Loading...
1/*
2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005 Intel Corporation. All rights reserved.
6 *
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
12 *
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
15 * conditions are met:
16 *
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer.
20 *
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
25 *
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 * SOFTWARE.
34 */
35
36#include <linux/mutex.h>
37#include <linux/inetdevice.h>
38#include <linux/slab.h>
39#include <linux/workqueue.h>
40#include <net/arp.h>
41#include <net/neighbour.h>
42#include <net/route.h>
43#include <net/netevent.h>
44#include <net/ipv6_stubs.h>
45#include <net/ip6_route.h>
46#include <rdma/ib_addr.h>
47#include <rdma/ib_cache.h>
48#include <rdma/ib_sa.h>
49#include <rdma/ib.h>
50#include <rdma/rdma_netlink.h>
51#include <net/netlink.h>
52
53#include "core_priv.h"
54
55struct addr_req {
56 struct list_head list;
57 struct sockaddr_storage src_addr;
58 struct sockaddr_storage dst_addr;
59 struct rdma_dev_addr *addr;
60 void *context;
61 void (*callback)(int status, struct sockaddr *src_addr,
62 struct rdma_dev_addr *addr, void *context);
63 unsigned long timeout;
64 struct delayed_work work;
65 bool resolve_by_gid_attr; /* Consider gid attr in resolve phase */
66 int status;
67 u32 seq;
68};
69
70static atomic_t ib_nl_addr_request_seq = ATOMIC_INIT(0);
71
72static DEFINE_SPINLOCK(lock);
73static LIST_HEAD(req_list);
74static struct workqueue_struct *addr_wq;
75
76static const struct nla_policy ib_nl_addr_policy[LS_NLA_TYPE_MAX] = {
77 [LS_NLA_TYPE_DGID] = {.type = NLA_BINARY,
78 .len = sizeof(struct rdma_nla_ls_gid),
79 .validation_type = NLA_VALIDATE_MIN,
80 .min = sizeof(struct rdma_nla_ls_gid)},
81};
82
83static inline bool ib_nl_is_good_ip_resp(const struct nlmsghdr *nlh)
84{
85 struct nlattr *tb[LS_NLA_TYPE_MAX] = {};
86 int ret;
87
88 if (nlh->nlmsg_flags & RDMA_NL_LS_F_ERR)
89 return false;
90
91 ret = nla_parse_deprecated(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
92 nlmsg_len(nlh), ib_nl_addr_policy, NULL);
93 if (ret)
94 return false;
95
96 return true;
97}
98
99static void ib_nl_process_good_ip_rsep(const struct nlmsghdr *nlh)
100{
101 const struct nlattr *head, *curr;
102 union ib_gid gid;
103 struct addr_req *req;
104 int len, rem;
105 int found = 0;
106
107 head = (const struct nlattr *)nlmsg_data(nlh);
108 len = nlmsg_len(nlh);
109
110 nla_for_each_attr(curr, head, len, rem) {
111 if (curr->nla_type == LS_NLA_TYPE_DGID)
112 memcpy(&gid, nla_data(curr), nla_len(curr));
113 }
114
115 spin_lock_bh(&lock);
116 list_for_each_entry(req, &req_list, list) {
117 if (nlh->nlmsg_seq != req->seq)
118 continue;
119 /* We set the DGID part, the rest was set earlier */
120 rdma_addr_set_dgid(req->addr, &gid);
121 req->status = 0;
122 found = 1;
123 break;
124 }
125 spin_unlock_bh(&lock);
126
127 if (!found)
128 pr_info("Couldn't find request waiting for DGID: %pI6\n",
129 &gid);
130}
131
132int ib_nl_handle_ip_res_resp(struct sk_buff *skb,
133 struct nlmsghdr *nlh,
134 struct netlink_ext_ack *extack)
135{
136 if ((nlh->nlmsg_flags & NLM_F_REQUEST) ||
137 !(NETLINK_CB(skb).sk))
138 return -EPERM;
139
140 if (ib_nl_is_good_ip_resp(nlh))
141 ib_nl_process_good_ip_rsep(nlh);
142
143 return 0;
144}
145
146static int ib_nl_ip_send_msg(struct rdma_dev_addr *dev_addr,
147 const void *daddr,
148 u32 seq, u16 family)
149{
150 struct sk_buff *skb = NULL;
151 struct nlmsghdr *nlh;
152 struct rdma_ls_ip_resolve_header *header;
153 void *data;
154 size_t size;
155 int attrtype;
156 int len;
157
158 if (family == AF_INET) {
159 size = sizeof(struct in_addr);
160 attrtype = RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_IPV4;
161 } else {
162 size = sizeof(struct in6_addr);
163 attrtype = RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_IPV6;
164 }
165
166 len = nla_total_size(sizeof(size));
167 len += NLMSG_ALIGN(sizeof(*header));
168
169 skb = nlmsg_new(len, GFP_KERNEL);
170 if (!skb)
171 return -ENOMEM;
172
173 data = ibnl_put_msg(skb, &nlh, seq, 0, RDMA_NL_LS,
174 RDMA_NL_LS_OP_IP_RESOLVE, NLM_F_REQUEST);
175 if (!data) {
176 nlmsg_free(skb);
177 return -ENODATA;
178 }
179
180 /* Construct the family header first */
181 header = skb_put(skb, NLMSG_ALIGN(sizeof(*header)));
182 header->ifindex = dev_addr->bound_dev_if;
183 nla_put(skb, attrtype, size, daddr);
184
185 /* Repair the nlmsg header length */
186 nlmsg_end(skb, nlh);
187 rdma_nl_multicast(&init_net, skb, RDMA_NL_GROUP_LS, GFP_KERNEL);
188
189 /* Make the request retry, so when we get the response from userspace
190 * we will have something.
191 */
192 return -ENODATA;
193}
194
195int rdma_addr_size(const struct sockaddr *addr)
196{
197 switch (addr->sa_family) {
198 case AF_INET:
199 return sizeof(struct sockaddr_in);
200 case AF_INET6:
201 return sizeof(struct sockaddr_in6);
202 case AF_IB:
203 return sizeof(struct sockaddr_ib);
204 default:
205 return 0;
206 }
207}
208EXPORT_SYMBOL(rdma_addr_size);
209
210int rdma_addr_size_in6(struct sockaddr_in6 *addr)
211{
212 int ret = rdma_addr_size((struct sockaddr *) addr);
213
214 return ret <= sizeof(*addr) ? ret : 0;
215}
216EXPORT_SYMBOL(rdma_addr_size_in6);
217
218int rdma_addr_size_kss(struct __kernel_sockaddr_storage *addr)
219{
220 int ret = rdma_addr_size((struct sockaddr *) addr);
221
222 return ret <= sizeof(*addr) ? ret : 0;
223}
224EXPORT_SYMBOL(rdma_addr_size_kss);
225
226/**
227 * rdma_copy_src_l2_addr - Copy netdevice source addresses
228 * @dev_addr: Destination address pointer where to copy the addresses
229 * @dev: Netdevice whose source addresses to copy
230 *
231 * rdma_copy_src_l2_addr() copies source addresses from the specified netdevice.
232 * This includes unicast address, broadcast address, device type and
233 * interface index.
