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
  45 #include <net/tcp.h>
  46 #include <net/ipv6.h>
  47
  48 #include <rdma/rdma_cm.h>
  49 #include <rdma/rdma_cm_ib.h>
  50 #include <rdma/ib_cache.h>
  51 #include <rdma/ib_cm.h>
  52 #include <rdma/ib_sa.h>
  53 #include <rdma/iw_cm.h>
  54
  55 MODULE_AUTHOR("Sean Hefty");
  56 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 #define CMA_CM_RESPONSE_TIMEOUT 20
  60 #define CMA_MAX_CM_RETRIES 15
  61 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  62 #define CMA_IBOE_PACKET_LIFETIME 18
  63
  64 static void cma_add_one(struct ib_device *device);
  65 static void cma_remove_one(struct ib_device *device);
  66
  67 static struct ib_client cma_client = {
  68         .name   = "cma",
  69         .add    = cma_add_one,
  70         .remove = cma_remove_one
  71 };
  72
  73 static struct ib_sa_client sa_client;
  74 static struct rdma_addr_client addr_client;
  75 static LIST_HEAD(dev_list);
  76 static LIST_HEAD(listen_any_list);
  77 static DEFINE_MUTEX(lock);
  78 static struct workqueue_struct *cma_wq;
  79 static DEFINE_IDR(sdp_ps);
  80 static DEFINE_IDR(tcp_ps);
  81 static DEFINE_IDR(udp_ps);
  82 static DEFINE_IDR(ipoib_ps);
  83
  84 struct cma_device {
  85         struct list_head        list;
  86         struct ib_device        *device;
  87         struct completion       comp;
  88         atomic_t                refcount;
  89         struct list_head        id_list;
  90 };
  91
  92 struct rdma_bind_list {
  93         struct idr              *ps;
  94         struct hlist_head       owners;
  95         unsigned short          port;
  96 };
  97
  98 /*
  99  * Device removal can occur at anytime, so we need extra handling to
 100  * serialize notifying the user of device removal with other callbacks.
 101  * We do this by disabling removal notification while a callback is in process,
 102  * and reporting it after the callback completes.
 103  */
 104 struct rdma_id_private {
 105         struct rdma_cm_id       id;
 106
 107         struct rdma_bind_list   *bind_list;
 108         struct hlist_node       node;
 109         struct list_head        list; /* listen_any_list or cma_device.list */
 110         struct list_head        listen_list; /* per device listens */
 111         struct cma_device       *cma_dev;
 112         struct list_head        mc_list;
 113
 114         int                     internal_id;
 115         enum rdma_cm_state      state;
 116         spinlock_t              lock;
 117         struct mutex            qp_mutex;
 118
 119         struct completion       comp;
 120         atomic_t                refcount;
 121         struct mutex            handler_mutex;
 122
 123         int                     backlog;
 124         int                     timeout_ms;
 125         struct ib_sa_query      *query;
 126         int                     query_id;
 127         union {
 128                 struct ib_cm_id *ib;
 129                 struct iw_cm_id *iw;
 130         } cm_id;
 131
 132         u32                     seq_num;
 133         u32                     qkey;
 134         u32                     qp_num;
 135         u8                      srq;
 136         u8                      tos;
 137         u8                      reuseaddr;
 138 };
 139
 140 struct cma_multicast {
 141         struct rdma_id_private *id_priv;
 142         union {
 143                 struct ib_sa_multicast *ib;
 144         } multicast;
 145         struct list_head        list;
 146         void                    *context;
 147         struct sockaddr_storage addr;
 148         struct kref             mcref;
 149 };
 150
 151 struct cma_work {
 152         struct work_struct      work;
 153         struct rdma_id_private  *id;
 154         enum rdma_cm_state      old_state;
 155         enum rdma_cm_state      new_state;
 156         struct rdma_cm_event    event;
 157 };
 158
 159 struct cma_ndev_work {
 160         struct work_struct      work;
 161         struct rdma_id_private  *id;
 162         struct rdma_cm_event    event;
 163 };
 164
 165 struct iboe_mcast_work {
 166         struct work_struct       work;
 167         struct rdma_id_private  *id;
 168         struct cma_multicast    *mc;
 169 };
 170
 171 union cma_ip_addr {
 172         struct in6_addr ip6;
 173         struct {
 174                 __be32 pad[3];
 175                 __be32 addr;
 176         } ip4;
 177 };
 178
 179 struct cma_hdr {
 180         u8 cma_version;
 181         u8 ip_version;  /* IP version: 7:4 */
 182         __be16 port;
 183         union cma_ip_addr src_addr;
 184         union cma_ip_addr dst_addr;
 185 };
 186
 187 struct sdp_hh {
 188         u8 bsdh[16];
 189         u8 sdp_version; /* Major version: 7:4 */
 190         u8 ip_version;  /* IP version: 7:4 */
 191         u8 sdp_specific1[10];
 192         __be16 port;
 193         __be16 sdp_specific2;
 194         union cma_ip_addr src_addr;
 195         union cma_ip_addr dst_addr;
 196 };
 197
 198 struct sdp_hah {
 199         u8 bsdh[16];
 200         u8 sdp_version;
 201 };
 202
 203 #define CMA_VERSION 0x00
 204 #define SDP_MAJ_VERSION 0x2
 205
 206 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 207 {
 208         unsigned long flags;
 209         int ret;
 210
 211         spin_lock_irqsave(&id_priv->lock, flags);
 212         ret = (id_priv->state == comp);
 213         spin_unlock_irqrestore(&id_priv->lock, flags);
 214         return ret;
 215 }
 216
 217 static int cma_comp_exch(struct rdma_id_private *id_priv,
 218                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 219 {
 220         unsigned long flags;
 221         int ret;
 222
 223         spin_lock_irqsave(&id_priv->lock, flags);
 224         if ((ret = (id_priv->state == comp)))
 225                 id_priv->state = exch;
 226         spin_unlock_irqrestore(&id_priv->lock, flags);
 227         return ret;
 228 }
 229
 230 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 231                                    enum rdma_cm_state exch)
 232 {
 233         unsigned long flags;
 234         enum rdma_cm_state old;
 235
 236         spin_lock_irqsave(&id_priv->lock, flags);
 237         old = id_priv->state;
 238         id_priv->state = exch;
 239         spin_unlock_irqrestore(&id_priv->lock, flags);
 240         return old;
 241 }
 242
 243 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 244 {
 245         return hdr->ip_version >> 4;
 246 }
 247
 248 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 249 {
 250         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 251 }
 252
 253 static inline u8 sdp_get_majv(u8 sdp_version)
 254 {
 255         return sdp_version >> 4;
 256 }
 257
 258 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 259 {
 260         return hh->ip_version >> 4;
 261 }
 262
 263 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 264 {
 265         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 266 }
 267
 268 static inline int cma_is_ud_ps(enum rdma_port_space ps)
 269 {
 270         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
 271 }
 272
 273 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 274                               struct cma_device *cma_dev)
 275 {
 276         atomic_inc(&cma_dev->refcount);
 277         id_priv->cma_dev = cma_dev;
 278         id_priv->id.device = cma_dev->device;
 279         id_priv->id.route.addr.dev_addr.transport =
 280                 rdma_node_get_transport(cma_dev->device->node_type);
 281         list_add_tail(&id_priv->list, &cma_dev->id_list);
 282 }
 283
 284 static inline void cma_deref_dev(struct cma_device *cma_dev)
 285 {
 286         if (atomic_dec_and_test(&cma_dev->refcount))
 287                 complete(&cma_dev->comp);
 288 }
 289
 290 static inline void release_mc(struct kref *kref)
 291 {
 292         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 293
 294         kfree(mc->multicast.ib);
 295         kfree(mc);
 296 }
 297
 298 static void cma_release_dev(struct rdma_id_private *id_priv)
 299 {
 300         mutex_lock(&lock);
 301         list_del(&id_priv->list);
 302         cma_deref_dev(id_priv->cma_dev);
 303         id_priv->cma_dev = NULL;
 304         mutex_unlock(&lock);
 305 }
 306
 307 static int cma_set_qkey(struct rdma_id_private *id_priv)
 308 {
 309         struct ib_sa_mcmember_rec rec;
 310         int ret = 0;
 311
 312         if (id_priv->qkey)
 313                 return 0;
 314
 315         switch (id_priv->id.ps) {
 316         case RDMA_PS_UDP:
 317                 id_priv->qkey = RDMA_UDP_QKEY;
 318                 break;
 319         case RDMA_PS_IPOIB:
 320                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 321                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 322                                              id_priv->id.port_num, &rec.mgid,
 323                                              &rec);
 324                 if (!ret)
 325                         id_priv->qkey = be32_to_cpu(rec.qkey);
 326                 break;
 327         default:
 328                 break;
 329         }
 330         return ret;
 331 }
 332
 333 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 334 {
 335         int i;
 336         int err;
 337         struct ib_port_attr props;
 338         union ib_gid tmp;
 339
 340         err = ib_query_port(device, port_num, &props);
 341         if (err)
 342                 return 1;
 343
 344         for (i = 0; i < props.gid_tbl_len; ++i) {
 345                 err = ib_query_gid(device, port_num, i, &tmp);
 346                 if (err)
 347                         return 1;
 348                 if (!memcmp(&tmp, gid, sizeof tmp))
 349                         return 0;
 350         }
 351
 352         return -EAGAIN;
 353 }
 354
 355 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 356 {
 357         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 358         struct cma_device *cma_dev;
 359         union ib_gid gid, iboe_gid;
 360         int ret = -ENODEV;
 361         u8 port;
 362         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 363                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 364
 365         mutex_lock(&lock);
 366         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 367         memcpy(&gid, dev_addr->src_dev_addr +
 368                rdma_addr_gid_offset(dev_addr), sizeof gid);
 369         list_for_each_entry(cma_dev, &dev_list, list) {
 370                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 371                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 372                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 373                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 374                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 375                                 else
 376                                         ret = find_gid_port(cma_dev->device, &gid, port);
 377
 378                                 if (!ret) {
 379                                         id_priv->id.port_num = port;
 380                                         goto out;
 381                                 } else if (ret == 1)
 382                                         break;
 383                         }
 384                 }
 385         }
 386
 387 out:
 388         if (!ret)
 389                 cma_attach_to_dev(id_priv, cma_dev);
 390
 391         mutex_unlock(&lock);
 392         return ret;
 393 }
 394
 395 static void cma_deref_id(struct rdma_id_private *id_priv)
 396 {
 397         if (atomic_dec_and_test(&id_priv->refcount))
 398                 complete(&id_priv->comp);
 399 }
 400
 401 static int cma_disable_callback(struct rdma_id_private *id_priv,
 402                                 enum rdma_cm_state state)
 403 {
 404         mutex_lock(&id_priv->handler_mutex);
 405         if (id_priv->state != state) {
 406                 mutex_unlock(&id_priv->handler_mutex);
 407                 return -EINVAL;
 408         }
 409         return 0;
 410 }
 411
 412 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
 413 {
 414         return (id_priv->id.device && id_priv->cm_id.ib);
 415 }
 416
 417 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 418                                   void *context, enum rdma_port_space ps)
 419 {
 420         struct rdma_id_private *id_priv;
 421
 422         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 423         if (!id_priv)
 424                 return ERR_PTR(-ENOMEM);
 425
 426         id_priv->state = RDMA_CM_IDLE;
 427         id_priv->id.context = context;
 428         id_priv->id.event_handler = event_handler;
 429         id_priv->id.ps = ps;
 430         spin_lock_init(&id_priv->lock);
 431         mutex_init(&id_priv->qp_mutex);
 432         init_completion(&id_priv->comp);
 433         atomic_set(&id_priv->refcount, 1);
 434         mutex_init(&id_priv->handler_mutex);
 435         INIT_LIST_HEAD(&id_priv->listen_list);
 436         INIT_LIST_HEAD(&id_priv->mc_list);
 437         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 438
 439         return &id_priv->id;
 440 }
 441 EXPORT_SYMBOL(rdma_create_id);
 442
 443 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 444 {
 445         struct ib_qp_attr qp_attr;
 446         int qp_attr_mask, ret;
 447
 448         qp_attr.qp_state = IB_QPS_INIT;
 449         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 450         if (ret)
 451                 return ret;
 452
 453         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 454         if (ret)
 455                 return ret;
 456
 457         qp_attr.qp_state = IB_QPS_RTR;
 458         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 459         if (ret)
 460                 return ret;
 461
 462         qp_attr.qp_state = IB_QPS_RTS;
 463         qp_attr.sq_psn = 0;
 464         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 465
 466         return ret;
 467 }
 468
 469 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 470 {
 471         struct ib_qp_attr qp_attr;
 472         int qp_attr_mask, ret;
 473
 474         qp_attr.qp_state = IB_QPS_INIT;
 475         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 476         if (ret)
 477                 return ret;
 478
 479         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 480 }
 481
 482 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 483                    struct ib_qp_init_attr *qp_init_attr)
 484 {
 485         struct rdma_id_private *id_priv;
 486         struct ib_qp *qp;
 487         int ret;
 488
 489         id_priv = container_of(id, struct rdma_id_private, id);
 490         if (id->device != pd->device)
 491                 return -EINVAL;
 492
 493         qp = ib_create_qp(pd, qp_init_attr);
 494         if (IS_ERR(qp))
 495                 return PTR_ERR(qp);
 496
 497         if (cma_is_ud_ps(id_priv->id.ps))
 498                 ret = cma_init_ud_qp(id_priv, qp);
 499         else
