]> asedeno.scripts.mit.edu Git - linux.git/blob - net/sched/cls_flower.c
Merge branch 'work.namei' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[linux.git] / net / sched / cls_flower.c
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * net/sched/cls_flower.c               Flower classifier
4  *
5  * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us>
6  */
7
8 #include <linux/kernel.h>
9 #include <linux/init.h>
10 #include <linux/module.h>
11 #include <linux/rhashtable.h>
12 #include <linux/workqueue.h>
13 #include <linux/refcount.h>
14
15 #include <linux/if_ether.h>
16 #include <linux/in6.h>
17 #include <linux/ip.h>
18 #include <linux/mpls.h>
19
20 #include <net/sch_generic.h>
21 #include <net/pkt_cls.h>
22 #include <net/ip.h>
23 #include <net/flow_dissector.h>
24 #include <net/geneve.h>
25
26 #include <net/dst.h>
27 #include <net/dst_metadata.h>
28
29 #include <uapi/linux/netfilter/nf_conntrack_common.h>
30
31 struct fl_flow_key {
32         struct flow_dissector_key_meta meta;
33         struct flow_dissector_key_control control;
34         struct flow_dissector_key_control enc_control;
35         struct flow_dissector_key_basic basic;
36         struct flow_dissector_key_eth_addrs eth;
37         struct flow_dissector_key_vlan vlan;
38         struct flow_dissector_key_vlan cvlan;
39         union {
40                 struct flow_dissector_key_ipv4_addrs ipv4;
41                 struct flow_dissector_key_ipv6_addrs ipv6;
42         };
43         struct flow_dissector_key_ports tp;
44         struct flow_dissector_key_icmp icmp;
45         struct flow_dissector_key_arp arp;
46         struct flow_dissector_key_keyid enc_key_id;
47         union {
48                 struct flow_dissector_key_ipv4_addrs enc_ipv4;
49                 struct flow_dissector_key_ipv6_addrs enc_ipv6;
50         };
51         struct flow_dissector_key_ports enc_tp;
52         struct flow_dissector_key_mpls mpls;
53         struct flow_dissector_key_tcp tcp;
54         struct flow_dissector_key_ip ip;
55         struct flow_dissector_key_ip enc_ip;
56         struct flow_dissector_key_enc_opts enc_opts;
57         struct flow_dissector_key_ports tp_min;
58         struct flow_dissector_key_ports tp_max;
59         struct flow_dissector_key_ct ct;
60 } __aligned(BITS_PER_LONG / 8); /* Ensure that we can do comparisons as longs. */
61
62 struct fl_flow_mask_range {
63         unsigned short int start;
64         unsigned short int end;
65 };
66
67 struct fl_flow_mask {
68         struct fl_flow_key key;
69         struct fl_flow_mask_range range;
70         u32 flags;
71         struct rhash_head ht_node;
72         struct rhashtable ht;
73         struct rhashtable_params filter_ht_params;
74         struct flow_dissector dissector;
75         struct list_head filters;
76         struct rcu_work rwork;
77         struct list_head list;
78         refcount_t refcnt;
79 };
80
81 struct fl_flow_tmplt {
82         struct fl_flow_key dummy_key;
83         struct fl_flow_key mask;
84         struct flow_dissector dissector;
85         struct tcf_chain *chain;
86 };
87
88 struct cls_fl_head {
89         struct rhashtable ht;
90         spinlock_t masks_lock; /* Protect masks list */
91         struct list_head masks;
92         struct list_head hw_filters;
93         struct rcu_work rwork;
94         struct idr handle_idr;
95 };
96
97 struct cls_fl_filter {
98         struct fl_flow_mask *mask;
99         struct rhash_head ht_node;
100         struct fl_flow_key mkey;
101         struct tcf_exts exts;
102         struct tcf_result res;
103         struct fl_flow_key key;
104         struct list_head list;
105         struct list_head hw_list;
106         u32 handle;
107         u32 flags;
108         u32 in_hw_count;
109         struct rcu_work rwork;
110         struct net_device *hw_dev;
111         /* Flower classifier is unlocked, which means that its reference counter
112          * can be changed concurrently without any kind of external
113          * synchronization. Use atomic reference counter to be concurrency-safe.
114          */
115         refcount_t refcnt;
116         bool deleted;
117 };
118
119 static const struct rhashtable_params mask_ht_params = {
120         .key_offset = offsetof(struct fl_flow_mask, key),
121         .key_len = sizeof(struct fl_flow_key),
122         .head_offset = offsetof(struct fl_flow_mask, ht_node),
123         .automatic_shrinking = true,
124 };
125
126 static unsigned short int fl_mask_range(const struct fl_flow_mask *mask)
127 {
128         return mask->range.end - mask->range.start;
129 }
130
131 static void fl_mask_update_range(struct fl_flow_mask *mask)
132 {
133         const u8 *bytes = (const u8 *) &mask->key;
134         size_t size = sizeof(mask->key);
135         size_t i, first = 0, last;
136
137         for (i = 0; i < size; i++) {
138                 if (bytes[i]) {
139                         first = i;
140                         break;
141                 }
142         }
143         last = first;
144         for (i = size - 1; i != first; i--) {
145                 if (bytes[i]) {
146                         last = i;
147                         break;
148                 }
149         }
150         mask->range.start = rounddown(first, sizeof(long));
151         mask->range.end = roundup(last + 1, sizeof(long));
152 }
153
154 static void *fl_key_get_start(struct fl_flow_key *key,
155                               const struct fl_flow_mask *mask)
156 {
157         return (u8 *) key + mask->range.start;
158 }
159
160 static void fl_set_masked_key(struct fl_flow_key *mkey, struct fl_flow_key *key,
161                               struct fl_flow_mask *mask)
162 {
163         const long *lkey = fl_key_get_start(key, mask);
164         const long *lmask = fl_key_get_start(&mask->key, mask);
165         long *lmkey = fl_key_get_start(mkey, mask);
166         int i;
167
168         for (i = 0; i < fl_mask_range(mask); i += sizeof(long))
169                 *lmkey++ = *lkey++ & *lmask++;
170 }
171
172 static bool fl_mask_fits_tmplt(struct fl_flow_tmplt *tmplt,
173                                struct fl_flow_mask *mask)
174 {
175         const long *lmask = fl_key_get_start(&mask->key, mask);
176         const long *ltmplt;
177         int i;
178
179         if (!tmplt)
180                 return true;
181         ltmplt = fl_key_get_start(&tmplt->mask, mask);
182         for (i = 0; i < fl_mask_range(mask); i += sizeof(long)) {
183                 if (~*ltmplt++ & *lmask++)
184                         return false;
185         }
186         return true;
187 }
188
189 static void fl_clear_masked_range(struct fl_flow_key *key,
190                                   struct fl_flow_mask *mask)
191 {
192         memset(fl_key_get_start(key, mask), 0, fl_mask_range(mask));
193 }
194
195 static bool fl_range_port_dst_cmp(struct cls_fl_filter *filter,
196                                   struct fl_flow_key *key,
197                                   struct fl_flow_key *mkey)
198 {
199         __be16 min_mask, max_mask, min_val, max_val;
200
201         min_mask = htons(filter->mask->key.tp_min.dst);
202         max_mask = htons(filter->mask->key.tp_max.dst);
203         min_val = htons(filter->key.tp_min.dst);
204         max_val = htons(filter->key.tp_max.dst);
205
206         if (min_mask && max_mask) {
207                 if (htons(key->tp.dst) < min_val ||
208                     htons(key->tp.dst) > max_val)
209                         return false;
210
211                 /* skb does not have min and max values */
212                 mkey->tp_min.dst = filter->mkey.tp_min.dst;
213                 mkey->tp_max.dst = filter->mkey.tp_max.dst;
214         }
215         return true;
216 }
217
218 static bool fl_range_port_src_cmp(struct cls_fl_filter *filter,
219                                   struct fl_flow_key *key,
220                                   struct fl_flow_key *mkey)
221 {
222         __be16 min_mask, max_mask, min_val, max_val;
223
224         min_mask = htons(filter->mask->key.tp_min.src);
225         max_mask = htons(filter->mask->key.tp_max.src);
226         min_val = htons(filter->key.tp_min.src);
227         max_val = htons(filter->key.tp_max.src);
228
229         if (min_mask && max_mask) {
230                 if (htons(key->tp.src) < min_val ||
231                     htons(key->tp.src) > max_val)
232                         return false;
233
234                 /* skb does not have min and max values */
235                 mkey->tp_min.src = filter->mkey.tp_min.src;
236                 mkey->tp_max.src = filter->mkey.tp_max.src;
237         }
238         return true;
239 }
240
241 static struct cls_fl_filter *__fl_lookup(struct fl_flow_mask *mask,
242                                          struct fl_flow_key *mkey)
243 {
244         return rhashtable_lookup_fast(&mask->ht, fl_key_get_start(mkey, mask),
245                                       mask->filter_ht_params);
246 }
247
248 static struct cls_fl_filter *fl_lookup_range(struct fl_flow_mask *mask,
249                                              struct fl_flow_key *mkey,
250                                              struct fl_flow_key *key)
251 {
252         struct cls_fl_filter *filter, *f;
253
254         list_for_each_entry_rcu(filter, &mask->filters, list) {
255                 if (!fl_range_port_dst_cmp(filter, key, mkey))
256                         continue;
257
258                 if (!fl_range_port_src_cmp(filter, key, mkey))
259                         continue;
260
261                 f = __fl_lookup(mask, mkey);
262                 if (f)
263                         return f;
264         }
265         return NULL;
266 }
267
268 static struct cls_fl_filter *fl_lookup(struct fl_flow_mask *mask,
269                                        struct fl_flow_key *mkey,
270                                        struct fl_flow_key *key)
271 {
272         if ((mask->flags & TCA_FLOWER_MASK_FLAGS_RANGE))
273                 return fl_lookup_range(mask, mkey, key);
274
275         return __fl_lookup(mask, mkey);
276 }
277
278 static u16 fl_ct_info_to_flower_map[] = {
279         [IP_CT_ESTABLISHED] =           TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
280                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
281         [IP_CT_RELATED] =               TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
282                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
283         [IP_CT_ESTABLISHED_REPLY] =     TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
284                                         TCA_FLOWER_KEY_CT_FLAGS_ESTABLISHED,
285         [IP_CT_RELATED_REPLY] =         TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
286                                         TCA_FLOWER_KEY_CT_FLAGS_RELATED,
287         [IP_CT_NEW] =                   TCA_FLOWER_KEY_CT_FLAGS_TRACKED |
288                                         TCA_FLOWER_KEY_CT_FLAGS_NEW,
289 };
290
291 static int fl_classify(struct sk_buff *skb, const struct tcf_proto *tp,
292                        struct tcf_result *res)
293 {
294         struct cls_fl_head *head = rcu_dereference_bh(tp->root);
295         struct fl_flow_key skb_mkey;
296         struct fl_flow_key skb_key;
297         struct fl_flow_mask *mask;
298         struct cls_fl_filter *f;
299
300         list_for_each_entry_rcu(mask, &head->masks, list) {
301                 fl_clear_masked_range(&skb_key, mask);
302
303                 skb_flow_dissect_meta(skb, &mask->dissector, &skb_key);
304                 /* skb_flow_dissect() does not set n_proto in case an unknown
305                  * protocol, so do it rather here.
306                  */
307                 skb_key.basic.n_proto = skb->protocol;
308                 skb_flow_dissect_tunnel_info(skb, &mask->dissector, &skb_key);
309                 skb_flow_dissect_ct(skb, &mask->dissector, &skb_key,
310                                     fl_ct_info_to_flower_map,
311                                     ARRAY_SIZE(fl_ct_info_to_flower_map));
312                 skb_flow_dissect(skb, &mask->dissector, &skb_key, 0);
313
314                 fl_set_masked_key(&skb_mkey, &skb_key, mask);
315
316                 f = fl_lookup(mask, &skb_mkey, &skb_key);
317                 if (f && !tc_skip_sw(f->flags)) {
318                         *res = f->res;
319                         return tcf_exts_exec(skb, &f->exts, res);
320                 }
321         }
322         return -1;
323 }
324
325 static int fl_init(struct tcf_proto *tp)
326 {
327         struct cls_fl_head *head;
328
329         head = kzalloc(sizeof(*head), GFP_KERNEL);
330         if (!head)
331                 return -ENOBUFS;
332
333         spin_lock_init(&head->masks_lock);
334         INIT_LIST_HEAD_RCU(&head->masks);
335         INIT_LIST_HEAD(&head->hw_filters);
336         rcu_assign_pointer(tp->root, head);
337         idr_init(&head->handle_idr);
338
339         return rhashtable_init(&head->ht, &mask_ht_params);
340 }
341
342 static void fl_mask_free(struct fl_flow_mask *mask, bool mask_init_done)
343 {
344         /* temporary masks don't have their filters list and ht initialized */
345         if (mask_init_done) {
346                 WARN_ON(!list_empty(&mask->filters));
347                 rhashtable_destroy(&mask->ht);
348         }
349         kfree(mask);
350 }
351
352 static void fl_mask_free_work(struct work_struct *work)
353 {
354         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
355                                                  struct fl_flow_mask, rwork);
356
357         fl_mask_free(mask, true);
358 }
359
360 static void fl_uninit_mask_free_work(struct work_struct *work)
361 {
362         struct fl_flow_mask *mask = container_of(to_rcu_work(work),
363                                                  struct fl_flow_mask, rwork);
364
365         fl_mask_free(mask, false);
366 }
367
368 static bool fl_mask_put(struct cls_fl_head *head, struct fl_flow_mask *mask)
369 {
370         if (!refcount_dec_and_test(&mask->refcnt))
371                 return false;
372
373         rhashtable_remove_fast(&head->ht, &mask->ht_node, mask_ht_params);
374
375         spin_lock(&head->masks_lock);
376         list_del_rcu(&mask->list);
377         spin_unlock(&head->masks_lock);
378
379         tcf_queue_work(&mask->rwork, fl_mask_free_work);
380
381         return true;
382 }
383
384 static struct cls_fl_head *fl_head_dereference(struct tcf_proto *tp)
385 {
386         /* Flower classifier only changes root pointer during init and destroy.