234 */
235void rdma_copy_src_l2_addr(struct rdma_dev_addr *dev_addr,
236 const struct net_device *dev)
237{
238 dev_addr->dev_type = dev->type;
239 memcpy(dev_addr->src_dev_addr, dev->dev_addr, MAX_ADDR_LEN);
240 memcpy(dev_addr->broadcast, dev->broadcast, MAX_ADDR_LEN);
241 dev_addr->bound_dev_if = dev->ifindex;
242}
243EXPORT_SYMBOL(rdma_copy_src_l2_addr);
244
245static struct net_device *
246rdma_find_ndev_for_src_ip_rcu(struct net *net, const struct sockaddr *src_in)
247{
248 struct net_device *dev = NULL;
249 int ret = -EADDRNOTAVAIL;
250
251 switch (src_in->sa_family) {
252 case AF_INET:
253 dev = __ip_dev_find(net,
254 ((const struct sockaddr_in *)src_in)->sin_addr.s_addr,
255 false);
256 if (dev)
257 ret = 0;
258 break;
259#if IS_ENABLED(CONFIG_IPV6)
260 case AF_INET6:
261 for_each_netdev_rcu(net, dev) {
262 if (ipv6_chk_addr(net,
263 &((const struct sockaddr_in6 *)src_in)->sin6_addr,
264 dev, 1)) {
265 ret = 0;
266 break;
267 }
268 }
269 break;
270#endif
271 }
272 return ret ? ERR_PTR(ret) : dev;
273}
274
275int rdma_translate_ip(const struct sockaddr *addr,
276 struct rdma_dev_addr *dev_addr)
277{
278 struct net_device *dev;
279
280 if (dev_addr->bound_dev_if) {
281 dev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
282 if (!dev)
283 return -ENODEV;
284 rdma_copy_src_l2_addr(dev_addr, dev);
285 dev_put(dev);
286 return 0;
287 }
288
289 rcu_read_lock();
290 dev = rdma_find_ndev_for_src_ip_rcu(dev_addr->net, addr);
291 if (!IS_ERR(dev))
292 rdma_copy_src_l2_addr(dev_addr, dev);
293 rcu_read_unlock();
294 return PTR_ERR_OR_ZERO(dev);
295}
296EXPORT_SYMBOL(rdma_translate_ip);
297
298static void set_timeout(struct addr_req *req, unsigned long time)
299{
300 unsigned long delay;
301
302 delay = time - jiffies;
303 if ((long)delay < 0)
304 delay = 0;
305
306 mod_delayed_work(addr_wq, &req->work, delay);
307}
308
309static void queue_req(struct addr_req *req)
310{
311 spin_lock_bh(&lock);
312 list_add_tail(&req->list, &req_list);
313 set_timeout(req, req->timeout);
314 spin_unlock_bh(&lock);
315}
316
317static int ib_nl_fetch_ha(struct rdma_dev_addr *dev_addr,
318 const void *daddr, u32 seq, u16 family)
319{
320 if (!rdma_nl_chk_listeners(RDMA_NL_GROUP_LS))
321 return -EADDRNOTAVAIL;
322
323 return ib_nl_ip_send_msg(dev_addr, daddr, seq, family);
324}
325
326static int dst_fetch_ha(const struct dst_entry *dst,
327 struct rdma_dev_addr *dev_addr,
328 const void *daddr)
329{
330 struct neighbour *n;
331 int ret = 0;
332
333 n = dst_neigh_lookup(dst, daddr);
334 if (!n)
335 return -ENODATA;
336
337 if (!(n->nud_state & NUD_VALID)) {
338 neigh_event_send(n, NULL);
339 ret = -ENODATA;
340 } else {
341 neigh_ha_snapshot(dev_addr->dst_dev_addr, n, dst->dev);
342 }
343
344 neigh_release(n);
345
346 return ret;
347}
348
349static bool has_gateway(const struct dst_entry *dst, sa_family_t family)
350{
351 struct rtable *rt;
352 struct rt6_info *rt6;
353
354 if (family == AF_INET) {
355 rt = container_of(dst, struct rtable, dst);
356 return rt->rt_uses_gateway;
357 }
358
359 rt6 = container_of(dst, struct rt6_info, dst);
360 return rt6->rt6i_flags & RTF_GATEWAY;
361}
362
363static int fetch_ha(const struct dst_entry *dst, struct rdma_dev_addr *dev_addr,
364 const struct sockaddr *dst_in, u32 seq)
365{
366 const struct sockaddr_in *dst_in4 =
367 (const struct sockaddr_in *)dst_in;
368 const struct sockaddr_in6 *dst_in6 =
369 (const struct sockaddr_in6 *)dst_in;
370 const void *daddr = (dst_in->sa_family == AF_INET) ?
371 (const void *)&dst_in4->sin_addr.s_addr :
372 (const void *)&dst_in6->sin6_addr;
373 sa_family_t family = dst_in->sa_family;
374
375 might_sleep();
376
377 /* If we have a gateway in IB mode then it must be an IB network */
378 if (has_gateway(dst, family) && dev_addr->network == RDMA_NETWORK_IB)
379 return ib_nl_fetch_ha(dev_addr, daddr, seq, family);
380 else
381 return dst_fetch_ha(dst, dev_addr, daddr);
382}
383
384static int addr4_resolve(struct sockaddr *src_sock,
385 const struct sockaddr *dst_sock,
386 struct rdma_dev_addr *addr,
387 struct rtable **prt)
388{
389 struct sockaddr_in *src_in = (struct sockaddr_in *)src_sock;
390 const struct sockaddr_in *dst_in =
391 (const struct sockaddr_in *)dst_sock;
392
393 __be32 src_ip = src_in->sin_addr.s_addr;
394 __be32 dst_ip = dst_in->sin_addr.s_addr;
395 struct rtable *rt;
396 struct flowi4 fl4;
397 int ret;
398
399 memset(&fl4, 0, sizeof(fl4));
400 fl4.daddr = dst_ip;
401 fl4.saddr = src_ip;
402 fl4.flowi4_oif = addr->bound_dev_if;
403 rt = ip_route_output_key(addr->net, &fl4);
404 ret = PTR_ERR_OR_ZERO(rt);
405 if (ret)
406 return ret;
407
408 src_in->sin_addr.s_addr = fl4.saddr;
409
410 addr->hoplimit = ip4_dst_hoplimit(&rt->dst);
411
412 *prt = rt;
413 return 0;
414}
415
416#if IS_ENABLED(CONFIG_IPV6)
417static int addr6_resolve(struct sockaddr *src_sock,
418 const struct sockaddr *dst_sock,
419 struct rdma_dev_addr *addr,
420 struct dst_entry **pdst)
421{
422 struct sockaddr_in6 *src_in = (struct sockaddr_in6 *)src_sock;
423 const struct sockaddr_in6 *dst_in =
424 (const struct sockaddr_in6 *)dst_sock;
425 struct flowi6 fl6;
426 struct dst_entry *dst;
427
428 memset(&fl6, 0, sizeof fl6);
429 fl6.daddr = dst_in->sin6_addr;
430 fl6.saddr = src_in->sin6_addr;
431 fl6.flowi6_oif = addr->bound_dev_if;
432
433 dst = ipv6_stub->ipv6_dst_lookup_flow(addr->net, NULL, &fl6, NULL);
434 if (IS_ERR(dst))
435 return PTR_ERR(dst);
436
437 if (ipv6_addr_any(&src_in->sin6_addr))
438 src_in->sin6_addr = fl6.saddr;
439
440 addr->hoplimit = ip6_dst_hoplimit(dst);
441
442 *pdst = dst;
443 return 0;
444}
445#else
446static int addr6_resolve(struct sockaddr *src_sock,
447 const struct sockaddr *dst_sock,
448 struct rdma_dev_addr *addr,
449 struct dst_entry **pdst)
450{
451 return -EADDRNOTAVAIL;
452}
453#endif
454
455static int addr_resolve_neigh(const struct dst_entry *dst,
456 const struct sockaddr *dst_in,
457 struct rdma_dev_addr *addr,
458 unsigned int ndev_flags,
459 u32 seq)
460{
461 int ret = 0;
462
463 if (ndev_flags & IFF_LOOPBACK) {
464 memcpy(addr->dst_dev_addr, addr->src_dev_addr, MAX_ADDR_LEN);
465 } else {
466 if (!(ndev_flags & IFF_NOARP)) {
467 /* If the device doesn't do ARP internally */
468 ret = fetch_ha(dst, addr, dst_in, seq);
469 }
470 }
471 return ret;
472}
473
474static int copy_src_l2_addr(struct rdma_dev_addr *dev_addr,
475 const struct sockaddr *dst_in,
476 const struct dst_entry *dst,
477 const struct net_device *ndev)
478{
479 int ret = 0;
480
481 if (dst->dev->flags & IFF_LOOPBACK)
482 ret = rdma_translate_ip(dst_in, dev_addr);
483 else
484 rdma_copy_src_l2_addr(dev_addr, dst->dev);
485
486 /*
487 * If there's a gateway and type of device not ARPHRD_INFINIBAND,
488 * we're definitely in RoCE v2 (as RoCE v1 isn't routable) set the
489 * network type accordingly.
