drivers/infiniband/core/addr.c

   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/addrconf.h>
  46 #include <net/ip6_route.h>
  47 #include <rdma/ib_addr.h>
  48 #include <rdma/ib.h>
  49 #include <rdma/rdma_netlink.h>
  50 #include <net/netlink.h>
  51
  52 #include "core_priv.h"
  53
  54 struct addr_req {
  55         struct list_head list;
  56         struct sockaddr_storage src_addr;
  57         struct sockaddr_storage dst_addr;
  58         struct rdma_dev_addr *addr;
  59         struct rdma_addr_client *client;
  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         int status;
  66         u32 seq;
  67 };
  68
  69 static atomic_t ib_nl_addr_request_seq = ATOMIC_INIT(0);
  70
  71 static void process_req(struct work_struct *work);
  72
  73 static DEFINE_MUTEX(lock);
  74 static LIST_HEAD(req_list);
  75 static DECLARE_DELAYED_WORK(work, process_req);
  76 static struct workqueue_struct *addr_wq;
  77
  78 static const struct nla_policy ib_nl_addr_policy[LS_NLA_TYPE_MAX] = {
  79         [LS_NLA_TYPE_DGID] = {.type = NLA_BINARY,
  80                 .len = sizeof(struct rdma_nla_ls_gid)},
  81 };
  82
  83 static 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(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
  99 static 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         mutex_lock(&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         mutex_unlock(&lock);
 126
 127         if (!found)
 128                 pr_info("Couldn't find request waiting for DGID: %pI6\n",
 129                         &gid);
 130 }
 131
 132 int 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 skb->len;
 144 }
 145
 146 static 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(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
 195 int rdma_addr_size(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 }
 208 EXPORT_SYMBOL(rdma_addr_size);
 209
 210 int 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 }
 216 EXPORT_SYMBOL(rdma_addr_size_in6);
 217
 218 int 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 }
 224 EXPORT_SYMBOL(rdma_addr_size_kss);
 225
 226 static struct rdma_addr_client self;
 227
 228 void rdma_addr_register_client(struct rdma_addr_client *client)
 229 {
 230         atomic_set(&client->refcount, 1);
 231         init_completion(&client->comp);
 232 }
 233 EXPORT_SYMBOL(rdma_addr_register_client);
 234
 235 static inline void put_client(struct rdma_addr_client *client)
 236 {
 237         if (atomic_dec_and_test(&client->refcount))
 238                 complete(&client->comp);
 239 }
 240
 241 void rdma_addr_unregister_client(struct rdma_addr_client *client)
 242 {
 243         put_client(client);
 244         wait_for_completion(&client->comp);
 245 }
 246 EXPORT_SYMBOL(rdma_addr_unregister_client);
 247
 248 int rdma_copy_addr(struct rdma_dev_addr *dev_addr, struct net_device *dev,
 249                      const unsigned char *dst_dev_addr)
 250 {
 251         dev_addr->dev_type = dev->type;
 252         memcpy(dev_addr->src_dev_addr, dev->dev_addr, MAX_ADDR_LEN);
 253         memcpy(dev_addr->broadcast, dev->broadcast, MAX_ADDR_LEN);
 254         if (dst_dev_addr)
 255                 memcpy(dev_addr->dst_dev_addr, dst_dev_addr, MAX_ADDR_LEN);
 256         dev_addr->bound_dev_if = dev->ifindex;
 257         return 0;
 258 }
 259 EXPORT_SYMBOL(rdma_copy_addr);
 260
 261 int rdma_translate_ip(const struct sockaddr *addr,
 262                       struct rdma_dev_addr *dev_addr,
 263                       u16 *vlan_id)
 264 {
 265         struct net_device *dev;
 266         int ret = -EADDRNOTAVAIL;
 267
 268         if (dev_addr->bound_dev_if) {
 269                 dev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
 270                 if (!dev)
 271                         return -ENODEV;
 272                 ret = rdma_copy_addr(dev_addr, dev, NULL);
 273                 dev_put(dev);
 274                 return ret;
 275         }
 276
 277         switch (addr->sa_family) {
 278         case AF_INET:
 279                 dev = ip_dev_find(dev_addr->net,
 280                         ((const struct sockaddr_in *)addr)->sin_addr.s_addr);
