xref: /openbmc/linux/net/ipv6/ila/ila_xlat.c (revision 83268fa6) 
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/jhash.h>
3 #include <linux/netfilter.h>
4 #include <linux/rcupdate.h>
5 #include <linux/rhashtable.h>
6 #include <linux/vmalloc.h>
7 #include <net/genetlink.h>
8 #include <net/ila.h>
9 #include <net/netns/generic.h>
10 #include <uapi/linux/genetlink.h>
11 #include "ila.h"
12 
13 struct ila_xlat_params {
14 	struct ila_params ip;
15 	int ifindex;
16 };
17 
18 struct ila_map {
19 	struct ila_xlat_params xp;
20 	struct rhash_head node;
21 	struct ila_map __rcu *next;
22 	struct rcu_head rcu;
23 };
24 
25 #define MAX_LOCKS 1024
26 #define	LOCKS_PER_CPU 10
27 
28 static int alloc_ila_locks(struct ila_net *ilan)
29 {
30 	return alloc_bucket_spinlocks(&ilan->xlat.locks, &ilan->xlat.locks_mask,
31 				      MAX_LOCKS, LOCKS_PER_CPU,
32 				      GFP_KERNEL);
33 }
34 
35 static u32 hashrnd __read_mostly;
36 static __always_inline void __ila_hash_secret_init(void)
37 {
38 	net_get_random_once(&hashrnd, sizeof(hashrnd));
39 }
40 
41 static inline u32 ila_locator_hash(struct ila_locator loc)
42 {
43 	u32 *v = (u32 *)loc.v32;
44 
45 	__ila_hash_secret_init();
46 	return jhash_2words(v[0], v[1], hashrnd);
47 }
48 
49 static inline spinlock_t *ila_get_lock(struct ila_net *ilan,
50 				       struct ila_locator loc)
51 {
52 	return &ilan->xlat.locks[ila_locator_hash(loc) & ilan->xlat.locks_mask];
53 }
54 
55 static inline int ila_cmp_wildcards(struct ila_map *ila,
56 				    struct ila_addr *iaddr, int ifindex)
57 {
58 	return (ila->xp.ifindex && ila->xp.ifindex != ifindex);
59 }
60 
61 static inline int ila_cmp_params(struct ila_map *ila,
62 				 struct ila_xlat_params *xp)
63 {
64 	return (ila->xp.ifindex != xp->ifindex);
65 }
66 
67 static int ila_cmpfn(struct rhashtable_compare_arg *arg,
68 		     const void *obj)
69 {
70 	const struct ila_map *ila = obj;
71 
72 	return (ila->xp.ip.locator_match.v64 != *(__be64 *)arg->key);
73 }
74 
75 static inline int ila_order(struct ila_map *ila)
76 {
77 	int score = 0;
78 
79 	if (ila->xp.ifindex)
80 		score += 1 << 1;
81 
82 	return score;
83 }
84 
85 static const struct rhashtable_params rht_params = {
86 	.nelem_hint = 1024,
87 	.head_offset = offsetof(struct ila_map, node),
88 	.key_offset = offsetof(struct ila_map, xp.ip.locator_match),
89 	.key_len = sizeof(u64), /* identifier */
90 	.max_size = 1048576,
91 	.min_size = 256,
92 	.automatic_shrinking = true,
93 	.obj_cmpfn = ila_cmpfn,
94 };
95 
96 static int parse_nl_config(struct genl_info *info,
97 			   struct ila_xlat_params *xp)
98 {
99 	memset(xp, 0, sizeof(*xp));
100 
101 	if (info->attrs[ILA_ATTR_LOCATOR])
102 		xp->ip.locator.v64 = (__force __be64)nla_get_u64(
103 			info->attrs[ILA_ATTR_LOCATOR]);
104 
105 	if (info->attrs[ILA_ATTR_LOCATOR_MATCH])
106 		xp->ip.locator_match.v64 = (__force __be64)nla_get_u64(