 500                 ret = cma_init_conn_qp(id_priv, qp);
 501         if (ret)
 502                 goto err;
 503
 504         id->qp = qp;
 505         id_priv->qp_num = qp->qp_num;
 506         id_priv->srq = (qp->srq != NULL);
 507         return 0;
 508 err:
 509         ib_destroy_qp(qp);
 510         return ret;
 511 }
 512 EXPORT_SYMBOL(rdma_create_qp);
 513
 514 void rdma_destroy_qp(struct rdma_cm_id *id)
 515 {
 516         struct rdma_id_private *id_priv;
 517
 518         id_priv = container_of(id, struct rdma_id_private, id);
 519         mutex_lock(&id_priv->qp_mutex);
 520         ib_destroy_qp(id_priv->id.qp);
 521         id_priv->id.qp = NULL;
 522         mutex_unlock(&id_priv->qp_mutex);
 523 }
 524 EXPORT_SYMBOL(rdma_destroy_qp);
 525
 526 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 527                              struct rdma_conn_param *conn_param)
 528 {
 529         struct ib_qp_attr qp_attr;
 530         int qp_attr_mask, ret;
 531
 532         mutex_lock(&id_priv->qp_mutex);
 533         if (!id_priv->id.qp) {
 534                 ret = 0;
 535                 goto out;
 536         }
 537
 538         /* Need to update QP attributes from default values. */
 539         qp_attr.qp_state = IB_QPS_INIT;
 540         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 541         if (ret)
 542                 goto out;
 543
 544         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 545         if (ret)
 546                 goto out;
 547
 548         qp_attr.qp_state = IB_QPS_RTR;
 549         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 550         if (ret)
 551                 goto out;
 552
 553         if (conn_param)
 554                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 555         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 556 out:
 557         mutex_unlock(&id_priv->qp_mutex);
 558         return ret;
 559 }
 560
 561 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 562                              struct rdma_conn_param *conn_param)
 563 {
 564         struct ib_qp_attr qp_attr;
 565         int qp_attr_mask, ret;
 566
 567         mutex_lock(&id_priv->qp_mutex);
 568         if (!id_priv->id.qp) {
 569                 ret = 0;
 570                 goto out;
 571         }
 572
 573         qp_attr.qp_state = IB_QPS_RTS;
 574         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 575         if (ret)
 576                 goto out;
 577
 578         if (conn_param)
 579                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 580         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 581 out:
 582         mutex_unlock(&id_priv->qp_mutex);
 583         return ret;
 584 }
 585
 586 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 587 {
 588         struct ib_qp_attr qp_attr;
 589         int ret;
 590
 591         mutex_lock(&id_priv->qp_mutex);
 592         if (!id_priv->id.qp) {
 593                 ret = 0;
 594                 goto out;
 595         }
 596
 597         qp_attr.qp_state = IB_QPS_ERR;
 598         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 599 out:
 600         mutex_unlock(&id_priv->qp_mutex);
 601         return ret;
 602 }
 603
 604 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 605                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 606 {
 607         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 608         int ret;
 609         u16 pkey;
 610
 611         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 612             IB_LINK_LAYER_INFINIBAND)
 613                 pkey = ib_addr_get_pkey(dev_addr);
 614         else
 615                 pkey = 0xffff;
 616
 617         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 618                                   pkey, &qp_attr->pkey_index);
 619         if (ret)
 620                 return ret;
 621
 622         qp_attr->port_num = id_priv->id.port_num;
 623         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 624
 625         if (cma_is_ud_ps(id_priv->id.ps)) {
 626                 ret = cma_set_qkey(id_priv);
 627                 if (ret)
 628                         return ret;
 629
 630                 qp_attr->qkey = id_priv->qkey;
 631                 *qp_attr_mask |= IB_QP_QKEY;
 632         } else {
 633                 qp_attr->qp_access_flags = 0;
 634                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 635         }
 636         return 0;
 637 }
 638
 639 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 640                        int *qp_attr_mask)
 641 {
 642         struct rdma_id_private *id_priv;
 643         int ret = 0;
 644
 645         id_priv = container_of(id, struct rdma_id_private, id);
 646         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 647         case RDMA_TRANSPORT_IB:
 648                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
 649                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 650                 else
 651                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 652                                                  qp_attr_mask);
 653                 if (qp_attr->qp_state == IB_QPS_RTR)
 654                         qp_attr->rq_psn = id_priv->seq_num;
 655                 break;
 656         case RDMA_TRANSPORT_IWARP:
 657                 if (!id_priv->cm_id.iw) {
 658                         qp_attr->qp_access_flags = 0;
 659                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 660                 } else
 661                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 662                                                  qp_attr_mask);
 663                 break;
 664         default:
 665                 ret = -ENOSYS;
 666                 break;
 667         }
 668
 669         return ret;
 670 }
 671 EXPORT_SYMBOL(rdma_init_qp_attr);
 672
 673 static inline int cma_zero_addr(struct sockaddr *addr)
 674 {
 675         struct in6_addr *ip6;
 676
 677         if (addr->sa_family == AF_INET)
 678                 return ipv4_is_zeronet(
 679                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 680         else {
 681                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 682                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 683                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 684         }
 685 }
 686
 687 static inline int cma_loopback_addr(struct sockaddr *addr)
 688 {
 689         if (addr->sa_family == AF_INET)
 690                 return ipv4_is_loopback(
 691                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 692         else
 693                 return ipv6_addr_loopback(
 694                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 695 }
 696
 697 static inline int cma_any_addr(struct sockaddr *addr)
 698 {
 699         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 700 }
 701
 702 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 703 {
 704         if (src->sa_family != dst->sa_family)
 705                 return -1;
 706
 707         switch (src->sa_family) {
 708         case AF_INET:
 709                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 710                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 711         default:
 712                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 713                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 714         }
 715 }
 716
 717 static inline __be16 cma_port(struct sockaddr *addr)
 718 {
 719         if (addr->sa_family == AF_INET)
 720                 return ((struct sockaddr_in *) addr)->sin_port;
 721         else
 722                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 723 }
 724
 725 static inline int cma_any_port(struct sockaddr *addr)
 726 {
 727         return !cma_port(addr);
 728 }
 729
 730 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 731                             u8 *ip_ver, __be16 *port,
 732                             union cma_ip_addr **src, union cma_ip_addr **dst)
 733 {
 734         switch (ps) {
 735         case RDMA_PS_SDP:
 736                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 737                     SDP_MAJ_VERSION)
 738                         return -EINVAL;
 739
 740                 *ip_ver = sdp_get_ip_ver(hdr);
 741                 *port   = ((struct sdp_hh *) hdr)->port;
 742                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 743                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 744                 break;
 745         default:
 746                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 747                         return -EINVAL;
 748
 749                 *ip_ver = cma_get_ip_ver(hdr);
 750                 *port   = ((struct cma_hdr *) hdr)->port;
 751                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 752                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 753                 break;
 754         }
 755
 756         if (*ip_ver != 4 && *ip_ver != 6)
 757                 return -EINVAL;
 758         return 0;
 759 }
 760
 761 static void cma_save_net_info(struct rdma_addr *addr,
 762                               struct rdma_addr *listen_addr,
 763                               u8 ip_ver, __be16 port,
 764                               union cma_ip_addr *src, union cma_ip_addr *dst)
 765 {
 766         struct sockaddr_in *listen4, *ip4;
 767         struct sockaddr_in6 *listen6, *ip6;
 768
 769         switch (ip_ver) {
 770         case 4:
 771                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 772                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 773                 ip4->sin_family = listen4->sin_family;
 774                 ip4->sin_addr.s_addr = dst->ip4.addr;
 775                 ip4->sin_port = listen4->sin_port;
 776
 777                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 778                 ip4->sin_family = listen4->sin_family;
 779                 ip4->sin_addr.s_addr = src->ip4.addr;
 780                 ip4->sin_port = port;
 781                 break;
 782         case 6:
 783                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 784                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 785                 ip6->sin6_family = listen6->sin6_family;
 786                 ip6->sin6_addr = dst->ip6;
 787                 ip6->sin6_port = listen6->sin6_port;
 788
 789                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 790                 ip6->sin6_family = listen6->sin6_family;
 791                 ip6->sin6_addr = src->ip6;
 792                 ip6->sin6_port = port;
 793                 break;
 794         default:
 795                 break;
 796         }
 797 }
 798
 799 static inline int cma_user_data_offset(enum rdma_port_space ps)
 800 {
 801         switch (ps) {
 802         case RDMA_PS_SDP:
 803                 return 0;
 804         default:
 805                 return sizeof(struct cma_hdr);
 806         }
 807 }
 808
 809 static void cma_cancel_route(struct rdma_id_private *id_priv)
 810 {
 811         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 812         case IB_LINK_LAYER_INFINIBAND:
 813                 if (id_priv->query)
 814                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 815                 break;
 816         default:
 817                 break;
 818         }
 819 }
 820
 821 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 822 {
 823         struct rdma_id_private *dev_id_priv;
 824
 825         /*
 826          * Remove from listen_any_list to prevent added devices from spawning
 827          * additional listen requests.
 828          */
 829         mutex_lock(&lock);
 830         list_del(&id_priv->list);
 831
 832         while (!list_empty(&id_priv->listen_list)) {
 833                 dev_id_priv = list_entry(id_priv->listen_list.next,
 834                                          struct rdma_id_private, listen_list);
 835                 /* sync with device removal to avoid duplicate destruction */
 836                 list_del_init(&dev_id_priv->list);
 837                 list_del(&dev_id_priv->listen_list);
 838                 mutex_unlock(&lock);
 839
 840                 rdma_destroy_id(&dev_id_priv->id);
 841                 mutex_lock(&lock);
 842         }
 843         mutex_unlock(&lock);
 844 }
 845
 846 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 847                                  enum rdma_cm_state state)
 848 {
 849         switch (state) {
 850         case RDMA_CM_ADDR_QUERY:
 851                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 852                 break;
 853         case RDMA_CM_ROUTE_QUERY:
 854                 cma_cancel_route(id_priv);
 855                 break;
 856         case RDMA_CM_LISTEN:
 857                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 858                                 && !id_priv->cma_dev)
 859                         cma_cancel_listens(id_priv);
 860                 break;
 861         default:
 862                 break;
 863         }
 864 }
 865
 866 static void cma_release_port(struct rdma_id_private *id_priv)
 867 {
 868         struct rdma_bind_list *bind_list = id_priv->bind_list;
 869
 870         if (!bind_list)
 871                 return;
 872
 873         mutex_lock(&lock);
 874         hlist_del(&id_priv->node);
 875         if (hlist_empty(&bind_list->owners)) {
 876                 idr_remove(bind_list->ps, bind_list->port);
 877                 kfree(bind_list);
 878         }
 879         mutex_unlock(&lock);
 880 }
 881
 882 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 883 {
 884         struct cma_multicast *mc;
 885
 886         while (!list_empty(&id_priv->mc_list)) {
 887                 mc = container_of(id_priv->mc_list.next,
 888                                   struct cma_multicast, list);
 889                 list_del(&mc->list);
 890                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 891                 case IB_LINK_LAYER_INFINIBAND:
 892                         ib_sa_free_multicast(mc->multicast.ib);
 893                         kfree(mc);
 894                         break;
 895                 case IB_LINK_LAYER_ETHERNET:
 896                         kref_put(&mc->mcref, release_mc);
 897                         break;
 898                 default:
 899                         break;
 900                 }
 901         }
 902 }
 903
 904 void rdma_destroy_id(struct rdma_cm_id *id)
 905 {
 906         struct rdma_id_private *id_priv;
 907         enum rdma_cm_state state;
 908
 909         id_priv = container_of(id, struct rdma_id_private, id);
 910         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 911         cma_cancel_operation(id_priv, state);
 912
 913         /*
 914          * Wait for any active callback to finish.  New callbacks will find
 915          * the id_priv state set to destroying and abort.
