFTA_OBJ_USERDATA]) {
7867                 obj->udata = nla_memdup(nla[NFTA_OBJ_USERDATA], GFP_KERNEL_ACCOUNT);
7868                 if (obj->udata == NULL)
7869                         goto err_userdata;
7870 
7871                 obj->udlen = nla_len(nla[NFTA_OBJ_USERDATA]);
7872         }
7873 
7874         err = nft_trans_obj_add(&ctx, NFT_MSG_NEWOBJ, obj);
7875         if (err < 0)
7876                 goto err_trans;
7877 
7878         err = rhltable_insert(&nft_objname_ht, &obj->rhlhead,
7879                               nft_objname_ht_params);
7880         if (err < 0)
7881                 goto err_obj_ht;
7882 
7883         list_add_tail_rcu(&obj->list, &table->objects);
7884 
7885         return 0;
7886 err_obj_ht:
7887         /* queued in transaction log */
7888         INIT_LIST_HEAD(&obj->list);
7889         return err;
7890 err_trans:
7891         kfree(obj->udata);
7892 err_userdata:
7893         kfree(obj->key.name);
7894 err_strdup:
7895         if (obj->ops->destroy)
7896                 obj->ops->destroy(&ctx, obj);
7897         kfree(obj);
7898 err_init:
7899         module_put(type->owner);
7900 err_type:
7901         nft_use_dec_restore(&table->use);
7902 
7903         return err;
7904 }
7905 
7906 static int nf_tables_fill_obj_info(struct sk_buff *skb, struct net *net,
7907                                    u32 portid, u32 seq, int event, u32 flags,
7908                                    int family, const struct nft_table *table,
7909                                    struct nft_object *obj, bool reset)
7910 {
7911         struct nlmsghdr *nlh;
7912 
7913         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
7914         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
7915                            NFNETLINK_V0, nft_base_seq(net));
7916         if (!nlh)
7917                 goto nla_put_failure;
7918 
7919         if (nla_put_string(skb, NFTA_OBJ_TABLE, table->name) ||
7920             nla_put_string(skb, NFTA_OBJ_NAME, obj->key.name) ||
7921             nla_put_be64(skb, NFTA_OBJ_HANDLE, cpu_to_be64(obj->handle),
7922                          NFTA_OBJ_PAD))
7923                 goto nla_put_failure;
7924 
7925         if (event == NFT_MSG_DELOBJ) {
7926                 nlmsg_end(skb, nlh);
7927                 return 0;
7928         }
7929 
7930         if (nla_put_be32(skb, NFTA_OBJ_TYPE, htonl(obj->ops->type->type)) ||
7931             nla_put_be32(skb, NFTA_OBJ_USE, htonl(obj->use)) ||
7932             nft_object_dump(skb, NFTA_OBJ_DATA, obj, reset))
7933                 goto nla_put_failure;
7934 
7935         if (obj->udata &&
7936             nla_put(skb, NFTA_OBJ_USERDATA, obj->udlen, obj->udata))
7937                 goto nla_put_failure;
7938 
7939         nlmsg_end(skb, nlh);
7940         return 0;
7941 
7942 nla_put_failure:
7943         nlmsg_trim(skb, nlh);
7944         return -1;
7945 }
7946 
7947 static void audit_log_obj_reset(const struct nft_table *table,
7948                                 unsigned int base_seq, unsigned int nentries)
7949 {
7950         char *buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, base_seq);
7951 
7952         audit_log_nfcfg(buf, table->family, nentries,
7953                         AUDIT_NFT_OP_OBJ_RESET, GFP_ATOMIC);
7954         kfree(buf);
7955 }
7956 
7957 struct nft_obj_dump_ctx {
7958         unsigned int    s_idx;
7959         char            *table;
7960         u32             type;
7961         bool            reset;
7962 };
7963 
7964 static int nf_tables_dump_obj(struct sk_buff *skb, struct netlink_callback *cb)
7965 {
7966         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
7967         struct nft_obj_dump_ctx *ctx = (void *)cb->ctx;
7968         struct net *net = sock_net(skb->sk);
7969         int family = nfmsg->nfgen_family;
7970         struct nftables_pernet *nft_net;
7971         const struct nft_table *table;
7972         unsigned int entries = 0;
7973         struct nft_object *obj;
7974         unsigned int idx = 0;
7975         int rc = 0;
7976 
7977         rcu_read_lock();
7978         nft_net = nft_pernet(net);
7979         cb->seq = READ_ONCE(nft_net->base_seq);
7980 
7981         list_for_each_entry_rcu(table, &nft_net->tables, list) {
7982                 if (family != NFPROTO_UNSPEC && family != table->family)
7983                         continue;
7984 
7985                 entries = 0;
7986                 list_for_each_entry_rcu(obj, &table->objects, list) {
7987                         if (!nft_is_active(net, obj))
7988                                 goto cont;
7989                         if (idx < ctx->s_idx)
7990                                 goto cont;
7991                         if (ctx->table && strcmp(ctx->table, table->name))
7992                                 goto cont;
7993                         if (ctx->type != NFT_OBJECT_UNSPEC &&
7994                             obj->ops->type->type != ctx->type)
7995                                 goto cont;
7996 
7997                         rc = nf_tables_fill_obj_info(skb, net,
7998                                                      NETLINK_CB(cb->skb).portid,
7999                                                      cb->nlh->nlmsg_seq,
8000                                                      NFT_MSG_NEWOBJ,
8001                                                      NLM_F_MULTI | NLM_F_APPEND,
8002                                                      table->family, table,
8003                                                      obj, ctx->reset);
8004                         if (rc < 0)
8005                                 break;
8006 
8007                         entries++;
8008                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
8009 cont:
8010                         idx++;
8011                 }
8012                 if (ctx->reset && entries)
8013                         audit_log_obj_reset(table, nft_net->base_seq, entries);
8014                 if (rc < 0)
8015                         break;
8016         }
8017         rcu_read_unlock();
8018 
8019         ctx->s_idx = idx;
8020         return skb->len;
8021 }
8022 
8023 static int nf_tables_dump_obj_start(struct netlink_callback *cb)
8024 {
8025         struct nft_obj_dump_ctx *ctx = (void *)cb->ctx;
8026         const struct nlattr * const *nla = cb->data;
8027 
8028         BUILD_BUG_ON(sizeof(*ctx) > sizeof(cb->ctx));
8029 
8030         if (nla[NFTA_OBJ_TABLE]) {
8031                 ctx->table = nla_strdup(nla[NFTA_OBJ_TABLE], GFP_ATOMIC);
8032                 if (!ctx->table)
8033                         return -ENOMEM;
8034         }
8035 
8036         if (nla[NFTA_OBJ_TYPE])
8037                 ctx->type = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
8038 
8039         if (NFNL_MSG_TYPE(cb->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
8040                 ctx->reset = true;
8041 
8042         return 0;
8043 }
8044 
8045 static int nf_tables_dump_obj_done(struct netlink_callback *cb)
8046 {
8047         struct nft_obj_dump_ctx *ctx = (void *)cb->ctx;
8048 
8049         kfree(ctx->table);
8050 
8051         return 0;
8052 }
8053 
8054 /* called with rcu_read_lock held */
8055 static int nf_tables_getobj(struct sk_buff *skb, const struct nfnl_info *info,
8056                             const struct nlattr * const nla[])
8057 {
8058         struct netlink_ext_ack *extack = info->extack;
8059         u8 genmask = nft_genmask_cur(info->net);
8060         u8 family = info->nfmsg->nfgen_family;
8061         const struct nft_table *table;
8062         struct net *net = info->net;
8063         struct nft_object *obj;
8064         struct sk_buff *skb2;
8065         bool reset = false;
8066         u32 objtype;
8067         int err;
8068 
8069         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
8070                 struct netlink_dump_control c = {
8071                         .start = nf_tables_dump_obj_start,
8072                         .dump = nf_tables_dump_obj,
8073                         .done = nf_tables_dump_obj_done,
8074                         .module = THIS_MODULE,
8075                         .data = (void *)nla,
8076                 };
8077 
8078                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
8079         }
8080 
8081         if (!nla[NFTA_OBJ_NAME] ||
8082             !nla[NFTA_OBJ_TYPE])
8083                 return -EINVAL;
8084 
8085         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask, 0);
8086         if (IS_ERR(table)) {
8087                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
8088                 return PTR_ERR(table);
8089         }
8090 
8091         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
8092         obj = nft_obj_lookup(net, table, nla[NFTA_OBJ_NAME], objtype, genmask);
8093         if (IS_ERR(obj)) {
8094                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_NAME]);
8095                 return PTR_ERR(obj);
8096         }
8097 
8098         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
8099         if (!skb2)
8100                 return -ENOMEM;
8101 
8102         if (NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_GETOBJ_RESET)
8103                 reset = true;
8104 
8105         if (reset) {
8106                 const struct nftables_pernet *nft_net;
8107                 char *buf;
8108 
8109                 nft_net = nft_pernet(net);
8110                 buf = kasprintf(GFP_ATOMIC, "%s:%u", table->name, nft_net->base_seq);
8111 
8112                 audit_log_nfcfg(buf,
8113                                 family,
8114                                 1,
8115                                 AUDIT_NFT_OP_OBJ_RESET,
8116                                 GFP_ATOMIC);
8117                 kfree(buf);
8118         }
8119 
8120         err = nf_tables_fill_obj_info(skb2, net, NETLINK_CB(skb).portid,
8121                                       info->nlh->nlmsg_seq, NFT_MSG_NEWOBJ, 0,
8122                                       family, table, obj, reset);
8123         if (err < 0)
8124                 goto err_fill_obj_info;
8125 
8126         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
8127 
8128 err_fill_obj_info:
8129         kfree_skb(skb2);
8130         return err;
8131 }
8132 
8133 static void nft_obj_destroy(const struct nft_ctx *ctx, struct nft_object *obj)
8134 {
8135         if (obj->ops->destroy)
8136                 obj->ops->destroy(ctx, obj);
8137 
8138         module_put(obj->ops->type->owner);
8139         kfree(obj->key.name);
8140         kfree(obj->udata);
8141         kfree(obj);
8142 }
8143 
8144 static int nf_tables_delobj(struct sk_buff *skb, const struct nfnl_info *info,
8145                             const struct nlattr * const nla[])
8146 {
8147         struct netlink_ext_ack *extack = info->extack;
8148         u8 genmask = nft_genmask_next(info->net);
8149         u8 family = info->nfmsg->nfgen_family;
8150         struct net *net = info->net;
8151         const struct nlattr *attr;
8152         struct nft_table *table;
8153         struct nft_object *obj;
8154         struct nft_ctx ctx;
8155         u32 objtype;
8156 
8157         if (!nla[NFTA_OBJ_TYPE] ||
8158             (!nla[NFTA_OBJ_NAME] && !nla[NFTA_OBJ_HANDLE]))
8159                 return -EINVAL;
8160 
8161         table = nft_table_lookup(net, nla[NFTA_OBJ_TABLE], family, genmask,
8162                                  NETLINK_CB(skb).portid);
8163         if (IS_ERR(table)) {
8164                 NL_SET_BAD_ATTR(extack, nla[NFTA_OBJ_TABLE]);
8165                 return PTR_ERR(table);
8166         }
8167 
8168         objtype = ntohl(nla_get_be32(nla[NFTA_OBJ_TYPE]));
8169         if (nla[NFTA_OBJ_HANDLE]) {
8170                 attr = nla[NFTA_OBJ_HANDLE];
8171                 obj = nft_obj_lookup_byhandle(table, attr, objtype, genmask);
8172         } else {
8173                 attr = nla[NFTA_OBJ_NAME];
8174                 obj = nft_obj_lookup(net, table, attr, objtype, genmask);
8175         }
8176 
8177         if (IS_ERR(obj)) {
8178                 if (PTR_ERR(obj) == -ENOENT &&
8179                     NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYOBJ)
8180                         return 0;
8181 
8182                 NL_SET_BAD_ATTR(extack, attr);
8183                 return PTR_ERR(obj);
8184         }
8185         if (obj->use > 0) {
8186                 NL_SET_BAD_ATTR(extack, attr);
8187                 return -EBUSY;
8188         }
8189 
8190         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
8191 
8192         return nft_delobj(&ctx, obj);
8193 }
8194 
8195 static void
8196 __nft_obj_notify(struct net *net, const struct nft_table *table,
8197                  struct nft_object *obj, u32 portid, u32 seq, int event,
8198                  u16 flags, int family, int report, gfp_t gfp)
8199 {
8200         struct nftables_pernet *nft_net = nft_pernet(net);
8201         struct sk_buff *skb;
8202         int err;
8203 
8204         if (!report &&
8205             !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
8206                 return;
8207 
8208         skb = nlmsg_new(NLMSG_GOODSIZE, gfp);
8209         if (skb == NULL)
8210                 goto err;
8211 
8212         err = nf_tables_fill_obj_info(skb, net, portid, seq, event,
8213                                       flags & (NLM_F_CREATE | NLM_F_EXCL),
8214                                       family, table, obj, false);
8215         if (err < 0) {
8216                 kfree_skb(skb);
8217                 goto err;
8218         }
8219 
8220         nft_notify_enqueue(skb, report, &nft_net->notify_list);
8221         return;
8222 err:
8223         nfnetlink_set_err(net, portid, NFNLGRP_NFTABLES, -ENOBUFS);
8224 }
8225 
8226 void nft_obj_notify(struct net *net, const struct nft_table *table,
8227                     struct nft_object *obj, u32 portid, u32 seq, int event,
8228                     u16 flags, int family, int report, gfp_t gfp)
8229 {
8230         struct nftables_pernet *nft_net = nft_pernet(net);
8231         char *buf = kasprintf(gfp, "%s:%u",
8232                               table->name, nft_net->base_seq);
8233 
8234         audit_log_nfcfg(buf,
8235                         family,
8236                         obj->handle,
8237                         event == NFT_MSG_NEWOBJ ?