 281
 282                 if (!dev)
 283                         return ret;
 284
 285                 ret = rdma_copy_addr(dev_addr, dev, NULL);
 286                 dev_addr->bound_dev_if = dev->ifindex;
 287                 if (vlan_id)
 288                         *vlan_id = rdma_vlan_dev_vlan_id(dev);
 289                 dev_put(dev);
 290                 break;
 291 #if IS_ENABLED(CONFIG_IPV6)
 292         case AF_INET6:
 293                 rcu_read_lock();
 294                 for_each_netdev_rcu(dev_addr->net, dev) {
 295                         if (ipv6_chk_addr(dev_addr->net,
 296                                           &((const struct sockaddr_in6 *)addr)->sin6_addr,
 297                                           dev, 1)) {
 298                                 ret = rdma_copy_addr(dev_addr, dev, NULL);
 299                                 dev_addr->bound_dev_if = dev->ifindex;
 300                                 if (vlan_id)
 301                                         *vlan_id = rdma_vlan_dev_vlan_id(dev);
 302                                 break;
 303                         }
 304                 }
 305                 rcu_read_unlock();
 306                 break;
 307 #endif
 308         }
 309         return ret;
 310 }
 311 EXPORT_SYMBOL(rdma_translate_ip);
 312
 313 static void set_timeout(struct delayed_work *delayed_work, unsigned long time)
 314 {
 315         unsigned long delay;
 316
 317         delay = time - jiffies;
 318         if ((long)delay < 0)
 319                 delay = 0;
 320
 321         mod_delayed_work(addr_wq, delayed_work, delay);
 322 }
 323
 324 static void queue_req(struct addr_req *req)
 325 {
 326         struct addr_req *temp_req;
 327
 328         mutex_lock(&lock);
 329         list_for_each_entry_reverse(temp_req, &req_list, list) {
 330                 if (time_after_eq(req->timeout, temp_req->timeout))
 331                         break;
 332         }
 333
 334         list_add(&req->list, &temp_req->list);
 335
 336         set_timeout(&req->work, req->timeout);
 337         mutex_unlock(&lock);
 338 }
 339
 340 static int ib_nl_fetch_ha(struct dst_entry *dst, struct rdma_dev_addr *dev_addr,
 341                           const void *daddr, u32 seq, u16 family)
 342 {
 343         if (rdma_nl_chk_listeners(RDMA_NL_GROUP_LS))
 344                 return -EADDRNOTAVAIL;
 345
 346         /* We fill in what we can, the response will fill the rest */
 347         rdma_copy_addr(dev_addr, dst->dev, NULL);
 348         return ib_nl_ip_send_msg(dev_addr, daddr, seq, family);
 349 }
 350
 351 static int dst_fetch_ha(struct dst_entry *dst, struct rdma_dev_addr *dev_addr,
 352                         const void *daddr)
 353 {
 354         struct neighbour *n;
 355         int ret;
 356
 357         n = dst_neigh_lookup(dst, daddr);
 358
 359         rcu_read_lock();
 360         if (!n || !(n->nud_state & NUD_VALID)) {
 361                 if (n)
 362                         neigh_event_send(n, NULL);
 363                 ret = -ENODATA;
 364         } else {
 365                 ret = rdma_copy_addr(dev_addr, dst->dev, n->ha);
 366         }
 367         rcu_read_unlock();
 368
 369         if (n)
 370                 neigh_release(n);
 371
 372         return ret;
 373 }
 374
 375 static bool has_gateway(struct dst_entry *dst, sa_family_t family)
 376 {
 377         struct rtable *rt;
 378         struct rt6_info *rt6;
 379
 380         if (family == AF_INET) {
 381                 rt = container_of(dst, struct rtable, dst);
 382                 return rt->rt_uses_gateway;
 383         }
 384
 385         rt6 = container_of(dst, struct rt6_info, dst);
 386         return rt6->rt6i_flags & RTF_GATEWAY;
 387 }
 388
 389 static int fetch_ha(struct dst_entry *dst, struct rdma_dev_addr *dev_addr,
 390                     const struct sockaddr *dst_in, u32 seq)
 391 {
 392         const struct sockaddr_in *dst_in4 =
 393                 (const struct sockaddr_in *)dst_in;
 394         const struct sockaddr_in6 *dst_in6 =
 395                 (const struct sockaddr_in6 *)dst_in;
 396         const void *daddr = (dst_in->sa_family == AF_INET) ?