 916          */
 917         mutex_lock(&id_priv->handler_mutex);
 918         mutex_unlock(&id_priv->handler_mutex);
 919
 920         if (id_priv->cma_dev) {
 921                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 922                 case RDMA_TRANSPORT_IB:
 923                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 924                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 925                         break;
 926                 case RDMA_TRANSPORT_IWARP:
 927                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 928                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 929                         break;
 930                 default:
 931                         break;
 932                 }
 933                 cma_leave_mc_groups(id_priv);
 934                 cma_release_dev(id_priv);
 935         }
 936
 937         cma_release_port(id_priv);
 938         cma_deref_id(id_priv);
 939         wait_for_completion(&id_priv->comp);
 940
 941         if (id_priv->internal_id)
 942                 cma_deref_id(id_priv->id.context);
 943
 944         kfree(id_priv->id.route.path_rec);
 945         kfree(id_priv);
 946 }
 947 EXPORT_SYMBOL(rdma_destroy_id);
 948
 949 static int cma_rep_recv(struct rdma_id_private *id_priv)
 950 {
 951         int ret;
 952
 953         ret = cma_modify_qp_rtr(id_priv, NULL);
 954         if (ret)
 955                 goto reject;
 956
 957         ret = cma_modify_qp_rts(id_priv, NULL);
 958         if (ret)
 959                 goto reject;
 960
 961         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 962         if (ret)
 963                 goto reject;
 964
 965         return 0;
 966 reject:
 967         cma_modify_qp_err(id_priv);
 968         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 969                        NULL, 0, NULL, 0);
 970         return ret;
 971 }
 972
 973 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 974 {
 975         if (id_priv->id.ps == RDMA_PS_SDP &&
 976             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 977             SDP_MAJ_VERSION)
 978                 return -EINVAL;
 979
 980         return 0;
 981 }
 982
 983 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 984                                    struct ib_cm_rep_event_param *rep_data,
 985                                    void *private_data)
 986 {
 987         event->param.conn.private_data = private_data;
 988         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 989         event->param.conn.responder_resources = rep_data->responder_resources;
 990         event->param.conn.initiator_depth = rep_data->initiator_depth;
 991         event->param.conn.flow_control = rep_data->flow_control;
 992         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 993         event->param.conn.srq = rep_data->srq;
 994         event->param.conn.qp_num = rep_data->remote_qpn;
 995 }
 996
 997 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 998 {
 999         struct rdma_id_private *id_priv = cm_id->context;
1000         struct rdma_cm_event event;
1001         int ret = 0;
1002
1003         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1004                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1005             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1006                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1007                 return 0;
1008
1009         memset(&event, 0, sizeof event);
1010         switch (ib_event->event) {
1011         case IB_CM_REQ_ERROR:
1012         case IB_CM_REP_ERROR:
1013                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1014                 event.status = -ETIMEDOUT;
1015                 break;
1016         case IB_CM_REP_RECEIVED:
1017                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1018                 if (event.status)
1019                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1020                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1021                         event.status = cma_rep_recv(id_priv);
1022                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1023                                                      RDMA_CM_EVENT_ESTABLISHED;
1024                 } else
1025                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1026                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1027                                        ib_event->private_data);
1028                 break;
1029         case IB_CM_RTU_RECEIVED:
1030         case IB_CM_USER_ESTABLISHED:
1031                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1032                 break;
1033         case IB_CM_DREQ_ERROR:
1034                 event.status = -ETIMEDOUT; /* fall through */
1035         case IB_CM_DREQ_RECEIVED:
1036         case IB_CM_DREP_RECEIVED:
1037                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1038                                    RDMA_CM_DISCONNECT))
1039                         goto out;
1040                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1041                 break;
1042         case IB_CM_TIMEWAIT_EXIT:
1043                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1044                 break;
1045         case IB_CM_MRA_RECEIVED:
1046                 /* ignore event */
1047                 goto out;
1048         case IB_CM_REJ_RECEIVED:
1049                 cma_modify_qp_err(id_priv);
1050                 event.status = ib_event->param.rej_rcvd.reason;
1051                 event.event = RDMA_CM_EVENT_REJECTED;
1052                 event.param.conn.private_data = ib_event->private_data;
1053                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1054                 break;
1055         default:
1056                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1057                        ib_event->event);
1058                 goto out;
1059         }
1060
1061         ret = id_priv->id.event_handler(&id_priv->id, &event);
1062         if (ret) {
1063                 /* Destroy the CM ID by returning a non-zero value. */
1064                 id_priv->cm_id.ib = NULL;
1065                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1066                 mutex_unlock(&id_priv->handler_mutex);
1067                 rdma_destroy_id(&id_priv->id);
1068                 return ret;
1069         }
1070 out:
1071         mutex_unlock(&id_priv->handler_mutex);
1072         return ret;
1073 }
1074
1075 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1076                                                struct ib_cm_event *ib_event)
1077 {
1078         struct rdma_id_private *id_priv;
1079         struct rdma_cm_id *id;
1080         struct rdma_route *rt;
1081         union cma_ip_addr *src, *dst;
1082         __be16 port;
1083         u8 ip_ver;
1084         int ret;
1085
1086         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1087                              &ip_ver, &port, &src, &dst))
1088                 goto err;
1089
1090         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1091                             listen_id->ps);
1092         if (IS_ERR(id))
1093                 goto err;
1094
1095         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1096                           ip_ver, port, src, dst);
1097
1098         rt = &id->route;
1099         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1100         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1101                                GFP_KERNEL);
1102         if (!rt->path_rec)
1103                 goto destroy_id;
1104
1105         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1106         if (rt->num_paths == 2)
1107                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1108
1109         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1110                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1111                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1112                 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1113         } else {
1114                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1115                                         &rt->addr.dev_addr);
1116                 if (ret)
1117                         goto destroy_id;
1118         }
1119         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1120
1121         id_priv = container_of(id, struct rdma_id_private, id);
1122         id_priv->state = RDMA_CM_CONNECT;
1123         return id_priv;
1124
1125 destroy_id:
1126         rdma_destroy_id(id);
1127 err:
1128         return NULL;
1129 }
1130
1131 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1132                                               struct ib_cm_event *ib_event)
1133 {
1134         struct rdma_id_private *id_priv;
1135         struct rdma_cm_id *id;
1136         union cma_ip_addr *src, *dst;
1137         __be16 port;
1138         u8 ip_ver;
1139         int ret;
1140
1141         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1142                             listen_id->ps);
1143         if (IS_ERR(id))
1144                 return NULL;
1145
1146
1147         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1148                              &ip_ver, &port, &src, &dst))
1149                 goto err;
1150
1151         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1152                           ip_ver, port, src, dst);
1153
1154         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1155                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1156                                         &id->route.addr.dev_addr);
1157                 if (ret)
1158                         goto err;
1159         }
1160
1161         id_priv = container_of(id, struct rdma_id_private, id);
1162         id_priv->state = RDMA_CM_CONNECT;
1163         return id_priv;
1164 err:
1165         rdma_destroy_id(id);
1166         return NULL;
1167 }
1168
1169 static void cma_set_req_event_data(struct rdma_cm_event *event,
1170                                    struct ib_cm_req_event_param *req_data,
1171                                    void *private_data, int offset)
1172 {
1173         event->param.conn.private_data = private_data + offset;
1174         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1175         event->param.conn.responder_resources = req_data->responder_resources;
1176         event->param.conn.initiator_depth = req_data->initiator_depth;
1177         event->param.conn.flow_control = req_data->flow_control;
1178         event->param.conn.retry_count = req_data->retry_count;
1179         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1180         event->param.conn.srq = req_data->srq;
1181         event->param.conn.qp_num = req_data->remote_qpn;
1182 }
1183
1184 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1185 {
1186         struct rdma_id_private *listen_id, *conn_id;
1187         struct rdma_cm_event event;
1188         int offset, ret;
1189
1190         listen_id = cm_id->context;
1191         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1192                 return -ECONNABORTED;
1193
1194         memset(&event, 0, sizeof event);
1195         offset = cma_user_data_offset(listen_id->id.ps);
1196         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1197         if (cma_is_ud_ps(listen_id->id.ps)) {
1198                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1199                 event.param.ud.private_data = ib_event->private_data + offset;
1200                 event.param.ud.private_data_len =
1201                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1202         } else {
1203                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1204                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1205                                        ib_event->private_data, offset);
1206         }
1207         if (!conn_id) {
1208                 ret = -ENOMEM;
1209                 goto out;
1210         }
1211
1212         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1213         ret = cma_acquire_dev(conn_id);
1214         if (ret)
1215                 goto release_conn_id;
1216
1217         conn_id->cm_id.ib = cm_id;
1218         cm_id->context = conn_id;
1219         cm_id->cm_handler = cma_ib_handler;
1220
1221         /*
1222          * Protect against the user destroying conn_id from another thread
1223          * until we're done accessing it.
1224          */
1225         atomic_inc(&conn_id->refcount);
1226         ret = conn_id->id.event_handler(&conn_id->id, &event);
1227         if (!ret) {
1228                 /*
1229                  * Acquire mutex to prevent user executing rdma_destroy_id()
1230                  * while we're accessing the cm_id.