8238                                  AUDIT_NFT_OP_OBJ_REGISTER :
8239                                  AUDIT_NFT_OP_OBJ_UNREGISTER,
8240                         gfp);
8241         kfree(buf);
8242 
8243         __nft_obj_notify(net, table, obj, portid, seq, event,
8244                          flags, family, report, gfp);
8245 }
8246 EXPORT_SYMBOL_GPL(nft_obj_notify);
8247 
8248 static void nf_tables_obj_notify(const struct nft_ctx *ctx,
8249                                  struct nft_object *obj, int event)
8250 {
8251         __nft_obj_notify(ctx->net, ctx->table, obj, ctx->portid,
8252                          ctx->seq, event, ctx->flags, ctx->family,
8253                          ctx->report, GFP_KERNEL);
8254 }
8255 
8256 /*
8257  * Flow tables
8258  */
8259 void nft_register_flowtable_type(struct nf_flowtable_type *type)
8260 {
8261         nfnl_lock(NFNL_SUBSYS_NFTABLES);
8262         list_add_tail_rcu(&type->list, &nf_tables_flowtables);
8263         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
8264 }
8265 EXPORT_SYMBOL_GPL(nft_register_flowtable_type);
8266 
8267 void nft_unregister_flowtable_type(struct nf_flowtable_type *type)
8268 {
8269         nfnl_lock(NFNL_SUBSYS_NFTABLES);
8270         list_del_rcu(&type->list);
8271         nfnl_unlock(NFNL_SUBSYS_NFTABLES);
8272 }
8273 EXPORT_SYMBOL_GPL(nft_unregister_flowtable_type);
8274 
8275 static const struct nla_policy nft_flowtable_policy[NFTA_FLOWTABLE_MAX + 1] = {
8276         [NFTA_FLOWTABLE_TABLE]          = { .type = NLA_STRING,
8277                                             .len = NFT_NAME_MAXLEN - 1 },
8278         [NFTA_FLOWTABLE_NAME]           = { .type = NLA_STRING,
8279                                             .len = NFT_NAME_MAXLEN - 1 },
8280         [NFTA_FLOWTABLE_HOOK]           = { .type = NLA_NESTED },
8281         [NFTA_FLOWTABLE_HANDLE]         = { .type = NLA_U64 },
8282         [NFTA_FLOWTABLE_FLAGS]          = { .type = NLA_U32 },
8283 };
8284 
8285 struct nft_flowtable *nft_flowtable_lookup(const struct nft_table *table,
8286                                            const struct nlattr *nla, u8 genmask)
8287 {
8288         struct nft_flowtable *flowtable;
8289 
8290         list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
8291                 if (!nla_strcmp(nla, flowtable->name) &&
8292                     nft_active_genmask(flowtable, genmask))
8293                         return flowtable;
8294         }
8295         return ERR_PTR(-ENOENT);
8296 }
8297 EXPORT_SYMBOL_GPL(nft_flowtable_lookup);
8298 
8299 void nf_tables_deactivate_flowtable(const struct nft_ctx *ctx,
8300                                     struct nft_flowtable *flowtable,
8301                                     enum nft_trans_phase phase)
8302 {
8303         switch (phase) {
8304         case NFT_TRANS_PREPARE_ERROR:
8305         case NFT_TRANS_PREPARE:
8306         case NFT_TRANS_ABORT:
8307         case NFT_TRANS_RELEASE:
8308                 nft_use_dec(&flowtable->use);
8309                 fallthrough;
8310         default:
8311                 return;
8312         }
8313 }
8314 EXPORT_SYMBOL_GPL(nf_tables_deactivate_flowtable);
8315 
8316 static struct nft_flowtable *
8317 nft_flowtable_lookup_byhandle(const struct nft_table *table,
8318                               const struct nlattr *nla, u8 genmask)
8319 {
8320        struct nft_flowtable *flowtable;
8321 
8322        list_for_each_entry(flowtable, &table->flowtables, list) {
8323                if (be64_to_cpu(nla_get_be64(nla)) == flowtable->handle &&
8324                    nft_active_genmask(flowtable, genmask))
8325                        return flowtable;
8326        }
8327        return ERR_PTR(-ENOENT);
8328 }
8329 
8330 struct nft_flowtable_hook {
8331         u32                     num;
8332         int                     priority;
8333         struct list_head        list;
8334 };
8335 
8336 static const struct nla_policy nft_flowtable_hook_policy[NFTA_FLOWTABLE_HOOK_MAX + 1] = {
8337         [NFTA_FLOWTABLE_HOOK_NUM]       = { .type = NLA_U32 },
8338         [NFTA_FLOWTABLE_HOOK_PRIORITY]  = { .type = NLA_U32 },
8339         [NFTA_FLOWTABLE_HOOK_DEVS]      = { .type = NLA_NESTED },
8340 };
8341 
8342 static int nft_flowtable_parse_hook(const struct nft_ctx *ctx,
8343                                     const struct nlattr * const nla[],
8344                                     struct nft_flowtable_hook *flowtable_hook,
8345                                     struct nft_flowtable *flowtable,
8346                                     struct netlink_ext_ack *extack, bool add)
8347 {
8348         struct nlattr *tb[NFTA_FLOWTABLE_HOOK_MAX + 1];
8349         struct nft_hook *hook;
8350         int hooknum, priority;
8351         int err;
8352 
8353         INIT_LIST_HEAD(&flowtable_hook->list);
8354 
8355         err = nla_parse_nested_deprecated(tb, NFTA_FLOWTABLE_HOOK_MAX,
8356                                           nla[NFTA_FLOWTABLE_HOOK],
8357                                           nft_flowtable_hook_policy, NULL);
8358         if (err < 0)
8359                 return err;
8360 
8361         if (add) {
8362                 if (!tb[NFTA_FLOWTABLE_HOOK_NUM] ||
8363                     !tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
8364                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
8365                         return -ENOENT;
8366                 }
8367 
8368                 hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
8369                 if (hooknum != NF_NETDEV_INGRESS)
8370                         return -EOPNOTSUPP;
8371 
8372                 priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
8373 
8374                 flowtable_hook->priority        = priority;
8375                 flowtable_hook->num             = hooknum;
8376         } else {
8377                 if (tb[NFTA_FLOWTABLE_HOOK_NUM]) {
8378                         hooknum = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_NUM]));
8379                         if (hooknum != flowtable->hooknum)
8380                                 return -EOPNOTSUPP;
8381                 }
8382 
8383                 if (tb[NFTA_FLOWTABLE_HOOK_PRIORITY]) {
8384                         priority = ntohl(nla_get_be32(tb[NFTA_FLOWTABLE_HOOK_PRIORITY]));
8385                         if (priority != flowtable->data.priority)
8386                                 return -EOPNOTSUPP;
8387                 }
8388 
8389                 flowtable_hook->priority        = flowtable->data.priority;
8390                 flowtable_hook->num             = flowtable->hooknum;
8391         }
8392 
8393         if (tb[NFTA_FLOWTABLE_HOOK_DEVS]) {
8394                 err = nf_tables_parse_netdev_hooks(ctx->net,
8395                                                    tb[NFTA_FLOWTABLE_HOOK_DEVS],
8396                                                    &flowtable_hook->list,
8397                                                    extack);
8398                 if (err < 0)
8399                         return err;
8400         }
8401 
8402         list_for_each_entry(hook, &flowtable_hook->list, list) {
8403                 hook->ops.pf            = NFPROTO_NETDEV;
8404                 hook->ops.hooknum       = flowtable_hook->num;
8405                 hook->ops.priority      = flowtable_hook->priority;
8406                 hook->ops.priv          = &flowtable->data;
8407                 hook->ops.hook          = flowtable->data.type->hook;
8408         }
8409 
8410         return err;
8411 }
8412 
8413 /* call under rcu_read_lock */
8414 static const struct nf_flowtable_type *__nft_flowtable_type_get(u8 family)
8415 {
8416         const struct nf_flowtable_type *type;
8417 
8418         list_for_each_entry_rcu(type, &nf_tables_flowtables, list) {
8419                 if (family == type->family)
8420                         return type;
8421         }
8422         return NULL;
8423 }
8424 
8425 static const struct nf_flowtable_type *
8426 nft_flowtable_type_get(struct net *net, u8 family)
8427 {
8428         const struct nf_flowtable_type *type;
8429 
8430         rcu_read_lock();
8431         type = __nft_flowtable_type_get(family);
8432         if (type != NULL && try_module_get(type->owner)) {
8433                 rcu_read_unlock();
8434                 return type;
8435         }
8436         rcu_read_unlock();
8437 
8438         lockdep_nfnl_nft_mutex_not_held();
8439 #ifdef CONFIG_MODULES
8440         if (type == NULL) {
8441                 if (nft_request_module(net, "nf-flowtable-%u", family) == -EAGAIN)
8442                         return ERR_PTR(-EAGAIN);
8443         }
8444 #endif
8445         return ERR_PTR(-ENOENT);
8446 }
8447 
8448 /* Only called from error and netdev event paths. */
8449 static void nft_unregister_flowtable_hook(struct net *net,
8450                                           struct nft_flowtable *flowtable,
8451                                           struct nft_hook *hook)
8452 {
8453         nf_unregister_net_hook(net, &hook->ops);
8454         flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
8455                                     FLOW_BLOCK_UNBIND);
8456 }
8457 
8458 static void __nft_unregister_flowtable_net_hooks(struct net *net,
8459                                                  struct list_head *hook_list,
8460                                                  bool release_netdev)
8461 {
8462         struct nft_hook *hook, *next;
8463 
8464         list_for_each_entry_safe(hook, next, hook_list, list) {
8465                 nf_unregister_net_hook(net, &hook->ops);
8466                 if (release_netdev) {
8467                         list_del(&hook->list);
8468                         kfree_rcu(hook, rcu);
8469                 }
8470         }
8471 }
8472 
8473 static void nft_unregister_flowtable_net_hooks(struct net *net,
8474                                                struct list_head *hook_list)
8475 {
8476         __nft_unregister_flowtable_net_hooks(net, hook_list, false);
8477 }
8478 
8479 static int nft_register_flowtable_net_hooks(struct net *net,
8480                                             struct nft_table *table,
8481                                             struct list_head *hook_list,
8482                                             struct nft_flowtable *flowtable)
8483 {
8484         struct nft_hook *hook, *hook2, *next;
8485         struct nft_flowtable *ft;
8486         int err, i = 0;
8487 
8488         list_for_each_entry(hook, hook_list, list) {
8489                 list_for_each_entry(ft, &table->flowtables, list) {
8490                         if (!nft_is_active_next(net, ft))
8491                                 continue;
8492 
8493                         list_for_each_entry(hook2, &ft->hook_list, list) {
8494                                 if (hook->ops.dev == hook2->ops.dev &&
8495                                     hook->ops.pf == hook2->ops.pf) {
8496                                         err = -EEXIST;
8497                                         goto err_unregister_net_hooks;
8498                                 }
8499                         }
8500                 }
8501 
8502                 err = flowtable->data.type->setup(&flowtable->data,
8503                                                   hook->ops.dev,
8504                                                   FLOW_BLOCK_BIND);
8505                 if (err < 0)
8506                         goto err_unregister_net_hooks;
8507 
8508                 err = nf_register_net_hook(net, &hook->ops);
8509                 if (err < 0) {
8510                         flowtable->data.type->setup(&flowtable->data,
8511                                                     hook->ops.dev,
8512                                                     FLOW_BLOCK_UNBIND);
8513                         goto err_unregister_net_hooks;
8514                 }
8515 
8516                 i++;
8517         }
8518 
8519         return 0;
8520 
8521 err_unregister_net_hooks:
8522         list_for_each_entry_safe(hook, next, hook_list, list) {
8523                 if (i-- <= 0)
8524                         break;
8525 
8526                 nft_unregister_flowtable_hook(net, flowtable, hook);
8527                 list_del_rcu(&hook->list);
8528                 kfree_rcu(hook, rcu);
8529         }
8530 
8531         return err;
8532 }
8533 
8534 static void nft_hooks_destroy(struct list_head *hook_list)
8535 {
8536         struct nft_hook *hook, *next;
8537 
8538         list_for_each_entry_safe(hook, next, hook_list, list) {
8539                 list_del_rcu(&hook->list);
8540                 kfree_rcu(hook, rcu);
8541         }
8542 }
8543 
8544 static int nft_flowtable_update(struct nft_ctx *ctx, const struct nlmsghdr *nlh,
8545                                 struct nft_flowtable *flowtable,
8546                                 struct netlink_ext_ack *extack)
8547 {
8548         const struct nlattr * const *nla = ctx->nla;
8549         struct nft_flowtable_hook flowtable_hook;
8550         struct nft_hook *hook, *next;
8551         struct nft_trans *trans;
8552         bool unregister = false;
8553         u32 flags;
8554         int err;
8555 
8556         err = nft_flowtable_parse_hook(ctx, nla, &flowtable_hook, flowtable,
8557                                        extack, false);
8558         if (err < 0)
8559                 return err;
8560 
8561         list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
8562                 if (nft_hook_list_find(&flowtable->hook_list, hook)) {
8563                         list_del(&hook->list);
8564                         kfree(hook);
8565                 }
8566         }
8567 
8568         if (nla[NFTA_FLOWTABLE_FLAGS]) {
8569                 flags = ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
8570                 if (flags & ~NFT_FLOWTABLE_MASK) {
8571                         err = -EOPNOTSUPP;
8572                         goto err_flowtable_update_hook;
8573                 }
8574                 if ((flowtable->data.flags & NFT_FLOWTABLE_HW_OFFLOAD) ^
8575                     (flags & NFT_FLOWTABLE_HW_OFFLOAD)) {
8576                         err = -EOPNOTSUPP;
8577                         goto err_flowtable_update_hook;
8578                 }
8579         } else {
8580                 flags = flowtable->data.flags;
8581         }
8582 
8583         err = nft_register_flowtable_net_hooks(ctx->net, ctx->table,
8584                                                &flowtable_hook.list, flowtable);
8585         if (err < 0)
8586                 goto err_flowtable_update_hook;
8587 
8588         trans = nft_trans_alloc(ctx, NFT_MSG_NEWFLOWTABLE,
8589                                 sizeof(struct nft_trans_flowtable));
8590         if (!trans) {
8591                 unregister = true;
8592                 err = -ENOMEM;
8593                 goto err_flowtable_update_hook;
8594         }
8595 
8596         nft_trans_flowtable_flags(trans) = flags;
8597         nft_trans_flowtable(trans) = flowtable;
8598         nft_trans_flowtable_update(trans) = true;
8599         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
8600         list_splice(&flowtable_hook.list, &nft_trans_flowtable_hooks(trans));
8601 
8602         nft_trans_commit_list_add_tail(ctx->net, trans);
8603 
8604         return 0;
8605 
8606 err_flowtable_update_hook:
8607         list_for_each_entry_safe(hook, next, &flowtable_hook.list, list) {
8608                 if (unregister)
8609                         nft_unregister_flowtable_hook(ctx->net, flowtable, hook);
8610                 list_del_rcu(&hook->list);
8611                 kfree_rcu(hook, rcu);
8612         }
8613 
8614         return err;
8615 
8616 }
8617 
8618 static int nf_tables_newflowtable(struct sk_buff *skb,
8619                                   const struct nfnl_info *info,
8620                                   const struct nlattr * const nla[])
8621 {
8622         struct netlink_ext_ack *extack = info->extack;
8623         struct nft_flowtable_hook flowtable_hook;
8624         u8 genmask = nft_genmask_next(info->net);
8625         u8 family = info->nfmsg->nfgen_family;
8626         const struct nf_flowtable_type *type;
8627         struct nft_flowtable *flowtable;
8628         struct net *net = info->net;
8629         struct nft_table *table;
8630         struct nft_trans *trans;
8631         struct nft_ctx ctx;
8632         int err;
8633 
8634         if (!nla[NFTA_FLOWTABLE_TABLE] ||
8635             !nla[NFTA_FLOWTABLE_NAME] ||
8636             !nla[NFTA_FLOWTABLE_HOOK])
8637                 return -EINVAL;
8638 
8639         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
8640                                  genmask, NETLINK_CB(skb).portid);
8641         if (IS_ERR(table)) {
8642                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
8643                 return PTR_ERR(table);
8644         }
8645 
8646         flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
8647                                          genmask);
8648         if (IS_ERR(flowtable)) {
8649                 err = PTR_ERR(flowtable);
8650                 if (err != -ENOENT) {
8651                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
8652                         return err;
8653                 }
8654         } else {
8655                 if (info->nlh->nlmsg_flags & NLM_F_EXCL) {
8656                         NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
8657                         return -EEXIST;
8658                 }
8659 
8660                 nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
8661 
8662                 return nft_flowtable_update(&ctx, info->nlh, flowtable, extack);
8663         }
8664 
8665         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
8666 
8667         if (!nft_use_inc(&table->use))
8668                 return -EMFILE;
8669 
8670         flowtable = kzalloc(sizeof(*flowtable), GFP_KERNEL_ACCOUNT);
8671         if (!flowtable) {
8672                 err = -ENOMEM;
8673                 goto flowtable_alloc;
8674         }
8675 
8676         flowtable->table = table;
8677         flowtable->handle = nf_tables_alloc_handle(table);
8678         INIT_LIST_HEAD(&flowtable->hook_list);
8679 
8680         flowtable->name = nla_strdup(nla[NFTA_FLOWTABLE_NAME], GFP_KERNEL_ACCOUNT);
8681         if (!flowtable->name) {
8682                 err = -ENOMEM;
8683                 goto err1;
8684         }
8685 
8686         type = nft_flowtable_type_get(net, family);
8687         if (IS_ERR(type)) {
8688                 err = PTR_ERR(type);
8689                 goto err2;
8690         }
8691 
8692         if (nla[NFTA_FLOWTABLE_FLAGS]) {
8693                 flowtable->data.flags =
8694                         ntohl(nla_get_be32(nla[NFTA_FLOWTABLE_FLAGS]));
8695                 if (flowtable->data.flags & ~NFT_FLOWTABLE_MASK) {
8696                         err = -EOPNOTSUPP;
8697                         goto err3;
8698                 }
8699         }
8700 
8701         write_pnet(&flowtable->data.net, net);
8702         flowtable->data.type = type;
8703         err = type->init(&flowtable->data);
8704         if (err < 0)
8705                 goto err3;
8706 
8707         err = nft_flowtable_parse_hook(&ctx, nla, &flowtable_hook, flowtable,
8708                                        extack, true);
8709         if (err < 0)
8710                 goto err_flowtable_parse_hooks;
8711 
8712         list_splice(&flowtable_hook.list, &flowtable->hook_list);
8713         flowtable->data.priority = flowtable_hook.priority;
8714         flowtable->hooknum = flowtable_hook.num;
8715 
8716         trans = nft_trans_flowtable_add(&ctx, NFT_MSG_NEWFLOWTABLE, flowtable);
8717         if (IS_ERR(trans)) {
8718                 err = PTR_ERR(trans);
8719                 goto err_flowtable_trans;
8720         }
8721 
8722         /* This must be LAST to ensure no packets are walking over this flowtable. */
8723         err = nft_register_flowtable_net_hooks(ctx.net, table,
8724                                                &flowtable->hook_list,
8725                                                flowtable);
8726         if (err < 0)
8727                 goto err_flowtable_hooks;
8728 
8729         list_add_tail_rcu(&flowtable->list, &table->flowtables);
8730 
8731         return 0;
8732 
8733 err_flowtable_hooks:
8734         nft_trans_destroy(trans);
