drivers/infiniband/core/cma.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) 2005-2006 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/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/idr.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/slab.h>
  44 #include <linux/module.h>
  45 #include <net/route.h>
  46
  47 #include <net/tcp.h>
  48 #include <net/ipv6.h>
  49
  50 #include <rdma/rdma_cm.h>
  51 #include <rdma/rdma_cm_ib.h>
  52 #include <rdma/rdma_netlink.h>
  53 #include <rdma/ib.h>
  54 #include <rdma/ib_cache.h>
  55 #include <rdma/ib_cm.h>
  56 #include <rdma/ib_sa.h>
  57 #include <rdma/iw_cm.h>
  58
  59 MODULE_AUTHOR("Sean Hefty");
  60 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 #define CMA_CM_RESPONSE_TIMEOUT 20
  64 #define CMA_MAX_CM_RETRIES 15
  65 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  66 #define CMA_IBOE_PACKET_LIFETIME 18
  67
  68 static void cma_add_one(struct ib_device *device);
  69 static void cma_remove_one(struct ib_device *device);
  70
  71 static struct ib_client cma_client = {
  72         .name   = "cma",
  73         .add    = cma_add_one,
  74         .remove = cma_remove_one
  75 };
  76
  77 static struct ib_sa_client sa_client;
  78 static struct rdma_addr_client addr_client;
  79 static LIST_HEAD(dev_list);
  80 static LIST_HEAD(listen_any_list);
  81 static DEFINE_MUTEX(lock);
  82 static struct workqueue_struct *cma_wq;
  83 static DEFINE_IDR(tcp_ps);
  84 static DEFINE_IDR(udp_ps);
  85 static DEFINE_IDR(ipoib_ps);
  86 static DEFINE_IDR(ib_ps);
  87
  88 struct cma_device {
  89         struct list_head        list;
  90         struct ib_device        *device;
  91         struct completion       comp;
  92         atomic_t                refcount;
  93         struct list_head        id_list;
  94 };
  95
  96 struct rdma_bind_list {
  97         struct idr              *ps;
  98         struct hlist_head       owners;
  99         unsigned short          port;
 100 };
 101
 102 enum {
 103         CMA_OPTION_AFONLY,
 104 };
 105
 106 /*
 107  * Device removal can occur at anytime, so we need extra handling to
 108  * serialize notifying the user of device removal with other callbacks.
 109  * We do this by disabling removal notification while a callback is in process,
 110  * and reporting it after the callback completes.
 111  */
 112 struct rdma_id_private {
 113         struct rdma_cm_id       id;
 114
 115         struct rdma_bind_list   *bind_list;
 116         struct hlist_node       node;
 117         struct list_head        list; /* listen_any_list or cma_device.list */
 118         struct list_head        listen_list; /* per device listens */
 119         struct cma_device       *cma_dev;
 120         struct list_head        mc_list;
 121
 122         int                     internal_id;
 123         enum rdma_cm_state      state;
 124         spinlock_t              lock;
 125         struct mutex            qp_mutex;
 126
 127         struct completion       comp;
 128         atomic_t                refcount;
 129         struct mutex            handler_mutex;
 130
 131         int                     backlog;
 132         int                     timeout_ms;
 133         struct ib_sa_query      *query;
 134         int                     query_id;
 135         union {
 136                 struct ib_cm_id *ib;
 137                 struct iw_cm_id *iw;
 138         } cm_id;
 139
 140         u32                     seq_num;
 141         u32                     qkey;
 142         u32                     qp_num;
 143         pid_t                   owner;
 144         u32                     options;
 145         u8                      srq;
 146         u8                      tos;
 147         u8                      reuseaddr;
 148         u8                      afonly;
 149 };
 150
 151 struct cma_multicast {
 152         struct rdma_id_private *id_priv;
 153         union {
 154                 struct ib_sa_multicast *ib;
 155         } multicast;
 156         struct list_head        list;
 157         void                    *context;
 158         struct sockaddr_storage addr;
 159         struct kref             mcref;
 160 };
 161
 162 struct cma_work {
 163         struct work_struct      work;
 164         struct rdma_id_private  *id;
 165         enum rdma_cm_state      old_state;
 166         enum rdma_cm_state      new_state;
 167         struct rdma_cm_event    event;
 168 };
 169
 170 struct cma_ndev_work {
 171         struct work_struct      work;
 172         struct rdma_id_private  *id;
 173         struct rdma_cm_event    event;
 174 };
 175
 176 struct iboe_mcast_work {
 177         struct work_struct       work;
 178         struct rdma_id_private  *id;
 179         struct cma_multicast    *mc;
 180 };
 181
 182 union cma_ip_addr {
 183         struct in6_addr ip6;
 184         struct {
 185                 __be32 pad[3];
 186                 __be32 addr;
 187         } ip4;
 188 };
 189
 190 struct cma_hdr {
 191         u8 cma_version;
 192         u8 ip_version;  /* IP version: 7:4 */
 193         __be16 port;
 194         union cma_ip_addr src_addr;
 195         union cma_ip_addr dst_addr;
 196 };
 197
 198 #define CMA_VERSION 0x00
 199
 200 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 201 {
 202         unsigned long flags;
 203         int ret;
 204
 205         spin_lock_irqsave(&id_priv->lock, flags);
 206         ret = (id_priv->state == comp);
 207         spin_unlock_irqrestore(&id_priv->lock, flags);
 208         return ret;
 209 }
 210
 211 static int cma_comp_exch(struct rdma_id_private *id_priv,
 212                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 213 {
 214         unsigned long flags;
 215         int ret;
 216
 217         spin_lock_irqsave(&id_priv->lock, flags);
 218         if ((ret = (id_priv->state == comp)))
 219                 id_priv->state = exch;
 220         spin_unlock_irqrestore(&id_priv->lock, flags);
 221         return ret;
 222 }
 223
 224 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 225                                    enum rdma_cm_state exch)
 226 {
 227         unsigned long flags;
 228         enum rdma_cm_state old;
 229
 230         spin_lock_irqsave(&id_priv->lock, flags);
 231         old = id_priv->state;
 232         id_priv->state = exch;
 233         spin_unlock_irqrestore(&id_priv->lock, flags);
 234         return old;
 235 }
 236
 237 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 238 {
 239         return hdr->ip_version >> 4;
 240 }
 241
 242 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 243 {
 244         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 245 }
 246
 247 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 248                               struct cma_device *cma_dev)
 249 {
 250         atomic_inc(&cma_dev->refcount);
 251         id_priv->cma_dev = cma_dev;
 252         id_priv->id.device = cma_dev->device;
 253         id_priv->id.route.addr.dev_addr.transport =
 254                 rdma_node_get_transport(cma_dev->device->node_type);
 255         list_add_tail(&id_priv->list, &cma_dev->id_list);
 256 }
 257
 258 static inline void cma_deref_dev(struct cma_device *cma_dev)
 259 {
 260         if (atomic_dec_and_test(&cma_dev->refcount))
 261                 complete(&cma_dev->comp);
 262 }
 263
 264 static inline void release_mc(struct kref *kref)
 265 {
 266         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 267
 268         kfree(mc->multicast.ib);
 269         kfree(mc);
 270 }
 271
 272 static void cma_release_dev(struct rdma_id_private *id_priv)
 273 {
 274         mutex_lock(&lock);
 275         list_del(&id_priv->list);
 276         cma_deref_dev(id_priv->cma_dev);
 277         id_priv->cma_dev = NULL;
 278         mutex_unlock(&lock);
 279 }
 280
 281 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 282 {
 283         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 284 }
 285
 286 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 287 {
 288         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 289 }
 290
 291 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 292 {
 293         return id_priv->id.route.addr.src_addr.ss_family;
 294 }
 295
 296 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 297 {
 298         struct ib_sa_mcmember_rec rec;
 299         int ret = 0;
 300
 301         if (id_priv->qkey) {
 302                 if (qkey && id_priv->qkey != qkey)
 303                         return -EINVAL;
 304                 return 0;
 305         }
 306
 307         if (qkey) {
 308                 id_priv->qkey = qkey;
 309                 return 0;
 310         }
 311
 312         switch (id_priv->id.ps) {
 313         case RDMA_PS_UDP:
 314         case RDMA_PS_IB:
 315                 id_priv->qkey = RDMA_UDP_QKEY;
 316                 break;
 317         case RDMA_PS_IPOIB:
 318                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 319                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 320                                              id_priv->id.port_num, &rec.mgid,
 321                                              &rec);
 322                 if (!ret)
 323                         id_priv->qkey = be32_to_cpu(rec.qkey);
 324                 break;
 325         default:
 326                 break;
 327         }
 328         return ret;
 329 }
 330
 331 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 332 {
 333         dev_addr->dev_type = ARPHRD_INFINIBAND;
 334         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 335         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 336 }
 337
 338 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 339 {
 340         int ret;
 341
 342         if (addr->sa_family != AF_IB) {
 343                 ret = rdma_translate_ip(addr, dev_addr);
 344         } else {
 345                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 346                 ret = 0;
 347         }
 348
 349         return ret;
 350 }
 351
 352 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 353 {
 354         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 355         struct cma_device *cma_dev;
 356         union ib_gid gid, iboe_gid;
 357         int ret = -ENODEV;
 358         u8 port, found_port;
 359         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 360                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 361
 362         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 363             id_priv->id.ps == RDMA_PS_IPOIB)
 364                 return -EINVAL;
 365
 366         mutex_lock(&lock);
 367         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 368         memcpy(&gid, dev_addr->src_dev_addr +
 369                rdma_addr_gid_offset(dev_addr), sizeof gid);
 370         list_for_each_entry(cma_dev, &dev_list, list) {
 371                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port)
 372                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 373                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 374                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 375                                         ret = ib_find_cached_gid(cma_dev->device, &iboe_gid, &found_port, NULL);
 376                                 else
 377                                         ret = ib_find_cached_gid(cma_dev->device, &gid, &found_port, NULL);
 378
 379                                 if (!ret && (port == found_port)) {
 380                                         id_priv->id.port_num = found_port;
 381                                         goto out;
 382                                 }
 383                         }
 384         }
 385
 386 out:
 387         if (!ret)
 388                 cma_attach_to_dev(id_priv, cma_dev);
 389
 390         mutex_unlock(&lock);
 391         return ret;
 392 }
 393
 394 /*
 395  * Select the source IB device and address to reach the destination IB address.