1231                  */
1232                 mutex_lock(&lock);
1233                 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1234                     !cma_is_ud_ps(conn_id->id.ps))
1235                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1236                 mutex_unlock(&lock);
1237                 mutex_unlock(&conn_id->handler_mutex);
1238                 cma_deref_id(conn_id);
1239                 goto out;
1240         }
1241         cma_deref_id(conn_id);
1242
1243         /* Destroy the CM ID by returning a non-zero value. */
1244         conn_id->cm_id.ib = NULL;
1245
1246 release_conn_id:
1247         cma_exch(conn_id, RDMA_CM_DESTROYING);
1248         mutex_unlock(&conn_id->handler_mutex);
1249         rdma_destroy_id(&conn_id->id);
1250
1251 out:
1252         mutex_unlock(&listen_id->handler_mutex);
1253         return ret;
1254 }
1255
1256 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1257 {
1258         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1259 }
1260
1261 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1262                                  struct ib_cm_compare_data *compare)
1263 {
1264         struct cma_hdr *cma_data, *cma_mask;
1265         struct sdp_hh *sdp_data, *sdp_mask;
1266         __be32 ip4_addr;
1267         struct in6_addr ip6_addr;
1268
1269         memset(compare, 0, sizeof *compare);
1270         cma_data = (void *) compare->data;
1271         cma_mask = (void *) compare->mask;
1272         sdp_data = (void *) compare->data;
1273         sdp_mask = (void *) compare->mask;
1274
1275         switch (addr->sa_family) {
1276         case AF_INET:
1277                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1278                 if (ps == RDMA_PS_SDP) {
1279                         sdp_set_ip_ver(sdp_data, 4);
1280                         sdp_set_ip_ver(sdp_mask, 0xF);
1281                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1282                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1283                 } else {
1284                         cma_set_ip_ver(cma_data, 4);
1285                         cma_set_ip_ver(cma_mask, 0xF);
1286                         cma_data->dst_addr.ip4.addr = ip4_addr;
1287                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1288                 }
1289                 break;
1290         case AF_INET6:
1291                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1292                 if (ps == RDMA_PS_SDP) {
1293                         sdp_set_ip_ver(sdp_data, 6);
1294                         sdp_set_ip_ver(sdp_mask, 0xF);
1295                         sdp_data->dst_addr.ip6 = ip6_addr;
1296                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1297                                sizeof sdp_mask->dst_addr.ip6);
1298                 } else {
1299                         cma_set_ip_ver(cma_data, 6);
1300                         cma_set_ip_ver(cma_mask, 0xF);
1301                         cma_data->dst_addr.ip6 = ip6_addr;
1302                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1303                                sizeof cma_mask->dst_addr.ip6);
1304                 }
1305                 break;
1306         default:
1307                 break;
1308         }
1309 }
1310
1311 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1312 {
1313         struct rdma_id_private *id_priv = iw_id->context;
1314         struct rdma_cm_event event;
1315         struct sockaddr_in *sin;
1316         int ret = 0;
1317
1318         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1319                 return 0;
1320
1321         memset(&event, 0, sizeof event);
1322         switch (iw_event->event) {
1323         case IW_CM_EVENT_CLOSE:
1324                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1325                 break;
1326         case IW_CM_EVENT_CONNECT_REPLY:
1327                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1328                 *sin = iw_event->local_addr;
1329                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1330                 *sin = iw_event->remote_addr;
1331                 switch (iw_event->status) {
1332                 case 0:
1333                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1334                         break;
1335                 case -ECONNRESET:
1336                 case -ECONNREFUSED:
1337                         event.event = RDMA_CM_EVENT_REJECTED;
1338                         break;
1339                 case -ETIMEDOUT:
1340                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1341                         break;
1342                 default:
1343                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1344                         break;
1345                 }
1346                 break;
1347         case IW_CM_EVENT_ESTABLISHED:
1348                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1349                 break;
1350         default:
1351                 BUG_ON(1);
1352         }
1353
1354         event.status = iw_event->status;
1355         event.param.conn.private_data = iw_event->private_data;
1356         event.param.conn.private_data_len = iw_event->private_data_len;
1357         ret = id_priv->id.event_handler(&id_priv->id, &event);
1358         if (ret) {
1359                 /* Destroy the CM ID by returning a non-zero value. */
1360                 id_priv->cm_id.iw = NULL;
1361                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1362                 mutex_unlock(&id_priv->handler_mutex);
1363                 rdma_destroy_id(&id_priv->id);
1364                 return ret;
1365         }
1366
1367         mutex_unlock(&id_priv->handler_mutex);
1368         return ret;
1369 }
1370
1371 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1372                                struct iw_cm_event *iw_event)
1373 {
1374         struct rdma_cm_id *new_cm_id;
1375         struct rdma_id_private *listen_id, *conn_id;
1376         struct sockaddr_in *sin;
1377         struct net_device *dev = NULL;
1378         struct rdma_cm_event event;
1379         int ret;
1380         struct ib_device_attr attr;
1381
1382         listen_id = cm_id->context;
1383         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1384                 return -ECONNABORTED;
1385
1386         /* Create a new RDMA id for the new IW CM ID */
1387         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1388                                    listen_id->id.context,
1389                                    RDMA_PS_TCP);
1390         if (IS_ERR(new_cm_id)) {
1391                 ret = -ENOMEM;
1392                 goto out;
1393         }
1394         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1395         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1396         conn_id->state = RDMA_CM_CONNECT;
1397
1398         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1399         if (!dev) {
1400                 ret = -EADDRNOTAVAIL;
1401                 mutex_unlock(&conn_id->handler_mutex);
1402                 rdma_destroy_id(new_cm_id);
1403                 goto out;
1404         }
1405         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1406         if (ret) {
1407                 mutex_unlock(&conn_id->handler_mutex);
1408                 rdma_destroy_id(new_cm_id);
1409                 goto out;
1410         }
1411
1412         ret = cma_acquire_dev(conn_id);
1413         if (ret) {
1414                 mutex_unlock(&conn_id->handler_mutex);
1415                 rdma_destroy_id(new_cm_id);
1416                 goto out;
1417         }
1418
1419         conn_id->cm_id.iw = cm_id;
1420         cm_id->context = conn_id;
1421         cm_id->cm_handler = cma_iw_handler;
1422
1423         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1424         *sin = iw_event->local_addr;
1425         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1426         *sin = iw_event->remote_addr;
1427
1428         ret = ib_query_device(conn_id->id.device, &attr);
1429         if (ret) {
1430                 mutex_unlock(&conn_id->handler_mutex);
1431                 rdma_destroy_id(new_cm_id);
1432                 goto out;
1433         }
1434
1435         memset(&event, 0, sizeof event);
1436         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1437         event.param.conn.private_data = iw_event->private_data;
1438         event.param.conn.private_data_len = iw_event->private_data_len;
1439         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1440         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1441
1442         /*
1443          * Protect against the user destroying conn_id from another thread
1444          * until we're done accessing it.
1445          */
1446         atomic_inc(&conn_id->refcount);
1447         ret = conn_id->id.event_handler(&conn_id->id, &event);
1448         if (ret) {
1449                 /* User wants to destroy the CM ID */
1450                 conn_id->cm_id.iw = NULL;
1451                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1452                 mutex_unlock(&conn_id->handler_mutex);
1453                 cma_deref_id(conn_id);
1454                 rdma_destroy_id(&conn_id->id);
1455                 goto out;
1456         }
1457
1458         mutex_unlock(&conn_id->handler_mutex);
1459         cma_deref_id(conn_id);
1460
1461 out:
1462         if (dev)
1463                 dev_put(dev);
1464         mutex_unlock(&listen_id->handler_mutex);
1465         return ret;
1466 }
1467
1468 static int cma_ib_listen(struct rdma_id_private *id_priv)
1469 {
1470         struct ib_cm_compare_data compare_data;
1471         struct sockaddr *addr;
1472         __be64 svc_id;
1473         int ret;
1474
1475         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1476                                             id_priv);
1477         if (IS_ERR(id_priv->cm_id.ib))
1478                 return PTR_ERR(id_priv->cm_id.ib);
1479
1480         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1481         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1482         if (cma_any_addr(addr))
1483                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1484         else {
1485                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1486                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1487         }
1488
1489         if (ret) {
1490                 ib_destroy_cm_id(id_priv->cm_id.ib);
1491                 id_priv->cm_id.ib = NULL;
1492         }
1493
1494         return ret;
1495 }
1496
1497 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1498 {
1499         int ret;
1500         struct sockaddr_in *sin;
1501
1502         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1503                                             iw_conn_req_handler,
1504                                             id_priv);
1505         if (IS_ERR(id_priv->cm_id.iw))
1506                 return PTR_ERR(id_priv->cm_id.iw);
1507
1508         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1509         id_priv->cm_id.iw->local_addr = *sin;
1510
1511         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1512
1513         if (ret) {
1514                 iw_destroy_cm_id(id_priv->cm_id.iw);
1515                 id_priv->cm_id.iw = NULL;
1516         }
1517
1518         return ret;
1519 }
1520
1521 static int cma_listen_handler(struct rdma_cm_id *id,
1522                               struct rdma_cm_event *event)
1523 {
1524         struct rdma_id_private *id_priv = id->context;
1525
1526         id->context = id_priv->id.context;
1527         id->event_handler = id_priv->id.event_handler;
1528         return id_priv->id.event_handler(id, event);
1529 }
1530
1531 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1532                               struct cma_device *cma_dev)
1533 {
1534         struct rdma_id_private *dev_id_priv;
1535         struct rdma_cm_id *id;
1536         int ret;
1537
1538         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1539         if (IS_ERR(id))
1540                 return;
1541
1542         dev_id_priv = container_of(id, struct rdma_id_private, id);
1543
1544         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1545         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1546                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1547
1548         cma_attach_to_dev(dev_id_priv, cma_dev);
1549         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1550         atomic_inc(&id_priv->refcount);
1551         dev_id_priv->internal_id = 1;
1552
1553         ret = rdma_listen(id, id_priv->backlog);
1554         if (ret)
1555                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1556                        "listening on device %s\n", ret, cma_dev->device->name);
1557 }
1558
1559 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1560 {
1561         struct cma_device *cma_dev;
1562
1563         mutex_lock(&lock);
1564         list_add_tail(&id_priv->list, &listen_any_list);
1565         list_for_each_entry(cma_dev, &dev_list, list)
1566                 cma_listen_on_dev(id_priv, cma_dev);
1567         mutex_unlock(&lock);
1568 }
1569
1570 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1571 {
1572         struct rdma_id_private *id_priv;
1573
1574         id_priv = container_of(id, struct rdma_id_private, id);
1575         id_priv->tos = (u8) tos;
1576 }
1577 EXPORT_SYMBOL(rdma_set_service_type);
1578
1579 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1580                               void *context)
1581 {
1582         struct cma_work *work = context;
1583         struct rdma_route *route;
1584
1585         route = &work->id->id.route;
1586
1587         if (!status) {
1588                 route->num_paths = 1;
1589                 *route->path_rec = *path_rec;
1590         } else {
1591                 work->old_state = RDMA_CM_ROUTE_QUERY;
1592                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1593                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1594                 work->event.status = status;
1595         }
1596
1597         queue_work(cma_wq, &work->work);
1598 }
1599
1600 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1601                               struct cma_work *work)
1602 {
1603         struct rdma_addr *addr = &id_priv->id.route.addr;
1604         struct ib_sa_path_rec path_rec;
1605         ib_sa_comp_mask comp_mask;
1606         struct sockaddr_in6 *sin6;
1607
1608         memset(&path_rec, 0, sizeof path_rec);
1609         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1610         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1611         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1612         path_rec.numb_path = 1;
1613         path_rec.reversible = 1;
1614         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1615                                                         (struct sockaddr *) &addr->dst_addr);
1616
1617         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1618                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1619                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1620
1621         if (addr->src_addr.ss_family == AF_INET) {
1622                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1623                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1624         } else {
1625                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1626                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1627                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1628         }
1629
1630         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1631                                                id_priv->id.port_num, &path_rec,
1632                                                comp_mask, timeout_ms,
1633                                                GFP_KERNEL, cma_query_handler,
1634                                                work, &id_priv->query);
1635
1636         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1637 }
1638
1639 static void cma_work_handler(struct work_struct *_work)
1640 {
1641         struct cma_work *work = container_of(_work, struct cma_work, work);
1642         struct rdma_id_private *id_priv = work->id;
1643         int destroy = 0;
1644
1645         mutex_lock(&id_priv->handler_mutex);
1646         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1647                 goto out;
1648
1649         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1650                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1651                 destroy = 1;
1652         }
1653 out:
1654         mutex_unlock(&id_priv->handler_mutex);
1655         cma_deref_id(id_priv);
1656         if (destroy)
1657                 rdma_destroy_id(&id_priv->id);
1658         kfree(work);
1659 }
1660
1661 static void cma_ndev_work_handler(struct work_struct *_work)
1662 {
1663         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1664         struct rdma_id_private *id_priv = work->id;
1665         int destroy = 0;
1666
1667         mutex_lock(&id_priv->handler_mutex);
