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
  44 #include <net/tcp.h>
  45 #include <net/ipv6.h>
  46
  47 #include <rdma/rdma_cm.h>
  48 #include <rdma/rdma_cm_ib.h>
  49 #include <rdma/ib_cache.h>
  50 #include <rdma/ib_cm.h>
  51 #include <rdma/ib_sa.h>
  52 #include <rdma/iw_cm.h>
  53
  54 MODULE_AUTHOR("Sean Hefty");
  55 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 #define CMA_CM_RESPONSE_TIMEOUT 20
  59 #define CMA_MAX_CM_RETRIES 15
  60 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  61
  62 static void cma_add_one(struct ib_device *device);
  63 static void cma_remove_one(struct ib_device *device);
  64
  65 static struct ib_client cma_client = {
  66         .name   = "cma",
  67         .add    = cma_add_one,
  68         .remove = cma_remove_one
  69 };
  70
  71 static struct ib_sa_client sa_client;
  72 static struct rdma_addr_client addr_client;
  73 static LIST_HEAD(dev_list);
  74 static LIST_HEAD(listen_any_list);
  75 static DEFINE_MUTEX(lock);
  76 static struct workqueue_struct *cma_wq;
  77 static DEFINE_IDR(sdp_ps);
  78 static DEFINE_IDR(tcp_ps);
  79 static DEFINE_IDR(udp_ps);
  80 static DEFINE_IDR(ipoib_ps);
  81 static int next_port;
  82
  83 struct cma_device {
  84         struct list_head        list;
  85         struct ib_device        *device;
  86         struct completion       comp;
  87         atomic_t                refcount;
  88         struct list_head        id_list;
  89 };
  90
  91 enum cma_state {
  92         CMA_IDLE,
  93         CMA_ADDR_QUERY,
  94         CMA_ADDR_RESOLVED,
  95         CMA_ROUTE_QUERY,
  96         CMA_ROUTE_RESOLVED,
  97         CMA_CONNECT,
  98         CMA_DISCONNECT,
  99         CMA_ADDR_BOUND,
 100         CMA_LISTEN,
 101         CMA_DEVICE_REMOVAL,
 102         CMA_DESTROYING
 103 };
 104
 105 struct rdma_bind_list {
 106         struct idr              *ps;
 107         struct hlist_head       owners;
 108         unsigned short          port;
 109 };
 110
 111 /*
 112  * Device removal can occur at anytime, so we need extra handling to
 113  * serialize notifying the user of device removal with other callbacks.
 114  * We do this by disabling removal notification while a callback is in process,
 115  * and reporting it after the callback completes.
 116  */
 117 struct rdma_id_private {
 118         struct rdma_cm_id       id;
 119
 120         struct rdma_bind_list   *bind_list;
 121         struct hlist_node       node;
 122         struct list_head        list; /* listen_any_list or cma_device.list */
 123         struct list_head        listen_list; /* per device listens */
 124         struct cma_device       *cma_dev;
 125         struct list_head        mc_list;
 126
 127         int                     internal_id;
 128         enum cma_state          state;
 129         spinlock_t              lock;
 130         struct mutex            qp_mutex;
 131
 132         struct completion       comp;
 133         atomic_t                refcount;
 134         struct mutex            handler_mutex;
 135
 136         int                     backlog;
 137         int                     timeout_ms;
 138         struct ib_sa_query      *query;
 139         int                     query_id;
 140         union {
 141                 struct ib_cm_id *ib;
 142                 struct iw_cm_id *iw;
 143         } cm_id;
 144
 145         u32                     seq_num;
 146         u32                     qkey;
 147         u32                     qp_num;
 148         u8                      srq;
 149         u8                      tos;
 150 };
 151
 152 struct cma_multicast {
 153         struct rdma_id_private *id_priv;
 154         union {
 155                 struct ib_sa_multicast *ib;
 156         } multicast;
 157         struct list_head        list;
 158         void                    *context;
 159         struct sockaddr_storage addr;
 160 };
 161
 162 struct cma_work {
 163         struct work_struct      work;
 164         struct rdma_id_private  *id;
 165         enum cma_state          old_state;
 166         enum cma_state          new_state;
 167         struct rdma_cm_event    event;
 168 };
 169
 170 struct cma_ndev_work {
 171         struct work_struct      work;
 172         struct rdma_id_private  *id;
 173         struct rdma_cm_event    event;
 174 };
 175
 176 union cma_ip_addr {
 177         struct in6_addr ip6;
 178         struct {
 179                 __be32 pad[3];
 180                 __be32 addr;
 181         } ip4;
 182 };
 183
 184 struct cma_hdr {
 185         u8 cma_version;
 186         u8 ip_version;  /* IP version: 7:4 */
 187         __be16 port;
 188         union cma_ip_addr src_addr;
 189         union cma_ip_addr dst_addr;
 190 };
 191
 192 struct sdp_hh {
 193         u8 bsdh[16];
 194         u8 sdp_version; /* Major version: 7:4 */
 195         u8 ip_version;  /* IP version: 7:4 */
 196         u8 sdp_specific1[10];
 197         __be16 port;
 198         __be16 sdp_specific2;
 199         union cma_ip_addr src_addr;
 200         union cma_ip_addr dst_addr;
 201 };
 202
 203 struct sdp_hah {
 204         u8 bsdh[16];
 205         u8 sdp_version;
 206 };
 207
 208 #define CMA_VERSION 0x00
 209 #define SDP_MAJ_VERSION 0x2
 210
 211 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 212 {
 213         unsigned long flags;
 214         int ret;
 215
 216         spin_lock_irqsave(&id_priv->lock, flags);
 217         ret = (id_priv->state == comp);
 218         spin_unlock_irqrestore(&id_priv->lock, flags);
 219         return ret;
 220 }
 221
 222 static int cma_comp_exch(struct rdma_id_private *id_priv,
 223                          enum cma_state comp, enum cma_state exch)
 224 {
 225         unsigned long flags;
 226         int ret;
 227
 228         spin_lock_irqsave(&id_priv->lock, flags);
 229         if ((ret = (id_priv->state == comp)))
 230                 id_priv->state = exch;
 231         spin_unlock_irqrestore(&id_priv->lock, flags);
 232         return ret;
 233 }
 234
 235 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 236                                enum cma_state exch)
 237 {
 238         unsigned long flags;
 239         enum cma_state old;
 240
 241         spin_lock_irqsave(&id_priv->lock, flags);
 242         old = id_priv->state;
 243         id_priv->state = exch;
 244         spin_unlock_irqrestore(&id_priv->lock, flags);
 245         return old;
 246 }
 247
 248 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 249 {
 250         return hdr->ip_version >> 4;
 251 }
 252
 253 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 254 {
 255         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 256 }
 257
 258 static inline u8 sdp_get_majv(u8 sdp_version)
 259 {
 260         return sdp_version >> 4;
 261 }
 262
 263 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 264 {
 265         return hh->ip_version >> 4;
 266 }
 267
 268 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 269 {
 270         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 271 }
 272
 273 static inline int cma_is_ud_ps(enum rdma_port_space ps)
 274 {
 275         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
 276 }
 277
 278 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 279                               struct cma_device *cma_dev)
 280 {
 281         atomic_inc(&cma_dev->refcount);
 282         id_priv->cma_dev = cma_dev;
 283         id_priv->id.device = cma_dev->device;
 284         list_add_tail(&id_priv->list, &cma_dev->id_list);
 285 }
 286
 287 static inline void cma_deref_dev(struct cma_device *cma_dev)
 288 {
 289         if (atomic_dec_and_test(&cma_dev->refcount))
 290                 complete(&cma_dev->comp);
 291 }
 292
 293 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 294 {
 295         list_del(&id_priv->list);
 296         cma_deref_dev(id_priv->cma_dev);
 297         id_priv->cma_dev = NULL;
 298 }
 299
 300 static int cma_set_qkey(struct rdma_id_private *id_priv)
 301 {
 302         struct ib_sa_mcmember_rec rec;
 303         int ret = 0;
 304
 305         if (id_priv->qkey)
 306                 return 0;
 307
 308         switch (id_priv->id.ps) {
 309         case RDMA_PS_UDP:
 310                 id_priv->qkey = RDMA_UDP_QKEY;
 311                 break;
 312         case RDMA_PS_IPOIB:
 313                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 314                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 315                                              id_priv->id.port_num, &rec.mgid,
 316                                              &rec);
 317                 if (!ret)
 318                         id_priv->qkey = be32_to_cpu(rec.qkey);
 319                 break;
 320         default:
 321                 break;
 322         }
 323         return ret;
 324 }
 325
 326 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 327 {
 328         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 329         struct cma_device *cma_dev;
 330         union ib_gid gid;
 331         int ret = -ENODEV;
 332
 333         switch (rdma_node_get_transport(dev_addr->dev_type)) {
 334         case RDMA_TRANSPORT_IB:
 335                 ib_addr_get_sgid(dev_addr, &gid);
 336                 break;
 337         case RDMA_TRANSPORT_IWARP:
 338                 iw_addr_get_sgid(dev_addr, &gid);
 339                 break;
 340         default:
 341                 return -ENODEV;
 342         }
 343
 344         list_for_each_entry(cma_dev, &dev_list, list) {
 345                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 346                                          &id_priv->id.port_num, NULL);
 347                 if (!ret) {
 348                         cma_attach_to_dev(id_priv, cma_dev);
 349                         break;
 350                 }
 351         }
 352         return ret;
 353 }
 354
 355 static void cma_deref_id(struct rdma_id_private *id_priv)
 356 {
 357         if (atomic_dec_and_test(&id_priv->refcount))
 358                 complete(&id_priv->comp);
 359 }
 360
 361 static int cma_disable_callback(struct rdma_id_private *id_priv,
 362                               enum cma_state state)
 363 {
 364         mutex_lock(&id_priv->handler_mutex);
 365         if (id_priv->state != state) {
 366                 mutex_unlock(&id_priv->handler_mutex);
 367                 return -EINVAL;
 368         }
 369         return 0;
 370 }
 371
 372 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
 373 {
 374         return (id_priv->id.device && id_priv->cm_id.ib);
 375 }
 376
 377 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 378                                   void *context, enum rdma_port_space ps)
 379 {
 380         struct rdma_id_private *id_priv;
 381
 382         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 383         if (!id_priv)
 384                 return ERR_PTR(-ENOMEM);
 385
 386         id_priv->state = CMA_IDLE;
 387         id_priv->id.context = context;
 388         id_priv->id.event_handler = event_handler;
 389         id_priv->id.ps = ps;
 390         spin_lock_init(&id_priv->lock);
 391         mutex_init(&id_priv->qp_mutex);
 392         init_completion(&id_priv->comp);
 393         atomic_set(&id_priv->refcount, 1);
 394         mutex_init(&id_priv->handler_mutex);
 395         INIT_LIST_HEAD(&id_priv->listen_list);
 396         INIT_LIST_HEAD(&id_priv->mc_list);
 397         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 398
 399         return &id_priv->id;
 400 }
 401 EXPORT_SYMBOL(rdma_create_id);
 402
 403 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 404 {
 405         struct ib_qp_attr qp_attr;
 406         int qp_attr_mask, ret;
 407
 408         qp_attr.qp_state = IB_QPS_INIT;
 409         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 410         if (ret)
 411                 return ret;
 412
 413         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 414         if (ret)
 415                 return ret;
 416
 417         qp_attr.qp_state = IB_QPS_RTR;
 418         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 419         if (ret)
 420                 return ret;
 421
 422         qp_attr.qp_state = IB_QPS_RTS;