107 			info->attrs[ILA_ATTR_LOCATOR_MATCH]);
108 
109 	if (info->attrs[ILA_ATTR_CSUM_MODE])
110 		xp->ip.csum_mode = nla_get_u8(info->attrs[ILA_ATTR_CSUM_MODE]);
111 	else
112 		xp->ip.csum_mode = ILA_CSUM_NO_ACTION;
113 
114 	if (info->attrs[ILA_ATTR_IDENT_TYPE])
115 		xp->ip.ident_type = nla_get_u8(
116 				info->attrs[ILA_ATTR_IDENT_TYPE]);
117 	else
118 		xp->ip.ident_type = ILA_ATYPE_USE_FORMAT;
119 
120 	if (info->attrs[ILA_ATTR_IFINDEX])
121 		xp->ifindex = nla_get_s32(info->attrs[ILA_ATTR_IFINDEX]);
122 
123 	return 0;
124 }
125 
126 /* Must be called with rcu readlock */
127 static inline struct ila_map *ila_lookup_wildcards(struct ila_addr *iaddr,
128 						   int ifindex,
129 						   struct ila_net *ilan)
130 {
131 	struct ila_map *ila;
132 
133 	ila = rhashtable_lookup_fast(&ilan->xlat.rhash_table, &iaddr->loc,
134 				     rht_params);
135 	while (ila) {
136 		if (!ila_cmp_wildcards(ila, iaddr, ifindex))
137 			return ila;
138 		ila = rcu_access_pointer(ila->next);
139 	}
140 
141 	return NULL;
142 }
143 
144 /* Must be called with rcu readlock */
145 static inline struct ila_map *ila_lookup_by_params(struct ila_xlat_params *xp,
146 						   struct ila_net *ilan)
147 {
148 	struct ila_map *ila;
149 
150 	ila = rhashtable_lookup_fast(&ilan->xlat.rhash_table,
151 				     &xp->ip.locator_match,
152 				     rht_params);
153 	while (ila) {
154 		if (!ila_cmp_params(ila, xp))
155 			return ila;
156 		ila = rcu_access_pointer(ila->next);
157 	}
158 
159 	return NULL;
160 }
161 
162 static inline void ila_release(struct ila_map *ila)
163 {
164 	kfree_rcu(ila, rcu);
165 }
166 
167 static void ila_free_node(struct ila_map *ila)
168 {
169 	struct ila_map *next;
170 
171 	/* Assume rcu_readlock held */
172 	while (ila) {
173 		next = rcu_access_pointer(ila->next);
174 		ila_release(ila);
175 		ila = next;
176 	}
177 }
178 
179 static void ila_free_cb(void *ptr, void *arg)
180 {
181 	ila_free_node((struct ila_map *)ptr);
182 }
183 
184 static int ila_xlat_addr(struct sk_buff *skb, bool sir2ila);
185 
186 static unsigned int
187 ila_nf_input(void *priv,
188 	     struct sk_buff *skb,
189 	     const struct nf_hook_state *state)
190 {
191 	ila_xlat_addr(skb, false);
192 	return NF_ACCEPT;
193 }
194 
195 static const struct nf_hook_ops ila_nf_hook_ops[] = {
196 	{
197 		.hook = ila_nf_input,
198 		.pf = NFPROTO_IPV6,
199 		.hooknum = NF_INET_PRE_ROUTING,
200 		.priority = -1,
201 	},
202 };
203 
204 static int ila_add_mapping(struct net *net, struct ila_xlat_params *xp)
205 {
206 	struct ila_net *ilan = net_generic(net, ila_net_id);
207 	struct ila_map *ila, *head;
208 	spinlock_t *lock = ila_get_lock(ilan, xp->ip.locator_match);
209 	int err = 0, order;
210 
211 	if (!ilan->xlat.hooks_registered) {
212 		/* We defer registering net hooks in the namespace until the
213 		 * first mapping is added.