8735 err_flowtable_trans:
8736         nft_hooks_destroy(&flowtable->hook_list);
8737 err_flowtable_parse_hooks:
8738         flowtable->data.type->free(&flowtable->data);
8739 err3:
8740         module_put(type->owner);
8741 err2:
8742         kfree(flowtable->name);
8743 err1:
8744         kfree(flowtable);
8745 flowtable_alloc:
8746         nft_use_dec_restore(&table->use);
8747 
8748         return err;
8749 }
8750 
8751 static void nft_flowtable_hook_release(struct nft_flowtable_hook *flowtable_hook)
8752 {
8753         struct nft_hook *this, *next;
8754 
8755         list_for_each_entry_safe(this, next, &flowtable_hook->list, list) {
8756                 list_del(&this->list);
8757                 kfree(this);
8758         }
8759 }
8760 
8761 static int nft_delflowtable_hook(struct nft_ctx *ctx,
8762                                  struct nft_flowtable *flowtable,
8763                                  struct netlink_ext_ack *extack)
8764 {
8765         const struct nlattr * const *nla = ctx->nla;
8766         struct nft_flowtable_hook flowtable_hook;
8767         LIST_HEAD(flowtable_del_list);
8768         struct nft_hook *this, *hook;
8769         struct nft_trans *trans;
8770         int err;
8771 
8772         err = nft_flowtable_parse_hook(ctx, nla, &flowtable_hook, flowtable,
8773                                        extack, false);
8774         if (err < 0)
8775                 return err;
8776 
8777         list_for_each_entry(this, &flowtable_hook.list, list) {
8778                 hook = nft_hook_list_find(&flowtable->hook_list, this);
8779                 if (!hook) {
8780                         err = -ENOENT;
8781                         goto err_flowtable_del_hook;
8782                 }
8783                 list_move(&hook->list, &flowtable_del_list);
8784         }
8785 
8786         trans = nft_trans_alloc(ctx, NFT_MSG_DELFLOWTABLE,
8787                                 sizeof(struct nft_trans_flowtable));
8788         if (!trans) {
8789                 err = -ENOMEM;
8790                 goto err_flowtable_del_hook;
8791         }
8792 
8793         nft_trans_flowtable(trans) = flowtable;
8794         nft_trans_flowtable_update(trans) = true;
8795         INIT_LIST_HEAD(&nft_trans_flowtable_hooks(trans));
8796         list_splice(&flowtable_del_list, &nft_trans_flowtable_hooks(trans));
8797         nft_flowtable_hook_release(&flowtable_hook);
8798 
8799         nft_trans_commit_list_add_tail(ctx->net, trans);
8800 
8801         return 0;
8802 
8803 err_flowtable_del_hook:
8804         list_splice(&flowtable_del_list, &flowtable->hook_list);
8805         nft_flowtable_hook_release(&flowtable_hook);
8806 
8807         return err;
8808 }
8809 
8810 static int nf_tables_delflowtable(struct sk_buff *skb,
8811                                   const struct nfnl_info *info,
8812                                   const struct nlattr * const nla[])
8813 {
8814         struct netlink_ext_ack *extack = info->extack;
8815         u8 genmask = nft_genmask_next(info->net);
8816         u8 family = info->nfmsg->nfgen_family;
8817         struct nft_flowtable *flowtable;
8818         struct net *net = info->net;
8819         const struct nlattr *attr;
8820         struct nft_table *table;
8821         struct nft_ctx ctx;
8822 
8823         if (!nla[NFTA_FLOWTABLE_TABLE] ||
8824             (!nla[NFTA_FLOWTABLE_NAME] &&
8825              !nla[NFTA_FLOWTABLE_HANDLE]))
8826                 return -EINVAL;
8827 
8828         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
8829                                  genmask, NETLINK_CB(skb).portid);
8830         if (IS_ERR(table)) {
8831                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
8832                 return PTR_ERR(table);
8833         }
8834 
8835         if (nla[NFTA_FLOWTABLE_HANDLE]) {
8836                 attr = nla[NFTA_FLOWTABLE_HANDLE];
8837                 flowtable = nft_flowtable_lookup_byhandle(table, attr, genmask);
8838         } else {
8839                 attr = nla[NFTA_FLOWTABLE_NAME];
8840                 flowtable = nft_flowtable_lookup(table, attr, genmask);
8841         }
8842 
8843         if (IS_ERR(flowtable)) {
8844                 if (PTR_ERR(flowtable) == -ENOENT &&
8845                     NFNL_MSG_TYPE(info->nlh->nlmsg_type) == NFT_MSG_DESTROYFLOWTABLE)
8846                         return 0;
8847 
8848                 NL_SET_BAD_ATTR(extack, attr);
8849                 return PTR_ERR(flowtable);
8850         }
8851 
8852         nft_ctx_init(&ctx, net, skb, info->nlh, family, table, NULL, nla);
8853 
8854         if (nla[NFTA_FLOWTABLE_HOOK])
8855                 return nft_delflowtable_hook(&ctx, flowtable, extack);
8856 
8857         if (flowtable->use > 0) {
8858                 NL_SET_BAD_ATTR(extack, attr);
8859                 return -EBUSY;
8860         }
8861 
8862         return nft_delflowtable(&ctx, flowtable);
8863 }
8864 
8865 static int nf_tables_fill_flowtable_info(struct sk_buff *skb, struct net *net,
8866                                          u32 portid, u32 seq, int event,
8867                                          u32 flags, int family,
8868                                          struct nft_flowtable *flowtable,
8869                                          struct list_head *hook_list)
8870 {
8871         struct nlattr *nest, *nest_devs;
8872         struct nft_hook *hook;
8873         struct nlmsghdr *nlh;
8874 
8875         event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, event);
8876         nlh = nfnl_msg_put(skb, portid, seq, event, flags, family,
8877                            NFNETLINK_V0, nft_base_seq(net));
8878         if (!nlh)
8879                 goto nla_put_failure;
8880 
8881         if (nla_put_string(skb, NFTA_FLOWTABLE_TABLE, flowtable->table->name) ||
8882             nla_put_string(skb, NFTA_FLOWTABLE_NAME, flowtable->name) ||
8883             nla_put_be64(skb, NFTA_FLOWTABLE_HANDLE, cpu_to_be64(flowtable->handle),
8884                          NFTA_FLOWTABLE_PAD))
8885                 goto nla_put_failure;
8886 
8887         if (event == NFT_MSG_DELFLOWTABLE && !hook_list) {
8888                 nlmsg_end(skb, nlh);
8889                 return 0;
8890         }
8891 
8892         if (nla_put_be32(skb, NFTA_FLOWTABLE_USE, htonl(flowtable->use)) ||
8893             nla_put_be32(skb, NFTA_FLOWTABLE_FLAGS, htonl(flowtable->data.flags)))
8894                 goto nla_put_failure;
8895 
8896         nest = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK);
8897         if (!nest)
8898                 goto nla_put_failure;
8899         if (nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_NUM, htonl(flowtable->hooknum)) ||
8900             nla_put_be32(skb, NFTA_FLOWTABLE_HOOK_PRIORITY, htonl(flowtable->data.priority)))
8901                 goto nla_put_failure;
8902 
8903         nest_devs = nla_nest_start_noflag(skb, NFTA_FLOWTABLE_HOOK_DEVS);
8904         if (!nest_devs)
8905                 goto nla_put_failure;
8906 
8907         if (!hook_list)
8908                 hook_list = &flowtable->hook_list;
8909 
8910         list_for_each_entry_rcu(hook, hook_list, list) {
8911                 if (nla_put_string(skb, NFTA_DEVICE_NAME, hook->ops.dev->name))
8912                         goto nla_put_failure;
8913         }
8914         nla_nest_end(skb, nest_devs);
8915         nla_nest_end(skb, nest);
8916 
8917         nlmsg_end(skb, nlh);
8918         return 0;
8919 
8920 nla_put_failure:
8921         nlmsg_trim(skb, nlh);
8922         return -1;
8923 }
8924 
8925 struct nft_flowtable_filter {
8926         char            *table;
8927 };
8928 
8929 static int nf_tables_dump_flowtable(struct sk_buff *skb,
8930                                     struct netlink_callback *cb)
8931 {
8932         const struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
8933         struct nft_flowtable_filter *filter = cb->data;
8934         unsigned int idx = 0, s_idx = cb->args[0];
8935         struct net *net = sock_net(skb->sk);
8936         int family = nfmsg->nfgen_family;
8937         struct nft_flowtable *flowtable;
8938         struct nftables_pernet *nft_net;
8939         const struct nft_table *table;
8940 
8941         rcu_read_lock();
8942         nft_net = nft_pernet(net);
8943         cb->seq = READ_ONCE(nft_net->base_seq);
8944 
8945         list_for_each_entry_rcu(table, &nft_net->tables, list) {
8946                 if (family != NFPROTO_UNSPEC && family != table->family)
8947                         continue;
8948 
8949                 list_for_each_entry_rcu(flowtable, &table->flowtables, list) {
8950                         if (!nft_is_active(net, flowtable))
8951                                 goto cont;
8952                         if (idx < s_idx)
8953                                 goto cont;
8954                         if (idx > s_idx)
8955                                 memset(&cb->args[1], 0,
8956                                        sizeof(cb->args) - sizeof(cb->args[0]));
8957                         if (filter && filter->table &&
8958                             strcmp(filter->table, table->name))
8959                                 goto cont;
8960 
8961                         if (nf_tables_fill_flowtable_info(skb, net, NETLINK_CB(cb->skb).portid,
8962                                                           cb->nlh->nlmsg_seq,
8963                                                           NFT_MSG_NEWFLOWTABLE,
8964                                                           NLM_F_MULTI | NLM_F_APPEND,
8965                                                           table->family,
8966                                                           flowtable, NULL) < 0)
8967                                 goto done;
8968 
8969                         nl_dump_check_consistent(cb, nlmsg_hdr(skb));
8970 cont:
8971                         idx++;
8972                 }
8973         }
8974 done:
8975         rcu_read_unlock();
8976 
8977         cb->args[0] = idx;
8978         return skb->len;
8979 }
8980 
8981 static int nf_tables_dump_flowtable_start(struct netlink_callback *cb)
8982 {
8983         const struct nlattr * const *nla = cb->data;
8984         struct nft_flowtable_filter *filter = NULL;
8985 
8986         if (nla[NFTA_FLOWTABLE_TABLE]) {
8987                 filter = kzalloc(sizeof(*filter), GFP_ATOMIC);
8988                 if (!filter)
8989                         return -ENOMEM;
8990 
8991                 filter->table = nla_strdup(nla[NFTA_FLOWTABLE_TABLE],
8992                                            GFP_ATOMIC);
8993                 if (!filter->table) {
8994                         kfree(filter);
8995                         return -ENOMEM;
8996                 }
8997         }
8998 
8999         cb->data = filter;
9000         return 0;
9001 }
9002 
9003 static int nf_tables_dump_flowtable_done(struct netlink_callback *cb)
9004 {
9005         struct nft_flowtable_filter *filter = cb->data;
9006 
9007         if (!filter)
9008                 return 0;
9009 
9010         kfree(filter->table);
9011         kfree(filter);
9012 
9013         return 0;
9014 }
9015 
9016 /* called with rcu_read_lock held */
9017 static int nf_tables_getflowtable(struct sk_buff *skb,
9018                                   const struct nfnl_info *info,
9019                                   const struct nlattr * const nla[])
9020 {
9021         struct netlink_ext_ack *extack = info->extack;
9022         u8 genmask = nft_genmask_cur(info->net);
9023         u8 family = info->nfmsg->nfgen_family;
9024         struct nft_flowtable *flowtable;
9025         const struct nft_table *table;
9026         struct net *net = info->net;
9027         struct sk_buff *skb2;
9028         int err;
9029 
9030         if (info->nlh->nlmsg_flags & NLM_F_DUMP) {
9031                 struct netlink_dump_control c = {
9032                         .start = nf_tables_dump_flowtable_start,
9033                         .dump = nf_tables_dump_flowtable,
9034                         .done = nf_tables_dump_flowtable_done,
9035                         .module = THIS_MODULE,
9036                         .data = (void *)nla,
9037                 };
9038 
9039                 return nft_netlink_dump_start_rcu(info->sk, skb, info->nlh, &c);
9040         }
9041 
9042         if (!nla[NFTA_FLOWTABLE_NAME])
9043                 return -EINVAL;
9044 
9045         table = nft_table_lookup(net, nla[NFTA_FLOWTABLE_TABLE], family,
9046                                  genmask, 0);
9047         if (IS_ERR(table)) {
9048                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_TABLE]);
9049                 return PTR_ERR(table);
9050         }
9051 
9052         flowtable = nft_flowtable_lookup(table, nla[NFTA_FLOWTABLE_NAME],
9053                                          genmask);
9054         if (IS_ERR(flowtable)) {
9055                 NL_SET_BAD_ATTR(extack, nla[NFTA_FLOWTABLE_NAME]);
9056                 return PTR_ERR(flowtable);
9057         }
9058 
9059         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
9060         if (!skb2)
9061                 return -ENOMEM;
9062 
9063         err = nf_tables_fill_flowtable_info(skb2, net, NETLINK_CB(skb).portid,
9064                                             info->nlh->nlmsg_seq,
9065                                             NFT_MSG_NEWFLOWTABLE, 0, family,
9066                                             flowtable, NULL);
9067         if (err < 0)
9068                 goto err_fill_flowtable_info;
9069 
9070         return nfnetlink_unicast(skb2, net, NETLINK_CB(skb).portid);
9071 
9072 err_fill_flowtable_info:
9073         kfree_skb(skb2);
9074         return err;
9075 }
9076 
9077 static void nf_tables_flowtable_notify(struct nft_ctx *ctx,
9078                                        struct nft_flowtable *flowtable,
9079                                        struct list_head *hook_list, int event)
9080 {
9081         struct nftables_pernet *nft_net = nft_pernet(ctx->net);
9082         struct sk_buff *skb;
9083         u16 flags = 0;
9084         int err;
9085 
9086         if (!ctx->report &&
9087             !nfnetlink_has_listeners(ctx->net, NFNLGRP_NFTABLES))
9088                 return;
9089 
9090         skb = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
9091         if (skb == NULL)
9092                 goto err;
9093 
9094         if (ctx->flags & (NLM_F_CREATE | NLM_F_EXCL))
9095                 flags |= ctx->flags & (NLM_F_CREATE | NLM_F_EXCL);
9096 
9097         err = nf_tables_fill_flowtable_info(skb, ctx->net, ctx->portid,
9098                                             ctx->seq, event, flags,
9099                                             ctx->family, flowtable, hook_list);
9100         if (err < 0) {
9101                 kfree_skb(skb);
9102                 goto err;
9103         }
9104 
9105         nft_notify_enqueue(skb, ctx->report, &nft_net->notify_list);
9106         return;
9107 err:
9108         nfnetlink_set_err(ctx->net, ctx->portid, NFNLGRP_NFTABLES, -ENOBUFS);
9109 }
9110 
9111 static void nf_tables_flowtable_destroy(struct nft_flowtable *flowtable)
9112 {
9113         struct nft_hook *hook, *next;
9114 
9115         flowtable->data.type->free(&flowtable->data);
9116         list_for_each_entry_safe(hook, next, &flowtable->hook_list, list) {
9117                 flowtable->data.type->setup(&flowtable->data, hook->ops.dev,
9118                                             FLOW_BLOCK_UNBIND);
9119                 list_del_rcu(&hook->list);
9120                 kfree(hook);
9121         }
9122         kfree(flowtable->name);
9123         module_put(flowtable->data.type->owner);
9124         kfree(flowtable);
9125 }
9126 
9127 static int nf_tables_fill_gen_info(struct sk_buff *skb, struct net *net,
9128                                    u32 portid, u32 seq)
9129 {
9130         struct nftables_pernet *nft_net = nft_pernet(net);
9131         struct nlmsghdr *nlh;
9132         char buf[TASK_COMM_LEN];
9133         int event = nfnl_msg_type(NFNL_SUBSYS_NFTABLES, NFT_MSG_NEWGEN);
9134 
9135         nlh = nfnl_msg_put(skb, portid, seq, event, 0, AF_UNSPEC,
9136                            NFNETLINK_V0, nft_base_seq(net));
9137         if (!nlh)
9138                 goto nla_put_failure;
9139 
9140         if (nla_put_be32(skb, NFTA_GEN_ID, htonl(nft_net->base_seq)) ||
9141             nla_put_be32(skb, NFTA_GEN_PROC_PID, htonl(task_pid_nr(current))) ||
9142             nla_put_string(skb, NFTA_GEN_PROC_NAME, get_task_comm(buf, current)))
9143                 goto nla_put_failure;
9144 
9145         nlmsg_end(skb, nlh);
9146         return 0;
9147 
9148 nla_put_failure:
9149         nlmsg_trim(skb, nlh);
9150         return -EMSGSIZE;
9151 }
9152 
9153 static void nft_flowtable_event(unsigned long event, struct net_device *dev,
9154                                 struct nft_flowtable *flowtable)
9155 {
9156         struct nft_hook *hook;
9157 
9158         list_for_each_entry(hook, &flowtable->hook_list, list) {
9159                 if (hook->ops.dev != dev)
9160                         continue;
9161 
9162                 /* flow_offload_netdev_event() cleans up entries for us. */
9163                 nft_unregister_flowtable_hook(dev_net(dev), flowtable, hook);
9164                 list_del_rcu(&hook->list);
9165                 kfree_rcu(hook, rcu);
9166                 break;
9167         }
9168 }
9169 
9170 static int nf_tables_flowtable_event(struct notifier_block *this,
9171                                      unsigned long event, void *ptr)
9172 {
9173         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
9174         struct nft_flowtable *flowtable;
9175         struct nftables_pernet *nft_net;
9176         struct nft_table *table;
9177         struct net *net;
9178 
9179         if (event != NETDEV_UNREGISTER)
9180                 return 0;
9181 
9182         net = dev_net(dev);
9183         nft_net = nft_pernet(net);
9184         mutex_lock(&nft_net->commit_mutex);
9185         list_for_each_entry(table, &nft_net->tables, list) {
9186                 list_for_each_entry(flowtable, &table->flowtables, list) {
9187                         nft_flowtable_event(event, dev, flowtable);
9188                 }
9189         }
9190         mutex_unlock(&nft_net->commit_mutex);
9191 
9192         return NOTIFY_DONE;
9193 }
9194 
9195 static struct notifier_block nf_tables_flowtable_notifier = {
9196         .notifier_call  = nf_tables_flowtable_event,
9197 };
9198 
9199 static void nf_tables_gen_notify(struct net *net, struct sk_buff *skb,
9200                                  int event)
9201 {
9202         struct nlmsghdr *nlh = nlmsg_hdr(skb);
9203         struct sk_buff *skb2;
9204         int err;
9205 
9206         if (!nlmsg_report(nlh) &&
9207             !nfnetlink_has_listeners(net, NFNLGRP_NFTABLES))
9208                 return;
9209 
9210         skb2 = nlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
9211         if (skb2 == NULL)
9212                 goto err;
9213 
9214         err = nf_tables_fill_gen_info(skb2, net, NETLINK_CB(skb).portid,
9215                                       nlh->nlmsg_seq);
9216         if (err < 0) {
9217                 kfree_skb(skb2);
9218                 goto err;
9219         }
9220 
9221         nfnetlink_send(skb2, net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
9222                        nlmsg_report(nlh), GFP_KERNEL);
9223         return;
9224 err:
9225         nfnetlink_set_err(net, NETLINK_CB(skb).portid, NFNLGRP_NFTABLES,
9226                           -ENOBUFS);
9227 }
9228 
9229 static int nf_tables_getgen(struct sk_buff *skb, const struct nfnl_info *info,
9230                             const struct nlattr * const nla[])
9231 {
9232         struct sk_buff *skb2;
9233         int err;
9234 
9235         skb2 = alloc_skb(NLMSG_GOODSIZE, GFP_ATOMIC);
9236         if (skb2 == NULL)
9237                 return -ENOMEM;
9238 
9239         err = nf_tables_fill_gen_info(skb2, info->net, NETLINK_CB(skb).portid,
9240                                       info->nlh->nlmsg_seq);
9241         if (err < 0)
9242                 goto err_fill_gen_info;
9243 
9244         return nfnetlink_unicast(skb2, info->net, NETLINK_CB(skb).portid);
9245 
9246 err_fill_gen_info:
9247         kfree_skb(skb2);
9248         return err;
9249 }
9250 
9251 static const struct nfnl_callback nf_tables_cb[NFT_MSG_MAX] = {
9252         [NFT_MSG_NEWTABLE] = {