490 */
491 if (has_gateway(dst, dst_in->sa_family) &&
492 ndev->type != ARPHRD_INFINIBAND)
493 dev_addr->network = dst_in->sa_family == AF_INET ?
494 RDMA_NETWORK_IPV4 :
495 RDMA_NETWORK_IPV6;
496 else
497 dev_addr->network = RDMA_NETWORK_IB;
498
499 return ret;
500}
501
502static int rdma_set_src_addr_rcu(struct rdma_dev_addr *dev_addr,
503 unsigned int *ndev_flags,
504 const struct sockaddr *dst_in,
505 const struct dst_entry *dst)
506{
507 struct net_device *ndev = READ_ONCE(dst->dev);
508
509 *ndev_flags = ndev->flags;
510 /* A physical device must be the RDMA device to use */
511 if (ndev->flags & IFF_LOOPBACK) {
512 /*
513 * RDMA (IB/RoCE, iWarp) doesn't run on lo interface or
514 * loopback IP address. So if route is resolved to loopback
515 * interface, translate that to a real ndev based on non
516 * loopback IP address.
517 */
518 ndev = rdma_find_ndev_for_src_ip_rcu(dev_net(ndev), dst_in);
519 if (IS_ERR(ndev))
520 return -ENODEV;
521 }
522
523 return copy_src_l2_addr(dev_addr, dst_in, dst, ndev);
524}
525
526static int set_addr_netns_by_gid_rcu(struct rdma_dev_addr *addr)
527{
528 struct net_device *ndev;
529
530 ndev = rdma_read_gid_attr_ndev_rcu(addr->sgid_attr);
531 if (IS_ERR(ndev))
532 return PTR_ERR(ndev);
533
534 /*
535 * Since we are holding the rcu, reading net and ifindex
536 * are safe without any additional reference; because
537 * change_net_namespace() in net/core/dev.c does rcu sync
538 * after it changes the state to IFF_DOWN and before
539 * updating netdev fields {net, ifindex}.
540 */
541 addr->net = dev_net(ndev);
542 addr->bound_dev_if = ndev->ifindex;
543 return 0;
544}
545
546static void rdma_addr_set_net_defaults(struct rdma_dev_addr *addr)
547{
548 addr->net = &init_net;
549 addr->bound_dev_if = 0;
550}
551
552static int addr_resolve(struct sockaddr *src_in,
553 const struct sockaddr *dst_in,
554 struct rdma_dev_addr *addr,
555 bool resolve_neigh,
556 bool resolve_by_gid_attr,
557 u32 seq)
558{
559 struct dst_entry *dst = NULL;
560 unsigned int ndev_flags = 0;
561 struct rtable *rt = NULL;
562 int ret;
563
564 if (!addr->net) {
565 pr_warn_ratelimited("%s: missing namespace\n", __func__);
566 return -EINVAL;
567 }
568
569 rcu_read_lock();
570 if (resolve_by_gid_attr) {
571 if (!addr->sgid_attr) {
572 rcu_read_unlock();
573 pr_warn_ratelimited("%s: missing gid_attr\n", __func__);
574 return -EINVAL;
575 }
576 /*
577 * If the request is for a specific gid attribute of the
578 * rdma_dev_addr, derive net from the netdevice of the
579 * GID attribute.
580 */
581 ret = set_addr_netns_by_gid_rcu(addr);
582 if (ret) {
583 rcu_read_unlock();
584 return ret;
585 }
586 }
587 if (src_in->sa_family == AF_INET) {
588 ret = addr4_resolve(src_in, dst_in, addr, &rt);
589 dst = &rt->dst;
590 } else {
591 ret = addr6_resolve(src_in, dst_in, addr, &dst);
592 }
593 if (ret) {
594 rcu_read_unlock();
595 goto done;
596 }
597 ret = rdma_set_src_addr_rcu(addr, &ndev_flags, dst_in, dst);
598 rcu_read_unlock();
599
600 /*
601 * Resolve neighbor destination address if requested and
602 * only if src addr translation didn't fail.
603 */
604 if (!ret && resolve_neigh)
605 ret = addr_resolve_neigh(dst, dst_in, addr, ndev_flags, seq);
606
607 if (src_in->sa_family == AF_INET)
608 ip_rt_put(rt);
609 else
610 dst_release(dst);
611done:
612 /*
613 * Clear the addr net to go back to its original state, only if it was
614 * derived from GID attribute in this context.
615 */
616 if (resolve_by_gid_attr)
617 rdma_addr_set_net_defaults(addr);
618 return ret;
619}
620
621static void process_one_req(struct work_struct *_work)
622{
623 struct addr_req *req;
624 struct sockaddr *src_in, *dst_in;
625
626 req = container_of(_work, struct addr_req, work.work);
627
628 if (req->status == -ENODATA) {
629 src_in = (struct sockaddr *)&req->src_addr;
630 dst_in = (struct sockaddr *)&req->dst_addr;
631 req->status = addr_resolve(src_in, dst_in, req->addr,
632 true, req->resolve_by_gid_attr,
633 req->seq);
634 if (req->status && time_after_eq(jiffies, req->timeout)) {
635 req->status = -ETIMEDOUT;
636 } else if (req->status == -ENODATA) {
637 /* requeue the work for retrying again */
638 spin_lock_bh(&lock);
639 if (!list_empty(&req->list))
640 set_timeout(req, req->timeout);
641 spin_unlock_bh(&lock);
642 return;
643 }
644 }
645
646 req->callback(req->status, (struct sockaddr *)&req->src_addr,
647 req->addr, req->context);
648 req->callback = NULL;
649
650 spin_lock_bh(&lock);
651 /*
652 * Although the work will normally have been canceled by the workqueue,
653 * it can still be requeued as long as it is on the req_list.
