git.oblomov.eu Git - linux-2.6/blob - 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 licensed under one of the following licenses:
   8  *
   9  * 1) under the terms of the "Common Public License 1.0" a copy of which is
  10  *    available from the Open Source Initiative, see
  11  *    http://www.opensource.org/licenses/cpl.php.
  12  *
  13  * 2) under the terms of the "The BSD License" a copy of which is
  14  *    available from the Open Source Initiative, see
  15  *    http://www.opensource.org/licenses/bsd-license.php.
  16  *
  17  * 3) under the terms of the "GNU General Public License (GPL) Version 2" a
  18  *    copy of which is available from the Open Source Initiative, see
  19  *    http://www.opensource.org/licenses/gpl-license.php.
  20  *
  21  * Licensee has the right to choose one of the above licenses.
  22  *
  23  * Redistributions of source code must retain the above copyright
  24  * notice and one of the license notices.
  25  *
  26  * Redistributions in binary form must reproduce both the above copyright
  27  * notice, one of the license notices in the documentation
  28  * and/or other materials provided with the distribution.
  29  *
  30  */
  31
  32 #include <linux/completion.h>
  33 #include <linux/in.h>
  34 #include <linux/in6.h>
  35 #include <linux/mutex.h>
  36 #include <linux/random.h>
  37 #include <linux/idr.h>
  38 #include <linux/inetdevice.h>
  39
  40 #include <net/tcp.h>
  41
  42 #include <rdma/rdma_cm.h>
  43 #include <rdma/rdma_cm_ib.h>
  44 #include <rdma/ib_cache.h>
  45 #include <rdma/ib_cm.h>
  46 #include <rdma/ib_sa.h>
  47 #include <rdma/iw_cm.h>
  48
  49 MODULE_AUTHOR("Sean Hefty");
  50 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  51 MODULE_LICENSE("Dual BSD/GPL");
  52
  53 #define CMA_CM_RESPONSE_TIMEOUT 20
  54 #define CMA_MAX_CM_RETRIES 15
  55
  56 static void cma_add_one(struct ib_device *device);
  57 static void cma_remove_one(struct ib_device *device);
  58
  59 static struct ib_client cma_client = {
  60         .name   = "cma",
  61         .add    = cma_add_one,
  62         .remove = cma_remove_one
  63 };
  64
  65 static LIST_HEAD(dev_list);
  66 static LIST_HEAD(listen_any_list);
  67 static DEFINE_MUTEX(lock);
  68 static struct workqueue_struct *cma_wq;
  69 static DEFINE_IDR(sdp_ps);
  70 static DEFINE_IDR(tcp_ps);
  71
  72 struct cma_device {
  73         struct list_head        list;
  74         struct ib_device        *device;
  75         __be64                  node_guid;
  76         struct completion       comp;
  77         atomic_t                refcount;
  78         struct list_head        id_list;
  79 };
  80
  81 enum cma_state {
  82         CMA_IDLE,
  83         CMA_ADDR_QUERY,
  84         CMA_ADDR_RESOLVED,
  85         CMA_ROUTE_QUERY,
  86         CMA_ROUTE_RESOLVED,
  87         CMA_CONNECT,
  88         CMA_DISCONNECT,
  89         CMA_ADDR_BOUND,
  90         CMA_LISTEN,
  91         CMA_DEVICE_REMOVAL,
  92         CMA_DESTROYING
  93 };
  94
  95 struct rdma_bind_list {
  96         struct idr              *ps;
  97         struct hlist_head       owners;
  98         unsigned short          port;
  99 };
 100
 101 /*
 102  * Device removal can occur at anytime, so we need extra handling to
 103  * serialize notifying the user of device removal with other callbacks.
 104  * We do this by disabling removal notification while a callback is in process,
 105  * and reporting it after the callback completes.
 106  */
 107 struct rdma_id_private {
 108         struct rdma_cm_id       id;
 109
 110         struct rdma_bind_list   *bind_list;
 111         struct hlist_node       node;
 112         struct list_head        list;
 113         struct list_head        listen_list;
 114         struct cma_device       *cma_dev;
 115
 116         enum cma_state          state;
 117         spinlock_t              lock;
 118         struct completion       comp;
 119         atomic_t                refcount;
 120         wait_queue_head_t       wait_remove;
 121         atomic_t                dev_remove;
 122
 123         int                     backlog;
 124         int                     timeout_ms;
 125         struct ib_sa_query      *query;
 126         int                     query_id;
 127         union {
 128                 struct ib_cm_id *ib;
 129                 struct iw_cm_id *iw;
 130         } cm_id;
 131
 132         u32                     seq_num;
 133         u32                     qp_num;
 134         enum ib_qp_type         qp_type;
 135         u8                      srq;
 136 };
 137
 138 struct cma_work {
 139         struct work_struct      work;
 140         struct rdma_id_private  *id;
 141         enum cma_state          old_state;
 142         enum cma_state          new_state;
 143         struct rdma_cm_event    event;
 144 };
 145
 146 union cma_ip_addr {
 147         struct in6_addr ip6;
 148         struct {
 149                 __u32 pad[3];
 150                 __u32 addr;
 151         } ip4;
 152 };
 153
 154 struct cma_hdr {
 155         u8 cma_version;
 156         u8 ip_version;  /* IP version: 7:4 */
 157         __u16 port;
 158         union cma_ip_addr src_addr;
 159         union cma_ip_addr dst_addr;
 160 };
 161
 162 struct sdp_hh {
 163         u8 bsdh[16];
 164         u8 sdp_version; /* Major version: 7:4 */
 165         u8 ip_version;  /* IP version: 7:4 */
 166         u8 sdp_specific1[10];
 167         __u16 port;
 168         __u16 sdp_specific2;
 169         union cma_ip_addr src_addr;
 170         union cma_ip_addr dst_addr;
 171 };
 172
 173 struct sdp_hah {
 174         u8 bsdh[16];
 175         u8 sdp_version;
 176 };
 177
 178 #define CMA_VERSION 0x00
 179 #define SDP_MAJ_VERSION 0x2
 180
 181 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 182 {
 183         unsigned long flags;
 184         int ret;
 185
 186         spin_lock_irqsave(&id_priv->lock, flags);
 187         ret = (id_priv->state == comp);
 188         spin_unlock_irqrestore(&id_priv->lock, flags);
 189         return ret;
 190 }
 191
 192 static int cma_comp_exch(struct rdma_id_private *id_priv,
 193                          enum cma_state comp, enum cma_state exch)
 194 {
 195         unsigned long flags;
 196         int ret;
 197
 198         spin_lock_irqsave(&id_priv->lock, flags);
 199         if ((ret = (id_priv->state == comp)))
 200                 id_priv->state = exch;
 201         spin_unlock_irqrestore(&id_priv->lock, flags);
 202         return ret;
 203 }
 204
 205 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 206                                enum cma_state exch)
 207 {
 208         unsigned long flags;
 209         enum cma_state old;
 210
 211         spin_lock_irqsave(&id_priv->lock, flags);
 212         old = id_priv->state;
 213         id_priv->state = exch;
 214         spin_unlock_irqrestore(&id_priv->lock, flags);
 215         return old;
 216 }
 217
 218 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 219 {
 220         return hdr->ip_version >> 4;
 221 }
 222
 223 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 224 {
 225         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 226 }
 227
 228 static inline u8 sdp_get_majv(u8 sdp_version)
 229 {
 230         return sdp_version >> 4;
 231 }
 232
 233 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 234 {
 235         return hh->ip_version >> 4;
 236 }
 237
 238 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 239 {
 240         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 241 }
 242
 243 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 244                               struct cma_device *cma_dev)
 245 {
 246         atomic_inc(&cma_dev->refcount);
 247         id_priv->cma_dev = cma_dev;
 248         id_priv->id.device = cma_dev->device;
 249         list_add_tail(&id_priv->list, &cma_dev->id_list);
 250 }
 251
 252 static inline void cma_deref_dev(struct cma_device *cma_dev)
 253 {
 254         if (atomic_dec_and_test(&cma_dev->refcount))
 255                 complete(&cma_dev->comp);
 256 }
 257
 258 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 259 {
 260         list_del(&id_priv->list);
 261         cma_deref_dev(id_priv->cma_dev);
 262         id_priv->cma_dev = NULL;
 263 }
 264
 265 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 266 {
 267         enum rdma_node_type dev_type = id_priv->id.route.addr.dev_addr.dev_type;
 268         struct cma_device *cma_dev;
 269         union ib_gid gid;
 270         int ret = -ENODEV;
 271
 272         switch (rdma_node_get_transport(dev_type)) {
 273         case RDMA_TRANSPORT_IB:
 274                 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
 275                 break;
 276         case RDMA_TRANSPORT_IWARP:
 277                 iw_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
 278                 break;
 279         default:
 280                 return -ENODEV;
 281         }
 282         mutex_lock(&lock);
 283         list_for_each_entry(cma_dev, &dev_list, list) {
 284                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 285                                          &id_priv->id.port_num, NULL);
 286                 if (!ret) {
 287                         cma_attach_to_dev(id_priv, cma_dev);
 288                         break;
 289                 }
 290         }
 291         mutex_unlock(&lock);
 292         return ret;
 293 }
 294
 295 static void cma_deref_id(struct rdma_id_private *id_priv)
 296 {
 297         if (atomic_dec_and_test(&id_priv->refcount))
 298                 complete(&id_priv->comp);
 299 }
 300
 301 static void cma_release_remove(struct rdma_id_private *id_priv)
 302 {
 303         if (atomic_dec_and_test(&id_priv->dev_remove))
 304                 wake_up(&id_priv->wait_remove);
 305 }
 306
 307 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 308                                   void *context, enum rdma_port_space ps)
 309 {
 310         struct rdma_id_private *id_priv;