214 		 */
215 		err = nf_register_net_hooks(net, ila_nf_hook_ops,
216 					    ARRAY_SIZE(ila_nf_hook_ops));
217 		if (err)
218 			return err;
219 
220 		ilan->xlat.hooks_registered = true;
221 	}
222 
223 	ila = kzalloc(sizeof(*ila), GFP_KERNEL);
224 	if (!ila)
225 		return -ENOMEM;
226 
227 	ila_init_saved_csum(&xp->ip);
228 
229 	ila->xp = *xp;
230 
231 	order = ila_order(ila);
232 
233 	spin_lock(lock);
234 
235 	head = rhashtable_lookup_fast(&ilan->xlat.rhash_table,
236 				      &xp->ip.locator_match,
237 				      rht_params);
238 	if (!head) {
239 		/* New entry for the rhash_table */
240 		err = rhashtable_lookup_insert_fast(&ilan->xlat.rhash_table,
241 						    &ila->node, rht_params);
242 	} else {
243 		struct ila_map *tila = head, *prev = NULL;
244 
245 		do {
246 			if (!ila_cmp_params(tila, xp)) {
247 				err = -EEXIST;
248 				goto out;
249 			}
250 
251 			if (order > ila_order(tila))
252 				break;
253 
254 			prev = tila;
255 			tila = rcu_dereference_protected(tila->next,
256 				lockdep_is_held(lock));
257 		} while (tila);
258 
259 		if (prev) {
260 			/* Insert in sub list of head */
261 			RCU_INIT_POINTER(ila->next, tila);
262 			rcu_assign_pointer(prev->next, ila);
263 		} else {
264 			/* Make this ila new head */
265 			RCU_INIT_POINTER(ila->next, head);
266 			err = rhashtable_replace_fast(&ilan->xlat.rhash_table,
267 						      &head->node,
268 						      &ila->node, rht_params);
269 			if (err)
270 				goto out;
271 		}
272 	}
273 
274 out:
275 	spin_unlock(lock);
276 
277 	if (err)
278 		kfree(ila);
279 
280 	return err;
281 }
282 
283 static int ila_del_mapping(struct net *net, struct ila_xlat_params *xp)
284 {
285 	struct ila_net *ilan = net_generic(net, ila_net_id);
286 	struct ila_map *ila, *head, *prev;
287 	spinlock_t *lock = ila_get_lock(ilan, xp->ip.locator_match);
288 	int err = -ENOENT;
289 
290 	spin_lock(lock);
291 
292 	head = rhashtable_lookup_fast(&ilan->xlat.rhash_table,
293 				      &xp->ip.locator_match, rht_params);
294 	ila = head;
295 
296 	prev = NULL;
297 
298 	while (ila) {
299 		if (ila_cmp_params(ila, xp)) {
300 			prev = ila;
301 			ila = rcu_dereference_protected(ila->next,
302 							lockdep_is_held(lock));
303 			continue;
304 		}
305 
306 		err = 0;
307 
308 		if (prev) {
309 			/* Not head, just delete from list */
310 			rcu_assign_pointer(prev->next, ila->next);
311 		} else {
312 			/* It is the head. If there is something in the
313 			 * sublist we need to make a new head.
314 			 */
315 			head = rcu_dereference_protected(ila->next,
316 							 lockdep_is_held(lock));
317 			if (head) {
318 				/* Put first entry in the sublist into the
319 				 * table
320 				 */
321 				err = rhashtable_replace_fast(
322 					&ilan->xlat.rhash_table, &ila->node,
323 					&head->node, rht_params);
324 				if (err)
325 					goto out;
326 			} else {
327 				/* Entry no longer used */
328 				err = rhashtable_remove_fast(
329 						&ilan->xlat.rhash_table,
330 						&ila->node, rht_params);
331 			}
332 		}
333 
334 		ila_release(ila);
335 
336 		break;
337 	}
338 
339 out:
340 	spin_unlock(lock);
341 
342 	return err;
343 }
344 
345 int ila_xlat_nl_cmd_add_mapping(struct sk_buff *skb, struct genl_info *info)
346 {
347 	struct net *net = genl_info_net(info);
348 	struct ila_xlat_params p;
349 	int err;
350 
351 	err = parse_nl_config(info, &p);
352 	if (err)
353 		return err;
354 
355 	return ila_add_mapping(net, &p);
356 }
357 
358 int ila_xlat_nl_cmd_del_mapping(struct sk_buff *skb, struct genl_info *info)
359 {
360 	struct net *net = genl_info_net(info);
361 	struct ila_xlat_params xp;
362 	int err;
363 
364 	err = parse_nl_config(info, &xp);
365 	if (err)
366 		return err;
367 
368 	ila_del_mapping(net, &xp);
369 
370 	return 0;
371 }
372 
373 static inline spinlock_t *lock_from_ila_map(struct ila_net *ilan,
374 					    struct ila_map *ila)
375 {
376 	return ila_get_lock(ilan, ila->xp.ip.locator_match);
377 }
378 
379 int ila_xlat_nl_cmd_flush(struct sk_buff *skb, struct genl_info *info)
380 {
381 	struct net *net = genl_info_net(info);
382 	struct ila_net *ilan = net_generic(net, ila_net_id);
383 	struct rhashtable_iter iter;
384 	struct ila_map *ila;
385 	spinlock_t *lock;