654 */
655 cancel_delayed_work(&req->work);
656 if (!list_empty(&req->list)) {
657 list_del_init(&req->list);
658 kfree(req);
659 }
660 spin_unlock_bh(&lock);
661}
662
663int rdma_resolve_ip(struct sockaddr *src_addr, const struct sockaddr *dst_addr,
664 struct rdma_dev_addr *addr, unsigned long timeout_ms,
665 void (*callback)(int status, struct sockaddr *src_addr,
666 struct rdma_dev_addr *addr, void *context),
667 bool resolve_by_gid_attr, void *context)
668{
669 struct sockaddr *src_in, *dst_in;
670 struct addr_req *req;
671 int ret = 0;
672
673 req = kzalloc(sizeof *req, GFP_KERNEL);
674 if (!req)
675 return -ENOMEM;
676
677 src_in = (struct sockaddr *) &req->src_addr;
678 dst_in = (struct sockaddr *) &req->dst_addr;
679
680 if (src_addr) {
681 if (src_addr->sa_family != dst_addr->sa_family) {
682 ret = -EINVAL;
683 goto err;
684 }
685
686 memcpy(src_in, src_addr, rdma_addr_size(src_addr));
687 } else {
688 src_in->sa_family = dst_addr->sa_family;
689 }
690
691 memcpy(dst_in, dst_addr, rdma_addr_size(dst_addr));
692 req->addr = addr;
693 req->callback = callback;
694 req->context = context;
695 req->resolve_by_gid_attr = resolve_by_gid_attr;
696 INIT_DELAYED_WORK(&req->work, process_one_req);
697 req->seq = (u32)atomic_inc_return(&ib_nl_addr_request_seq);
698
699 req->status = addr_resolve(src_in, dst_in, addr, true,
700 req->resolve_by_gid_attr, req->seq);
701 switch (req->status) {
702 case 0:
703 req->timeout = jiffies;
704 queue_req(req);
705 break;
706 case -ENODATA:
707 req->timeout = msecs_to_jiffies(timeout_ms) + jiffies;
708 queue_req(req);
709 break;
710 default:
711 ret = req->status;
712 goto err;
713 }
714 return ret;
715err:
716 kfree(req);
717 return ret;
718}
719EXPORT_SYMBOL(rdma_resolve_ip);
720
721int roce_resolve_route_from_path(struct sa_path_rec *rec,
722 const struct ib_gid_attr *attr)
723{
724 union {
725 struct sockaddr _sockaddr;
726 struct sockaddr_in _sockaddr_in;
727 struct sockaddr_in6 _sockaddr_in6;
728 } sgid, dgid;
729 struct rdma_dev_addr dev_addr = {};
730 int ret;
731
732 might_sleep();
733
734 if (rec->roce.route_resolved)
735 return 0;
736
737 rdma_gid2ip((struct sockaddr *)&sgid, &rec->sgid);
738 rdma_gid2ip((struct sockaddr *)&dgid, &rec->dgid);
739
740 if (sgid._sockaddr.sa_family != dgid._sockaddr.sa_family)
741 return -EINVAL;
742
743 if (!attr || !attr->ndev)
744 return -EINVAL;
745
746 dev_addr.net = &init_net;
747 dev_addr.sgid_attr = attr;
748
749 ret = addr_resolve((struct sockaddr *)&sgid, (struct sockaddr *)&dgid,
750 &dev_addr, false, true, 0);
751 if (ret)
752 return ret;
753
754 if ((dev_addr.network == RDMA_NETWORK_IPV4 ||
755 dev_addr.network == RDMA_NETWORK_IPV6) &&
756 rec->rec_type != SA_PATH_REC_TYPE_ROCE_V2)
757 return -EINVAL;
758
759 rec->roce.route_resolved = true;
760 return 0;
761}
762
763/**
764 * rdma_addr_cancel - Cancel resolve ip request
765 * @addr: Pointer to address structure given previously
766 * during rdma_resolve_ip().
767 * rdma_addr_cancel() is synchronous function which cancels any pending
768 * request if there is any.
769 */
770void rdma_addr_cancel(struct rdma_dev_addr *addr)
771{
772 struct addr_req *req, *temp_req;
773 struct addr_req *found = NULL;
774
775 spin_lock_bh(&lock);
776 list_for_each_entry_safe(req, temp_req, &req_list, list) {
777 if (req->addr == addr) {
778 /*
779 * Removing from the list means we take ownership of
780 * the req
781 */
782 list_del_init(&req->list);
783 found = req;
784 break;
785 }
786 }
787 spin_unlock_bh(&lock);
788
789 if (!found)
790 return;
791
792 /*
793 * sync canceling the work after removing it from the req_list
794 * guarentees no work is running and none will be started.
795 */
796 cancel_delayed_work_sync(&found->work);
797 kfree(found);
798}
799EXPORT_SYMBOL(rdma_addr_cancel);
800
801struct resolve_cb_context {
802 struct completion comp;
803 int status;
804};
805
806static void resolve_cb(int status, struct sockaddr *src_addr,
807 struct rdma_dev_addr *addr, void *context)
808{
809 ((struct resolve_cb_context *)context)->status = status;
810 complete(&((struct resolve_cb_context *)context)->comp);
811}
812
813int rdma_addr_find_l2_eth_by_grh(const union ib_gid *sgid,
814 const union ib_gid *dgid,
815 u8 *dmac, const struct ib_gid_attr *sgid_attr,
816 int *hoplimit)
817{
818 struct rdma_dev_addr dev_addr;
819 struct resolve_cb_context ctx;
820 union {
821 struct sockaddr_in _sockaddr_in;
822 struct sockaddr_in6 _sockaddr_in6;
823 } sgid_addr, dgid_addr;
824 int ret;
825
826 rdma_gid2ip((struct sockaddr *)&sgid_addr, sgid);
827 rdma_gid2ip((struct sockaddr *)&dgid_addr, dgid);
828
829 memset(&dev_addr, 0, sizeof(dev_addr));
830 dev_addr.net = &init_net;
831 dev_addr.sgid_attr = sgid_attr;
832
833 init_completion(&ctx.comp);
834 ret = rdma_resolve_ip((struct sockaddr *)&sgid_addr,
835 (struct sockaddr *)&dgid_addr, &dev_addr, 1000,
836 resolve_cb, true, &ctx);
837 if (ret)
838 return ret;
839
840 wait_for_completion(&ctx.comp);
841
842 ret = ctx.status;
843 if (ret)
844 return ret;
845
846 memcpy(dmac, dev_addr.dst_dev_addr, ETH_ALEN);
847 *hoplimit = dev_addr.hoplimit;
848 return 0;
849}
850
851static int netevent_callback(struct notifier_block *self, unsigned long event,
852 void *ctx)
853{
854 struct addr_req *req;
855
856 if (event == NETEVENT_NEIGH_UPDATE) {
857 struct neighbour *neigh = ctx;
858
859 if (neigh->nud_state & NUD_VALID) {
860 spin_lock_bh(&lock);
861 list_for_each_entry(req, &req_list, list)
862 set_timeout(req, jiffies);
863 spin_unlock_bh(&lock);
864 }
865 }
866 return 0;
867}
868
869static struct notifier_block nb = {
870 .notifier_call = netevent_callback
871};
872
873int addr_init(void)
874{
875 addr_wq = alloc_ordered_workqueue("ib_addr", 0);
876 if (!addr_wq)
877 return -ENOMEM;
878
879 register_netevent_notifier(&nb);
880
881 return 0;
882}
883
884void addr_cleanup(void)
885{
886 unregister_netevent_notifier(&nb);
887 destroy_workqueue(addr_wq);
888 WARN_ON(!list_empty(&req_list));
889}
1/*
2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005 Intel Corporation. All rights reserved.