 311
 312         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 313         if (!id_priv)
 314                 return ERR_PTR(-ENOMEM);
 315
 316         id_priv->state = CMA_IDLE;
 317         id_priv->id.context = context;
 318         id_priv->id.event_handler = event_handler;
 319         id_priv->id.ps = ps;
 320         spin_lock_init(&id_priv->lock);
 321         init_completion(&id_priv->comp);
 322         atomic_set(&id_priv->refcount, 1);
 323         init_waitqueue_head(&id_priv->wait_remove);
 324         atomic_set(&id_priv->dev_remove, 0);
 325         INIT_LIST_HEAD(&id_priv->listen_list);
 326         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 327
 328         return &id_priv->id;
 329 }
 330 EXPORT_SYMBOL(rdma_create_id);
 331
 332 static int cma_init_ib_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 333 {
 334         struct ib_qp_attr qp_attr;
 335         struct rdma_dev_addr *dev_addr;
 336         int ret;
 337
 338         dev_addr = &id_priv->id.route.addr.dev_addr;
 339         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 340                                   ib_addr_get_pkey(dev_addr),
 341                                   &qp_attr.pkey_index);
 342         if (ret)
 343                 return ret;
 344
 345         qp_attr.qp_state = IB_QPS_INIT;
 346         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 347         qp_attr.port_num = id_priv->id.port_num;
 348         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS |
 349                                           IB_QP_PKEY_INDEX | IB_QP_PORT);
 350 }
 351
 352 static int cma_init_iw_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 353 {
 354         struct ib_qp_attr qp_attr;
 355
 356         qp_attr.qp_state = IB_QPS_INIT;
 357         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 358
 359         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS);
 360 }
 361
 362 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 363                    struct ib_qp_init_attr *qp_init_attr)
 364 {
 365         struct rdma_id_private *id_priv;
 366         struct ib_qp *qp;
 367         int ret;
 368
 369         id_priv = container_of(id, struct rdma_id_private, id);
 370         if (id->device != pd->device)
 371                 return -EINVAL;
 372
 373         qp = ib_create_qp(pd, qp_init_attr);
 374         if (IS_ERR(qp))
 375                 return PTR_ERR(qp);
 376
 377         switch (rdma_node_get_transport(id->device->node_type)) {
 378         case RDMA_TRANSPORT_IB:
 379                 ret = cma_init_ib_qp(id_priv, qp);
 380                 break;
 381         case RDMA_TRANSPORT_IWARP:
 382                 ret = cma_init_iw_qp(id_priv, qp);
 383                 break;
 384         default:
 385                 ret = -ENOSYS;
 386                 break;
 387         }
 388
 389         if (ret)
 390                 goto err;
 391
 392         id->qp = qp;
 393         id_priv->qp_num = qp->qp_num;
 394         id_priv->qp_type = qp->qp_type;
 395         id_priv->srq = (qp->srq != NULL);
 396         return 0;
 397 err:
 398         ib_destroy_qp(qp);
 399         return ret;
 400 }
 401 EXPORT_SYMBOL(rdma_create_qp);
 402
 403 void rdma_destroy_qp(struct rdma_cm_id *id)
 404 {
 405         ib_destroy_qp(id->qp);
 406 }
 407 EXPORT_SYMBOL(rdma_destroy_qp);
 408
 409 static int cma_modify_qp_rtr(struct rdma_cm_id *id)
 410 {
 411         struct ib_qp_attr qp_attr;
 412         int qp_attr_mask, ret;
 413
 414         if (!id->qp)
 415                 return 0;
 416
 417         /* Need to update QP attributes from default values. */
 418         qp_attr.qp_state = IB_QPS_INIT;
 419         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 420         if (ret)
 421                 return ret;
 422
 423         ret = ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 424         if (ret)
 425                 return ret;
 426
 427         qp_attr.qp_state = IB_QPS_RTR;
 428         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 429         if (ret)
 430                 return ret;
 431
 432         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 433 }
 434
 435 static int cma_modify_qp_rts(struct rdma_cm_id *id)
 436 {
 437         struct ib_qp_attr qp_attr;
 438         int qp_attr_mask, ret;
 439
 440         if (!id->qp)
 441                 return 0;
 442
 443         qp_attr.qp_state = IB_QPS_RTS;
 444         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 445         if (ret)
 446                 return ret;
 447
 448         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 449 }
 450
 451 static int cma_modify_qp_err(struct rdma_cm_id *id)
 452 {
 453         struct ib_qp_attr qp_attr;
 454
 455         if (!id->qp)
 456                 return 0;
 457
 458         qp_attr.qp_state = IB_QPS_ERR;
 459         return ib_modify_qp(id->qp, &qp_attr, IB_QP_STATE);
 460 }
 461
 462 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 463                        int *qp_attr_mask)
 464 {
 465         struct rdma_id_private *id_priv;
 466         int ret;
 467
 468         id_priv = container_of(id, struct rdma_id_private, id);
 469         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 470         case RDMA_TRANSPORT_IB:
 471                 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 472                                          qp_attr_mask);
 473                 if (qp_attr->qp_state == IB_QPS_RTR)
 474                         qp_attr->rq_psn = id_priv->seq_num;
 475                 break;
 476         case RDMA_TRANSPORT_IWARP:
 477                 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 478                                         qp_attr_mask);
 479                 break;
 480         default:
 481                 ret = -ENOSYS;
 482                 break;
 483         }
 484
 485         return ret;
 486 }
 487 EXPORT_SYMBOL(rdma_init_qp_attr);
 488
 489 static inline int cma_zero_addr(struct sockaddr *addr)
 490 {
 491         struct in6_addr *ip6;
 492
 493         if (addr->sa_family == AF_INET)
 494                 return ZERONET(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 495         else {
 496                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 497                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 498                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 499         }
 500 }
 501
 502 static inline int cma_loopback_addr(struct sockaddr *addr)
 503 {
 504         return LOOPBACK(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 505 }
 506
 507 static inline int cma_any_addr(struct sockaddr *addr)
 508 {
 509         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 510 }
 511
 512 static inline int cma_any_port(struct sockaddr *addr)
 513 {
 514         return !((struct sockaddr_in *) addr)->sin_port;
 515 }
 516
 517 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 518                             u8 *ip_ver, __u16 *port,
 519                             union cma_ip_addr **src, union cma_ip_addr **dst)
 520 {
 521         switch (ps) {
 522         case RDMA_PS_SDP:
 523                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 524                     SDP_MAJ_VERSION)
 525                         return -EINVAL;
 526
 527                 *ip_ver = sdp_get_ip_ver(hdr);
 528                 *port   = ((struct sdp_hh *) hdr)->port;
 529                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 530                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 531                 break;
 532         default:
 533                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 534                         return -EINVAL;
 535
 536                 *ip_ver = cma_get_ip_ver(hdr);
 537                 *port   = ((struct cma_hdr *) hdr)->port;
 538                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 539                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 540                 break;
 541         }
 542
 543         if (*ip_ver != 4 && *ip_ver != 6)
 544                 return -EINVAL;
 545         return 0;
 546 }
 547
 548 static void cma_save_net_info(struct rdma_addr *addr,
 549                               struct rdma_addr *listen_addr,
 550                               u8 ip_ver, __u16 port,
 551                               union cma_ip_addr *src, union cma_ip_addr *dst)
 552 {
 553         struct sockaddr_in *listen4, *ip4;
 554         struct sockaddr_in6 *listen6, *ip6;
 555
 556         switch (ip_ver) {
 557         case 4:
 558                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 559                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 560                 ip4->sin_family = listen4->sin_family;
 561                 ip4->sin_addr.s_addr = dst->ip4.addr;
 562                 ip4->sin_port = listen4->sin_port;
 563
 564                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 565                 ip4->sin_family = listen4->sin_family;
 566                 ip4->sin_addr.s_addr = src->ip4.addr;
 567                 ip4->sin_port = port;
 568                 break;
 569         case 6:
 570                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 571                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 572                 ip6->sin6_family = listen6->sin6_family;
 573                 ip6->sin6_addr = dst->ip6;
 574                 ip6->sin6_port = listen6->sin6_port;
 575
 576                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 577                 ip6->sin6_family = listen6->sin6_family;
 578                 ip6->sin6_addr = src->ip6;
 579                 ip6->sin6_port = port;
 580                 break;
 581         default:
 582                 break;
 583         }
 584 }
 585
 586 static inline int cma_user_data_offset(enum rdma_port_space ps)
 587 {
 588         switch (ps) {