1668         if (id_priv->state == RDMA_CM_DESTROYING ||
1669             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1670                 goto out;
1671
1672         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1673                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1674                 destroy = 1;
1675         }
1676
1677 out:
1678         mutex_unlock(&id_priv->handler_mutex);
1679         cma_deref_id(id_priv);
1680         if (destroy)
1681                 rdma_destroy_id(&id_priv->id);
1682         kfree(work);
1683 }
1684
1685 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1686 {
1687         struct rdma_route *route = &id_priv->id.route;
1688         struct cma_work *work;
1689         int ret;
1690
1691         work = kzalloc(sizeof *work, GFP_KERNEL);
1692         if (!work)
1693                 return -ENOMEM;
1694
1695         work->id = id_priv;
1696         INIT_WORK(&work->work, cma_work_handler);
1697         work->old_state = RDMA_CM_ROUTE_QUERY;
1698         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1699         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1700
1701         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1702         if (!route->path_rec) {
1703                 ret = -ENOMEM;
1704                 goto err1;
1705         }
1706
1707         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1708         if (ret)
1709                 goto err2;
1710
1711         return 0;
1712 err2:
1713         kfree(route->path_rec);
1714         route->path_rec = NULL;
1715 err1:
1716         kfree(work);
1717         return ret;
1718 }
1719
1720 int rdma_set_ib_paths(struct rdma_cm_id *id,
1721                       struct ib_sa_path_rec *path_rec, int num_paths)
1722 {
1723         struct rdma_id_private *id_priv;
1724         int ret;
1725
1726         id_priv = container_of(id, struct rdma_id_private, id);
1727         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1728                            RDMA_CM_ROUTE_RESOLVED))
1729                 return -EINVAL;
1730
1731         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1732                                      GFP_KERNEL);
1733         if (!id->route.path_rec) {
1734                 ret = -ENOMEM;
1735                 goto err;
1736         }
1737
1738         id->route.num_paths = num_paths;
1739         return 0;
1740 err:
1741         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1742         return ret;
1743 }
1744 EXPORT_SYMBOL(rdma_set_ib_paths);
1745
1746 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1747 {
1748         struct cma_work *work;
1749
1750         work = kzalloc(sizeof *work, GFP_KERNEL);
1751         if (!work)
1752                 return -ENOMEM;
1753
1754         work->id = id_priv;
1755         INIT_WORK(&work->work, cma_work_handler);
1756         work->old_state = RDMA_CM_ROUTE_QUERY;
1757         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1758         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1759         queue_work(cma_wq, &work->work);
1760         return 0;
1761 }
1762
1763 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1764 {
1765         struct rdma_route *route = &id_priv->id.route;
1766         struct rdma_addr *addr = &route->addr;
1767         struct cma_work *work;
1768         int ret;
1769         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1770         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1771         struct net_device *ndev = NULL;
1772         u16 vid;
1773
1774         if (src_addr->sin_family != dst_addr->sin_family)
1775                 return -EINVAL;
1776
1777         work = kzalloc(sizeof *work, GFP_KERNEL);
1778         if (!work)
1779                 return -ENOMEM;
1780
1781         work->id = id_priv;
1782         INIT_WORK(&work->work, cma_work_handler);
1783
1784         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1785         if (!route->path_rec) {
1786                 ret = -ENOMEM;
1787                 goto err1;
1788         }
1789
1790         route->num_paths = 1;
1791
1792         if (addr->dev_addr.bound_dev_if)
1793                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1794         if (!ndev) {
1795                 ret = -ENODEV;
1796                 goto err2;
1797         }
1798
1799         vid = rdma_vlan_dev_vlan_id(ndev);
1800
1801         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1802         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1803
1804         route->path_rec->hop_limit = 1;
1805         route->path_rec->reversible = 1;
1806         route->path_rec->pkey = cpu_to_be16(0xffff);
1807         route->path_rec->mtu_selector = IB_SA_EQ;
1808         route->path_rec->sl = id_priv->tos >> 5;
1809
1810         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1811         route->path_rec->rate_selector = IB_SA_EQ;
1812         route->path_rec->rate = iboe_get_rate(ndev);
1813         dev_put(ndev);
1814         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1815         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1816         if (!route->path_rec->mtu) {
1817                 ret = -EINVAL;
1818                 goto err2;
1819         }
1820
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         work->event.status = 0;
1825
1826         queue_work(cma_wq, &work->work);
1827
1828         return 0;
1829
1830 err2:
1831         kfree(route->path_rec);
1832         route->path_rec = NULL;
1833 err1:
1834         kfree(work);
1835         return ret;
1836 }
1837
1838 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1839 {
1840         struct rdma_id_private *id_priv;
1841         int ret;
1842
1843         id_priv = container_of(id, struct rdma_id_private, id);
1844         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1845                 return -EINVAL;
1846
1847         atomic_inc(&id_priv->refcount);
1848         switch (rdma_node_get_transport(id->device->node_type)) {
1849         case RDMA_TRANSPORT_IB:
1850                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1851                 case IB_LINK_LAYER_INFINIBAND:
1852                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1853                         break;
1854                 case IB_LINK_LAYER_ETHERNET:
1855                         ret = cma_resolve_iboe_route(id_priv);
1856                         break;
1857                 default:
1858                         ret = -ENOSYS;
1859                 }
1860                 break;
1861         case RDMA_TRANSPORT_IWARP:
1862                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1863                 break;
1864         default:
1865                 ret = -ENOSYS;
1866                 break;
1867         }
1868         if (ret)
1869                 goto err;
1870
1871         return 0;
1872 err:
1873         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1874         cma_deref_id(id_priv);
1875         return ret;
1876 }
1877 EXPORT_SYMBOL(rdma_resolve_route);
1878
1879 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1880 {
1881         struct cma_device *cma_dev;
1882         struct ib_port_attr port_attr;
1883         union ib_gid gid;
1884         u16 pkey;
1885         int ret;
1886         u8 p;
1887
1888         mutex_lock(&lock);
1889         if (list_empty(&dev_list)) {
1890                 ret = -ENODEV;
1891                 goto out;
1892         }
1893         list_for_each_entry(cma_dev, &dev_list, list)
1894                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1895                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1896                             port_attr.state == IB_PORT_ACTIVE)
1897                                 goto port_found;
1898
1899         p = 1;
1900         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1901
1902 port_found:
1903         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1904         if (ret)
1905                 goto out;
1906
1907         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1908         if (ret)
1909                 goto out;
1910
1911         id_priv->id.route.addr.dev_addr.dev_type =
1912                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1913                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1914
1915         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1916         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1917         id_priv->id.port_num = p;
1918         cma_attach_to_dev(id_priv, cma_dev);
1919 out:
1920         mutex_unlock(&lock);
1921         return ret;
1922 }
1923
1924 static void addr_handler(int status, struct sockaddr *src_addr,
1925                          struct rdma_dev_addr *dev_addr, void *context)
1926 {
1927         struct rdma_id_private *id_priv = context;
1928         struct rdma_cm_event event;
1929
1930         memset(&event, 0, sizeof event);
1931         mutex_lock(&id_priv->handler_mutex);
1932         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1933                            RDMA_CM_ADDR_RESOLVED))
1934                 goto out;
1935
1936         if (!status && !id_priv->cma_dev)
1937                 status = cma_acquire_dev(id_priv);
1938
1939         if (status) {
1940                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1941                                    RDMA_CM_ADDR_BOUND))
1942                         goto out;
1943                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1944                 event.status = status;
1945         } else {
1946                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1947                        ip_addr_size(src_addr));
1948                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1949         }
1950
1951         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1952                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1953                 mutex_unlock(&id_priv->handler_mutex);
1954                 cma_deref_id(id_priv);
1955                 rdma_destroy_id(&id_priv->id);
1956                 return;
1957         }
1958 out:
1959         mutex_unlock(&id_priv->handler_mutex);
1960         cma_deref_id(id_priv);
1961 }
1962
1963 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1964 {
1965         struct cma_work *work;
1966         struct sockaddr *src, *dst;
1967         union ib_gid gid;
1968         int ret;
1969
1970         work = kzalloc(sizeof *work, GFP_KERNEL);
1971         if (!work)
1972                 return -ENOMEM;
1973
1974         if (!id_priv->cma_dev) {
1975                 ret = cma_bind_loopback(id_priv);
1976                 if (ret)
1977                         goto err;
1978         }
1979
1980         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1981         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1982
1983         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1984         if (cma_zero_addr(src)) {
1985                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
1986                 if ((src->sa_family = dst->sa_family) == AF_INET) {
1987                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
1988                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1989                 } else {
1990                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
1991                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
1992                 }
1993         }
1994
1995         work->id = id_priv;
1996         INIT_WORK(&work->work, cma_work_handler);
1997         work->old_state = RDMA_CM_ADDR_QUERY;
1998         work->new_state = RDMA_CM_ADDR_RESOLVED;
1999         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2000         queue_work(cma_wq, &work->work);
2001         return 0;
2002 err:
2003         kfree(work);
2004         return ret;
2005 }
2006
2007 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2008                          struct sockaddr *dst_addr)
2009 {
2010         if (!src_addr || !src_addr->sa_family) {
2011                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2012                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2013                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2014                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2015                 }
2016         }
2017         return rdma_bind_addr(id, src_addr);
2018 }
2019
2020 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2021                       struct sockaddr *dst_addr, int timeout_ms)
2022 {
2023         struct rdma_id_private *id_priv;
2024         int ret;
2025
2026         id_priv = container_of(id, struct rdma_id_private, id);
2027         if (id_priv->state == RDMA_CM_IDLE) {
2028                 ret = cma_bind_addr(id, src_addr, dst_addr);
2029                 if (ret)
2030                         return ret;
2031         }
2032
2033         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2034                 return -EINVAL;
2035
2036         atomic_inc(&id_priv->refcount);
2037         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2038         if (cma_any_addr(dst_addr))
2039                 ret = cma_resolve_loopback(id_priv);
2040         else
2041                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2042                                       dst_addr, &id->route.addr.dev_addr,
2043                                       timeout_ms, addr_handler, id_priv);
2044         if (ret)
2045                 goto err;
2046
2047         return 0;
2048 err:
2049         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2050         cma_deref_id(id_priv);
2051         return ret;
2052 }
2053 EXPORT_SYMBOL(rdma_resolve_addr);
2054
2055 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2056 {
2057         struct rdma_id_private *id_priv;
2058         unsigned long flags;
2059         int ret;
2060
2061         id_priv = container_of(id, struct rdma_id_private, id);
2062         spin_lock_irqsave(&id_priv->lock, flags);
2063         if (id_priv->state == RDMA_CM_IDLE) {
2064                 id_priv->reuseaddr = reuse;
2065                 ret = 0;
2066         } else {
2067                 ret = -EINVAL;
2068         }
2069         spin_unlock_irqrestore(&id_priv->lock, flags);
2070         return ret;
2071 }
2072 EXPORT_SYMBOL(rdma_set_reuseaddr);
2073
2074 static void cma_bind_port(struct rdma_bind_list *bind_list,
2075                           struct rdma_id_private *id_priv)
2076 {
2077         struct sockaddr_in *sin;
2078
2079         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2080         sin->sin_port = htons(bind_list->port);
2081         id_priv->bind_list = bind_list;
2082         hlist_add_head(&id_priv->node, &bind_list->owners);
2083 }
2084
2085 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2086                           unsigned short snum)
2087 {
2088         struct rdma_bind_list *bind_list;
2089         int port, ret;
2090
2091         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2092         if (!bind_list)
2093                 return -ENOMEM;
2094
2095         do {
2096                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2097         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2098
2099         if (ret)
2100                 goto err1;
2101
2102         if (port != snum) {
2103                 ret = -EADDRNOTAVAIL;
2104                 goto err2;
2105         }
2106
2107         bind_list->ps = ps;
2108         bind_list->port = (unsigned short) port;
2109         cma_bind_port(bind_list, id_priv);
2110         return 0;
2111 err2:
2112         idr_remove(ps, port);
2113 err1:
2114         kfree(bind_list);
2115         return ret;
2116 }
2117
2118 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2119 {
2120         static unsigned int last_used_port;
2121         int low, high, remaining;
2122         unsigned int rover;
2123
2124         inet_get_local_port_range(&low, &high);
2125         remaining = (high - low) + 1;
2126         rover = net_random() % remaining + low;
2127 retry:
2128         if (last_used_port != rover &&
2129             !idr_find(ps, (unsigned short) rover)) {
2130                 int ret = cma_alloc_port(ps, id_priv, rover);
2131                 /*
2132                  * Remember previously used port number in order to avoid
2133                  * re-using same port immediately after it is closed.
2134                  */
2135                 if (!ret)
2136                         last_used_port = rover;
2137                 if (ret != -EADDRNOTAVAIL)
2138                         return ret;
2139         }
2140         if (--remaining) {
2141                 rover++;
2142                 if ((rover < low) || (rover > high))
2143                         rover = low;
2144                 goto retry;
2145         }
2146         return -EADDRNOTAVAIL;
2147 }
2148
2149 /*
2150  * Check that the requested port is available.  This is called when trying to
2151  * bind to a specific port, or when trying to listen on a bound port.  In
2152  * the latter case, the provided id_priv may already be on the bind_list, but
2153  * we still need to check that it's okay to start listening.