 396  */
 397 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 398 {
 399         struct cma_device *cma_dev, *cur_dev;
 400         struct sockaddr_ib *addr;
 401         union ib_gid gid, sgid, *dgid;
 402         u16 pkey, index;
 403         u8 p;
 404         int i;
 405
 406         cma_dev = NULL;
 407         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 408         dgid = (union ib_gid *) &addr->sib_addr;
 409         pkey = ntohs(addr->sib_pkey);
 410
 411         list_for_each_entry(cur_dev, &dev_list, list) {
 412                 if (rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
 413                         continue;
 414
 415                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
 416                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 417                                 continue;
 418
 419                         for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, &gid); i++) {
 420                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 421                                         cma_dev = cur_dev;
 422                                         sgid = gid;
 423                                         id_priv->id.port_num = p;
 424                                         goto found;
 425                                 }
 426
 427                                 if (!cma_dev && (gid.global.subnet_prefix ==
 428                                                  dgid->global.subnet_prefix)) {
 429                                         cma_dev = cur_dev;
 430                                         sgid = gid;
 431                                         id_priv->id.port_num = p;
 432                                 }
 433                         }
 434                 }
 435         }
 436
 437         if (!cma_dev)
 438                 return -ENODEV;
 439
 440 found:
 441         cma_attach_to_dev(id_priv, cma_dev);
 442         addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
 443         memcpy(&addr->sib_addr, &sgid, sizeof sgid);
 444         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 445         return 0;
 446 }
 447
 448 static void cma_deref_id(struct rdma_id_private *id_priv)
 449 {
 450         if (atomic_dec_and_test(&id_priv->refcount))
 451                 complete(&id_priv->comp);
 452 }
 453
 454 static int cma_disable_callback(struct rdma_id_private *id_priv,
 455                                 enum rdma_cm_state state)
 456 {
 457         mutex_lock(&id_priv->handler_mutex);
 458         if (id_priv->state != state) {
 459                 mutex_unlock(&id_priv->handler_mutex);
 460                 return -EINVAL;
 461         }
 462         return 0;
 463 }
 464
 465 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 466                                   void *context, enum rdma_port_space ps,
 467                                   enum ib_qp_type qp_type)
 468 {
 469         struct rdma_id_private *id_priv;
 470
 471         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 472         if (!id_priv)
 473                 return ERR_PTR(-ENOMEM);
 474
 475         id_priv->owner = task_pid_nr(current);
 476         id_priv->state = RDMA_CM_IDLE;
 477         id_priv->id.context = context;
 478         id_priv->id.event_handler = event_handler;
 479         id_priv->id.ps = ps;
 480         id_priv->id.qp_type = qp_type;
 481         spin_lock_init(&id_priv->lock);
 482         mutex_init(&id_priv->qp_mutex);
 483         init_completion(&id_priv->comp);
 484         atomic_set(&id_priv->refcount, 1);
 485         mutex_init(&id_priv->handler_mutex);
 486         INIT_LIST_HEAD(&id_priv->listen_list);
 487         INIT_LIST_HEAD(&id_priv->mc_list);
 488         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 489
 490         return &id_priv->id;
 491 }
 492 EXPORT_SYMBOL(rdma_create_id);
 493
 494 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 495 {
 496         struct ib_qp_attr qp_attr;
 497         int qp_attr_mask, ret;
 498
 499         qp_attr.qp_state = IB_QPS_INIT;
 500         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 501         if (ret)
 502                 return ret;
 503
 504         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 505         if (ret)
 506                 return ret;
 507
 508         qp_attr.qp_state = IB_QPS_RTR;
 509         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 510         if (ret)
 511                 return ret;
 512
 513         qp_attr.qp_state = IB_QPS_RTS;
 514         qp_attr.sq_psn = 0;
 515         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 516
 517         return ret;
 518 }
 519
 520 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 521 {
 522         struct ib_qp_attr qp_attr;
 523         int qp_attr_mask, ret;
 524
 525         qp_attr.qp_state = IB_QPS_INIT;
 526         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 527         if (ret)
 528                 return ret;
 529
 530         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 531 }
 532
 533 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 534                    struct ib_qp_init_attr *qp_init_attr)
 535 {
 536         struct rdma_id_private *id_priv;
 537         struct ib_qp *qp;
 538         int ret;
 539
 540         id_priv = container_of(id, struct rdma_id_private, id);
 541         if (id->device != pd->device)
 542                 return -EINVAL;
 543
 544         qp = ib_create_qp(pd, qp_init_attr);
 545         if (IS_ERR(qp))
 546                 return PTR_ERR(qp);
 547
 548         if (id->qp_type == IB_QPT_UD)
 549                 ret = cma_init_ud_qp(id_priv, qp);
 550         else
 551                 ret = cma_init_conn_qp(id_priv, qp);
 552         if (ret)
 553                 goto err;
 554
 555         id->qp = qp;
 556         id_priv->qp_num = qp->qp_num;
 557         id_priv->srq = (qp->srq != NULL);
 558         return 0;
 559 err:
 560         ib_destroy_qp(qp);
 561         return ret;
 562 }
 563 EXPORT_SYMBOL(rdma_create_qp);
 564
 565 void rdma_destroy_qp(struct rdma_cm_id *id)
 566 {
 567         struct rdma_id_private *id_priv;
 568
 569         id_priv = container_of(id, struct rdma_id_private, id);
 570         mutex_lock(&id_priv->qp_mutex);
 571         ib_destroy_qp(id_priv->id.qp);
 572         id_priv->id.qp = NULL;
 573         mutex_unlock(&id_priv->qp_mutex);
 574 }
 575 EXPORT_SYMBOL(rdma_destroy_qp);
 576
 577 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 578                              struct rdma_conn_param *conn_param)
 579 {
 580         struct ib_qp_attr qp_attr;
 581         int qp_attr_mask, ret;
 582
 583         mutex_lock(&id_priv->qp_mutex);
 584         if (!id_priv->id.qp) {
 585                 ret = 0;
 586                 goto out;
 587         }
 588
 589         /* Need to update QP attributes from default values. */
 590         qp_attr.qp_state = IB_QPS_INIT;
 591         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 592         if (ret)
 593                 goto out;
 594
 595         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 596         if (ret)
 597                 goto out;
 598
 599         qp_attr.qp_state = IB_QPS_RTR;
 600         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 601         if (ret)
 602                 goto out;
 603
 604         if (conn_param)
 605                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 606         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 607 out:
 608         mutex_unlock(&id_priv->qp_mutex);
 609         return ret;
 610 }
 611
 612 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 613                              struct rdma_conn_param *conn_param)
 614 {
 615         struct ib_qp_attr qp_attr;
 616         int qp_attr_mask, ret;
 617
 618         mutex_lock(&id_priv->qp_mutex);
 619         if (!id_priv->id.qp) {
 620                 ret = 0;
 621                 goto out;
 622         }
 623
 624         qp_attr.qp_state = IB_QPS_RTS;
 625         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 626         if (ret)
 627                 goto out;
 628
 629         if (conn_param)
 630                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 631         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 632 out:
 633         mutex_unlock(&id_priv->qp_mutex);
 634         return ret;
 635 }
 636
 637 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 638 {
 639         struct ib_qp_attr qp_attr;
 640         int ret;
 641
 642         mutex_lock(&id_priv->qp_mutex);
 643         if (!id_priv->id.qp) {
 644                 ret = 0;
 645                 goto out;
 646         }
 647
 648         qp_attr.qp_state = IB_QPS_ERR;
 649         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 650 out:
 651         mutex_unlock(&id_priv->qp_mutex);
 652         return ret;
 653 }
 654
 655 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 656                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 657 {
 658         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 659         int ret;
 660         u16 pkey;
 661
 662         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 663             IB_LINK_LAYER_INFINIBAND)
 664                 pkey = ib_addr_get_pkey(dev_addr);
 665         else
 666                 pkey = 0xffff;
 667
 668         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 669                                   pkey, &qp_attr->pkey_index);
 670         if (ret)
 671                 return ret;
 672
 673         qp_attr->port_num = id_priv->id.port_num;
 674         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 675
 676         if (id_priv->id.qp_type == IB_QPT_UD) {
 677                 ret = cma_set_qkey(id_priv, 0);
 678                 if (ret)
 679                         return ret;
 680
 681                 qp_attr->qkey = id_priv->qkey;
 682                 *qp_attr_mask |= IB_QP_QKEY;
 683         } else {
 684                 qp_attr->qp_access_flags = 0;
 685                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 686         }
 687         return 0;
 688 }
 689
 690 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 691                        int *qp_attr_mask)
 692 {
 693         struct rdma_id_private *id_priv;
 694         int ret = 0;
 695
 696         id_priv = container_of(id, struct rdma_id_private, id);
 697         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 698         case RDMA_TRANSPORT_IB:
 699                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 700                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 701                 else
 702                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 703                                                  qp_attr_mask);
 704                 if (qp_attr->qp_state == IB_QPS_RTR)
 705                         qp_attr->rq_psn = id_priv->seq_num;
 706                 break;
 707         case RDMA_TRANSPORT_IWARP:
 708                 if (!id_priv->cm_id.iw) {
 709                         qp_attr->qp_access_flags = 0;
 710                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 711                 } else
 712                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 713                                                  qp_attr_mask);
 714                 break;
 715         default:
 716                 ret = -ENOSYS;
 717                 break;
 718         }
 719
 720         return ret;
 721 }
 722 EXPORT_SYMBOL(rdma_init_qp_attr);
 723
 724 static inline int cma_zero_addr(struct sockaddr *addr)
 725 {
 726         switch (addr->sa_family) {
 727         case AF_INET:
 728                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
 729         case AF_INET6:
 730                 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
 731         case AF_IB:
 732                 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
 733         default:
 734                 return 0;
 735         }
 736 }
 737
 738 static inline int cma_loopback_addr(struct sockaddr *addr)
 739 {
 740         switch (addr->sa_family) {
 741         case AF_INET:
 742                 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 743         case AF_INET6:
 744                 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
 745         case AF_IB:
 746                 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
 747         default:
 748                 return 0;
 749         }
 750 }
 751
 752 static inline int cma_any_addr(struct sockaddr *addr)
 753 {
 754         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 755 }
 756
 757 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 758 {
 759         if (src->sa_family != dst->sa_family)
 760                 return -1;
 761
 762         switch (src->sa_family) {
 763         case AF_INET:
 764                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 765                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 766         case AF_INET6:
 767                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 768                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 769         default:
 770                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
 771                                    &((struct sockaddr_ib *) dst)->sib_addr);
 772         }
 773 }
 774
 775 static __be16 cma_port(struct sockaddr *addr)
 776 {
 777         struct sockaddr_ib *sib;
 778
 779         switch (addr->sa_family) {
 780         case AF_INET:
 781                 return ((struct sockaddr_in *) addr)->sin_port;
 782         case AF_INET6:
 783                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 784         case AF_IB:
 785                 sib = (struct sockaddr_ib *) addr;
 786                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
 787                                     be64_to_cpu(sib->sib_sid_mask)));
 788         default:
 789                 return 0;
 790         }
 791 }
 792
 793 static inline int cma_any_port(struct sockaddr *addr)
 794 {
 795         return !cma_port(addr);
 796 }
 797
 798 static void cma_save_ib_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
 799                              struct ib_sa_path_rec *path)
 800 {
 801         struct sockaddr_ib *listen_ib, *ib;
 802
 803         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
 804         ib = (struct sockaddr_ib *) &id->route.addr.src_addr;
 805         ib->sib_family = listen_ib->sib_family;
 806         ib->sib_pkey = path->pkey;
 807         ib->sib_flowinfo = path->flow_label;
 808         memcpy(&ib->sib_addr, &path->sgid, 16);
 809         ib->sib_sid = listen_ib->sib_sid;
 810         ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
 811         ib->sib_scope_id = listen_ib->sib_scope_id;
 812
 813         ib = (struct sockaddr_ib *) &id->route.addr.dst_addr;
 814         ib->sib_family = listen_ib->sib_family;
 815         ib->sib_pkey = path->pkey;
 816         ib->sib_flowinfo = path->flow_label;
 817         memcpy(&ib->sib_addr, &path->dgid, 16);
 818 }
 819
 820 static void cma_save_ip4_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
 821                               struct cma_hdr *hdr)
 822 {
 823         struct sockaddr_in *listen4, *ip4;
 824
 825         listen4 = (struct sockaddr_in *) &listen_id->route.addr.src_addr;
 826         ip4 = (struct sockaddr_in *) &id->route.addr.src_addr;
 827         ip4->sin_family = listen4->sin_family;
 828         ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
 829         ip4->sin_port = listen4->sin_port;
 830
 831         ip4 = (struct sockaddr_in *) &id->route.addr.dst_addr;
 832         ip4->sin_family = listen4->sin_family;
 833         ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
 834         ip4->sin_port = hdr->port;
 835 }
 836
 837 static void cma_save_ip6_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
 838                               struct cma_hdr *hdr)
 839 {
 840         struct sockaddr_in6 *listen6, *ip6;
 841
 842         listen6 = (struct sockaddr_in6 *) &listen_id->route.addr.src_addr;
 843         ip6 = (struct sockaddr_in6 *) &id->route.addr.src_addr;
 844         ip6->sin6_family = listen6->sin6_family;
 845         ip6->sin6_addr = hdr->dst_addr.ip6;
 846         ip6->sin6_port = listen6->sin6_port;
 847
 848         ip6 = (struct sockaddr_in6 *) &id->route.addr.dst_addr;
 849         ip6->sin6_family = listen6->sin6_family;
 850         ip6->sin6_addr = hdr->src_addr.ip6;
 851         ip6->sin6_port = hdr->port;
 852 }
 853
 854 static int cma_save_net_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
 855                              struct ib_cm_event *ib_event)
 856 {
 857         struct cma_hdr *hdr;
 858
 859         if ((listen_id->route.addr.src_addr.ss_family == AF_IB) &&
 860             (ib_event->event == IB_CM_REQ_RECEIVED)) {
 861                 cma_save_ib_info(id, listen_id, ib_event->param.req_rcvd.primary_path);
 862                 return 0;
 863         }
 864
 865         hdr = ib_event->private_data;
 866         if (hdr->cma_version != CMA_VERSION)
 867                 return -EINVAL;
 868
 869         switch (cma_get_ip_ver(hdr)) {
 870         case 4:
 871                 cma_save_ip4_info(id, listen_id, hdr);
 872                 break;
 873         case 6:
 874                 cma_save_ip6_info(id, listen_id, hdr);
 875                 break;
 876         default:
 877                 return -EINVAL;
 878         }
 879         return 0;
 880 }
 881
 882 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
 883 {
 884         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
 885 }
 886
 887 static void cma_cancel_route(struct rdma_id_private *id_priv)
 888 {
 889         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 890         case IB_LINK_LAYER_INFINIBAND:
 891                 if (id_priv->query)
 892                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 893                 break;
 894         default:
 895                 break;
 896         }
 897 }
 898
 899 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 900 {
 901         struct rdma_id_private *dev_id_priv;
 902
 903         /*
 904          * Remove from listen_any_list to prevent added devices from spawning
 905          * additional listen requests.