 589         case RDMA_PS_SDP:
 590                 return 0;
 591         default:
 592                 return sizeof(struct cma_hdr);
 593         }
 594 }
 595
 596 static int cma_notify_user(struct rdma_id_private *id_priv,
 597                            enum rdma_cm_event_type type, int status,
 598                            void *data, u8 data_len)
 599 {
 600         struct rdma_cm_event event;
 601
 602         event.event = type;
 603         event.status = status;
 604         event.private_data = data;
 605         event.private_data_len = data_len;
 606
 607         return id_priv->id.event_handler(&id_priv->id, &event);
 608 }
 609
 610 static void cma_cancel_route(struct rdma_id_private *id_priv)
 611 {
 612         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 613         case RDMA_TRANSPORT_IB:
 614                 if (id_priv->query)
 615                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 616                 break;
 617         default:
 618                 break;
 619         }
 620 }
 621
 622 static inline int cma_internal_listen(struct rdma_id_private *id_priv)
 623 {
 624         return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
 625                cma_any_addr(&id_priv->id.route.addr.src_addr);
 626 }
 627
 628 static void cma_destroy_listen(struct rdma_id_private *id_priv)
 629 {
 630         cma_exch(id_priv, CMA_DESTROYING);
 631
 632         if (id_priv->cma_dev) {
 633                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 634                 case RDMA_TRANSPORT_IB:
 635                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 636                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 637                         break;
 638                 case RDMA_TRANSPORT_IWARP:
 639                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 640                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 641                         break;
 642                 default:
 643                         break;
 644                 }
 645                 cma_detach_from_dev(id_priv);
 646         }
 647         list_del(&id_priv->listen_list);
 648
 649         cma_deref_id(id_priv);
 650         wait_for_completion(&id_priv->comp);
 651
 652         kfree(id_priv);
 653 }
 654
 655 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 656 {
 657         struct rdma_id_private *dev_id_priv;
 658
 659         mutex_lock(&lock);
 660         list_del(&id_priv->list);
 661
 662         while (!list_empty(&id_priv->listen_list)) {
 663                 dev_id_priv = list_entry(id_priv->listen_list.next,
 664                                          struct rdma_id_private, listen_list);
 665                 cma_destroy_listen(dev_id_priv);
 666         }
 667         mutex_unlock(&lock);
 668 }
 669
 670 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 671                                  enum cma_state state)
 672 {
 673         switch (state) {
 674         case CMA_ADDR_QUERY:
 675                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 676                 break;
 677         case CMA_ROUTE_QUERY:
 678                 cma_cancel_route(id_priv);
 679                 break;
 680         case CMA_LISTEN:
 681                 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
 682                     !id_priv->cma_dev)
 683                         cma_cancel_listens(id_priv);
 684                 break;
 685         default:
 686                 break;
 687         }
 688 }
 689
 690 static void cma_release_port(struct rdma_id_private *id_priv)
 691 {
 692         struct rdma_bind_list *bind_list = id_priv->bind_list;
 693
 694         if (!bind_list)
 695                 return;
 696
 697         mutex_lock(&lock);
 698         hlist_del(&id_priv->node);
 699         if (hlist_empty(&bind_list->owners)) {
 700                 idr_remove(bind_list->ps, bind_list->port);
 701                 kfree(bind_list);
 702         }
 703         mutex_unlock(&lock);
 704 }
 705
 706 void rdma_destroy_id(struct rdma_cm_id *id)
 707 {
 708         struct rdma_id_private *id_priv;
 709         enum cma_state state;
 710
 711         id_priv = container_of(id, struct rdma_id_private, id);
 712         state = cma_exch(id_priv, CMA_DESTROYING);
 713         cma_cancel_operation(id_priv, state);
 714
 715         if (id_priv->cma_dev) {
 716                 switch (rdma_node_get_transport(id->device->node_type)) {
 717                 case RDMA_TRANSPORT_IB:
 718                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 719                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 720                         break;
 721                 case RDMA_TRANSPORT_IWARP:
 722                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 723                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 724                         break;
 725                 default:
 726                         break;
 727                 }
 728                 mutex_lock(&lock);
 729                 cma_detach_from_dev(id_priv);
 730                 mutex_unlock(&lock);
 731         }
 732
 733         cma_release_port(id_priv);
 734         cma_deref_id(id_priv);
 735         wait_for_completion(&id_priv->comp);
 736
 737         kfree(id_priv->id.route.path_rec);
 738         kfree(id_priv);
 739 }
 740 EXPORT_SYMBOL(rdma_destroy_id);
 741
 742 static int cma_rep_recv(struct rdma_id_private *id_priv)
 743 {
 744         int ret;
 745
 746         ret = cma_modify_qp_rtr(&id_priv->id);
 747         if (ret)
 748                 goto reject;
 749
 750         ret = cma_modify_qp_rts(&id_priv->id);
 751         if (ret)
 752                 goto reject;
 753
 754         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 755         if (ret)
 756                 goto reject;
 757
 758         return 0;
 759 reject:
 760         cma_modify_qp_err(&id_priv->id);
 761         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 762                        NULL, 0, NULL, 0);
 763         return ret;
 764 }
 765
 766 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 767 {
 768         if (id_priv->id.ps == RDMA_PS_SDP &&
 769             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 770             SDP_MAJ_VERSION)
 771                 return -EINVAL;
 772
 773         return 0;
 774 }
 775
 776 static int cma_rtu_recv(struct rdma_id_private *id_priv)
 777 {
 778         int ret;
 779
 780         ret = cma_modify_qp_rts(&id_priv->id);
 781         if (ret)
 782                 goto reject;
 783
 784         return 0;
 785 reject:
 786         cma_modify_qp_err(&id_priv->id);
 787         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 788                        NULL, 0, NULL, 0);
 789         return ret;
 790 }
 791
 792 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 793 {
 794         struct rdma_id_private *id_priv = cm_id->context;
 795         enum rdma_cm_event_type event;
 796         u8 private_data_len = 0;
 797         int ret = 0, status = 0;
 798
 799         atomic_inc(&id_priv->dev_remove);
 800         if (!cma_comp(id_priv, CMA_CONNECT))
 801                 goto out;
 802
 803         switch (ib_event->event) {
 804         case IB_CM_REQ_ERROR:
 805         case IB_CM_REP_ERROR:
 806                 event = RDMA_CM_EVENT_UNREACHABLE;
 807                 status = -ETIMEDOUT;
 808                 break;
 809         case IB_CM_REP_RECEIVED:
 810                 status = cma_verify_rep(id_priv, ib_event->private_data);
 811                 if (status)
 812                         event = RDMA_CM_EVENT_CONNECT_ERROR;
 813                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 814                         status = cma_rep_recv(id_priv);
 815                         event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
 816                                          RDMA_CM_EVENT_ESTABLISHED;
 817                 } else
 818                         event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 819                 private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 820                 break;
 821         case IB_CM_RTU_RECEIVED:
 822                 status = cma_rtu_recv(id_priv);
 823                 event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
 824                                  RDMA_CM_EVENT_ESTABLISHED;
 825                 break;
 826         case IB_CM_DREQ_ERROR:
 827                 status = -ETIMEDOUT; /* fall through */
 828         case IB_CM_DREQ_RECEIVED:
 829         case IB_CM_DREP_RECEIVED:
 830                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 831                         goto out;
 832                 event = RDMA_CM_EVENT_DISCONNECTED;
 833                 break;
 834         case IB_CM_TIMEWAIT_EXIT:
 835         case IB_CM_MRA_RECEIVED:
 836                 /* ignore event */
 837                 goto out;
 838         case IB_CM_REJ_RECEIVED:
 839                 cma_modify_qp_err(&id_priv->id);
 840                 status = ib_event->param.rej_rcvd.reason;
 841                 event = RDMA_CM_EVENT_REJECTED;
 842                 private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 843                 break;
 844         default:
 845                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
 846                        ib_event->event);
 847                 goto out;
 848         }
 849
 850         ret = cma_notify_user(id_priv, event, status, ib_event->private_data,
 851                               private_data_len);
 852         if (ret) {
 853                 /* Destroy the CM ID by returning a non-zero value. */
 854                 id_priv->cm_id.ib = NULL;
 855                 cma_exch(id_priv, CMA_DESTROYING);
 856                 cma_release_remove(id_priv);
 857                 rdma_destroy_id(&id_priv->id);
 858                 return ret;
 859         }
 860 out:
 861         cma_release_remove(id_priv);