387          * Users must obtain reference to tcf_proto instance before calling its
388          * API, so tp->root pointer is protected from concurrent call to
389          * fl_destroy() by reference counting.
390          */
391         return rcu_dereference_raw(tp->root);
392 }
393
394 static void __fl_destroy_filter(struct cls_fl_filter *f)
395 {
396         tcf_exts_destroy(&f->exts);
397         tcf_exts_put_net(&f->exts);
398         kfree(f);
399 }
400
401 static void fl_destroy_filter_work(struct work_struct *work)
402 {
403         struct cls_fl_filter *f = container_of(to_rcu_work(work),
404                                         struct cls_fl_filter, rwork);
405
406         __fl_destroy_filter(f);
407 }
408
409 static void fl_hw_destroy_filter(struct tcf_proto *tp, struct cls_fl_filter *f,
410                                  bool rtnl_held, struct netlink_ext_ack *extack)
411 {
412         struct tcf_block *block = tp->chain->block;
413         struct flow_cls_offload cls_flower = {};
414
415         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
416         cls_flower.command = FLOW_CLS_DESTROY;
417         cls_flower.cookie = (unsigned long) f;
418
419         tc_setup_cb_destroy(block, tp, TC_SETUP_CLSFLOWER, &cls_flower, false,
420                             &f->flags, &f->in_hw_count, rtnl_held);
421
422 }
423
424 static int fl_hw_replace_filter(struct tcf_proto *tp,
425                                 struct cls_fl_filter *f, bool rtnl_held,
426                                 struct netlink_ext_ack *extack)
427 {
428         struct tcf_block *block = tp->chain->block;
429         struct flow_cls_offload cls_flower = {};
430         bool skip_sw = tc_skip_sw(f->flags);
431         int err = 0;
432
433         cls_flower.rule = flow_rule_alloc(tcf_exts_num_actions(&f->exts));
434         if (!cls_flower.rule)
435                 return -ENOMEM;
436
437         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, extack);
438         cls_flower.command = FLOW_CLS_REPLACE;
439         cls_flower.cookie = (unsigned long) f;
440         cls_flower.rule->match.dissector = &f->mask->dissector;
441         cls_flower.rule->match.mask = &f->mask->key;
442         cls_flower.rule->match.key = &f->mkey;
443         cls_flower.classid = f->res.classid;
444
445         err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts,
446                                    rtnl_held);
447         if (err) {
448                 kfree(cls_flower.rule);
449                 if (skip_sw) {
450                         NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
451                         return err;
452                 }
453                 return 0;
454         }
455
456         err = tc_setup_cb_add(block, tp, TC_SETUP_CLSFLOWER, &cls_flower,
457                               skip_sw, &f->flags, &f->in_hw_count, rtnl_held);
458         tc_cleanup_flow_action(&cls_flower.rule->action);
459         kfree(cls_flower.rule);
460
461         if (err) {
462                 fl_hw_destroy_filter(tp, f, rtnl_held, NULL);
463                 return err;
464         }
465
466         if (skip_sw && !(f->flags & TCA_CLS_FLAGS_IN_HW))
467                 return -EINVAL;
468
469         return 0;
470 }
471
472 static void fl_hw_update_stats(struct tcf_proto *tp, struct cls_fl_filter *f,
473                                bool rtnl_held)
474 {
475         struct tcf_block *block = tp->chain->block;
476         struct flow_cls_offload cls_flower = {};
477
478         tc_cls_common_offload_init(&cls_flower.common, tp, f->flags, NULL);
479         cls_flower.command = FLOW_CLS_STATS;
480         cls_flower.cookie = (unsigned long) f;
481         cls_flower.classid = f->res.classid;
482
483         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false,
484                          rtnl_held);
485
486         tcf_exts_stats_update(&f->exts, cls_flower.stats.bytes,
487                               cls_flower.stats.pkts,
488                               cls_flower.stats.lastused);
489 }
490
491 static void __fl_put(struct cls_fl_filter *f)
492 {
493         if (!refcount_dec_and_test(&f->refcnt))
494                 return;
495
496         if (tcf_exts_get_net(&f->exts))
497                 tcf_queue_work(&f->rwork, fl_destroy_filter_work);
498         else
499                 __fl_destroy_filter(f);
500 }
501
502 static struct cls_fl_filter *__fl_get(struct cls_fl_head *head, u32 handle)
503 {
504         struct cls_fl_filter *f;
505
506         rcu_read_lock();
507         f = idr_find(&head->handle_idr, handle);
508         if (f && !refcount_inc_not_zero(&f->refcnt))
509                 f = NULL;
510         rcu_read_unlock();
511
512         return f;
513 }
514
515 static int __fl_delete(struct tcf_proto *tp, struct cls_fl_filter *f,
516                        bool *last, bool rtnl_held,
517                        struct netlink_ext_ack *extack)
518 {
519         struct cls_fl_head *head = fl_head_dereference(tp);
520
521         *last = false;
522
523         spin_lock(&tp->lock);
524         if (f->deleted) {
525                 spin_unlock(&tp->lock);
526                 return -ENOENT;
527         }
528
529         f->deleted = true;
530         rhashtable_remove_fast(&f->mask->ht, &f->ht_node,
531                                f->mask->filter_ht_params);
532         idr_remove(&head->handle_idr, f->handle);
533         list_del_rcu(&f->list);
534         spin_unlock(&tp->lock);
535
536         *last = fl_mask_put(head, f->mask);
537         if (!tc_skip_hw(f->flags))
538                 fl_hw_destroy_filter(tp, f, rtnl_held, extack);
539         tcf_unbind_filter(tp, &f->res);
540         __fl_put(f);
541
542         return 0;
543 }
544
545 static void fl_destroy_sleepable(struct work_struct *work)
546 {
547         struct cls_fl_head *head = container_of(to_rcu_work(work),
548                                                 struct cls_fl_head,
549                                                 rwork);
550
551         rhashtable_destroy(&head->ht);
552         kfree(head);
553         module_put(THIS_MODULE);
554 }
555
556 static void fl_destroy(struct tcf_proto *tp, bool rtnl_held,
557                        struct netlink_ext_ack *extack)
558 {
559         struct cls_fl_head *head = fl_head_dereference(tp);
560         struct fl_flow_mask *mask, *next_mask;
561         struct cls_fl_filter *f, *next;
562         bool last;
563
564         list_for_each_entry_safe(mask, next_mask, &head->masks, list) {
565                 list_for_each_entry_safe(f, next, &mask->filters, list) {
566                         __fl_delete(tp, f, &last, rtnl_held, extack);
567                         if (last)
568                                 break;
569                 }
570         }
571         idr_destroy(&head->handle_idr);
572
573         __module_get(THIS_MODULE);
574         tcf_queue_work(&head->rwork, fl_destroy_sleepable);
575 }
576
577 static void fl_put(struct tcf_proto *tp, void *arg)
578 {
579         struct cls_fl_filter *f = arg;
580
581         __fl_put(f);
582 }
583
584 static void *fl_get(struct tcf_proto *tp, u32 handle)
585 {
586         struct cls_fl_head *head = fl_head_dereference(tp);
587
588         return __fl_get(head, handle);
589 }
590
591 static const struct nla_policy fl_policy[TCA_FLOWER_MAX + 1] = {
592         [TCA_FLOWER_UNSPEC]             = { .type = NLA_UNSPEC },
593         [TCA_FLOWER_CLASSID]            = { .type = NLA_U32 },
594         [TCA_FLOWER_INDEV]              = { .type = NLA_STRING,
595                                             .len = IFNAMSIZ },
596         [TCA_FLOWER_KEY_ETH_DST]        = { .len = ETH_ALEN },
597         [TCA_FLOWER_KEY_ETH_DST_MASK]   = { .len = ETH_ALEN },
598         [TCA_FLOWER_KEY_ETH_SRC]        = { .len = ETH_ALEN },
599         [TCA_FLOWER_KEY_ETH_SRC_MASK]   = { .len = ETH_ALEN },
600         [TCA_FLOWER_KEY_ETH_TYPE]       = { .type = NLA_U16 },
601         [TCA_FLOWER_KEY_IP_PROTO]       = { .type = NLA_U8 },
602         [TCA_FLOWER_KEY_IPV4_SRC]       = { .type = NLA_U32 },
603         [TCA_FLOWER_KEY_IPV4_SRC_MASK]  = { .type = NLA_U32 },
604         [TCA_FLOWER_KEY_IPV4_DST]       = { .type = NLA_U32 },
605         [TCA_FLOWER_KEY_IPV4_DST_MASK]  = { .type = NLA_U32 },
606         [TCA_FLOWER_KEY_IPV6_SRC]       = { .len = sizeof(struct in6_addr) },
607         [TCA_FLOWER_KEY_IPV6_SRC_MASK]  = { .len = sizeof(struct in6_addr) },
608         [TCA_FLOWER_KEY_IPV6_DST]       = { .len = sizeof(struct in6_addr) },
609         [TCA_FLOWER_KEY_IPV6_DST_MASK]  = { .len = sizeof(struct in6_addr) },
610         [TCA_FLOWER_KEY_TCP_SRC]        = { .type = NLA_U16 },
611         [TCA_FLOWER_KEY_TCP_DST]        = { .type = NLA_U16 },
612         [TCA_FLOWER_KEY_UDP_SRC]        = { .type = NLA_U16 },
613         [TCA_FLOWER_KEY_UDP_DST]        = { .type = NLA_U16 },
614         [TCA_FLOWER_KEY_VLAN_ID]        = { .type = NLA_U16 },
615         [TCA_FLOWER_KEY_VLAN_PRIO]      = { .type = NLA_U8 },
616         [TCA_FLOWER_KEY_VLAN_ETH_TYPE]  = { .type = NLA_U16 },
617         [TCA_FLOWER_KEY_ENC_KEY_ID]     = { .type = NLA_U32 },
618         [TCA_FLOWER_KEY_ENC_IPV4_SRC]   = { .type = NLA_U32 },
619         [TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK] = { .type = NLA_U32 },
620         [TCA_FLOWER_KEY_ENC_IPV4_DST]   = { .type = NLA_U32 },
621         [TCA_FLOWER_KEY_ENC_IPV4_DST_MASK] = { .type = NLA_U32 },
622         [TCA_FLOWER_KEY_ENC_IPV6_SRC]   = { .len = sizeof(struct in6_addr) },
623         [TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK] = { .len = sizeof(struct in6_addr) },
624         [TCA_FLOWER_KEY_ENC_IPV6_DST]   = { .len = sizeof(struct in6_addr) },
625         [TCA_FLOWER_KEY_ENC_IPV6_DST_MASK] = { .len = sizeof(struct in6_addr) },
626         [TCA_FLOWER_KEY_TCP_SRC_MASK]   = { .type = NLA_U16 },
627         [TCA_FLOWER_KEY_TCP_DST_MASK]   = { .type = NLA_U16 },
628         [TCA_FLOWER_KEY_UDP_SRC_MASK]   = { .type = NLA_U16 },
629         [TCA_FLOWER_KEY_UDP_DST_MASK]   = { .type = NLA_U16 },
630         [TCA_FLOWER_KEY_SCTP_SRC_MASK]  = { .type = NLA_U16 },
631         [TCA_FLOWER_KEY_SCTP_DST_MASK]  = { .type = NLA_U16 },
632         [TCA_FLOWER_KEY_SCTP_SRC]       = { .type = NLA_U16 },
633         [TCA_FLOWER_KEY_SCTP_DST]       = { .type = NLA_U16 },
634         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT]       = { .type = NLA_U16 },
635         [TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK]  = { .type = NLA_U16 },
636         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT]       = { .type = NLA_U16 },
637         [TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK]  = { .type = NLA_U16 },
638         [TCA_FLOWER_KEY_FLAGS]          = { .type = NLA_U32 },
639         [TCA_FLOWER_KEY_FLAGS_MASK]     = { .type = NLA_U32 },
640         [TCA_FLOWER_KEY_ICMPV4_TYPE]    = { .type = NLA_U8 },
641         [TCA_FLOWER_KEY_ICMPV4_TYPE_MASK] = { .type = NLA_U8 },
642         [TCA_FLOWER_KEY_ICMPV4_CODE]    = { .type = NLA_U8 },
643         [TCA_FLOWER_KEY_ICMPV4_CODE_MASK] = { .type = NLA_U8 },
644         [TCA_FLOWER_KEY_ICMPV6_TYPE]    = { .type = NLA_U8 },
645         [TCA_FLOWER_KEY_ICMPV6_TYPE_MASK] = { .type = NLA_U8 },
646         [TCA_FLOWER_KEY_ICMPV6_CODE]    = { .type = NLA_U8 },
647         [TCA_FLOWER_KEY_ICMPV6_CODE_MASK] = { .type = NLA_U8 },
648         [TCA_FLOWER_KEY_ARP_SIP]        = { .type = NLA_U32 },
649         [TCA_FLOWER_KEY_ARP_SIP_MASK]   = { .type = NLA_U32 },
650         [TCA_FLOWER_KEY_ARP_TIP]        = { .type = NLA_U32 },
651         [TCA_FLOWER_KEY_ARP_TIP_MASK]   = { .type = NLA_U32 },
652         [TCA_FLOWER_KEY_ARP_OP]         = { .type = NLA_U8 },
653         [TCA_FLOWER_KEY_ARP_OP_MASK]    = { .type = NLA_U8 },
654         [TCA_FLOWER_KEY_ARP_SHA]        = { .len = ETH_ALEN },
655         [TCA_FLOWER_KEY_ARP_SHA_MASK]   = { .len = ETH_ALEN },
656         [TCA_FLOWER_KEY_ARP_THA]        = { .len = ETH_ALEN },
657         [TCA_FLOWER_KEY_ARP_THA_MASK]   = { .len = ETH_ALEN },
658         [TCA_FLOWER_KEY_MPLS_TTL]       = { .type = NLA_U8 },
659         [TCA_FLOWER_KEY_MPLS_BOS]       = { .type = NLA_U8 },
660         [TCA_FLOWER_KEY_MPLS_TC]        = { .type = NLA_U8 },
661         [TCA_FLOWER_KEY_MPLS_LABEL]     = { .type = NLA_U32 },
662         [TCA_FLOWER_KEY_TCP_FLAGS]      = { .type = NLA_U16 },
663         [TCA_FLOWER_KEY_TCP_FLAGS_MASK] = { .type = NLA_U16 },
664         [TCA_FLOWER_KEY_IP_TOS]         = { .type = NLA_U8 },
665         [TCA_FLOWER_KEY_IP_TOS_MASK]    = { .type = NLA_U8 },
666         [TCA_FLOWER_KEY_IP_TTL]         = { .type = NLA_U8 },
667         [TCA_FLOWER_KEY_IP_TTL_MASK]    = { .type = NLA_U8 },
668         [TCA_FLOWER_KEY_CVLAN_ID]       = { .type = NLA_U16 },
669         [TCA_FLOWER_KEY_CVLAN_PRIO]     = { .type = NLA_U8 },
670         [TCA_FLOWER_KEY_CVLAN_ETH_TYPE] = { .type = NLA_U16 },
671         [TCA_FLOWER_KEY_ENC_IP_TOS]     = { .type = NLA_U8 },
672         [TCA_FLOWER_KEY_ENC_IP_TOS_MASK] = { .type = NLA_U8 },
673         [TCA_FLOWER_KEY_ENC_IP_TTL]      = { .type = NLA_U8 },
674         [TCA_FLOWER_KEY_ENC_IP_TTL_MASK] = { .type = NLA_U8 },
675         [TCA_FLOWER_KEY_ENC_OPTS]       = { .type = NLA_NESTED },
676         [TCA_FLOWER_KEY_ENC_OPTS_MASK]  = { .type = NLA_NESTED },
677         [TCA_FLOWER_KEY_CT_STATE]       = { .type = NLA_U16 },
678         [TCA_FLOWER_KEY_CT_STATE_MASK]  = { .type = NLA_U16 },