 423         qp_attr.sq_psn = 0;
 424         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 425
 426         return ret;
 427 }
 428
 429 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 430 {
 431         struct ib_qp_attr qp_attr;
 432         int qp_attr_mask, ret;
 433
 434         qp_attr.qp_state = IB_QPS_INIT;
 435         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 436         if (ret)
 437                 return ret;
 438
 439         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 440 }
 441
 442 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 443                    struct ib_qp_init_attr *qp_init_attr)
 444 {
 445         struct rdma_id_private *id_priv;
 446         struct ib_qp *qp;
 447         int ret;
 448
 449         id_priv = container_of(id, struct rdma_id_private, id);
 450         if (id->device != pd->device)
 451                 return -EINVAL;
 452
 453         qp = ib_create_qp(pd, qp_init_attr);
 454         if (IS_ERR(qp))
 455                 return PTR_ERR(qp);
 456
 457         if (cma_is_ud_ps(id_priv->id.ps))
 458                 ret = cma_init_ud_qp(id_priv, qp);
 459         else
 460                 ret = cma_init_conn_qp(id_priv, qp);
 461         if (ret)
 462                 goto err;
 463
 464         id->qp = qp;
 465         id_priv->qp_num = qp->qp_num;
 466         id_priv->srq = (qp->srq != NULL);
 467         return 0;
 468 err:
 469         ib_destroy_qp(qp);
 470         return ret;
 471 }
 472 EXPORT_SYMBOL(rdma_create_qp);
 473
 474 void rdma_destroy_qp(struct rdma_cm_id *id)
 475 {
 476         struct rdma_id_private *id_priv;
 477
 478         id_priv = container_of(id, struct rdma_id_private, id);
 479         mutex_lock(&id_priv->qp_mutex);
 480         ib_destroy_qp(id_priv->id.qp);
 481         id_priv->id.qp = NULL;
 482         mutex_unlock(&id_priv->qp_mutex);
 483 }
 484 EXPORT_SYMBOL(rdma_destroy_qp);
 485
 486 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 487                              struct rdma_conn_param *conn_param)
 488 {
 489         struct ib_qp_attr qp_attr;
 490         int qp_attr_mask, ret;
 491
 492         mutex_lock(&id_priv->qp_mutex);
 493         if (!id_priv->id.qp) {
 494                 ret = 0;
 495                 goto out;
 496         }
 497
 498         /* Need to update QP attributes from default values. */
 499         qp_attr.qp_state = IB_QPS_INIT;
 500         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 501         if (ret)
 502                 goto out;
 503
 504         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 505         if (ret)
 506                 goto out;
 507
 508         qp_attr.qp_state = IB_QPS_RTR;
 509         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 510         if (ret)
 511                 goto out;
 512
 513         if (conn_param)
 514                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 515         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 516 out:
 517         mutex_unlock(&id_priv->qp_mutex);
 518         return ret;
 519 }
 520
 521 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 522                              struct rdma_conn_param *conn_param)
 523 {
 524         struct ib_qp_attr qp_attr;
 525         int qp_attr_mask, ret;
 526
 527         mutex_lock(&id_priv->qp_mutex);
 528         if (!id_priv->id.qp) {
 529                 ret = 0;
 530                 goto out;
 531         }
 532
 533         qp_attr.qp_state = IB_QPS_RTS;
 534         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 535         if (ret)
 536                 goto out;
 537
 538         if (conn_param)
 539                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 540         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 541 out:
 542         mutex_unlock(&id_priv->qp_mutex);
 543         return ret;
 544 }
 545
 546 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 547 {
 548         struct ib_qp_attr qp_attr;
 549         int ret;
 550
 551         mutex_lock(&id_priv->qp_mutex);
 552         if (!id_priv->id.qp) {
 553                 ret = 0;
 554                 goto out;
 555         }
 556
 557         qp_attr.qp_state = IB_QPS_ERR;
 558         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 559 out:
 560         mutex_unlock(&id_priv->qp_mutex);
 561         return ret;
 562 }
 563
 564 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 565                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 566 {
 567         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 568         int ret;
 569
 570         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 571                                   ib_addr_get_pkey(dev_addr),
 572                                   &qp_attr->pkey_index);
 573         if (ret)
 574                 return ret;
 575
 576         qp_attr->port_num = id_priv->id.port_num;
 577         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 578
 579         if (cma_is_ud_ps(id_priv->id.ps)) {
 580                 ret = cma_set_qkey(id_priv);
 581                 if (ret)
 582                         return ret;
 583
 584                 qp_attr->qkey = id_priv->qkey;
 585                 *qp_attr_mask |= IB_QP_QKEY;
 586         } else {
 587                 qp_attr->qp_access_flags = 0;
 588                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 589         }
 590         return 0;
 591 }
 592
 593 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 594                        int *qp_attr_mask)
 595 {
 596         struct rdma_id_private *id_priv;
 597         int ret = 0;
 598
 599         id_priv = container_of(id, struct rdma_id_private, id);
 600         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 601         case RDMA_TRANSPORT_IB:
 602                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
 603                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 604                 else
 605                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 606                                                  qp_attr_mask);
 607                 if (qp_attr->qp_state == IB_QPS_RTR)
 608                         qp_attr->rq_psn = id_priv->seq_num;
 609                 break;
 610         case RDMA_TRANSPORT_IWARP:
 611                 if (!id_priv->cm_id.iw) {
 612                         qp_attr->qp_access_flags = 0;
 613                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 614                 } else
 615                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 616                                                  qp_attr_mask);
 617                 break;
 618         default:
 619                 ret = -ENOSYS;
 620                 break;
 621         }
 622
 623         return ret;
 624 }
 625 EXPORT_SYMBOL(rdma_init_qp_attr);
 626
 627 static inline int cma_zero_addr(struct sockaddr *addr)
 628 {
 629         struct in6_addr *ip6;
 630
 631         if (addr->sa_family == AF_INET)
 632                 return ipv4_is_zeronet(
 633                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 634         else {
 635                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 636                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 637                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 638         }
 639 }
 640
 641 static inline int cma_loopback_addr(struct sockaddr *addr)
 642 {
 643         if (addr->sa_family == AF_INET)
 644                 return ipv4_is_loopback(
 645                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 646         else
 647                 return ipv6_addr_loopback(
 648                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 649 }
 650
 651 static inline int cma_any_addr(struct sockaddr *addr)
 652 {
 653         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 654 }
 655
 656 static inline __be16 cma_port(struct sockaddr *addr)
 657 {
 658         if (addr->sa_family == AF_INET)
 659                 return ((struct sockaddr_in *) addr)->sin_port;
 660         else
 661                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 662 }
 663
 664 static inline int cma_any_port(struct sockaddr *addr)
 665 {
 666         return !cma_port(addr);
 667 }
 668
 669 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 670                             u8 *ip_ver, __be16 *port,
 671                             union cma_ip_addr **src, union cma_ip_addr **dst)
 672 {
 673         switch (ps) {
 674         case RDMA_PS_SDP:
 675                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 676                     SDP_MAJ_VERSION)
 677                         return -EINVAL;
 678
 679                 *ip_ver = sdp_get_ip_ver(hdr);
 680                 *port   = ((struct sdp_hh *) hdr)->port;
 681                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 682                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 683                 break;
 684         default:
 685                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 686                         return -EINVAL;
 687
 688                 *ip_ver = cma_get_ip_ver(hdr);
 689                 *port   = ((struct cma_hdr *) hdr)->port;
 690                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 691                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 692                 break;
 693         }
 694
 695         if (*ip_ver != 4 && *ip_ver != 6)
 696                 return -EINVAL;
 697         return 0;
 698 }
 699
 700 static void cma_save_net_info(struct rdma_addr *addr,
 701                               struct rdma_addr *listen_addr,
 702                               u8 ip_ver, __be16 port,
 703                               union cma_ip_addr *src, union cma_ip_addr *dst)
 704 {
 705         struct sockaddr_in *listen4, *ip4;
 706         struct sockaddr_in6 *listen6, *ip6;
 707
 708         switch (ip_ver) {
 709         case 4:
 710                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 711                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 712                 ip4->sin_family = listen4->sin_family;
 713                 ip4->sin_addr.s_addr = dst->ip4.addr;
 714                 ip4->sin_port = listen4->sin_port;
 715
 716                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 717                 ip4->sin_family = listen4->sin_family;
 718                 ip4->sin_addr.s_addr = src->ip4.addr;
 719                 ip4->sin_port = port;
 720                 break;
 721         case 6:
 722                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 723                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 724                 ip6->sin6_family = listen6->sin6_family;
 725                 ip6->sin6_addr = dst->ip6;
 726                 ip6->sin6_port = listen6->sin6_port;
 727
 728                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 729                 ip6->sin6_family = listen6->sin6_family;
 730                 ip6->sin6_addr = src->ip6;
 731                 ip6->sin6_port = port;
 732                 break;
 733         default:
 734                 break;
 735         }
 736 }
 737
 738 static inline int cma_user_data_offset(enum rdma_port_space ps)
 739 {
 740         switch (ps) {
 741         case RDMA_PS_SDP:
 742                 return 0;
 743         default:
 744                 return sizeof(struct cma_hdr);
 745         }
 746 }
 747
 748 static void cma_cancel_route(struct rdma_id_private *id_priv)
 749 {
 750         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 751         case RDMA_TRANSPORT_IB:
 752                 if (id_priv->query)
 753                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 754                 break;
 755         default:
 756                 break;
 757         }
 758 }
 759
 760 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 761 {
 762         struct rdma_id_private *dev_id_priv;
 763
 764         /*
 765          * Remove from listen_any_list to prevent added devices from spawning
 766          * additional listen requests.