 397                 (const void *)&dst_in4->sin_addr.s_addr :
 398                 (const void *)&dst_in6->sin6_addr;
 399         sa_family_t family = dst_in->sa_family;
 400
 401         /* Gateway + ARPHRD_INFINIBAND -> IB router */
 402         if (has_gateway(dst, family) && dst->dev->type == ARPHRD_INFINIBAND)
 403                 return ib_nl_fetch_ha(dst, dev_addr, daddr, seq, family);
 404         else
 405                 return dst_fetch_ha(dst, dev_addr, daddr);
 406 }
 407
 408 static int addr4_resolve(struct sockaddr_in *src_in,
 409                          const struct sockaddr_in *dst_in,
 410                          struct rdma_dev_addr *addr,
 411                          struct rtable **prt)
 412 {
 413         __be32 src_ip = src_in->sin_addr.s_addr;
 414         __be32 dst_ip = dst_in->sin_addr.s_addr;
 415         struct rtable *rt;
 416         struct flowi4 fl4;
 417         int ret;
 418
 419         memset(&fl4, 0, sizeof(fl4));
 420         fl4.daddr = dst_ip;
 421         fl4.saddr = src_ip;
 422         fl4.flowi4_oif = addr->bound_dev_if;
 423         rt = ip_route_output_key(addr->net, &fl4);
 424         ret = PTR_ERR_OR_ZERO(rt);
 425         if (ret)
 426                 return ret;
 427
 428         src_in->sin_family = AF_INET;
 429         src_in->sin_addr.s_addr = fl4.saddr;
 430
 431         /* If there's a gateway and type of device not ARPHRD_INFINIBAND, we're
 432          * definitely in RoCE v2 (as RoCE v1 isn't routable) set the network
 433          * type accordingly.
 434          */
 435         if (rt->rt_uses_gateway && rt->dst.dev->type != ARPHRD_INFINIBAND)
 436                 addr->network = RDMA_NETWORK_IPV4;
 437
 438         addr->hoplimit = ip4_dst_hoplimit(&rt->dst);
 439
 440         *prt = rt;
 441         return 0;
 442 }
 443
 444 #if IS_ENABLED(CONFIG_IPV6)
 445 static int addr6_resolve(struct sockaddr_in6 *src_in,
 446                          const struct sockaddr_in6 *dst_in,
 447                          struct rdma_dev_addr *addr,
 448                          struct dst_entry **pdst)
 449 {
 450         struct flowi6 fl6;
 451         struct dst_entry *dst;
 452         struct rt6_info *rt;
 453         int ret;
 454
 455         memset(&fl6, 0, sizeof fl6);
 456         fl6.daddr = dst_in->sin6_addr;
 457         fl6.saddr = src_in->sin6_addr;
 458         fl6.flowi6_oif = addr->bound_dev_if;
 459
 460         ret = ipv6_stub->ipv6_dst_lookup(addr->net, NULL, &dst, &fl6);
 461         if (ret < 0)
 462                 return ret;
 463
 464         rt = (struct rt6_info *)dst;
 465         if (ipv6_addr_any(&src_in->sin6_addr)) {
 466                 src_in->sin6_family = AF_INET6;
 467                 src_in->sin6_addr = fl6.saddr;
 468         }
 469
 470         /* If there's a gateway and type of device not ARPHRD_INFINIBAND, we're
 471          * definitely in RoCE v2 (as RoCE v1 isn't routable) set the network
 472          * type accordingly.