679         [TCA_FLOWER_KEY_CT_ZONE]        = { .type = NLA_U16 },
680         [TCA_FLOWER_KEY_CT_ZONE_MASK]   = { .type = NLA_U16 },
681         [TCA_FLOWER_KEY_CT_MARK]        = { .type = NLA_U32 },
682         [TCA_FLOWER_KEY_CT_MARK_MASK]   = { .type = NLA_U32 },
683         [TCA_FLOWER_KEY_CT_LABELS]      = { .type = NLA_BINARY,
684                                             .len = 128 / BITS_PER_BYTE },
685         [TCA_FLOWER_KEY_CT_LABELS_MASK] = { .type = NLA_BINARY,
686                                             .len = 128 / BITS_PER_BYTE },
687 };
688
689 static const struct nla_policy
690 enc_opts_policy[TCA_FLOWER_KEY_ENC_OPTS_MAX + 1] = {
691         [TCA_FLOWER_KEY_ENC_OPTS_GENEVE]        = { .type = NLA_NESTED },
692 };
693
694 static const struct nla_policy
695 geneve_opt_policy[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1] = {
696         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]      = { .type = NLA_U16 },
697         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]       = { .type = NLA_U8 },
698         [TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]       = { .type = NLA_BINARY,
699                                                        .len = 128 },
700 };
701
702 static void fl_set_key_val(struct nlattr **tb,
703                            void *val, int val_type,
704                            void *mask, int mask_type, int len)
705 {
706         if (!tb[val_type])
707                 return;
708         nla_memcpy(val, tb[val_type], len);
709         if (mask_type == TCA_FLOWER_UNSPEC || !tb[mask_type])
710                 memset(mask, 0xff, len);
711         else
712                 nla_memcpy(mask, tb[mask_type], len);
713 }
714
715 static int fl_set_key_port_range(struct nlattr **tb, struct fl_flow_key *key,
716                                  struct fl_flow_key *mask)
717 {
718         fl_set_key_val(tb, &key->tp_min.dst,
719                        TCA_FLOWER_KEY_PORT_DST_MIN, &mask->tp_min.dst,
720                        TCA_FLOWER_UNSPEC, sizeof(key->tp_min.dst));
721         fl_set_key_val(tb, &key->tp_max.dst,
722                        TCA_FLOWER_KEY_PORT_DST_MAX, &mask->tp_max.dst,
723                        TCA_FLOWER_UNSPEC, sizeof(key->tp_max.dst));
724         fl_set_key_val(tb, &key->tp_min.src,
725                        TCA_FLOWER_KEY_PORT_SRC_MIN, &mask->tp_min.src,
726                        TCA_FLOWER_UNSPEC, sizeof(key->tp_min.src));
727         fl_set_key_val(tb, &key->tp_max.src,
728                        TCA_FLOWER_KEY_PORT_SRC_MAX, &mask->tp_max.src,
729                        TCA_FLOWER_UNSPEC, sizeof(key->tp_max.src));
730
731         if ((mask->tp_min.dst && mask->tp_max.dst &&
732              htons(key->tp_max.dst) <= htons(key->tp_min.dst)) ||
733              (mask->tp_min.src && mask->tp_max.src &&
734               htons(key->tp_max.src) <= htons(key->tp_min.src)))
735                 return -EINVAL;
736
737         return 0;
738 }
739
740 static int fl_set_key_mpls(struct nlattr **tb,
741                            struct flow_dissector_key_mpls *key_val,
742                            struct flow_dissector_key_mpls *key_mask)
743 {
744         if (tb[TCA_FLOWER_KEY_MPLS_TTL]) {
745                 key_val->mpls_ttl = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TTL]);
746                 key_mask->mpls_ttl = MPLS_TTL_MASK;
747         }
748         if (tb[TCA_FLOWER_KEY_MPLS_BOS]) {
749                 u8 bos = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_BOS]);
750
751                 if (bos & ~MPLS_BOS_MASK)
752                         return -EINVAL;
753                 key_val->mpls_bos = bos;
754                 key_mask->mpls_bos = MPLS_BOS_MASK;
755         }
756         if (tb[TCA_FLOWER_KEY_MPLS_TC]) {
757                 u8 tc = nla_get_u8(tb[TCA_FLOWER_KEY_MPLS_TC]);
758
759                 if (tc & ~MPLS_TC_MASK)
760                         return -EINVAL;
761                 key_val->mpls_tc = tc;
762                 key_mask->mpls_tc = MPLS_TC_MASK;
763         }
764         if (tb[TCA_FLOWER_KEY_MPLS_LABEL]) {
765                 u32 label = nla_get_u32(tb[TCA_FLOWER_KEY_MPLS_LABEL]);
766
767                 if (label & ~MPLS_LABEL_MASK)
768                         return -EINVAL;
769                 key_val->mpls_label = label;
770                 key_mask->mpls_label = MPLS_LABEL_MASK;
771         }
772         return 0;
773 }
774
775 static void fl_set_key_vlan(struct nlattr **tb,
776                             __be16 ethertype,
777                             int vlan_id_key, int vlan_prio_key,
778                             struct flow_dissector_key_vlan *key_val,
779                             struct flow_dissector_key_vlan *key_mask)
780 {
781 #define VLAN_PRIORITY_MASK      0x7
782
783         if (tb[vlan_id_key]) {
784                 key_val->vlan_id =
785                         nla_get_u16(tb[vlan_id_key]) & VLAN_VID_MASK;
786                 key_mask->vlan_id = VLAN_VID_MASK;
787         }
788         if (tb[vlan_prio_key]) {
789                 key_val->vlan_priority =
790                         nla_get_u8(tb[vlan_prio_key]) &
791                         VLAN_PRIORITY_MASK;
792                 key_mask->vlan_priority = VLAN_PRIORITY_MASK;
793         }
794         key_val->vlan_tpid = ethertype;
795         key_mask->vlan_tpid = cpu_to_be16(~0);
796 }
797
798 static void fl_set_key_flag(u32 flower_key, u32 flower_mask,
799                             u32 *dissector_key, u32 *dissector_mask,
800                             u32 flower_flag_bit, u32 dissector_flag_bit)
801 {
802         if (flower_mask & flower_flag_bit) {
803                 *dissector_mask |= dissector_flag_bit;
804                 if (flower_key & flower_flag_bit)
805                         *dissector_key |= dissector_flag_bit;
806         }
807 }
808
809 static int fl_set_key_flags(struct nlattr **tb,
810                             u32 *flags_key, u32 *flags_mask)
811 {
812         u32 key, mask;
813
814         /* mask is mandatory for flags */
815         if (!tb[TCA_FLOWER_KEY_FLAGS_MASK])
816                 return -EINVAL;
817
818         key = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS]));
819         mask = be32_to_cpu(nla_get_u32(tb[TCA_FLOWER_KEY_FLAGS_MASK]));
820
821         *flags_key  = 0;
822         *flags_mask = 0;
823
824         fl_set_key_flag(key, mask, flags_key, flags_mask,
825                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
826         fl_set_key_flag(key, mask, flags_key, flags_mask,
827                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
828                         FLOW_DIS_FIRST_FRAG);
829
830         return 0;
831 }
832
833 static void fl_set_key_ip(struct nlattr **tb, bool encap,
834                           struct flow_dissector_key_ip *key,
835                           struct flow_dissector_key_ip *mask)
836 {
837         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
838         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
839         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
840         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
841
842         fl_set_key_val(tb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos));
843         fl_set_key_val(tb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl));
844 }
845
846 static int fl_set_geneve_opt(const struct nlattr *nla, struct fl_flow_key *key,
847                              int depth, int option_len,
848                              struct netlink_ext_ack *extack)
849 {
850         struct nlattr *tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX + 1];
851         struct nlattr *class = NULL, *type = NULL, *data = NULL;
852         struct geneve_opt *opt;
853         int err, data_len = 0;
854
855         if (option_len > sizeof(struct geneve_opt))
856                 data_len = option_len - sizeof(struct geneve_opt);
857
858         opt = (struct geneve_opt *)&key->enc_opts.data[key->enc_opts.len];
859         memset(opt, 0xff, option_len);
860         opt->length = data_len / 4;
861         opt->r1 = 0;
862         opt->r2 = 0;
863         opt->r3 = 0;
864
865         /* If no mask has been prodived we assume an exact match. */
866         if (!depth)
867                 return sizeof(struct geneve_opt) + data_len;
868
869         if (nla_type(nla) != TCA_FLOWER_KEY_ENC_OPTS_GENEVE) {
870                 NL_SET_ERR_MSG(extack, "Non-geneve option type for mask");
871                 return -EINVAL;
872         }
873
874         err = nla_parse_nested_deprecated(tb,
875                                           TCA_FLOWER_KEY_ENC_OPT_GENEVE_MAX,
876                                           nla, geneve_opt_policy, extack);
877         if (err < 0)
878                 return err;
879
880         /* We are not allowed to omit any of CLASS, TYPE or DATA
881          * fields from the key.
882          */
883         if (!option_len &&
884             (!tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS] ||
885              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE] ||
886              !tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA])) {
887                 NL_SET_ERR_MSG(extack, "Missing tunnel key geneve option class, type or data");
888                 return -EINVAL;
889         }
890
891         /* Omitting any of CLASS, TYPE or DATA fields is allowed
892          * for the mask.
893          */
894         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA]) {
895                 int new_len = key->enc_opts.len;
896
897                 data = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA];
898                 data_len = nla_len(data);
899                 if (data_len < 4) {
900                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is less than 4 bytes long");
901                         return -ERANGE;
902                 }
903                 if (data_len % 4) {
904                         NL_SET_ERR_MSG(extack, "Tunnel key geneve option data is not a multiple of 4 bytes long");
905                         return -ERANGE;
906                 }
907
908                 new_len += sizeof(struct geneve_opt) + data_len;
909                 BUILD_BUG_ON(FLOW_DIS_TUN_OPTS_MAX != IP_TUNNEL_OPTS_MAX);
910                 if (new_len > FLOW_DIS_TUN_OPTS_MAX) {
911                         NL_SET_ERR_MSG(extack, "Tunnel options exceeds max size");
912                         return -ERANGE;
913                 }
914                 opt->length = data_len / 4;
915                 memcpy(opt->opt_data, nla_data(data), data_len);
916         }
917
918         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS]) {
919                 class = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS];
920                 opt->opt_class = nla_get_be16(class);
921         }
922
923         if (tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE]) {
924                 type = tb[TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE];
925                 opt->type = nla_get_u8(type);
926         }
927
928         return sizeof(struct geneve_opt) + data_len;
929 }
930
931 static int fl_set_enc_opt(struct nlattr **tb, struct fl_flow_key *key,
932                           struct fl_flow_key *mask,
933                           struct netlink_ext_ack *extack)
934 {
935         const struct nlattr *nla_enc_key, *nla_opt_key, *nla_opt_msk = NULL;
936         int err, option_len, key_depth, msk_depth = 0;
937
938         err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS],
939                                              TCA_FLOWER_KEY_ENC_OPTS_MAX,
940                                              enc_opts_policy, extack);
941         if (err)
942                 return err;
943
944         nla_enc_key = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS]);
945
946         if (tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]) {
947                 err = nla_validate_nested_deprecated(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK],
948                                                      TCA_FLOWER_KEY_ENC_OPTS_MAX,
949                                                      enc_opts_policy, extack);
950                 if (err)
951                         return err;
952
953                 nla_opt_msk = nla_data(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
954                 msk_depth = nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS_MASK]);
955         }
956
957         nla_for_each_attr(nla_opt_key, nla_enc_key,
958                           nla_len(tb[TCA_FLOWER_KEY_ENC_OPTS]), key_depth) {
959                 switch (nla_type(nla_opt_key)) {
960                 case TCA_FLOWER_KEY_ENC_OPTS_GENEVE:
961                         option_len = 0;
962                         key->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
963                         option_len = fl_set_geneve_opt(nla_opt_key, key,
964                                                        key_depth, option_len,
965                                                        extack);
966                         if (option_len < 0)
967                                 return option_len;
968
969                         key->enc_opts.len += option_len;
970                         /* At the same time we need to parse through the mask
971                          * in order to verify exact and mask attribute lengths.