9253                 .call           = nf_tables_newtable,
9254                 .type           = NFNL_CB_BATCH,
9255                 .attr_count     = NFTA_TABLE_MAX,
9256                 .policy         = nft_table_policy,
9257         },
9258         [NFT_MSG_GETTABLE] = {
9259                 .call           = nf_tables_gettable,
9260                 .type           = NFNL_CB_RCU,
9261                 .attr_count     = NFTA_TABLE_MAX,
9262                 .policy         = nft_table_policy,
9263         },
9264         [NFT_MSG_DELTABLE] = {
9265                 .call           = nf_tables_deltable,
9266                 .type           = NFNL_CB_BATCH,
9267                 .attr_count     = NFTA_TABLE_MAX,
9268                 .policy         = nft_table_policy,
9269         },
9270         [NFT_MSG_DESTROYTABLE] = {
9271                 .call           = nf_tables_deltable,
9272                 .type           = NFNL_CB_BATCH,
9273                 .attr_count     = NFTA_TABLE_MAX,
9274                 .policy         = nft_table_policy,
9275         },
9276         [NFT_MSG_NEWCHAIN] = {
9277                 .call           = nf_tables_newchain,
9278                 .type           = NFNL_CB_BATCH,
9279                 .attr_count     = NFTA_CHAIN_MAX,
9280                 .policy         = nft_chain_policy,
9281         },
9282         [NFT_MSG_GETCHAIN] = {
9283                 .call           = nf_tables_getchain,
9284                 .type           = NFNL_CB_RCU,
9285                 .attr_count     = NFTA_CHAIN_MAX,
9286                 .policy         = nft_chain_policy,
9287         },
9288         [NFT_MSG_DELCHAIN] = {
9289                 .call           = nf_tables_delchain,
9290                 .type           = NFNL_CB_BATCH,
9291                 .attr_count     = NFTA_CHAIN_MAX,
9292                 .policy         = nft_chain_policy,
9293         },
9294         [NFT_MSG_DESTROYCHAIN] = {
9295                 .call           = nf_tables_delchain,
9296                 .type           = NFNL_CB_BATCH,
9297                 .attr_count     = NFTA_CHAIN_MAX,
9298                 .policy         = nft_chain_policy,
9299         },
9300         [NFT_MSG_NEWRULE] = {
9301                 .call           = nf_tables_newrule,
9302                 .type           = NFNL_CB_BATCH,
9303                 .attr_count     = NFTA_RULE_MAX,
9304                 .policy         = nft_rule_policy,
9305         },
9306         [NFT_MSG_GETRULE] = {
9307                 .call           = nf_tables_getrule,
9308                 .type           = NFNL_CB_RCU,
9309                 .attr_count     = NFTA_RULE_MAX,
9310                 .policy         = nft_rule_policy,
9311         },
9312         [NFT_MSG_GETRULE_RESET] = {
9313                 .call           = nf_tables_getrule_reset,
9314                 .type           = NFNL_CB_RCU,
9315                 .attr_count     = NFTA_RULE_MAX,
9316                 .policy         = nft_rule_policy,
9317         },
9318         [NFT_MSG_DELRULE] = {
9319                 .call           = nf_tables_delrule,
9320                 .type           = NFNL_CB_BATCH,
9321                 .attr_count     = NFTA_RULE_MAX,
9322                 .policy         = nft_rule_policy,
9323         },
9324         [NFT_MSG_DESTROYRULE] = {
9325                 .call           = nf_tables_delrule,
9326                 .type           = NFNL_CB_BATCH,
9327                 .attr_count     = NFTA_RULE_MAX,
9328                 .policy         = nft_rule_policy,
9329         },
9330         [NFT_MSG_NEWSET] = {
9331                 .call           = nf_tables_newset,
9332                 .type           = NFNL_CB_BATCH,
9333                 .attr_count     = NFTA_SET_MAX,
9334                 .policy         = nft_set_policy,
9335         },
9336         [NFT_MSG_GETSET] = {
9337                 .call           = nf_tables_getset,
9338                 .type           = NFNL_CB_RCU,
9339                 .attr_count     = NFTA_SET_MAX,
9340                 .policy         = nft_set_policy,
9341         },
9342         [NFT_MSG_DELSET] = {
9343                 .call           = nf_tables_delset,
9344                 .type           = NFNL_CB_BATCH,
9345                 .attr_count     = NFTA_SET_MAX,
9346                 .policy         = nft_set_policy,
9347         },
9348         [NFT_MSG_DESTROYSET] = {
9349                 .call           = nf_tables_delset,
9350                 .type           = NFNL_CB_BATCH,
9351                 .attr_count     = NFTA_SET_MAX,
9352                 .policy         = nft_set_policy,
9353         },
9354         [NFT_MSG_NEWSETELEM] = {
9355                 .call           = nf_tables_newsetelem,
9356                 .type           = NFNL_CB_BATCH,
9357                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
9358                 .policy         = nft_set_elem_list_policy,
9359         },
9360         [NFT_MSG_GETSETELEM] = {
9361                 .call           = nf_tables_getsetelem,
9362                 .type           = NFNL_CB_RCU,
9363                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
9364                 .policy         = nft_set_elem_list_policy,
9365         },
9366         [NFT_MSG_GETSETELEM_RESET] = {
9367                 .call           = nf_tables_getsetelem_reset,
9368                 .type           = NFNL_CB_RCU,
9369                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
9370                 .policy         = nft_set_elem_list_policy,
9371         },
9372         [NFT_MSG_DELSETELEM] = {
9373                 .call           = nf_tables_delsetelem,
9374                 .type           = NFNL_CB_BATCH,
9375                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
9376                 .policy         = nft_set_elem_list_policy,
9377         },
9378         [NFT_MSG_DESTROYSETELEM] = {
9379                 .call           = nf_tables_delsetelem,
9380                 .type           = NFNL_CB_BATCH,
9381                 .attr_count     = NFTA_SET_ELEM_LIST_MAX,
9382                 .policy         = nft_set_elem_list_policy,
9383         },
9384         [NFT_MSG_GETGEN] = {
9385                 .call           = nf_tables_getgen,
9386                 .type           = NFNL_CB_RCU,
9387         },
9388         [NFT_MSG_NEWOBJ] = {
9389                 .call           = nf_tables_newobj,
9390                 .type           = NFNL_CB_BATCH,
9391                 .attr_count     = NFTA_OBJ_MAX,
9392                 .policy         = nft_obj_policy,
9393         },
9394         [NFT_MSG_GETOBJ] = {
9395                 .call           = nf_tables_getobj,
9396                 .type           = NFNL_CB_RCU,
9397                 .attr_count     = NFTA_OBJ_MAX,
9398                 .policy         = nft_obj_policy,
9399         },
9400         [NFT_MSG_DELOBJ] = {
9401                 .call           = nf_tables_delobj,
9402                 .type           = NFNL_CB_BATCH,
9403                 .attr_count     = NFTA_OBJ_MAX,
9404                 .policy         = nft_obj_policy,
9405         },
9406         [NFT_MSG_DESTROYOBJ] = {
9407                 .call           = nf_tables_delobj,
9408                 .type           = NFNL_CB_BATCH,
9409                 .attr_count     = NFTA_OBJ_MAX,
9410                 .policy         = nft_obj_policy,
9411         },
9412         [NFT_MSG_GETOBJ_RESET] = {
9413                 .call           = nf_tables_getobj,
9414                 .type           = NFNL_CB_RCU,
9415                 .attr_count     = NFTA_OBJ_MAX,
9416                 .policy         = nft_obj_policy,
9417         },
9418         [NFT_MSG_NEWFLOWTABLE] = {
9419                 .call           = nf_tables_newflowtable,
9420                 .type           = NFNL_CB_BATCH,
9421                 .attr_count     = NFTA_FLOWTABLE_MAX,
9422                 .policy         = nft_flowtable_policy,
9423         },
9424         [NFT_MSG_GETFLOWTABLE] = {
9425                 .call           = nf_tables_getflowtable,
9426                 .type           = NFNL_CB_RCU,
9427                 .attr_count     = NFTA_FLOWTABLE_MAX,
9428                 .policy         = nft_flowtable_policy,
9429         },
9430         [NFT_MSG_DELFLOWTABLE] = {
9431                 .call           = nf_tables_delflowtable,
9432                 .type           = NFNL_CB_BATCH,
9433                 .attr_count     = NFTA_FLOWTABLE_MAX,
9434                 .policy         = nft_flowtable_policy,
9435         },
9436         [NFT_MSG_DESTROYFLOWTABLE] = {
9437                 .call           = nf_tables_delflowtable,
9438                 .type           = NFNL_CB_BATCH,
9439                 .attr_count     = NFTA_FLOWTABLE_MAX,
9440                 .policy         = nft_flowtable_policy,
9441         },
9442 };
9443 
9444 static int nf_tables_validate(struct net *net)
9445 {
9446         struct nftables_pernet *nft_net = nft_pernet(net);
9447         struct nft_table *table;
9448 
9449         list_for_each_entry(table, &nft_net->tables, list) {
9450                 switch (table->validate_state) {
9451                 case NFT_VALIDATE_SKIP:
9452                         continue;
9453                 case NFT_VALIDATE_NEED:
9454                         nft_validate_state_update(table, NFT_VALIDATE_DO);
9455                         fallthrough;
9456                 case NFT_VALIDATE_DO:
9457                         if (nft_table_validate(net, table) < 0)
9458                                 return -EAGAIN;
9459 
9460                         nft_validate_state_update(table, NFT_VALIDATE_SKIP);
9461                         break;
9462                 }
9463         }
9464 
9465         return 0;
9466 }
9467 
9468 /* a drop policy has to be deferred until all rules have been activated,
9469  * otherwise a large ruleset that contains a drop-policy base chain will
9470  * cause all packets to get dropped until the full transaction has been
9471  * processed.
9472  *
9473  * We defer the drop policy until the transaction has been finalized.
9474  */
9475 static void nft_chain_commit_drop_policy(struct nft_trans_chain *trans)
9476 {
9477         struct nft_base_chain *basechain;
9478 
9479         if (trans->policy != NF_DROP)
9480                 return;
9481 
9482         if (!nft_is_base_chain(trans->chain))
9483                 return;
9484 
9485         basechain = nft_base_chain(trans->chain);
9486         basechain->policy = NF_DROP;
9487 }
9488 
9489 static void nft_chain_commit_update(struct nft_trans_chain *trans)
9490 {
9491         struct nft_table *table = trans->nft_trans_binding.nft_trans.table;
9492         struct nft_base_chain *basechain;
9493 
9494         if (trans->name) {
9495                 rhltable_remove(&table->chains_ht,
9496                                 &trans->chain->rhlhead,
9497                                 nft_chain_ht_params);
9498                 swap(trans->chain->name, trans->name);
9499                 rhltable_insert_key(&table->chains_ht,
9500                                     trans->chain->name,
9501                                     &trans->chain->rhlhead,
9502                                     nft_chain_ht_params);
9503         }
9504 
9505         if (!nft_is_base_chain(trans->chain))
9506                 return;
9507 
9508         nft_chain_stats_replace(trans);
9509 
9510         basechain = nft_base_chain(trans->chain);
9511 
9512         switch (trans->policy) {
9513         case NF_DROP:
9514         case NF_ACCEPT:
9515                 basechain->policy = trans->policy;
9516                 break;
9517         }
9518 }
9519 
9520 static void nft_obj_commit_update(const struct nft_ctx *ctx,
9521                                   struct nft_trans *trans)
9522 {
9523         struct nft_object *newobj;
9524         struct nft_object *obj;
9525 
9526         obj = nft_trans_obj(trans);
9527         newobj = nft_trans_obj_newobj(trans);
9528 
9529         if (WARN_ON_ONCE(!obj->ops->update))
9530                 return;
9531 
9532         obj->ops->update(obj, newobj);
9533         nft_obj_destroy(ctx, newobj);
9534 }
9535 
9536 static void nft_commit_release(struct nft_trans *trans)
9537 {
9538         struct nft_ctx ctx = {
9539                 .net = trans->net,
9540         };
9541 
9542         nft_ctx_update(&ctx, trans);
9543 
9544         switch (trans->msg_type) {
9545         case NFT_MSG_DELTABLE:
9546         case NFT_MSG_DESTROYTABLE:
9547                 nf_tables_table_destroy(trans->table);
9548                 break;
9549         case NFT_MSG_NEWCHAIN:
9550                 free_percpu(nft_trans_chain_stats(trans));
9551                 kfree(nft_trans_chain_name(trans));
9552                 break;
9553         case NFT_MSG_DELCHAIN:
9554         case NFT_MSG_DESTROYCHAIN:
9555                 if (nft_trans_chain_update(trans))
9556                         nft_hooks_destroy(&nft_trans_chain_hooks(trans));
9557                 else
9558                         nf_tables_chain_destroy(nft_trans_chain(trans));
9559                 break;
9560         case NFT_MSG_DELRULE:
9561         case NFT_MSG_DESTROYRULE:
9562                 nf_tables_rule_destroy(&ctx, nft_trans_rule(trans));
9563                 break;
9564         case NFT_MSG_DELSET:
9565         case NFT_MSG_DESTROYSET:
9566                 nft_set_destroy(&ctx, nft_trans_set(trans));
9567                 break;
9568         case NFT_MSG_DELSETELEM:
9569         case NFT_MSG_DESTROYSETELEM:
9570                 nf_tables_set_elem_destroy(&ctx,
9571                                            nft_trans_elem_set(trans),
9572                                            nft_trans_elem_priv(trans));
9573                 break;
9574         case NFT_MSG_DELOBJ:
9575         case NFT_MSG_DESTROYOBJ:
9576                 nft_obj_destroy(&ctx, nft_trans_obj(trans));
9577                 break;
9578         case NFT_MSG_DELFLOWTABLE:
9579         case NFT_MSG_DESTROYFLOWTABLE:
9580                 if (nft_trans_flowtable_update(trans))
9581                         nft_hooks_destroy(&nft_trans_flowtable_hooks(trans));
9582                 else
9583                         nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
9584                 break;
9585         }
9586 
9587         if (trans->put_net)
9588                 put_net(trans->net);
9589 
9590         kfree(trans);
9591 }
9592 
9593 static void nf_tables_trans_destroy_work(struct work_struct *w)
9594 {
9595         struct nft_trans *trans, *next;
9596         LIST_HEAD(head);
9597 
9598         spin_lock(&nf_tables_destroy_list_lock);
9599         list_splice_init(&nf_tables_destroy_list, &head);
9600         spin_unlock(&nf_tables_destroy_list_lock);
9601 
9602         if (list_empty(&head))
9603                 return;
9604 
9605         synchronize_rcu();
9606 
9607         list_for_each_entry_safe(trans, next, &head, list) {
9608                 nft_trans_list_del(trans);
9609                 nft_commit_release(trans);
9610         }
9611 }
9612 
9613 void nf_tables_trans_destroy_flush_work(void)
9614 {
9615         flush_work(&trans_destroy_work);
9616 }
9617 EXPORT_SYMBOL_GPL(nf_tables_trans_destroy_flush_work);
9618 
9619 static bool nft_expr_reduce(struct nft_regs_track *track,
9620                             const struct nft_expr *expr)
9621 {
9622         return false;
9623 }
9624 
9625 static int nf_tables_commit_chain_prepare(struct net *net, struct nft_chain *chain)
9626 {
9627         const struct nft_expr *expr, *last;
9628         struct nft_regs_track track = {};
9629         unsigned int size, data_size;
9630         void *data, *data_boundary;
9631         struct nft_rule_dp *prule;
9632         struct nft_rule *rule;
9633 
9634         /* already handled or inactive chain? */
9635         if (chain->blob_next || !nft_is_active_next(net, chain))
9636                 return 0;
9637 
9638         data_size = 0;
9639         list_for_each_entry(rule, &chain->rules, list) {
9640                 if (nft_is_active_next(net, rule)) {
9641                         data_size += sizeof(*prule) + rule->dlen;
9642                         if (data_size > INT_MAX)
9643                                 return -ENOMEM;
9644                 }
9645         }
9646 
9647         chain->blob_next = nf_tables_chain_alloc_rules(chain, data_size);
9648         if (!chain->blob_next)
9649                 return -ENOMEM;
9650 
9651         data = (void *)chain->blob_next->data;
9652         data_boundary = data + data_size;
9653         size = 0;
9654 
9655         list_for_each_entry(rule, &chain->rules, list) {
9656                 if (!nft_is_active_next(net, rule))
9657                         continue;
9658 
9659                 prule = (struct nft_rule_dp *)data;
9660                 data += offsetof(struct nft_rule_dp, data);
9661                 if (WARN_ON_ONCE(data > data_boundary))
9662                         return -ENOMEM;
9663 
9664                 size = 0;
9665                 track.last = nft_expr_last(rule);
9666                 nft_rule_for_each_expr(expr, last, rule) {
9667                         track.cur = expr;
9668 
9669                         if (nft_expr_reduce(&track, expr)) {
9670                                 expr = track.cur;
9671                                 continue;
9672                         }
9673 
9674                         if (WARN_ON_ONCE(data + size + expr->ops->size > data_boundary))
9675                                 return -ENOMEM;
9676 
9677                         memcpy(data + size, expr, expr->ops->size);
9678                         size += expr->ops->size;
9679                 }
9680                 if (WARN_ON_ONCE(size >= 1 << 12))
9681                         return -ENOMEM;
9682 
9683                 prule->handle = rule->handle;
9684                 prule->dlen = size;
9685                 prule->is_last = 0;
9686 
9687                 data += size;
9688                 size = 0;
9689                 chain->blob_next->size += (unsigned long)(data - (void *)prule);
9690         }
9691 
9692         if (WARN_ON_ONCE(data > data_boundary))
9693                 return -ENOMEM;
9694 
9695         prule = (struct nft_rule_dp *)data;
9696         nft_last_rule(chain, prule);
9697 
9698         return 0;
9699 }
9700 
9701 static void nf_tables_commit_chain_prepare_cancel(struct net *net)
9702 {
9703         struct nftables_pernet *nft_net = nft_pernet(net);
9704         struct nft_trans *trans, *next;
9705 
9706         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
9707                 if (trans->msg_type == NFT_MSG_NEWRULE ||
9708                     trans->msg_type == NFT_MSG_DELRULE) {
9709                         struct nft_chain *chain = nft_trans_rule_chain(trans);
9710 
9711                         kvfree(chain->blob_next);
9712                         chain->blob_next = NULL;
9713                 }
9714         }
9715 }
9716 
9717 static void __nf_tables_commit_chain_free_rules(struct rcu_head *h)
9718 {
9719         struct nft_rule_dp_last *l = container_of(h, struct nft_rule_dp_last, h);
9720 
9721         kvfree(l->blob);
9722 }
9723 
9724 static void nf_tables_commit_chain_free_rules_old(struct nft_rule_blob *blob)
9725 {
9726         struct nft_rule_dp_last *last;
9727 
9728         /* last rule trailer is after end marker */
9729         last = (void *)blob + sizeof(*blob) + blob->size;
9730         last->blob = blob;
9731 
9732         call_rcu(&last->h, __nf_tables_commit_chain_free_rules);
9733 }
9734 
9735 static void nf_tables_commit_chain(struct net *net, struct nft_chain *chain)
9736 {
9737         struct nft_rule_blob *g0, *g1;
9738         bool next_genbit;
9739 
9740         next_genbit = nft_gencursor_next(net);
9741 
9742         g0 = rcu_dereference_protected(chain->blob_gen_0,
9743                                        lockdep_commit_lock_is_held(net));
9744         g1 = rcu_dereference_protected(chain->blob_gen_1,
9745                                        lockdep_commit_lock_is_held(net));
9746 
9747         /* No changes to this chain? */
9748         if (chain->blob_next == NULL) {
9749                 /* chain had no change in last or next generation */
9750                 if (g0 == g1)
9751                         return;
9752                 /*
9753                  * chain had no change in this generation; make sure next
9754                  * one uses same rules as current generation.