 906          */
 907         mutex_lock(&lock);
 908         list_del(&id_priv->list);
 909
 910         while (!list_empty(&id_priv->listen_list)) {
 911                 dev_id_priv = list_entry(id_priv->listen_list.next,
 912                                          struct rdma_id_private, listen_list);
 913                 /* sync with device removal to avoid duplicate destruction */
 914                 list_del_init(&dev_id_priv->list);
 915                 list_del(&dev_id_priv->listen_list);
 916                 mutex_unlock(&lock);
 917
 918                 rdma_destroy_id(&dev_id_priv->id);
 919                 mutex_lock(&lock);
 920         }
 921         mutex_unlock(&lock);
 922 }
 923
 924 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 925                                  enum rdma_cm_state state)
 926 {
 927         switch (state) {
 928         case RDMA_CM_ADDR_QUERY:
 929                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 930                 break;
 931         case RDMA_CM_ROUTE_QUERY:
 932                 cma_cancel_route(id_priv);
 933                 break;
 934         case RDMA_CM_LISTEN:
 935                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
 936                         cma_cancel_listens(id_priv);
 937                 break;
 938         default:
 939                 break;
 940         }
 941 }
 942
 943 static void cma_release_port(struct rdma_id_private *id_priv)
 944 {
 945         struct rdma_bind_list *bind_list = id_priv->bind_list;
 946
 947         if (!bind_list)
 948                 return;
 949
 950         mutex_lock(&lock);
 951         hlist_del(&id_priv->node);
 952         if (hlist_empty(&bind_list->owners)) {
 953                 idr_remove(bind_list->ps, bind_list->port);
 954                 kfree(bind_list);
 955         }
 956         mutex_unlock(&lock);
 957 }
 958
 959 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 960 {
 961         struct cma_multicast *mc;
 962
 963         while (!list_empty(&id_priv->mc_list)) {
 964                 mc = container_of(id_priv->mc_list.next,
 965                                   struct cma_multicast, list);
 966                 list_del(&mc->list);
 967                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 968                 case IB_LINK_LAYER_INFINIBAND:
 969                         ib_sa_free_multicast(mc->multicast.ib);
 970                         kfree(mc);
 971                         break;
 972                 case IB_LINK_LAYER_ETHERNET:
 973                         kref_put(&mc->mcref, release_mc);
 974                         break;
 975                 default:
 976                         break;
 977                 }
 978         }
 979 }
 980
 981 void rdma_destroy_id(struct rdma_cm_id *id)
 982 {
 983         struct rdma_id_private *id_priv;
 984         enum rdma_cm_state state;
 985
 986         id_priv = container_of(id, struct rdma_id_private, id);
 987         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 988         cma_cancel_operation(id_priv, state);
 989
 990         /*
 991          * Wait for any active callback to finish.  New callbacks will find
 992          * the id_priv state set to destroying and abort.
 993          */
 994         mutex_lock(&id_priv->handler_mutex);
 995         mutex_unlock(&id_priv->handler_mutex);
 996
 997         if (id_priv->cma_dev) {
 998                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 999                 case RDMA_TRANSPORT_IB:
1000                         if (id_priv->cm_id.ib)
1001                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1002                         break;
1003                 case RDMA_TRANSPORT_IWARP:
1004                         if (id_priv->cm_id.iw)
1005                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1006                         break;
1007                 default:
1008                         break;
1009                 }
1010                 cma_leave_mc_groups(id_priv);
1011                 cma_release_dev(id_priv);
1012         }
1013
1014         cma_release_port(id_priv);
1015         cma_deref_id(id_priv);
1016         wait_for_completion(&id_priv->comp);
1017
1018         if (id_priv->internal_id)
1019                 cma_deref_id(id_priv->id.context);
1020
1021         kfree(id_priv->id.route.path_rec);
1022         kfree(id_priv);
1023 }
1024 EXPORT_SYMBOL(rdma_destroy_id);
1025
1026 static int cma_rep_recv(struct rdma_id_private *id_priv)
1027 {
1028         int ret;
1029
1030         ret = cma_modify_qp_rtr(id_priv, NULL);
1031         if (ret)
1032                 goto reject;
1033
1034         ret = cma_modify_qp_rts(id_priv, NULL);
1035         if (ret)
1036                 goto reject;
1037
1038         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1039         if (ret)
1040                 goto reject;
1041
1042         return 0;
1043 reject:
1044         cma_modify_qp_err(id_priv);
1045         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1046                        NULL, 0, NULL, 0);
1047         return ret;
1048 }
1049
1050 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1051                                    struct ib_cm_rep_event_param *rep_data,
1052                                    void *private_data)
1053 {
1054         event->param.conn.private_data = private_data;
1055         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1056         event->param.conn.responder_resources = rep_data->responder_resources;
1057         event->param.conn.initiator_depth = rep_data->initiator_depth;
1058         event->param.conn.flow_control = rep_data->flow_control;
1059         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1060         event->param.conn.srq = rep_data->srq;
1061         event->param.conn.qp_num = rep_data->remote_qpn;
1062 }
1063
1064 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1065 {
1066         struct rdma_id_private *id_priv = cm_id->context;
1067         struct rdma_cm_event event;
1068         int ret = 0;
1069
1070         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1071                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1072             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1073                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1074                 return 0;
1075
1076         memset(&event, 0, sizeof event);
1077         switch (ib_event->event) {
1078         case IB_CM_REQ_ERROR:
1079         case IB_CM_REP_ERROR:
1080                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1081                 event.status = -ETIMEDOUT;
1082                 break;
1083         case IB_CM_REP_RECEIVED:
1084                 if (id_priv->id.qp) {
1085                         event.status = cma_rep_recv(id_priv);
1086                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1087                                                      RDMA_CM_EVENT_ESTABLISHED;
1088                 } else {
1089                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1090                 }
1091                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1092                                        ib_event->private_data);
1093                 break;
1094         case IB_CM_RTU_RECEIVED:
1095         case IB_CM_USER_ESTABLISHED:
1096                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1097                 break;
1098         case IB_CM_DREQ_ERROR:
1099                 event.status = -ETIMEDOUT; /* fall through */
1100         case IB_CM_DREQ_RECEIVED:
1101         case IB_CM_DREP_RECEIVED:
1102                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1103                                    RDMA_CM_DISCONNECT))
1104                         goto out;
1105                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1106                 break;
1107         case IB_CM_TIMEWAIT_EXIT:
1108                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1109                 break;
1110         case IB_CM_MRA_RECEIVED:
1111                 /* ignore event */
1112                 goto out;
1113         case IB_CM_REJ_RECEIVED:
1114                 cma_modify_qp_err(id_priv);
1115                 event.status = ib_event->param.rej_rcvd.reason;
1116                 event.event = RDMA_CM_EVENT_REJECTED;
1117                 event.param.conn.private_data = ib_event->private_data;
1118                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1119                 break;
1120         default:
1121                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1122                        ib_event->event);
1123                 goto out;
1124         }
1125
1126         ret = id_priv->id.event_handler(&id_priv->id, &event);
1127         if (ret) {
1128                 /* Destroy the CM ID by returning a non-zero value. */
1129                 id_priv->cm_id.ib = NULL;
1130                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1131                 mutex_unlock(&id_priv->handler_mutex);
1132                 rdma_destroy_id(&id_priv->id);
1133                 return ret;
1134         }
1135 out:
1136         mutex_unlock(&id_priv->handler_mutex);
1137         return ret;
1138 }
1139
1140 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1141                                                struct ib_cm_event *ib_event)
1142 {
1143         struct rdma_id_private *id_priv;
1144         struct rdma_cm_id *id;
1145         struct rdma_route *rt;
1146         int ret;
1147
1148         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1149                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1150         if (IS_ERR(id))
1151                 return NULL;
1152
1153         id_priv = container_of(id, struct rdma_id_private, id);
1154         if (cma_save_net_info(id, listen_id, ib_event))
1155                 goto err;
1156
1157         rt = &id->route;
1158         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1159         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1160                                GFP_KERNEL);
1161         if (!rt->path_rec)
1162                 goto err;
1163
1164         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1165         if (rt->num_paths == 2)
1166                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1167
1168         if (cma_any_addr(cma_src_addr(id_priv))) {
1169                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1170                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1171                 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1172         } else {
1173                 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1174                 if (ret)
1175                         goto err;
1176         }
1177         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1178
1179         id_priv->state = RDMA_CM_CONNECT;
1180         return id_priv;
1181
1182 err:
1183         rdma_destroy_id(id);
1184         return NULL;
1185 }
1186
1187 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1188                                               struct ib_cm_event *ib_event)
1189 {
1190         struct rdma_id_private *id_priv;
1191         struct rdma_cm_id *id;
1192         int ret;
1193
1194         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1195                             listen_id->ps, IB_QPT_UD);
1196         if (IS_ERR(id))
1197                 return NULL;
1198
1199         id_priv = container_of(id, struct rdma_id_private, id);
1200         if (cma_save_net_info(id, listen_id, ib_event))
1201                 goto err;
1202
1203         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1204                 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1205                 if (ret)
1206                         goto err;
1207         }
1208
1209         id_priv->state = RDMA_CM_CONNECT;
1210         return id_priv;
1211 err:
1212         rdma_destroy_id(id);
1213         return NULL;
1214 }
1215
1216 static void cma_set_req_event_data(struct rdma_cm_event *event,
1217                                    struct ib_cm_req_event_param *req_data,
1218                                    void *private_data, int offset)
1219 {
1220         event->param.conn.private_data = private_data + offset;
1221         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1222         event->param.conn.responder_resources = req_data->responder_resources;
1223         event->param.conn.initiator_depth = req_data->initiator_depth;
1224         event->param.conn.flow_control = req_data->flow_control;
1225         event->param.conn.retry_count = req_data->retry_count;
1226         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1227         event->param.conn.srq = req_data->srq;
1228         event->param.conn.qp_num = req_data->remote_qpn;
1229 }
1230
1231 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1232 {
1233         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1234                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1235                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1236                  (id->qp_type == IB_QPT_UD)) ||
1237                 (!id->qp_type));
1238 }
1239
1240 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1241 {
1242         struct rdma_id_private *listen_id, *conn_id;
1243         struct rdma_cm_event event;
1244         int offset, ret;
1245
1246         listen_id = cm_id->context;
1247         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1248                 return -EINVAL;
1249
1250         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1251                 return -ECONNABORTED;
1252
1253         memset(&event, 0, sizeof event);
1254         offset = cma_user_data_offset(listen_id);
1255         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1256         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1257                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1258                 event.param.ud.private_data = ib_event->private_data + offset;
1259                 event.param.ud.private_data_len =
1260                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1261         } else {
1262                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1263                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1264                                        ib_event->private_data, offset);
1265         }
1266         if (!conn_id) {
1267                 ret = -ENOMEM;
1268                 goto err1;
1269         }
1270
1271         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1272         ret = cma_acquire_dev(conn_id);
1273         if (ret)
1274                 goto err2;
1275
1276         conn_id->cm_id.ib = cm_id;
1277         cm_id->context = conn_id;
1278         cm_id->cm_handler = cma_ib_handler;
1279
1280         /*
1281          * Protect against the user destroying conn_id from another thread
1282          * until we're done accessing it.
1283          */
1284         atomic_inc(&conn_id->refcount);
1285         ret = conn_id->id.event_handler(&conn_id->id, &event);
1286         if (ret)
1287                 goto err3;
1288
1289         /*
1290          * Acquire mutex to prevent user executing rdma_destroy_id()
1291          * while we're accessing the cm_id.