 767          */
 768         mutex_lock(&lock);
 769         list_del(&id_priv->list);
 770
 771         while (!list_empty(&id_priv->listen_list)) {
 772                 dev_id_priv = list_entry(id_priv->listen_list.next,
 773                                          struct rdma_id_private, listen_list);
 774                 /* sync with device removal to avoid duplicate destruction */
 775                 list_del_init(&dev_id_priv->list);
 776                 list_del(&dev_id_priv->listen_list);
 777                 mutex_unlock(&lock);
 778
 779                 rdma_destroy_id(&dev_id_priv->id);
 780                 mutex_lock(&lock);
 781         }
 782         mutex_unlock(&lock);
 783 }
 784
 785 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 786                                  enum cma_state state)
 787 {
 788         switch (state) {
 789         case CMA_ADDR_QUERY:
 790                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 791                 break;
 792         case CMA_ROUTE_QUERY:
 793                 cma_cancel_route(id_priv);
 794                 break;
 795         case CMA_LISTEN:
 796                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 797                                 && !id_priv->cma_dev)
 798                         cma_cancel_listens(id_priv);
 799                 break;
 800         default:
 801                 break;
 802         }
 803 }
 804
 805 static void cma_release_port(struct rdma_id_private *id_priv)
 806 {
 807         struct rdma_bind_list *bind_list = id_priv->bind_list;
 808
 809         if (!bind_list)
 810                 return;
 811
 812         mutex_lock(&lock);
 813         hlist_del(&id_priv->node);
 814         if (hlist_empty(&bind_list->owners)) {
 815                 idr_remove(bind_list->ps, bind_list->port);
 816                 kfree(bind_list);
 817         }
 818         mutex_unlock(&lock);
 819 }
 820
 821 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 822 {
 823         struct cma_multicast *mc;
 824
 825         while (!list_empty(&id_priv->mc_list)) {
 826                 mc = container_of(id_priv->mc_list.next,
 827                                   struct cma_multicast, list);
 828                 list_del(&mc->list);
 829                 ib_sa_free_multicast(mc->multicast.ib);
 830                 kfree(mc);
 831         }
 832 }
 833
 834 void rdma_destroy_id(struct rdma_cm_id *id)
 835 {
 836         struct rdma_id_private *id_priv;
 837         enum cma_state state;
 838
 839         id_priv = container_of(id, struct rdma_id_private, id);
 840         state = cma_exch(id_priv, CMA_DESTROYING);
 841         cma_cancel_operation(id_priv, state);
 842
 843         mutex_lock(&lock);
 844         if (id_priv->cma_dev) {
 845                 mutex_unlock(&lock);
 846                 switch (rdma_node_get_transport(id->device->node_type)) {
 847                 case RDMA_TRANSPORT_IB:
 848                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 849                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 850                         break;
 851                 case RDMA_TRANSPORT_IWARP:
 852                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 853                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 854                         break;
 855                 default:
 856                         break;
 857                 }
 858                 cma_leave_mc_groups(id_priv);
 859                 mutex_lock(&lock);
 860                 cma_detach_from_dev(id_priv);
 861         }
 862         mutex_unlock(&lock);
 863
 864         cma_release_port(id_priv);
 865         cma_deref_id(id_priv);
 866         wait_for_completion(&id_priv->comp);
 867
 868         if (id_priv->internal_id)
 869                 cma_deref_id(id_priv->id.context);
 870
 871         kfree(id_priv->id.route.path_rec);
 872         kfree(id_priv);
 873 }
 874 EXPORT_SYMBOL(rdma_destroy_id);
 875
 876 static int cma_rep_recv(struct rdma_id_private *id_priv)
 877 {
 878         int ret;
 879
 880         ret = cma_modify_qp_rtr(id_priv, NULL);
 881         if (ret)
 882                 goto reject;
 883
 884         ret = cma_modify_qp_rts(id_priv, NULL);
 885         if (ret)
 886                 goto reject;
 887
 888         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 889         if (ret)
 890                 goto reject;
 891
 892         return 0;
 893 reject:
 894         cma_modify_qp_err(id_priv);
 895         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 896                        NULL, 0, NULL, 0);
 897         return ret;
 898 }
 899
 900 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 901 {
 902         if (id_priv->id.ps == RDMA_PS_SDP &&
 903             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 904             SDP_MAJ_VERSION)
 905                 return -EINVAL;
 906
 907         return 0;
 908 }
 909
 910 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 911                                    struct ib_cm_rep_event_param *rep_data,
 912                                    void *private_data)
 913 {
 914         event->param.conn.private_data = private_data;
 915         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 916         event->param.conn.responder_resources = rep_data->responder_resources;
 917         event->param.conn.initiator_depth = rep_data->initiator_depth;
 918         event->param.conn.flow_control = rep_data->flow_control;
 919         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 920         event->param.conn.srq = rep_data->srq;
 921         event->param.conn.qp_num = rep_data->remote_qpn;
 922 }
 923
 924 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 925 {
 926         struct rdma_id_private *id_priv = cm_id->context;
 927         struct rdma_cm_event event;
 928         int ret = 0;
 929
 930         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
 931                 cma_disable_callback(id_priv, CMA_CONNECT)) ||
 932             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
 933                 cma_disable_callback(id_priv, CMA_DISCONNECT)))
 934                 return 0;
 935
 936         memset(&event, 0, sizeof event);
 937         switch (ib_event->event) {
 938         case IB_CM_REQ_ERROR:
 939         case IB_CM_REP_ERROR:
 940                 event.event = RDMA_CM_EVENT_UNREACHABLE;
 941                 event.status = -ETIMEDOUT;
 942                 break;
 943         case IB_CM_REP_RECEIVED:
 944                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
 945                 if (event.status)
 946                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
 947                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 948                         event.status = cma_rep_recv(id_priv);
 949                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
 950                                                      RDMA_CM_EVENT_ESTABLISHED;
 951                 } else
 952                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 953                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
 954                                        ib_event->private_data);
 955                 break;
 956         case IB_CM_RTU_RECEIVED:
 957         case IB_CM_USER_ESTABLISHED:
 958                 event.event = RDMA_CM_EVENT_ESTABLISHED;
 959                 break;
 960         case IB_CM_DREQ_ERROR:
 961                 event.status = -ETIMEDOUT; /* fall through */
 962         case IB_CM_DREQ_RECEIVED:
 963         case IB_CM_DREP_RECEIVED:
 964                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 965                         goto out;
 966                 event.event = RDMA_CM_EVENT_DISCONNECTED;
 967                 break;
 968         case IB_CM_TIMEWAIT_EXIT:
 969                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
 970                 break;
 971         case IB_CM_MRA_RECEIVED:
 972                 /* ignore event */
 973                 goto out;
 974         case IB_CM_REJ_RECEIVED:
 975                 cma_modify_qp_err(id_priv);
 976                 event.status = ib_event->param.rej_rcvd.reason;
 977                 event.event = RDMA_CM_EVENT_REJECTED;
 978                 event.param.conn.private_data = ib_event->private_data;
 979                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 980                 break;
 981         default:
 982                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
 983                        ib_event->event);
 984                 goto out;
 985         }
 986
 987         ret = id_priv->id.event_handler(&id_priv->id, &event);
 988         if (ret) {
 989                 /* Destroy the CM ID by returning a non-zero value. */
 990                 id_priv->cm_id.ib = NULL;
 991                 cma_exch(id_priv, CMA_DESTROYING);
 992                 mutex_unlock(&id_priv->handler_mutex);
 993                 rdma_destroy_id(&id_priv->id);
 994                 return ret;
 995         }
 996 out:
 997         mutex_unlock(&id_priv->handler_mutex);
 998         return ret;
 999 }
1000
1001 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1002                                                struct ib_cm_event *ib_event)
1003 {
1004         struct rdma_id_private *id_priv;
1005         struct rdma_cm_id *id;
1006         struct rdma_route *rt;
1007         union cma_ip_addr *src, *dst;
1008         __be16 port;
1009         u8 ip_ver;
1010         int ret;
1011
1012         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1013                              &ip_ver, &port, &src, &dst))
1014                 goto err;
1015
1016         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1017                             listen_id->ps);
1018         if (IS_ERR(id))
1019                 goto err;
1020
1021         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1022                           ip_ver, port, src, dst);
1023
1024         rt = &id->route;
1025         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1026         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1027                                GFP_KERNEL);
1028         if (!rt->path_rec)
1029                 goto destroy_id;
1030
1031         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1032         if (rt->num_paths == 2)
1033                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1034
1035         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1036         ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1037                                 &id->route.addr.dev_addr);
1038         if (ret)
1039                 goto destroy_id;
1040
1041         id_priv = container_of(id, struct rdma_id_private, id);
1042         id_priv->state = CMA_CONNECT;
1043         return id_priv;
1044
1045 destroy_id:
1046         rdma_destroy_id(id);
1047 err:
1048         return NULL;
1049 }
1050
1051 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1052                                               struct ib_cm_event *ib_event)
1053 {
1054         struct rdma_id_private *id_priv;
1055         struct rdma_cm_id *id;
1056         union cma_ip_addr *src, *dst;
1057         __be16 port;
1058         u8 ip_ver;
1059         int ret;
1060
1061         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1062                             listen_id->ps);
1063         if (IS_ERR(id))
1064                 return NULL;
1065
1066
1067         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1068                              &ip_ver, &port, &src, &dst))
1069                 goto err;
1070
1071         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1072                           ip_ver, port, src, dst);
1073
1074         ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1075                                 &id->route.addr.dev_addr);
1076         if (ret)
1077                 goto err;
1078
1079         id_priv = container_of(id, struct rdma_id_private, id);
1080         id_priv->state = CMA_CONNECT;
1081         return id_priv;
1082 err:
1083         rdma_destroy_id(id);
1084         return NULL;
1085 }
1086
1087 static void cma_set_req_event_data(struct rdma_cm_event *event,
1088                                    struct ib_cm_req_event_param *req_data,
1089                                    void *private_data, int offset)
1090 {
1091         event->param.conn.private_data = private_data + offset;
1092         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1093         event->param.conn.responder_resources = req_data->responder_resources;
1094         event->param.conn.initiator_depth = req_data->initiator_depth;
1095         event->param.conn.flow_control = req_data->flow_control;
1096         event->param.conn.retry_count = req_data->retry_count;
1097         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1098         event->param.conn.srq = req_data->srq;
1099         event->param.conn.qp_num = req_data->remote_qpn;
1100 }
1101
1102 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1103 {
1104         struct rdma_id_private *listen_id, *conn_id;
1105         struct rdma_cm_event event;
1106         int offset, ret;
1107
1108         listen_id = cm_id->context;
1109         if (cma_disable_callback(listen_id, CMA_LISTEN))
1110                 return -ECONNABORTED;
1111
1112         memset(&event, 0, sizeof event);
1113         offset = cma_user_data_offset(listen_id->id.ps);
1114         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1115         if (cma_is_ud_ps(listen_id->id.ps)) {
1116                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1117                 event.param.ud.private_data = ib_event->private_data + offset;
1118                 event.param.ud.private_data_len =
1119                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1120         } else {
1121                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1122                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1123                                        ib_event->private_data, offset);
1124         }
1125         if (!conn_id) {
1126                 ret = -ENOMEM;
1127                 goto out;
1128         }
1129
1130         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1131         mutex_lock(&lock);
1132         ret = cma_acquire_dev(conn_id);
1133         mutex_unlock(&lock);
1134         if (ret)
1135                 goto release_conn_id;
1136
1137         conn_id->cm_id.ib = cm_id;
1138         cm_id->context = conn_id;
1139         cm_id->cm_handler = cma_ib_handler;
1140
1141         ret = conn_id->id.event_handler(&conn_id->id, &event);
1142         if (!ret) {
1143                 /*
1144                  * Acquire mutex to prevent user executing rdma_destroy_id()
1145                  * while we're accessing the cm_id.