386 	int ret;
387 
388 	ret = rhashtable_walk_init(&ilan->xlat.rhash_table, &iter, GFP_KERNEL);
389 	if (ret)
390 		goto done;
391 
392 	rhashtable_walk_start(&iter);
393 
394 	for (;;) {
395 		ila = rhashtable_walk_next(&iter);
396 
397 		if (IS_ERR(ila)) {
398 			if (PTR_ERR(ila) == -EAGAIN)
399 				continue;
400 			ret = PTR_ERR(ila);
401 			goto done;
402 		} else if (!ila) {
403 			break;
404 		}
405 
406 		lock = lock_from_ila_map(ilan, ila);
407 
408 		spin_lock(lock);
409 
410 		ret = rhashtable_remove_fast(&ilan->xlat.rhash_table,
411 					     &ila->node, rht_params);
412 		if (!ret)
413 			ila_free_node(ila);
414 
415 		spin_unlock(lock);
416 
417 		if (ret)
418 			break;
419 	}
420 
421 done:
422 	rhashtable_walk_stop(&iter);
423 	return ret;
424 }
425 
426 static int ila_fill_info(struct ila_map *ila, struct sk_buff *msg)
427 {
428 	if (nla_put_u64_64bit(msg, ILA_ATTR_LOCATOR,
429 			      (__force u64)ila->xp.ip.locator.v64,
430 			      ILA_ATTR_PAD) ||
431 	    nla_put_u64_64bit(msg, ILA_ATTR_LOCATOR_MATCH,
432 			      (__force u64)ila->xp.ip.locator_match.v64,
433 			      ILA_ATTR_PAD) ||
434 	    nla_put_s32(msg, ILA_ATTR_IFINDEX, ila->xp.ifindex) ||
435 	    nla_put_u8(msg, ILA_ATTR_CSUM_MODE, ila->xp.ip.csum_mode) ||
436 	    nla_put_u8(msg, ILA_ATTR_IDENT_TYPE, ila->xp.ip.ident_type))
437 		return -1;
438 
439 	return 0;
440 }
441 
442 static int ila_dump_info(struct ila_map *ila,
443 			 u32 portid, u32 seq, u32 flags,
444 			 struct sk_buff *skb, u8 cmd)
445 {
446 	void *hdr;
447 
448 	hdr = genlmsg_put(skb, portid, seq, &ila_nl_family, flags, cmd);
449 	if (!hdr)
450 		return -ENOMEM;
451 
452 	if (ila_fill_info(ila, skb) < 0)
453 		goto nla_put_failure;
454 
455 	genlmsg_end(skb, hdr);
456 	return 0;
457 
458 nla_put_failure:
459 	genlmsg_cancel(skb, hdr);
460 	return -EMSGSIZE;
461 }
462 
463 int ila_xlat_nl_cmd_get_mapping(struct sk_buff *skb, struct genl_info *info)
464 {
465 	struct net *net = genl_info_net(info);
466 	struct ila_net *ilan = net_generic(net, ila_net_id);
467 	struct sk_buff *msg;
468 	struct ila_xlat_params xp;
469 	struct ila_map *ila;
470 	int ret;
471 
472 	ret = parse_nl_config(info, &xp);
473 	if (ret)
474 		return ret;
475 
476 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
477 	if (!msg)
478 		return -ENOMEM;
479 
480 	rcu_read_lock();
481 
482 	ila = ila_lookup_by_params(&xp, ilan);
483 	if (ila) {
484 		ret = ila_dump_info(ila,
485 				    info->snd_portid,
486 				    info->snd_seq, 0, msg,
487 				    info->genlhdr->cmd);
488 	}
489 
490 	rcu_read_unlock();
491 
492 	if (ret < 0)
493 		goto out_free;
494 
495 	return genlmsg_reply(msg, info);
496 
497 out_free:
498 	nlmsg_free(msg);
499 	return ret;
500 }
501 
502 struct ila_dump_iter {
503 	struct rhashtable_iter rhiter;
504 	int skip;
505 };
506 
507 int ila_xlat_nl_dump_start(struct netlink_callback *cb)
508 {
509 	struct net *net = sock_net(cb->skb->sk);
510 	struct ila_net *ilan = net_generic(net, ila_net_id);
511 	struct ila_dump_iter *iter;
512 	int ret;
513 
514 	iter = kmalloc(sizeof(*iter), GFP_KERNEL);
515 	if (!iter)
516 		return -ENOMEM;
517 
518 	ret = rhashtable_walk_init(&ilan->xlat.rhash_table, &iter->rhiter,
519 				   GFP_KERNEL);
520 	if (ret) {
521 		kfree(iter);
522 		return ret;
523 	}
524 
525 	iter->skip = 0;
526 	cb->args[0] = (long)iter;
527 
528 	return ret;
529 }
530 
531 int ila_xlat_nl_dump_done(struct netlink_callback *cb)
532 {
533 	struct ila_dump_iter *iter = (struct ila_dump_iter *)cb->args[0];
534 
535 	rhashtable_walk_exit(&iter->rhiter);
536 
537 	kfree(iter);
538 
539 	return 0;
540 }
541 
542 int ila_xlat_nl_dump(struct sk_buff *skb, struct netlink_callback *cb)
543 {
544 	struct ila_dump_iter *iter = (struct ila_dump_iter *)cb->args[0];
545 	struct rhashtable_iter *rhiter = &iter->rhiter;
546 	int skip = iter->skip;
547 	struct ila_map *ila;
548 	int ret;
549 
550 	rhashtable_walk_start(rhiter);
551 
552 	/* Get first entry */
553 	ila = rhashtable_walk_peek(rhiter);
554 
555 	if (ila && !IS_ERR(ila) && skip) {
556 		/* Skip over visited entries */
557 
558 		while (ila && skip) {
559 			/* Skip over any ila entries in this list that we
560 			 * have already dumped.