 862         return ret;
 863 }
 864
 865 static struct rdma_id_private *cma_new_id(struct rdma_cm_id *listen_id,
 866                                           struct ib_cm_event *ib_event)
 867 {
 868         struct rdma_id_private *id_priv;
 869         struct rdma_cm_id *id;
 870         struct rdma_route *rt;
 871         union cma_ip_addr *src, *dst;
 872         __u16 port;
 873         u8 ip_ver;
 874
 875         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 876                             listen_id->ps);
 877         if (IS_ERR(id))
 878                 return NULL;
 879
 880         rt = &id->route;
 881         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
 882         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, GFP_KERNEL);
 883         if (!rt->path_rec)
 884                 goto err;
 885
 886         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 887                              &ip_ver, &port, &src, &dst))
 888                 goto err;
 889
 890         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 891                           ip_ver, port, src, dst);
 892         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
 893         if (rt->num_paths == 2)
 894                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
 895
 896         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
 897         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
 898         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
 899         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
 900
 901         id_priv = container_of(id, struct rdma_id_private, id);
 902         id_priv->state = CMA_CONNECT;
 903         return id_priv;
 904 err:
 905         rdma_destroy_id(id);
 906         return NULL;
 907 }
 908
 909 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 910 {
 911         struct rdma_id_private *listen_id, *conn_id;
 912         int offset, ret;
 913
 914         listen_id = cm_id->context;
 915         atomic_inc(&listen_id->dev_remove);
 916         if (!cma_comp(listen_id, CMA_LISTEN)) {
 917                 ret = -ECONNABORTED;
 918                 goto out;
 919         }
 920
 921         conn_id = cma_new_id(&listen_id->id, ib_event);
 922         if (!conn_id) {
 923                 ret = -ENOMEM;
 924                 goto out;
 925         }
 926
 927         atomic_inc(&conn_id->dev_remove);
 928         ret = cma_acquire_dev(conn_id);
 929         if (ret) {
 930                 ret = -ENODEV;
 931                 cma_release_remove(conn_id);
 932                 rdma_destroy_id(&conn_id->id);
 933                 goto out;
 934         }
 935
 936         conn_id->cm_id.ib = cm_id;
 937         cm_id->context = conn_id;
 938         cm_id->cm_handler = cma_ib_handler;
 939
 940         offset = cma_user_data_offset(listen_id->id.ps);
 941         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
 942                               ib_event->private_data + offset,
 943                               IB_CM_REQ_PRIVATE_DATA_SIZE - offset);
 944         if (ret) {
 945                 /* Destroy the CM ID by returning a non-zero value. */
 946                 conn_id->cm_id.ib = NULL;
 947                 cma_exch(conn_id, CMA_DESTROYING);
 948                 cma_release_remove(conn_id);
 949                 rdma_destroy_id(&conn_id->id);
 950         }
 951 out:
 952         cma_release_remove(listen_id);
 953         return ret;
 954 }
 955
 956 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
 957 {
 958         return cpu_to_be64(((u64)ps << 16) +
 959                be16_to_cpu(((struct sockaddr_in *) addr)->sin_port));
 960 }
 961
 962 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
 963                                  struct ib_cm_compare_data *compare)
 964 {
 965         struct cma_hdr *cma_data, *cma_mask;
 966         struct sdp_hh *sdp_data, *sdp_mask;
 967         __u32 ip4_addr;
 968         struct in6_addr ip6_addr;
 969
 970         memset(compare, 0, sizeof *compare);
 971         cma_data = (void *) compare->data;
 972         cma_mask = (void *) compare->mask;
 973         sdp_data = (void *) compare->data;
 974         sdp_mask = (void *) compare->mask;
 975
 976         switch (addr->sa_family) {
 977         case AF_INET:
 978                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
 979                 if (ps == RDMA_PS_SDP) {
 980                         sdp_set_ip_ver(sdp_data, 4);
 981                         sdp_set_ip_ver(sdp_mask, 0xF);
 982                         sdp_data->dst_addr.ip4.addr = ip4_addr;
 983                         sdp_mask->dst_addr.ip4.addr = ~0;
 984                 } else {
 985                         cma_set_ip_ver(cma_data, 4);
 986                         cma_set_ip_ver(cma_mask, 0xF);
 987                         cma_data->dst_addr.ip4.addr = ip4_addr;
 988                         cma_mask->dst_addr.ip4.addr = ~0;
 989                 }
 990                 break;
 991         case AF_INET6:
 992                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
 993                 if (ps == RDMA_PS_SDP) {
 994                         sdp_set_ip_ver(sdp_data, 6);
 995                         sdp_set_ip_ver(sdp_mask, 0xF);
 996                         sdp_data->dst_addr.ip6 = ip6_addr;
 997                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
 998                                sizeof sdp_mask->dst_addr.ip6);
 999                 } else {
1000                         cma_set_ip_ver(cma_data, 6);
1001                         cma_set_ip_ver(cma_mask, 0xF);
1002                         cma_data->dst_addr.ip6 = ip6_addr;
1003                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1004                                sizeof cma_mask->dst_addr.ip6);
1005                 }
1006                 break;
1007         default:
1008                 break;
1009         }
1010 }
1011
1012 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1013 {
1014         struct rdma_id_private *id_priv = iw_id->context;
1015         enum rdma_cm_event_type event = 0;
1016         struct sockaddr_in *sin;
1017         int ret = 0;
1018
1019         atomic_inc(&id_priv->dev_remove);
1020
1021         switch (iw_event->event) {
1022         case IW_CM_EVENT_CLOSE:
1023                 event = RDMA_CM_EVENT_DISCONNECTED;
1024                 break;
1025         case IW_CM_EVENT_CONNECT_REPLY:
1026                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1027                 *sin = iw_event->local_addr;
1028                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1029                 *sin = iw_event->remote_addr;
1030                 if (iw_event->status)
1031                         event = RDMA_CM_EVENT_REJECTED;
1032                 else
1033                         event = RDMA_CM_EVENT_ESTABLISHED;
1034                 break;
1035         case IW_CM_EVENT_ESTABLISHED:
1036                 event = RDMA_CM_EVENT_ESTABLISHED;
1037                 break;
1038         default:
1039                 BUG_ON(1);
1040         }
1041
1042         ret = cma_notify_user(id_priv, event, iw_event->status,
1043                               iw_event->private_data,
1044                               iw_event->private_data_len);
1045         if (ret) {
1046                 /* Destroy the CM ID by returning a non-zero value. */
1047                 id_priv->cm_id.iw = NULL;
1048                 cma_exch(id_priv, CMA_DESTROYING);
1049                 cma_release_remove(id_priv);
1050                 rdma_destroy_id(&id_priv->id);
1051                 return ret;
1052         }
1053
1054         cma_release_remove(id_priv);
1055         return ret;
1056 }
1057
1058 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1059                                struct iw_cm_event *iw_event)
1060 {
1061         struct rdma_cm_id *new_cm_id;
1062         struct rdma_id_private *listen_id, *conn_id;
1063         struct sockaddr_in *sin;
1064         struct net_device *dev = NULL;
1065         int ret;
1066
1067         listen_id = cm_id->context;
1068         atomic_inc(&listen_id->dev_remove);
1069         if (!cma_comp(listen_id, CMA_LISTEN)) {
1070                 ret = -ECONNABORTED;
1071                 goto out;
1072         }
1073
1074         /* Create a new RDMA id for the new IW CM ID */
1075         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1076                                    listen_id->id.context,
1077                                    RDMA_PS_TCP);
1078         if (!new_cm_id) {
1079                 ret = -ENOMEM;
1080                 goto out;
1081         }
1082         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1083         atomic_inc(&conn_id->dev_remove);
1084         conn_id->state = CMA_CONNECT;
1085
1086         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1087         if (!dev) {
1088                 ret = -EADDRNOTAVAIL;
1089                 cma_release_remove(conn_id);
1090                 rdma_destroy_id(new_cm_id);
1091                 goto out;
1092         }
1093         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1094         if (ret) {
1095                 cma_release_remove(conn_id);
1096                 rdma_destroy_id(new_cm_id);
1097                 goto out;
1098         }
1099
1100         ret = cma_acquire_dev(conn_id);
1101         if (ret) {
1102                 cma_release_remove(conn_id);
1103                 rdma_destroy_id(new_cm_id);
1104                 goto out;
1105         }
1106
1107         conn_id->cm_id.iw = cm_id;
1108         cm_id->context = conn_id;
1109         cm_id->cm_handler = cma_iw_handler;
1110
1111         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1112         *sin = iw_event->local_addr;
1113         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1114         *sin = iw_event->remote_addr;
1115
1116         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