1146                  */
1147                 mutex_lock(&lock);
1148                 if (cma_comp(conn_id, CMA_CONNECT) &&
1149                     !cma_is_ud_ps(conn_id->id.ps))
1150                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1151                 mutex_unlock(&lock);
1152                 mutex_unlock(&conn_id->handler_mutex);
1153                 goto out;
1154         }
1155
1156         /* Destroy the CM ID by returning a non-zero value. */
1157         conn_id->cm_id.ib = NULL;
1158
1159 release_conn_id:
1160         cma_exch(conn_id, CMA_DESTROYING);
1161         mutex_unlock(&conn_id->handler_mutex);
1162         rdma_destroy_id(&conn_id->id);
1163
1164 out:
1165         mutex_unlock(&listen_id->handler_mutex);
1166         return ret;
1167 }
1168
1169 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1170 {
1171         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1172 }
1173
1174 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1175                                  struct ib_cm_compare_data *compare)
1176 {
1177         struct cma_hdr *cma_data, *cma_mask;
1178         struct sdp_hh *sdp_data, *sdp_mask;
1179         __be32 ip4_addr;
1180         struct in6_addr ip6_addr;
1181
1182         memset(compare, 0, sizeof *compare);
1183         cma_data = (void *) compare->data;
1184         cma_mask = (void *) compare->mask;
1185         sdp_data = (void *) compare->data;
1186         sdp_mask = (void *) compare->mask;
1187
1188         switch (addr->sa_family) {
1189         case AF_INET:
1190                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1191                 if (ps == RDMA_PS_SDP) {
1192                         sdp_set_ip_ver(sdp_data, 4);
1193                         sdp_set_ip_ver(sdp_mask, 0xF);
1194                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1195                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1196                 } else {
1197                         cma_set_ip_ver(cma_data, 4);
1198                         cma_set_ip_ver(cma_mask, 0xF);
1199                         cma_data->dst_addr.ip4.addr = ip4_addr;
1200                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1201                 }
1202                 break;
1203         case AF_INET6:
1204                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1205                 if (ps == RDMA_PS_SDP) {
1206                         sdp_set_ip_ver(sdp_data, 6);
1207                         sdp_set_ip_ver(sdp_mask, 0xF);
1208                         sdp_data->dst_addr.ip6 = ip6_addr;
1209                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1210                                sizeof sdp_mask->dst_addr.ip6);
1211                 } else {
1212                         cma_set_ip_ver(cma_data, 6);
1213                         cma_set_ip_ver(cma_mask, 0xF);
1214                         cma_data->dst_addr.ip6 = ip6_addr;
1215                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1216                                sizeof cma_mask->dst_addr.ip6);
1217                 }
1218                 break;
1219         default:
1220                 break;
1221         }
1222 }
1223
1224 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1225 {
1226         struct rdma_id_private *id_priv = iw_id->context;
1227         struct rdma_cm_event event;
1228         struct sockaddr_in *sin;
1229         int ret = 0;
1230
1231         if (cma_disable_callback(id_priv, CMA_CONNECT))
1232                 return 0;
1233
1234         memset(&event, 0, sizeof event);
1235         switch (iw_event->event) {
1236         case IW_CM_EVENT_CLOSE:
1237                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1238                 break;
1239         case IW_CM_EVENT_CONNECT_REPLY:
1240                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1241                 *sin = iw_event->local_addr;
1242                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1243                 *sin = iw_event->remote_addr;
1244                 switch (iw_event->status) {
1245                 case 0:
1246                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1247                         break;
1248                 case -ECONNRESET:
1249                 case -ECONNREFUSED:
1250                         event.event = RDMA_CM_EVENT_REJECTED;
1251                         break;
1252                 case -ETIMEDOUT:
1253                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1254                         break;
1255                 default:
1256                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1257                         break;
1258                 }
1259                 break;
1260         case IW_CM_EVENT_ESTABLISHED:
1261                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1262                 break;
1263         default:
1264                 BUG_ON(1);
1265         }
1266
1267         event.status = iw_event->status;
1268         event.param.conn.private_data = iw_event->private_data;
1269         event.param.conn.private_data_len = iw_event->private_data_len;
1270         ret = id_priv->id.event_handler(&id_priv->id, &event);
1271         if (ret) {
1272                 /* Destroy the CM ID by returning a non-zero value. */
1273                 id_priv->cm_id.iw = NULL;
1274                 cma_exch(id_priv, CMA_DESTROYING);
1275                 mutex_unlock(&id_priv->handler_mutex);
1276                 rdma_destroy_id(&id_priv->id);
1277                 return ret;
1278         }
1279
1280         mutex_unlock(&id_priv->handler_mutex);
1281         return ret;
1282 }
1283
1284 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1285                                struct iw_cm_event *iw_event)
1286 {
1287         struct rdma_cm_id *new_cm_id;
1288         struct rdma_id_private *listen_id, *conn_id;
1289         struct sockaddr_in *sin;
1290         struct net_device *dev = NULL;
1291         struct rdma_cm_event event;
1292         int ret;
1293         struct ib_device_attr attr;
1294
1295         listen_id = cm_id->context;
1296         if (cma_disable_callback(listen_id, CMA_LISTEN))
1297                 return -ECONNABORTED;
1298
1299         /* Create a new RDMA id for the new IW CM ID */
1300         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1301                                    listen_id->id.context,
1302                                    RDMA_PS_TCP);
1303         if (IS_ERR(new_cm_id)) {
1304                 ret = -ENOMEM;
1305                 goto out;
1306         }
1307         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1308         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1309         conn_id->state = CMA_CONNECT;
1310
1311         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1312         if (!dev) {
1313                 ret = -EADDRNOTAVAIL;
1314                 mutex_unlock(&conn_id->handler_mutex);
1315                 rdma_destroy_id(new_cm_id);
1316                 goto out;
1317         }
1318         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1319         if (ret) {
1320                 mutex_unlock(&conn_id->handler_mutex);
1321                 rdma_destroy_id(new_cm_id);
1322                 goto out;
1323         }
1324
1325         mutex_lock(&lock);
1326         ret = cma_acquire_dev(conn_id);
1327         mutex_unlock(&lock);
1328         if (ret) {
1329                 mutex_unlock(&conn_id->handler_mutex);
1330                 rdma_destroy_id(new_cm_id);
1331                 goto out;
1332         }
1333
1334         conn_id->cm_id.iw = cm_id;
1335         cm_id->context = conn_id;
1336         cm_id->cm_handler = cma_iw_handler;
1337
1338         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1339         *sin = iw_event->local_addr;
1340         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1341         *sin = iw_event->remote_addr;
1342
1343         ret = ib_query_device(conn_id->id.device, &attr);
1344         if (ret) {
1345                 mutex_unlock(&conn_id->handler_mutex);
1346                 rdma_destroy_id(new_cm_id);
1347                 goto out;
1348         }
1349
1350         memset(&event, 0, sizeof event);
1351         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1352         event.param.conn.private_data = iw_event->private_data;
1353         event.param.conn.private_data_len = iw_event->private_data_len;
1354         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1355         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1356         ret = conn_id->id.event_handler(&conn_id->id, &event);
1357         if (ret) {
1358                 /* User wants to destroy the CM ID */
1359                 conn_id->cm_id.iw = NULL;
1360                 cma_exch(conn_id, CMA_DESTROYING);
1361                 mutex_unlock(&conn_id->handler_mutex);
1362                 rdma_destroy_id(&conn_id->id);
1363                 goto out;
1364         }
1365
1366         mutex_unlock(&conn_id->handler_mutex);
1367
1368 out:
1369         if (dev)
1370                 dev_put(dev);
1371         mutex_unlock(&listen_id->handler_mutex);
1372         return ret;
1373 }
1374
1375 static int cma_ib_listen(struct rdma_id_private *id_priv)
1376 {
1377         struct ib_cm_compare_data compare_data;
1378         struct sockaddr *addr;
1379         __be64 svc_id;
1380         int ret;
1381
1382         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1383                                             id_priv);
1384         if (IS_ERR(id_priv->cm_id.ib))
1385                 return PTR_ERR(id_priv->cm_id.ib);
1386
1387         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1388         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1389         if (cma_any_addr(addr))
1390                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1391         else {
1392                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1393                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1394         }
1395
1396         if (ret) {
1397                 ib_destroy_cm_id(id_priv->cm_id.ib);
1398                 id_priv->cm_id.ib = NULL;
1399         }
1400
1401         return ret;
1402 }
1403
1404 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1405 {
1406         int ret;
1407         struct sockaddr_in *sin;
1408
1409         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1410                                             iw_conn_req_handler,
1411                                             id_priv);
1412         if (IS_ERR(id_priv->cm_id.iw))
1413                 return PTR_ERR(id_priv->cm_id.iw);
1414
1415         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1416         id_priv->cm_id.iw->local_addr = *sin;
1417
1418         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1419
1420         if (ret) {
1421                 iw_destroy_cm_id(id_priv->cm_id.iw);
1422                 id_priv->cm_id.iw = NULL;
1423         }
1424
1425         return ret;
1426 }
1427
1428 static int cma_listen_handler(struct rdma_cm_id *id,
1429                               struct rdma_cm_event *event)
1430 {
1431         struct rdma_id_private *id_priv = id->context;
1432
1433         id->context = id_priv->id.context;
1434         id->event_handler = id_priv->id.event_handler;
1435         return id_priv->id.event_handler(id, event);
1436 }
1437
1438 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1439                               struct cma_device *cma_dev)
1440 {
1441         struct rdma_id_private *dev_id_priv;
1442         struct rdma_cm_id *id;
1443         int ret;
1444
1445         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1446         if (IS_ERR(id))
1447                 return;
1448
1449         dev_id_priv = container_of(id, struct rdma_id_private, id);
1450
1451         dev_id_priv->state = CMA_ADDR_BOUND;
1452         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1453                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1454
1455         cma_attach_to_dev(dev_id_priv, cma_dev);
1456         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1457         atomic_inc(&id_priv->refcount);
1458         dev_id_priv->internal_id = 1;
1459
1460         ret = rdma_listen(id, id_priv->backlog);
1461         if (ret)
1462                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1463                        "listening on device %s\n", ret, cma_dev->device->name);
1464 }
1465
1466 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1467 {
1468         struct cma_device *cma_dev;
1469
1470         mutex_lock(&lock);
1471         list_add_tail(&id_priv->list, &listen_any_list);
1472         list_for_each_entry(cma_dev, &dev_list, list)
1473                 cma_listen_on_dev(id_priv, cma_dev);
1474         mutex_unlock(&lock);
1475 }
1476
1477 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1478 {
1479         struct sockaddr_storage addr_in;
1480
1481         memset(&addr_in, 0, sizeof addr_in);
1482         addr_in.ss_family = af;
1483         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1484 }
1485
1486 int rdma_listen(struct rdma_cm_id *id, int backlog)
1487 {
1488         struct rdma_id_private *id_priv;
1489         int ret;
1490
1491         id_priv = container_of(id, struct rdma_id_private, id);
1492         if (id_priv->state == CMA_IDLE) {
1493                 ret = cma_bind_any(id, AF_INET);
1494                 if (ret)
1495                         return ret;
1496         }
1497
1498         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1499                 return -EINVAL;
1500
1501         id_priv->backlog = backlog;
1502         if (id->device) {
1503                 switch (rdma_node_get_transport(id->device->node_type)) {
1504                 case RDMA_TRANSPORT_IB:
1505                         ret = cma_ib_listen(id_priv);
1506                         if (ret)
1507                                 goto err;
1508                         break;
1509                 case RDMA_TRANSPORT_IWARP:
1510                         ret = cma_iw_listen(id_priv, backlog);
1511                         if (ret)
1512                                 goto err;
1513                         break;
1514                 default:
1515                         ret = -ENOSYS;
1516                         goto err;
1517                 }
1518         } else
1519                 cma_listen_on_all(id_priv);
1520
1521         return 0;
1522 err:
1523         id_priv->backlog = 0;
1524         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1525         return ret;
1526 }
1527 EXPORT_SYMBOL(rdma_listen);
1528
1529 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1530 {
1531         struct rdma_id_private *id_priv;
1532
1533         id_priv = container_of(id, struct rdma_id_private, id);
1534         id_priv->tos = (u8) tos;
1535 }
1536 EXPORT_SYMBOL(rdma_set_service_type);
1537
1538 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1539                               void *context)
1540 {
1541         struct cma_work *work = context;
1542         struct rdma_route *route;
1543
1544         route = &work->id->id.route;
1545
1546         if (!status) {
1547                 route->num_paths = 1;
1548                 *route->path_rec = *path_rec;
1549         } else {
1550                 work->old_state = CMA_ROUTE_QUERY;
1551                 work->new_state = CMA_ADDR_RESOLVED;
1552                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1553                 work->event.status = status;
1554         }
1555
1556         queue_work(cma_wq, &work->work);
1557 }
1558
1559 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1560                               struct cma_work *work)
1561 {
1562         struct rdma_addr *addr = &id_priv->id.route.addr;
1563         struct ib_sa_path_rec path_rec;
1564         ib_sa_comp_mask comp_mask;
1565         struct sockaddr_in6 *sin6;
1566
1567         memset(&path_rec, 0, sizeof path_rec);
1568         ib_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1569         ib_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1570         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1571         path_rec.numb_path = 1;