9755                  */
9756                 if (next_genbit) {
9757                         rcu_assign_pointer(chain->blob_gen_1, g0);
9758                         nf_tables_commit_chain_free_rules_old(g1);
9759                 } else {
9760                         rcu_assign_pointer(chain->blob_gen_0, g1);
9761                         nf_tables_commit_chain_free_rules_old(g0);
9762                 }
9763 
9764                 return;
9765         }
9766 
9767         if (next_genbit)
9768                 rcu_assign_pointer(chain->blob_gen_1, chain->blob_next);
9769         else
9770                 rcu_assign_pointer(chain->blob_gen_0, chain->blob_next);
9771 
9772         chain->blob_next = NULL;
9773 
9774         if (g0 == g1)
9775                 return;
9776 
9777         if (next_genbit)
9778                 nf_tables_commit_chain_free_rules_old(g1);
9779         else
9780                 nf_tables_commit_chain_free_rules_old(g0);
9781 }
9782 
9783 static void nft_obj_del(struct nft_object *obj)
9784 {
9785         rhltable_remove(&nft_objname_ht, &obj->rhlhead, nft_objname_ht_params);
9786         list_del_rcu(&obj->list);
9787 }
9788 
9789 void nft_chain_del(struct nft_chain *chain)
9790 {
9791         struct nft_table *table = chain->table;
9792 
9793         WARN_ON_ONCE(rhltable_remove(&table->chains_ht, &chain->rhlhead,
9794                                      nft_chain_ht_params));
9795         list_del_rcu(&chain->list);
9796 }
9797 
9798 static void nft_trans_gc_setelem_remove(struct nft_ctx *ctx,
9799                                         struct nft_trans_gc *trans)
9800 {
9801         struct nft_elem_priv **priv = trans->priv;
9802         unsigned int i;
9803 
9804         for (i = 0; i < trans->count; i++) {
9805                 nft_setelem_data_deactivate(ctx->net, trans->set, priv[i]);
9806                 nft_setelem_remove(ctx->net, trans->set, priv[i]);
9807         }
9808 }
9809 
9810 void nft_trans_gc_destroy(struct nft_trans_gc *trans)
9811 {
9812         nft_set_put(trans->set);
9813         put_net(trans->net);
9814         kfree(trans);
9815 }
9816 
9817 static void nft_trans_gc_trans_free(struct rcu_head *rcu)
9818 {
9819         struct nft_elem_priv *elem_priv;
9820         struct nft_trans_gc *trans;
9821         struct nft_ctx ctx = {};
9822         unsigned int i;
9823 
9824         trans = container_of(rcu, struct nft_trans_gc, rcu);
9825         ctx.net = read_pnet(&trans->set->net);
9826 
9827         for (i = 0; i < trans->count; i++) {
9828                 elem_priv = trans->priv[i];
9829                 if (!nft_setelem_is_catchall(trans->set, elem_priv))
9830                         atomic_dec(&trans->set->nelems);
9831 
9832                 nf_tables_set_elem_destroy(&ctx, trans->set, elem_priv);
9833         }
9834 
9835         nft_trans_gc_destroy(trans);
9836 }
9837 
9838 static bool nft_trans_gc_work_done(struct nft_trans_gc *trans)
9839 {
9840         struct nftables_pernet *nft_net;
9841         struct nft_ctx ctx = {};
9842 
9843         nft_net = nft_pernet(trans->net);
9844 
9845         mutex_lock(&nft_net->commit_mutex);
9846 
9847         /* Check for race with transaction, otherwise this batch refers to
9848          * stale objects that might not be there anymore. Skip transaction if
9849          * set has been destroyed from control plane transaction in case gc
9850          * worker loses race.
9851          */
9852         if (READ_ONCE(nft_net->gc_seq) != trans->seq || trans->set->dead) {
9853                 mutex_unlock(&nft_net->commit_mutex);
9854                 return false;
9855         }
9856 
9857         ctx.net = trans->net;
9858         ctx.table = trans->set->table;
9859 
9860         nft_trans_gc_setelem_remove(&ctx, trans);
9861         mutex_unlock(&nft_net->commit_mutex);
9862 
9863         return true;
9864 }
9865 
9866 static void nft_trans_gc_work(struct work_struct *work)
9867 {
9868         struct nft_trans_gc *trans, *next;
9869         LIST_HEAD(trans_gc_list);
9870 
9871         spin_lock(&nf_tables_gc_list_lock);
9872         list_splice_init(&nf_tables_gc_list, &trans_gc_list);
9873         spin_unlock(&nf_tables_gc_list_lock);
9874 
9875         list_for_each_entry_safe(trans, next, &trans_gc_list, list) {
9876                 list_del(&trans->list);
9877                 if (!nft_trans_gc_work_done(trans)) {
9878                         nft_trans_gc_destroy(trans);
9879                         continue;
9880                 }
9881                 call_rcu(&trans->rcu, nft_trans_gc_trans_free);
9882         }
9883 }
9884 
9885 struct nft_trans_gc *nft_trans_gc_alloc(struct nft_set *set,
9886                                         unsigned int gc_seq, gfp_t gfp)
9887 {
9888         struct net *net = read_pnet(&set->net);
9889         struct nft_trans_gc *trans;
9890 
9891         trans = kzalloc(sizeof(*trans), gfp);
9892         if (!trans)
9893                 return NULL;
9894 
9895         trans->net = maybe_get_net(net);
9896         if (!trans->net) {
9897                 kfree(trans);
9898                 return NULL;
9899         }
9900 
9901         refcount_inc(&set->refs);
9902         trans->set = set;
9903         trans->seq = gc_seq;
9904 
9905         return trans;
9906 }
9907 
9908 void nft_trans_gc_elem_add(struct nft_trans_gc *trans, void *priv)
9909 {
9910         trans->priv[trans->count++] = priv;
9911 }
9912 
9913 static void nft_trans_gc_queue_work(struct nft_trans_gc *trans)
9914 {
9915         spin_lock(&nf_tables_gc_list_lock);
9916         list_add_tail(&trans->list, &nf_tables_gc_list);
9917         spin_unlock(&nf_tables_gc_list_lock);
9918 
9919         schedule_work(&trans_gc_work);
9920 }
9921 
9922 static int nft_trans_gc_space(struct nft_trans_gc *trans)
9923 {
9924         return NFT_TRANS_GC_BATCHCOUNT - trans->count;
9925 }
9926 
9927 struct nft_trans_gc *nft_trans_gc_queue_async(struct nft_trans_gc *gc,
9928                                               unsigned int gc_seq, gfp_t gfp)
9929 {
9930         struct nft_set *set;
9931 
9932         if (nft_trans_gc_space(gc))
9933                 return gc;
9934 
9935         set = gc->set;
9936         nft_trans_gc_queue_work(gc);
9937 
9938         return nft_trans_gc_alloc(set, gc_seq, gfp);
9939 }
9940 
9941 void nft_trans_gc_queue_async_done(struct nft_trans_gc *trans)
9942 {
9943         if (trans->count == 0) {
9944                 nft_trans_gc_destroy(trans);
9945                 return;
9946         }
9947 
9948         nft_trans_gc_queue_work(trans);
9949 }
9950 
9951 struct nft_trans_gc *nft_trans_gc_queue_sync(struct nft_trans_gc *gc, gfp_t gfp)
9952 {
9953         struct nft_set *set;
9954 
9955         if (WARN_ON_ONCE(!lockdep_commit_lock_is_held(gc->net)))
9956                 return NULL;
9957 
9958         if (nft_trans_gc_space(gc))
9959                 return gc;
9960 
9961         set = gc->set;
9962         call_rcu(&gc->rcu, nft_trans_gc_trans_free);
9963 
9964         return nft_trans_gc_alloc(set, 0, gfp);
9965 }
9966 
9967 void nft_trans_gc_queue_sync_done(struct nft_trans_gc *trans)
9968 {
9969         WARN_ON_ONCE(!lockdep_commit_lock_is_held(trans->net));
9970 
9971         if (trans->count == 0) {
9972                 nft_trans_gc_destroy(trans);
9973                 return;
9974         }
9975 
9976         call_rcu(&trans->rcu, nft_trans_gc_trans_free);
9977 }
9978 
9979 struct nft_trans_gc *nft_trans_gc_catchall_async(struct nft_trans_gc *gc,
9980                                                  unsigned int gc_seq)
9981 {
9982         struct nft_set_elem_catchall *catchall;
9983         const struct nft_set *set = gc->set;
9984         struct nft_set_ext *ext;
9985 
9986         list_for_each_entry_rcu(catchall, &set->catchall_list, list) {
9987                 ext = nft_set_elem_ext(set, catchall->elem);
9988 
9989                 if (!nft_set_elem_expired(ext))
9990                         continue;
9991                 if (nft_set_elem_is_dead(ext))
9992                         goto dead_elem;
9993 
9994                 nft_set_elem_dead(ext);
9995 dead_elem:
9996                 gc = nft_trans_gc_queue_async(gc, gc_seq, GFP_ATOMIC);
9997                 if (!gc)
9998                         return NULL;
9999 
10000                 nft_trans_gc_elem_add(gc, catchall->elem);
10001         }
10002 
10003         return gc;
10004 }
10005 
10006 struct nft_trans_gc *nft_trans_gc_catchall_sync(struct nft_trans_gc *gc)
10007 {
10008         struct nft_set_elem_catchall *catchall, *next;
10009         u64 tstamp = nft_net_tstamp(gc->net);
10010         const struct nft_set *set = gc->set;
10011         struct nft_elem_priv *elem_priv;
10012         struct nft_set_ext *ext;
10013 
10014         WARN_ON_ONCE(!lockdep_commit_lock_is_held(gc->net));
10015 
10016         list_for_each_entry_safe(catchall, next, &set->catchall_list, list) {
10017                 ext = nft_set_elem_ext(set, catchall->elem);
10018 
10019                 if (!__nft_set_elem_expired(ext, tstamp))
10020                         continue;
10021 
10022                 gc = nft_trans_gc_queue_sync(gc, GFP_KERNEL);
10023                 if (!gc)
10024                         return NULL;
10025 
10026                 elem_priv = catchall->elem;
10027                 nft_setelem_data_deactivate(gc->net, gc->set, elem_priv);
10028                 nft_setelem_catchall_destroy(catchall);
10029                 nft_trans_gc_elem_add(gc, elem_priv);
10030         }
10031 
10032         return gc;
10033 }
10034 
10035 static void nf_tables_module_autoload_cleanup(struct net *net)
10036 {
10037         struct nftables_pernet *nft_net = nft_pernet(net);
10038         struct nft_module_request *req, *next;
10039 
10040         WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
10041         list_for_each_entry_safe(req, next, &nft_net->module_list, list) {
10042                 WARN_ON_ONCE(!req->done);
10043                 list_del(&req->list);
10044                 kfree(req);
10045         }
10046 }
10047 
10048 static void nf_tables_commit_release(struct net *net)
10049 {
10050         struct nftables_pernet *nft_net = nft_pernet(net);
10051         struct nft_trans *trans;
10052 
10053         /* all side effects have to be made visible.
10054          * For example, if a chain named 'foo' has been deleted, a
10055          * new transaction must not find it anymore.
10056          *
10057          * Memory reclaim happens asynchronously from work queue
10058          * to prevent expensive synchronize_rcu() in commit phase.
10059          */
10060         if (list_empty(&nft_net->commit_list)) {
10061                 nf_tables_module_autoload_cleanup(net);
10062                 mutex_unlock(&nft_net->commit_mutex);
10063                 return;
10064         }
10065 
10066         trans = list_last_entry(&nft_net->commit_list,
10067                                 struct nft_trans, list);
10068         get_net(trans->net);
10069         WARN_ON_ONCE(trans->put_net);
10070 
10071         trans->put_net = true;
10072         spin_lock(&nf_tables_destroy_list_lock);
10073         list_splice_tail_init(&nft_net->commit_list, &nf_tables_destroy_list);
10074         spin_unlock(&nf_tables_destroy_list_lock);
10075 
10076         nf_tables_module_autoload_cleanup(net);
10077         schedule_work(&trans_destroy_work);
10078 
10079         mutex_unlock(&nft_net->commit_mutex);
10080 }
10081 
10082 static void nft_commit_notify(struct net *net, u32 portid)
10083 {
10084         struct nftables_pernet *nft_net = nft_pernet(net);
10085         struct sk_buff *batch_skb = NULL, *nskb, *skb;
10086         unsigned char *data;
10087         int len;
10088 
10089         list_for_each_entry_safe(skb, nskb, &nft_net->notify_list, list) {
10090                 if (!batch_skb) {
10091 new_batch:
10092                         batch_skb = skb;
10093                         len = NLMSG_GOODSIZE - skb->len;
10094                         list_del(&skb->list);
10095                         continue;
10096                 }
10097                 len -= skb->len;
10098                 if (len > 0 && NFT_CB(skb).report == NFT_CB(batch_skb).report) {
10099                         data = skb_put(batch_skb, skb->len);
10100                         memcpy(data, skb->data, skb->len);
10101                         list_del(&skb->list);
10102                         kfree_skb(skb);
10103                         continue;
10104                 }
10105                 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
10106                                NFT_CB(batch_skb).report, GFP_KERNEL);
10107                 goto new_batch;
10108         }
10109 
10110         if (batch_skb) {
10111                 nfnetlink_send(batch_skb, net, portid, NFNLGRP_NFTABLES,
10112                                NFT_CB(batch_skb).report, GFP_KERNEL);
10113         }
10114 
10115         WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
10116 }
10117 
10118 static int nf_tables_commit_audit_alloc(struct list_head *adl,
10119                                         struct nft_table *table)
10120 {
10121         struct nft_audit_data *adp;
10122 
10123         list_for_each_entry(adp, adl, list) {
10124                 if (adp->table == table)
10125                         return 0;
10126         }
10127         adp = kzalloc(sizeof(*adp), GFP_KERNEL);
10128         if (!adp)
10129                 return -ENOMEM;
10130         adp->table = table;
10131         list_add(&adp->list, adl);
10132         return 0;
10133 }
10134 
10135 static void nf_tables_commit_audit_free(struct list_head *adl)
10136 {
10137         struct nft_audit_data *adp, *adn;
10138 
10139         list_for_each_entry_safe(adp, adn, adl, list) {
10140                 list_del(&adp->list);
10141                 kfree(adp);
10142         }
10143 }
10144 
10145 static void nf_tables_commit_audit_collect(struct list_head *adl,
10146                                            struct nft_table *table, u32 op)
10147 {
10148         struct nft_audit_data *adp;
10149 
10150         list_for_each_entry(adp, adl, list) {
10151                 if (adp->table == table)
10152                         goto found;
10153         }
10154         WARN_ONCE(1, "table=%s not expected in commit list", table->name);
10155         return;
10156 found:
10157         adp->entries++;
10158         if (!adp->op || adp->op > op)
10159                 adp->op = op;
10160 }
10161 
10162 #define AUNFTABLENAMELEN (NFT_TABLE_MAXNAMELEN + 22)
10163 
10164 static void nf_tables_commit_audit_log(struct list_head *adl, u32 generation)
10165 {
10166         struct nft_audit_data *adp, *adn;
10167         char aubuf[AUNFTABLENAMELEN];
10168 
10169         list_for_each_entry_safe(adp, adn, adl, list) {
10170                 snprintf(aubuf, AUNFTABLENAMELEN, "%s:%u", adp->table->name,
10171                          generation);
10172                 audit_log_nfcfg(aubuf, adp->table->family, adp->entries,
10173                                 nft2audit_op[adp->op], GFP_KERNEL);
10174                 list_del(&adp->list);
10175                 kfree(adp);
10176         }
10177 }
10178 
10179 static void nft_set_commit_update(struct list_head *set_update_list)
10180 {
10181         struct nft_set *set, *next;
10182 
10183         list_for_each_entry_safe(set, next, set_update_list, pending_update) {
10184                 list_del_init(&set->pending_update);
10185 
10186                 if (!set->ops->commit || set->dead)
10187                         continue;
10188 
10189                 set->ops->commit(set);
10190         }
10191 }
10192 
10193 static unsigned int nft_gc_seq_begin(struct nftables_pernet *nft_net)
10194 {
10195         unsigned int gc_seq;
10196 
10197         /* Bump gc counter, it becomes odd, this is the busy mark. */
10198         gc_seq = READ_ONCE(nft_net->gc_seq);
10199         WRITE_ONCE(nft_net->gc_seq, ++gc_seq);
10200 
10201         return gc_seq;
10202 }
10203 
10204 static void nft_gc_seq_end(struct nftables_pernet *nft_net, unsigned int gc_seq)
10205 {
10206         WRITE_ONCE(nft_net->gc_seq, ++gc_seq);
10207 }
10208 
10209 static int nf_tables_commit(struct net *net, struct sk_buff *skb)
10210 {
10211         struct nftables_pernet *nft_net = nft_pernet(net);
10212         const struct nlmsghdr *nlh = nlmsg_hdr(skb);
10213         struct nft_trans_binding *trans_binding;
10214         struct nft_trans *trans, *next;
10215         unsigned int base_seq, gc_seq;
10216         LIST_HEAD(set_update_list);
10217         struct nft_trans_elem *te;
10218         struct nft_chain *chain;
10219         struct nft_table *table;
10220         struct nft_ctx ctx;
10221         LIST_HEAD(adl);
10222         int err;
10223 
10224         if (list_empty(&nft_net->commit_list)) {
10225                 mutex_unlock(&nft_net->commit_mutex);
10226                 return 0;
10227         }
10228 
10229         nft_ctx_init(&ctx, net, skb, nlh, NFPROTO_UNSPEC, NULL, NULL, NULL);
10230 
10231         list_for_each_entry(trans_binding, &nft_net->binding_list, binding_list) {
10232                 trans = &trans_binding->nft_trans;
10233                 switch (trans->msg_type) {
10234                 case NFT_MSG_NEWSET:
10235                         if (!nft_trans_set_update(trans) &&
10236                             nft_set_is_anonymous(nft_trans_set(trans)) &&
10237                             !nft_trans_set_bound(trans)) {
10238                                 pr_warn_once("nftables ruleset with unbound set\n");
10239                                 return -EINVAL;
10240                         }
10241                         break;
10242                 case NFT_MSG_NEWCHAIN:
10243                         if (!nft_trans_chain_update(trans) &&
10244                             nft_chain_binding(nft_trans_chain(trans)) &&
10245                             !nft_trans_chain_bound(trans)) {
10246                                 pr_warn_once("nftables ruleset with unbound chain\n");
10247                                 return -EINVAL;
10248                         }
10249                         break;
10250                 default:
10251                         WARN_ONCE(1, "Unhandled bind type %d", trans->msg_type);
10252                         break;
10253                 }
10254         }
10255 
10256         /* 0. Validate ruleset, otherwise roll back for error reporting. */
10257         if (nf_tables_validate(net) < 0) {
10258                 nft_net->validate_state = NFT_VALIDATE_DO;
10259                 return -EAGAIN;
10260         }
10261 
10262         err = nft_flow_rule_offload_commit(net);
10263         if (err < 0)
10264                 return err;
10265 
10266         /* 1.  Allocate space for next generation rules_gen_X[] */
10267         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
10268                 struct nft_table *table = trans->table;
10269                 int ret;
10270 
10271                 ret = nf_tables_commit_audit_alloc(&adl, table);
10272                 if (ret) {
10273                         nf_tables_commit_chain_prepare_cancel(net);
10274                         nf_tables_commit_audit_free(&adl);
10275                         return ret;
10276                 }
10277                 if (trans->msg_type == NFT_MSG_NEWRULE ||
10278                     trans->msg_type == NFT_MSG_DELRULE) {
10279                         chain = nft_trans_rule_chain(trans);
10280 
10281                         ret = nf_tables_commit_chain_prepare(net, chain);
10282                         if (ret < 0) {
10283                                 nf_tables_commit_chain_prepare_cancel(net);
10284                                 nf_tables_commit_audit_free(&adl);
10285                                 return ret;
10286                         }
10287                 }
10288         }
10289 
10290         /* step 2.  Make rules_gen_X visible to packet path */
10291         list_for_each_entry(table, &nft_net->tables, list) {
10292                 list_for_each_entry(chain, &table->chains, list)
10293                         nf_tables_commit_chain(net, chain);
10294         }
10295 
10296         /*
10297          * Bump generation counter, invalidate any dump in progress.