 473          */
 474         if (rt->rt6i_flags & RTF_GATEWAY &&
 475             ip6_dst_idev(dst)->dev->type != ARPHRD_INFINIBAND)
 476                 addr->network = RDMA_NETWORK_IPV6;
 477
 478         addr->hoplimit = ip6_dst_hoplimit(dst);
 479
 480         *pdst = dst;
 481         return 0;
 482 }
 483 #else
 484 static int addr6_resolve(struct sockaddr_in6 *src_in,
 485                          const struct sockaddr_in6 *dst_in,
 486                          struct rdma_dev_addr *addr,
 487                          struct dst_entry **pdst)
 488 {
 489         return -EADDRNOTAVAIL;
 490 }
 491 #endif
 492
 493 static int addr_resolve_neigh(struct dst_entry *dst,
 494                               const struct sockaddr *dst_in,
 495                               struct rdma_dev_addr *addr,
 496                               u32 seq)
 497 {
 498         if (dst->dev->flags & IFF_LOOPBACK) {
 499                 int ret;
 500
 501                 ret = rdma_translate_ip(dst_in, addr, NULL);
 502                 if (!ret)
 503                         memcpy(addr->dst_dev_addr, addr->src_dev_addr,
 504                                MAX_ADDR_LEN);
 505
 506                 return ret;
 507         }
 508
 509         /* If the device doesn't do ARP internally */
 510         if (!(dst->dev->flags & IFF_NOARP))
 511                 return fetch_ha(dst, addr, dst_in, seq);
 512
 513         return rdma_copy_addr(addr, dst->dev, NULL);
 514 }
 515
 516 static int addr_resolve(struct sockaddr *src_in,
 517                         const struct sockaddr *dst_in,
 518                         struct rdma_dev_addr *addr,
 519                         bool resolve_neigh,
 520                         u32 seq)
 521 {
 522         struct net_device *ndev;
 523         struct dst_entry *dst;
 524         int ret;
 525
 526         if (!addr->net) {
 527                 pr_warn_ratelimited("%s: missing namespace\n", __func__);
 528                 return -EINVAL;
 529         }
 530
 531         if (src_in->sa_family == AF_INET) {
 532                 struct rtable *rt = NULL;
 533                 const struct sockaddr_in *dst_in4 =
 534                         (const struct sockaddr_in *)dst_in;
 535
 536                 ret = addr4_resolve((struct sockaddr_in *)src_in,
 537                                     dst_in4, addr, &rt);
 538                 if (ret)
 539                         return ret;
 540
 541                 if (resolve_neigh)
 542                         ret = addr_resolve_neigh(&rt->dst, dst_in, addr, seq);
 543
 544                 if (addr->bound_dev_if) {
 545                         ndev = dev_get_by_index(addr->net, addr->bound_dev_if);
 546                 } else {
 547                         ndev = rt->dst.dev;
 548                         dev_hold(ndev);
 549                 }
 550
 551                 ip_rt_put(rt);
 552         } else {
 553                 const struct sockaddr_in6 *dst_in6 =
 554                         (const struct sockaddr_in6 *)dst_in;
 555
 556                 ret = addr6_resolve((struct sockaddr_in6 *)src_in,
 557                                     dst_in6, addr,
 558                                     &dst);
 559                 if (ret)
 560                         return ret;
 561
 562                 if (resolve_neigh)
 563                         ret = addr_resolve_neigh(dst, dst_in, addr, seq);
 564
 565                 if (addr->bound_dev_if) {
 566                         ndev = dev_get_by_index(addr->net, addr->bound_dev_if);
 567                 } else {
 568                         ndev = dst->dev;
 569                         dev_hold(ndev);
 570                 }
 571
 572                 dst_release(dst);
 573         }
 574
 575         if (ndev->flags & IFF_LOOPBACK) {
 576                 ret = rdma_translate_ip(dst_in, addr, NULL);
 577                 /*
 578                  * Put the loopback device and get the translated
 579                  * device instead.