561 			 */
562 			ila = rcu_access_pointer(ila->next);
563 			skip--;
564 		}
565 	}
566 
567 	skip = 0;
568 
569 	for (;;) {
570 		if (IS_ERR(ila)) {
571 			ret = PTR_ERR(ila);
572 			if (ret == -EAGAIN) {
573 				/* Table has changed and iter has reset. Return
574 				 * -EAGAIN to the application even if we have
575 				 * written data to the skb. The application
576 				 * needs to deal with this.
577 				 */
578 
579 				goto out_ret;
580 			} else {
581 				break;
582 			}
583 		} else if (!ila) {
584 			ret = 0;
585 			break;
586 		}
587 
588 		while (ila) {
589 			ret =  ila_dump_info(ila, NETLINK_CB(cb->skb).portid,
590 					     cb->nlh->nlmsg_seq, NLM_F_MULTI,
591 					     skb, ILA_CMD_GET);
592 			if (ret)
593 				goto out;
594 
595 			skip++;
596 			ila = rcu_access_pointer(ila->next);
597 		}
598 
599 		skip = 0;
600 		ila = rhashtable_walk_next(rhiter);
601 	}
602 
603 out:
604 	iter->skip = skip;
605 	ret = (skb->len ? : ret);
606 
607 out_ret:
608 	rhashtable_walk_stop(rhiter);
609 	return ret;
610 }
611 
612 #define ILA_HASH_TABLE_SIZE 1024
613 
614 int ila_xlat_init_net(struct net *net)
615 {
616 	struct ila_net *ilan = net_generic(net, ila_net_id);
617 	int err;
618 
619 	err = alloc_ila_locks(ilan);
620 	if (err)
621 		return err;
622 
623 	rhashtable_init(&ilan->xlat.rhash_table, &rht_params);
624 
625 	return 0;
626 }
627 
628 void ila_xlat_exit_net(struct net *net)
629 {
630 	struct ila_net *ilan = net_generic(net, ila_net_id);
631 
632 	rhashtable_free_and_destroy(&ilan->xlat.rhash_table, ila_free_cb, NULL);
633 
634 	free_bucket_spinlocks(ilan->xlat.locks);
635 
636 	if (ilan->xlat.hooks_registered)
637 		nf_unregister_net_hooks(net, ila_nf_hook_ops,
638 					ARRAY_SIZE(ila_nf_hook_ops));
639 }
640 
641 static int ila_xlat_addr(struct sk_buff *skb, bool sir2ila)
642 {
643 	struct ila_map *ila;
644 	struct ipv6hdr *ip6h = ipv6_hdr(skb);
645 	struct net *net = dev_net(skb->dev);
646 	struct ila_net *ilan = net_generic(net, ila_net_id);
647 	struct ila_addr *iaddr = ila_a2i(&ip6h->daddr);
648 
649 	/* Assumes skb contains a valid IPv6 header that is pulled */
650 
651 	/* No check here that ILA type in the mapping matches what is in the
652 	 * address. We assume that whatever sender gaves us can be translated.
653 	 * The checksum mode however is relevant.
654 	 */
655 
656 	rcu_read_lock();
657 
658 	ila = ila_lookup_wildcards(iaddr, skb->dev->ifindex, ilan);
659 	if (ila)
660 		ila_update_ipv6_locator(skb, &ila->xp.ip, sir2ila);
661 
662 	rcu_read_unlock();
663 
664 	return 0;
665 }
666