10298          * Cannot fail after this point.
10299          */
10300         base_seq = READ_ONCE(nft_net->base_seq);
10301         while (++base_seq == 0)
10302                 ;
10303 
10304         WRITE_ONCE(nft_net->base_seq, base_seq);
10305 
10306         gc_seq = nft_gc_seq_begin(nft_net);
10307 
10308         /* step 3. Start new generation, rules_gen_X now in use. */
10309         net->nft.gencursor = nft_gencursor_next(net);
10310 
10311         list_for_each_entry_safe(trans, next, &nft_net->commit_list, list) {
10312                 struct nft_table *table = trans->table;
10313 
10314                 nft_ctx_update(&ctx, trans);
10315 
10316                 nf_tables_commit_audit_collect(&adl, table, trans->msg_type);
10317                 switch (trans->msg_type) {
10318                 case NFT_MSG_NEWTABLE:
10319                         if (nft_trans_table_update(trans)) {
10320                                 if (!(table->flags & __NFT_TABLE_F_UPDATE)) {
10321                                         nft_trans_destroy(trans);
10322                                         break;
10323                                 }
10324                                 if (table->flags & NFT_TABLE_F_DORMANT)
10325                                         nf_tables_table_disable(net, table);
10326 
10327                                 table->flags &= ~__NFT_TABLE_F_UPDATE;
10328                         } else {
10329                                 nft_clear(net, table);
10330                         }
10331                         nf_tables_table_notify(&ctx, NFT_MSG_NEWTABLE);
10332                         nft_trans_destroy(trans);
10333                         break;
10334                 case NFT_MSG_DELTABLE:
10335                 case NFT_MSG_DESTROYTABLE:
10336                         list_del_rcu(&table->list);
10337                         nf_tables_table_notify(&ctx, trans->msg_type);
10338                         break;
10339                 case NFT_MSG_NEWCHAIN:
10340                         if (nft_trans_chain_update(trans)) {
10341                                 nft_chain_commit_update(nft_trans_container_chain(trans));
10342                                 nf_tables_chain_notify(&ctx, NFT_MSG_NEWCHAIN,
10343                                                        &nft_trans_chain_hooks(trans));
10344                                 list_splice(&nft_trans_chain_hooks(trans),
10345                                             &nft_trans_basechain(trans)->hook_list);
10346                                 /* trans destroyed after rcu grace period */
10347                         } else {
10348                                 nft_chain_commit_drop_policy(nft_trans_container_chain(trans));
10349                                 nft_clear(net, nft_trans_chain(trans));
10350                                 nf_tables_chain_notify(&ctx, NFT_MSG_NEWCHAIN, NULL);
10351                                 nft_trans_destroy(trans);
10352                         }
10353                         break;
10354                 case NFT_MSG_DELCHAIN:
10355                 case NFT_MSG_DESTROYCHAIN:
10356                         if (nft_trans_chain_update(trans)) {
10357                                 nf_tables_chain_notify(&ctx, NFT_MSG_DELCHAIN,
10358                                                        &nft_trans_chain_hooks(trans));
10359                                 if (!(table->flags & NFT_TABLE_F_DORMANT)) {
10360                                         nft_netdev_unregister_hooks(net,
10361                                                                     &nft_trans_chain_hooks(trans),
10362                                                                     true);
10363                                 }
10364                         } else {
10365                                 nft_chain_del(nft_trans_chain(trans));
10366                                 nf_tables_chain_notify(&ctx, NFT_MSG_DELCHAIN,
10367                                                        NULL);
10368                                 nf_tables_unregister_hook(ctx.net, ctx.table,
10369                                                           nft_trans_chain(trans));
10370                         }
10371                         break;
10372                 case NFT_MSG_NEWRULE:
10373                         nft_clear(net, nft_trans_rule(trans));
10374                         nf_tables_rule_notify(&ctx, nft_trans_rule(trans),
10375                                               NFT_MSG_NEWRULE);
10376                         if (nft_trans_rule_chain(trans)->flags & NFT_CHAIN_HW_OFFLOAD)
10377                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
10378 
10379                         nft_trans_destroy(trans);
10380                         break;
10381                 case NFT_MSG_DELRULE:
10382                 case NFT_MSG_DESTROYRULE:
10383                         list_del_rcu(&nft_trans_rule(trans)->list);
10384                         nf_tables_rule_notify(&ctx, nft_trans_rule(trans),
10385                                               trans->msg_type);
10386                         nft_rule_expr_deactivate(&ctx, nft_trans_rule(trans),
10387                                                  NFT_TRANS_COMMIT);
10388 
10389                         if (nft_trans_rule_chain(trans)->flags & NFT_CHAIN_HW_OFFLOAD)
10390                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
10391                         break;
10392                 case NFT_MSG_NEWSET:
10393                         if (nft_trans_set_update(trans)) {
10394                                 struct nft_set *set = nft_trans_set(trans);
10395 
10396                                 WRITE_ONCE(set->timeout, nft_trans_set_timeout(trans));
10397                                 WRITE_ONCE(set->gc_int, nft_trans_set_gc_int(trans));
10398 
10399                                 if (nft_trans_set_size(trans))
10400                                         WRITE_ONCE(set->size, nft_trans_set_size(trans));
10401                         } else {
10402                                 nft_clear(net, nft_trans_set(trans));
10403                                 /* This avoids hitting -EBUSY when deleting the table
10404                                  * from the transaction.
10405                                  */
10406                                 if (nft_set_is_anonymous(nft_trans_set(trans)) &&
10407                                     !list_empty(&nft_trans_set(trans)->bindings))
10408                                         nft_use_dec(&table->use);
10409                         }
10410                         nf_tables_set_notify(&ctx, nft_trans_set(trans),
10411                                              NFT_MSG_NEWSET, GFP_KERNEL);
10412                         nft_trans_destroy(trans);
10413                         break;
10414                 case NFT_MSG_DELSET:
10415                 case NFT_MSG_DESTROYSET:
10416                         nft_trans_set(trans)->dead = 1;
10417                         list_del_rcu(&nft_trans_set(trans)->list);
10418                         nf_tables_set_notify(&ctx, nft_trans_set(trans),
10419                                              trans->msg_type, GFP_KERNEL);
10420                         break;
10421                 case NFT_MSG_NEWSETELEM:
10422                         te = nft_trans_container_elem(trans);
10423 
10424                         nft_setelem_activate(net, te->set, te->elem_priv);
10425                         nf_tables_setelem_notify(&ctx, te->set,
10426                                                  te->elem_priv,
10427                                                  NFT_MSG_NEWSETELEM);
10428                         if (te->set->ops->commit &&
10429                             list_empty(&te->set->pending_update)) {
10430                                 list_add_tail(&te->set->pending_update,
10431                                               &set_update_list);
10432                         }
10433                         nft_trans_destroy(trans);
10434                         break;
10435                 case NFT_MSG_DELSETELEM:
10436                 case NFT_MSG_DESTROYSETELEM:
10437                         te = nft_trans_container_elem(trans);
10438 
10439                         nf_tables_setelem_notify(&ctx, te->set,
10440                                                  te->elem_priv,
10441                                                  trans->msg_type);
10442                         nft_setelem_remove(net, te->set, te->elem_priv);
10443                         if (!nft_setelem_is_catchall(te->set, te->elem_priv)) {
10444                                 atomic_dec(&te->set->nelems);
10445                                 te->set->ndeact--;
10446                         }
10447                         if (te->set->ops->commit &&
10448                             list_empty(&te->set->pending_update)) {
10449                                 list_add_tail(&te->set->pending_update,
10450                                               &set_update_list);
10451                         }
10452                         break;
10453                 case NFT_MSG_NEWOBJ:
10454                         if (nft_trans_obj_update(trans)) {
10455                                 nft_obj_commit_update(&ctx, trans);
10456                                 nf_tables_obj_notify(&ctx,
10457                                                      nft_trans_obj(trans),
10458                                                      NFT_MSG_NEWOBJ);
10459                         } else {
10460                                 nft_clear(net, nft_trans_obj(trans));
10461                                 nf_tables_obj_notify(&ctx,
10462                                                      nft_trans_obj(trans),
10463                                                      NFT_MSG_NEWOBJ);
10464                                 nft_trans_destroy(trans);
10465                         }
10466                         break;
10467                 case NFT_MSG_DELOBJ:
10468                 case NFT_MSG_DESTROYOBJ:
10469                         nft_obj_del(nft_trans_obj(trans));
10470                         nf_tables_obj_notify(&ctx, nft_trans_obj(trans),
10471                                              trans->msg_type);
10472                         break;
10473                 case NFT_MSG_NEWFLOWTABLE:
10474                         if (nft_trans_flowtable_update(trans)) {
10475                                 nft_trans_flowtable(trans)->data.flags =
10476                                         nft_trans_flowtable_flags(trans);
10477                                 nf_tables_flowtable_notify(&ctx,
10478                                                            nft_trans_flowtable(trans),
10479                                                            &nft_trans_flowtable_hooks(trans),
10480                                                            NFT_MSG_NEWFLOWTABLE);
10481                                 list_splice(&nft_trans_flowtable_hooks(trans),
10482                                             &nft_trans_flowtable(trans)->hook_list);
10483                         } else {
10484                                 nft_clear(net, nft_trans_flowtable(trans));
10485                                 nf_tables_flowtable_notify(&ctx,
10486                                                            nft_trans_flowtable(trans),
10487                                                            NULL,
10488                                                            NFT_MSG_NEWFLOWTABLE);
10489                         }
10490                         nft_trans_destroy(trans);
10491                         break;
10492                 case NFT_MSG_DELFLOWTABLE:
10493                 case NFT_MSG_DESTROYFLOWTABLE:
10494                         if (nft_trans_flowtable_update(trans)) {
10495                                 nf_tables_flowtable_notify(&ctx,
10496                                                            nft_trans_flowtable(trans),
10497                                                            &nft_trans_flowtable_hooks(trans),
10498                                                            trans->msg_type);
10499                                 nft_unregister_flowtable_net_hooks(net,
10500                                                                    &nft_trans_flowtable_hooks(trans));
10501                         } else {
10502                                 list_del_rcu(&nft_trans_flowtable(trans)->list);
10503                                 nf_tables_flowtable_notify(&ctx,
10504                                                            nft_trans_flowtable(trans),
10505                                                            NULL,
10506                                                            trans->msg_type);
10507                                 nft_unregister_flowtable_net_hooks(net,
10508                                                 &nft_trans_flowtable(trans)->hook_list);
10509                         }
10510                         break;
10511                 }
10512         }
10513 
10514         nft_set_commit_update(&set_update_list);
10515 
10516         nft_commit_notify(net, NETLINK_CB(skb).portid);
10517         nf_tables_gen_notify(net, skb, NFT_MSG_NEWGEN);
10518         nf_tables_commit_audit_log(&adl, nft_net->base_seq);
10519 
10520         nft_gc_seq_end(nft_net, gc_seq);
10521         nft_net->validate_state = NFT_VALIDATE_SKIP;
10522         nf_tables_commit_release(net);
10523 
10524         return 0;
10525 }
10526 
10527 static void nf_tables_module_autoload(struct net *net)
10528 {
10529         struct nftables_pernet *nft_net = nft_pernet(net);
10530         struct nft_module_request *req, *next;
10531         LIST_HEAD(module_list);
10532 
10533         list_splice_init(&nft_net->module_list, &module_list);
10534         mutex_unlock(&nft_net->commit_mutex);
10535         list_for_each_entry_safe(req, next, &module_list, list) {
10536                 request_module("%s", req->module);
10537                 req->done = true;
10538         }
10539         mutex_lock(&nft_net->commit_mutex);
10540         list_splice(&module_list, &nft_net->module_list);
10541 }
10542 
10543 static void nf_tables_abort_release(struct nft_trans *trans)
10544 {
10545         struct nft_ctx ctx = { };
10546 
10547         nft_ctx_update(&ctx, trans);
10548 
10549         switch (trans->msg_type) {
10550         case NFT_MSG_NEWTABLE:
10551                 nf_tables_table_destroy(trans->table);
10552                 break;
10553         case NFT_MSG_NEWCHAIN:
10554                 if (nft_trans_chain_update(trans))
10555                         nft_hooks_destroy(&nft_trans_chain_hooks(trans));
10556                 else
10557                         nf_tables_chain_destroy(nft_trans_chain(trans));
10558                 break;
10559         case NFT_MSG_NEWRULE:
10560                 nf_tables_rule_destroy(&ctx, nft_trans_rule(trans));
10561                 break;
10562         case NFT_MSG_NEWSET:
10563                 nft_set_destroy(&ctx, nft_trans_set(trans));
10564                 break;
10565         case NFT_MSG_NEWSETELEM:
10566                 nft_set_elem_destroy(nft_trans_elem_set(trans),
10567                                      nft_trans_elem_priv(trans), true);
10568                 break;
10569         case NFT_MSG_NEWOBJ:
10570                 nft_obj_destroy(&ctx, nft_trans_obj(trans));
10571                 break;
10572         case NFT_MSG_NEWFLOWTABLE:
10573                 if (nft_trans_flowtable_update(trans))
10574                         nft_hooks_destroy(&nft_trans_flowtable_hooks(trans));
10575                 else
10576                         nf_tables_flowtable_destroy(nft_trans_flowtable(trans));
10577                 break;
10578         }
10579         kfree(trans);
10580 }
10581 
10582 static void nft_set_abort_update(struct list_head *set_update_list)
10583 {
10584         struct nft_set *set, *next;
10585 
10586         list_for_each_entry_safe(set, next, set_update_list, pending_update) {
10587                 list_del_init(&set->pending_update);
10588 
10589                 if (!set->ops->abort)
10590                         continue;
10591 
10592                 set->ops->abort(set);
10593         }
10594 }
10595 
10596 static int __nf_tables_abort(struct net *net, enum nfnl_abort_action action)
10597 {
10598         struct nftables_pernet *nft_net = nft_pernet(net);
10599         struct nft_trans *trans, *next;
10600         LIST_HEAD(set_update_list);
10601         struct nft_trans_elem *te;
10602         struct nft_ctx ctx = {
10603                 .net = net,
10604         };
10605         int err = 0;
10606 
10607         if (action == NFNL_ABORT_VALIDATE &&
10608             nf_tables_validate(net) < 0)
10609                 err = -EAGAIN;
10610 
10611         list_for_each_entry_safe_reverse(trans, next, &nft_net->commit_list,
10612                                          list) {
10613                 struct nft_table *table = trans->table;
10614 
10615                 nft_ctx_update(&ctx, trans);
10616 
10617                 switch (trans->msg_type) {
10618                 case NFT_MSG_NEWTABLE:
10619                         if (nft_trans_table_update(trans)) {
10620                                 if (!(table->flags & __NFT_TABLE_F_UPDATE)) {
10621                                         nft_trans_destroy(trans);
10622                                         break;
10623                                 }
10624                                 if (table->flags & __NFT_TABLE_F_WAS_DORMANT) {
10625                                         nf_tables_table_disable(net, table);
10626                                         table->flags |= NFT_TABLE_F_DORMANT;
10627                                 } else if (table->flags & __NFT_TABLE_F_WAS_AWAKEN) {
10628                                         table->flags &= ~NFT_TABLE_F_DORMANT;
10629                                 }
10630                                 if (table->flags & __NFT_TABLE_F_WAS_ORPHAN) {
10631                                         table->flags &= ~NFT_TABLE_F_OWNER;
10632                                         table->nlpid = 0;
10633                                 }
10634                                 table->flags &= ~__NFT_TABLE_F_UPDATE;
10635                                 nft_trans_destroy(trans);
10636                         } else {
10637                                 list_del_rcu(&table->list);
10638                         }
10639                         break;
10640                 case NFT_MSG_DELTABLE:
10641                 case NFT_MSG_DESTROYTABLE:
10642                         nft_clear(trans->net, table);
10643                         nft_trans_destroy(trans);
10644                         break;
10645                 case NFT_MSG_NEWCHAIN:
10646                         if (nft_trans_chain_update(trans)) {
10647                                 if (!(table->flags & NFT_TABLE_F_DORMANT)) {
10648                                         nft_netdev_unregister_hooks(net,
10649                                                                     &nft_trans_chain_hooks(trans),
10650                                                                     true);
10651                                 }
10652                                 free_percpu(nft_trans_chain_stats(trans));
10653                                 kfree(nft_trans_chain_name(trans));
10654                                 nft_trans_destroy(trans);
10655                         } else {
10656                                 if (nft_trans_chain_bound(trans)) {
10657                                         nft_trans_destroy(trans);
10658                                         break;
10659                                 }
10660                                 nft_use_dec_restore(&table->use);
10661                                 nft_chain_del(nft_trans_chain(trans));
10662                                 nf_tables_unregister_hook(trans->net, table,
10663                                                           nft_trans_chain(trans));
10664                         }
10665                         break;
10666                 case NFT_MSG_DELCHAIN:
10667                 case NFT_MSG_DESTROYCHAIN:
10668                         if (nft_trans_chain_update(trans)) {
10669                                 list_splice(&nft_trans_chain_hooks(trans),
10670                                             &nft_trans_basechain(trans)->hook_list);
10671                         } else {
10672                                 nft_use_inc_restore(&table->use);
10673                                 nft_clear(trans->net, nft_trans_chain(trans));
10674                         }
10675                         nft_trans_destroy(trans);
10676                         break;
10677                 case NFT_MSG_NEWRULE:
10678                         if (nft_trans_rule_bound(trans)) {
10679                                 nft_trans_destroy(trans);
10680                                 break;
10681                         }
10682                         nft_use_dec_restore(&nft_trans_rule_chain(trans)->use);
10683                         list_del_rcu(&nft_trans_rule(trans)->list);
10684                         nft_rule_expr_deactivate(&ctx,