1292          */
1293         mutex_lock(&lock);
1294         if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1295                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1296         mutex_unlock(&lock);
1297         mutex_unlock(&conn_id->handler_mutex);
1298         mutex_unlock(&listen_id->handler_mutex);
1299         cma_deref_id(conn_id);
1300         return 0;
1301
1302 err3:
1303         cma_deref_id(conn_id);
1304         /* Destroy the CM ID by returning a non-zero value. */
1305         conn_id->cm_id.ib = NULL;
1306 err2:
1307         cma_exch(conn_id, RDMA_CM_DESTROYING);
1308         mutex_unlock(&conn_id->handler_mutex);
1309 err1:
1310         mutex_unlock(&listen_id->handler_mutex);
1311         if (conn_id)
1312                 rdma_destroy_id(&conn_id->id);
1313         return ret;
1314 }
1315
1316 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1317 {
1318         if (addr->sa_family == AF_IB)
1319                 return ((struct sockaddr_ib *) addr)->sib_sid;
1320
1321         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1322 }
1323 EXPORT_SYMBOL(rdma_get_service_id);
1324
1325 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1326                                  struct ib_cm_compare_data *compare)
1327 {
1328         struct cma_hdr *cma_data, *cma_mask;
1329         __be32 ip4_addr;
1330         struct in6_addr ip6_addr;
1331
1332         memset(compare, 0, sizeof *compare);
1333         cma_data = (void *) compare->data;
1334         cma_mask = (void *) compare->mask;
1335
1336         switch (addr->sa_family) {
1337         case AF_INET:
1338                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1339                 cma_set_ip_ver(cma_data, 4);
1340                 cma_set_ip_ver(cma_mask, 0xF);
1341                 if (!cma_any_addr(addr)) {
1342                         cma_data->dst_addr.ip4.addr = ip4_addr;
1343                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1344                 }
1345                 break;
1346         case AF_INET6:
1347                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1348                 cma_set_ip_ver(cma_data, 6);
1349                 cma_set_ip_ver(cma_mask, 0xF);
1350                 if (!cma_any_addr(addr)) {
1351                         cma_data->dst_addr.ip6 = ip6_addr;
1352                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1353                                sizeof cma_mask->dst_addr.ip6);
1354                 }
1355                 break;
1356         default:
1357                 break;
1358         }
1359 }
1360
1361 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1362 {
1363         struct rdma_id_private *id_priv = iw_id->context;
1364         struct rdma_cm_event event;
1365         int ret = 0;
1366         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1367         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1368
1369         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1370                 return 0;
1371
1372         memset(&event, 0, sizeof event);
1373         switch (iw_event->event) {
1374         case IW_CM_EVENT_CLOSE:
1375                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1376                 break;
1377         case IW_CM_EVENT_CONNECT_REPLY:
1378                 memcpy(cma_src_addr(id_priv), laddr,
1379                        rdma_addr_size(laddr));
1380                 memcpy(cma_dst_addr(id_priv), raddr,
1381                        rdma_addr_size(raddr));
1382                 switch (iw_event->status) {
1383                 case 0:
1384                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1385                         event.param.conn.initiator_depth = iw_event->ird;
1386                         event.param.conn.responder_resources = iw_event->ord;
1387                         break;
1388                 case -ECONNRESET:
1389                 case -ECONNREFUSED:
1390                         event.event = RDMA_CM_EVENT_REJECTED;
1391                         break;
1392                 case -ETIMEDOUT:
1393                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1394                         break;
1395                 default:
1396                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1397                         break;
1398                 }
1399                 break;
1400         case IW_CM_EVENT_ESTABLISHED:
1401                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1402                 event.param.conn.initiator_depth = iw_event->ird;
1403                 event.param.conn.responder_resources = iw_event->ord;
1404                 break;
1405         default:
1406                 BUG_ON(1);
1407         }
1408
1409         event.status = iw_event->status;
1410         event.param.conn.private_data = iw_event->private_data;
1411         event.param.conn.private_data_len = iw_event->private_data_len;
1412         ret = id_priv->id.event_handler(&id_priv->id, &event);
1413         if (ret) {
1414                 /* Destroy the CM ID by returning a non-zero value. */
1415                 id_priv->cm_id.iw = NULL;
1416                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1417                 mutex_unlock(&id_priv->handler_mutex);
1418                 rdma_destroy_id(&id_priv->id);
1419                 return ret;
1420         }
1421
1422         mutex_unlock(&id_priv->handler_mutex);
1423         return ret;
1424 }
1425
1426 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1427                                struct iw_cm_event *iw_event)
1428 {
1429         struct rdma_cm_id *new_cm_id;
1430         struct rdma_id_private *listen_id, *conn_id;
1431         struct net_device *dev = NULL;
1432         struct rdma_cm_event event;
1433         int ret;
1434         struct ib_device_attr attr;
1435         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1436         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1437
1438         listen_id = cm_id->context;
1439         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1440                 return -ECONNABORTED;
1441
1442         /* Create a new RDMA id for the new IW CM ID */
1443         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1444                                    listen_id->id.context,
1445                                    RDMA_PS_TCP, IB_QPT_RC);
1446         if (IS_ERR(new_cm_id)) {
1447                 ret = -ENOMEM;
1448                 goto out;
1449         }
1450         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1451         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1452         conn_id->state = RDMA_CM_CONNECT;
1453
1454         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
1455         if (ret) {
1456                 mutex_unlock(&conn_id->handler_mutex);
1457                 rdma_destroy_id(new_cm_id);
1458                 goto out;
1459         }
1460
1461         ret = cma_acquire_dev(conn_id);
1462         if (ret) {
1463                 mutex_unlock(&conn_id->handler_mutex);
1464                 rdma_destroy_id(new_cm_id);
1465                 goto out;
1466         }
1467
1468         conn_id->cm_id.iw = cm_id;
1469         cm_id->context = conn_id;
1470         cm_id->cm_handler = cma_iw_handler;
1471
1472         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
1473         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
1474
1475         ret = ib_query_device(conn_id->id.device, &attr);
1476         if (ret) {
1477                 mutex_unlock(&conn_id->handler_mutex);
1478                 rdma_destroy_id(new_cm_id);
1479                 goto out;
1480         }
1481
1482         memset(&event, 0, sizeof event);
1483         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1484         event.param.conn.private_data = iw_event->private_data;
1485         event.param.conn.private_data_len = iw_event->private_data_len;
1486         event.param.conn.initiator_depth = iw_event->ird;
1487         event.param.conn.responder_resources = iw_event->ord;
1488
1489         /*
1490          * Protect against the user destroying conn_id from another thread
1491          * until we're done accessing it.
1492          */
1493         atomic_inc(&conn_id->refcount);
1494         ret = conn_id->id.event_handler(&conn_id->id, &event);
1495         if (ret) {
1496                 /* User wants to destroy the CM ID */
1497                 conn_id->cm_id.iw = NULL;
1498                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1499                 mutex_unlock(&conn_id->handler_mutex);
1500                 cma_deref_id(conn_id);
1501                 rdma_destroy_id(&conn_id->id);
1502                 goto out;
1503         }
1504
1505         mutex_unlock(&conn_id->handler_mutex);
1506         cma_deref_id(conn_id);
1507
1508 out:
1509         if (dev)
1510                 dev_put(dev);
1511         mutex_unlock(&listen_id->handler_mutex);
1512         return ret;
1513 }
1514
1515 static int cma_ib_listen(struct rdma_id_private *id_priv)
1516 {
1517         struct ib_cm_compare_data compare_data;
1518         struct sockaddr *addr;
1519         struct ib_cm_id *id;
1520         __be64 svc_id;
1521         int ret;
1522
1523         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1524         if (IS_ERR(id))
1525                 return PTR_ERR(id);
1526
1527         id_priv->cm_id.ib = id;
1528
1529         addr = cma_src_addr(id_priv);
1530         svc_id = rdma_get_service_id(&id_priv->id, addr);
1531         if (cma_any_addr(addr) && !id_priv->afonly)
1532                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1533         else {
1534                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1535                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1536         }
1537
1538         if (ret) {
1539                 ib_destroy_cm_id(id_priv->cm_id.ib);
1540                 id_priv->cm_id.ib = NULL;
1541         }
1542
1543         return ret;
1544 }
1545
1546 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1547 {
1548         int ret;
1549         struct iw_cm_id *id;
1550
1551         id = iw_create_cm_id(id_priv->id.device,
1552                              iw_conn_req_handler,
1553                              id_priv);
1554         if (IS_ERR(id))
1555                 return PTR_ERR(id);
1556
1557         id_priv->cm_id.iw = id;
1558
1559         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
1560                rdma_addr_size(cma_src_addr(id_priv)));
1561
1562         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1563
1564         if (ret) {
1565                 iw_destroy_cm_id(id_priv->cm_id.iw);
1566                 id_priv->cm_id.iw = NULL;
1567         }
1568
1569         return ret;
1570 }
1571
1572 static int cma_listen_handler(struct rdma_cm_id *id,
1573                               struct rdma_cm_event *event)
1574 {
1575         struct rdma_id_private *id_priv = id->context;
1576
1577         id->context = id_priv->id.context;
1578         id->event_handler = id_priv->id.event_handler;
1579         return id_priv->id.event_handler(id, event);
1580 }
1581
1582 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1583                               struct cma_device *cma_dev)
1584 {
1585         struct rdma_id_private *dev_id_priv;
1586         struct rdma_cm_id *id;
1587         int ret;
1588
1589         if (cma_family(id_priv) == AF_IB &&
1590             rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB)
1591                 return;
1592
1593         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1594                             id_priv->id.qp_type);
1595         if (IS_ERR(id))
1596                 return;
1597
1598         dev_id_priv = container_of(id, struct rdma_id_private, id);
1599
1600         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1601         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1602                rdma_addr_size(cma_src_addr(id_priv)));
1603
1604         cma_attach_to_dev(dev_id_priv, cma_dev);
1605         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1606         atomic_inc(&id_priv->refcount);
1607         dev_id_priv->internal_id = 1;
1608         dev_id_priv->afonly = id_priv->afonly;
1609
1610         ret = rdma_listen(id, id_priv->backlog);
1611         if (ret)
1612                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1613                        "listening on device %s\n", ret, cma_dev->device->name);
1614 }
1615
1616 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1617 {
1618         struct cma_device *cma_dev;
1619
1620         mutex_lock(&lock);
1621         list_add_tail(&id_priv->list, &listen_any_list);
1622         list_for_each_entry(cma_dev, &dev_list, list)
1623                 cma_listen_on_dev(id_priv, cma_dev);
1624         mutex_unlock(&lock);
1625 }
1626
1627 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1628 {
1629         struct rdma_id_private *id_priv;
1630
1631         id_priv = container_of(id, struct rdma_id_private, id);
1632         id_priv->tos = (u8) tos;
1633 }
1634 EXPORT_SYMBOL(rdma_set_service_type);
1635
1636 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1637                               void *context)
1638 {
1639         struct cma_work *work = context;
1640         struct rdma_route *route;
1641
1642         route = &work->id->id.route;
1643
1644         if (!status) {
1645                 route->num_paths = 1;
1646                 *route->path_rec = *path_rec;
1647         } else {
1648                 work->old_state = RDMA_CM_ROUTE_QUERY;
1649                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1650                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1651                 work->event.status = status;
1652         }
1653
1654         queue_work(cma_wq, &work->work);
1655 }
1656
1657 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1658                               struct cma_work *work)
1659 {
1660         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1661         struct ib_sa_path_rec path_rec;
1662         ib_sa_comp_mask comp_mask;
1663         struct sockaddr_in6 *sin6;
1664         struct sockaddr_ib *sib;
1665
1666         memset(&path_rec, 0, sizeof path_rec);
1667         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1668         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1669         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1670         path_rec.numb_path = 1;
1671         path_rec.reversible = 1;
1672         path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
1673
1674         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1675                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1676                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1677
1678         switch (cma_family(id_priv)) {
1679         case AF_INET:
1680                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1681                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1682                 break;
1683         case AF_INET6:
1684                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1685                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1686                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1687                 break;
1688         case AF_IB:
1689                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1690                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1691                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1692                 break;
1693         }
1694
1695         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1696                                                id_priv->id.port_num, &path_rec,
1697                                                comp_mask, timeout_ms,
1698                                                GFP_KERNEL, cma_query_handler,
1699                                                work, &id_priv->query);
1700
1701         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1702 }
1703
1704 static void cma_work_handler(struct work_struct *_work)
1705 {
1706         struct cma_work *work = container_of(_work, struct cma_work, work);
1707         struct rdma_id_private *id_priv = work->id;
1708         int destroy = 0;
1709
1710         mutex_lock(&id_priv->handler_mutex);
1711         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1712                 goto out;
1713
1714         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1715                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1716                 destroy = 1;
1717         }
1718 out:
1719         mutex_unlock(&id_priv->handler_mutex);
1720         cma_deref_id(id_priv);
1721         if (destroy)
1722                 rdma_destroy_id(&id_priv->id);
1723         kfree(work);
1724 }
1725
1726 static void cma_ndev_work_handler(struct work_struct *_work)
1727 {
1728         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1729         struct rdma_id_private *id_priv = work->id;
1730         int destroy = 0;
1731
1732         mutex_lock(&id_priv->handler_mutex);
1733         if (id_priv->state == RDMA_CM_DESTROYING ||
1734             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1735                 goto out;
1736
1737         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1738                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1739                 destroy = 1;
1740         }
1741
1742 out:
1743         mutex_unlock(&id_priv->handler_mutex);
1744         cma_deref_id(id_priv);
1745         if (destroy)
1746                 rdma_destroy_id(&id_priv->id);
1747         kfree(work);
1748 }
1749
1750 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1751 {
1752         struct rdma_route *route = &id_priv->id.route;
1753         struct cma_work *work;
1754         int ret;
1755
1756         work = kzalloc(sizeof *work, GFP_KERNEL);
1757         if (!work)
1758                 return -ENOMEM;
1759
1760         work->id = id_priv;
1761         INIT_WORK(&work->work, cma_work_handler);
1762         work->old_state = RDMA_CM_ROUTE_QUERY;
1763         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1764         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1765
1766         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1767         if (!route->path_rec) {
1768                 ret = -ENOMEM;
1769                 goto err1;
1770         }
1771
1772         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1773         if (ret)
1774                 goto err2;
1775
1776         return 0;
1777 err2:
1778         kfree(route->path_rec);
1779         route->path_rec = NULL;
1780 err1:
1781         kfree(work);
1782         return ret;
1783 }
1784
1785 int rdma_set_ib_paths(struct rdma_cm_id *id,
1786                       struct ib_sa_path_rec *path_rec, int num_paths)
1787 {
1788         struct rdma_id_private *id_priv;
1789         int ret;
1790
1791         id_priv = container_of(id, struct rdma_id_private, id);
1792         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1793                            RDMA_CM_ROUTE_RESOLVED))
1794                 return -EINVAL;
1795
1796         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1797                                      GFP_KERNEL);
1798         if (!id->route.path_rec) {
1799                 ret = -ENOMEM;
1800                 goto err;
1801         }
1802
1803         id->route.num_paths = num_paths;
1804         return 0;
1805 err:
1806         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1807         return ret;
1808 }
1809 EXPORT_SYMBOL(rdma_set_ib_paths);
1810
1811 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1812 {
1813         struct cma_work *work;
1814
1815         work = kzalloc(sizeof *work, GFP_KERNEL);
1816         if (!work)
1817                 return -ENOMEM;
1818
1819         work->id = id_priv;
1820         INIT_WORK(&work->work, cma_work_handler);
1821         work->old_state = RDMA_CM_ROUTE_QUERY;
1822         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1823         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1824         queue_work(cma_wq, &work->work);
1825         return 0;
1826 }
1827
1828 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1829 {
1830         struct rdma_route *route = &id_priv->id.route;
1831         struct rdma_addr *addr = &route->addr;
1832         struct cma_work *work;
1833         int ret;
1834         struct net_device *ndev = NULL;
1835         u16 vid;
1836
1837         work = kzalloc(sizeof *work, GFP_KERNEL);
1838         if (!work)
1839                 return -ENOMEM;
1840
1841         work->id = id_priv;
1842         INIT_WORK(&work->work, cma_work_handler);
1843
1844         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1845         if (!route->path_rec) {
1846                 ret = -ENOMEM;
1847                 goto err1;
1848         }
1849
1850         route->num_paths = 1;
1851
1852         if (addr->dev_addr.bound_dev_if)
1853                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1854         if (!ndev) {
1855                 ret = -ENODEV;
1856                 goto err2;
1857         }
1858
1859         vid = rdma_vlan_dev_vlan_id(ndev);
1860
1861         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1862         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1863
1864         route->path_rec->hop_limit = 1;
1865         route->path_rec->reversible = 1;
1866         route->path_rec->pkey = cpu_to_be16(0xffff);
1867         route->path_rec->mtu_selector = IB_SA_EQ;
1868         route->path_rec->sl = netdev_get_prio_tc_map(
1869                         ndev->priv_flags & IFF_802_1Q_VLAN ?