 580                  */
 581                 dev_put(ndev);
 582                 ndev = dev_get_by_index(addr->net, addr->bound_dev_if);
 583         } else {
 584                 addr->bound_dev_if = ndev->ifindex;
 585         }
 586         dev_put(ndev);
 587
 588         return ret;
 589 }
 590
 591 static void process_one_req(struct work_struct *_work)
 592 {
 593         struct addr_req *req;
 594         struct sockaddr *src_in, *dst_in;
 595
 596         mutex_lock(&lock);
 597         req = container_of(_work, struct addr_req, work.work);
 598
 599         if (req->status == -ENODATA) {
 600                 src_in = (struct sockaddr *)&req->src_addr;
 601                 dst_in = (struct sockaddr *)&req->dst_addr;
 602                 req->status = addr_resolve(src_in, dst_in, req->addr,
 603                                            true, req->seq);
 604                 if (req->status && time_after_eq(jiffies, req->timeout)) {
 605                         req->status = -ETIMEDOUT;
 606                 } else if (req->status == -ENODATA) {
 607                         /* requeue the work for retrying again */
 608                         set_timeout(&req->work, req->timeout);
 609                         mutex_unlock(&lock);
 610                         return;
 611                 }
 612         }
 613         list_del(&req->list);
 614         mutex_unlock(&lock);
 615
 616         /*
 617          * Although the work will normally have been canceled by the
 618          * workqueue, it can still be requeued as long as it is on the
 619          * req_list, so it could have been requeued before we grabbed &lock.
 620          * We need to cancel it after it is removed from req_list to really be
 621          * sure it is safe to free.
 622          */
 623         cancel_delayed_work(&req->work);
 624
 625         req->callback(req->status, (struct sockaddr *)&req->src_addr,
 626                 req->addr, req->context);
 627         put_client(req->client);
 628         kfree(req);
 629 }
 630
 631 static void process_req(struct work_struct *work)
 632 {
 633         struct addr_req *req, *temp_req;
 634         struct sockaddr *src_in, *dst_in;
 635         struct list_head done_list;
 636
 637         INIT_LIST_HEAD(&done_list);
 638
 639         mutex_lock(&lock);
 640         list_for_each_entry_safe(req, temp_req, &req_list, list) {
 641                 if (req->status == -ENODATA) {
 642                         src_in = (struct sockaddr *) &req->src_addr;
 643                         dst_in = (struct sockaddr *) &req->dst_addr;
 644                         req->status = addr_resolve(src_in, dst_in, req->addr,
 645                                                    true, req->seq);
 646                         if (req->status && time_after_eq(jiffies, req->timeout))
 647                                 req->status = -ETIMEDOUT;
 648                         else if (req->status == -ENODATA) {
 649                                 set_timeout(&req->work, req->timeout);
 650                                 continue;
 651                         }
 652                 }
 653                 list_move_tail(&req->list, &done_list);
 654         }
 655
 656         mutex_unlock(&lock);
 657
 658         list_for_each_entry_safe(req, temp_req, &done_list, list) {
 659                 list_del(&req->list);
 660                 /* It is safe to cancel other work items from this work item
 661                  * because at a time there can be only one work item running
 662                  * with this single threaded work queue.