972                          */
973                         mask->enc_opts.dst_opt_type = TUNNEL_GENEVE_OPT;
974                         option_len = fl_set_geneve_opt(nla_opt_msk, mask,
975                                                        msk_depth, option_len,
976                                                        extack);
977                         if (option_len < 0)
978                                 return option_len;
979
980                         mask->enc_opts.len += option_len;
981                         if (key->enc_opts.len != mask->enc_opts.len) {
982                                 NL_SET_ERR_MSG(extack, "Key and mask miss aligned");
983                                 return -EINVAL;
984                         }
985
986                         if (msk_depth)
987                                 nla_opt_msk = nla_next(nla_opt_msk, &msk_depth);
988                         break;
989                 default:
990                         NL_SET_ERR_MSG(extack, "Unknown tunnel option type");
991                         return -EINVAL;
992                 }
993         }
994
995         return 0;
996 }
997
998 static int fl_set_key_ct(struct nlattr **tb,
999                          struct flow_dissector_key_ct *key,
1000                          struct flow_dissector_key_ct *mask,
1001                          struct netlink_ext_ack *extack)
1002 {
1003         if (tb[TCA_FLOWER_KEY_CT_STATE]) {
1004                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK)) {
1005                         NL_SET_ERR_MSG(extack, "Conntrack isn't enabled");
1006                         return -EOPNOTSUPP;
1007                 }
1008                 fl_set_key_val(tb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
1009                                &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
1010                                sizeof(key->ct_state));
1011         }
1012         if (tb[TCA_FLOWER_KEY_CT_ZONE]) {
1013                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES)) {
1014                         NL_SET_ERR_MSG(extack, "Conntrack zones isn't enabled");
1015                         return -EOPNOTSUPP;
1016                 }
1017                 fl_set_key_val(tb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
1018                                &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
1019                                sizeof(key->ct_zone));
1020         }
1021         if (tb[TCA_FLOWER_KEY_CT_MARK]) {
1022                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_MARK)) {
1023                         NL_SET_ERR_MSG(extack, "Conntrack mark isn't enabled");
1024                         return -EOPNOTSUPP;
1025                 }
1026                 fl_set_key_val(tb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
1027                                &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
1028                                sizeof(key->ct_mark));
1029         }
1030         if (tb[TCA_FLOWER_KEY_CT_LABELS]) {
1031                 if (!IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS)) {
1032                         NL_SET_ERR_MSG(extack, "Conntrack labels aren't enabled");
1033                         return -EOPNOTSUPP;
1034                 }
1035                 fl_set_key_val(tb, key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
1036                                mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
1037                                sizeof(key->ct_labels));
1038         }
1039
1040         return 0;
1041 }
1042
1043 static int fl_set_key(struct net *net, struct nlattr **tb,
1044                       struct fl_flow_key *key, struct fl_flow_key *mask,
1045                       struct netlink_ext_ack *extack)
1046 {
1047         __be16 ethertype;
1048         int ret = 0;
1049
1050         if (tb[TCA_FLOWER_INDEV]) {
1051                 int err = tcf_change_indev(net, tb[TCA_FLOWER_INDEV], extack);
1052                 if (err < 0)
1053                         return err;
1054                 key->meta.ingress_ifindex = err;
1055                 mask->meta.ingress_ifindex = 0xffffffff;
1056         }
1057
1058         fl_set_key_val(tb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
1059                        mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
1060                        sizeof(key->eth.dst));
1061         fl_set_key_val(tb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
1062                        mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
1063                        sizeof(key->eth.src));
1064
1065         if (tb[TCA_FLOWER_KEY_ETH_TYPE]) {
1066                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_ETH_TYPE]);
1067
1068                 if (eth_type_vlan(ethertype)) {
1069                         fl_set_key_vlan(tb, ethertype, TCA_FLOWER_KEY_VLAN_ID,
1070                                         TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan,
1071                                         &mask->vlan);
1072
1073                         if (tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]) {
1074                                 ethertype = nla_get_be16(tb[TCA_FLOWER_KEY_VLAN_ETH_TYPE]);
1075                                 if (eth_type_vlan(ethertype)) {
1076                                         fl_set_key_vlan(tb, ethertype,
1077                                                         TCA_FLOWER_KEY_CVLAN_ID,
1078                                                         TCA_FLOWER_KEY_CVLAN_PRIO,
1079                                                         &key->cvlan, &mask->cvlan);
1080                                         fl_set_key_val(tb, &key->basic.n_proto,
1081                                                        TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
1082                                                        &mask->basic.n_proto,
1083                                                        TCA_FLOWER_UNSPEC,
1084                                                        sizeof(key->basic.n_proto));
1085                                 } else {
1086                                         key->basic.n_proto = ethertype;
1087                                         mask->basic.n_proto = cpu_to_be16(~0);
1088                                 }
1089                         }
1090                 } else {
1091                         key->basic.n_proto = ethertype;
1092                         mask->basic.n_proto = cpu_to_be16(~0);
1093                 }
1094         }
1095
1096         if (key->basic.n_proto == htons(ETH_P_IP) ||
1097             key->basic.n_proto == htons(ETH_P_IPV6)) {
1098                 fl_set_key_val(tb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
1099                                &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
1100                                sizeof(key->basic.ip_proto));
1101                 fl_set_key_ip(tb, false, &key->ip, &mask->ip);
1102         }
1103
1104         if (tb[TCA_FLOWER_KEY_IPV4_SRC] || tb[TCA_FLOWER_KEY_IPV4_DST]) {
1105                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1106                 mask->control.addr_type = ~0;
1107                 fl_set_key_val(tb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
1108                                &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
1109                                sizeof(key->ipv4.src));
1110                 fl_set_key_val(tb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
1111                                &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
1112                                sizeof(key->ipv4.dst));
1113         } else if (tb[TCA_FLOWER_KEY_IPV6_SRC] || tb[TCA_FLOWER_KEY_IPV6_DST]) {
1114                 key->control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1115                 mask->control.addr_type = ~0;
1116                 fl_set_key_val(tb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
1117                                &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
1118                                sizeof(key->ipv6.src));
1119                 fl_set_key_val(tb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
1120                                &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
1121                                sizeof(key->ipv6.dst));
1122         }
1123
1124         if (key->basic.ip_proto == IPPROTO_TCP) {
1125                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
1126                                &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
1127                                sizeof(key->tp.src));
1128                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
1129                                &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
1130                                sizeof(key->tp.dst));
1131                 fl_set_key_val(tb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
1132                                &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
1133                                sizeof(key->tcp.flags));
1134         } else if (key->basic.ip_proto == IPPROTO_UDP) {
1135                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
1136                                &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
1137                                sizeof(key->tp.src));
1138                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
1139                                &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
1140                                sizeof(key->tp.dst));
1141         } else if (key->basic.ip_proto == IPPROTO_SCTP) {
1142                 fl_set_key_val(tb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
1143                                &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
1144                                sizeof(key->tp.src));
1145                 fl_set_key_val(tb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
1146                                &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
1147                                sizeof(key->tp.dst));
1148         } else if (key->basic.n_proto == htons(ETH_P_IP) &&
1149                    key->basic.ip_proto == IPPROTO_ICMP) {
1150                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV4_TYPE,
1151                                &mask->icmp.type,
1152                                TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
1153                                sizeof(key->icmp.type));
1154                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV4_CODE,
1155                                &mask->icmp.code,
1156                                TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
1157                                sizeof(key->icmp.code));
1158         } else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
1159                    key->basic.ip_proto == IPPROTO_ICMPV6) {
1160                 fl_set_key_val(tb, &key->icmp.type, TCA_FLOWER_KEY_ICMPV6_TYPE,
1161                                &mask->icmp.type,
1162                                TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
1163                                sizeof(key->icmp.type));
1164                 fl_set_key_val(tb, &key->icmp.code, TCA_FLOWER_KEY_ICMPV6_CODE,
1165                                &mask->icmp.code,
1166                                TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
1167                                sizeof(key->icmp.code));
1168         } else if (key->basic.n_proto == htons(ETH_P_MPLS_UC) ||
1169                    key->basic.n_proto == htons(ETH_P_MPLS_MC)) {
1170                 ret = fl_set_key_mpls(tb, &key->mpls, &mask->mpls);
1171                 if (ret)
1172                         return ret;
1173         } else if (key->basic.n_proto == htons(ETH_P_ARP) ||
1174                    key->basic.n_proto == htons(ETH_P_RARP)) {
1175                 fl_set_key_val(tb, &key->arp.sip, TCA_FLOWER_KEY_ARP_SIP,
1176                                &mask->arp.sip, TCA_FLOWER_KEY_ARP_SIP_MASK,
1177                                sizeof(key->arp.sip));
1178                 fl_set_key_val(tb, &key->arp.tip, TCA_FLOWER_KEY_ARP_TIP,
1179                                &mask->arp.tip, TCA_FLOWER_KEY_ARP_TIP_MASK,
1180                                sizeof(key->arp.tip));
1181                 fl_set_key_val(tb, &key->arp.op, TCA_FLOWER_KEY_ARP_OP,
1182                                &mask->arp.op, TCA_FLOWER_KEY_ARP_OP_MASK,
1183                                sizeof(key->arp.op));
1184                 fl_set_key_val(tb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
1185                                mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
1186                                sizeof(key->arp.sha));
1187                 fl_set_key_val(tb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
1188                                mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
1189                                sizeof(key->arp.tha));
1190         }
1191
1192         if (key->basic.ip_proto == IPPROTO_TCP ||
1193             key->basic.ip_proto == IPPROTO_UDP ||
1194             key->basic.ip_proto == IPPROTO_SCTP) {
1195                 ret = fl_set_key_port_range(tb, key, mask);
1196                 if (ret)
1197                         return ret;
1198         }
1199
1200         if (tb[TCA_FLOWER_KEY_ENC_IPV4_SRC] ||
1201             tb[TCA_FLOWER_KEY_ENC_IPV4_DST]) {
1202                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV4_ADDRS;
1203                 mask->enc_control.addr_type = ~0;
1204                 fl_set_key_val(tb, &key->enc_ipv4.src,
1205                                TCA_FLOWER_KEY_ENC_IPV4_SRC,
1206                                &mask->enc_ipv4.src,
1207                                TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
1208                                sizeof(key->enc_ipv4.src));
1209                 fl_set_key_val(tb, &key->enc_ipv4.dst,
1210                                TCA_FLOWER_KEY_ENC_IPV4_DST,
1211                                &mask->enc_ipv4.dst,
1212                                TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
1213                                sizeof(key->enc_ipv4.dst));
1214         }
1215
1216         if (tb[TCA_FLOWER_KEY_ENC_IPV6_SRC] ||
1217             tb[TCA_FLOWER_KEY_ENC_IPV6_DST]) {
1218                 key->enc_control.addr_type = FLOW_DISSECTOR_KEY_IPV6_ADDRS;
1219                 mask->enc_control.addr_type = ~0;
1220                 fl_set_key_val(tb, &key->enc_ipv6.src,
1221                                TCA_FLOWER_KEY_ENC_IPV6_SRC,
1222                                &mask->enc_ipv6.src,
1223                                TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
1224                                sizeof(key->enc_ipv6.src));
1225                 fl_set_key_val(tb, &key->enc_ipv6.dst,
1226                                TCA_FLOWER_KEY_ENC_IPV6_DST,
1227                                &mask->enc_ipv6.dst,
1228                                TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
1229                                sizeof(key->enc_ipv6.dst));
1230         }
1231
1232         fl_set_key_val(tb, &key->enc_key_id.keyid, TCA_FLOWER_KEY_ENC_KEY_ID,
1233                        &mask->enc_key_id.keyid, TCA_FLOWER_UNSPEC,
