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 struct ib_sa_client sa_client;
  66 static LIST_HEAD(dev_list);
  67 static LIST_HEAD(listen_any_list);
  68 static DEFINE_MUTEX(lock);
  69 static struct workqueue_struct *cma_wq;
  70 static DEFINE_IDR(sdp_ps);
  71 static DEFINE_IDR(tcp_ps);
  72
  73 struct cma_device {
  74         struct list_head        list;
  75         struct ib_device        *device;
  76         __be64                  node_guid;
  77         struct completion       comp;
  78         atomic_t                refcount;
  79         struct list_head        id_list;
  80 };
  81
  82 enum cma_state {
  83         CMA_IDLE,
  84         CMA_ADDR_QUERY,
  85         CMA_ADDR_RESOLVED,
  86         CMA_ROUTE_QUERY,
  87         CMA_ROUTE_RESOLVED,
  88         CMA_CONNECT,
  89         CMA_DISCONNECT,
  90         CMA_ADDR_BOUND,
  91         CMA_LISTEN,
  92         CMA_DEVICE_REMOVAL,
  93         CMA_DESTROYING
  94 };
  95
  96 struct rdma_bind_list {
  97         struct idr              *ps;
  98         struct hlist_head       owners;
  99         unsigned short          port;
 100 };
 101
 102 /*
 103  * Device removal can occur at anytime, so we need extra handling to
 104  * serialize notifying the user of device removal with other callbacks.
 105  * We do this by disabling removal notification while a callback is in process,
 106  * and reporting it after the callback completes.
 107  */
 108 struct rdma_id_private {
 109         struct rdma_cm_id       id;
 110
 111         struct rdma_bind_list   *bind_list;
 112         struct hlist_node       node;
 113         struct list_head        list;
 114         struct list_head        listen_list;
 115         struct cma_device       *cma_dev;
 116
 117         enum cma_state          state;
 118         spinlock_t              lock;
 119         struct completion       comp;
 120         atomic_t                refcount;
 121         wait_queue_head_t       wait_remove;
 122         atomic_t                dev_remove;
 123
 124         int                     backlog;
 125         int                     timeout_ms;
 126         struct ib_sa_query      *query;
 127         int                     query_id;
 128         union {
 129                 struct ib_cm_id *ib;
 130                 struct iw_cm_id *iw;
 131         } cm_id;
 132
 133         u32                     seq_num;
 134         u32                     qp_num;
 135         enum ib_qp_type         qp_type;
 136         u8                      srq;
 137 };
 138
 139 struct cma_work {
 140         struct work_struct      work;
 141         struct rdma_id_private  *id;
 142         enum cma_state          old_state;
 143         enum cma_state          new_state;
 144         struct rdma_cm_event    event;
 145 };
 146
 147 union cma_ip_addr {
 148         struct in6_addr ip6;
 149         struct {
 150                 __u32 pad[3];
 151                 __u32 addr;
 152         } ip4;
 153 };
 154
 155 struct cma_hdr {
 156         u8 cma_version;
 157         u8 ip_version;  /* IP version: 7:4 */
 158         __u16 port;
 159         union cma_ip_addr src_addr;
 160         union cma_ip_addr dst_addr;
 161 };
 162
 163 struct sdp_hh {
 164         u8 bsdh[16];
 165         u8 sdp_version; /* Major version: 7:4 */
 166         u8 ip_version;  /* IP version: 7:4 */
 167         u8 sdp_specific1[10];
 168         __u16 port;
 169         __u16 sdp_specific2;
 170         union cma_ip_addr src_addr;
 171         union cma_ip_addr dst_addr;
 172 };
 173
 174 struct sdp_hah {
 175         u8 bsdh[16];
 176         u8 sdp_version;
 177 };
 178
 179 #define CMA_VERSION 0x00
 180 #define SDP_MAJ_VERSION 0x2
 181
 182 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 183 {
 184         unsigned long flags;
 185         int ret;
 186
 187         spin_lock_irqsave(&id_priv->lock, flags);
 188         ret = (id_priv->state == comp);
 189         spin_unlock_irqrestore(&id_priv->lock, flags);
 190         return ret;
 191 }
 192
 193 static int cma_comp_exch(struct rdma_id_private *id_priv,
 194                          enum cma_state comp, enum cma_state exch)
 195 {
 196         unsigned long flags;
 197         int ret;
 198
 199         spin_lock_irqsave(&id_priv->lock, flags);
 200         if ((ret = (id_priv->state == comp)))
 201                 id_priv->state = exch;
 202         spin_unlock_irqrestore(&id_priv->lock, flags);
 203         return ret;
 204 }
 205
 206 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 207                                enum cma_state exch)
 208 {
 209         unsigned long flags;
 210         enum cma_state old;
 211
 212         spin_lock_irqsave(&id_priv->lock, flags);
 213         old = id_priv->state;
 214         id_priv->state = exch;
 215         spin_unlock_irqrestore(&id_priv->lock, flags);
 216         return old;
 217 }
 218
 219 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 220 {
 221         return hdr->ip_version >> 4;
 222 }
 223
 224 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 225 {
 226         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 227 }
 228
 229 static inline u8 sdp_get_majv(u8 sdp_version)
 230 {
 231         return sdp_version >> 4;
 232 }
 233
 234 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 235 {
 236         return hh->ip_version >> 4;
 237 }
 238
 239 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 240 {
 241         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 242 }
 243
 244 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 245                               struct cma_device *cma_dev)
 246 {
 247         atomic_inc(&cma_dev->refcount);
 248         id_priv->cma_dev = cma_dev;