2154  */
2155 static int cma_check_port(struct rdma_bind_list *bind_list,
2156                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2157 {
2158         struct rdma_id_private *cur_id;
2159         struct sockaddr *addr, *cur_addr;
2160         struct hlist_node *node;
2161
2162         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2163         if (cma_any_addr(addr) && !reuseaddr)
2164                 return -EADDRNOTAVAIL;
2165
2166         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2167                 if (id_priv == cur_id)
2168                         continue;
2169
2170                 if ((cur_id->state == RDMA_CM_LISTEN) ||
2171                     !reuseaddr || !cur_id->reuseaddr) {
2172                         cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2173                         if (cma_any_addr(cur_addr))
2174                                 return -EADDRNOTAVAIL;
2175
2176                         if (!cma_addr_cmp(addr, cur_addr))
2177                                 return -EADDRINUSE;
2178                 }
2179         }
2180         return 0;
2181 }
2182
2183 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2184 {
2185         struct rdma_bind_list *bind_list;
2186         unsigned short snum;
2187         int ret;
2188
2189         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2190         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2191                 return -EACCES;
2192
2193         bind_list = idr_find(ps, snum);
2194         if (!bind_list) {
2195                 ret = cma_alloc_port(ps, id_priv, snum);
2196         } else {
2197                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2198                 if (!ret)
2199                         cma_bind_port(bind_list, id_priv);
2200         }
2201         return ret;
2202 }
2203
2204 static int cma_bind_listen(struct rdma_id_private *id_priv)
2205 {
2206         struct rdma_bind_list *bind_list = id_priv->bind_list;
2207         int ret = 0;
2208
2209         mutex_lock(&lock);
2210         if (bind_list->owners.first->next)
2211                 ret = cma_check_port(bind_list, id_priv, 0);
2212         mutex_unlock(&lock);
2213         return ret;
2214 }
2215
2216 static int cma_get_port(struct rdma_id_private *id_priv)
2217 {
2218         struct idr *ps;
2219         int ret;
2220
2221         switch (id_priv->id.ps) {
2222         case RDMA_PS_SDP:
2223                 ps = &sdp_ps;
2224                 break;
2225         case RDMA_PS_TCP:
2226                 ps = &tcp_ps;
2227                 break;
2228         case RDMA_PS_UDP:
2229                 ps = &udp_ps;
2230                 break;
2231         case RDMA_PS_IPOIB:
2232                 ps = &ipoib_ps;
2233                 break;
2234         default:
2235                 return -EPROTONOSUPPORT;
2236         }
2237
2238         mutex_lock(&lock);
2239         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2240                 ret = cma_alloc_any_port(ps, id_priv);
2241         else
2242                 ret = cma_use_port(ps, id_priv);
2243         mutex_unlock(&lock);
2244
2245         return ret;
2246 }
2247
2248 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2249                                struct sockaddr *addr)
2250 {
2251 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2252         struct sockaddr_in6 *sin6;
2253
2254         if (addr->sa_family != AF_INET6)
2255                 return 0;
2256
2257         sin6 = (struct sockaddr_in6 *) addr;
2258         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2259             !sin6->sin6_scope_id)
2260                         return -EINVAL;
2261
2262         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2263 #endif
2264         return 0;
2265 }
2266
2267 int rdma_listen(struct rdma_cm_id *id, int backlog)
2268 {
2269         struct rdma_id_private *id_priv;
2270         int ret;
2271
2272         id_priv = container_of(id, struct rdma_id_private, id);
2273         if (id_priv->state == RDMA_CM_IDLE) {
2274                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2275                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2276                 if (ret)
2277                         return ret;
2278         }
2279
2280         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2281                 return -EINVAL;
2282
2283         if (id_priv->reuseaddr) {
2284                 ret = cma_bind_listen(id_priv);
2285                 if (ret)
2286                         goto err;
2287         }
2288
2289         id_priv->backlog = backlog;
2290         if (id->device) {
2291                 switch (rdma_node_get_transport(id->device->node_type)) {
2292                 case RDMA_TRANSPORT_IB:
2293                         ret = cma_ib_listen(id_priv);
2294                         if (ret)
2295                                 goto err;
2296                         break;
2297                 case RDMA_TRANSPORT_IWARP:
2298                         ret = cma_iw_listen(id_priv, backlog);
2299                         if (ret)
2300                                 goto err;
2301                         break;
2302                 default:
2303                         ret = -ENOSYS;
2304                         goto err;
2305                 }
2306         } else
2307                 cma_listen_on_all(id_priv);
2308
2309         return 0;
2310 err:
2311         id_priv->backlog = 0;
2312         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2313         return ret;
2314 }
2315 EXPORT_SYMBOL(rdma_listen);
2316
2317 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2318 {
2319         struct rdma_id_private *id_priv;
2320         int ret;
2321
2322         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2323                 return -EAFNOSUPPORT;
2324
2325         id_priv = container_of(id, struct rdma_id_private, id);
2326         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2327                 return -EINVAL;
2328
2329         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2330         if (ret)
2331                 goto err1;
2332
2333         if (!cma_any_addr(addr)) {
2334                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2335                 if (ret)
2336                         goto err1;
2337
2338                 ret = cma_acquire_dev(id_priv);
2339                 if (ret)
2340                         goto err1;
2341         }
2342
2343         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2344         ret = cma_get_port(id_priv);
2345         if (ret)
2346                 goto err2;
2347
2348         return 0;
2349 err2:
2350         if (id_priv->cma_dev)
2351                 cma_release_dev(id_priv);
2352 err1:
2353         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2354         return ret;
2355 }
2356 EXPORT_SYMBOL(rdma_bind_addr);
2357
2358 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2359                           struct rdma_route *route)
2360 {
2361         struct cma_hdr *cma_hdr;
2362         struct sdp_hh *sdp_hdr;
2363
2364         if (route->addr.src_addr.ss_family == AF_INET) {
2365                 struct sockaddr_in *src4, *dst4;
2366
2367                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2368                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2369
2370                 switch (ps) {
2371                 case RDMA_PS_SDP:
2372                         sdp_hdr = hdr;
2373                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2374                                 return -EINVAL;
2375                         sdp_set_ip_ver(sdp_hdr, 4);
2376                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2377                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2378                         sdp_hdr->port = src4->sin_port;
2379                         break;
2380                 default:
2381                         cma_hdr = hdr;
2382                         cma_hdr->cma_version = CMA_VERSION;
2383                         cma_set_ip_ver(cma_hdr, 4);
2384                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2385                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2386                         cma_hdr->port = src4->sin_port;
2387                         break;
2388                 }
2389         } else {
2390                 struct sockaddr_in6 *src6, *dst6;
2391
2392                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2393                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2394
2395                 switch (ps) {
2396                 case RDMA_PS_SDP:
2397                         sdp_hdr = hdr;
2398                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2399                                 return -EINVAL;
2400                         sdp_set_ip_ver(sdp_hdr, 6);
2401                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2402                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2403                         sdp_hdr->port = src6->sin6_port;
2404                         break;
2405                 default:
2406                         cma_hdr = hdr;
2407                         cma_hdr->cma_version = CMA_VERSION;
2408                         cma_set_ip_ver(cma_hdr, 6);
2409                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2410                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2411                         cma_hdr->port = src6->sin6_port;
2412                         break;
2413                 }
2414         }
2415         return 0;
2416 }
2417
2418 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2419                                 struct ib_cm_event *ib_event)
2420 {
2421         struct rdma_id_private *id_priv = cm_id->context;
2422         struct rdma_cm_event event;
2423         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2424         int ret = 0;
2425
2426         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2427                 return 0;
2428
2429         memset(&event, 0, sizeof event);
2430         switch (ib_event->event) {
2431         case IB_CM_SIDR_REQ_ERROR:
2432                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2433                 event.status = -ETIMEDOUT;
2434                 break;
2435         case IB_CM_SIDR_REP_RECEIVED:
2436                 event.param.ud.private_data = ib_event->private_data;
2437                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2438                 if (rep->status != IB_SIDR_SUCCESS) {
2439                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2440                         event.status = ib_event->param.sidr_rep_rcvd.status;
2441                         break;
2442                 }
2443                 ret = cma_set_qkey(id_priv);
2444                 if (ret) {
2445                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2446                         event.status = -EINVAL;
2447                         break;
2448                 }
2449                 if (id_priv->qkey != rep->qkey) {
2450                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2451                         event.status = -EINVAL;
2452                         break;
2453                 }
2454                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2455                                      id_priv->id.route.path_rec,
2456                                      &event.param.ud.ah_attr);
2457                 event.param.ud.qp_num = rep->qpn;
2458                 event.param.ud.qkey = rep->qkey;
2459                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2460                 event.status = 0;
2461                 break;
2462         default:
2463                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2464                        ib_event->event);
2465                 goto out;
2466         }
2467
2468         ret = id_priv->id.event_handler(&id_priv->id, &event);
2469         if (ret) {
2470                 /* Destroy the CM ID by returning a non-zero value. */
2471                 id_priv->cm_id.ib = NULL;
2472                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2473                 mutex_unlock(&id_priv->handler_mutex);
2474                 rdma_destroy_id(&id_priv->id);
2475                 return ret;
2476         }
2477 out:
2478         mutex_unlock(&id_priv->handler_mutex);
2479         return ret;
2480 }
2481
2482 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2483                               struct rdma_conn_param *conn_param)
2484 {
2485         struct ib_cm_sidr_req_param req;
2486         struct rdma_route *route;
2487         int ret;
2488
2489         req.private_data_len = sizeof(struct cma_hdr) +
2490                                conn_param->private_data_len;
2491         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2492         if (!req.private_data)
2493                 return -ENOMEM;
2494
2495         if (conn_param->private_data && conn_param->private_data_len)
2496                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2497                        conn_param->private_data, conn_param->private_data_len);
2498
2499         route = &id_priv->id.route;
2500         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2501         if (ret)
2502                 goto out;
2503
2504         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2505                                             cma_sidr_rep_handler, id_priv);
2506         if (IS_ERR(id_priv->cm_id.ib)) {
2507                 ret = PTR_ERR(id_priv->cm_id.ib);
2508                 goto out;
2509         }
2510
2511         req.path = route->path_rec;
2512         req.service_id = cma_get_service_id(id_priv->id.ps,
2513                                             (struct sockaddr *) &route->addr.dst_addr);
2514         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2515         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2516
2517         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2518         if (ret) {
2519                 ib_destroy_cm_id(id_priv->cm_id.ib);
2520                 id_priv->cm_id.ib = NULL;
2521         }
2522 out:
2523         kfree(req.private_data);
2524         return ret;
2525 }
2526
2527 static int cma_connect_ib(struct rdma_id_private *id_priv,
2528                           struct rdma_conn_param *conn_param)
2529 {
2530         struct ib_cm_req_param req;
2531         struct rdma_route *route;
2532         void *private_data;
2533         int offset, ret;
2534
2535         memset(&req, 0, sizeof req);
2536         offset = cma_user_data_offset(id_priv->id.ps);
2537         req.private_data_len = offset + conn_param->private_data_len;
2538         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2539         if (!private_data)
2540                 return -ENOMEM;
2541
2542         if (conn_param->private_data && conn_param->private_data_len)
2543                 memcpy(private_data + offset, conn_param->private_data,
2544                        conn_param->private_data_len);
2545
2546         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2547                                             id_priv);