1870                                 vlan_dev_real_dev(ndev) : ndev,
1871                         rt_tos2priority(id_priv->tos));
1872
1873         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1874         route->path_rec->rate_selector = IB_SA_EQ;
1875         route->path_rec->rate = iboe_get_rate(ndev);
1876         dev_put(ndev);
1877         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1878         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1879         if (!route->path_rec->mtu) {
1880                 ret = -EINVAL;
1881                 goto err2;
1882         }
1883
1884         work->old_state = RDMA_CM_ROUTE_QUERY;
1885         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1886         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1887         work->event.status = 0;
1888
1889         queue_work(cma_wq, &work->work);
1890
1891         return 0;
1892
1893 err2:
1894         kfree(route->path_rec);
1895         route->path_rec = NULL;
1896 err1:
1897         kfree(work);
1898         return ret;
1899 }
1900
1901 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1902 {
1903         struct rdma_id_private *id_priv;
1904         int ret;
1905
1906         id_priv = container_of(id, struct rdma_id_private, id);
1907         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1908                 return -EINVAL;
1909
1910         atomic_inc(&id_priv->refcount);
1911         switch (rdma_node_get_transport(id->device->node_type)) {
1912         case RDMA_TRANSPORT_IB:
1913                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1914                 case IB_LINK_LAYER_INFINIBAND:
1915                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1916                         break;
1917                 case IB_LINK_LAYER_ETHERNET:
1918                         ret = cma_resolve_iboe_route(id_priv);
1919                         break;
1920                 default:
1921                         ret = -ENOSYS;
1922                 }
1923                 break;
1924         case RDMA_TRANSPORT_IWARP:
1925                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1926                 break;
1927         default:
1928                 ret = -ENOSYS;
1929                 break;
1930         }
1931         if (ret)
1932                 goto err;
1933
1934         return 0;
1935 err:
1936         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1937         cma_deref_id(id_priv);
1938         return ret;
1939 }
1940 EXPORT_SYMBOL(rdma_resolve_route);
1941
1942 static void cma_set_loopback(struct sockaddr *addr)
1943 {
1944         switch (addr->sa_family) {
1945         case AF_INET:
1946                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1947                 break;
1948         case AF_INET6:
1949                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
1950                               0, 0, 0, htonl(1));
1951                 break;
1952         default:
1953                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
1954                             0, 0, 0, htonl(1));
1955                 break;
1956         }
1957 }
1958
1959 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1960 {
1961         struct cma_device *cma_dev, *cur_dev;
1962         struct ib_port_attr port_attr;
1963         union ib_gid gid;
1964         u16 pkey;
1965         int ret;
1966         u8 p;
1967
1968         cma_dev = NULL;
1969         mutex_lock(&lock);
1970         list_for_each_entry(cur_dev, &dev_list, list) {
1971                 if (cma_family(id_priv) == AF_IB &&
1972                     rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
1973                         continue;
1974
1975                 if (!cma_dev)
1976                         cma_dev = cur_dev;
1977
1978                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
1979                         if (!ib_query_port(cur_dev->device, p, &port_attr) &&
1980                             port_attr.state == IB_PORT_ACTIVE) {
1981                                 cma_dev = cur_dev;
1982                                 goto port_found;
1983                         }
1984                 }
1985         }
1986
1987         if (!cma_dev) {
1988                 ret = -ENODEV;
1989                 goto out;
1990         }
1991
1992         p = 1;
1993
1994 port_found:
1995         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1996         if (ret)
1997                 goto out;
1998
1999         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2000         if (ret)
2001                 goto out;
2002
2003         id_priv->id.route.addr.dev_addr.dev_type =
2004                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2005                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2006
2007         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2008         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2009         id_priv->id.port_num = p;
2010         cma_attach_to_dev(id_priv, cma_dev);
2011         cma_set_loopback(cma_src_addr(id_priv));
2012 out:
2013         mutex_unlock(&lock);
2014         return ret;
2015 }
2016
2017 static void addr_handler(int status, struct sockaddr *src_addr,
2018                          struct rdma_dev_addr *dev_addr, void *context)
2019 {
2020         struct rdma_id_private *id_priv = context;
2021         struct rdma_cm_event event;
2022
2023         memset(&event, 0, sizeof event);
2024         mutex_lock(&id_priv->handler_mutex);
2025         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2026                            RDMA_CM_ADDR_RESOLVED))
2027                 goto out;
2028
2029         if (!status && !id_priv->cma_dev)
2030                 status = cma_acquire_dev(id_priv);
2031
2032         if (status) {
2033                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2034                                    RDMA_CM_ADDR_BOUND))
2035                         goto out;
2036                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2037                 event.status = status;
2038         } else {
2039                 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2040                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2041         }
2042
2043         if (id_priv->id.event_handler(&id_priv->id, &event)) {
2044                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2045                 mutex_unlock(&id_priv->handler_mutex);
2046                 cma_deref_id(id_priv);
2047                 rdma_destroy_id(&id_priv->id);
2048                 return;
2049         }
2050 out:
2051         mutex_unlock(&id_priv->handler_mutex);
2052         cma_deref_id(id_priv);
2053 }
2054
2055 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2056 {
2057         struct cma_work *work;
2058         union ib_gid gid;
2059         int ret;
2060
2061         work = kzalloc(sizeof *work, GFP_KERNEL);
2062         if (!work)
2063                 return -ENOMEM;
2064
2065         if (!id_priv->cma_dev) {
2066                 ret = cma_bind_loopback(id_priv);
2067                 if (ret)
2068                         goto err;
2069         }
2070
2071         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2072         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2073
2074         work->id = id_priv;
2075         INIT_WORK(&work->work, cma_work_handler);
2076         work->old_state = RDMA_CM_ADDR_QUERY;
2077         work->new_state = RDMA_CM_ADDR_RESOLVED;
2078         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2079         queue_work(cma_wq, &work->work);
2080         return 0;
2081 err:
2082         kfree(work);
2083         return ret;
2084 }
2085
2086 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2087 {
2088         struct cma_work *work;
2089         int ret;
2090
2091         work = kzalloc(sizeof *work, GFP_KERNEL);
2092         if (!work)
2093                 return -ENOMEM;
2094
2095         if (!id_priv->cma_dev) {
2096                 ret = cma_resolve_ib_dev(id_priv);
2097                 if (ret)
2098                         goto err;
2099         }
2100
2101         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2102                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2103
2104         work->id = id_priv;
2105         INIT_WORK(&work->work, cma_work_handler);
2106         work->old_state = RDMA_CM_ADDR_QUERY;
2107         work->new_state = RDMA_CM_ADDR_RESOLVED;
2108         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2109         queue_work(cma_wq, &work->work);
2110         return 0;
2111 err:
2112         kfree(work);
2113         return ret;
2114 }
2115
2116 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2117                          struct sockaddr *dst_addr)
2118 {
2119         if (!src_addr || !src_addr->sa_family) {
2120                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2121                 src_addr->sa_family = dst_addr->sa_family;
2122                 if (dst_addr->sa_family == AF_INET6) {
2123                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2124                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2125                 } else if (dst_addr->sa_family == AF_IB) {
2126                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
2127                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2128                 }
2129         }
2130         return rdma_bind_addr(id, src_addr);
2131 }
2132
2133 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2134                       struct sockaddr *dst_addr, int timeout_ms)
2135 {
2136         struct rdma_id_private *id_priv;
2137         int ret;
2138
2139         id_priv = container_of(id, struct rdma_id_private, id);
2140         if (id_priv->state == RDMA_CM_IDLE) {
2141                 ret = cma_bind_addr(id, src_addr, dst_addr);
2142                 if (ret)
2143                         return ret;
2144         }
2145
2146         if (cma_family(id_priv) != dst_addr->sa_family)
2147                 return -EINVAL;
2148
2149         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2150                 return -EINVAL;
2151
2152         atomic_inc(&id_priv->refcount);
2153         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2154         if (cma_any_addr(dst_addr)) {
2155                 ret = cma_resolve_loopback(id_priv);
2156         } else {
2157                 if (dst_addr->sa_family == AF_IB) {
2158                         ret = cma_resolve_ib_addr(id_priv);
2159                 } else {
2160                         ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2161                                               dst_addr, &id->route.addr.dev_addr,
2162                                               timeout_ms, addr_handler, id_priv);
2163                 }
2164         }
2165         if (ret)
2166                 goto err;
2167
2168         return 0;
2169 err:
2170         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2171         cma_deref_id(id_priv);
2172         return ret;
2173 }
2174 EXPORT_SYMBOL(rdma_resolve_addr);
2175
2176 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2177 {
2178         struct rdma_id_private *id_priv;
2179         unsigned long flags;
2180         int ret;
2181
2182         id_priv = container_of(id, struct rdma_id_private, id);
2183         spin_lock_irqsave(&id_priv->lock, flags);
2184         if (reuse || id_priv->state == RDMA_CM_IDLE) {
2185                 id_priv->reuseaddr = reuse;
2186                 ret = 0;
2187         } else {
2188                 ret = -EINVAL;
2189         }
2190         spin_unlock_irqrestore(&id_priv->lock, flags);
2191         return ret;
2192 }
2193 EXPORT_SYMBOL(rdma_set_reuseaddr);
2194
2195 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2196 {
2197         struct rdma_id_private *id_priv;
2198         unsigned long flags;
2199         int ret;
2200
2201         id_priv = container_of(id, struct rdma_id_private, id);
2202         spin_lock_irqsave(&id_priv->lock, flags);
2203         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2204                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2205                 id_priv->afonly = afonly;
2206                 ret = 0;
2207         } else {
2208                 ret = -EINVAL;
2209         }
2210         spin_unlock_irqrestore(&id_priv->lock, flags);
2211         return ret;
2212 }
2213 EXPORT_SYMBOL(rdma_set_afonly);
2214
2215 static void cma_bind_port(struct rdma_bind_list *bind_list,
2216                           struct rdma_id_private *id_priv)
2217 {
2218         struct sockaddr *addr;
2219         struct sockaddr_ib *sib;
2220         u64 sid, mask;
2221         __be16 port;
2222
2223         addr = cma_src_addr(id_priv);
2224         port = htons(bind_list->port);
2225
2226         switch (addr->sa_family) {
2227         case AF_INET:
2228                 ((struct sockaddr_in *) addr)->sin_port = port;
2229                 break;
2230         case AF_INET6:
2231                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2232                 break;
2233         case AF_IB:
2234                 sib = (struct sockaddr_ib *) addr;
2235                 sid = be64_to_cpu(sib->sib_sid);
2236                 mask = be64_to_cpu(sib->sib_sid_mask);
2237                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2238                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2239                 break;
2240         }
2241         id_priv->bind_list = bind_list;
2242         hlist_add_head(&id_priv->node, &bind_list->owners);
2243 }
2244
2245 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2246                           unsigned short snum)
2247 {
2248         struct rdma_bind_list *bind_list;
2249         int ret;
2250
2251         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2252         if (!bind_list)
2253                 return -ENOMEM;
2254
2255         ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2256         if (ret < 0)
2257                 goto err;
2258
2259         bind_list->ps = ps;
2260         bind_list->port = (unsigned short)ret;
2261         cma_bind_port(bind_list, id_priv);
2262         return 0;
2263 err:
2264         kfree(bind_list);
2265         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2266 }
2267
2268 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2269 {
2270         static unsigned int last_used_port;
2271         int low, high, remaining;
2272         unsigned int rover;
2273
2274         inet_get_local_port_range(&low, &high);
2275         remaining = (high - low) + 1;
2276         rover = net_random() % remaining + low;
2277 retry:
2278         if (last_used_port != rover &&
2279             !idr_find(ps, (unsigned short) rover)) {
2280                 int ret = cma_alloc_port(ps, id_priv, rover);
2281                 /*
2282                  * Remember previously used port number in order to avoid
2283                  * re-using same port immediately after it is closed.
2284                  */
2285                 if (!ret)
2286                         last_used_port = rover;
2287                 if (ret != -EADDRNOTAVAIL)
2288                         return ret;
2289         }
2290         if (--remaining) {
2291                 rover++;
2292                 if ((rover < low) || (rover > high))
2293                         rover = low;
2294                 goto retry;
2295         }
2296         return -EADDRNOTAVAIL;
2297 }
2298
2299 /*
2300  * Check that the requested port is available.  This is called when trying to
2301  * bind to a specific port, or when trying to listen on a bound port.  In
2302  * the latter case, the provided id_priv may already be on the bind_list, but
2303  * we still need to check that it's okay to start listening.