1117                               iw_event->private_data,
1118                               iw_event->private_data_len);
1119         if (ret) {
1120                 /* User wants to destroy the CM ID */
1121                 conn_id->cm_id.iw = NULL;
1122                 cma_exch(conn_id, CMA_DESTROYING);
1123                 cma_release_remove(conn_id);
1124                 rdma_destroy_id(&conn_id->id);
1125         }
1126
1127 out:
1128         if (dev)
1129                 dev_put(dev);
1130         cma_release_remove(listen_id);
1131         return ret;
1132 }
1133
1134 static int cma_ib_listen(struct rdma_id_private *id_priv)
1135 {
1136         struct ib_cm_compare_data compare_data;
1137         struct sockaddr *addr;
1138         __be64 svc_id;
1139         int ret;
1140
1141         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1142                                             id_priv);
1143         if (IS_ERR(id_priv->cm_id.ib))
1144                 return PTR_ERR(id_priv->cm_id.ib);
1145
1146         addr = &id_priv->id.route.addr.src_addr;
1147         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1148         if (cma_any_addr(addr))
1149                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1150         else {
1151                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1152                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1153         }
1154
1155         if (ret) {
1156                 ib_destroy_cm_id(id_priv->cm_id.ib);
1157                 id_priv->cm_id.ib = NULL;
1158         }
1159
1160         return ret;
1161 }
1162
1163 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1164 {
1165         int ret;
1166         struct sockaddr_in *sin;
1167
1168         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1169                                             iw_conn_req_handler,
1170                                             id_priv);
1171         if (IS_ERR(id_priv->cm_id.iw))
1172                 return PTR_ERR(id_priv->cm_id.iw);
1173
1174         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1175         id_priv->cm_id.iw->local_addr = *sin;
1176
1177         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1178
1179         if (ret) {
1180                 iw_destroy_cm_id(id_priv->cm_id.iw);
1181                 id_priv->cm_id.iw = NULL;
1182         }
1183
1184         return ret;
1185 }
1186
1187 static int cma_listen_handler(struct rdma_cm_id *id,
1188                               struct rdma_cm_event *event)
1189 {
1190         struct rdma_id_private *id_priv = id->context;
1191
1192         id->context = id_priv->id.context;
1193         id->event_handler = id_priv->id.event_handler;
1194         return id_priv->id.event_handler(id, event);
1195 }
1196
1197 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1198                               struct cma_device *cma_dev)
1199 {
1200         struct rdma_id_private *dev_id_priv;
1201         struct rdma_cm_id *id;
1202         int ret;
1203
1204         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1205         if (IS_ERR(id))
1206                 return;
1207
1208         dev_id_priv = container_of(id, struct rdma_id_private, id);
1209
1210         dev_id_priv->state = CMA_ADDR_BOUND;
1211         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1212                ip_addr_size(&id_priv->id.route.addr.src_addr));
1213
1214         cma_attach_to_dev(dev_id_priv, cma_dev);
1215         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1216
1217         ret = rdma_listen(id, id_priv->backlog);
1218         if (ret)
1219                 goto err;
1220
1221         return;
1222 err:
1223         cma_destroy_listen(dev_id_priv);
1224 }
1225
1226 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1227 {
1228         struct cma_device *cma_dev;
1229
1230         mutex_lock(&lock);
1231         list_add_tail(&id_priv->list, &listen_any_list);
1232         list_for_each_entry(cma_dev, &dev_list, list)
1233                 cma_listen_on_dev(id_priv, cma_dev);
1234         mutex_unlock(&lock);
1235 }
1236
1237 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1238 {
1239         struct sockaddr_in addr_in;
1240
1241         memset(&addr_in, 0, sizeof addr_in);
1242         addr_in.sin_family = af;
1243         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1244 }
1245
1246 int rdma_listen(struct rdma_cm_id *id, int backlog)
1247 {
1248         struct rdma_id_private *id_priv;
1249         int ret;
1250
1251         id_priv = container_of(id, struct rdma_id_private, id);
1252         if (id_priv->state == CMA_IDLE) {
1253                 ret = cma_bind_any(id, AF_INET);
1254                 if (ret)
1255                         return ret;
1256         }
1257
1258         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1259                 return -EINVAL;
1260
1261         id_priv->backlog = backlog;
1262         if (id->device) {
1263                 switch (rdma_node_get_transport(id->device->node_type)) {
1264                 case RDMA_TRANSPORT_IB:
1265                         ret = cma_ib_listen(id_priv);
1266                         if (ret)
1267                                 goto err;
1268                         break;
1269                 case RDMA_TRANSPORT_IWARP:
1270                         ret = cma_iw_listen(id_priv, backlog);
1271                         if (ret)
1272                                 goto err;
1273                         break;
1274                 default:
1275                         ret = -ENOSYS;
1276                         goto err;
1277                 }
1278         } else
1279                 cma_listen_on_all(id_priv);
1280
1281         return 0;
1282 err:
1283         id_priv->backlog = 0;
1284         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1285         return ret;
1286 }
1287 EXPORT_SYMBOL(rdma_listen);
1288
1289 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1290                               void *context)
1291 {
1292         struct cma_work *work = context;
1293         struct rdma_route *route;
1294
1295         route = &work->id->id.route;
1296
1297         if (!status) {
1298                 route->num_paths = 1;
1299                 *route->path_rec = *path_rec;
1300         } else {
1301                 work->old_state = CMA_ROUTE_QUERY;
1302                 work->new_state = CMA_ADDR_RESOLVED;
1303                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1304         }
1305
1306         queue_work(cma_wq, &work->work);
1307 }
1308
1309 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1310                               struct cma_work *work)
1311 {
1312         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1313         struct ib_sa_path_rec path_rec;
1314
1315         memset(&path_rec, 0, sizeof path_rec);
1316         ib_addr_get_sgid(addr, &path_rec.sgid);
1317         ib_addr_get_dgid(addr, &path_rec.dgid);
1318         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1319         path_rec.numb_path = 1;
1320
1321         id_priv->query_id = ib_sa_path_rec_get(id_priv->id.device,
1322                                 id_priv->id.port_num, &path_rec,
1323                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1324                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
1325                                 timeout_ms, GFP_KERNEL,
1326                                 cma_query_handler, work, &id_priv->query);
1327
1328         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1329 }
1330
1331 static void cma_work_handler(void *data)
1332 {
1333         struct cma_work *work = data;
1334         struct rdma_id_private *id_priv = work->id;
1335         int destroy = 0;
1336
1337         atomic_inc(&id_priv->dev_remove);
1338         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1339                 goto out;
1340
1341         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1342                 cma_exch(id_priv, CMA_DESTROYING);
1343                 destroy = 1;
1344         }
1345 out:
1346         cma_release_remove(id_priv);
1347         cma_deref_id(id_priv);
1348         if (destroy)
1349                 rdma_destroy_id(&id_priv->id);
1350         kfree(work);
1351 }
1352
1353 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1354 {
1355         struct rdma_route *route = &id_priv->id.route;
1356         struct cma_work *work;
1357         int ret;
1358
1359         work = kzalloc(sizeof *work, GFP_KERNEL);
1360         if (!work)
1361                 return -ENOMEM;
1362
1363         work->id = id_priv;
1364         INIT_WORK(&work->work, cma_work_handler, work);
1365         work->old_state = CMA_ROUTE_QUERY;
1366         work->new_state = CMA_ROUTE_RESOLVED;
1367         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1368
1369         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1370         if (!route->path_rec) {
1371                 ret = -ENOMEM;
1372                 goto err1;
1373         }
1374
1375         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1376         if (ret)
1377                 goto err2;
1378
1379         return 0;
1380 err2:
1381         kfree(route->path_rec);
1382         route->path_rec = NULL;
1383 err1:
1384         kfree(work);
1385         return ret;
1386 }
1387
1388 int rdma_set_ib_paths(struct rdma_cm_id *id,
1389                       struct ib_sa_path_rec *path_rec, int num_paths)
1390 {
1391         struct rdma_id_private *id_priv;
1392         int ret;
1393
1394         id_priv = container_of(id, struct rdma_id_private, id);
1395         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1396                 return -EINVAL;
1397
1398         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1399         if (!id->route.path_rec) {
1400                 ret = -ENOMEM;
1401                 goto err;
1402         }
1403