10685                                                  nft_trans_rule(trans),
10686                                                  NFT_TRANS_ABORT);
10687                         if (nft_trans_rule_chain(trans)->flags & NFT_CHAIN_HW_OFFLOAD)
10688                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
10689                         break;
10690                 case NFT_MSG_DELRULE:
10691                 case NFT_MSG_DESTROYRULE:
10692                         nft_use_inc_restore(&nft_trans_rule_chain(trans)->use);
10693                         nft_clear(trans->net, nft_trans_rule(trans));
10694                         nft_rule_expr_activate(&ctx, nft_trans_rule(trans));
10695                         if (nft_trans_rule_chain(trans)->flags & NFT_CHAIN_HW_OFFLOAD)
10696                                 nft_flow_rule_destroy(nft_trans_flow_rule(trans));
10697 
10698                         nft_trans_destroy(trans);
10699                         break;
10700                 case NFT_MSG_NEWSET:
10701                         if (nft_trans_set_update(trans)) {
10702                                 nft_trans_destroy(trans);
10703                                 break;
10704                         }
10705                         nft_use_dec_restore(&table->use);
10706                         if (nft_trans_set_bound(trans)) {
10707                                 nft_trans_destroy(trans);
10708                                 break;
10709                         }
10710                         nft_trans_set(trans)->dead = 1;
10711                         list_del_rcu(&nft_trans_set(trans)->list);
10712                         break;
10713                 case NFT_MSG_DELSET:
10714                 case NFT_MSG_DESTROYSET:
10715                         nft_use_inc_restore(&table->use);
10716                         nft_clear(trans->net, nft_trans_set(trans));
10717                         if (nft_trans_set(trans)->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
10718                                 nft_map_activate(&ctx, nft_trans_set(trans));
10719 
10720                         nft_trans_destroy(trans);
10721                         break;
10722                 case NFT_MSG_NEWSETELEM:
10723                         if (nft_trans_elem_set_bound(trans)) {
10724                                 nft_trans_destroy(trans);
10725                                 break;
10726                         }
10727                         te = nft_trans_container_elem(trans);
10728                         nft_setelem_remove(net, te->set, te->elem_priv);
10729                         if (!nft_setelem_is_catchall(te->set, te->elem_priv))
10730                                 atomic_dec(&te->set->nelems);
10731 
10732                         if (te->set->ops->abort &&
10733                             list_empty(&te->set->pending_update)) {
10734                                 list_add_tail(&te->set->pending_update,
10735                                               &set_update_list);
10736                         }
10737                         break;
10738                 case NFT_MSG_DELSETELEM:
10739                 case NFT_MSG_DESTROYSETELEM:
10740                         te = nft_trans_container_elem(trans);
10741 
10742                         if (!nft_setelem_active_next(net, te->set, te->elem_priv)) {
10743                                 nft_setelem_data_activate(net, te->set, te->elem_priv);
10744                                 nft_setelem_activate(net, te->set, te->elem_priv);
10745                         }
10746                         if (!nft_setelem_is_catchall(te->set, te->elem_priv))
10747                                 te->set->ndeact--;
10748 
10749                         if (te->set->ops->abort &&
10750                             list_empty(&te->set->pending_update)) {
10751                                 list_add_tail(&te->set->pending_update,
10752                                               &set_update_list);
10753                         }
10754                         nft_trans_destroy(trans);
10755                         break;
10756                 case NFT_MSG_NEWOBJ:
10757                         if (nft_trans_obj_update(trans)) {
10758                                 nft_obj_destroy(&ctx, nft_trans_obj_newobj(trans));
10759                                 nft_trans_destroy(trans);
10760                         } else {
10761                                 nft_use_dec_restore(&table->use);
10762                                 nft_obj_del(nft_trans_obj(trans));
10763                         }
10764                         break;
10765                 case NFT_MSG_DELOBJ:
10766                 case NFT_MSG_DESTROYOBJ:
10767                         nft_use_inc_restore(&table->use);
10768                         nft_clear(trans->net, nft_trans_obj(trans));
10769                         nft_trans_destroy(trans);
10770                         break;
10771                 case NFT_MSG_NEWFLOWTABLE:
10772                         if (nft_trans_flowtable_update(trans)) {
10773                                 nft_unregister_flowtable_net_hooks(net,
10774                                                 &nft_trans_flowtable_hooks(trans));
10775                         } else {
10776                                 nft_use_dec_restore(&table->use);
10777                                 list_del_rcu(&nft_trans_flowtable(trans)->list);
10778                                 nft_unregister_flowtable_net_hooks(net,
10779                                                 &nft_trans_flowtable(trans)->hook_list);
10780                         }
10781                         break;
10782                 case NFT_MSG_DELFLOWTABLE:
10783                 case NFT_MSG_DESTROYFLOWTABLE:
10784                         if (nft_trans_flowtable_update(trans)) {
10785                                 list_splice(&nft_trans_flowtable_hooks(trans),
10786                                             &nft_trans_flowtable(trans)->hook_list);
10787                         } else {
10788                                 nft_use_inc_restore(&table->use);
10789                                 nft_clear(trans->net, nft_trans_flowtable(trans));
10790                         }
10791                         nft_trans_destroy(trans);
10792                         break;
10793                 }
10794         }
10795 
10796         nft_set_abort_update(&set_update_list);
10797 
10798         synchronize_rcu();
10799 
10800         list_for_each_entry_safe_reverse(trans, next,
10801                                          &nft_net->commit_list, list) {
10802                 nft_trans_list_del(trans);
10803                 nf_tables_abort_release(trans);
10804         }
10805 
10806         return err;
10807 }
10808 
10809 static int nf_tables_abort(struct net *net, struct sk_buff *skb,
10810                            enum nfnl_abort_action action)
10811 {
10812         struct nftables_pernet *nft_net = nft_pernet(net);
10813         unsigned int gc_seq;
10814         int ret;
10815 
10816         gc_seq = nft_gc_seq_begin(nft_net);
10817         ret = __nf_tables_abort(net, action);
10818         nft_gc_seq_end(nft_net, gc_seq);
10819 
10820         WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
10821 
10822         /* module autoload needs to happen after GC sequence update because it
10823          * temporarily releases and grabs mutex again.
10824          */
10825         if (action == NFNL_ABORT_AUTOLOAD)
10826                 nf_tables_module_autoload(net);
10827         else
10828                 nf_tables_module_autoload_cleanup(net);
10829 
10830         mutex_unlock(&nft_net->commit_mutex);
10831 
10832         return ret;
10833 }
10834 
10835 static bool nf_tables_valid_genid(struct net *net, u32 genid)
10836 {
10837         struct nftables_pernet *nft_net = nft_pernet(net);
10838         bool genid_ok;
10839 
10840         mutex_lock(&nft_net->commit_mutex);
10841         nft_net->tstamp = get_jiffies_64();
10842 
10843         genid_ok = genid == 0 || nft_net->base_seq == genid;
10844         if (!genid_ok)
10845                 mutex_unlock(&nft_net->commit_mutex);
10846 
10847         /* else, commit mutex has to be released by commit or abort function */
10848         return genid_ok;
10849 }
10850 
10851 static const struct nfnetlink_subsystem nf_tables_subsys = {
10852         .name           = "nf_tables",
10853         .subsys_id      = NFNL_SUBSYS_NFTABLES,
10854         .cb_count       = NFT_MSG_MAX,
10855         .cb             = nf_tables_cb,
10856         .commit         = nf_tables_commit,
10857         .abort          = nf_tables_abort,
10858         .valid_genid    = nf_tables_valid_genid,
10859         .owner          = THIS_MODULE,
10860 };
10861 
10862 int nft_chain_validate_dependency(const struct nft_chain *chain,
10863                                   enum nft_chain_types type)
10864 {
10865         const struct nft_base_chain *basechain;
10866 
10867         if (nft_is_base_chain(chain)) {
10868                 basechain = nft_base_chain(chain);
10869                 if (basechain->type->type != type)
10870                         return -EOPNOTSUPP;
10871         }
10872         return 0;
10873 }
10874 EXPORT_SYMBOL_GPL(nft_chain_validate_dependency);
10875 
10876 int nft_chain_validate_hooks(const struct nft_chain *chain,
10877                              unsigned int hook_flags)
10878 {
10879         struct nft_base_chain *basechain;
10880 
10881         if (nft_is_base_chain(chain)) {
10882                 basechain = nft_base_chain(chain);
10883 
10884                 if ((1 << basechain->ops.hooknum) & hook_flags)
10885                         return 0;
10886 
10887                 return -EOPNOTSUPP;
10888         }
10889 
10890         return 0;
10891 }
10892 EXPORT_SYMBOL_GPL(nft_chain_validate_hooks);
10893 
10894 /**
10895  *      nft_parse_u32_check - fetch u32 attribute and check for maximum value
10896  *
10897  *      @attr: netlink attribute to fetch value from
10898  *      @max: maximum value to be stored in dest
10899  *      @dest: pointer to the variable
10900  *
10901  *      Parse, check and store a given u32 netlink attribute into variable.
10902  *      This function returns -ERANGE if the value goes over maximum value.
10903  *      Otherwise a 0 is returned and the attribute value is stored in the
10904  *      destination variable.
10905  */
10906 int nft_parse_u32_check(const struct nlattr *attr, int max, u32 *dest)
10907 {
10908         u32 val;
10909 
10910         val = ntohl(nla_get_be32(attr));
10911         if (val > max)
10912                 return -ERANGE;
10913 
10914         *dest = val;
10915         return 0;
10916 }
10917 EXPORT_SYMBOL_GPL(nft_parse_u32_check);
10918 
10919 static int nft_parse_register(const struct nlattr *attr, u32 *preg)
10920 {
10921         unsigned int reg;
10922 
10923         reg = ntohl(nla_get_be32(attr));
10924         switch (reg) {
10925         case NFT_REG_VERDICT...NFT_REG_4:
10926                 *preg = reg * NFT_REG_SIZE / NFT_REG32_SIZE;
10927                 break;
10928         case NFT_REG32_00...NFT_REG32_15:
10929                 *preg = reg + NFT_REG_SIZE / NFT_REG32_SIZE - NFT_REG32_00;
10930                 break;
10931         default:
10932                 return -ERANGE;
10933         }
10934 
10935         return 0;
10936 }
10937 
10938 /**
10939  *      nft_dump_register - dump a register value to a netlink attribute
10940  *
10941  *      @skb: socket buffer
10942  *      @attr: attribute number
10943  *      @reg: register number
10944  *
10945  *      Construct a netlink attribute containing the register number. For
10946  *      compatibility reasons, register numbers being a multiple of 4 are
10947  *      translated to the corresponding 128 bit register numbers.
10948  */
10949 int nft_dump_register(struct sk_buff *skb, unsigned int attr, unsigned int reg)
10950 {
10951         if (reg % (NFT_REG_SIZE / NFT_REG32_SIZE) == 0)
10952                 reg = reg / (NFT_REG_SIZE / NFT_REG32_SIZE);
10953         else
10954                 reg = reg - NFT_REG_SIZE / NFT_REG32_SIZE + NFT_REG32_00;
10955 
10956         return nla_put_be32(skb, attr, htonl(reg));
10957 }
10958 EXPORT_SYMBOL_GPL(nft_dump_register);
10959 
10960 static int nft_validate_register_load(enum nft_registers reg, unsigned int len)
10961 {
10962         if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
10963                 return -EINVAL;
10964         if (len == 0)
10965                 return -EINVAL;
10966         if (reg * NFT_REG32_SIZE + len > sizeof_field(struct nft_regs, data))
10967                 return -ERANGE;
10968 
10969         return 0;
10970 }
10971 
10972 int nft_parse_register_load(const struct nlattr *attr, u8 *sreg, u32 len)
10973 {
10974         u32 reg;
10975         int err;
10976 
10977         err = nft_parse_register(attr, &reg);
10978         if (err < 0)
10979                 return err;
10980 
10981         err = nft_validate_register_load(reg, len);
10982         if (err < 0)
10983                 return err;
10984 
10985         *sreg = reg;
10986         return 0;
10987 }
10988 EXPORT_SYMBOL_GPL(nft_parse_register_load);
10989 
10990 static int nft_validate_register_store(const struct nft_ctx *ctx,
10991                                        enum nft_registers reg,
10992                                        const struct nft_data *data,
10993                                        enum nft_data_types type,
10994                                        unsigned int len)
10995 {
10996         int err;
10997 
10998         switch (reg) {
10999         case NFT_REG_VERDICT:
11000                 if (type != NFT_DATA_VERDICT)
11001                         return -EINVAL;
11002 
11003                 if (data != NULL &&
11004                     (data->verdict.code == NFT_GOTO ||
11005                      data->verdict.code == NFT_JUMP)) {
11006                         err = nft_chain_validate(ctx, data->verdict.chain);
11007                         if (err < 0)
11008                                 return err;
11009                 }
11010 
11011                 return 0;
11012         default:
11013                 if (type != NFT_DATA_VALUE)
11014                         return -EINVAL;
11015 
11016                 if (reg < NFT_REG_1 * NFT_REG_SIZE / NFT_REG32_SIZE)
11017                         return -EINVAL;
11018                 if (len == 0)
11019                         return -EINVAL;
11020                 if (reg * NFT_REG32_SIZE + len >
11021                     sizeof_field(struct nft_regs, data))
11022                         return -ERANGE;
11023 
11024                 return 0;
11025         }
11026 }
11027 
11028 int nft_parse_register_store(const struct nft_ctx *ctx,
11029                              const struct nlattr *attr, u8 *dreg,
11030                              const struct nft_data *data,
11031                              enum nft_data_types type, unsigned int len)
11032 {
11033         int err;
11034         u32 reg;
11035 
11036         err = nft_parse_register(attr, &reg);
11037         if (err < 0)
11038                 return err;
11039 
11040         err = nft_validate_register_store(ctx, reg, data, type, len);
11041         if (err < 0)
11042                 return err;
11043 
11044         *dreg = reg;
11045         return 0;
11046 }
11047 EXPORT_SYMBOL_GPL(nft_parse_register_store);
11048 
11049 static const struct nla_policy nft_verdict_policy[NFTA_VERDICT_MAX + 1] = {
11050         [NFTA_VERDICT_CODE]     = { .type = NLA_U32 },
11051         [NFTA_VERDICT_CHAIN]    = { .type = NLA_STRING,
11052                                     .len = NFT_CHAIN_MAXNAMELEN - 1 },
11053         [NFTA_VERDICT_CHAIN_ID] = { .type = NLA_U32 },
11054 };
11055 
11056 static int nft_verdict_init(const struct nft_ctx *ctx, struct nft_data *data,
11057                             struct nft_data_desc *desc, const struct nlattr *nla)
11058 {
11059         u8 genmask = nft_genmask_next(ctx->net);
11060         struct nlattr *tb[NFTA_VERDICT_MAX + 1];
11061         struct nft_chain *chain;
11062         int err;
11063 
11064         err = nla_parse_nested_deprecated(tb, NFTA_VERDICT_MAX, nla,
11065                                           nft_verdict_policy, NULL);
11066         if (err < 0)
11067                 return err;
11068 
11069         if (!tb[NFTA_VERDICT_CODE])
11070                 return -EINVAL;
11071 
11072         /* zero padding hole for memcmp */
11073         memset(data, 0, sizeof(*data));
11074         data->verdict.code = ntohl(nla_get_be32(tb[NFTA_VERDICT_CODE]));
11075 
11076         switch (data->verdict.code) {
11077         case NF_ACCEPT:
11078         case NF_DROP:
11079         case NF_QUEUE:
11080                 break;
11081         case NFT_CONTINUE:
11082         case NFT_BREAK:
11083         case NFT_RETURN:
11084                 break;
11085         case NFT_JUMP:
11086         case NFT_GOTO:
11087                 if (tb[NFTA_VERDICT_CHAIN]) {
11088                         chain = nft_chain_lookup(ctx->net, ctx->table,
11089                                                  tb[NFTA_VERDICT_CHAIN],
11090                                                  genmask);
11091                 } else if (tb[NFTA_VERDICT_CHAIN_ID]) {
11092                         chain = nft_chain_lookup_byid(ctx->net, ctx->table,
11093                                                       tb[NFTA_VERDICT_CHAIN_ID],
11094                                                       genmask);
11095                         if (IS_ERR(chain))
11096                                 return PTR_ERR(chain);
11097                 } else {
11098                         return -EINVAL;
11099                 }
11100 
11101                 if (IS_ERR(chain))
11102                         return PTR_ERR(chain);
11103                 if (nft_is_base_chain(chain))
11104                         return -EOPNOTSUPP;
11105                 if (nft_chain_is_bound(chain))
11106                         return -EINVAL;
11107                 if (desc->flags & NFT_DATA_DESC_SETELEM &&
11108                     chain->flags & NFT_CHAIN_BINDING)
11109                         return -EINVAL;
11110                 if (!nft_use_inc(&chain->use))
11111                         return -EMFILE;
11112 
11113                 data->verdict.chain = chain;
11114                 break;
11115         default:
11116                 return -EINVAL;
11117         }
11118 
11119         desc->len = sizeof(data->verdict);
11120 
11121         return 0;
11122 }
11123 
11124 static void nft_verdict_uninit(const struct nft_data *data)
11125 {
11126         struct nft_chain *chain;
11127 
11128         switch (data->verdict.code) {
11129         case NFT_JUMP:
11130         case NFT_GOTO:
11131                 chain = data->verdict.chain;
11132                 nft_use_dec(&chain->use);
11133                 break;
11134         }
11135 }
11136 
11137 int nft_verdict_dump(struct sk_buff *skb, int type, const struct nft_verdict *v)
11138 {
11139         struct nlattr *nest;
11140 
11141         nest = nla_nest_start_noflag(skb, type);
11142         if (!nest)
11143                 goto nla_put_failure;
11144 
11145         if (nla_put_be32(skb, NFTA_VERDICT_CODE, htonl(v->code)))
11146                 goto nla_put_failure;
11147 
11148         switch (v->code) {
11149         case NFT_JUMP:
11150         case NFT_GOTO:
11151                 if (nla_put_string(skb, NFTA_VERDICT_CHAIN,
11152                                    v->chain->name))
11153                         goto nla_put_failure;
11154         }
11155         nla_nest_end(skb, nest);
11156         return 0;
11157 
11158 nla_put_failure:
11159         return -1;
11160 }
11161 
11162 static int nft_value_init(const struct nft_ctx *ctx,
11163                           struct nft_data *data, struct nft_data_desc *desc,
11164                           const struct nlattr *nla)
11165 {
11166         unsigned int len;
11167 
11168         len = nla_len(nla);
11169         if (len == 0)
11170                 return -EINVAL;
11171         if (len > desc->size)
11172                 return -EOVERFLOW;
11173         if (desc->len) {
11174                 if (len != desc->len)
11175                         return -EINVAL;
11176         } else {
11177                 desc->len = len;
11178         }
11179 
11180         nla_memcpy(data->data, nla, len);
11181 
11182         return 0;
11183 }
11184 
11185 static int nft_value_dump(struct sk_buff *skb, const struct nft_data *data,
11186                           unsigned int len)
11187 {
11188         return nla_put(skb, NFTA_DATA_VALUE, len, data->data);
11189 }
11190 
11191 static const struct nla_policy nft_data_policy[NFTA_DATA_MAX + 1] = {
11192         [NFTA_DATA_VALUE]       = { .type = NLA_BINARY },
11193         [NFTA_DATA_VERDICT]     = { .type = NLA_NESTED },
11194 };
11195 
11196 /**
11197  *      nft_data_init - parse nf_tables data netlink attributes
11198  *
11199  *      @ctx: context of the expression using the data
11200  *      @data: destination struct nft_data
11201  *      @desc: data description
11202  *      @nla: netlink attribute containing data
11203  *
11204  *      Parse the netlink data attributes and initialize a struct nft_data.