2548         if (IS_ERR(id_priv->cm_id.ib)) {
2549                 ret = PTR_ERR(id_priv->cm_id.ib);
2550                 goto out;
2551         }
2552
2553         route = &id_priv->id.route;
2554         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2555         if (ret)
2556                 goto out;
2557         req.private_data = private_data;
2558
2559         req.primary_path = &route->path_rec[0];
2560         if (route->num_paths == 2)
2561                 req.alternate_path = &route->path_rec[1];
2562
2563         req.service_id = cma_get_service_id(id_priv->id.ps,
2564                                             (struct sockaddr *) &route->addr.dst_addr);
2565         req.qp_num = id_priv->qp_num;
2566         req.qp_type = IB_QPT_RC;
2567         req.starting_psn = id_priv->seq_num;
2568         req.responder_resources = conn_param->responder_resources;
2569         req.initiator_depth = conn_param->initiator_depth;
2570         req.flow_control = conn_param->flow_control;
2571         req.retry_count = conn_param->retry_count;
2572         req.rnr_retry_count = conn_param->rnr_retry_count;
2573         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2574         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2575         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2576         req.srq = id_priv->srq ? 1 : 0;
2577
2578         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2579 out:
2580         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2581                 ib_destroy_cm_id(id_priv->cm_id.ib);
2582                 id_priv->cm_id.ib = NULL;
2583         }
2584
2585         kfree(private_data);
2586         return ret;
2587 }
2588
2589 static int cma_connect_iw(struct rdma_id_private *id_priv,
2590                           struct rdma_conn_param *conn_param)
2591 {
2592         struct iw_cm_id *cm_id;
2593         struct sockaddr_in* sin;
2594         int ret;
2595         struct iw_cm_conn_param iw_param;
2596
2597         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2598         if (IS_ERR(cm_id)) {
2599                 ret = PTR_ERR(cm_id);
2600                 goto out;
2601         }
2602
2603         id_priv->cm_id.iw = cm_id;
2604
2605         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2606         cm_id->local_addr = *sin;
2607
2608         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2609         cm_id->remote_addr = *sin;
2610
2611         ret = cma_modify_qp_rtr(id_priv, conn_param);
2612         if (ret)
2613                 goto out;
2614
2615         iw_param.ord = conn_param->initiator_depth;
2616         iw_param.ird = conn_param->responder_resources;
2617         iw_param.private_data = conn_param->private_data;
2618         iw_param.private_data_len = conn_param->private_data_len;
2619         if (id_priv->id.qp)
2620                 iw_param.qpn = id_priv->qp_num;
2621         else
2622                 iw_param.qpn = conn_param->qp_num;
2623         ret = iw_cm_connect(cm_id, &iw_param);
2624 out:
2625         if (ret && !IS_ERR(cm_id)) {
2626                 iw_destroy_cm_id(cm_id);
2627                 id_priv->cm_id.iw = NULL;
2628         }
2629         return ret;
2630 }
2631
2632 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2633 {
2634         struct rdma_id_private *id_priv;
2635         int ret;
2636
2637         id_priv = container_of(id, struct rdma_id_private, id);
2638         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2639                 return -EINVAL;
2640
2641         if (!id->qp) {
2642                 id_priv->qp_num = conn_param->qp_num;
2643                 id_priv->srq = conn_param->srq;
2644         }
2645
2646         switch (rdma_node_get_transport(id->device->node_type)) {
2647         case RDMA_TRANSPORT_IB:
2648                 if (cma_is_ud_ps(id->ps))
2649                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2650                 else
2651                         ret = cma_connect_ib(id_priv, conn_param);
2652                 break;
2653         case RDMA_TRANSPORT_IWARP:
2654                 ret = cma_connect_iw(id_priv, conn_param);
2655                 break;
2656         default:
2657                 ret = -ENOSYS;
2658                 break;
2659         }
2660         if (ret)
2661                 goto err;
2662
2663         return 0;
2664 err:
2665         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2666         return ret;
2667 }
2668 EXPORT_SYMBOL(rdma_connect);
2669
2670 static int cma_accept_ib(struct rdma_id_private *id_priv,
2671                          struct rdma_conn_param *conn_param)
2672 {
2673         struct ib_cm_rep_param rep;
2674         int ret;
2675
2676         ret = cma_modify_qp_rtr(id_priv, conn_param);
2677         if (ret)
2678                 goto out;
2679
2680         ret = cma_modify_qp_rts(id_priv, conn_param);
2681         if (ret)
2682                 goto out;
2683
2684         memset(&rep, 0, sizeof rep);
2685         rep.qp_num = id_priv->qp_num;
2686         rep.starting_psn = id_priv->seq_num;
2687         rep.private_data = conn_param->private_data;
2688         rep.private_data_len = conn_param->private_data_len;
2689         rep.responder_resources = conn_param->responder_resources;
2690         rep.initiator_depth = conn_param->initiator_depth;
2691         rep.failover_accepted = 0;
2692         rep.flow_control = conn_param->flow_control;
2693         rep.rnr_retry_count = conn_param->rnr_retry_count;
2694         rep.srq = id_priv->srq ? 1 : 0;
2695
2696         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2697 out:
2698         return ret;
2699 }
2700
2701 static int cma_accept_iw(struct rdma_id_private *id_priv,
2702                   struct rdma_conn_param *conn_param)
2703 {
2704         struct iw_cm_conn_param iw_param;
2705         int ret;
2706
2707         ret = cma_modify_qp_rtr(id_priv, conn_param);
2708         if (ret)
2709                 return ret;
2710
2711         iw_param.ord = conn_param->initiator_depth;
2712         iw_param.ird = conn_param->responder_resources;
2713         iw_param.private_data = conn_param->private_data;
2714         iw_param.private_data_len = conn_param->private_data_len;
2715         if (id_priv->id.qp) {
2716                 iw_param.qpn = id_priv->qp_num;
2717         } else
2718                 iw_param.qpn = conn_param->qp_num;
2719
2720         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2721 }
2722
2723 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2724                              enum ib_cm_sidr_status status,
2725                              const void *private_data, int private_data_len)
2726 {
2727         struct ib_cm_sidr_rep_param rep;
2728         int ret;
2729
2730         memset(&rep, 0, sizeof rep);
2731         rep.status = status;
2732         if (status == IB_SIDR_SUCCESS) {
2733                 ret = cma_set_qkey(id_priv);
2734                 if (ret)
2735                         return ret;
2736                 rep.qp_num = id_priv->qp_num;
2737                 rep.qkey = id_priv->qkey;
2738         }
2739         rep.private_data = private_data;
2740         rep.private_data_len = private_data_len;
2741
2742         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2743 }
2744
2745 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2746 {
2747         struct rdma_id_private *id_priv;
2748         int ret;
2749
2750         id_priv = container_of(id, struct rdma_id_private, id);
2751         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2752                 return -EINVAL;
2753
2754         if (!id->qp && conn_param) {
2755                 id_priv->qp_num = conn_param->qp_num;
2756                 id_priv->srq = conn_param->srq;
2757         }
2758
2759         switch (rdma_node_get_transport(id->device->node_type)) {
2760         case RDMA_TRANSPORT_IB:
2761                 if (cma_is_ud_ps(id->ps))
2762                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2763                                                 conn_param->private_data,
2764                                                 conn_param->private_data_len);
2765                 else if (conn_param)
2766                         ret = cma_accept_ib(id_priv, conn_param);
2767                 else
2768                         ret = cma_rep_recv(id_priv);
2769                 break;
2770         case RDMA_TRANSPORT_IWARP:
2771                 ret = cma_accept_iw(id_priv, conn_param);
2772                 break;
2773         default:
2774                 ret = -ENOSYS;
2775                 break;
2776         }
2777
2778         if (ret)
2779                 goto reject;
2780
2781         return 0;
2782 reject:
2783         cma_modify_qp_err(id_priv);
2784         rdma_reject(id, NULL, 0);
2785         return ret;
2786 }
2787 EXPORT_SYMBOL(rdma_accept);
2788
2789 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2790 {
2791         struct rdma_id_private *id_priv;
2792         int ret;
2793
2794         id_priv = container_of(id, struct rdma_id_private, id);
2795         if (!cma_has_cm_dev(id_priv))
2796                 return -EINVAL;
2797
2798         switch (id->device->node_type) {
2799         case RDMA_NODE_IB_CA:
2800                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2801                 break;
2802         default:
2803                 ret = 0;
2804                 break;
2805         }
2806         return ret;
2807 }
2808 EXPORT_SYMBOL(rdma_notify);
2809
2810 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2811                 u8 private_data_len)
2812 {
2813         struct rdma_id_private *id_priv;
2814         int ret;
2815
2816         id_priv = container_of(id, struct rdma_id_private, id);
2817         if (!cma_has_cm_dev(id_priv))
2818                 return -EINVAL;
2819
2820         switch (rdma_node_get_transport(id->device->node_type)) {
2821         case RDMA_TRANSPORT_IB:
2822                 if (cma_is_ud_ps(id->ps))
2823                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2824                                                 private_data, private_data_len);
2825                 else
2826                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2827                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2828                                              0, private_data, private_data_len);
2829                 break;
2830         case RDMA_TRANSPORT_IWARP:
2831                 ret = iw_cm_reject(id_priv->cm_id.iw,
2832                                    private_data, private_data_len);
2833                 break;
2834         default:
2835                 ret = -ENOSYS;
2836                 break;
2837         }
2838         return ret;
2839 }
2840 EXPORT_SYMBOL(rdma_reject);
2841
2842 int rdma_disconnect(struct rdma_cm_id *id)
2843 {
2844         struct rdma_id_private *id_priv;
2845         int ret;
2846
2847         id_priv = container_of(id, struct rdma_id_private, id);
2848         if (!cma_has_cm_dev(id_priv))
2849                 return -EINVAL;
2850
2851         switch (rdma_node_get_transport(id->device->node_type)) {
2852         case RDMA_TRANSPORT_IB:
2853                 ret = cma_modify_qp_err(id_priv);
2854                 if (ret)
2855                         goto out;
2856                 /* Initiate or respond to a disconnect. */
2857                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2858                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2859                 break;
2860         case RDMA_TRANSPORT_IWARP:
2861                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2862                 break;
2863         default:
2864                 ret = -EINVAL;
2865                 break;
2866         }
2867 out:
2868         return ret;
2869 }
2870 EXPORT_SYMBOL(rdma_disconnect);
2871
2872 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2873 {
2874         struct rdma_id_private *id_priv;
2875         struct cma_multicast *mc = multicast->context;
2876         struct rdma_cm_event event;
2877         int ret;
2878
2879         id_priv = mc->id_priv;
2880         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2881             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2882                 return 0;
2883
2884         mutex_lock(&id_priv->qp_mutex);
2885         if (!status && id_priv->id.qp)
2886                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2887                                          multicast->rec.mlid);
2888         mutex_unlock(&id_priv->qp_mutex);
2889
2890         memset(&event, 0, sizeof event);
2891         event.status = status;
2892         event.param.ud.private_data = mc->context;
2893         if (!status) {
2894                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2895                 ib_init_ah_from_mcmember(id_priv->id.device,
2896                                          id_priv->id.port_num, &multicast->rec,
2897                                          &event.param.ud.ah_attr);
2898                 event.param.ud.qp_num = 0xFFFFFF;
2899                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2900         } else
2901                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2902
2903         ret = id_priv->id.event_handler(&id_priv->id, &event);
2904         if (ret) {
2905                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2906                 mutex_unlock(&id_priv->handler_mutex);
2907                 rdma_destroy_id(&id_priv->id);
2908                 return 0;
2909         }
2910
2911         mutex_unlock(&id_priv->handler_mutex);
2912         return 0;
2913 }
2914
2915 static void cma_set_mgid(struct rdma_id_private *id_priv,
2916                          struct sockaddr *addr, union ib_gid *mgid)
2917 {
2918         unsigned char mc_map[MAX_ADDR_LEN];
2919         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2920         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2921         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2922
2923         if (cma_any_addr(addr)) {
2924                 memset(mgid, 0, sizeof *mgid);
2925         } else if ((addr->sa_family == AF_INET6) &&
2926                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2927                                                                  0xFF10A01B)) {
2928                 /* IPv6 address is an SA assigned MGID. */
2929                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2930         } else if ((addr->sa_family == AF_INET6)) {
2931                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2932                 if (id_priv->id.ps == RDMA_PS_UDP)
2933                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2934                 *mgid = *(union ib_gid *) (mc_map + 4);
2935         } else {
2936                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2937                 if (id_priv->id.ps == RDMA_PS_UDP)
2938                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2939                 *mgid = *(union ib_gid *) (mc_map + 4);
2940         }
2941 }
2942
2943 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2944                                  struct cma_multicast *mc)
2945 {
2946         struct ib_sa_mcmember_rec rec;