1404         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1405         return 0;
1406 err:
1407         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1408         return ret;
1409 }
1410 EXPORT_SYMBOL(rdma_set_ib_paths);
1411
1412 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1413 {
1414         struct cma_work *work;
1415
1416         work = kzalloc(sizeof *work, GFP_KERNEL);
1417         if (!work)
1418                 return -ENOMEM;
1419
1420         work->id = id_priv;
1421         INIT_WORK(&work->work, cma_work_handler, work);
1422         work->old_state = CMA_ROUTE_QUERY;
1423         work->new_state = CMA_ROUTE_RESOLVED;
1424         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1425         queue_work(cma_wq, &work->work);
1426         return 0;
1427 }
1428
1429 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1430 {
1431         struct rdma_id_private *id_priv;
1432         int ret;
1433
1434         id_priv = container_of(id, struct rdma_id_private, id);
1435         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1436                 return -EINVAL;
1437
1438         atomic_inc(&id_priv->refcount);
1439         switch (rdma_node_get_transport(id->device->node_type)) {
1440         case RDMA_TRANSPORT_IB:
1441                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1442                 break;
1443         case RDMA_TRANSPORT_IWARP:
1444                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1445                 break;
1446         default:
1447                 ret = -ENOSYS;
1448                 break;
1449         }
1450         if (ret)
1451                 goto err;
1452
1453         return 0;
1454 err:
1455         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1456         cma_deref_id(id_priv);
1457         return ret;
1458 }
1459 EXPORT_SYMBOL(rdma_resolve_route);
1460
1461 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1462 {
1463         struct cma_device *cma_dev;
1464         struct ib_port_attr port_attr;
1465         union ib_gid gid;
1466         u16 pkey;
1467         int ret;
1468         u8 p;
1469
1470         mutex_lock(&lock);
1471         list_for_each_entry(cma_dev, &dev_list, list)
1472                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1473                         if (!ib_query_port (cma_dev->device, p, &port_attr) &&
1474                             port_attr.state == IB_PORT_ACTIVE)
1475                                 goto port_found;
1476
1477         if (!list_empty(&dev_list)) {
1478                 p = 1;
1479                 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1480         } else {
1481                 ret = -ENODEV;
1482                 goto out;
1483         }
1484
1485 port_found:
1486         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1487         if (ret)
1488                 goto out;
1489
1490         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1491         if (ret)
1492                 goto out;
1493
1494         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1495         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1496         id_priv->id.port_num = p;
1497         cma_attach_to_dev(id_priv, cma_dev);
1498 out:
1499         mutex_unlock(&lock);
1500         return ret;
1501 }
1502
1503 static void addr_handler(int status, struct sockaddr *src_addr,
1504                          struct rdma_dev_addr *dev_addr, void *context)
1505 {
1506         struct rdma_id_private *id_priv = context;
1507         enum rdma_cm_event_type event;
1508
1509         atomic_inc(&id_priv->dev_remove);
1510         if (!id_priv->cma_dev && !status)
1511                 status = cma_acquire_dev(id_priv);
1512
1513         if (status) {
1514                 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND))
1515                         goto out;
1516                 event = RDMA_CM_EVENT_ADDR_ERROR;
1517         } else {
1518                 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED))
1519                         goto out;
1520                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1521                        ip_addr_size(src_addr));
1522                 event = RDMA_CM_EVENT_ADDR_RESOLVED;
1523         }
1524
1525         if (cma_notify_user(id_priv, event, status, NULL, 0)) {
1526                 cma_exch(id_priv, CMA_DESTROYING);
1527                 cma_release_remove(id_priv);
1528                 cma_deref_id(id_priv);
1529                 rdma_destroy_id(&id_priv->id);
1530                 return;
1531         }
1532 out:
1533         cma_release_remove(id_priv);
1534         cma_deref_id(id_priv);
1535 }
1536
1537 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1538 {
1539         struct cma_work *work;
1540         struct sockaddr_in *src_in, *dst_in;
1541         union ib_gid gid;
1542         int ret;
1543
1544         work = kzalloc(sizeof *work, GFP_KERNEL);
1545         if (!work)
1546                 return -ENOMEM;
1547
1548         if (!id_priv->cma_dev) {
1549                 ret = cma_bind_loopback(id_priv);
1550                 if (ret)
1551                         goto err;
1552         }
1553
1554         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1555         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1556
1557         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1558                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1559                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1560                 src_in->sin_family = dst_in->sin_family;
1561                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1562         }
1563
1564         work->id = id_priv;
1565         INIT_WORK(&work->work, cma_work_handler, work);
1566         work->old_state = CMA_ADDR_QUERY;
1567         work->new_state = CMA_ADDR_RESOLVED;
1568         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1569         queue_work(cma_wq, &work->work);
1570         return 0;
1571 err:
1572         kfree(work);
1573         return ret;
1574 }
1575
1576 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1577                          struct sockaddr *dst_addr)
1578 {
1579         if (src_addr && src_addr->sa_family)
1580                 return rdma_bind_addr(id, src_addr);
1581         else
1582                 return cma_bind_any(id, dst_addr->sa_family);
1583 }
1584
1585 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1586                       struct sockaddr *dst_addr, int timeout_ms)
1587 {
1588         struct rdma_id_private *id_priv;
1589         int ret;
1590
1591         id_priv = container_of(id, struct rdma_id_private, id);
1592         if (id_priv->state == CMA_IDLE) {
1593                 ret = cma_bind_addr(id, src_addr, dst_addr);
1594                 if (ret)
1595                         return ret;
1596         }
1597
1598         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1599                 return -EINVAL;
1600
1601         atomic_inc(&id_priv->refcount);
1602         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1603         if (cma_any_addr(dst_addr))
1604                 ret = cma_resolve_loopback(id_priv);
1605         else
1606                 ret = rdma_resolve_ip(&id->route.addr.src_addr, dst_addr,
1607                                       &id->route.addr.dev_addr,
1608                                       timeout_ms, addr_handler, id_priv);
1609         if (ret)
1610                 goto err;
1611
1612         return 0;
1613 err:
1614         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1615         cma_deref_id(id_priv);
1616         return ret;
1617 }
1618 EXPORT_SYMBOL(rdma_resolve_addr);
1619
1620 static void cma_bind_port(struct rdma_bind_list *bind_list,
1621                           struct rdma_id_private *id_priv)
1622 {
1623         struct sockaddr_in *sin;
1624
1625         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1626         sin->sin_port = htons(bind_list->port);
1627         id_priv->bind_list = bind_list;
1628         hlist_add_head(&id_priv->node, &bind_list->owners);
1629 }
1630
1631 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1632                           unsigned short snum)
1633 {
1634         struct rdma_bind_list *bind_list;
1635         int port, start, ret;
1636
1637         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1638         if (!bind_list)
1639                 return -ENOMEM;
1640
1641         start = snum ? snum : sysctl_local_port_range[0];
1642
1643         do {
1644                 ret = idr_get_new_above(ps, bind_list, start, &port);
1645         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1646
1647         if (ret)
1648                 goto err;
1649
1650         if ((snum && port != snum) ||
1651             (!snum && port > sysctl_local_port_range[1])) {
1652                 idr_remove(ps, port);
1653                 ret = -EADDRNOTAVAIL;
1654                 goto err;
1655         }
1656
1657         bind_list->ps = ps;
1658         bind_list->port = (unsigned short) port;
1659         cma_bind_port(bind_list, id_priv);
1660         return 0;
1661 err:
1662         kfree(bind_list);
1663         return ret;
1664 }
1665
1666 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1667 {
1668         struct rdma_id_private *cur_id;
1669         struct sockaddr_in *sin, *cur_sin;
1670         struct rdma_bind_list *bind_list;
1671         struct hlist_node *node;
1672         unsigned short snum;
1673
1674         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1675         snum = ntohs(sin->sin_port);
1676         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1677                 return -EACCES;
1678
1679         bind_list = idr_find(ps, snum);
1680         if (!bind_list)
1681                 return cma_alloc_port(ps, id_priv, snum);
1682
1683         /*
1684          * We don't support binding to any address if anyone is bound to
1685          * a specific address on the same port.