11205  *      The type and length of data are returned in the data description.
11206  *
11207  *      The caller can indicate that it only wants to accept data of type
11208  *      NFT_DATA_VALUE by passing NULL for the ctx argument.
11209  */
11210 int nft_data_init(const struct nft_ctx *ctx, struct nft_data *data,
11211                   struct nft_data_desc *desc, const struct nlattr *nla)
11212 {
11213         struct nlattr *tb[NFTA_DATA_MAX + 1];
11214         int err;
11215 
11216         if (WARN_ON_ONCE(!desc->size))
11217                 return -EINVAL;
11218 
11219         err = nla_parse_nested_deprecated(tb, NFTA_DATA_MAX, nla,
11220                                           nft_data_policy, NULL);
11221         if (err < 0)
11222                 return err;
11223 
11224         if (tb[NFTA_DATA_VALUE]) {
11225                 if (desc->type != NFT_DATA_VALUE)
11226                         return -EINVAL;
11227 
11228                 err = nft_value_init(ctx, data, desc, tb[NFTA_DATA_VALUE]);
11229         } else if (tb[NFTA_DATA_VERDICT] && ctx != NULL) {
11230                 if (desc->type != NFT_DATA_VERDICT)
11231                         return -EINVAL;
11232 
11233                 err = nft_verdict_init(ctx, data, desc, tb[NFTA_DATA_VERDICT]);
11234         } else {
11235                 err = -EINVAL;
11236         }
11237 
11238         return err;
11239 }
11240 EXPORT_SYMBOL_GPL(nft_data_init);
11241 
11242 /**
11243  *      nft_data_release - release a nft_data item
11244  *
11245  *      @data: struct nft_data to release
11246  *      @type: type of data
11247  *
11248  *      Release a nft_data item. NFT_DATA_VALUE types can be silently discarded,
11249  *      all others need to be released by calling this function.
11250  */
11251 void nft_data_release(const struct nft_data *data, enum nft_data_types type)
11252 {
11253         if (type < NFT_DATA_VERDICT)
11254                 return;
11255         switch (type) {
11256         case NFT_DATA_VERDICT:
11257                 return nft_verdict_uninit(data);
11258         default:
11259                 WARN_ON(1);
11260         }
11261 }
11262 EXPORT_SYMBOL_GPL(nft_data_release);
11263 
11264 int nft_data_dump(struct sk_buff *skb, int attr, const struct nft_data *data,
11265                   enum nft_data_types type, unsigned int len)
11266 {
11267         struct nlattr *nest;
11268         int err;
11269 
11270         nest = nla_nest_start_noflag(skb, attr);
11271         if (nest == NULL)
11272                 return -1;
11273 
11274         switch (type) {
11275         case NFT_DATA_VALUE:
11276                 err = nft_value_dump(skb, data, len);
11277                 break;
11278         case NFT_DATA_VERDICT:
11279                 err = nft_verdict_dump(skb, NFTA_DATA_VERDICT, &data->verdict);
11280                 break;
11281         default:
11282                 err = -EINVAL;
11283                 WARN_ON(1);
11284         }
11285 
11286         nla_nest_end(skb, nest);
11287         return err;
11288 }
11289 EXPORT_SYMBOL_GPL(nft_data_dump);
11290 
11291 int __nft_release_basechain(struct nft_ctx *ctx)
11292 {
11293         struct nft_rule *rule, *nr;
11294 
11295         if (WARN_ON(!nft_is_base_chain(ctx->chain)))
11296                 return 0;
11297 
11298         nf_tables_unregister_hook(ctx->net, ctx->chain->table, ctx->chain);
11299         list_for_each_entry_safe(rule, nr, &ctx->chain->rules, list) {
11300                 list_del(&rule->list);
11301                 nft_use_dec(&ctx->chain->use);
11302                 nf_tables_rule_release(ctx, rule);
11303         }
11304         nft_chain_del(ctx->chain);
11305         nft_use_dec(&ctx->table->use);
11306         nf_tables_chain_destroy(ctx->chain);
11307 
11308         return 0;
11309 }
11310 EXPORT_SYMBOL_GPL(__nft_release_basechain);
11311 
11312 static void __nft_release_hook(struct net *net, struct nft_table *table)
11313 {
11314         struct nft_flowtable *flowtable;
11315         struct nft_chain *chain;
11316 
11317         list_for_each_entry(chain, &table->chains, list)
11318                 __nf_tables_unregister_hook(net, table, chain, true);
11319         list_for_each_entry(flowtable, &table->flowtables, list)
11320                 __nft_unregister_flowtable_net_hooks(net, &flowtable->hook_list,
11321                                                      true);
11322 }
11323 
11324 static void __nft_release_hooks(struct net *net)
11325 {
11326         struct nftables_pernet *nft_net = nft_pernet(net);
11327         struct nft_table *table;
11328 
11329         list_for_each_entry(table, &nft_net->tables, list) {
11330                 if (nft_table_has_owner(table))
11331                         continue;
11332 
11333                 __nft_release_hook(net, table);
11334         }
11335 }
11336 
11337 static void __nft_release_table(struct net *net, struct nft_table *table)
11338 {
11339         struct nft_flowtable *flowtable, *nf;
11340         struct nft_chain *chain, *nc;
11341         struct nft_object *obj, *ne;
11342         struct nft_rule *rule, *nr;
11343         struct nft_set *set, *ns;
11344         struct nft_ctx ctx = {
11345                 .net    = net,
11346                 .family = NFPROTO_NETDEV,
11347         };
11348 
11349         ctx.family = table->family;
11350         ctx.table = table;
11351         list_for_each_entry(chain, &table->chains, list) {
11352                 if (nft_chain_binding(chain))
11353                         continue;
11354 
11355                 ctx.chain = chain;
11356                 list_for_each_entry_safe(rule, nr, &chain->rules, list) {
11357                         list_del(&rule->list);
11358                         nft_use_dec(&chain->use);
11359                         nf_tables_rule_release(&ctx, rule);
11360                 }
11361         }
11362         list_for_each_entry_safe(flowtable, nf, &table->flowtables, list) {
11363                 list_del(&flowtable->list);
11364                 nft_use_dec(&table->use);
11365                 nf_tables_flowtable_destroy(flowtable);
11366         }
11367         list_for_each_entry_safe(set, ns, &table->sets, list) {
11368                 list_del(&set->list);
11369                 nft_use_dec(&table->use);
11370                 if (set->flags & (NFT_SET_MAP | NFT_SET_OBJECT))
11371                         nft_map_deactivate(&ctx, set);
11372 
11373                 nft_set_destroy(&ctx, set);
11374         }
11375         list_for_each_entry_safe(obj, ne, &table->objects, list) {
11376                 nft_obj_del(obj);
11377                 nft_use_dec(&table->use);
11378                 nft_obj_destroy(&ctx, obj);
11379         }
11380         list_for_each_entry_safe(chain, nc, &table->chains, list) {
11381                 nft_chain_del(chain);
11382                 nft_use_dec(&table->use);
11383                 nf_tables_chain_destroy(chain);
11384         }
11385         nf_tables_table_destroy(table);
11386 }
11387 
11388 static void __nft_release_tables(struct net *net)
11389 {
11390         struct nftables_pernet *nft_net = nft_pernet(net);
11391         struct nft_table *table, *nt;
11392 
11393         list_for_each_entry_safe(table, nt, &nft_net->tables, list) {
11394                 if (nft_table_has_owner(table))
11395                         continue;
11396 
11397                 list_del(&table->list);
11398 
11399                 __nft_release_table(net, table);
11400         }
11401 }
11402 
11403 static int nft_rcv_nl_event(struct notifier_block *this, unsigned long event,
11404                             void *ptr)
11405 {
11406         struct nft_table *table, *to_delete[8];
11407         struct nftables_pernet *nft_net;
11408         struct netlink_notify *n = ptr;
11409         struct net *net = n->net;
11410         unsigned int deleted;
11411         bool restart = false;
11412         unsigned int gc_seq;
11413 
11414         if (event != NETLINK_URELEASE || n->protocol != NETLINK_NETFILTER)
11415                 return NOTIFY_DONE;
11416 
11417         nft_net = nft_pernet(net);
11418         deleted = 0;
11419         mutex_lock(&nft_net->commit_mutex);
11420 
11421         gc_seq = nft_gc_seq_begin(nft_net);
11422 
11423         nf_tables_trans_destroy_flush_work();
11424 again:
11425         list_for_each_entry(table, &nft_net->tables, list) {
11426                 if (nft_table_has_owner(table) &&
11427                     n->portid == table->nlpid) {
11428                         if (table->flags & NFT_TABLE_F_PERSIST) {
11429                                 table->flags &= ~NFT_TABLE_F_OWNER;
11430                                 continue;
11431                         }
11432                         __nft_release_hook(net, table);
11433                         list_del_rcu(&table->list);
11434                         to_delete[deleted++] = table;
11435                         if (deleted >= ARRAY_SIZE(to_delete))
11436                                 break;
11437                 }
11438         }
11439         if (deleted) {
11440                 restart = deleted >= ARRAY_SIZE(to_delete);
11441                 synchronize_rcu();
11442                 while (deleted)
11443                         __nft_release_table(net, to_delete[--deleted]);
11444 
11445                 if (restart)
11446                         goto again;
11447         }
11448         nft_gc_seq_end(nft_net, gc_seq);
11449 
11450         mutex_unlock(&nft_net->commit_mutex);
11451 
11452         return NOTIFY_DONE;
11453 }
11454 
11455 static struct notifier_block nft_nl_notifier = {
11456         .notifier_call  = nft_rcv_nl_event,
11457 };
11458 
11459 static int __net_init nf_tables_init_net(struct net *net)
11460 {
11461         struct nftables_pernet *nft_net = nft_pernet(net);
11462 
11463         INIT_LIST_HEAD(&nft_net->tables);
11464         INIT_LIST_HEAD(&nft_net->commit_list);
11465         INIT_LIST_HEAD(&nft_net->binding_list);
11466         INIT_LIST_HEAD(&nft_net->module_list);
11467         INIT_LIST_HEAD(&nft_net->notify_list);
11468         mutex_init(&nft_net->commit_mutex);
11469         nft_net->base_seq = 1;
11470         nft_net->gc_seq = 0;
11471         nft_net->validate_state = NFT_VALIDATE_SKIP;
11472 
11473         return 0;
11474 }
11475 
11476 static void __net_exit nf_tables_pre_exit_net(struct net *net)
11477 {
11478         struct nftables_pernet *nft_net = nft_pernet(net);
11479 
11480         mutex_lock(&nft_net->commit_mutex);
11481         __nft_release_hooks(net);
11482         mutex_unlock(&nft_net->commit_mutex);
11483 }
11484 
11485 static void __net_exit nf_tables_exit_net(struct net *net)
11486 {
11487         struct nftables_pernet *nft_net = nft_pernet(net);
11488         unsigned int gc_seq;
11489 
11490         mutex_lock(&nft_net->commit_mutex);
11491 
11492         gc_seq = nft_gc_seq_begin(nft_net);
11493 
11494         WARN_ON_ONCE(!list_empty(&nft_net->commit_list));
11495 
11496         if (!list_empty(&nft_net->module_list))
11497                 nf_tables_module_autoload_cleanup(net);
11498 
11499         __nft_release_tables(net);
11500 
11501         nft_gc_seq_end(nft_net, gc_seq);
11502 
11503         mutex_unlock(&nft_net->commit_mutex);
11504         WARN_ON_ONCE(!list_empty(&nft_net->tables));
11505         WARN_ON_ONCE(!list_empty(&nft_net->module_list));
11506         WARN_ON_ONCE(!list_empty(&nft_net->notify_list));
11507 }
11508 
11509 static void nf_tables_exit_batch(struct list_head *net_exit_list)
11510 {
11511         flush_work(&trans_gc_work);
11512 }
11513 
11514 static struct pernet_operations nf_tables_net_ops = {
11515         .init           = nf_tables_init_net,
11516         .pre_exit       = nf_tables_pre_exit_net,
11517         .exit           = nf_tables_exit_net,
11518         .exit_batch     = nf_tables_exit_batch,
11519         .id             = &nf_tables_net_id,
11520         .size           = sizeof(struct nftables_pernet),
11521 };
11522 
11523 static int __init nf_tables_module_init(void)
11524 {
11525         int err;
11526 
11527         BUILD_BUG_ON(offsetof(struct nft_trans_table, nft_trans) != 0);
11528         BUILD_BUG_ON(offsetof(struct nft_trans_chain, nft_trans_binding.nft_trans) != 0);
11529         BUILD_BUG_ON(offsetof(struct nft_trans_rule, nft_trans) != 0);
11530         BUILD_BUG_ON(offsetof(struct nft_trans_set, nft_trans_binding.nft_trans) != 0);
11531         BUILD_BUG_ON(offsetof(struct nft_trans_elem, nft_trans) != 0);
11532         BUILD_BUG_ON(offsetof(struct nft_trans_obj, nft_trans) != 0);
11533         BUILD_BUG_ON(offsetof(struct nft_trans_flowtable, nft_trans) != 0);
11534 
11535         err = register_pernet_subsys(&nf_tables_net_ops);
11536         if (err < 0)
11537                 return err;
11538 
11539         err = nft_chain_filter_init();
11540         if (err < 0)
11541                 goto err_chain_filter;
11542 
11543         err = nf_tables_core_module_init();
11544         if (err < 0)
11545                 goto err_core_module;
11546 
11547         err = register_netdevice_notifier(&nf_tables_flowtable_notifier);
11548         if (err < 0)
11549                 goto err_netdev_notifier;
11550 
11551         err = rhltable_init(&nft_objname_ht, &nft_objname_ht_params);
11552         if (err < 0)
11553                 goto err_rht_objname;
11554 
11555         err = nft_offload_init();
11556         if (err < 0)
11557                 goto err_offload;
11558 
11559         err = netlink_register_notifier(&nft_nl_notifier);
11560         if (err < 0)
11561                 goto err_netlink_notifier;
11562 
11563         /* must be last */
11564         err = nfnetlink_subsys_register(&nf_tables_subsys);
11565         if (err < 0)
11566                 goto err_nfnl_subsys;
11567 
11568         nft_chain_route_init();
11569 
11570         return err;
11571 
11572 err_nfnl_subsys:
11573         netlink_unregister_notifier(&nft_nl_notifier);
11574 err_netlink_notifier:
11575         nft_offload_exit();
11576 err_offload:
11577         rhltable_destroy(&nft_objname_ht);
11578 err_rht_objname:
11579         unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
11580 err_netdev_notifier:
11581         nf_tables_core_module_exit();
11582 err_core_module:
11583         nft_chain_filter_fini();
11584 err_chain_filter:
11585         unregister_pernet_subsys(&nf_tables_net_ops);
11586         return err;
11587 }
11588 
11589 static void __exit nf_tables_module_exit(void)
11590 {
11591         nfnetlink_subsys_unregister(&nf_tables_subsys);
11592         netlink_unregister_notifier(&nft_nl_notifier);
11593         nft_offload_exit();
11594         unregister_netdevice_notifier(&nf_tables_flowtable_notifier);
11595         nft_chain_filter_fini();
11596         nft_chain_route_fini();
11597         nf_tables_trans_destroy_flush_work();
11598         unregister_pernet_subsys(&nf_tables_net_ops);
11599         cancel_work_sync(&trans_gc_work);
11600         cancel_work_sync(&trans_destroy_work);
11601         rcu_barrier();
11602         rhltable_destroy(&nft_objname_ht);
11603         nf_tables_core_module_exit();
11604 }
11605 
11606 module_init(nf_tables_module_init);
11607 module_exit(nf_tables_module_exit);
11608 
11609 MODULE_LICENSE("GPL");
11610 MODULE_AUTHOR("Patrick McHardy <kaber@trash.net>");
11611 MODULE_DESCRIPTION("Framework for packet filtering and classification");
11612 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_NFTABLES);
11613 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

sflogo.php