6 *
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
12 *
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
15 * conditions are met:
16 *
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer.
20 *
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
25 *
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 * SOFTWARE.
34 */
35
36#include <linux/mutex.h>
37#include <linux/inetdevice.h>
38#include <linux/slab.h>
39#include <linux/workqueue.h>
40#include <linux/module.h>
41#include <net/arp.h>
42#include <net/neighbour.h>
43#include <net/route.h>
44#include <net/netevent.h>
45#include <net/ipv6_stubs.h>
46#include <net/ip6_route.h>
47#include <rdma/ib_addr.h>
48#include <rdma/ib_cache.h>
49#include <rdma/ib_sa.h>
50#include <rdma/ib.h>
51#include <rdma/rdma_netlink.h>
52#include <net/netlink.h>
53
54#include "core_priv.h"
55
56struct addr_req {
57 struct list_head list;
58 struct sockaddr_storage src_addr;
59 struct sockaddr_storage dst_addr;
60 struct rdma_dev_addr *addr;
61 void *context;
62 void (*callback)(int status, struct sockaddr *src_addr,
63 struct rdma_dev_addr *addr, void *context);
64 unsigned long timeout;
65 struct delayed_work work;
66 bool resolve_by_gid_attr; /* Consider gid attr in resolve phase */
67 int status;
68 u32 seq;
69};
70
71static atomic_t ib_nl_addr_request_seq = ATOMIC_INIT(0);
72
73static DEFINE_SPINLOCK(lock);
74static LIST_HEAD(req_list);
75static struct workqueue_struct *addr_wq;
76
77static const struct nla_policy ib_nl_addr_policy[LS_NLA_TYPE_MAX] = {
78 [LS_NLA_TYPE_DGID] = {.type = NLA_BINARY,
79 .len = sizeof(struct rdma_nla_ls_gid)},
80};
81
82static inline bool ib_nl_is_good_ip_resp(const struct nlmsghdr *nlh)
83{
84 struct nlattr *tb[LS_NLA_TYPE_MAX] = {};
85 int ret;
86
87 if (nlh->nlmsg_flags & RDMA_NL_LS_F_ERR)
88 return false;
89
90 ret = nla_parse_deprecated(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
91 nlmsg_len(nlh), ib_nl_addr_policy, NULL);
92 if (ret)
93 return false;
94
95 return true;
96}
97
98static void ib_nl_process_good_ip_rsep(const struct nlmsghdr *nlh)
99{
100 const struct nlattr *head, *curr;
101 union ib_gid gid;
102 struct addr_req *req;
103 int len, rem;
104 int found = 0;
105
106 head = (const struct nlattr *)nlmsg_data(nlh);
107 len = nlmsg_len(nlh);
108
109 nla_for_each_attr(curr, head, len, rem) {
110 if (curr->nla_type == LS_NLA_TYPE_DGID)
111 memcpy(&gid, nla_data(curr), nla_len(curr));
112 }
113
114 spin_lock_bh(&lock);
115 list_for_each_entry(req, &req_list, list) {
116 if (nlh->nlmsg_seq != req->seq)
117 continue;
118 /* We set the DGID part, the rest was set earlier */
119 rdma_addr_set_dgid(req->addr, &gid);
120 req->status = 0;
121 found = 1;
122 break;
123 }
124 spin_unlock_bh(&lock);
125
126 if (!found)
127 pr_info("Couldn't find request waiting for DGID: %pI6\n",
128 &gid);
129}
130
131int ib_nl_handle_ip_res_resp(struct sk_buff *skb,
132 struct nlmsghdr *nlh,
133 struct netlink_ext_ack *extack)
134{
135 if ((nlh->nlmsg_flags & NLM_F_REQUEST) ||
136 !(NETLINK_CB(skb).sk))
137 return -EPERM;
138
139 if (ib_nl_is_good_ip_resp(nlh))
140 ib_nl_process_good_ip_rsep(nlh);
141
142 return skb->len;
143}
144
145static int ib_nl_ip_send_msg(struct rdma_dev_addr *dev_addr,
146 const void *daddr,
147 u32 seq, u16 family)
148{
149 struct sk_buff *skb = NULL;
150 struct nlmsghdr *nlh;
151 struct rdma_ls_ip_resolve_header *header;
152 void *data;
153 size_t size;
154 int attrtype;
155 int len;
156
157 if (family == AF_INET) {
158 size = sizeof(struct in_addr);
159 attrtype = RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_IPV4;
160 } else {
161 size = sizeof(struct in6_addr);
162 attrtype = RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_IPV6;
163 }
164
165 len = nla_total_size(sizeof(size));
166 len += NLMSG_ALIGN(sizeof(*header));
167
168 skb = nlmsg_new(len, GFP_KERNEL);
169 if (!skb)
170 return -ENOMEM;
171
172 data = ibnl_put_msg(skb, &nlh, seq, 0, RDMA_NL_LS,
173 RDMA_NL_LS_OP_IP_RESOLVE, NLM_F_REQUEST);
174 if (!data) {
175 nlmsg_free(skb);
176 return -ENODATA;
177 }
178
179 /* Construct the family header first */
180 header = skb_put(skb, NLMSG_ALIGN(sizeof(*header)));
181 header->ifindex = dev_addr->bound_dev_if;
182 nla_put(skb, attrtype, size, daddr);
183
184 /* Repair the nlmsg header length */
185 nlmsg_end(skb, nlh);
186 rdma_nl_multicast(&init_net, skb, RDMA_NL_GROUP_LS, GFP_KERNEL);
187
188 /* Make the request retry, so when we get the response from userspace
189 * we will have something.
190 */
191 return -ENODATA;
192}
193
194int rdma_addr_size(const struct sockaddr *addr)
195{
196 switch (addr->sa_family) {
197 case AF_INET:
198 return sizeof(struct sockaddr_in);
199 case AF_INET6:
200 return sizeof(struct sockaddr_in6);
201 case AF_IB:
202 return sizeof(struct sockaddr_ib);
203 default:
204 return 0;
205 }
206}
207EXPORT_SYMBOL(rdma_addr_size);
208
209int rdma_addr_size_in6(struct sockaddr_in6 *addr)
210{
211 int ret = rdma_addr_size((struct sockaddr *) addr);
212
213 return ret <= sizeof(*addr) ? ret : 0;
214}
215EXPORT_SYMBOL(rdma_addr_size_in6);
216
217int rdma_addr_size_kss(struct __kernel_sockaddr_storage *addr)
218{
219 int ret = rdma_addr_size((struct sockaddr *) addr);
220
221 return ret <= sizeof(*addr) ? ret : 0;
222}
223EXPORT_SYMBOL(rdma_addr_size_kss);
224
225/**
226 * rdma_copy_src_l2_addr - Copy netdevice source addresses
227 * @dev_addr: Destination address pointer where to copy the addresses
228 * @dev: Netdevice whose source addresses to copy
229 *
230 * rdma_copy_src_l2_addr() copies source addresses from the specified netdevice.
231 * This includes unicast address, broadcast address, device type and
232 * interface index.