1686          */
1687         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1688                 return -EADDRNOTAVAIL;
1689
1690         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1691                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1692                         return -EADDRNOTAVAIL;
1693
1694                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1695                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1696                         return -EADDRINUSE;
1697         }
1698
1699         cma_bind_port(bind_list, id_priv);
1700         return 0;
1701 }
1702
1703 static int cma_get_port(struct rdma_id_private *id_priv)
1704 {
1705         struct idr *ps;
1706         int ret;
1707
1708         switch (id_priv->id.ps) {
1709         case RDMA_PS_SDP:
1710                 ps = &sdp_ps;
1711                 break;
1712         case RDMA_PS_TCP:
1713                 ps = &tcp_ps;
1714                 break;
1715         default:
1716                 return -EPROTONOSUPPORT;
1717         }
1718
1719         mutex_lock(&lock);
1720         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1721                 ret = cma_alloc_port(ps, id_priv, 0);
1722         else
1723                 ret = cma_use_port(ps, id_priv);
1724         mutex_unlock(&lock);
1725
1726         return ret;
1727 }
1728
1729 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1730 {
1731         struct rdma_id_private *id_priv;
1732         int ret;
1733
1734         if (addr->sa_family != AF_INET)
1735                 return -EAFNOSUPPORT;
1736
1737         id_priv = container_of(id, struct rdma_id_private, id);
1738         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1739                 return -EINVAL;
1740
1741         if (!cma_any_addr(addr)) {
1742                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1743                 if (!ret)
1744                         ret = cma_acquire_dev(id_priv);
1745                 if (ret)
1746                         goto err;
1747         }
1748
1749         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
1750         ret = cma_get_port(id_priv);
1751         if (ret)
1752                 goto err;
1753
1754         return 0;
1755 err:
1756         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
1757         return ret;
1758 }
1759 EXPORT_SYMBOL(rdma_bind_addr);
1760
1761 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
1762                           struct rdma_route *route)
1763 {
1764         struct sockaddr_in *src4, *dst4;
1765         struct cma_hdr *cma_hdr;
1766         struct sdp_hh *sdp_hdr;
1767
1768         src4 = (struct sockaddr_in *) &route->addr.src_addr;
1769         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
1770
1771         switch (ps) {
1772         case RDMA_PS_SDP:
1773                 sdp_hdr = hdr;
1774                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
1775                         return -EINVAL;
1776                 sdp_set_ip_ver(sdp_hdr, 4);
1777                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1778                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1779                 sdp_hdr->port = src4->sin_port;
1780                 break;
1781         default:
1782                 cma_hdr = hdr;
1783                 cma_hdr->cma_version = CMA_VERSION;
1784                 cma_set_ip_ver(cma_hdr, 4);
1785                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1786                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1787                 cma_hdr->port = src4->sin_port;
1788                 break;
1789         }
1790         return 0;
1791 }
1792
1793 static int cma_connect_ib(struct rdma_id_private *id_priv,
1794                           struct rdma_conn_param *conn_param)
1795 {
1796         struct ib_cm_req_param req;
1797         struct rdma_route *route;
1798         void *private_data;
1799         int offset, ret;
1800
1801         memset(&req, 0, sizeof req);
1802         offset = cma_user_data_offset(id_priv->id.ps);
1803         req.private_data_len = offset + conn_param->private_data_len;
1804         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
1805         if (!private_data)
1806                 return -ENOMEM;
1807
1808         if (conn_param->private_data && conn_param->private_data_len)
1809                 memcpy(private_data + offset, conn_param->private_data,
1810                        conn_param->private_data_len);
1811
1812         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
1813                                             id_priv);
1814         if (IS_ERR(id_priv->cm_id.ib)) {
1815                 ret = PTR_ERR(id_priv->cm_id.ib);
1816                 goto out;
1817         }
1818
1819         route = &id_priv->id.route;
1820         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
1821         if (ret)
1822                 goto out;
1823         req.private_data = private_data;
1824
1825         req.primary_path = &route->path_rec[0];
1826         if (route->num_paths == 2)
1827                 req.alternate_path = &route->path_rec[1];
1828
1829         req.service_id = cma_get_service_id(id_priv->id.ps,
1830                                             &route->addr.dst_addr);
1831         req.qp_num = id_priv->qp_num;
1832         req.qp_type = id_priv->qp_type;
1833         req.starting_psn = id_priv->seq_num;
1834         req.responder_resources = conn_param->responder_resources;
1835         req.initiator_depth = conn_param->initiator_depth;
1836         req.flow_control = conn_param->flow_control;
1837         req.retry_count = conn_param->retry_count;
1838         req.rnr_retry_count = conn_param->rnr_retry_count;
1839         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1840         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1841         req.max_cm_retries = CMA_MAX_CM_RETRIES;
1842         req.srq = id_priv->srq ? 1 : 0;
1843
1844         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
1845 out:
1846         kfree(private_data);
1847         return ret;
1848 }
1849
1850 static int cma_connect_iw(struct rdma_id_private *id_priv,
1851                           struct rdma_conn_param *conn_param)
1852 {
1853         struct iw_cm_id *cm_id;
1854         struct sockaddr_in* sin;
1855         int ret;
1856         struct iw_cm_conn_param iw_param;
1857
1858         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
1859         if (IS_ERR(cm_id)) {
1860                 ret = PTR_ERR(cm_id);
1861                 goto out;
1862         }
1863
1864         id_priv->cm_id.iw = cm_id;
1865
1866         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
1867         cm_id->local_addr = *sin;
1868
1869         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
1870         cm_id->remote_addr = *sin;
1871
1872         ret = cma_modify_qp_rtr(&id_priv->id);
1873         if (ret) {
1874                 iw_destroy_cm_id(cm_id);
1875                 return ret;
1876         }
1877
1878         iw_param.ord = conn_param->initiator_depth;
1879         iw_param.ird = conn_param->responder_resources;
1880         iw_param.private_data = conn_param->private_data;
1881         iw_param.private_data_len = conn_param->private_data_len;
1882         if (id_priv->id.qp)
1883                 iw_param.qpn = id_priv->qp_num;
1884         else
1885                 iw_param.qpn = conn_param->qp_num;
1886         ret = iw_cm_connect(cm_id, &iw_param);
1887 out:
1888         return ret;
1889 }
1890
1891 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1892 {
1893         struct rdma_id_private *id_priv;
1894         int ret;
1895
1896         id_priv = container_of(id, struct rdma_id_private, id);
1897         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
1898                 return -EINVAL;
1899
1900         if (!id->qp) {
1901                 id_priv->qp_num = conn_param->qp_num;
1902                 id_priv->qp_type = conn_param->qp_type;
1903                 id_priv->srq = conn_param->srq;
1904         }
1905
1906         switch (rdma_node_get_transport(id->device->node_type)) {
1907         case RDMA_TRANSPORT_IB:
1908                 ret = cma_connect_ib(id_priv, conn_param);
1909                 break;
1910         case RDMA_TRANSPORT_IWARP:
1911                 ret = cma_connect_iw(id_priv, conn_param);
1912                 break;
1913         default:
1914                 ret = -ENOSYS;
1915                 break;
1916         }
1917         if (ret)
1918                 goto err;
1919
1920         return 0;
1921 err:
1922         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
1923         return ret;
1924 }
1925 EXPORT_SYMBOL(rdma_connect);
1926
1927 static int cma_accept_ib(struct rdma_id_private *id_priv,
1928                          struct rdma_conn_param *conn_param)
1929 {
1930         struct ib_cm_rep_param rep;
1931         int ret;
1932
1933         ret = cma_modify_qp_rtr(&id_priv->id);
1934         if (ret)
1935                 return ret;
1936
1937         memset(&rep, 0, sizeof rep);