1572         path_rec.reversible = 1;
1573         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1574                                                         (struct sockaddr *) &addr->dst_addr);
1575
1576         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1577                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1578                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1579
1580         if (addr->src_addr.ss_family == AF_INET) {
1581                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1582                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1583         } else {
1584                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1585                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1586                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1587         }
1588
1589         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1590                                                id_priv->id.port_num, &path_rec,
1591                                                comp_mask, timeout_ms,
1592                                                GFP_KERNEL, cma_query_handler,
1593                                                work, &id_priv->query);
1594
1595         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1596 }
1597
1598 static void cma_work_handler(struct work_struct *_work)
1599 {
1600         struct cma_work *work = container_of(_work, struct cma_work, work);
1601         struct rdma_id_private *id_priv = work->id;
1602         int destroy = 0;
1603
1604         mutex_lock(&id_priv->handler_mutex);
1605         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1606                 goto out;
1607
1608         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1609                 cma_exch(id_priv, CMA_DESTROYING);
1610                 destroy = 1;
1611         }
1612 out:
1613         mutex_unlock(&id_priv->handler_mutex);
1614         cma_deref_id(id_priv);
1615         if (destroy)
1616                 rdma_destroy_id(&id_priv->id);
1617         kfree(work);
1618 }
1619
1620 static void cma_ndev_work_handler(struct work_struct *_work)
1621 {
1622         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1623         struct rdma_id_private *id_priv = work->id;
1624         int destroy = 0;
1625
1626         mutex_lock(&id_priv->handler_mutex);
1627         if (id_priv->state == CMA_DESTROYING ||
1628             id_priv->state == CMA_DEVICE_REMOVAL)
1629                 goto out;
1630
1631         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1632                 cma_exch(id_priv, CMA_DESTROYING);
1633                 destroy = 1;
1634         }
1635
1636 out:
1637         mutex_unlock(&id_priv->handler_mutex);
1638         cma_deref_id(id_priv);
1639         if (destroy)
1640                 rdma_destroy_id(&id_priv->id);
1641         kfree(work);
1642 }
1643
1644 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1645 {
1646         struct rdma_route *route = &id_priv->id.route;
1647         struct cma_work *work;
1648         int ret;
1649
1650         work = kzalloc(sizeof *work, GFP_KERNEL);
1651         if (!work)
1652                 return -ENOMEM;
1653
1654         work->id = id_priv;
1655         INIT_WORK(&work->work, cma_work_handler);
1656         work->old_state = CMA_ROUTE_QUERY;
1657         work->new_state = CMA_ROUTE_RESOLVED;
1658         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1659
1660         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1661         if (!route->path_rec) {
1662                 ret = -ENOMEM;
1663                 goto err1;
1664         }
1665
1666         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1667         if (ret)
1668                 goto err2;
1669
1670         return 0;
1671 err2:
1672         kfree(route->path_rec);
1673         route->path_rec = NULL;
1674 err1:
1675         kfree(work);
1676         return ret;
1677 }
1678
1679 int rdma_set_ib_paths(struct rdma_cm_id *id,
1680                       struct ib_sa_path_rec *path_rec, int num_paths)
1681 {
1682         struct rdma_id_private *id_priv;
1683         int ret;
1684
1685         id_priv = container_of(id, struct rdma_id_private, id);
1686         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1687                 return -EINVAL;
1688
1689         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1690         if (!id->route.path_rec) {
1691                 ret = -ENOMEM;
1692                 goto err;
1693         }
1694
1695         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1696         return 0;
1697 err:
1698         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1699         return ret;
1700 }
1701 EXPORT_SYMBOL(rdma_set_ib_paths);
1702
1703 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1704 {
1705         struct cma_work *work;
1706
1707         work = kzalloc(sizeof *work, GFP_KERNEL);
1708         if (!work)
1709                 return -ENOMEM;
1710
1711         work->id = id_priv;
1712         INIT_WORK(&work->work, cma_work_handler);
1713         work->old_state = CMA_ROUTE_QUERY;
1714         work->new_state = CMA_ROUTE_RESOLVED;
1715         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1716         queue_work(cma_wq, &work->work);
1717         return 0;
1718 }
1719
1720 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1721 {
1722         struct rdma_id_private *id_priv;
1723         int ret;
1724
1725         id_priv = container_of(id, struct rdma_id_private, id);
1726         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1727                 return -EINVAL;
1728
1729         atomic_inc(&id_priv->refcount);
1730         switch (rdma_node_get_transport(id->device->node_type)) {
1731         case RDMA_TRANSPORT_IB:
1732                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1733                 break;
1734         case RDMA_TRANSPORT_IWARP:
1735                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1736                 break;
1737         default:
1738                 ret = -ENOSYS;
1739                 break;
1740         }
1741         if (ret)
1742                 goto err;
1743
1744         return 0;
1745 err:
1746         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1747         cma_deref_id(id_priv);
1748         return ret;
1749 }
1750 EXPORT_SYMBOL(rdma_resolve_route);
1751
1752 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1753 {
1754         struct cma_device *cma_dev;
1755         struct ib_port_attr port_attr;
1756         union ib_gid gid;
1757         u16 pkey;
1758         int ret;
1759         u8 p;
1760
1761         mutex_lock(&lock);
1762         if (list_empty(&dev_list)) {
1763                 ret = -ENODEV;
1764                 goto out;
1765         }
1766         list_for_each_entry(cma_dev, &dev_list, list)
1767                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1768                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1769                             port_attr.state == IB_PORT_ACTIVE)
1770                                 goto port_found;
1771
1772         p = 1;
1773         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1774
1775 port_found:
1776         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1777         if (ret)
1778                 goto out;
1779
1780         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1781         if (ret)
1782                 goto out;
1783
1784         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1785         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1786         id_priv->id.port_num = p;
1787         cma_attach_to_dev(id_priv, cma_dev);
1788 out:
1789         mutex_unlock(&lock);
1790         return ret;
1791 }
1792
1793 static void addr_handler(int status, struct sockaddr *src_addr,
1794                          struct rdma_dev_addr *dev_addr, void *context)
1795 {
1796         struct rdma_id_private *id_priv = context;
1797         struct rdma_cm_event event;
1798
1799         memset(&event, 0, sizeof event);
1800         mutex_lock(&id_priv->handler_mutex);
1801
1802         /*
1803          * Grab mutex to block rdma_destroy_id() from removing the device while
1804          * we're trying to acquire it.
1805          */
1806         mutex_lock(&lock);
1807         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1808                 mutex_unlock(&lock);
1809                 goto out;
1810         }
1811
1812         if (!status && !id_priv->cma_dev)
1813                 status = cma_acquire_dev(id_priv);
1814         mutex_unlock(&lock);
1815
1816         if (status) {
1817                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1818                         goto out;
1819                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1820                 event.status = status;
1821         } else {
1822                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1823                        ip_addr_size(src_addr));
1824                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1825         }
1826
1827         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1828                 cma_exch(id_priv, CMA_DESTROYING);
1829                 mutex_unlock(&id_priv->handler_mutex);
1830                 cma_deref_id(id_priv);
1831                 rdma_destroy_id(&id_priv->id);
1832                 return;
1833         }
1834 out:
1835         mutex_unlock(&id_priv->handler_mutex);
1836         cma_deref_id(id_priv);
1837 }
1838
1839 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1840 {
1841         struct cma_work *work;
1842         struct sockaddr_in *src_in, *dst_in;
1843         union ib_gid gid;
1844         int ret;
1845
1846         work = kzalloc(sizeof *work, GFP_KERNEL);
1847         if (!work)
1848                 return -ENOMEM;
1849
1850         if (!id_priv->cma_dev) {
1851                 ret = cma_bind_loopback(id_priv);
1852                 if (ret)
1853                         goto err;
1854         }
1855
1856         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1857         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1858
1859         if (cma_zero_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)) {
1860                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1861                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1862                 src_in->sin_family = dst_in->sin_family;
1863                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1864         }
1865
1866         work->id = id_priv;
1867         INIT_WORK(&work->work, cma_work_handler);
1868         work->old_state = CMA_ADDR_QUERY;
1869         work->new_state = CMA_ADDR_RESOLVED;
1870         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1871         queue_work(cma_wq, &work->work);
1872         return 0;
1873 err:
1874         kfree(work);
1875         return ret;
1876 }
1877
1878 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1879                          struct sockaddr *dst_addr)
1880 {
1881         if (src_addr && src_addr->sa_family)
1882                 return rdma_bind_addr(id, src_addr);
1883         else
1884                 return cma_bind_any(id, dst_addr->sa_family);
1885 }
1886
1887 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1888                       struct sockaddr *dst_addr, int timeout_ms)
1889 {
1890         struct rdma_id_private *id_priv;
1891         int ret;
1892
1893         id_priv = container_of(id, struct rdma_id_private, id);
1894         if (id_priv->state == CMA_IDLE) {
1895                 ret = cma_bind_addr(id, src_addr, dst_addr);
1896                 if (ret)
1897                         return ret;
1898         }
1899
1900         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1901                 return -EINVAL;
1902
1903         atomic_inc(&id_priv->refcount);
1904         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1905         if (cma_any_addr(dst_addr))
1906                 ret = cma_resolve_loopback(id_priv);
1907         else
1908                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
1909                                       dst_addr, &id->route.addr.dev_addr,
1910                                       timeout_ms, addr_handler, id_priv);
1911         if (ret)
1912                 goto err;
1913
1914         return 0;
1915 err:
1916         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1917         cma_deref_id(id_priv);
1918         return ret;
1919 }
1920 EXPORT_SYMBOL(rdma_resolve_addr);
1921
1922 static void cma_bind_port(struct rdma_bind_list *bind_list,
1923                           struct rdma_id_private *id_priv)
1924 {
1925         struct sockaddr_in *sin;
1926
1927         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1928         sin->sin_port = htons(bind_list->port);
1929         id_priv->bind_list = bind_list;
1930         hlist_add_head(&id_priv->node, &bind_list->owners);
1931 }
1932
1933 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1934                           unsigned short snum)
1935 {
1936         struct rdma_bind_list *bind_list;
1937         int port, ret;
1938
1939         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1940         if (!bind_list)
1941                 return -ENOMEM;
1942
1943         do {
1944                 ret = idr_get_new_above(ps, bind_list, snum, &port);
1945         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1946
1947         if (ret)
1948                 goto err1;
1949
1950         if (port != snum) {
1951                 ret = -EADDRNOTAVAIL;
1952                 goto err2;
1953         }
1954
1955         bind_list->ps = ps;
1956         bind_list->port = (unsigned short) port;
1957         cma_bind_port(bind_list, id_priv);
1958         return 0;
1959 err2:
1960         idr_remove(ps, port);
1961 err1:
1962         kfree(bind_list);
1963         return ret;
1964 }
1965
1966 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1967 {
1968         struct rdma_bind_list *bind_list;
1969         int port, ret, low, high;
1970
1971         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1972         if (!bind_list)
1973                 return -ENOMEM;
1974
1975 retry:
1976         /* FIXME: add proper port randomization per like inet_csk_get_port */
1977         do {
1978                 ret = idr_get_new_above(ps, bind_list, next_port, &port);
1979         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1980
1981         if (ret)
1982                 goto err1;
1983
1984         inet_get_local_port_range(&low, &high);
1985         if (port > high) {
1986                 if (next_port != low) {
1987                         idr_remove(ps, port);
1988                         next_port = low;
1989                         goto retry;
1990                 }
1991                 ret = -EADDRNOTAVAIL;
1992                 goto err2;
1993         }
1994
1995         if (port == high)
1996                 next_port = low;
1997         else
1998                 next_port = port + 1;
1999
2000         bind_list->ps = ps;
2001         bind_list->port = (unsigned short) port;
2002         cma_bind_port(bind_list, id_priv);
2003         return 0;
2004 err2:
2005         idr_remove(ps, port);
2006 err1:
2007         kfree(bind_list);
2008         return ret;
2009 }
2010
2011 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2012 {
2013         struct rdma_id_private *cur_id;
2014         struct sockaddr_in *sin, *cur_sin;
2015         struct rdma_bind_list *bind_list;
2016         struct hlist_node *node;
2017         unsigned short snum;
2018
2019         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2020         snum = ntohs(sin->sin_port);
2021         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2022                 return -EACCES;
2023
2024         bind_list = idr_find(ps, snum);
2025         if (!bind_list)
2026                 return cma_alloc_port(ps, id_priv, snum);
2027
2028         /*
2029          * We don't support binding to any address if anyone is bound to
2030          * a specific address on the same port.