233 */
234void rdma_copy_src_l2_addr(struct rdma_dev_addr *dev_addr,
235 const struct net_device *dev)
236{
237 dev_addr->dev_type = dev->type;
238 memcpy(dev_addr->src_dev_addr, dev->dev_addr, MAX_ADDR_LEN);
239 memcpy(dev_addr->broadcast, dev->broadcast, MAX_ADDR_LEN);
240 dev_addr->bound_dev_if = dev->ifindex;
241}
242EXPORT_SYMBOL(rdma_copy_src_l2_addr);
243
244static struct net_device *
245rdma_find_ndev_for_src_ip_rcu(struct net *net, const struct sockaddr *src_in)
246{
247 struct net_device *dev = NULL;
248 int ret = -EADDRNOTAVAIL;
249
250 switch (src_in->sa_family) {
251 case AF_INET:
252 dev = __ip_dev_find(net,
253 ((const struct sockaddr_in *)src_in)->sin_addr.s_addr,
254 false);
255 if (dev)
256 ret = 0;
257 break;
258#if IS_ENABLED(CONFIG_IPV6)
259 case AF_INET6:
260 for_each_netdev_rcu(net, dev) {
261 if (ipv6_chk_addr(net,
262 &((const struct sockaddr_in6 *)src_in)->sin6_addr,
263 dev, 1)) {
264 ret = 0;
265 break;
266 }
267 }
268 break;
269#endif
270 }
271 return ret ? ERR_PTR(ret) : dev;
272}
273
274int rdma_translate_ip(const struct sockaddr *addr,
275 struct rdma_dev_addr *dev_addr)
276{
277 struct net_device *dev;
278
279 if (dev_addr->bound_dev_if) {
280 dev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
281 if (!dev)
282 return -ENODEV;
283 rdma_copy_src_l2_addr(dev_addr, dev);
284 dev_put(dev);
285 return 0;
286 }
287
288 rcu_read_lock();
289 dev = rdma_find_ndev_for_src_ip_rcu(dev_addr->net, addr);
290 if (!IS_ERR(dev))
291 rdma_copy_src_l2_addr(dev_addr, dev);
292 rcu_read_unlock();
293 return PTR_ERR_OR_ZERO(dev);
294}
295EXPORT_SYMBOL(rdma_translate_ip);
296
297static void set_timeout(struct addr_req *req, unsigned long time)
298{
299 unsigned long delay;
300
301 delay = time - jiffies;
302 if ((long)delay < 0)
303 delay = 0;
304
305 mod_delayed_work(addr_wq, &req->work, delay);
306}
307
308static void queue_req(struct addr_req *req)
309{
310 spin_lock_bh(&lock);
311 list_add_tail(&req->list, &req_list);
312 set_timeout(req, req->timeout);
313 spin_unlock_bh(&lock);
314}
315
316static int ib_nl_fetch_ha(struct rdma_dev_addr *dev_addr,
317 const void *daddr, u32 seq, u16 family)
318{
319 if (!rdma_nl_chk_listeners(RDMA_NL_GROUP_LS))
320 return -EADDRNOTAVAIL;
321
322 return ib_nl_ip_send_msg(dev_addr, daddr, seq, family);
323}
324
325static int dst_fetch_ha(const struct dst_entry *dst,
326 struct rdma_dev_addr *dev_addr,
327 const void *daddr)
328{
329 struct neighbour *n;
330 int ret = 0;
331
332 n = dst_neigh_lookup(dst, daddr);
333 if (!n)
334 return -ENODATA;
335
336 if (!(n->nud_state & NUD_VALID)) {
337 neigh_event_send(n, NULL);
338 ret = -ENODATA;
339 } else {
340 neigh_ha_snapshot(dev_addr->dst_dev_addr, n, dst->dev);
341 }
342
343 neigh_release(n);
344
345 return ret;
346}
347
348static bool has_gateway(const struct dst_entry *dst, sa_family_t family)
349{
350 struct rtable *rt;
351 struct rt6_info *rt6;
352
353 if (family == AF_INET) {
354 rt = container_of(dst, struct rtable, dst);
355 return rt->rt_uses_gateway;
356 }
357
358 rt6 = container_of(dst, struct rt6_info, dst);
359 return rt6->rt6i_flags & RTF_GATEWAY;
360}
361
362static int fetch_ha(const struct dst_entry *dst, struct rdma_dev_addr *dev_addr,
363 const struct sockaddr *dst_in, u32 seq)
364{
365 const struct sockaddr_in *dst_in4 =
366 (const struct sockaddr_in *)dst_in;
367 const struct sockaddr_in6 *dst_in6 =
368 (const struct sockaddr_in6 *)dst_in;
369 const void *daddr = (dst_in->sa_family == AF_INET) ?
370 (const void *)&dst_in4->sin_addr.s_addr :
371 (const void *)&dst_in6->sin6_addr;
372 sa_family_t family = dst_in->sa_family;
373
374 /* If we have a gateway in IB mode then it must be an IB network */
375 if (has_gateway(dst, family) && dev_addr->network == RDMA_NETWORK_IB)
376 return ib_nl_fetch_ha(dev_addr, daddr, seq, family);
377 else
378 return dst_fetch_ha(dst, dev_addr, daddr);
379}
380
381static int addr4_resolve(struct sockaddr *src_sock,
382 const struct sockaddr *dst_sock,
383 struct rdma_dev_addr *addr,
384 struct rtable **prt)
385{
386 struct sockaddr_in *src_in = (struct sockaddr_in *)src_sock;
387 const struct sockaddr_in *dst_in =
388 (const struct sockaddr_in *)dst_sock;
389
390 __be32 src_ip = src_in->sin_addr.s_addr;
391 __be32 dst_ip = dst_in->sin_addr.s_addr;
392 struct rtable *rt;
393 struct flowi4 fl4;
394 int ret;
395
396 memset(&fl4, 0, sizeof(fl4));
397 fl4.daddr = dst_ip;
398 fl4.saddr = src_ip;
399 fl4.flowi4_oif = addr->bound_dev_if;
400 rt = ip_route_output_key(addr->net, &fl4);
401 ret = PTR_ERR_OR_ZERO(rt);
402 if (ret)
403 return ret;
404
405 src_in->sin_addr.s_addr = fl4.saddr;
406
407 addr->hoplimit = ip4_dst_hoplimit(&rt->dst);
408
409 *prt = rt;
410 return 0;
411}
412
413#if IS_ENABLED(CONFIG_IPV6)
414static int addr6_resolve(struct sockaddr *src_sock,
415 const struct sockaddr *dst_sock,
416 struct rdma_dev_addr *addr,
417 struct dst_entry **pdst)
418{
419 struct sockaddr_in6 *src_in = (struct sockaddr_in6 *)src_sock;
420 const struct sockaddr_in6 *dst_in =
421 (const struct sockaddr_in6 *)dst_sock;
422 struct flowi6 fl6;
423 struct dst_entry *dst;
424 int ret;
425
426 memset(&fl6, 0, sizeof fl6);
427 fl6.daddr = dst_in->sin6_addr;
428 fl6.saddr = src_in->sin6_addr;
429 fl6.flowi6_oif = addr->bound_dev_if;
430
431 ret = ipv6_stub->ipv6_dst_lookup(addr->net, NULL, &dst, &fl6);
432 if (ret < 0)
433 return ret;
434
435 if (ipv6_addr_any(&src_in->sin6_addr))
436 src_in->sin6_addr = fl6.saddr;
437
438 addr->hoplimit = ip6_dst_hoplimit(dst);
439
440 *pdst = dst;
441 return 0;
442}
443#else
444static int addr6_resolve(struct sockaddr *src_sock,
445 const struct sockaddr *dst_sock,
446 struct rdma_dev_addr *addr,
447 struct dst_entry **pdst)
448{
449 return -EADDRNOTAVAIL;
450}
451#endif
452
453static int addr_resolve_neigh(const struct dst_entry *dst,
454 const struct sockaddr *dst_in,
455 struct rdma_dev_addr *addr,
456 unsigned int ndev_flags,
457 u32 seq)
458{
459 int ret = 0;
460
461 if (ndev_flags & IFF_LOOPBACK) {
462 memcpy(addr->dst_dev_addr, addr->src_dev_addr, MAX_ADDR_LEN);
463 } else {
464 if (!(ndev_flags & IFF_NOARP)) {
465 /* If the device doesn't do ARP internally */
466 ret = fetch_ha(dst, addr, dst_in, seq);
467 }
468 }
469 return ret;
470}
471
472static int copy_src_l2_addr(struct rdma_dev_addr *dev_addr,
473 const struct sockaddr *dst_in,
474 const struct dst_entry *dst,
475 const struct net_device *ndev)
476{
477 int ret = 0;
478
479 if (dst->dev->flags & IFF_LOOPBACK)
480 ret = rdma_translate_ip(dst_in, dev_addr);
481 else
482 rdma_copy_src_l2_addr(dev_addr, dst->dev);
483
484 /*
485 * If there's a gateway and type of device not ARPHRD_INFINIBAND,
486 * we're definitely in RoCE v2 (as RoCE v1 isn't routable) set the
487 * network type accordingly.