2304  */
2305 static int cma_check_port(struct rdma_bind_list *bind_list,
2306                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2307 {
2308         struct rdma_id_private *cur_id;
2309         struct sockaddr *addr, *cur_addr;
2310
2311         addr = cma_src_addr(id_priv);
2312         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2313                 if (id_priv == cur_id)
2314                         continue;
2315
2316                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2317                     cur_id->reuseaddr)
2318                         continue;
2319
2320                 cur_addr = cma_src_addr(cur_id);
2321                 if (id_priv->afonly && cur_id->afonly &&
2322                     (addr->sa_family != cur_addr->sa_family))
2323                         continue;
2324
2325                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2326                         return -EADDRNOTAVAIL;
2327
2328                 if (!cma_addr_cmp(addr, cur_addr))
2329                         return -EADDRINUSE;
2330         }
2331         return 0;
2332 }
2333
2334 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2335 {
2336         struct rdma_bind_list *bind_list;
2337         unsigned short snum;
2338         int ret;
2339
2340         snum = ntohs(cma_port(cma_src_addr(id_priv)));
2341         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2342                 return -EACCES;
2343
2344         bind_list = idr_find(ps, snum);
2345         if (!bind_list) {
2346                 ret = cma_alloc_port(ps, id_priv, snum);
2347         } else {
2348                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2349                 if (!ret)
2350                         cma_bind_port(bind_list, id_priv);
2351         }
2352         return ret;
2353 }
2354
2355 static int cma_bind_listen(struct rdma_id_private *id_priv)
2356 {
2357         struct rdma_bind_list *bind_list = id_priv->bind_list;
2358         int ret = 0;
2359
2360         mutex_lock(&lock);
2361         if (bind_list->owners.first->next)
2362                 ret = cma_check_port(bind_list, id_priv, 0);
2363         mutex_unlock(&lock);
2364         return ret;
2365 }
2366
2367 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2368 {
2369         switch (id_priv->id.ps) {
2370         case RDMA_PS_TCP:
2371                 return &tcp_ps;
2372         case RDMA_PS_UDP:
2373                 return &udp_ps;
2374         case RDMA_PS_IPOIB:
2375                 return &ipoib_ps;
2376         case RDMA_PS_IB:
2377                 return &ib_ps;
2378         default:
2379                 return NULL;
2380         }
2381 }
2382
2383 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2384 {
2385         struct idr *ps = NULL;
2386         struct sockaddr_ib *sib;
2387         u64 sid_ps, mask, sid;
2388
2389         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2390         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2391         sid = be64_to_cpu(sib->sib_sid) & mask;
2392
2393         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2394                 sid_ps = RDMA_IB_IP_PS_IB;
2395                 ps = &ib_ps;
2396         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2397                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2398                 sid_ps = RDMA_IB_IP_PS_TCP;
2399                 ps = &tcp_ps;
2400         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2401                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2402                 sid_ps = RDMA_IB_IP_PS_UDP;
2403                 ps = &udp_ps;
2404         }
2405
2406         if (ps) {
2407                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2408                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2409                                                 be64_to_cpu(sib->sib_sid_mask));
2410         }
2411         return ps;
2412 }
2413
2414 static int cma_get_port(struct rdma_id_private *id_priv)
2415 {
2416         struct idr *ps;
2417         int ret;
2418
2419         if (cma_family(id_priv) != AF_IB)
2420                 ps = cma_select_inet_ps(id_priv);
2421         else
2422                 ps = cma_select_ib_ps(id_priv);
2423         if (!ps)
2424                 return -EPROTONOSUPPORT;
2425
2426         mutex_lock(&lock);
2427         if (cma_any_port(cma_src_addr(id_priv)))
2428                 ret = cma_alloc_any_port(ps, id_priv);
2429         else
2430                 ret = cma_use_port(ps, id_priv);
2431         mutex_unlock(&lock);
2432
2433         return ret;
2434 }
2435
2436 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2437                                struct sockaddr *addr)
2438 {
2439 #if IS_ENABLED(CONFIG_IPV6)
2440         struct sockaddr_in6 *sin6;
2441
2442         if (addr->sa_family != AF_INET6)
2443                 return 0;
2444
2445         sin6 = (struct sockaddr_in6 *) addr;
2446         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2447             !sin6->sin6_scope_id)
2448                         return -EINVAL;
2449
2450         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2451 #endif
2452         return 0;
2453 }
2454
2455 int rdma_listen(struct rdma_cm_id *id, int backlog)
2456 {
2457         struct rdma_id_private *id_priv;
2458         int ret;
2459
2460         id_priv = container_of(id, struct rdma_id_private, id);
2461         if (id_priv->state == RDMA_CM_IDLE) {
2462                 id->route.addr.src_addr.ss_family = AF_INET;
2463                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2464                 if (ret)
2465                         return ret;
2466         }
2467
2468         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2469                 return -EINVAL;
2470
2471         if (id_priv->reuseaddr) {
2472                 ret = cma_bind_listen(id_priv);
2473                 if (ret)
2474                         goto err;
2475         }
2476
2477         id_priv->backlog = backlog;
2478         if (id->device) {
2479                 switch (rdma_node_get_transport(id->device->node_type)) {
2480                 case RDMA_TRANSPORT_IB:
2481                         ret = cma_ib_listen(id_priv);
2482                         if (ret)
2483                                 goto err;
2484                         break;
2485                 case RDMA_TRANSPORT_IWARP:
2486                         ret = cma_iw_listen(id_priv, backlog);
2487                         if (ret)
2488                                 goto err;
2489                         break;
2490                 default:
2491                         ret = -ENOSYS;
2492                         goto err;
2493                 }
2494         } else
2495                 cma_listen_on_all(id_priv);
2496
2497         return 0;
2498 err:
2499         id_priv->backlog = 0;
2500         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2501         return ret;
2502 }
2503 EXPORT_SYMBOL(rdma_listen);
2504
2505 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2506 {
2507         struct rdma_id_private *id_priv;
2508         int ret;
2509
2510         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2511             addr->sa_family != AF_IB)
2512                 return -EAFNOSUPPORT;
2513
2514         id_priv = container_of(id, struct rdma_id_private, id);
2515         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2516                 return -EINVAL;
2517
2518         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2519         if (ret)
2520                 goto err1;
2521
2522         if (!cma_any_addr(addr)) {
2523                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2524                 if (ret)
2525                         goto err1;
2526
2527                 ret = cma_acquire_dev(id_priv);
2528                 if (ret)
2529                         goto err1;
2530         }
2531
2532         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2533         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2534                 if (addr->sa_family == AF_INET)
2535                         id_priv->afonly = 1;
2536 #if IS_ENABLED(CONFIG_IPV6)
2537                 else if (addr->sa_family == AF_INET6)
2538                         id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2539 #endif
2540         }
2541         ret = cma_get_port(id_priv);
2542         if (ret)
2543                 goto err2;
2544
2545         return 0;
2546 err2:
2547         if (id_priv->cma_dev)
2548                 cma_release_dev(id_priv);
2549 err1:
2550         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2551         return ret;
2552 }
2553 EXPORT_SYMBOL(rdma_bind_addr);
2554
2555 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2556 {
2557         struct cma_hdr *cma_hdr;
2558
2559         cma_hdr = hdr;
2560         cma_hdr->cma_version = CMA_VERSION;
2561         if (cma_family(id_priv) == AF_INET) {
2562                 struct sockaddr_in *src4, *dst4;
2563
2564                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2565                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2566
2567                 cma_set_ip_ver(cma_hdr, 4);
2568                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2569                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2570                 cma_hdr->port = src4->sin_port;
2571         } else if (cma_family(id_priv) == AF_INET6) {
2572                 struct sockaddr_in6 *src6, *dst6;
2573
2574                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2575                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2576
2577                 cma_set_ip_ver(cma_hdr, 6);
2578                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2579                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2580                 cma_hdr->port = src6->sin6_port;
2581         }
2582         return 0;
2583 }
2584
2585 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2586                                 struct ib_cm_event *ib_event)
2587 {
2588         struct rdma_id_private *id_priv = cm_id->context;
2589         struct rdma_cm_event event;
2590         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2591         int ret = 0;
2592
2593         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2594                 return 0;
2595
2596         memset(&event, 0, sizeof event);
2597         switch (ib_event->event) {
2598         case IB_CM_SIDR_REQ_ERROR:
2599                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2600                 event.status = -ETIMEDOUT;
2601                 break;
2602         case IB_CM_SIDR_REP_RECEIVED:
2603                 event.param.ud.private_data = ib_event->private_data;
2604                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2605                 if (rep->status != IB_SIDR_SUCCESS) {
2606                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2607                         event.status = ib_event->param.sidr_rep_rcvd.status;
2608                         break;
2609                 }
2610                 ret = cma_set_qkey(id_priv, rep->qkey);
2611                 if (ret) {
2612                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2613                         event.status = ret;
2614                         break;
2615                 }
2616                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2617                                      id_priv->id.route.path_rec,
2618                                      &event.param.ud.ah_attr);
2619                 event.param.ud.qp_num = rep->qpn;
2620                 event.param.ud.qkey = rep->qkey;
2621                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2622                 event.status = 0;
2623                 break;
2624         default:
2625                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2626                        ib_event->event);
2627                 goto out;
2628         }
2629
2630         ret = id_priv->id.event_handler(&id_priv->id, &event);
2631         if (ret) {
2632                 /* Destroy the CM ID by returning a non-zero value. */
2633                 id_priv->cm_id.ib = NULL;
2634                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2635                 mutex_unlock(&id_priv->handler_mutex);
2636                 rdma_destroy_id(&id_priv->id);
2637                 return ret;
2638         }
2639 out:
2640         mutex_unlock(&id_priv->handler_mutex);
2641         return ret;
2642 }
2643
2644 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2645                               struct rdma_conn_param *conn_param)
2646 {
2647         struct ib_cm_sidr_req_param req;
2648         struct ib_cm_id *id;
2649         void *private_data;
2650         int offset, ret;
2651
2652         memset(&req, 0, sizeof req);
2653         offset = cma_user_data_offset(id_priv);
2654         req.private_data_len = offset + conn_param->private_data_len;
2655         if (req.private_data_len < conn_param->private_data_len)
2656                 return -EINVAL;
2657
2658         if (req.private_data_len) {
2659                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2660                 if (!private_data)
2661                         return -ENOMEM;
2662         } else {
2663                 private_data = NULL;
2664         }
2665
2666         if (conn_param->private_data && conn_param->private_data_len)
2667                 memcpy(private_data + offset, conn_param->private_data,
2668                        conn_param->private_data_len);
2669
2670         if (private_data) {
2671                 ret = cma_format_hdr(private_data, id_priv);
2672                 if (ret)
2673                         goto out;
2674                 req.private_data = private_data;
2675         }
2676
2677         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2678                              id_priv);
2679         if (IS_ERR(id)) {
2680                 ret = PTR_ERR(id);
2681                 goto out;
2682         }
2683         id_priv->cm_id.ib = id;
2684
2685         req.path = id_priv->id.route.path_rec;
2686         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2687         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2688         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2689
2690         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2691         if (ret) {
2692                 ib_destroy_cm_id(id_priv->cm_id.ib);
2693                 id_priv->cm_id.ib = NULL;
2694         }
2695 out:
2696         kfree(private_data);
2697         return ret;
2698 }
2699
2700 static int cma_connect_ib(struct rdma_id_private *id_priv,
2701                           struct rdma_conn_param *conn_param)
2702 {
2703         struct ib_cm_req_param req;
2704         struct rdma_route *route;
2705         void *private_data;
2706         struct ib_cm_id *id;
2707         int offset, ret;
2708
2709         memset(&req, 0, sizeof req);
2710         offset = cma_user_data_offset(id_priv);
2711         req.private_data_len = offset + conn_param->private_data_len;
2712         if (req.private_data_len < conn_param->private_data_len)
2713                 return -EINVAL;
2714
2715         if (req.private_data_len) {
2716                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2717                 if (!private_data)
2718                         return -ENOMEM;
2719         } else {
2720                 private_data = NULL;
2721         }
2722
2723         if (conn_param->private_data && conn_param->private_data_len)
2724                 memcpy(private_data + offset, conn_param->private_data,
2725                        conn_param->private_data_len);
2726
2727         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2728         if (IS_ERR(id)) {
2729                 ret = PTR_ERR(id);
2730                 goto out;
2731         }
2732         id_priv->cm_id.ib = id;
2733
2734         route = &id_priv->id.route;
2735         if (private_data) {
2736                 ret = cma_format_hdr(private_data, id_priv);
2737                 if (ret)
2738                         goto out;
2739                 req.private_data = private_data;
2740         }
2741
2742         req.primary_path = &route->path_rec[0];
2743         if (route->num_paths == 2)
2744                 req.alternate_path = &route->path_rec[1];
2745
2746         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2747         req.qp_num = id_priv->qp_num;
2748         req.qp_type = id_priv->id.qp_type;
2749         req.starting_psn = id_priv->seq_num;
2750         req.responder_resources = conn_param->responder_resources;
2751         req.initiator_depth = conn_param->initiator_depth;
2752         req.flow_control = conn_param->flow_control;
2753         req.retry_count = min_t(u8, 7, conn_param->retry_count);