2947         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2948         ib_sa_comp_mask comp_mask;
2949         int ret;
2950
2951         ib_addr_get_mgid(dev_addr, &rec.mgid);
2952         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2953                                      &rec.mgid, &rec);
2954         if (ret)
2955                 return ret;
2956
2957         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2958         if (id_priv->id.ps == RDMA_PS_UDP)
2959                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2960         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2961         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2962         rec.join_state = 1;
2963
2964         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2965                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2966                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2967                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2968                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2969
2970         if (id_priv->id.ps == RDMA_PS_IPOIB)
2971                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2972                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2973
2974         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2975                                                 id_priv->id.port_num, &rec,
2976                                                 comp_mask, GFP_KERNEL,
2977                                                 cma_ib_mc_handler, mc);
2978         if (IS_ERR(mc->multicast.ib))
2979                 return PTR_ERR(mc->multicast.ib);
2980
2981         return 0;
2982 }
2983
2984 static void iboe_mcast_work_handler(struct work_struct *work)
2985 {
2986         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
2987         struct cma_multicast *mc = mw->mc;
2988         struct ib_sa_multicast *m = mc->multicast.ib;
2989
2990         mc->multicast.ib->context = mc;
2991         cma_ib_mc_handler(0, m);
2992         kref_put(&mc->mcref, release_mc);
2993         kfree(mw);
2994 }
2995
2996 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
2997 {
2998         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
2999         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3000
3001         if (cma_any_addr(addr)) {
3002                 memset(mgid, 0, sizeof *mgid);
3003         } else if (addr->sa_family == AF_INET6) {
3004                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3005         } else {
3006                 mgid->raw[0] = 0xff;
3007                 mgid->raw[1] = 0x0e;
3008                 mgid->raw[2] = 0;
3009                 mgid->raw[3] = 0;
3010                 mgid->raw[4] = 0;
3011                 mgid->raw[5] = 0;
3012                 mgid->raw[6] = 0;
3013                 mgid->raw[7] = 0;
3014                 mgid->raw[8] = 0;
3015                 mgid->raw[9] = 0;
3016                 mgid->raw[10] = 0xff;
3017                 mgid->raw[11] = 0xff;
3018                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3019         }
3020 }
3021
3022 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3023                                    struct cma_multicast *mc)
3024 {
3025         struct iboe_mcast_work *work;
3026         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3027         int err;
3028         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3029         struct net_device *ndev = NULL;
3030
3031         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3032                 return -EINVAL;
3033
3034         work = kzalloc(sizeof *work, GFP_KERNEL);
3035         if (!work)
3036                 return -ENOMEM;
3037
3038         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3039         if (!mc->multicast.ib) {
3040                 err = -ENOMEM;
3041                 goto out1;
3042         }
3043
3044         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3045
3046         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3047         if (id_priv->id.ps == RDMA_PS_UDP)
3048                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3049
3050         if (dev_addr->bound_dev_if)
3051                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3052         if (!ndev) {
3053                 err = -ENODEV;
3054                 goto out2;
3055         }
3056         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3057         mc->multicast.ib->rec.hop_limit = 1;
3058         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3059         dev_put(ndev);
3060         if (!mc->multicast.ib->rec.mtu) {
3061                 err = -EINVAL;
3062                 goto out2;
3063         }
3064         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3065         work->id = id_priv;
3066         work->mc = mc;
3067         INIT_WORK(&work->work, iboe_mcast_work_handler);
3068         kref_get(&mc->mcref);
3069         queue_work(cma_wq, &work->work);
3070
3071         return 0;
3072
3073 out2:
3074         kfree(mc->multicast.ib);
3075 out1:
3076         kfree(work);
3077         return err;
3078 }
3079
3080 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3081                         void *context)
3082 {
3083         struct rdma_id_private *id_priv;
3084         struct cma_multicast *mc;
3085         int ret;
3086
3087         id_priv = container_of(id, struct rdma_id_private, id);
3088         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3089             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3090                 return -EINVAL;
3091
3092         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3093         if (!mc)
3094                 return -ENOMEM;
3095
3096         memcpy(&mc->addr, addr, ip_addr_size(addr));
3097         mc->context = context;
3098         mc->id_priv = id_priv;
3099
3100         spin_lock(&id_priv->lock);
3101         list_add(&mc->list, &id_priv->mc_list);
3102         spin_unlock(&id_priv->lock);
3103
3104         switch (rdma_node_get_transport(id->device->node_type)) {
3105         case RDMA_TRANSPORT_IB:
3106                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3107                 case IB_LINK_LAYER_INFINIBAND:
3108                         ret = cma_join_ib_multicast(id_priv, mc);
3109                         break;
3110                 case IB_LINK_LAYER_ETHERNET:
3111                         kref_init(&mc->mcref);
3112                         ret = cma_iboe_join_multicast(id_priv, mc);
3113                         break;
3114                 default:
3115                         ret = -EINVAL;
3116                 }
3117                 break;
3118         default:
3119                 ret = -ENOSYS;
3120                 break;
3121         }
3122
3123         if (ret) {
3124                 spin_lock_irq(&id_priv->lock);
3125                 list_del(&mc->list);
3126                 spin_unlock_irq(&id_priv->lock);
3127                 kfree(mc);
3128         }
3129         return ret;
3130 }
3131 EXPORT_SYMBOL(rdma_join_multicast);
3132
3133 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3134 {
3135         struct rdma_id_private *id_priv;
3136         struct cma_multicast *mc;
3137
3138         id_priv = container_of(id, struct rdma_id_private, id);
3139         spin_lock_irq(&id_priv->lock);
3140         list_for_each_entry(mc, &id_priv->mc_list, list) {
3141                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3142                         list_del(&mc->list);
3143                         spin_unlock_irq(&id_priv->lock);
3144
3145                         if (id->qp)
3146                                 ib_detach_mcast(id->qp,
3147                                                 &mc->multicast.ib->rec.mgid,
3148                                                 mc->multicast.ib->rec.mlid);
3149                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3150                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3151                                 case IB_LINK_LAYER_INFINIBAND:
3152                                         ib_sa_free_multicast(mc->multicast.ib);
3153                                         kfree(mc);
3154                                         break;
3155                                 case IB_LINK_LAYER_ETHERNET:
3156                                         kref_put(&mc->mcref, release_mc);
3157                                         break;
3158                                 default:
3159                                         break;
3160                                 }
3161                         }
3162                         return;
3163                 }
3164         }
3165         spin_unlock_irq(&id_priv->lock);
3166 }
3167 EXPORT_SYMBOL(rdma_leave_multicast);
3168
3169 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3170 {
3171         struct rdma_dev_addr *dev_addr;
3172         struct cma_ndev_work *work;
3173
3174         dev_addr = &id_priv->id.route.addr.dev_addr;
3175
3176         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3177             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3178                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3179                        ndev->name, &id_priv->id);
3180                 work = kzalloc(sizeof *work, GFP_KERNEL);
3181                 if (!work)
3182                         return -ENOMEM;
3183
3184                 INIT_WORK(&work->work, cma_ndev_work_handler);
3185                 work->id = id_priv;
3186                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3187                 atomic_inc(&id_priv->refcount);
3188                 queue_work(cma_wq, &work->work);
3189         }
3190
3191         return 0;
3192 }
3193
3194 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3195                                void *ctx)
3196 {
3197         struct net_device *ndev = (struct net_device *)ctx;
3198         struct cma_device *cma_dev;
3199         struct rdma_id_private *id_priv;
3200         int ret = NOTIFY_DONE;
3201
3202         if (dev_net(ndev) != &init_net)
3203                 return NOTIFY_DONE;
3204
3205         if (event != NETDEV_BONDING_FAILOVER)
3206                 return NOTIFY_DONE;
3207
3208         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3209                 return NOTIFY_DONE;
3210
3211         mutex_lock(&lock);
3212         list_for_each_entry(cma_dev, &dev_list, list)
3213                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3214                         ret = cma_netdev_change(ndev, id_priv);
3215                         if (ret)
3216                                 goto out;
3217                 }
3218
3219 out:
3220         mutex_unlock(&lock);
3221         return ret;
3222 }
3223
3224 static struct notifier_block cma_nb = {
3225         .notifier_call = cma_netdev_callback
3226 };
3227
3228 static void cma_add_one(struct ib_device *device)
3229 {
3230         struct cma_device *cma_dev;
3231         struct rdma_id_private *id_priv;
3232
3233         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3234         if (!cma_dev)
3235                 return;
3236
3237         cma_dev->device = device;
3238
3239         init_completion(&cma_dev->comp);
3240         atomic_set(&cma_dev->refcount, 1);
3241         INIT_LIST_HEAD(&cma_dev->id_list);
3242         ib_set_client_data(device, &cma_client, cma_dev);
3243
3244         mutex_lock(&lock);
3245         list_add_tail(&cma_dev->list, &dev_list);
3246         list_for_each_entry(id_priv, &listen_any_list, list)
3247                 cma_listen_on_dev(id_priv, cma_dev);
3248         mutex_unlock(&lock);
3249 }
3250
3251 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3252 {
3253         struct rdma_cm_event event;
3254         enum rdma_cm_state state;
3255         int ret = 0;
3256
3257         /* Record that we want to remove the device */
3258         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3259         if (state == RDMA_CM_DESTROYING)
3260                 return 0;
3261
3262         cma_cancel_operation(id_priv, state);
3263         mutex_lock(&id_priv->handler_mutex);
3264
3265         /* Check for destruction from another callback. */
3266         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3267                 goto out;
3268
3269         memset(&event, 0, sizeof event);
3270         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3271         ret = id_priv->id.event_handler(&id_priv->id, &event);
3272 out:
3273         mutex_unlock(&id_priv->handler_mutex);
3274         return ret;
3275 }
3276
3277 static void cma_process_remove(struct cma_device *cma_dev)
3278 {
3279         struct rdma_id_private *id_priv;
3280         int ret;
3281
3282         mutex_lock(&lock);
3283         while (!list_empty(&cma_dev->id_list)) {
3284                 id_priv = list_entry(cma_dev->id_list.next,
3285                                      struct rdma_id_private, list);
3286
3287                 list_del(&id_priv->listen_list);
3288                 list_del_init(&id_priv->list);
3289                 atomic_inc(&id_priv->refcount);
3290                 mutex_unlock(&lock);
3291
3292                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3293                 cma_deref_id(id_priv);
3294                 if (ret)
3295                         rdma_destroy_id(&id_priv->id);
3296
3297                 mutex_lock(&lock);
3298         }
3299         mutex_unlock(&lock);
3300
3301         cma_deref_dev(cma_dev);
3302         wait_for_completion(&cma_dev->comp);
3303 }
3304
3305 static void cma_remove_one(struct ib_device *device)
3306 {
3307         struct cma_device *cma_dev;
3308
3309         cma_dev = ib_get_client_data(device, &cma_client);
3310         if (!cma_dev)
3311                 return;
3312
3313         mutex_lock(&lock);
3314         list_del(&cma_dev->list);
3315         mutex_unlock(&lock);
3316
3317         cma_process_remove(cma_dev);
3318         kfree(cma_dev);
3319 }
3320
3321 static int __init cma_init(void)
3322 {
3323         int ret;
3324
3325         cma_wq = create_singlethread_workqueue("rdma_cm");
3326         if (!cma_wq)
3327                 return -ENOMEM;
3328
3329         ib_sa_register_client(&sa_client);
3330         rdma_addr_register_client(&addr_client);
3331         register_netdevice_notifier(&cma_nb);
3332
3333         ret = ib_register_client(&cma_client);
3334         if (ret)
3335                 goto err;
3336         return 0;
3337
3338 err:
3339         unregister_netdevice_notifier(&cma_nb);
3340         rdma_addr_unregister_client(&addr_client);
3341         ib_sa_unregister_client(&sa_client);
3342         destroy_workqueue(cma_wq);
3343         return ret;
3344 }
3345
3346 static void __exit cma_cleanup(void)
3347 {
3348         ib_unregister_client(&cma_client);
3349         unregister_netdevice_notifier(&cma_nb);
3350         rdma_addr_unregister_client(&addr_client);
3351         ib_sa_unregister_client(&sa_client);
3352         destroy_workqueue(cma_wq);
3353         idr_destroy(&sdp_ps);
3354         idr_destroy(&tcp_ps);
3355         idr_destroy(&udp_ps);
3356         idr_destroy(&ipoib_ps);
3357 }
3358
3359 module_init(cma_init);
3360 module_exit(cma_cleanup);