1234                        sizeof(key->enc_key_id.keyid));
1235
1236         fl_set_key_val(tb, &key->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
1237                        &mask->enc_tp.src, TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
1238                        sizeof(key->enc_tp.src));
1239
1240         fl_set_key_val(tb, &key->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
1241                        &mask->enc_tp.dst, TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
1242                        sizeof(key->enc_tp.dst));
1243
1244         fl_set_key_ip(tb, true, &key->enc_ip, &mask->enc_ip);
1245
1246         if (tb[TCA_FLOWER_KEY_ENC_OPTS]) {
1247                 ret = fl_set_enc_opt(tb, key, mask, extack);
1248                 if (ret)
1249                         return ret;
1250         }
1251
1252         ret = fl_set_key_ct(tb, &key->ct, &mask->ct, extack);
1253         if (ret)
1254                 return ret;
1255
1256         if (tb[TCA_FLOWER_KEY_FLAGS])
1257                 ret = fl_set_key_flags(tb, &key->control.flags, &mask->control.flags);
1258
1259         return ret;
1260 }
1261
1262 static void fl_mask_copy(struct fl_flow_mask *dst,
1263                          struct fl_flow_mask *src)
1264 {
1265         const void *psrc = fl_key_get_start(&src->key, src);
1266         void *pdst = fl_key_get_start(&dst->key, src);
1267
1268         memcpy(pdst, psrc, fl_mask_range(src));
1269         dst->range = src->range;
1270 }
1271
1272 static const struct rhashtable_params fl_ht_params = {
1273         .key_offset = offsetof(struct cls_fl_filter, mkey), /* base offset */
1274         .head_offset = offsetof(struct cls_fl_filter, ht_node),
1275         .automatic_shrinking = true,
1276 };
1277
1278 static int fl_init_mask_hashtable(struct fl_flow_mask *mask)
1279 {
1280         mask->filter_ht_params = fl_ht_params;
1281         mask->filter_ht_params.key_len = fl_mask_range(mask);
1282         mask->filter_ht_params.key_offset += mask->range.start;
1283
1284         return rhashtable_init(&mask->ht, &mask->filter_ht_params);
1285 }
1286
1287 #define FL_KEY_MEMBER_OFFSET(member) offsetof(struct fl_flow_key, member)
1288 #define FL_KEY_MEMBER_SIZE(member) FIELD_SIZEOF(struct fl_flow_key, member)
1289
1290 #define FL_KEY_IS_MASKED(mask, member)                                          \
1291         memchr_inv(((char *)mask) + FL_KEY_MEMBER_OFFSET(member),               \
1292                    0, FL_KEY_MEMBER_SIZE(member))                               \
1293
1294 #define FL_KEY_SET(keys, cnt, id, member)                                       \
1295         do {                                                                    \
1296                 keys[cnt].key_id = id;                                          \
1297                 keys[cnt].offset = FL_KEY_MEMBER_OFFSET(member);                \
1298                 cnt++;                                                          \
1299         } while(0);
1300
1301 #define FL_KEY_SET_IF_MASKED(mask, keys, cnt, id, member)                       \
1302         do {                                                                    \
1303                 if (FL_KEY_IS_MASKED(mask, member))                             \
1304                         FL_KEY_SET(keys, cnt, id, member);                      \
1305         } while(0);
1306
1307 static void fl_init_dissector(struct flow_dissector *dissector,
1308                               struct fl_flow_key *mask)
1309 {
1310         struct flow_dissector_key keys[FLOW_DISSECTOR_KEY_MAX];
1311         size_t cnt = 0;
1312
1313         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1314                              FLOW_DISSECTOR_KEY_META, meta);
1315         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_CONTROL, control);
1316         FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_BASIC, basic);
1317         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1318                              FLOW_DISSECTOR_KEY_ETH_ADDRS, eth);
1319         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1320                              FLOW_DISSECTOR_KEY_IPV4_ADDRS, ipv4);
1321         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1322                              FLOW_DISSECTOR_KEY_IPV6_ADDRS, ipv6);
1323         if (FL_KEY_IS_MASKED(mask, tp) ||
1324             FL_KEY_IS_MASKED(mask, tp_min) || FL_KEY_IS_MASKED(mask, tp_max))
1325                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_PORTS, tp);
1326         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1327                              FLOW_DISSECTOR_KEY_IP, ip);
1328         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1329                              FLOW_DISSECTOR_KEY_TCP, tcp);
1330         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1331                              FLOW_DISSECTOR_KEY_ICMP, icmp);
1332         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1333                              FLOW_DISSECTOR_KEY_ARP, arp);
1334         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1335                              FLOW_DISSECTOR_KEY_MPLS, mpls);
1336         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1337                              FLOW_DISSECTOR_KEY_VLAN, vlan);
1338         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1339                              FLOW_DISSECTOR_KEY_CVLAN, cvlan);
1340         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1341                              FLOW_DISSECTOR_KEY_ENC_KEYID, enc_key_id);
1342         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1343                              FLOW_DISSECTOR_KEY_ENC_IPV4_ADDRS, enc_ipv4);
1344         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1345                              FLOW_DISSECTOR_KEY_ENC_IPV6_ADDRS, enc_ipv6);
1346         if (FL_KEY_IS_MASKED(mask, enc_ipv4) ||
1347             FL_KEY_IS_MASKED(mask, enc_ipv6))
1348                 FL_KEY_SET(keys, cnt, FLOW_DISSECTOR_KEY_ENC_CONTROL,
1349                            enc_control);
1350         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1351                              FLOW_DISSECTOR_KEY_ENC_PORTS, enc_tp);
1352         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1353                              FLOW_DISSECTOR_KEY_ENC_IP, enc_ip);
1354         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1355                              FLOW_DISSECTOR_KEY_ENC_OPTS, enc_opts);
1356         FL_KEY_SET_IF_MASKED(mask, keys, cnt,
1357                              FLOW_DISSECTOR_KEY_CT, ct);
1358
1359         skb_flow_dissector_init(dissector, keys, cnt);
1360 }
1361
1362 static struct fl_flow_mask *fl_create_new_mask(struct cls_fl_head *head,
1363                                                struct fl_flow_mask *mask)
1364 {
1365         struct fl_flow_mask *newmask;
1366         int err;
1367
1368         newmask = kzalloc(sizeof(*newmask), GFP_KERNEL);
1369         if (!newmask)
1370                 return ERR_PTR(-ENOMEM);
1371
1372         fl_mask_copy(newmask, mask);
1373
1374         if ((newmask->key.tp_min.dst && newmask->key.tp_max.dst) ||
1375             (newmask->key.tp_min.src && newmask->key.tp_max.src))
1376                 newmask->flags |= TCA_FLOWER_MASK_FLAGS_RANGE;
1377
1378         err = fl_init_mask_hashtable(newmask);
1379         if (err)
1380                 goto errout_free;
1381
1382         fl_init_dissector(&newmask->dissector, &newmask->key);
1383
1384         INIT_LIST_HEAD_RCU(&newmask->filters);
1385
1386         refcount_set(&newmask->refcnt, 1);
1387         err = rhashtable_replace_fast(&head->ht, &mask->ht_node,
1388                                       &newmask->ht_node, mask_ht_params);
1389         if (err)
1390                 goto errout_destroy;
1391
1392         spin_lock(&head->masks_lock);
1393         list_add_tail_rcu(&newmask->list, &head->masks);
1394         spin_unlock(&head->masks_lock);
1395
1396         return newmask;
1397
1398 errout_destroy:
1399         rhashtable_destroy(&newmask->ht);
1400 errout_free:
1401         kfree(newmask);
1402
1403         return ERR_PTR(err);
1404 }
1405
1406 static int fl_check_assign_mask(struct cls_fl_head *head,
1407                                 struct cls_fl_filter *fnew,
1408                                 struct cls_fl_filter *fold,
1409                                 struct fl_flow_mask *mask)
1410 {
1411         struct fl_flow_mask *newmask;
1412         int ret = 0;
1413
1414         rcu_read_lock();
1415
1416         /* Insert mask as temporary node to prevent concurrent creation of mask
1417          * with same key. Any concurrent lookups with same key will return
1418          * -EAGAIN because mask's refcnt is zero.
1419          */
1420         fnew->mask = rhashtable_lookup_get_insert_fast(&head->ht,
1421                                                        &mask->ht_node,
1422                                                        mask_ht_params);
1423         if (!fnew->mask) {
1424                 rcu_read_unlock();
1425
1426                 if (fold) {
1427                         ret = -EINVAL;
1428                         goto errout_cleanup;
1429                 }
1430
1431                 newmask = fl_create_new_mask(head, mask);
1432                 if (IS_ERR(newmask)) {
1433                         ret = PTR_ERR(newmask);
1434                         goto errout_cleanup;
1435                 }
1436
1437                 fnew->mask = newmask;
1438                 return 0;
1439         } else if (IS_ERR(fnew->mask)) {
1440                 ret = PTR_ERR(fnew->mask);
1441         } else if (fold && fold->mask != fnew->mask) {
1442                 ret = -EINVAL;
1443         } else if (!refcount_inc_not_zero(&fnew->mask->refcnt)) {
1444                 /* Mask was deleted concurrently, try again */
1445                 ret = -EAGAIN;
1446         }
1447         rcu_read_unlock();
1448         return ret;
1449
1450 errout_cleanup:
1451         rhashtable_remove_fast(&head->ht, &mask->ht_node,
1452                                mask_ht_params);
1453         return ret;
1454 }
1455
1456 static int fl_set_parms(struct net *net, struct tcf_proto *tp,
1457                         struct cls_fl_filter *f, struct fl_flow_mask *mask,
1458                         unsigned long base, struct nlattr **tb,
1459                         struct nlattr *est, bool ovr,
1460                         struct fl_flow_tmplt *tmplt, bool rtnl_held,
1461                         struct netlink_ext_ack *extack)
1462 {
1463         int err;
1464
1465         err = tcf_exts_validate(net, tp, tb, est, &f->exts, ovr, rtnl_held,
1466                                 extack);
1467         if (err < 0)
1468                 return err;
1469
1470         if (tb[TCA_FLOWER_CLASSID]) {
1471                 f->res.classid = nla_get_u32(tb[TCA_FLOWER_CLASSID]);
1472                 if (!rtnl_held)
1473                         rtnl_lock();
1474                 tcf_bind_filter(tp, &f->res, base);
1475                 if (!rtnl_held)
1476                         rtnl_unlock();
1477         }
1478
1479         err = fl_set_key(net, tb, &f->key, &mask->key, extack);
1480         if (err)
1481                 return err;
1482
1483         fl_mask_update_range(mask);
1484         fl_set_masked_key(&f->mkey, &f->key, mask);
1485
1486         if (!fl_mask_fits_tmplt(tmplt, mask)) {
1487                 NL_SET_ERR_MSG_MOD(extack, "Mask does not fit the template");
1488                 return -EINVAL;
1489         }
1490
1491         return 0;
1492 }
1493
1494 static int fl_ht_insert_unique(struct cls_fl_filter *fnew,
1495                                struct cls_fl_filter *fold,
1496                                bool *in_ht)
1497 {
1498         struct fl_flow_mask *mask = fnew->mask;
1499         int err;
1500
1501         err = rhashtable_lookup_insert_fast(&mask->ht,
1502                                             &fnew->ht_node,
1503                                             mask->filter_ht_params);
1504         if (err) {
1505                 *in_ht = false;
1506                 /* It is okay if filter with same key exists when
1507                  * overwriting.
1508                  */
1509                 return fold && err == -EEXIST ? 0 : err;
1510         }
1511
1512         *in_ht = true;
1513         return 0;
1514 }
1515
1516 static int fl_change(struct net *net, struct sk_buff *in_skb,
1517                      struct tcf_proto *tp, unsigned long base,
1518                      u32 handle, struct nlattr **tca,
1519                      void **arg, bool ovr, bool rtnl_held,
1520                      struct netlink_ext_ack *extack)
1521 {
1522         struct cls_fl_head *head = fl_head_dereference(tp);
1523         struct cls_fl_filter *fold = *arg;
1524         struct cls_fl_filter *fnew;
1525         struct fl_flow_mask *mask;
1526         struct nlattr **tb;
1527         bool in_ht;
1528         int err;
1529
1530         if (!tca[TCA_OPTIONS]) {
1531                 err = -EINVAL;
1532                 goto errout_fold;
1533         }
1534
1535         mask = kzalloc(sizeof(struct fl_flow_mask), GFP_KERNEL);
1536         if (!mask) {
1537                 err = -ENOBUFS;
1538                 goto errout_fold;
1539         }
1540
1541         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1542         if (!tb) {
1543                 err = -ENOBUFS;
1544                 goto errout_mask_alloc;
1545         }
1546
1547         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1548                                           tca[TCA_OPTIONS], fl_policy, NULL);
1549         if (err < 0)
1550                 goto errout_tb;
1551
1552         if (fold && handle && fold->handle != handle) {
1553                 err = -EINVAL;
1554                 goto errout_tb;
1555         }
1556
1557         fnew = kzalloc(sizeof(*fnew), GFP_KERNEL);
1558         if (!fnew) {
1559                 err = -ENOBUFS;
1560                 goto errout_tb;
1561         }
1562         INIT_LIST_HEAD(&fnew->hw_list);
1563         refcount_set(&fnew->refcnt, 1);
1564
1565         err = tcf_exts_init(&fnew->exts, net, TCA_FLOWER_ACT, 0);
1566         if (err < 0)
1567                 goto errout;
1568
1569         if (tb[TCA_FLOWER_FLAGS]) {
1570                 fnew->flags = nla_get_u32(tb[TCA_FLOWER_FLAGS]);
1571
1572                 if (!tc_flags_valid(fnew->flags)) {
1573                         err = -EINVAL;
1574                         goto errout;
1575                 }
1576         }
1577
1578         err = fl_set_parms(net, tp, fnew, mask, base, tb, tca[TCA_RATE], ovr,
1579                            tp->chain->tmplt_priv, rtnl_held, extack);
1580         if (err)
1581                 goto errout;
1582
1583         err = fl_check_assign_mask(head, fnew, fold, mask);
1584         if (err)
1585                 goto errout;
1586
1587         err = fl_ht_insert_unique(fnew, fold, &in_ht);
1588         if (err)
1589                 goto errout_mask;
1590
1591         if (!tc_skip_hw(fnew->flags)) {
1592                 err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
1593                 if (err)
1594                         goto errout_ht;
1595         }
1596
1597         if (!tc_in_hw(fnew->flags))
1598                 fnew->flags |= TCA_CLS_FLAGS_NOT_IN_HW;
1599
1600         spin_lock(&tp->lock);
1601
1602         /* tp was deleted concurrently. -EAGAIN will cause caller to lookup
1603          * proto again or create new one, if necessary.