2754         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2755         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2756         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2757         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2758         req.srq = id_priv->srq ? 1 : 0;
2759
2760         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2761 out:
2762         if (ret && !IS_ERR(id)) {
2763                 ib_destroy_cm_id(id);
2764                 id_priv->cm_id.ib = NULL;
2765         }
2766
2767         kfree(private_data);
2768         return ret;
2769 }
2770
2771 static int cma_connect_iw(struct rdma_id_private *id_priv,
2772                           struct rdma_conn_param *conn_param)
2773 {
2774         struct iw_cm_id *cm_id;
2775         int ret;
2776         struct iw_cm_conn_param iw_param;
2777
2778         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2779         if (IS_ERR(cm_id))
2780                 return PTR_ERR(cm_id);
2781
2782         id_priv->cm_id.iw = cm_id;
2783
2784         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
2785                rdma_addr_size(cma_src_addr(id_priv)));
2786         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
2787                rdma_addr_size(cma_dst_addr(id_priv)));
2788
2789         ret = cma_modify_qp_rtr(id_priv, conn_param);
2790         if (ret)
2791                 goto out;
2792
2793         if (conn_param) {
2794                 iw_param.ord = conn_param->initiator_depth;
2795                 iw_param.ird = conn_param->responder_resources;
2796                 iw_param.private_data = conn_param->private_data;
2797                 iw_param.private_data_len = conn_param->private_data_len;
2798                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2799         } else {
2800                 memset(&iw_param, 0, sizeof iw_param);
2801                 iw_param.qpn = id_priv->qp_num;
2802         }
2803         ret = iw_cm_connect(cm_id, &iw_param);
2804 out:
2805         if (ret) {
2806                 iw_destroy_cm_id(cm_id);
2807                 id_priv->cm_id.iw = NULL;
2808         }
2809         return ret;
2810 }
2811
2812 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2813 {
2814         struct rdma_id_private *id_priv;
2815         int ret;
2816
2817         id_priv = container_of(id, struct rdma_id_private, id);
2818         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2819                 return -EINVAL;
2820
2821         if (!id->qp) {
2822                 id_priv->qp_num = conn_param->qp_num;
2823                 id_priv->srq = conn_param->srq;
2824         }
2825
2826         switch (rdma_node_get_transport(id->device->node_type)) {
2827         case RDMA_TRANSPORT_IB:
2828                 if (id->qp_type == IB_QPT_UD)
2829                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2830                 else
2831                         ret = cma_connect_ib(id_priv, conn_param);
2832                 break;
2833         case RDMA_TRANSPORT_IWARP:
2834                 ret = cma_connect_iw(id_priv, conn_param);
2835                 break;
2836         default:
2837                 ret = -ENOSYS;
2838                 break;
2839         }
2840         if (ret)
2841                 goto err;
2842
2843         return 0;
2844 err:
2845         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2846         return ret;
2847 }
2848 EXPORT_SYMBOL(rdma_connect);
2849
2850 static int cma_accept_ib(struct rdma_id_private *id_priv,
2851                          struct rdma_conn_param *conn_param)
2852 {
2853         struct ib_cm_rep_param rep;
2854         int ret;
2855
2856         ret = cma_modify_qp_rtr(id_priv, conn_param);
2857         if (ret)
2858                 goto out;
2859
2860         ret = cma_modify_qp_rts(id_priv, conn_param);
2861         if (ret)
2862                 goto out;
2863
2864         memset(&rep, 0, sizeof rep);
2865         rep.qp_num = id_priv->qp_num;
2866         rep.starting_psn = id_priv->seq_num;
2867         rep.private_data = conn_param->private_data;
2868         rep.private_data_len = conn_param->private_data_len;
2869         rep.responder_resources = conn_param->responder_resources;
2870         rep.initiator_depth = conn_param->initiator_depth;
2871         rep.failover_accepted = 0;
2872         rep.flow_control = conn_param->flow_control;
2873         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2874         rep.srq = id_priv->srq ? 1 : 0;
2875
2876         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2877 out:
2878         return ret;
2879 }
2880
2881 static int cma_accept_iw(struct rdma_id_private *id_priv,
2882                   struct rdma_conn_param *conn_param)
2883 {
2884         struct iw_cm_conn_param iw_param;
2885         int ret;
2886
2887         ret = cma_modify_qp_rtr(id_priv, conn_param);
2888         if (ret)
2889                 return ret;
2890
2891         iw_param.ord = conn_param->initiator_depth;
2892         iw_param.ird = conn_param->responder_resources;
2893         iw_param.private_data = conn_param->private_data;
2894         iw_param.private_data_len = conn_param->private_data_len;
2895         if (id_priv->id.qp) {
2896                 iw_param.qpn = id_priv->qp_num;
2897         } else
2898                 iw_param.qpn = conn_param->qp_num;
2899
2900         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2901 }
2902
2903 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2904                              enum ib_cm_sidr_status status, u32 qkey,
2905                              const void *private_data, int private_data_len)
2906 {
2907         struct ib_cm_sidr_rep_param rep;
2908         int ret;
2909
2910         memset(&rep, 0, sizeof rep);
2911         rep.status = status;
2912         if (status == IB_SIDR_SUCCESS) {
2913                 ret = cma_set_qkey(id_priv, qkey);
2914                 if (ret)
2915                         return ret;
2916                 rep.qp_num = id_priv->qp_num;
2917                 rep.qkey = id_priv->qkey;
2918         }
2919         rep.private_data = private_data;
2920         rep.private_data_len = private_data_len;
2921
2922         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2923 }
2924
2925 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2926 {
2927         struct rdma_id_private *id_priv;
2928         int ret;
2929
2930         id_priv = container_of(id, struct rdma_id_private, id);
2931
2932         id_priv->owner = task_pid_nr(current);
2933
2934         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2935                 return -EINVAL;
2936
2937         if (!id->qp && conn_param) {
2938                 id_priv->qp_num = conn_param->qp_num;
2939                 id_priv->srq = conn_param->srq;
2940         }
2941
2942         switch (rdma_node_get_transport(id->device->node_type)) {
2943         case RDMA_TRANSPORT_IB:
2944                 if (id->qp_type == IB_QPT_UD) {
2945                         if (conn_param)
2946                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2947                                                         conn_param->qkey,
2948                                                         conn_param->private_data,
2949                                                         conn_param->private_data_len);
2950                         else
2951                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2952                                                         0, NULL, 0);
2953                 } else {
2954                         if (conn_param)
2955                                 ret = cma_accept_ib(id_priv, conn_param);
2956                         else
2957                                 ret = cma_rep_recv(id_priv);
2958                 }
2959                 break;
2960         case RDMA_TRANSPORT_IWARP:
2961                 ret = cma_accept_iw(id_priv, conn_param);
2962                 break;
2963         default:
2964                 ret = -ENOSYS;
2965                 break;
2966         }
2967
2968         if (ret)
2969                 goto reject;
2970
2971         return 0;
2972 reject:
2973         cma_modify_qp_err(id_priv);
2974         rdma_reject(id, NULL, 0);
2975         return ret;
2976 }
2977 EXPORT_SYMBOL(rdma_accept);
2978
2979 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2980 {
2981         struct rdma_id_private *id_priv;
2982         int ret;
2983
2984         id_priv = container_of(id, struct rdma_id_private, id);
2985         if (!id_priv->cm_id.ib)
2986                 return -EINVAL;
2987
2988         switch (id->device->node_type) {
2989         case RDMA_NODE_IB_CA:
2990                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2991                 break;
2992         default:
2993                 ret = 0;
2994                 break;
2995         }
2996         return ret;
2997 }
2998 EXPORT_SYMBOL(rdma_notify);
2999
3000 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3001                 u8 private_data_len)
3002 {
3003         struct rdma_id_private *id_priv;
3004         int ret;
3005
3006         id_priv = container_of(id, struct rdma_id_private, id);
3007         if (!id_priv->cm_id.ib)
3008                 return -EINVAL;
3009
3010         switch (rdma_node_get_transport(id->device->node_type)) {
3011         case RDMA_TRANSPORT_IB:
3012                 if (id->qp_type == IB_QPT_UD)
3013                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3014                                                 private_data, private_data_len);
3015                 else
3016                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
3017                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
3018                                              0, private_data, private_data_len);
3019                 break;
3020         case RDMA_TRANSPORT_IWARP:
3021                 ret = iw_cm_reject(id_priv->cm_id.iw,
3022                                    private_data, private_data_len);
3023                 break;
3024         default:
3025                 ret = -ENOSYS;
3026                 break;
3027         }
3028         return ret;
3029 }
3030 EXPORT_SYMBOL(rdma_reject);
3031
3032 int rdma_disconnect(struct rdma_cm_id *id)
3033 {
3034         struct rdma_id_private *id_priv;
3035         int ret;
3036
3037         id_priv = container_of(id, struct rdma_id_private, id);
3038         if (!id_priv->cm_id.ib)
3039                 return -EINVAL;
3040
3041         switch (rdma_node_get_transport(id->device->node_type)) {
3042         case RDMA_TRANSPORT_IB:
3043                 ret = cma_modify_qp_err(id_priv);
3044                 if (ret)
3045                         goto out;
3046                 /* Initiate or respond to a disconnect. */
3047                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3048                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3049                 break;
3050         case RDMA_TRANSPORT_IWARP:
3051                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3052                 break;
3053         default:
3054                 ret = -EINVAL;
3055                 break;
3056         }
3057 out:
3058         return ret;
3059 }
3060 EXPORT_SYMBOL(rdma_disconnect);
3061
3062 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3063 {
3064         struct rdma_id_private *id_priv;
3065         struct cma_multicast *mc = multicast->context;
3066         struct rdma_cm_event event;
3067         int ret;
3068
3069         id_priv = mc->id_priv;
3070         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3071             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3072                 return 0;
3073
3074         if (!status)
3075                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3076         mutex_lock(&id_priv->qp_mutex);
3077         if (!status && id_priv->id.qp)
3078                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3079                                          be16_to_cpu(multicast->rec.mlid));
3080         mutex_unlock(&id_priv->qp_mutex);
3081
3082         memset(&event, 0, sizeof event);
3083         event.status = status;
3084         event.param.ud.private_data = mc->context;
3085         if (!status) {
3086                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3087                 ib_init_ah_from_mcmember(id_priv->id.device,
3088                                          id_priv->id.port_num, &multicast->rec,
3089                                          &event.param.ud.ah_attr);
3090                 event.param.ud.qp_num = 0xFFFFFF;
3091                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3092         } else
3093                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3094
3095         ret = id_priv->id.event_handler(&id_priv->id, &event);
3096         if (ret) {
3097                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3098                 mutex_unlock(&id_priv->handler_mutex);
3099                 rdma_destroy_id(&id_priv->id);
3100                 return 0;
3101         }
3102
3103         mutex_unlock(&id_priv->handler_mutex);
3104         return 0;
3105 }
3106
3107 static void cma_set_mgid(struct rdma_id_private *id_priv,
3108                          struct sockaddr *addr, union ib_gid *mgid)
3109 {
3110         unsigned char mc_map[MAX_ADDR_LEN];
3111         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3112         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3113         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3114
3115         if (cma_any_addr(addr)) {
3116                 memset(mgid, 0, sizeof *mgid);
3117         } else if ((addr->sa_family == AF_INET6) &&
3118                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3119                                                                  0xFF10A01B)) {
3120                 /* IPv6 address is an SA assigned MGID. */
3121                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3122         } else if (addr->sa_family == AF_IB) {
3123                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3124         } else if ((addr->sa_family == AF_INET6)) {
3125                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3126                 if (id_priv->id.ps == RDMA_PS_UDP)
3127                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3128                 *mgid = *(union ib_gid *) (mc_map + 4);
3129         } else {
3130                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3131                 if (id_priv->id.ps == RDMA_PS_UDP)
3132                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3133                 *mgid = *(union ib_gid *) (mc_map + 4);
3134         }
3135 }
3136
3137 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3138                                  struct cma_multicast *mc)
3139 {
3140         struct ib_sa_mcmember_rec rec;
3141         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3142         ib_sa_comp_mask comp_mask;
3143         int ret;
3144
3145         ib_addr_get_mgid(dev_addr, &rec.mgid);
3146         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3147                                      &rec.mgid, &rec);
3148         if (ret)
3149                 return ret;
3150
3151         ret = cma_set_qkey(id_priv, 0);
3152         if (ret)
3153                 return ret;
3154
3155         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3156         rec.qkey = cpu_to_be32(id_priv->qkey);
3157         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3158         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3159         rec.join_state = 1;
3160
3161         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3162                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3163                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3164                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3165                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3166
3167         if (id_priv->id.ps == RDMA_PS_IPOIB)
3168                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3169                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3170                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3171                              IB_SA_MCMEMBER_REC_MTU |
3172                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
3173
3174         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3175                                                 id_priv->id.port_num, &rec,
3176                                                 comp_mask, GFP_KERNEL,
3177                                                 cma_ib_mc_handler, mc);
3178         return PTR_ERR_OR_ZERO(mc->multicast.ib);
3179 }
3180
3181 static void iboe_mcast_work_handler(struct work_struct *work)
3182 {