1938         rep.qp_num = id_priv->qp_num;
1939         rep.starting_psn = id_priv->seq_num;
1940         rep.private_data = conn_param->private_data;
1941         rep.private_data_len = conn_param->private_data_len;
1942         rep.responder_resources = conn_param->responder_resources;
1943         rep.initiator_depth = conn_param->initiator_depth;
1944         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
1945         rep.failover_accepted = 0;
1946         rep.flow_control = conn_param->flow_control;
1947         rep.rnr_retry_count = conn_param->rnr_retry_count;
1948         rep.srq = id_priv->srq ? 1 : 0;
1949
1950         return ib_send_cm_rep(id_priv->cm_id.ib, &rep);
1951 }
1952
1953 static int cma_accept_iw(struct rdma_id_private *id_priv,
1954                   struct rdma_conn_param *conn_param)
1955 {
1956         struct iw_cm_conn_param iw_param;
1957         int ret;
1958
1959         ret = cma_modify_qp_rtr(&id_priv->id);
1960         if (ret)
1961                 return ret;
1962
1963         iw_param.ord = conn_param->initiator_depth;
1964         iw_param.ird = conn_param->responder_resources;
1965         iw_param.private_data = conn_param->private_data;
1966         iw_param.private_data_len = conn_param->private_data_len;
1967         if (id_priv->id.qp) {
1968                 iw_param.qpn = id_priv->qp_num;
1969         } else
1970                 iw_param.qpn = conn_param->qp_num;
1971
1972         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
1973 }
1974
1975 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1976 {
1977         struct rdma_id_private *id_priv;
1978         int ret;
1979
1980         id_priv = container_of(id, struct rdma_id_private, id);
1981         if (!cma_comp(id_priv, CMA_CONNECT))
1982                 return -EINVAL;
1983
1984         if (!id->qp && conn_param) {
1985                 id_priv->qp_num = conn_param->qp_num;
1986                 id_priv->qp_type = conn_param->qp_type;
1987                 id_priv->srq = conn_param->srq;
1988         }
1989
1990         switch (rdma_node_get_transport(id->device->node_type)) {
1991         case RDMA_TRANSPORT_IB:
1992                 if (conn_param)
1993                         ret = cma_accept_ib(id_priv, conn_param);
1994                 else
1995                         ret = cma_rep_recv(id_priv);
1996                 break;
1997         case RDMA_TRANSPORT_IWARP:
1998                 ret = cma_accept_iw(id_priv, conn_param);
1999                 break;
2000         default:
2001                 ret = -ENOSYS;
2002                 break;
2003         }
2004
2005         if (ret)
2006                 goto reject;
2007
2008         return 0;
2009 reject:
2010         cma_modify_qp_err(id);
2011         rdma_reject(id, NULL, 0);
2012         return ret;
2013 }
2014 EXPORT_SYMBOL(rdma_accept);
2015
2016 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2017                 u8 private_data_len)
2018 {
2019         struct rdma_id_private *id_priv;
2020         int ret;
2021
2022         id_priv = container_of(id, struct rdma_id_private, id);
2023         if (!cma_comp(id_priv, CMA_CONNECT))
2024                 return -EINVAL;
2025
2026         switch (rdma_node_get_transport(id->device->node_type)) {
2027         case RDMA_TRANSPORT_IB:
2028                 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2029                                      IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2030                                      private_data, private_data_len);
2031                 break;
2032         case RDMA_TRANSPORT_IWARP:
2033                 ret = iw_cm_reject(id_priv->cm_id.iw,
2034                                    private_data, private_data_len);
2035                 break;
2036         default:
2037                 ret = -ENOSYS;
2038                 break;
2039         }
2040         return ret;
2041 }
2042 EXPORT_SYMBOL(rdma_reject);
2043
2044 int rdma_disconnect(struct rdma_cm_id *id)
2045 {
2046         struct rdma_id_private *id_priv;
2047         int ret;
2048
2049         id_priv = container_of(id, struct rdma_id_private, id);
2050         if (!cma_comp(id_priv, CMA_CONNECT) &&
2051             !cma_comp(id_priv, CMA_DISCONNECT))
2052                 return -EINVAL;
2053
2054         switch (rdma_node_get_transport(id->device->node_type)) {
2055         case RDMA_TRANSPORT_IB:
2056                 ret = cma_modify_qp_err(id);
2057                 if (ret)
2058                         goto out;
2059                 /* Initiate or respond to a disconnect. */
2060                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2061                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2062                 break;
2063         case RDMA_TRANSPORT_IWARP:
2064                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2065                 break;
2066         default:
2067                 ret = -EINVAL;
2068                 break;
2069         }
2070 out:
2071         return ret;
2072 }
2073 EXPORT_SYMBOL(rdma_disconnect);
2074
2075 static void cma_add_one(struct ib_device *device)
2076 {
2077         struct cma_device *cma_dev;
2078         struct rdma_id_private *id_priv;
2079
2080         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2081         if (!cma_dev)
2082                 return;
2083
2084         cma_dev->device = device;
2085         cma_dev->node_guid = device->node_guid;
2086         if (!cma_dev->node_guid)
2087                 goto err;
2088
2089         init_completion(&cma_dev->comp);
2090         atomic_set(&cma_dev->refcount, 1);
2091         INIT_LIST_HEAD(&cma_dev->id_list);
2092         ib_set_client_data(device, &cma_client, cma_dev);
2093
2094         mutex_lock(&lock);
2095         list_add_tail(&cma_dev->list, &dev_list);
2096         list_for_each_entry(id_priv, &listen_any_list, list)
2097                 cma_listen_on_dev(id_priv, cma_dev);
2098         mutex_unlock(&lock);
2099         return;
2100 err:
2101         kfree(cma_dev);
2102 }
2103
2104 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2105 {
2106         enum cma_state state;
2107
2108         /* Record that we want to remove the device */
2109         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2110         if (state == CMA_DESTROYING)
2111                 return 0;
2112
2113         cma_cancel_operation(id_priv, state);
2114         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2115
2116         /* Check for destruction from another callback. */
2117         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2118                 return 0;
2119
2120         return cma_notify_user(id_priv, RDMA_CM_EVENT_DEVICE_REMOVAL,
2121                                0, NULL, 0);
2122 }
2123
2124 static void cma_process_remove(struct cma_device *cma_dev)
2125 {
2126         struct list_head remove_list;
2127         struct rdma_id_private *id_priv;
2128         int ret;
2129
2130         INIT_LIST_HEAD(&remove_list);
2131
2132         mutex_lock(&lock);
2133         while (!list_empty(&cma_dev->id_list)) {
2134                 id_priv = list_entry(cma_dev->id_list.next,
2135                                      struct rdma_id_private, list);
2136
2137                 if (cma_internal_listen(id_priv)) {
2138                         cma_destroy_listen(id_priv);
2139                         continue;
2140                 }
2141
2142                 list_del(&id_priv->list);
2143                 list_add_tail(&id_priv->list, &remove_list);
2144                 atomic_inc(&id_priv->refcount);
2145                 mutex_unlock(&lock);
2146
2147                 ret = cma_remove_id_dev(id_priv);
2148                 cma_deref_id(id_priv);
2149                 if (ret)
2150                         rdma_destroy_id(&id_priv->id);
2151
2152                 mutex_lock(&lock);
2153         }
2154         mutex_unlock(&lock);
2155
2156         cma_deref_dev(cma_dev);
2157         wait_for_completion(&cma_dev->comp);
2158 }
2159
2160 static void cma_remove_one(struct ib_device *device)
2161 {
2162         struct cma_device *cma_dev;
2163
2164         cma_dev = ib_get_client_data(device, &cma_client);
2165         if (!cma_dev)
2166                 return;
2167
2168         mutex_lock(&lock);
2169         list_del(&cma_dev->list);
2170         mutex_unlock(&lock);
2171
2172         cma_process_remove(cma_dev);
2173         kfree(cma_dev);
2174 }
2175
2176 static int cma_init(void)
2177 {
2178         int ret;
2179
2180         cma_wq = create_singlethread_workqueue("rdma_cm_wq");
2181         if (!cma_wq)
2182                 return -ENOMEM;
2183
2184         ret = ib_register_client(&cma_client);
2185         if (ret)
2186                 goto err;
2187         return 0;
2188
2189 err:
2190         destroy_workqueue(cma_wq);
2191         return ret;
2192 }
2193
2194 static void cma_cleanup(void)
2195 {
2196         ib_unregister_client(&cma_client);
2197         destroy_workqueue(cma_wq);
2198         idr_destroy(&sdp_ps);
2199         idr_destroy(&tcp_ps);
2200 }
2201
2202 module_init(cma_init);
2203 module_exit(cma_cleanup);