1604          */
1605         if (tp->deleting) {
1606                 err = -EAGAIN;
1607                 goto errout_hw;
1608         }
1609
1610         if (fold) {
1611                 /* Fold filter was deleted concurrently. Retry lookup. */
1612                 if (fold->deleted) {
1613                         err = -EAGAIN;
1614                         goto errout_hw;
1615                 }
1616
1617                 fnew->handle = handle;
1618
1619                 if (!in_ht) {
1620                         struct rhashtable_params params =
1621                                 fnew->mask->filter_ht_params;
1622
1623                         err = rhashtable_insert_fast(&fnew->mask->ht,
1624                                                      &fnew->ht_node,
1625                                                      params);
1626                         if (err)
1627                                 goto errout_hw;
1628                         in_ht = true;
1629                 }
1630
1631                 refcount_inc(&fnew->refcnt);
1632                 rhashtable_remove_fast(&fold->mask->ht,
1633                                        &fold->ht_node,
1634                                        fold->mask->filter_ht_params);
1635                 idr_replace(&head->handle_idr, fnew, fnew->handle);
1636                 list_replace_rcu(&fold->list, &fnew->list);
1637                 fold->deleted = true;
1638
1639                 spin_unlock(&tp->lock);
1640
1641                 fl_mask_put(head, fold->mask);
1642                 if (!tc_skip_hw(fold->flags))
1643                         fl_hw_destroy_filter(tp, fold, rtnl_held, NULL);
1644                 tcf_unbind_filter(tp, &fold->res);
1645                 /* Caller holds reference to fold, so refcnt is always > 0
1646                  * after this.
1647                  */
1648                 refcount_dec(&fold->refcnt);
1649                 __fl_put(fold);
1650         } else {
1651                 if (handle) {
1652                         /* user specifies a handle and it doesn't exist */
1653                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1654                                             handle, GFP_ATOMIC);
1655
1656                         /* Filter with specified handle was concurrently
1657                          * inserted after initial check in cls_api. This is not
1658                          * necessarily an error if NLM_F_EXCL is not set in
1659                          * message flags. Returning EAGAIN will cause cls_api to
1660                          * try to update concurrently inserted rule.
1661                          */
1662                         if (err == -ENOSPC)
1663                                 err = -EAGAIN;
1664                 } else {
1665                         handle = 1;
1666                         err = idr_alloc_u32(&head->handle_idr, fnew, &handle,
1667                                             INT_MAX, GFP_ATOMIC);
1668                 }
1669                 if (err)
1670                         goto errout_hw;
1671
1672                 refcount_inc(&fnew->refcnt);
1673                 fnew->handle = handle;
1674                 list_add_tail_rcu(&fnew->list, &fnew->mask->filters);
1675                 spin_unlock(&tp->lock);
1676         }
1677
1678         *arg = fnew;
1679
1680         kfree(tb);
1681         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
1682         return 0;
1683
1684 errout_ht:
1685         spin_lock(&tp->lock);
1686 errout_hw:
1687         fnew->deleted = true;
1688         spin_unlock(&tp->lock);
1689         if (!tc_skip_hw(fnew->flags))
1690                 fl_hw_destroy_filter(tp, fnew, rtnl_held, NULL);
1691         if (in_ht)
1692                 rhashtable_remove_fast(&fnew->mask->ht, &fnew->ht_node,
1693                                        fnew->mask->filter_ht_params);
1694 errout_mask:
1695         fl_mask_put(head, fnew->mask);
1696 errout:
1697         __fl_put(fnew);
1698 errout_tb:
1699         kfree(tb);
1700 errout_mask_alloc:
1701         tcf_queue_work(&mask->rwork, fl_uninit_mask_free_work);
1702 errout_fold:
1703         if (fold)
1704                 __fl_put(fold);
1705         return err;
1706 }
1707
1708 static int fl_delete(struct tcf_proto *tp, void *arg, bool *last,
1709                      bool rtnl_held, struct netlink_ext_ack *extack)
1710 {
1711         struct cls_fl_head *head = fl_head_dereference(tp);
1712         struct cls_fl_filter *f = arg;
1713         bool last_on_mask;
1714         int err = 0;
1715
1716         err = __fl_delete(tp, f, &last_on_mask, rtnl_held, extack);
1717         *last = list_empty(&head->masks);
1718         __fl_put(f);
1719
1720         return err;
1721 }
1722
1723 static void fl_walk(struct tcf_proto *tp, struct tcf_walker *arg,
1724                     bool rtnl_held)
1725 {
1726         struct cls_fl_head *head = fl_head_dereference(tp);
1727         unsigned long id = arg->cookie, tmp;
1728         struct cls_fl_filter *f;
1729
1730         arg->count = arg->skip;
1731
1732         idr_for_each_entry_continue_ul(&head->handle_idr, f, tmp, id) {
1733                 /* don't return filters that are being deleted */
1734                 if (!refcount_inc_not_zero(&f->refcnt))
1735                         continue;
1736                 if (arg->fn(tp, f, arg) < 0) {
1737                         __fl_put(f);
1738                         arg->stop = 1;
1739                         break;
1740                 }
1741                 __fl_put(f);
1742                 arg->count++;
1743         }
1744         arg->cookie = id;
1745 }
1746
1747 static struct cls_fl_filter *
1748 fl_get_next_hw_filter(struct tcf_proto *tp, struct cls_fl_filter *f, bool add)
1749 {
1750         struct cls_fl_head *head = fl_head_dereference(tp);
1751
1752         spin_lock(&tp->lock);
1753         if (list_empty(&head->hw_filters)) {
1754                 spin_unlock(&tp->lock);
1755                 return NULL;
1756         }
1757
1758         if (!f)
1759                 f = list_entry(&head->hw_filters, struct cls_fl_filter,
1760                                hw_list);
1761         list_for_each_entry_continue(f, &head->hw_filters, hw_list) {
1762                 if (!(add && f->deleted) && refcount_inc_not_zero(&f->refcnt)) {
1763                         spin_unlock(&tp->lock);
1764                         return f;
1765                 }
1766         }
1767
1768         spin_unlock(&tp->lock);
1769         return NULL;
1770 }
1771
1772 static int fl_reoffload(struct tcf_proto *tp, bool add, flow_setup_cb_t *cb,
1773                         void *cb_priv, struct netlink_ext_ack *extack)
1774 {
1775         struct tcf_block *block = tp->chain->block;
1776         struct flow_cls_offload cls_flower = {};
1777         struct cls_fl_filter *f = NULL;
1778         int err;
1779
1780         /* hw_filters list can only be changed by hw offload functions after
1781          * obtaining rtnl lock. Make sure it is not changed while reoffload is
1782          * iterating it.
1783          */
1784         ASSERT_RTNL();
1785
1786         while ((f = fl_get_next_hw_filter(tp, f, add))) {
1787                 cls_flower.rule =
1788                         flow_rule_alloc(tcf_exts_num_actions(&f->exts));
1789                 if (!cls_flower.rule) {
1790                         __fl_put(f);
1791                         return -ENOMEM;
1792                 }
1793
1794                 tc_cls_common_offload_init(&cls_flower.common, tp, f->flags,
1795                                            extack);
1796                 cls_flower.command = add ?
1797                         FLOW_CLS_REPLACE : FLOW_CLS_DESTROY;
1798                 cls_flower.cookie = (unsigned long)f;
1799                 cls_flower.rule->match.dissector = &f->mask->dissector;
1800                 cls_flower.rule->match.mask = &f->mask->key;
1801                 cls_flower.rule->match.key = &f->mkey;
1802
1803                 err = tc_setup_flow_action(&cls_flower.rule->action, &f->exts,
1804                                            true);
1805                 if (err) {
1806                         kfree(cls_flower.rule);
1807                         if (tc_skip_sw(f->flags)) {
1808                                 NL_SET_ERR_MSG_MOD(extack, "Failed to setup flow action");
1809                                 __fl_put(f);
1810                                 return err;
1811                         }
1812                         goto next_flow;
1813                 }
1814
1815                 cls_flower.classid = f->res.classid;
1816
1817                 err = tc_setup_cb_reoffload(block, tp, add, cb,
1818                                             TC_SETUP_CLSFLOWER, &cls_flower,
1819                                             cb_priv, &f->flags,
1820                                             &f->in_hw_count);
1821                 tc_cleanup_flow_action(&cls_flower.rule->action);
1822                 kfree(cls_flower.rule);
1823
1824                 if (err) {
1825                         __fl_put(f);
1826                         return err;
1827                 }
1828 next_flow:
1829                 __fl_put(f);
1830         }
1831
1832         return 0;
1833 }
1834
1835 static void fl_hw_add(struct tcf_proto *tp, void *type_data)
1836 {
1837         struct flow_cls_offload *cls_flower = type_data;
1838         struct cls_fl_filter *f =
1839                 (struct cls_fl_filter *) cls_flower->cookie;
1840         struct cls_fl_head *head = fl_head_dereference(tp);
1841
1842         spin_lock(&tp->lock);
1843         list_add(&f->hw_list, &head->hw_filters);
1844         spin_unlock(&tp->lock);
1845 }
1846
1847 static void fl_hw_del(struct tcf_proto *tp, void *type_data)
1848 {
1849         struct flow_cls_offload *cls_flower = type_data;
1850         struct cls_fl_filter *f =
1851                 (struct cls_fl_filter *) cls_flower->cookie;
1852
1853         spin_lock(&tp->lock);
1854         if (!list_empty(&f->hw_list))
1855                 list_del_init(&f->hw_list);
1856         spin_unlock(&tp->lock);
1857 }
1858
1859 static int fl_hw_create_tmplt(struct tcf_chain *chain,
1860                               struct fl_flow_tmplt *tmplt)
1861 {
1862         struct flow_cls_offload cls_flower = {};
1863         struct tcf_block *block = chain->block;
1864
1865         cls_flower.rule = flow_rule_alloc(0);
1866         if (!cls_flower.rule)
1867                 return -ENOMEM;
1868
1869         cls_flower.common.chain_index = chain->index;
1870         cls_flower.command = FLOW_CLS_TMPLT_CREATE;
1871         cls_flower.cookie = (unsigned long) tmplt;
1872         cls_flower.rule->match.dissector = &tmplt->dissector;
1873         cls_flower.rule->match.mask = &tmplt->mask;
1874         cls_flower.rule->match.key = &tmplt->dummy_key;
1875
1876         /* We don't care if driver (any of them) fails to handle this
1877          * call. It serves just as a hint for it.