 663                  */
 664                 cancel_delayed_work(&req->work);
 665                 req->callback(req->status, (struct sockaddr *) &req->src_addr,
 666                         req->addr, req->context);
 667                 put_client(req->client);
 668                 kfree(req);
 669         }
 670 }
 671
 672 int rdma_resolve_ip(struct rdma_addr_client *client,
 673                     struct sockaddr *src_addr, struct sockaddr *dst_addr,
 674                     struct rdma_dev_addr *addr, int timeout_ms,
 675                     void (*callback)(int status, struct sockaddr *src_addr,
 676                                      struct rdma_dev_addr *addr, void *context),
 677                     void *context)
 678 {
 679         struct sockaddr *src_in, *dst_in;
 680         struct addr_req *req;
 681         int ret = 0;
 682
 683         req = kzalloc(sizeof *req, GFP_KERNEL);
 684         if (!req)
 685                 return -ENOMEM;
 686
 687         src_in = (struct sockaddr *) &req->src_addr;
 688         dst_in = (struct sockaddr *) &req->dst_addr;
 689
 690         if (src_addr) {
 691                 if (src_addr->sa_family != dst_addr->sa_family) {
 692                         ret = -EINVAL;
 693                         goto err;
 694                 }
 695
 696                 memcpy(src_in, src_addr, rdma_addr_size(src_addr));
 697         } else {
 698                 src_in->sa_family = dst_addr->sa_family;
 699         }
 700
 701         memcpy(dst_in, dst_addr, rdma_addr_size(dst_addr));
 702         req->addr = addr;
 703         req->callback = callback;
 704         req->context = context;
 705         req->client = client;
 706         atomic_inc(&client->refcount);
 707         INIT_DELAYED_WORK(&req->work, process_one_req);
 708         req->seq = (u32)atomic_inc_return(&ib_nl_addr_request_seq);
 709
 710         req->status = addr_resolve(src_in, dst_in, addr, true, req->seq);
 711         switch (req->status) {
 712         case 0:
 713                 req->timeout = jiffies;
 714                 queue_req(req);
 715                 break;
 716         case -ENODATA:
 717                 req->timeout = msecs_to_jiffies(timeout_ms) + jiffies;
 718                 queue_req(req);
 719                 break;
 720         default:
 721                 ret = req->status;
 722                 atomic_dec(&client->refcount);
 723                 goto err;
 724         }
 725         return ret;
 726 err:
 727         kfree(req);
 728         return ret;
 729 }
 730 EXPORT_SYMBOL(rdma_resolve_ip);
 731
 732 int rdma_resolve_ip_route(struct sockaddr *src_addr,
 733                           const struct sockaddr *dst_addr,
 734                           struct rdma_dev_addr *addr)
 735 {
 736         struct sockaddr_storage ssrc_addr = {};
 737         struct sockaddr *src_in = (struct sockaddr *)&ssrc_addr;
 738
 739         if (src_addr) {
 740                 if (src_addr->sa_family != dst_addr->sa_family)
 741                         return -EINVAL;
 742
 743                 memcpy(src_in, src_addr, rdma_addr_size(src_addr));
 744         } else {
 745                 src_in->sa_family = dst_addr->sa_family;
 746         }
 747
 748         return addr_resolve(src_in, dst_addr, addr, false, 0);
 749 }
 750 EXPORT_SYMBOL(rdma_resolve_ip_route);
 751
 752 void rdma_addr_cancel(struct rdma_dev_addr *addr)
 753 {
 754         struct addr_req *req, *temp_req;
 755
 756         mutex_lock(&lock);
 757         list_for_each_entry_safe(req, temp_req, &req_list, list) {
 758                 if (req->addr == addr) {
 759                         req->status = -ECANCELED;
 760                         req->timeout = jiffies;
 761                         list_move(&req->list, &req_list);
 762                         set_timeout(&req->work, req->timeout);
 763                         break;
 764                 }
 765         }
 766         mutex_unlock(&lock);
 767 }
 768 EXPORT_SYMBOL(rdma_addr_cancel);
 769
 770 struct resolve_cb_context {
 771         struct rdma_dev_addr *addr;
 772         struct completion comp;
 773         int status;
 774 };