2031          */
2032         if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2033                 return -EADDRNOTAVAIL;
2034
2035         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2036                 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2037                         return -EADDRNOTAVAIL;
2038
2039                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2040                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2041                         return -EADDRINUSE;
2042         }
2043
2044         cma_bind_port(bind_list, id_priv);
2045         return 0;
2046 }
2047
2048 static int cma_get_port(struct rdma_id_private *id_priv)
2049 {
2050         struct idr *ps;
2051         int ret;
2052
2053         switch (id_priv->id.ps) {
2054         case RDMA_PS_SDP:
2055                 ps = &sdp_ps;
2056                 break;
2057         case RDMA_PS_TCP:
2058                 ps = &tcp_ps;
2059                 break;
2060         case RDMA_PS_UDP:
2061                 ps = &udp_ps;
2062                 break;
2063         case RDMA_PS_IPOIB:
2064                 ps = &ipoib_ps;
2065                 break;
2066         default:
2067                 return -EPROTONOSUPPORT;
2068         }
2069
2070         mutex_lock(&lock);
2071         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2072                 ret = cma_alloc_any_port(ps, id_priv);
2073         else
2074                 ret = cma_use_port(ps, id_priv);
2075         mutex_unlock(&lock);
2076
2077         return ret;
2078 }
2079
2080 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2081 {
2082         struct rdma_id_private *id_priv;
2083         int ret;
2084
2085         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2086                 return -EAFNOSUPPORT;
2087
2088         id_priv = container_of(id, struct rdma_id_private, id);
2089         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2090                 return -EINVAL;
2091
2092         if (!cma_any_addr(addr)) {
2093                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2094                 if (ret)
2095                         goto err1;
2096
2097                 mutex_lock(&lock);
2098                 ret = cma_acquire_dev(id_priv);
2099                 mutex_unlock(&lock);
2100                 if (ret)
2101                         goto err1;
2102         }
2103
2104         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2105         ret = cma_get_port(id_priv);
2106         if (ret)
2107                 goto err2;
2108
2109         return 0;
2110 err2:
2111         if (!cma_any_addr(addr)) {
2112                 mutex_lock(&lock);
2113                 cma_detach_from_dev(id_priv);
2114                 mutex_unlock(&lock);
2115         }
2116 err1:
2117         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2118         return ret;
2119 }
2120 EXPORT_SYMBOL(rdma_bind_addr);
2121
2122 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2123                           struct rdma_route *route)
2124 {
2125         struct cma_hdr *cma_hdr;
2126         struct sdp_hh *sdp_hdr;
2127
2128         if (route->addr.src_addr.ss_family == AF_INET) {
2129                 struct sockaddr_in *src4, *dst4;
2130
2131                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2132                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2133
2134                 switch (ps) {
2135                 case RDMA_PS_SDP:
2136                         sdp_hdr = hdr;
2137                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2138                                 return -EINVAL;
2139                         sdp_set_ip_ver(sdp_hdr, 4);
2140                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2141                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2142                         sdp_hdr->port = src4->sin_port;
2143                         break;
2144                 default:
2145                         cma_hdr = hdr;
2146                         cma_hdr->cma_version = CMA_VERSION;
2147                         cma_set_ip_ver(cma_hdr, 4);
2148                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2149                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2150                         cma_hdr->port = src4->sin_port;
2151                         break;
2152                 }
2153         } else {
2154                 struct sockaddr_in6 *src6, *dst6;
2155
2156                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2157                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2158
2159                 switch (ps) {
2160                 case RDMA_PS_SDP:
2161                         sdp_hdr = hdr;
2162                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2163                                 return -EINVAL;
2164                         sdp_set_ip_ver(sdp_hdr, 6);
2165                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2166                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2167                         sdp_hdr->port = src6->sin6_port;
2168                         break;
2169                 default:
2170                         cma_hdr = hdr;
2171                         cma_hdr->cma_version = CMA_VERSION;
2172                         cma_set_ip_ver(cma_hdr, 6);
2173                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2174                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2175                         cma_hdr->port = src6->sin6_port;
2176                         break;
2177                 }
2178         }
2179         return 0;
2180 }
2181
2182 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2183                                 struct ib_cm_event *ib_event)
2184 {
2185         struct rdma_id_private *id_priv = cm_id->context;
2186         struct rdma_cm_event event;
2187         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2188         int ret = 0;
2189
2190         if (cma_disable_callback(id_priv, CMA_CONNECT))
2191                 return 0;
2192
2193         memset(&event, 0, sizeof event);
2194         switch (ib_event->event) {
2195         case IB_CM_SIDR_REQ_ERROR:
2196                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2197                 event.status = -ETIMEDOUT;
2198                 break;
2199         case IB_CM_SIDR_REP_RECEIVED:
2200                 event.param.ud.private_data = ib_event->private_data;
2201                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2202                 if (rep->status != IB_SIDR_SUCCESS) {
2203                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2204                         event.status = ib_event->param.sidr_rep_rcvd.status;
2205                         break;
2206                 }
2207                 ret = cma_set_qkey(id_priv);
2208                 if (ret) {
2209                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2210                         event.status = -EINVAL;
2211                         break;
2212                 }
2213                 if (id_priv->qkey != rep->qkey) {
2214                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2215                         event.status = -EINVAL;
2216                         break;
2217                 }
2218                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2219                                      id_priv->id.route.path_rec,
2220                                      &event.param.ud.ah_attr);
2221                 event.param.ud.qp_num = rep->qpn;
2222                 event.param.ud.qkey = rep->qkey;
2223                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2224                 event.status = 0;
2225                 break;
2226         default:
2227                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2228                        ib_event->event);
2229                 goto out;
2230         }
2231
2232         ret = id_priv->id.event_handler(&id_priv->id, &event);
2233         if (ret) {
2234                 /* Destroy the CM ID by returning a non-zero value. */
2235                 id_priv->cm_id.ib = NULL;
2236                 cma_exch(id_priv, CMA_DESTROYING);
2237                 mutex_unlock(&id_priv->handler_mutex);
2238                 rdma_destroy_id(&id_priv->id);
2239                 return ret;
2240         }
2241 out:
2242         mutex_unlock(&id_priv->handler_mutex);
2243         return ret;
2244 }
2245
2246 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2247                               struct rdma_conn_param *conn_param)
2248 {
2249         struct ib_cm_sidr_req_param req;
2250         struct rdma_route *route;
2251         int ret;
2252
2253         req.private_data_len = sizeof(struct cma_hdr) +
2254                                conn_param->private_data_len;
2255         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2256         if (!req.private_data)
2257                 return -ENOMEM;
2258
2259         if (conn_param->private_data && conn_param->private_data_len)
2260                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2261                        conn_param->private_data, conn_param->private_data_len);
2262
2263         route = &id_priv->id.route;
2264         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2265         if (ret)
2266                 goto out;
2267
2268         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2269                                             cma_sidr_rep_handler, id_priv);
2270         if (IS_ERR(id_priv->cm_id.ib)) {
2271                 ret = PTR_ERR(id_priv->cm_id.ib);
2272                 goto out;
2273         }
2274
2275         req.path = route->path_rec;
2276         req.service_id = cma_get_service_id(id_priv->id.ps,
2277                                             (struct sockaddr *) &route->addr.dst_addr);
2278         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2279         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2280
2281         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2282         if (ret) {
2283                 ib_destroy_cm_id(id_priv->cm_id.ib);
2284                 id_priv->cm_id.ib = NULL;
2285         }
2286 out:
2287         kfree(req.private_data);
2288         return ret;
2289 }
2290
2291 static int cma_connect_ib(struct rdma_id_private *id_priv,
2292                           struct rdma_conn_param *conn_param)
2293 {
2294         struct ib_cm_req_param req;
2295         struct rdma_route *route;
2296         void *private_data;
2297         int offset, ret;
2298
2299         memset(&req, 0, sizeof req);
2300         offset = cma_user_data_offset(id_priv->id.ps);
2301         req.private_data_len = offset + conn_param->private_data_len;
2302         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2303         if (!private_data)
2304                 return -ENOMEM;
2305
2306         if (conn_param->private_data && conn_param->private_data_len)
2307                 memcpy(private_data + offset, conn_param->private_data,
2308                        conn_param->private_data_len);
2309
2310         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2311                                             id_priv);
2312         if (IS_ERR(id_priv->cm_id.ib)) {
2313                 ret = PTR_ERR(id_priv->cm_id.ib);
2314                 goto out;
2315         }
2316
2317         route = &id_priv->id.route;
2318         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2319         if (ret)
2320                 goto out;
2321         req.private_data = private_data;
2322
2323         req.primary_path = &route->path_rec[0];
2324         if (route->num_paths == 2)
2325                 req.alternate_path = &route->path_rec[1];
2326
2327         req.service_id = cma_get_service_id(id_priv->id.ps,
2328                                             (struct sockaddr *) &route->addr.dst_addr);
2329         req.qp_num = id_priv->qp_num;
2330         req.qp_type = IB_QPT_RC;
2331         req.starting_psn = id_priv->seq_num;
2332         req.responder_resources = conn_param->responder_resources;
2333         req.initiator_depth = conn_param->initiator_depth;
2334         req.flow_control = conn_param->flow_control;
2335         req.retry_count = conn_param->retry_count;
2336         req.rnr_retry_count = conn_param->rnr_retry_count;