1878          */
1879         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
1880         kfree(cls_flower.rule);
1881
1882         return 0;
1883 }
1884
1885 static void fl_hw_destroy_tmplt(struct tcf_chain *chain,
1886                                 struct fl_flow_tmplt *tmplt)
1887 {
1888         struct flow_cls_offload cls_flower = {};
1889         struct tcf_block *block = chain->block;
1890
1891         cls_flower.common.chain_index = chain->index;
1892         cls_flower.command = FLOW_CLS_TMPLT_DESTROY;
1893         cls_flower.cookie = (unsigned long) tmplt;
1894
1895         tc_setup_cb_call(block, TC_SETUP_CLSFLOWER, &cls_flower, false, true);
1896 }
1897
1898 static void *fl_tmplt_create(struct net *net, struct tcf_chain *chain,
1899                              struct nlattr **tca,
1900                              struct netlink_ext_ack *extack)
1901 {
1902         struct fl_flow_tmplt *tmplt;
1903         struct nlattr **tb;
1904         int err;
1905
1906         if (!tca[TCA_OPTIONS])
1907                 return ERR_PTR(-EINVAL);
1908
1909         tb = kcalloc(TCA_FLOWER_MAX + 1, sizeof(struct nlattr *), GFP_KERNEL);
1910         if (!tb)
1911                 return ERR_PTR(-ENOBUFS);
1912         err = nla_parse_nested_deprecated(tb, TCA_FLOWER_MAX,
1913                                           tca[TCA_OPTIONS], fl_policy, NULL);
1914         if (err)
1915                 goto errout_tb;
1916
1917         tmplt = kzalloc(sizeof(*tmplt), GFP_KERNEL);
1918         if (!tmplt) {
1919                 err = -ENOMEM;
1920                 goto errout_tb;
1921         }
1922         tmplt->chain = chain;
1923         err = fl_set_key(net, tb, &tmplt->dummy_key, &tmplt->mask, extack);
1924         if (err)
1925                 goto errout_tmplt;
1926
1927         fl_init_dissector(&tmplt->dissector, &tmplt->mask);
1928
1929         err = fl_hw_create_tmplt(chain, tmplt);
1930         if (err)
1931                 goto errout_tmplt;
1932
1933         kfree(tb);
1934         return tmplt;
1935
1936 errout_tmplt:
1937         kfree(tmplt);
1938 errout_tb:
1939         kfree(tb);
1940         return ERR_PTR(err);
1941 }
1942
1943 static void fl_tmplt_destroy(void *tmplt_priv)
1944 {
1945         struct fl_flow_tmplt *tmplt = tmplt_priv;
1946
1947         fl_hw_destroy_tmplt(tmplt->chain, tmplt);
1948         kfree(tmplt);
1949 }
1950
1951 static int fl_dump_key_val(struct sk_buff *skb,
1952                            void *val, int val_type,
1953                            void *mask, int mask_type, int len)
1954 {
1955         int err;
1956
1957         if (!memchr_inv(mask, 0, len))
1958                 return 0;
1959         err = nla_put(skb, val_type, len, val);
1960         if (err)
1961                 return err;
1962         if (mask_type != TCA_FLOWER_UNSPEC) {
1963                 err = nla_put(skb, mask_type, len, mask);
1964                 if (err)
1965                         return err;
1966         }
1967         return 0;
1968 }
1969
1970 static int fl_dump_key_port_range(struct sk_buff *skb, struct fl_flow_key *key,
1971                                   struct fl_flow_key *mask)
1972 {
1973         if (fl_dump_key_val(skb, &key->tp_min.dst, TCA_FLOWER_KEY_PORT_DST_MIN,
1974                             &mask->tp_min.dst, TCA_FLOWER_UNSPEC,
1975                             sizeof(key->tp_min.dst)) ||
1976             fl_dump_key_val(skb, &key->tp_max.dst, TCA_FLOWER_KEY_PORT_DST_MAX,
1977                             &mask->tp_max.dst, TCA_FLOWER_UNSPEC,
1978                             sizeof(key->tp_max.dst)) ||
1979             fl_dump_key_val(skb, &key->tp_min.src, TCA_FLOWER_KEY_PORT_SRC_MIN,
1980                             &mask->tp_min.src, TCA_FLOWER_UNSPEC,
1981                             sizeof(key->tp_min.src)) ||
1982             fl_dump_key_val(skb, &key->tp_max.src, TCA_FLOWER_KEY_PORT_SRC_MAX,
1983                             &mask->tp_max.src, TCA_FLOWER_UNSPEC,
1984                             sizeof(key->tp_max.src)))
1985                 return -1;
1986
1987         return 0;
1988 }
1989
1990 static int fl_dump_key_mpls(struct sk_buff *skb,
1991                             struct flow_dissector_key_mpls *mpls_key,
1992                             struct flow_dissector_key_mpls *mpls_mask)
1993 {
1994         int err;
1995
1996         if (!memchr_inv(mpls_mask, 0, sizeof(*mpls_mask)))
1997                 return 0;
1998         if (mpls_mask->mpls_ttl) {
1999                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TTL,
2000                                  mpls_key->mpls_ttl);
2001                 if (err)
2002                         return err;
2003         }
2004         if (mpls_mask->mpls_tc) {
2005                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_TC,
2006                                  mpls_key->mpls_tc);
2007                 if (err)
2008                         return err;
2009         }
2010         if (mpls_mask->mpls_label) {
2011                 err = nla_put_u32(skb, TCA_FLOWER_KEY_MPLS_LABEL,
2012                                   mpls_key->mpls_label);
2013                 if (err)
2014                         return err;
2015         }
2016         if (mpls_mask->mpls_bos) {
2017                 err = nla_put_u8(skb, TCA_FLOWER_KEY_MPLS_BOS,
2018                                  mpls_key->mpls_bos);
2019                 if (err)
2020                         return err;
2021         }
2022         return 0;
2023 }
2024
2025 static int fl_dump_key_ip(struct sk_buff *skb, bool encap,
2026                           struct flow_dissector_key_ip *key,
2027                           struct flow_dissector_key_ip *mask)
2028 {
2029         int tos_key = encap ? TCA_FLOWER_KEY_ENC_IP_TOS : TCA_FLOWER_KEY_IP_TOS;
2030         int ttl_key = encap ? TCA_FLOWER_KEY_ENC_IP_TTL : TCA_FLOWER_KEY_IP_TTL;
2031         int tos_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TOS_MASK : TCA_FLOWER_KEY_IP_TOS_MASK;
2032         int ttl_mask = encap ? TCA_FLOWER_KEY_ENC_IP_TTL_MASK : TCA_FLOWER_KEY_IP_TTL_MASK;
2033
2034         if (fl_dump_key_val(skb, &key->tos, tos_key, &mask->tos, tos_mask, sizeof(key->tos)) ||
2035             fl_dump_key_val(skb, &key->ttl, ttl_key, &mask->ttl, ttl_mask, sizeof(key->ttl)))
2036                 return -1;
2037
2038         return 0;
2039 }
2040
2041 static int fl_dump_key_vlan(struct sk_buff *skb,
2042                             int vlan_id_key, int vlan_prio_key,
2043                             struct flow_dissector_key_vlan *vlan_key,
2044                             struct flow_dissector_key_vlan *vlan_mask)
2045 {
2046         int err;
2047
2048         if (!memchr_inv(vlan_mask, 0, sizeof(*vlan_mask)))
2049                 return 0;
2050         if (vlan_mask->vlan_id) {
2051                 err = nla_put_u16(skb, vlan_id_key,
2052                                   vlan_key->vlan_id);
2053                 if (err)
2054                         return err;
2055         }
2056         if (vlan_mask->vlan_priority) {
2057                 err = nla_put_u8(skb, vlan_prio_key,
2058                                  vlan_key->vlan_priority);
2059                 if (err)
2060                         return err;
2061         }
2062         return 0;
2063 }
2064
2065 static void fl_get_key_flag(u32 dissector_key, u32 dissector_mask,
2066                             u32 *flower_key, u32 *flower_mask,
2067                             u32 flower_flag_bit, u32 dissector_flag_bit)
2068 {
2069         if (dissector_mask & dissector_flag_bit) {
2070                 *flower_mask |= flower_flag_bit;
2071                 if (dissector_key & dissector_flag_bit)
2072                         *flower_key |= flower_flag_bit;
2073         }
2074 }
2075
2076 static int fl_dump_key_flags(struct sk_buff *skb, u32 flags_key, u32 flags_mask)
2077 {
2078         u32 key, mask;
2079         __be32 _key, _mask;
2080         int err;
2081
2082         if (!memchr_inv(&flags_mask, 0, sizeof(flags_mask)))
2083                 return 0;
2084
2085         key = 0;
2086         mask = 0;
2087
2088         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2089                         TCA_FLOWER_KEY_FLAGS_IS_FRAGMENT, FLOW_DIS_IS_FRAGMENT);
2090         fl_get_key_flag(flags_key, flags_mask, &key, &mask,
2091                         TCA_FLOWER_KEY_FLAGS_FRAG_IS_FIRST,
2092                         FLOW_DIS_FIRST_FRAG);
2093
2094         _key = cpu_to_be32(key);
2095         _mask = cpu_to_be32(mask);
2096
2097         err = nla_put(skb, TCA_FLOWER_KEY_FLAGS, 4, &_key);
2098         if (err)
2099                 return err;
2100
2101         return nla_put(skb, TCA_FLOWER_KEY_FLAGS_MASK, 4, &_mask);
2102 }
2103
2104 static int fl_dump_key_geneve_opt(struct sk_buff *skb,
2105                                   struct flow_dissector_key_enc_opts *enc_opts)
2106 {
2107         struct geneve_opt *opt;
2108         struct nlattr *nest;
2109         int opt_off = 0;
2110
2111         nest = nla_nest_start_noflag(skb, TCA_FLOWER_KEY_ENC_OPTS_GENEVE);
2112         if (!nest)
2113                 goto nla_put_failure;
2114
2115         while (enc_opts->len > opt_off) {
2116                 opt = (struct geneve_opt *)&enc_opts->data[opt_off];
2117
2118                 if (nla_put_be16(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_CLASS,
2119                                  opt->opt_class))
2120                         goto nla_put_failure;
2121                 if (nla_put_u8(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_TYPE,
2122                                opt->type))
2123                         goto nla_put_failure;
2124                 if (nla_put(skb, TCA_FLOWER_KEY_ENC_OPT_GENEVE_DATA,
2125                             opt->length * 4, opt->opt_data))
2126                         goto nla_put_failure;
2127
2128                 opt_off += sizeof(struct geneve_opt) + opt->length * 4;
2129         }
2130         nla_nest_end(skb, nest);
2131         return 0;
2132
2133 nla_put_failure:
2134         nla_nest_cancel(skb, nest);
2135         return -EMSGSIZE;
2136 }
2137
2138 static int fl_dump_key_ct(struct sk_buff *skb,
2139                           struct flow_dissector_key_ct *key,
2140                           struct flow_dissector_key_ct *mask)
2141 {
2142         if (IS_ENABLED(CONFIG_NF_CONNTRACK) &&
2143             fl_dump_key_val(skb, &key->ct_state, TCA_FLOWER_KEY_CT_STATE,
2144                             &mask->ct_state, TCA_FLOWER_KEY_CT_STATE_MASK,
2145                             sizeof(key->ct_state)))
2146                 goto nla_put_failure;
2147
2148         if (IS_ENABLED(CONFIG_NF_CONNTRACK_ZONES) &&
2149             fl_dump_key_val(skb, &key->ct_zone, TCA_FLOWER_KEY_CT_ZONE,
2150                             &mask->ct_zone, TCA_FLOWER_KEY_CT_ZONE_MASK,
2151                             sizeof(key->ct_zone)))
2152                 goto nla_put_failure;
2153
2154         if (IS_ENABLED(CONFIG_NF_CONNTRACK_MARK) &&
2155             fl_dump_key_val(skb, &key->ct_mark, TCA_FLOWER_KEY_CT_MARK,
2156                             &mask->ct_mark, TCA_FLOWER_KEY_CT_MARK_MASK,
2157                             sizeof(key->ct_mark)))
2158                 goto nla_put_failure;
2159
2160         if (IS_ENABLED(CONFIG_NF_CONNTRACK_LABELS) &&
2161             fl_dump_key_val(skb, &key->ct_labels, TCA_FLOWER_KEY_CT_LABELS,
2162                             &mask->ct_labels, TCA_FLOWER_KEY_CT_LABELS_MASK,
2163                             sizeof(key->ct_labels)))
2164                 goto nla_put_failure;
2165
2166         return 0;
2167
2168 nla_put_failure:
2169         return -EMSGSIZE;
2170 }
2171
2172 static int fl_dump_key_options(struct sk_buff *skb, int enc_opt_type,
2173                                struct flow_dissector_key_enc_opts *enc_opts)
2174 {
2175         struct nlattr *nest;
2176         int err;
2177
2178         if (!enc_opts->len)
2179                 return 0;
2180
2181         nest = nla_nest_start_noflag(skb, enc_opt_type);
2182         if (!nest)
2183                 goto nla_put_failure;
2184
2185         switch (enc_opts->dst_opt_type) {
2186         case TUNNEL_GENEVE_OPT:
2187                 err = fl_dump_key_geneve_opt(skb, enc_opts);
2188                 if (err)
2189                         goto nla_put_failure;
2190                 break;
2191         default:
2192                 goto nla_put_failure;
2193         }
2194         nla_nest_end(skb, nest);
2195         return 0;
2196
2197 nla_put_failure:
2198         nla_nest_cancel(skb, nest);
2199         return -EMSGSIZE;
2200 }
2201
2202 static int fl_dump_key_enc_opt(struct sk_buff *skb,
2203                                struct flow_dissector_key_enc_opts *key_opts,
2204                                struct flow_dissector_key_enc_opts *msk_opts)
2205 {
2206         int err;
2207
2208         err = fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS, key_opts);
2209         if (err)
2210                 return err;
2211
2212         return fl_dump_key_options(skb, TCA_FLOWER_KEY_ENC_OPTS_MASK, msk_opts);
2213 }
2214
2215 static int fl_dump_key(struct sk_buff *skb, struct net *net,
2216                        struct fl_flow_key *key, struct fl_flow_key *mask)
2217 {
2218         if (mask->meta.ingress_ifindex) {
2219                 struct net_device *dev;
2220
2221                 dev = __dev_get_by_index(net, key->meta.ingress_ifindex);
2222                 if (dev && nla_put_string(skb, TCA_FLOWER_INDEV, dev->name))
2223                         goto nla_put_failure;
2224         }
2225
2226         if (fl_dump_key_val(skb, key->eth.dst, TCA_FLOWER_KEY_ETH_DST,
2227                             mask->eth.dst, TCA_FLOWER_KEY_ETH_DST_MASK,
2228                             sizeof(key->eth.dst)) ||
2229             fl_dump_key_val(skb, key->eth.src, TCA_FLOWER_KEY_ETH_SRC,
2230                             mask->eth.src, TCA_FLOWER_KEY_ETH_SRC_MASK,
2231                             sizeof(key->eth.src)) ||
2232             fl_dump_key_val(skb, &key->basic.n_proto, TCA_FLOWER_KEY_ETH_TYPE,
2233                             &mask->basic.n_proto, TCA_FLOWER_UNSPEC,
2234                             sizeof(key->basic.n_proto)))
2235                 goto nla_put_failure;
2236
2237         if (fl_dump_key_mpls(skb, &key->mpls, &mask->mpls))
2238                 goto nla_put_failure;
2239
2240         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_VLAN_ID,
2241                              TCA_FLOWER_KEY_VLAN_PRIO, &key->vlan, &mask->vlan))
2242                 goto nla_put_failure;
2243