3183         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3184         struct cma_multicast *mc = mw->mc;
3185         struct ib_sa_multicast *m = mc->multicast.ib;
3186
3187         mc->multicast.ib->context = mc;
3188         cma_ib_mc_handler(0, m);
3189         kref_put(&mc->mcref, release_mc);
3190         kfree(mw);
3191 }
3192
3193 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3194 {
3195         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3196         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3197
3198         if (cma_any_addr(addr)) {
3199                 memset(mgid, 0, sizeof *mgid);
3200         } else if (addr->sa_family == AF_INET6) {
3201                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3202         } else {
3203                 mgid->raw[0] = 0xff;
3204                 mgid->raw[1] = 0x0e;
3205                 mgid->raw[2] = 0;
3206                 mgid->raw[3] = 0;
3207                 mgid->raw[4] = 0;
3208                 mgid->raw[5] = 0;
3209                 mgid->raw[6] = 0;
3210                 mgid->raw[7] = 0;
3211                 mgid->raw[8] = 0;
3212                 mgid->raw[9] = 0;
3213                 mgid->raw[10] = 0xff;
3214                 mgid->raw[11] = 0xff;
3215                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3216         }
3217 }
3218
3219 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3220                                    struct cma_multicast *mc)
3221 {
3222         struct iboe_mcast_work *work;
3223         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3224         int err;
3225         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3226         struct net_device *ndev = NULL;
3227
3228         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3229                 return -EINVAL;
3230
3231         work = kzalloc(sizeof *work, GFP_KERNEL);
3232         if (!work)
3233                 return -ENOMEM;
3234
3235         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3236         if (!mc->multicast.ib) {
3237                 err = -ENOMEM;
3238                 goto out1;
3239         }
3240
3241         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3242
3243         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3244         if (id_priv->id.ps == RDMA_PS_UDP)
3245                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3246
3247         if (dev_addr->bound_dev_if)
3248                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3249         if (!ndev) {
3250                 err = -ENODEV;
3251                 goto out2;
3252         }
3253         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3254         mc->multicast.ib->rec.hop_limit = 1;
3255         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3256         dev_put(ndev);
3257         if (!mc->multicast.ib->rec.mtu) {
3258                 err = -EINVAL;
3259                 goto out2;
3260         }
3261         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3262         work->id = id_priv;
3263         work->mc = mc;
3264         INIT_WORK(&work->work, iboe_mcast_work_handler);
3265         kref_get(&mc->mcref);
3266         queue_work(cma_wq, &work->work);
3267
3268         return 0;
3269
3270 out2:
3271         kfree(mc->multicast.ib);
3272 out1:
3273         kfree(work);
3274         return err;
3275 }
3276
3277 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3278                         void *context)
3279 {
3280         struct rdma_id_private *id_priv;
3281         struct cma_multicast *mc;
3282         int ret;
3283
3284         id_priv = container_of(id, struct rdma_id_private, id);
3285         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3286             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3287                 return -EINVAL;
3288
3289         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3290         if (!mc)
3291                 return -ENOMEM;
3292
3293         memcpy(&mc->addr, addr, rdma_addr_size(addr));
3294         mc->context = context;
3295         mc->id_priv = id_priv;
3296
3297         spin_lock(&id_priv->lock);
3298         list_add(&mc->list, &id_priv->mc_list);
3299         spin_unlock(&id_priv->lock);
3300
3301         switch (rdma_node_get_transport(id->device->node_type)) {
3302         case RDMA_TRANSPORT_IB:
3303                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3304                 case IB_LINK_LAYER_INFINIBAND:
3305                         ret = cma_join_ib_multicast(id_priv, mc);
3306                         break;
3307                 case IB_LINK_LAYER_ETHERNET:
3308                         kref_init(&mc->mcref);
3309                         ret = cma_iboe_join_multicast(id_priv, mc);
3310                         break;
3311                 default:
3312                         ret = -EINVAL;
3313                 }
3314                 break;
3315         default:
3316                 ret = -ENOSYS;
3317                 break;
3318         }
3319
3320         if (ret) {
3321                 spin_lock_irq(&id_priv->lock);
3322                 list_del(&mc->list);
3323                 spin_unlock_irq(&id_priv->lock);
3324                 kfree(mc);
3325         }
3326         return ret;
3327 }
3328 EXPORT_SYMBOL(rdma_join_multicast);
3329
3330 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3331 {
3332         struct rdma_id_private *id_priv;
3333         struct cma_multicast *mc;
3334
3335         id_priv = container_of(id, struct rdma_id_private, id);
3336         spin_lock_irq(&id_priv->lock);
3337         list_for_each_entry(mc, &id_priv->mc_list, list) {
3338                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3339                         list_del(&mc->list);
3340                         spin_unlock_irq(&id_priv->lock);
3341
3342                         if (id->qp)
3343                                 ib_detach_mcast(id->qp,
3344                                                 &mc->multicast.ib->rec.mgid,
3345                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3346                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3347                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3348                                 case IB_LINK_LAYER_INFINIBAND:
3349                                         ib_sa_free_multicast(mc->multicast.ib);
3350                                         kfree(mc);
3351                                         break;
3352                                 case IB_LINK_LAYER_ETHERNET:
3353                                         kref_put(&mc->mcref, release_mc);
3354                                         break;
3355                                 default:
3356                                         break;
3357                                 }
3358                         }
3359                         return;
3360                 }
3361         }
3362         spin_unlock_irq(&id_priv->lock);
3363 }
3364 EXPORT_SYMBOL(rdma_leave_multicast);
3365
3366 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3367 {
3368         struct rdma_dev_addr *dev_addr;
3369         struct cma_ndev_work *work;
3370
3371         dev_addr = &id_priv->id.route.addr.dev_addr;
3372
3373         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3374             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3375                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3376                        ndev->name, &id_priv->id);
3377                 work = kzalloc(sizeof *work, GFP_KERNEL);
3378                 if (!work)
3379                         return -ENOMEM;
3380
3381                 INIT_WORK(&work->work, cma_ndev_work_handler);
3382                 work->id = id_priv;
3383                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3384                 atomic_inc(&id_priv->refcount);
3385                 queue_work(cma_wq, &work->work);
3386         }
3387
3388         return 0;
3389 }
3390
3391 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3392                                void *ptr)
3393 {
3394         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3395         struct cma_device *cma_dev;
3396         struct rdma_id_private *id_priv;
3397         int ret = NOTIFY_DONE;
3398
3399         if (dev_net(ndev) != &init_net)
3400                 return NOTIFY_DONE;
3401
3402         if (event != NETDEV_BONDING_FAILOVER)
3403                 return NOTIFY_DONE;
3404
3405         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3406                 return NOTIFY_DONE;
3407
3408         mutex_lock(&lock);
3409         list_for_each_entry(cma_dev, &dev_list, list)
3410                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3411                         ret = cma_netdev_change(ndev, id_priv);
3412                         if (ret)
3413                                 goto out;
3414                 }
3415
3416 out:
3417         mutex_unlock(&lock);
3418         return ret;
3419 }
3420
3421 static struct notifier_block cma_nb = {
3422         .notifier_call = cma_netdev_callback
3423 };
3424
3425 static void cma_add_one(struct ib_device *device)
3426 {
3427         struct cma_device *cma_dev;
3428         struct rdma_id_private *id_priv;
3429
3430         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3431         if (!cma_dev)
3432                 return;
3433
3434         cma_dev->device = device;
3435
3436         init_completion(&cma_dev->comp);
3437         atomic_set(&cma_dev->refcount, 1);
3438         INIT_LIST_HEAD(&cma_dev->id_list);
3439         ib_set_client_data(device, &cma_client, cma_dev);
3440
3441         mutex_lock(&lock);
3442         list_add_tail(&cma_dev->list, &dev_list);
3443         list_for_each_entry(id_priv, &listen_any_list, list)
3444                 cma_listen_on_dev(id_priv, cma_dev);
3445         mutex_unlock(&lock);
3446 }
3447
3448 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3449 {
3450         struct rdma_cm_event event;
3451         enum rdma_cm_state state;
3452         int ret = 0;
3453
3454         /* Record that we want to remove the device */
3455         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3456         if (state == RDMA_CM_DESTROYING)
3457                 return 0;
3458
3459         cma_cancel_operation(id_priv, state);
3460         mutex_lock(&id_priv->handler_mutex);
3461
3462         /* Check for destruction from another callback. */
3463         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3464                 goto out;
3465
3466         memset(&event, 0, sizeof event);
3467         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3468         ret = id_priv->id.event_handler(&id_priv->id, &event);
3469 out:
3470         mutex_unlock(&id_priv->handler_mutex);
3471         return ret;
3472 }
3473
3474 static void cma_process_remove(struct cma_device *cma_dev)
3475 {
3476         struct rdma_id_private *id_priv;
3477         int ret;
3478
3479         mutex_lock(&lock);
3480         while (!list_empty(&cma_dev->id_list)) {
3481                 id_priv = list_entry(cma_dev->id_list.next,
3482                                      struct rdma_id_private, list);
3483
3484                 list_del(&id_priv->listen_list);
3485                 list_del_init(&id_priv->list);
3486                 atomic_inc(&id_priv->refcount);
3487                 mutex_unlock(&lock);
3488
3489                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3490                 cma_deref_id(id_priv);
3491                 if (ret)
3492                         rdma_destroy_id(&id_priv->id);
3493
3494                 mutex_lock(&lock);
3495         }
3496         mutex_unlock(&lock);
3497
3498         cma_deref_dev(cma_dev);
3499         wait_for_completion(&cma_dev->comp);
3500 }
3501
3502 static void cma_remove_one(struct ib_device *device)
3503 {
3504         struct cma_device *cma_dev;
3505
3506         cma_dev = ib_get_client_data(device, &cma_client);
3507         if (!cma_dev)
3508                 return;
3509
3510         mutex_lock(&lock);
3511         list_del(&cma_dev->list);
3512         mutex_unlock(&lock);
3513
3514         cma_process_remove(cma_dev);
3515         kfree(cma_dev);
3516 }
3517
3518 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3519 {
3520         struct nlmsghdr *nlh;
3521         struct rdma_cm_id_stats *id_stats;
3522         struct rdma_id_private *id_priv;
3523         struct rdma_cm_id *id = NULL;
3524         struct cma_device *cma_dev;
3525         int i_dev = 0, i_id = 0;
3526
3527         /*
3528          * We export all of the IDs as a sequence of messages.  Each
3529          * ID gets its own netlink message.
3530          */
3531         mutex_lock(&lock);
3532
3533         list_for_each_entry(cma_dev, &dev_list, list) {
3534                 if (i_dev < cb->args[0]) {
3535                         i_dev++;
3536                         continue;
3537                 }
3538
3539                 i_id = 0;
3540                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3541                         if (i_id < cb->args[1]) {
3542                                 i_id++;
3543                                 continue;
3544                         }
3545
3546                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3547                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3548                                                 RDMA_NL_RDMA_CM_ID_STATS);
3549                         if (!id_stats)
3550                                 goto out;
3551
3552                         memset(id_stats, 0, sizeof *id_stats);
3553                         id = &id_priv->id;
3554                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3555                         id_stats->port_num = id->port_num;
3556                         id_stats->bound_dev_if =
3557                                 id->route.addr.dev_addr.bound_dev_if;
3558
3559                         if (ibnl_put_attr(skb, nlh,
3560                                           rdma_addr_size(cma_src_addr(id_priv)),
3561                                           cma_src_addr(id_priv),
3562                                           RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3563                                 goto out;
3564                         if (ibnl_put_attr(skb, nlh,
3565                                           rdma_addr_size(cma_src_addr(id_priv)),
3566                                           cma_dst_addr(id_priv),
3567                                           RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3568                                 goto out;
3569
3570                         id_stats->pid           = id_priv->owner;
3571                         id_stats->port_space    = id->ps;
3572                         id_stats->cm_state      = id_priv->state;
3573                         id_stats->qp_num        = id_priv->qp_num;
3574                         id_stats->qp_type       = id->qp_type;
3575
3576                         i_id++;
3577                 }
3578
3579                 cb->args[1] = 0;
3580                 i_dev++;
3581         }
3582
3583 out:
3584         mutex_unlock(&lock);
3585         cb->args[0] = i_dev;
3586         cb->args[1] = i_id;
3587
3588         return skb->len;
3589 }
3590
3591 static const struct ibnl_client_cbs cma_cb_table[] = {
3592         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3593                                        .module = THIS_MODULE },
3594 };
3595
3596 static int __init cma_init(void)
3597 {
3598         int ret;
3599
3600         cma_wq = create_singlethread_workqueue("rdma_cm");
3601         if (!cma_wq)
3602                 return -ENOMEM;
3603
3604         ib_sa_register_client(&sa_client);
3605         rdma_addr_register_client(&addr_client);
3606         register_netdevice_notifier(&cma_nb);
3607
3608         ret = ib_register_client(&cma_client);
3609         if (ret)
3610                 goto err;
3611
3612         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3613                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3614
3615         return 0;
3616
3617 err:
3618         unregister_netdevice_notifier(&cma_nb);
3619         rdma_addr_unregister_client(&addr_client);
3620         ib_sa_unregister_client(&sa_client);
3621         destroy_workqueue(cma_wq);
3622         return ret;
3623 }
3624
3625 static void __exit cma_cleanup(void)
3626 {
3627         ibnl_remove_client(RDMA_NL_RDMA_CM);
3628         ib_unregister_client(&cma_client);
3629         unregister_netdevice_notifier(&cma_nb);
3630         rdma_addr_unregister_client(&addr_client);
3631         ib_sa_unregister_client(&sa_client);
3632         destroy_workqueue(cma_wq);
3633         idr_destroy(&tcp_ps);
3634         idr_destroy(&udp_ps);
3635         idr_destroy(&ipoib_ps);
3636         idr_destroy(&ib_ps);
3637 }
3638
3639 module_init(cma_init);
3640 module_exit(cma_cleanup);