 775
 776 static void resolve_cb(int status, struct sockaddr *src_addr,
 777              struct rdma_dev_addr *addr, void *context)
 778 {
 779         if (!status)
 780                 memcpy(((struct resolve_cb_context *)context)->addr,
 781                        addr, sizeof(struct rdma_dev_addr));
 782         ((struct resolve_cb_context *)context)->status = status;
 783         complete(&((struct resolve_cb_context *)context)->comp);
 784 }
 785
 786 int rdma_addr_find_l2_eth_by_grh(const union ib_gid *sgid,
 787                                  const union ib_gid *dgid,
 788                                  u8 *dmac, u16 *vlan_id, int *if_index,
 789                                  int *hoplimit)
 790 {
 791         int ret = 0;
 792         struct rdma_dev_addr dev_addr;
 793         struct resolve_cb_context ctx;
 794         struct net_device *dev;
 795
 796         union {
 797                 struct sockaddr     _sockaddr;
 798                 struct sockaddr_in  _sockaddr_in;
 799                 struct sockaddr_in6 _sockaddr_in6;
 800         } sgid_addr, dgid_addr;
 801
 802
 803         rdma_gid2ip(&sgid_addr._sockaddr, sgid);
 804         rdma_gid2ip(&dgid_addr._sockaddr, dgid);
 805
 806         memset(&dev_addr, 0, sizeof(dev_addr));
 807         if (if_index)
 808                 dev_addr.bound_dev_if = *if_index;
 809         dev_addr.net = &init_net;
 810
 811         ctx.addr = &dev_addr;
 812         init_completion(&ctx.comp);
 813         ret = rdma_resolve_ip(&self, &sgid_addr._sockaddr, &dgid_addr._sockaddr,
 814                         &dev_addr, 1000, resolve_cb, &ctx);
 815         if (ret)
 816                 return ret;
 817
 818         wait_for_completion(&ctx.comp);
 819
 820         ret = ctx.status;
 821         if (ret)
 822                 return ret;
 823
 824         memcpy(dmac, dev_addr.dst_dev_addr, ETH_ALEN);
 825         dev = dev_get_by_index(&init_net, dev_addr.bound_dev_if);
 826         if (!dev)
 827                 return -ENODEV;
 828         if (if_index)
 829                 *if_index = dev_addr.bound_dev_if;
 830         if (vlan_id)
 831                 *vlan_id = rdma_vlan_dev_vlan_id(dev);
 832         if (hoplimit)
 833                 *hoplimit = dev_addr.hoplimit;
 834         dev_put(dev);
 835         return ret;
 836 }
 837 EXPORT_SYMBOL(rdma_addr_find_l2_eth_by_grh);
 838
 839 int rdma_addr_find_smac_by_sgid(union ib_gid *sgid, u8 *smac, u16 *vlan_id)
 840 {
 841         int ret = 0;
 842         struct rdma_dev_addr dev_addr;
 843         union {
 844                 struct sockaddr     _sockaddr;
 845                 struct sockaddr_in  _sockaddr_in;
 846                 struct sockaddr_in6 _sockaddr_in6;
 847         } gid_addr;
 848
 849         rdma_gid2ip(&gid_addr._sockaddr, sgid);
 850
 851         memset(&dev_addr, 0, sizeof(dev_addr));
 852         dev_addr.net = &init_net;
 853         ret = rdma_translate_ip(&gid_addr._sockaddr, &dev_addr, vlan_id);
 854         if (ret)
 855                 return ret;
 856
 857         memcpy(smac, dev_addr.src_dev_addr, ETH_ALEN);
 858         return ret;
 859 }
 860 EXPORT_SYMBOL(rdma_addr_find_smac_by_sgid);
 861
 862 static int netevent_callback(struct notifier_block *self, unsigned long event,
 863         void *ctx)
 864 {
 865         if (event == NETEVENT_NEIGH_UPDATE) {
 866                 struct neighbour *neigh = ctx;
 867
 868                 if (neigh->nud_state & NUD_VALID)
 869                         set_timeout(&work, jiffies);
 870         }
 871         return 0;
 872 }
 873
 874 static struct notifier_block nb = {
 875         .notifier_call = netevent_callback
 876 };
 877
 878 int addr_init(void)
 879 {
 880         addr_wq = alloc_ordered_workqueue("ib_addr", 0);
 881         if (!addr_wq)
 882                 return -ENOMEM;
 883
 884         register_netevent_notifier(&nb);
 885         rdma_addr_register_client(&self);
 886
 887         return 0;
 888 }
 889
 890 void addr_cleanup(void)
 891 {
 892         rdma_addr_unregister_client(&self);
 893         unregister_netevent_notifier(&nb);
 894         destroy_workqueue(addr_wq);
 895 }