488 */
489 if (has_gateway(dst, dst_in->sa_family) &&
490 ndev->type != ARPHRD_INFINIBAND)
491 dev_addr->network = dst_in->sa_family == AF_INET ?
492 RDMA_NETWORK_IPV4 :
493 RDMA_NETWORK_IPV6;
494 else
495 dev_addr->network = RDMA_NETWORK_IB;
496
497 return ret;
498}
499
500static int rdma_set_src_addr_rcu(struct rdma_dev_addr *dev_addr,
501 unsigned int *ndev_flags,
502 const struct sockaddr *dst_in,
503 const struct dst_entry *dst)
504{
505 struct net_device *ndev = READ_ONCE(dst->dev);
506
507 *ndev_flags = ndev->flags;
508 /* A physical device must be the RDMA device to use */
509 if (ndev->flags & IFF_LOOPBACK) {
510 /*
511 * RDMA (IB/RoCE, iWarp) doesn't run on lo interface or
512 * loopback IP address. So if route is resolved to loopback
513 * interface, translate that to a real ndev based on non
514 * loopback IP address.
515 */
516 ndev = rdma_find_ndev_for_src_ip_rcu(dev_net(ndev), dst_in);
517 if (IS_ERR(ndev))
518 return -ENODEV;
519 }
520
521 return copy_src_l2_addr(dev_addr, dst_in, dst, ndev);
522}
523
524static int set_addr_netns_by_gid_rcu(struct rdma_dev_addr *addr)
525{
526 struct net_device *ndev;
527
528 ndev = rdma_read_gid_attr_ndev_rcu(addr->sgid_attr);
529 if (IS_ERR(ndev))
530 return PTR_ERR(ndev);
531
532 /*
533 * Since we are holding the rcu, reading net and ifindex
534 * are safe without any additional reference; because
535 * change_net_namespace() in net/core/dev.c does rcu sync
536 * after it changes the state to IFF_DOWN and before
537 * updating netdev fields {net, ifindex}.
538 */
539 addr->net = dev_net(ndev);
540 addr->bound_dev_if = ndev->ifindex;
541 return 0;
542}
543
544static void rdma_addr_set_net_defaults(struct rdma_dev_addr *addr)
545{
546 addr->net = &init_net;
547 addr->bound_dev_if = 0;
548}
549
550static int addr_resolve(struct sockaddr *src_in,
551 const struct sockaddr *dst_in,
552 struct rdma_dev_addr *addr,
553 bool resolve_neigh,
554 bool resolve_by_gid_attr,
555 u32 seq)
556{
557 struct dst_entry *dst = NULL;
558 unsigned int ndev_flags = 0;
559 struct rtable *rt = NULL;
560 int ret;
561
562 if (!addr->net) {
563 pr_warn_ratelimited("%s: missing namespace\n", __func__);
564 return -EINVAL;
565 }
566
567 rcu_read_lock();
568 if (resolve_by_gid_attr) {
569 if (!addr->sgid_attr) {
570 rcu_read_unlock();
571 pr_warn_ratelimited("%s: missing gid_attr\n", __func__);
572 return -EINVAL;
573 }
574 /*
575 * If the request is for a specific gid attribute of the
576 * rdma_dev_addr, derive net from the netdevice of the
577 * GID attribute.
578 */
579 ret = set_addr_netns_by_gid_rcu(addr);
580 if (ret) {
581 rcu_read_unlock();
582 return ret;
583 }
584 }
585 if (src_in->sa_family == AF_INET) {
586 ret = addr4_resolve(src_in, dst_in, addr, &rt);
587 dst = &rt->dst;
588 } else {
589 ret = addr6_resolve(src_in, dst_in, addr, &dst);
590 }
591 if (ret) {
592 rcu_read_unlock();
593 goto done;
594 }
595 ret = rdma_set_src_addr_rcu(addr, &ndev_flags, dst_in, dst);
596 rcu_read_unlock();
597
598 /*
599 * Resolve neighbor destination address if requested and
600 * only if src addr translation didn't fail.
601 */
602 if (!ret && resolve_neigh)
603 ret = addr_resolve_neigh(dst, dst_in, addr, ndev_flags, seq);
604
605 if (src_in->sa_family == AF_INET)
606 ip_rt_put(rt);
607 else
608 dst_release(dst);
609done:
610 /*
611 * Clear the addr net to go back to its original state, only if it was
612 * derived from GID attribute in this context.
613 */
614 if (resolve_by_gid_attr)
615 rdma_addr_set_net_defaults(addr);
616 return ret;
617}
618
619static void process_one_req(struct work_struct *_work)
620{
621 struct addr_req *req;
622 struct sockaddr *src_in, *dst_in;
623
624 req = container_of(_work, struct addr_req, work.work);
625
626 if (req->status == -ENODATA) {
627 src_in = (struct sockaddr *)&req->src_addr;
628 dst_in = (struct sockaddr *)&req->dst_addr;
629 req->status = addr_resolve(src_in, dst_in, req->addr,
630 true, req->resolve_by_gid_attr,
631 req->seq);
632 if (req->status && time_after_eq(jiffies, req->timeout)) {
633 req->status = -ETIMEDOUT;
634 } else if (req->status == -ENODATA) {
635 /* requeue the work for retrying again */
636 spin_lock_bh(&lock);
637 if (!list_empty(&req->list))
638 set_timeout(req, req->timeout);
639 spin_unlock_bh(&lock);
640 return;
641 }
642 }
643
644 req->callback(req->status, (struct sockaddr *)&req->src_addr,
645 req->addr, req->context);
646 req->callback = NULL;
647
648 spin_lock_bh(&lock);
649 if (!list_empty(&req->list)) {
650 /*
651 * Although the work will normally have been canceled by the
652 * workqueue, it can still be requeued as long as it is on the
653 * req_list.