2337         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2338         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2339         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2340         req.srq = id_priv->srq ? 1 : 0;
2341
2342         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2343 out:
2344         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2345                 ib_destroy_cm_id(id_priv->cm_id.ib);
2346                 id_priv->cm_id.ib = NULL;
2347         }
2348
2349         kfree(private_data);
2350         return ret;
2351 }
2352
2353 static int cma_connect_iw(struct rdma_id_private *id_priv,
2354                           struct rdma_conn_param *conn_param)
2355 {
2356         struct iw_cm_id *cm_id;
2357         struct sockaddr_in* sin;
2358         int ret;
2359         struct iw_cm_conn_param iw_param;
2360
2361         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2362         if (IS_ERR(cm_id)) {
2363                 ret = PTR_ERR(cm_id);
2364                 goto out;
2365         }
2366
2367         id_priv->cm_id.iw = cm_id;
2368
2369         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2370         cm_id->local_addr = *sin;
2371
2372         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2373         cm_id->remote_addr = *sin;
2374
2375         ret = cma_modify_qp_rtr(id_priv, conn_param);
2376         if (ret)
2377                 goto out;
2378
2379         iw_param.ord = conn_param->initiator_depth;
2380         iw_param.ird = conn_param->responder_resources;
2381         iw_param.private_data = conn_param->private_data;
2382         iw_param.private_data_len = conn_param->private_data_len;
2383         if (id_priv->id.qp)
2384                 iw_param.qpn = id_priv->qp_num;
2385         else
2386                 iw_param.qpn = conn_param->qp_num;
2387         ret = iw_cm_connect(cm_id, &iw_param);
2388 out:
2389         if (ret && !IS_ERR(cm_id)) {
2390                 iw_destroy_cm_id(cm_id);
2391                 id_priv->cm_id.iw = NULL;
2392         }
2393         return ret;
2394 }
2395
2396 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2397 {
2398         struct rdma_id_private *id_priv;
2399         int ret;
2400
2401         id_priv = container_of(id, struct rdma_id_private, id);
2402         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2403                 return -EINVAL;
2404
2405         if (!id->qp) {
2406                 id_priv->qp_num = conn_param->qp_num;
2407                 id_priv->srq = conn_param->srq;
2408         }
2409
2410         switch (rdma_node_get_transport(id->device->node_type)) {
2411         case RDMA_TRANSPORT_IB:
2412                 if (cma_is_ud_ps(id->ps))
2413                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2414                 else
2415                         ret = cma_connect_ib(id_priv, conn_param);
2416                 break;
2417         case RDMA_TRANSPORT_IWARP:
2418                 ret = cma_connect_iw(id_priv, conn_param);
2419                 break;
2420         default:
2421                 ret = -ENOSYS;
2422                 break;
2423         }
2424         if (ret)
2425                 goto err;
2426
2427         return 0;
2428 err:
2429         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2430         return ret;
2431 }
2432 EXPORT_SYMBOL(rdma_connect);
2433
2434 static int cma_accept_ib(struct rdma_id_private *id_priv,
2435                          struct rdma_conn_param *conn_param)
2436 {
2437         struct ib_cm_rep_param rep;
2438         int ret;
2439
2440         ret = cma_modify_qp_rtr(id_priv, conn_param);
2441         if (ret)
2442                 goto out;
2443
2444         ret = cma_modify_qp_rts(id_priv, conn_param);
2445         if (ret)
2446                 goto out;
2447
2448         memset(&rep, 0, sizeof rep);
2449         rep.qp_num = id_priv->qp_num;
2450         rep.starting_psn = id_priv->seq_num;
2451         rep.private_data = conn_param->private_data;
2452         rep.private_data_len = conn_param->private_data_len;
2453         rep.responder_resources = conn_param->responder_resources;
2454         rep.initiator_depth = conn_param->initiator_depth;
2455         rep.failover_accepted = 0;
2456         rep.flow_control = conn_param->flow_control;
2457         rep.rnr_retry_count = conn_param->rnr_retry_count;
2458         rep.srq = id_priv->srq ? 1 : 0;
2459
2460         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2461 out:
2462         return ret;
2463 }
2464
2465 static int cma_accept_iw(struct rdma_id_private *id_priv,
2466                   struct rdma_conn_param *conn_param)
2467 {
2468         struct iw_cm_conn_param iw_param;
2469         int ret;
2470
2471         ret = cma_modify_qp_rtr(id_priv, conn_param);
2472         if (ret)
2473                 return ret;
2474
2475         iw_param.ord = conn_param->initiator_depth;
2476         iw_param.ird = conn_param->responder_resources;
2477         iw_param.private_data = conn_param->private_data;
2478         iw_param.private_data_len = conn_param->private_data_len;
2479         if (id_priv->id.qp) {
2480                 iw_param.qpn = id_priv->qp_num;
2481         } else
2482                 iw_param.qpn = conn_param->qp_num;
2483
2484         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2485 }
2486
2487 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2488                              enum ib_cm_sidr_status status,
2489                              const void *private_data, int private_data_len)
2490 {
2491         struct ib_cm_sidr_rep_param rep;
2492         int ret;
2493
2494         memset(&rep, 0, sizeof rep);
2495         rep.status = status;
2496         if (status == IB_SIDR_SUCCESS) {
2497                 ret = cma_set_qkey(id_priv);
2498                 if (ret)
2499                         return ret;
2500                 rep.qp_num = id_priv->qp_num;
2501                 rep.qkey = id_priv->qkey;
2502         }
2503         rep.private_data = private_data;
2504         rep.private_data_len = private_data_len;
2505
2506         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2507 }
2508
2509 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2510 {
2511         struct rdma_id_private *id_priv;
2512         int ret;
2513
2514         id_priv = container_of(id, struct rdma_id_private, id);
2515         if (!cma_comp(id_priv, CMA_CONNECT))
2516                 return -EINVAL;
2517
2518         if (!id->qp && conn_param) {
2519                 id_priv->qp_num = conn_param->qp_num;
2520                 id_priv->srq = conn_param->srq;
2521         }
2522
2523         switch (rdma_node_get_transport(id->device->node_type)) {
2524         case RDMA_TRANSPORT_IB:
2525                 if (cma_is_ud_ps(id->ps))
2526                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2527                                                 conn_param->private_data,
2528                                                 conn_param->private_data_len);
2529                 else if (conn_param)
2530                         ret = cma_accept_ib(id_priv, conn_param);
2531                 else
2532                         ret = cma_rep_recv(id_priv);
2533                 break;
2534         case RDMA_TRANSPORT_IWARP:
2535                 ret = cma_accept_iw(id_priv, conn_param);
2536                 break;
2537         default:
2538                 ret = -ENOSYS;
2539                 break;
2540         }
2541
2542         if (ret)
2543                 goto reject;
2544
2545         return 0;
2546 reject:
2547         cma_modify_qp_err(id_priv);
2548         rdma_reject(id, NULL, 0);
2549         return ret;
2550 }
2551 EXPORT_SYMBOL(rdma_accept);
2552
2553 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2554 {
2555         struct rdma_id_private *id_priv;
2556         int ret;
2557
2558         id_priv = container_of(id, struct rdma_id_private, id);
2559         if (!cma_has_cm_dev(id_priv))
2560                 return -EINVAL;
2561
2562         switch (id->device->node_type) {
2563         case RDMA_NODE_IB_CA:
2564                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2565                 break;
2566         default:
2567                 ret = 0;
2568                 break;
2569         }
2570         return ret;
2571 }
2572 EXPORT_SYMBOL(rdma_notify);
2573
2574 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2575                 u8 private_data_len)
2576 {
2577         struct rdma_id_private *id_priv;
2578         int ret;
2579
2580         id_priv = container_of(id, struct rdma_id_private, id);
2581         if (!cma_has_cm_dev(id_priv))
2582                 return -EINVAL;
2583
2584         switch (rdma_node_get_transport(id->device->node_type)) {
2585         case RDMA_TRANSPORT_IB:
2586                 if (cma_is_ud_ps(id->ps))
2587                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2588                                                 private_data, private_data_len);
2589                 else
2590                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2591                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2592                                              0, private_data, private_data_len);
2593                 break;
2594         case RDMA_TRANSPORT_IWARP:
2595                 ret = iw_cm_reject(id_priv->cm_id.iw,
2596                                    private_data, private_data_len);
2597                 break;
2598         default:
2599                 ret = -ENOSYS;
2600                 break;
2601         }
2602         return ret;
2603 }
2604 EXPORT_SYMBOL(rdma_reject);
2605
2606 int rdma_disconnect(struct rdma_cm_id *id)
2607 {
2608         struct rdma_id_private *id_priv;
2609         int ret;
2610
2611         id_priv = container_of(id, struct rdma_id_private, id);
2612         if (!cma_has_cm_dev(id_priv))
2613                 return -EINVAL;
2614
2615         switch (rdma_node_get_transport(id->device->node_type)) {
2616         case RDMA_TRANSPORT_IB:
2617                 ret = cma_modify_qp_err(id_priv);
2618                 if (ret)
2619                         goto out;
2620                 /* Initiate or respond to a disconnect. */
2621                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2622                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2623                 break;
2624         case RDMA_TRANSPORT_IWARP:
2625                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2626                 break;
2627         default:
2628                 ret = -EINVAL;
2629                 break;
2630         }
2631 out:
2632         return ret;
2633 }
2634 EXPORT_SYMBOL(rdma_disconnect);
2635
2636 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2637 {
2638         struct rdma_id_private *id_priv;
2639         struct cma_multicast *mc = multicast->context;
2640         struct rdma_cm_event event;
2641         int ret;
2642
2643         id_priv = mc->id_priv;
2644         if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2645             cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2646                 return 0;
2647
2648         mutex_lock(&id_priv->qp_mutex);
2649         if (!status && id_priv->id.qp)
2650                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2651                                          multicast->rec.mlid);
2652         mutex_unlock(&id_priv->qp_mutex);
2653
2654         memset(&event, 0, sizeof event);
2655         event.status = status;
2656         event.param.ud.private_data = mc->context;
2657         if (!status) {
2658                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2659                 ib_init_ah_from_mcmember(id_priv->id.device,
2660                                          id_priv->id.port_num, &multicast->rec,
2661                                          &event.param.ud.ah_attr);
2662                 event.param.ud.qp_num = 0xFFFFFF;
2663                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2664         } else
2665                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2666
2667         ret = id_priv->id.event_handler(&id_priv->id, &event);
2668         if (ret) {
2669                 cma_exch(id_priv, CMA_DESTROYING);
2670                 mutex_unlock(&id_priv->handler_mutex);
2671                 rdma_destroy_id(&id_priv->id);