2244         if (fl_dump_key_vlan(skb, TCA_FLOWER_KEY_CVLAN_ID,
2245                              TCA_FLOWER_KEY_CVLAN_PRIO,
2246                              &key->cvlan, &mask->cvlan) ||
2247             (mask->cvlan.vlan_tpid &&
2248              nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2249                           key->cvlan.vlan_tpid)))
2250                 goto nla_put_failure;
2251
2252         if (mask->basic.n_proto) {
2253                 if (mask->cvlan.vlan_tpid) {
2254                         if (nla_put_be16(skb, TCA_FLOWER_KEY_CVLAN_ETH_TYPE,
2255                                          key->basic.n_proto))
2256                                 goto nla_put_failure;
2257                 } else if (mask->vlan.vlan_tpid) {
2258                         if (nla_put_be16(skb, TCA_FLOWER_KEY_VLAN_ETH_TYPE,
2259                                          key->basic.n_proto))
2260                                 goto nla_put_failure;
2261                 }
2262         }
2263
2264         if ((key->basic.n_proto == htons(ETH_P_IP) ||
2265              key->basic.n_proto == htons(ETH_P_IPV6)) &&
2266             (fl_dump_key_val(skb, &key->basic.ip_proto, TCA_FLOWER_KEY_IP_PROTO,
2267                             &mask->basic.ip_proto, TCA_FLOWER_UNSPEC,
2268                             sizeof(key->basic.ip_proto)) ||
2269             fl_dump_key_ip(skb, false, &key->ip, &mask->ip)))
2270                 goto nla_put_failure;
2271
2272         if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2273             (fl_dump_key_val(skb, &key->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC,
2274                              &mask->ipv4.src, TCA_FLOWER_KEY_IPV4_SRC_MASK,
2275                              sizeof(key->ipv4.src)) ||
2276              fl_dump_key_val(skb, &key->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST,
2277                              &mask->ipv4.dst, TCA_FLOWER_KEY_IPV4_DST_MASK,
2278                              sizeof(key->ipv4.dst))))
2279                 goto nla_put_failure;
2280         else if (key->control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2281                  (fl_dump_key_val(skb, &key->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC,
2282                                   &mask->ipv6.src, TCA_FLOWER_KEY_IPV6_SRC_MASK,
2283                                   sizeof(key->ipv6.src)) ||
2284                   fl_dump_key_val(skb, &key->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST,
2285                                   &mask->ipv6.dst, TCA_FLOWER_KEY_IPV6_DST_MASK,
2286                                   sizeof(key->ipv6.dst))))
2287                 goto nla_put_failure;
2288
2289         if (key->basic.ip_proto == IPPROTO_TCP &&
2290             (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_TCP_SRC,
2291                              &mask->tp.src, TCA_FLOWER_KEY_TCP_SRC_MASK,
2292                              sizeof(key->tp.src)) ||
2293              fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_TCP_DST,
2294                              &mask->tp.dst, TCA_FLOWER_KEY_TCP_DST_MASK,
2295                              sizeof(key->tp.dst)) ||
2296              fl_dump_key_val(skb, &key->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS,
2297                              &mask->tcp.flags, TCA_FLOWER_KEY_TCP_FLAGS_MASK,
2298                              sizeof(key->tcp.flags))))
2299                 goto nla_put_failure;
2300         else if (key->basic.ip_proto == IPPROTO_UDP &&
2301                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_UDP_SRC,
2302                                   &mask->tp.src, TCA_FLOWER_KEY_UDP_SRC_MASK,
2303                                   sizeof(key->tp.src)) ||
2304                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_UDP_DST,
2305                                   &mask->tp.dst, TCA_FLOWER_KEY_UDP_DST_MASK,
2306                                   sizeof(key->tp.dst))))
2307                 goto nla_put_failure;
2308         else if (key->basic.ip_proto == IPPROTO_SCTP &&
2309                  (fl_dump_key_val(skb, &key->tp.src, TCA_FLOWER_KEY_SCTP_SRC,
2310                                   &mask->tp.src, TCA_FLOWER_KEY_SCTP_SRC_MASK,
2311                                   sizeof(key->tp.src)) ||
2312                   fl_dump_key_val(skb, &key->tp.dst, TCA_FLOWER_KEY_SCTP_DST,
2313                                   &mask->tp.dst, TCA_FLOWER_KEY_SCTP_DST_MASK,
2314                                   sizeof(key->tp.dst))))
2315                 goto nla_put_failure;
2316         else if (key->basic.n_proto == htons(ETH_P_IP) &&
2317                  key->basic.ip_proto == IPPROTO_ICMP &&
2318                  (fl_dump_key_val(skb, &key->icmp.type,
2319                                   TCA_FLOWER_KEY_ICMPV4_TYPE, &mask->icmp.type,
2320                                   TCA_FLOWER_KEY_ICMPV4_TYPE_MASK,
2321                                   sizeof(key->icmp.type)) ||
2322                   fl_dump_key_val(skb, &key->icmp.code,
2323                                   TCA_FLOWER_KEY_ICMPV4_CODE, &mask->icmp.code,
2324                                   TCA_FLOWER_KEY_ICMPV4_CODE_MASK,
2325                                   sizeof(key->icmp.code))))
2326                 goto nla_put_failure;
2327         else if (key->basic.n_proto == htons(ETH_P_IPV6) &&
2328                  key->basic.ip_proto == IPPROTO_ICMPV6 &&
2329                  (fl_dump_key_val(skb, &key->icmp.type,
2330                                   TCA_FLOWER_KEY_ICMPV6_TYPE, &mask->icmp.type,
2331                                   TCA_FLOWER_KEY_ICMPV6_TYPE_MASK,
2332                                   sizeof(key->icmp.type)) ||
2333                   fl_dump_key_val(skb, &key->icmp.code,
2334                                   TCA_FLOWER_KEY_ICMPV6_CODE, &mask->icmp.code,
2335                                   TCA_FLOWER_KEY_ICMPV6_CODE_MASK,
2336                                   sizeof(key->icmp.code))))
2337                 goto nla_put_failure;
2338         else if ((key->basic.n_proto == htons(ETH_P_ARP) ||
2339                   key->basic.n_proto == htons(ETH_P_RARP)) &&
2340                  (fl_dump_key_val(skb, &key->arp.sip,
2341                                   TCA_FLOWER_KEY_ARP_SIP, &mask->arp.sip,
2342                                   TCA_FLOWER_KEY_ARP_SIP_MASK,
2343                                   sizeof(key->arp.sip)) ||
2344                   fl_dump_key_val(skb, &key->arp.tip,
2345                                   TCA_FLOWER_KEY_ARP_TIP, &mask->arp.tip,
2346                                   TCA_FLOWER_KEY_ARP_TIP_MASK,
2347                                   sizeof(key->arp.tip)) ||
2348                   fl_dump_key_val(skb, &key->arp.op,
2349                                   TCA_FLOWER_KEY_ARP_OP, &mask->arp.op,
2350                                   TCA_FLOWER_KEY_ARP_OP_MASK,
2351                                   sizeof(key->arp.op)) ||
2352                   fl_dump_key_val(skb, key->arp.sha, TCA_FLOWER_KEY_ARP_SHA,
2353                                   mask->arp.sha, TCA_FLOWER_KEY_ARP_SHA_MASK,
2354                                   sizeof(key->arp.sha)) ||
2355                   fl_dump_key_val(skb, key->arp.tha, TCA_FLOWER_KEY_ARP_THA,
2356                                   mask->arp.tha, TCA_FLOWER_KEY_ARP_THA_MASK,
2357                                   sizeof(key->arp.tha))))
2358                 goto nla_put_failure;
2359
2360         if ((key->basic.ip_proto == IPPROTO_TCP ||
2361              key->basic.ip_proto == IPPROTO_UDP ||
2362              key->basic.ip_proto == IPPROTO_SCTP) &&
2363              fl_dump_key_port_range(skb, key, mask))
2364                 goto nla_put_failure;
2365
2366         if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV4_ADDRS &&
2367             (fl_dump_key_val(skb, &key->enc_ipv4.src,
2368                             TCA_FLOWER_KEY_ENC_IPV4_SRC, &mask->enc_ipv4.src,
2369                             TCA_FLOWER_KEY_ENC_IPV4_SRC_MASK,
2370                             sizeof(key->enc_ipv4.src)) ||
2371              fl_dump_key_val(skb, &key->enc_ipv4.dst,
2372                              TCA_FLOWER_KEY_ENC_IPV4_DST, &mask->enc_ipv4.dst,
2373                              TCA_FLOWER_KEY_ENC_IPV4_DST_MASK,
2374                              sizeof(key->enc_ipv4.dst))))
2375                 goto nla_put_failure;
2376         else if (key->enc_control.addr_type == FLOW_DISSECTOR_KEY_IPV6_ADDRS &&
2377                  (fl_dump_key_val(skb, &key->enc_ipv6.src,
2378                             TCA_FLOWER_KEY_ENC_IPV6_SRC, &mask->enc_ipv6.src,
2379                             TCA_FLOWER_KEY_ENC_IPV6_SRC_MASK,
2380                             sizeof(key->enc_ipv6.src)) ||
2381                  fl_dump_key_val(skb, &key->enc_ipv6.dst,
2382                                  TCA_FLOWER_KEY_ENC_IPV6_DST,
2383                                  &mask->enc_ipv6.dst,
2384                                  TCA_FLOWER_KEY_ENC_IPV6_DST_MASK,
2385                             sizeof(key->enc_ipv6.dst))))
2386                 goto nla_put_failure;
2387
2388         if (fl_dump_key_val(skb, &key->enc_key_id, TCA_FLOWER_KEY_ENC_KEY_ID,
2389                             &mask->enc_key_id, TCA_FLOWER_UNSPEC,
2390                             sizeof(key->enc_key_id)) ||
2391             fl_dump_key_val(skb, &key->enc_tp.src,
2392                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT,
2393                             &mask->enc_tp.src,
2394                             TCA_FLOWER_KEY_ENC_UDP_SRC_PORT_MASK,
2395                             sizeof(key->enc_tp.src)) ||
2396             fl_dump_key_val(skb, &key->enc_tp.dst,
2397                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT,
2398                             &mask->enc_tp.dst,
2399                             TCA_FLOWER_KEY_ENC_UDP_DST_PORT_MASK,
2400                             sizeof(key->enc_tp.dst)) ||
2401             fl_dump_key_ip(skb, true, &key->enc_ip, &mask->enc_ip) ||
2402             fl_dump_key_enc_opt(skb, &key->enc_opts, &mask->enc_opts))
2403                 goto nla_put_failure;
2404
2405         if (fl_dump_key_ct(skb, &key->ct, &mask->ct))
2406                 goto nla_put_failure;
2407
2408         if (fl_dump_key_flags(skb, key->control.flags, mask->control.flags))
2409                 goto nla_put_failure;
2410
2411         return 0;
2412
2413 nla_put_failure:
2414         return -EMSGSIZE;
2415 }
2416
2417 static int fl_dump(struct net *net, struct tcf_proto *tp, void *fh,
2418                    struct sk_buff *skb, struct tcmsg *t, bool rtnl_held)
2419 {
2420         struct cls_fl_filter *f = fh;
2421         struct nlattr *nest;
2422         struct fl_flow_key *key, *mask;
2423         bool skip_hw;
2424
2425         if (!f)
2426                 return skb->len;
2427
2428         t->tcm_handle = f->handle;
2429
2430         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2431         if (!nest)
2432                 goto nla_put_failure;
2433
2434         spin_lock(&tp->lock);
2435
2436         if (f->res.classid &&
2437             nla_put_u32(skb, TCA_FLOWER_CLASSID, f->res.classid))
2438                 goto nla_put_failure_locked;
2439
2440         key = &f->key;
2441         mask = &f->mask->key;
2442         skip_hw = tc_skip_hw(f->flags);
2443
2444         if (fl_dump_key(skb, net, key, mask))
2445                 goto nla_put_failure_locked;
2446
2447         if (f->flags && nla_put_u32(skb, TCA_FLOWER_FLAGS, f->flags))
2448                 goto nla_put_failure_locked;
2449
2450         spin_unlock(&tp->lock);
2451
2452         if (!skip_hw)
2453                 fl_hw_update_stats(tp, f, rtnl_held);
2454
2455         if (nla_put_u32(skb, TCA_FLOWER_IN_HW_COUNT, f->in_hw_count))
2456                 goto nla_put_failure;
2457
2458         if (tcf_exts_dump(skb, &f->exts))
2459                 goto nla_put_failure;
2460
2461         nla_nest_end(skb, nest);
2462
2463         if (tcf_exts_dump_stats(skb, &f->exts) < 0)
2464                 goto nla_put_failure;
2465
2466         return skb->len;
2467
2468 nla_put_failure_locked:
2469         spin_unlock(&tp->lock);
2470 nla_put_failure:
2471         nla_nest_cancel(skb, nest);
2472         return -1;
2473 }
2474
2475 static int fl_tmplt_dump(struct sk_buff *skb, struct net *net, void *tmplt_priv)
2476 {
2477         struct fl_flow_tmplt *tmplt = tmplt_priv;
2478         struct fl_flow_key *key, *mask;
2479         struct nlattr *nest;
2480
2481         nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
2482         if (!nest)
2483                 goto nla_put_failure;
2484
2485         key = &tmplt->dummy_key;
2486         mask = &tmplt->mask;
2487
2488         if (fl_dump_key(skb, net, key, mask))
2489                 goto nla_put_failure;
2490
2491         nla_nest_end(skb, nest);
2492
2493         return skb->len;
2494
2495 nla_put_failure:
2496         nla_nest_cancel(skb, nest);
2497         return -EMSGSIZE;
2498 }
2499
2500 static void fl_bind_class(void *fh, u32 classid, unsigned long cl)
2501 {
2502         struct cls_fl_filter *f = fh;
2503
2504         if (f && f->res.classid == classid)
2505                 f->res.class = cl;
2506 }
2507
2508 static struct tcf_proto_ops cls_fl_ops __read_mostly = {
2509         .kind           = "flower",
2510         .classify       = fl_classify,
2511         .init           = fl_init,
2512         .destroy        = fl_destroy,
2513         .get            = fl_get,
2514         .put            = fl_put,
2515         .change         = fl_change,
2516         .delete         = fl_delete,
2517         .walk           = fl_walk,
2518         .reoffload      = fl_reoffload,
2519         .hw_add         = fl_hw_add,
2520         .hw_del         = fl_hw_del,
2521         .dump           = fl_dump,
2522         .bind_class     = fl_bind_class,
2523         .tmplt_create   = fl_tmplt_create,
2524         .tmplt_destroy  = fl_tmplt_destroy,
2525         .tmplt_dump     = fl_tmplt_dump,
2526         .owner          = THIS_MODULE,
2527         .flags          = TCF_PROTO_OPS_DOIT_UNLOCKED,
2528 };
2529
2530 static int __init cls_fl_init(void)
2531 {
2532         return register_tcf_proto_ops(&cls_fl_ops);
2533 }
2534
2535 static void __exit cls_fl_exit(void)
2536 {
2537         unregister_tcf_proto_ops(&cls_fl_ops);
2538 }
2539
2540 module_init(cls_fl_init);
2541 module_exit(cls_fl_exit);
2542
2543 MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>");
2544 MODULE_DESCRIPTION("Flower classifier");
2545 MODULE_LICENSE("GPL v2");