654 */
655 cancel_delayed_work(&req->work);
656 list_del_init(&req->list);
657 kfree(req);
658 }
659 spin_unlock_bh(&lock);
660}
661
662int rdma_resolve_ip(struct sockaddr *src_addr, const struct sockaddr *dst_addr,
663 struct rdma_dev_addr *addr, unsigned long timeout_ms,
664 void (*callback)(int status, struct sockaddr *src_addr,
665 struct rdma_dev_addr *addr, void *context),
666 bool resolve_by_gid_attr, void *context)
667{
668 struct sockaddr *src_in, *dst_in;
669 struct addr_req *req;
670 int ret = 0;
671
672 req = kzalloc(sizeof *req, GFP_KERNEL);
673 if (!req)
674 return -ENOMEM;
675
676 src_in = (struct sockaddr *) &req->src_addr;
677 dst_in = (struct sockaddr *) &req->dst_addr;
678
679 if (src_addr) {
680 if (src_addr->sa_family != dst_addr->sa_family) {
681 ret = -EINVAL;
682 goto err;
683 }
684
685 memcpy(src_in, src_addr, rdma_addr_size(src_addr));
686 } else {
687 src_in->sa_family = dst_addr->sa_family;
688 }
689
690 memcpy(dst_in, dst_addr, rdma_addr_size(dst_addr));
691 req->addr = addr;
692 req->callback = callback;
693 req->context = context;
694 req->resolve_by_gid_attr = resolve_by_gid_attr;
695 INIT_DELAYED_WORK(&req->work, process_one_req);
696 req->seq = (u32)atomic_inc_return(&ib_nl_addr_request_seq);
697
698 req->status = addr_resolve(src_in, dst_in, addr, true,
699 req->resolve_by_gid_attr, req->seq);
700 switch (req->status) {
701 case 0:
702 req->timeout = jiffies;
703 queue_req(req);
704 break;
705 case -ENODATA:
706 req->timeout = msecs_to_jiffies(timeout_ms) + jiffies;
707 queue_req(req);
708 break;
709 default:
710 ret = req->status;
711 goto err;
712 }
713 return ret;
714err:
715 kfree(req);
716 return ret;
717}
718EXPORT_SYMBOL(rdma_resolve_ip);
719
720int roce_resolve_route_from_path(struct sa_path_rec *rec,
721 const struct ib_gid_attr *attr)
722{
723 union {
724 struct sockaddr _sockaddr;
725 struct sockaddr_in _sockaddr_in;
726 struct sockaddr_in6 _sockaddr_in6;
727 } sgid, dgid;
728 struct rdma_dev_addr dev_addr = {};
729 int ret;
730
731 if (rec->roce.route_resolved)
732 return 0;
733
734 rdma_gid2ip((struct sockaddr *)&sgid, &rec->sgid);
735 rdma_gid2ip((struct sockaddr *)&dgid, &rec->dgid);
736
737 if (sgid._sockaddr.sa_family != dgid._sockaddr.sa_family)
738 return -EINVAL;
739
740 if (!attr || !attr->ndev)
741 return -EINVAL;
742
743 dev_addr.net = &init_net;
744 dev_addr.sgid_attr = attr;
745
746 ret = addr_resolve((struct sockaddr *)&sgid, (struct sockaddr *)&dgid,
747 &dev_addr, false, true, 0);
748 if (ret)
749 return ret;
750
751 if ((dev_addr.network == RDMA_NETWORK_IPV4 ||
752 dev_addr.network == RDMA_NETWORK_IPV6) &&
753 rec->rec_type != SA_PATH_REC_TYPE_ROCE_V2)
754 return -EINVAL;
755
756 rec->roce.route_resolved = true;
757 return 0;
758}
759
760/**
761 * rdma_addr_cancel - Cancel resolve ip request
762 * @addr: Pointer to address structure given previously
763 * during rdma_resolve_ip().
764 * rdma_addr_cancel() is synchronous function which cancels any pending
765 * request if there is any.
766 */
767void rdma_addr_cancel(struct rdma_dev_addr *addr)
768{
769 struct addr_req *req, *temp_req;
770 struct addr_req *found = NULL;
771
772 spin_lock_bh(&lock);
773 list_for_each_entry_safe(req, temp_req, &req_list, list) {
774 if (req->addr == addr) {
775 /*
776 * Removing from the list means we take ownership of
777 * the req
778 */
779 list_del_init(&req->list);
780 found = req;
781 break;
782 }
783 }
784 spin_unlock_bh(&lock);
785
786 if (!found)
787 return;
788
789 /*
790 * sync canceling the work after removing it from the req_list
791 * guarentees no work is running and none will be started.
792 */
793 cancel_delayed_work_sync(&found->work);
794 kfree(found);
795}
796EXPORT_SYMBOL(rdma_addr_cancel);
797
798struct resolve_cb_context {
799 struct completion comp;
800 int status;
801};
802
803static void resolve_cb(int status, struct sockaddr *src_addr,
804 struct rdma_dev_addr *addr, void *context)
805{
806 ((struct resolve_cb_context *)context)->status = status;
807 complete(&((struct resolve_cb_context *)context)->comp);
808}
809
810int rdma_addr_find_l2_eth_by_grh(const union ib_gid *sgid,
811 const union ib_gid *dgid,
812 u8 *dmac, const struct ib_gid_attr *sgid_attr,
813 int *hoplimit)
814{
815 struct rdma_dev_addr dev_addr;
816 struct resolve_cb_context ctx;
817 union {
818 struct sockaddr_in _sockaddr_in;
819 struct sockaddr_in6 _sockaddr_in6;
820 } sgid_addr, dgid_addr;
821 int ret;
822
823 rdma_gid2ip((struct sockaddr *)&sgid_addr, sgid);
824 rdma_gid2ip((struct sockaddr *)&dgid_addr, dgid);
825
826 memset(&dev_addr, 0, sizeof(dev_addr));
827 dev_addr.net = &init_net;
828 dev_addr.sgid_attr = sgid_attr;
829
830 init_completion(&ctx.comp);
831 ret = rdma_resolve_ip((struct sockaddr *)&sgid_addr,
832 (struct sockaddr *)&dgid_addr, &dev_addr, 1000,
833 resolve_cb, true, &ctx);
834 if (ret)
835 return ret;
836
837 wait_for_completion(&ctx.comp);
838
839 ret = ctx.status;
840 if (ret)
841 return ret;
842
843 memcpy(dmac, dev_addr.dst_dev_addr, ETH_ALEN);
844 *hoplimit = dev_addr.hoplimit;
845 return 0;
846}
847
848static int netevent_callback(struct notifier_block *self, unsigned long event,
849 void *ctx)
850{
851 struct addr_req *req;
852
853 if (event == NETEVENT_NEIGH_UPDATE) {
854 struct neighbour *neigh = ctx;
855
856 if (neigh->nud_state & NUD_VALID) {
857 spin_lock_bh(&lock);
858 list_for_each_entry(req, &req_list, list)
859 set_timeout(req, jiffies);
860 spin_unlock_bh(&lock);
861 }
862 }
863 return 0;
864}
865
866static struct notifier_block nb = {
867 .notifier_call = netevent_callback
868};
869
870int addr_init(void)
871{
872 addr_wq = alloc_ordered_workqueue("ib_addr", 0);
873 if (!addr_wq)
874 return -ENOMEM;
875
876 register_netevent_notifier(&nb);
877
878 return 0;
879}
880
881void addr_cleanup(void)
882{
883 unregister_netevent_notifier(&nb);
884 destroy_workqueue(addr_wq);
885 WARN_ON(!list_empty(&req_list));
886}