2672                 return 0;
2673         }
2674
2675         mutex_unlock(&id_priv->handler_mutex);
2676         return 0;
2677 }
2678
2679 static void cma_set_mgid(struct rdma_id_private *id_priv,
2680                          struct sockaddr *addr, union ib_gid *mgid)
2681 {
2682         unsigned char mc_map[MAX_ADDR_LEN];
2683         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2684         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2685         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2686
2687         if (cma_any_addr(addr)) {
2688                 memset(mgid, 0, sizeof *mgid);
2689         } else if ((addr->sa_family == AF_INET6) &&
2690                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2691                                                                  0xFF10A01B)) {
2692                 /* IPv6 address is an SA assigned MGID. */
2693                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2694         } else {
2695                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2696                 if (id_priv->id.ps == RDMA_PS_UDP)
2697                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2698                 *mgid = *(union ib_gid *) (mc_map + 4);
2699         }
2700 }
2701
2702 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2703                                  struct cma_multicast *mc)
2704 {
2705         struct ib_sa_mcmember_rec rec;
2706         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2707         ib_sa_comp_mask comp_mask;
2708         int ret;
2709
2710         ib_addr_get_mgid(dev_addr, &rec.mgid);
2711         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2712                                      &rec.mgid, &rec);
2713         if (ret)
2714                 return ret;
2715
2716         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2717         if (id_priv->id.ps == RDMA_PS_UDP)
2718                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2719         ib_addr_get_sgid(dev_addr, &rec.port_gid);
2720         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2721         rec.join_state = 1;
2722
2723         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2724                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2725                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2726                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2727                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2728
2729         if (id_priv->id.ps == RDMA_PS_IPOIB)
2730                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2731                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2732
2733         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2734                                                 id_priv->id.port_num, &rec,
2735                                                 comp_mask, GFP_KERNEL,
2736                                                 cma_ib_mc_handler, mc);
2737         if (IS_ERR(mc->multicast.ib))
2738                 return PTR_ERR(mc->multicast.ib);
2739
2740         return 0;
2741 }
2742
2743 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2744                         void *context)
2745 {
2746         struct rdma_id_private *id_priv;
2747         struct cma_multicast *mc;
2748         int ret;
2749
2750         id_priv = container_of(id, struct rdma_id_private, id);
2751         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2752             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2753                 return -EINVAL;
2754
2755         mc = kmalloc(sizeof *mc, GFP_KERNEL);
2756         if (!mc)
2757                 return -ENOMEM;
2758
2759         memcpy(&mc->addr, addr, ip_addr_size(addr));
2760         mc->context = context;
2761         mc->id_priv = id_priv;
2762
2763         spin_lock(&id_priv->lock);
2764         list_add(&mc->list, &id_priv->mc_list);
2765         spin_unlock(&id_priv->lock);
2766
2767         switch (rdma_node_get_transport(id->device->node_type)) {
2768         case RDMA_TRANSPORT_IB:
2769                 ret = cma_join_ib_multicast(id_priv, mc);
2770                 break;
2771         default:
2772                 ret = -ENOSYS;
2773                 break;
2774         }
2775
2776         if (ret) {
2777                 spin_lock_irq(&id_priv->lock);
2778                 list_del(&mc->list);
2779                 spin_unlock_irq(&id_priv->lock);
2780                 kfree(mc);
2781         }
2782         return ret;
2783 }
2784 EXPORT_SYMBOL(rdma_join_multicast);
2785
2786 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2787 {
2788         struct rdma_id_private *id_priv;
2789         struct cma_multicast *mc;
2790
2791         id_priv = container_of(id, struct rdma_id_private, id);
2792         spin_lock_irq(&id_priv->lock);
2793         list_for_each_entry(mc, &id_priv->mc_list, list) {
2794                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2795                         list_del(&mc->list);
2796                         spin_unlock_irq(&id_priv->lock);
2797
2798                         if (id->qp)
2799                                 ib_detach_mcast(id->qp,
2800                                                 &mc->multicast.ib->rec.mgid,
2801                                                 mc->multicast.ib->rec.mlid);
2802                         ib_sa_free_multicast(mc->multicast.ib);
2803                         kfree(mc);
2804                         return;
2805                 }
2806         }
2807         spin_unlock_irq(&id_priv->lock);
2808 }
2809 EXPORT_SYMBOL(rdma_leave_multicast);
2810
2811 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
2812 {
2813         struct rdma_dev_addr *dev_addr;
2814         struct cma_ndev_work *work;
2815
2816         dev_addr = &id_priv->id.route.addr.dev_addr;
2817
2818         if ((dev_addr->src_dev == ndev) &&
2819             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
2820                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
2821                        ndev->name, &id_priv->id);
2822                 work = kzalloc(sizeof *work, GFP_KERNEL);
2823                 if (!work)
2824                         return -ENOMEM;
2825
2826                 INIT_WORK(&work->work, cma_ndev_work_handler);
2827                 work->id = id_priv;
2828                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
2829                 atomic_inc(&id_priv->refcount);
2830                 queue_work(cma_wq, &work->work);
2831         }
2832
2833         return 0;
2834 }
2835
2836 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
2837                                void *ctx)
2838 {
2839         struct net_device *ndev = (struct net_device *)ctx;
2840         struct cma_device *cma_dev;
2841         struct rdma_id_private *id_priv;
2842         int ret = NOTIFY_DONE;
2843
2844         if (dev_net(ndev) != &init_net)
2845                 return NOTIFY_DONE;
2846
2847         if (event != NETDEV_BONDING_FAILOVER)
2848                 return NOTIFY_DONE;
2849
2850         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
2851                 return NOTIFY_DONE;
2852
2853         mutex_lock(&lock);
2854         list_for_each_entry(cma_dev, &dev_list, list)
2855                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
2856                         ret = cma_netdev_change(ndev, id_priv);
2857                         if (ret)
2858                                 goto out;
2859                 }
2860
2861 out:
2862         mutex_unlock(&lock);
2863         return ret;
2864 }
2865
2866 static struct notifier_block cma_nb = {
2867         .notifier_call = cma_netdev_callback
2868 };
2869
2870 static void cma_add_one(struct ib_device *device)
2871 {
2872         struct cma_device *cma_dev;
2873         struct rdma_id_private *id_priv;
2874
2875         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2876         if (!cma_dev)
2877                 return;
2878
2879         cma_dev->device = device;
2880
2881         init_completion(&cma_dev->comp);
2882         atomic_set(&cma_dev->refcount, 1);
2883         INIT_LIST_HEAD(&cma_dev->id_list);
2884         ib_set_client_data(device, &cma_client, cma_dev);
2885
2886         mutex_lock(&lock);
2887         list_add_tail(&cma_dev->list, &dev_list);
2888         list_for_each_entry(id_priv, &listen_any_list, list)
2889                 cma_listen_on_dev(id_priv, cma_dev);
2890         mutex_unlock(&lock);
2891 }
2892
2893 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2894 {
2895         struct rdma_cm_event event;
2896         enum cma_state state;
2897         int ret = 0;
2898
2899         /* Record that we want to remove the device */
2900         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2901         if (state == CMA_DESTROYING)
2902                 return 0;
2903
2904         cma_cancel_operation(id_priv, state);
2905         mutex_lock(&id_priv->handler_mutex);
2906
2907         /* Check for destruction from another callback. */
2908         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2909                 goto out;
2910
2911         memset(&event, 0, sizeof event);
2912         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2913         ret = id_priv->id.event_handler(&id_priv->id, &event);
2914 out:
2915         mutex_unlock(&id_priv->handler_mutex);
2916         return ret;
2917 }
2918
2919 static void cma_process_remove(struct cma_device *cma_dev)
2920 {
2921         struct rdma_id_private *id_priv;
2922         int ret;
2923
2924         mutex_lock(&lock);
2925         while (!list_empty(&cma_dev->id_list)) {
2926                 id_priv = list_entry(cma_dev->id_list.next,
2927                                      struct rdma_id_private, list);
2928
2929                 list_del(&id_priv->listen_list);
2930                 list_del_init(&id_priv->list);
2931                 atomic_inc(&id_priv->refcount);
2932                 mutex_unlock(&lock);
2933
2934                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
2935                 cma_deref_id(id_priv);
2936                 if (ret)
2937                         rdma_destroy_id(&id_priv->id);
2938
2939                 mutex_lock(&lock);
2940         }
2941         mutex_unlock(&lock);
2942
2943         cma_deref_dev(cma_dev);
2944         wait_for_completion(&cma_dev->comp);
2945 }
2946
2947 static void cma_remove_one(struct ib_device *device)
2948 {
2949         struct cma_device *cma_dev;
2950
2951         cma_dev = ib_get_client_data(device, &cma_client);
2952         if (!cma_dev)
2953                 return;
2954
2955         mutex_lock(&lock);
2956         list_del(&cma_dev->list);
2957         mutex_unlock(&lock);
2958
2959         cma_process_remove(cma_dev);
2960         kfree(cma_dev);
2961 }
2962
2963 static int __init cma_init(void)
2964 {
2965         int ret, low, high, remaining;
2966
2967         get_random_bytes(&next_port, sizeof next_port);
2968         inet_get_local_port_range(&low, &high);
2969         remaining = (high - low) + 1;
2970         next_port = ((unsigned int) next_port % remaining) + low;
2971
2972         cma_wq = create_singlethread_workqueue("rdma_cm");
2973         if (!cma_wq)
2974                 return -ENOMEM;
2975
2976         ib_sa_register_client(&sa_client);
2977         rdma_addr_register_client(&addr_client);
2978         register_netdevice_notifier(&cma_nb);
2979
2980         ret = ib_register_client(&cma_client);
2981         if (ret)
2982                 goto err;
2983         return 0;
2984
2985 err:
2986         unregister_netdevice_notifier(&cma_nb);
2987         rdma_addr_unregister_client(&addr_client);
2988         ib_sa_unregister_client(&sa_client);
2989         destroy_workqueue(cma_wq);
2990         return ret;
2991 }
2992
2993 static void __exit cma_cleanup(void)
2994 {
2995         ib_unregister_client(&cma_client);
2996         unregister_netdevice_notifier(&cma_nb);
2997         rdma_addr_unregister_client(&addr_client);
2998         ib_sa_unregister_client(&sa_client);
2999         destroy_workqueue(cma_wq);
3000         idr_destroy(&sdp_ps);
3001         idr_destroy(&tcp_ps);
3002         idr_destroy(&udp_ps);
3003         idr_destroy(&ipoib_ps);
3004 }
3005
3006 module_init(cma_init);
3007 module_exit(cma_cleanup);