 249         id_priv->id.device = cma_dev->device;
 250         list_add_tail(&id_priv->list, &cma_dev->id_list);
 251 }
 252
 253 static inline void cma_deref_dev(struct cma_device *cma_dev)
 254 {
 255         if (atomic_dec_and_test(&cma_dev->refcount))
 256                 complete(&cma_dev->comp);
 257 }
 258
 259 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 260 {
 261         list_del(&id_priv->list);
 262         cma_deref_dev(id_priv->cma_dev);
 263         id_priv->cma_dev = NULL;
 264 }
 265
 266 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 267 {
 268         enum rdma_node_type dev_type = id_priv->id.route.addr.dev_addr.dev_type;
 269         struct cma_device *cma_dev;
 270         union ib_gid gid;
 271         int ret = -ENODEV;
 272
 273         switch (rdma_node_get_transport(dev_type)) {
 274         case RDMA_TRANSPORT_IB:
 275                 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
 276                 break;
 277         case RDMA_TRANSPORT_IWARP:
 278                 iw_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
 279                 break;
 280         default:
 281                 return -ENODEV;
 282         }
 283
 284         list_for_each_entry(cma_dev, &dev_list, list) {
 285                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 286                                          &id_priv->id.port_num, NULL);
 287                 if (!ret) {
 288                         cma_attach_to_dev(id_priv, cma_dev);
 289                         break;
 290                 }
 291         }
 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         mutex_lock(&lock);
 716         if (id_priv->cma_dev) {
 717                 mutex_unlock(&lock);
 718                 switch (rdma_node_get_transport(id->device->node_type)) {
 719                 case RDMA_TRANSPORT_IB:
 720                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 721                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 722                         break;
 723                 case RDMA_TRANSPORT_IWARP:
 724                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 725                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 726                         break;
 727                 default:
 728                         break;
 729                 }
 730                 mutex_lock(&lock);
 731                 cma_detach_from_dev(id_priv);
 732         }
 733         mutex_unlock(&lock);
 734
 735         cma_release_port(id_priv);
 736         cma_deref_id(id_priv);
 737         wait_for_completion(&id_priv->comp);
 738
 739         kfree(id_priv->id.route.path_rec);
 740         kfree(id_priv);
 741 }
 742 EXPORT_SYMBOL(rdma_destroy_id);
 743
 744 static int cma_rep_recv(struct rdma_id_private *id_priv)
 745 {
 746         int ret;
 747
 748         ret = cma_modify_qp_rtr(&id_priv->id);
 749         if (ret)
 750                 goto reject;
 751
 752         ret = cma_modify_qp_rts(&id_priv->id);
 753         if (ret)
 754                 goto reject;
 755
 756         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 757         if (ret)
 758                 goto reject;
 759
 760         return 0;
 761 reject:
 762         cma_modify_qp_err(&id_priv->id);
 763         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 764                        NULL, 0, NULL, 0);
 765         return ret;
 766 }
 767
 768 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 769 {
 770         if (id_priv->id.ps == RDMA_PS_SDP &&
 771             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 772             SDP_MAJ_VERSION)
 773                 return -EINVAL;
 774
 775         return 0;
 776 }
 777
 778 static int cma_rtu_recv(struct rdma_id_private *id_priv)
 779 {
 780         int ret;
 781
 782         ret = cma_modify_qp_rts(&id_priv->id);
 783         if (ret)
 784                 goto reject;
 785
 786         return 0;
 787 reject:
 788         cma_modify_qp_err(&id_priv->id);
 789         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 790                        NULL, 0, NULL, 0);
 791         return ret;
 792 }
 793
 794 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 795 {
 796         struct rdma_id_private *id_priv = cm_id->context;
 797         enum rdma_cm_event_type event;
 798         u8 private_data_len = 0;
 799         int ret = 0, status = 0;
 800
 801         atomic_inc(&id_priv->dev_remove);
 802         if (!cma_comp(id_priv, CMA_CONNECT))
 803                 goto out;
 804
 805         switch (ib_event->event) {
 806         case IB_CM_REQ_ERROR:
 807         case IB_CM_REP_ERROR:
 808                 event = RDMA_CM_EVENT_UNREACHABLE;
 809                 status = -ETIMEDOUT;
 810                 break;
 811         case IB_CM_REP_RECEIVED:
 812                 status = cma_verify_rep(id_priv, ib_event->private_data);
 813                 if (status)
 814                         event = RDMA_CM_EVENT_CONNECT_ERROR;
 815                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 816                         status = cma_rep_recv(id_priv);
 817                         event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
 818                                          RDMA_CM_EVENT_ESTABLISHED;
 819                 } else
 820                         event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 821                 private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 822                 break;
 823         case IB_CM_RTU_RECEIVED:
 824                 status = cma_rtu_recv(id_priv);
 825                 event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
 826                                  RDMA_CM_EVENT_ESTABLISHED;
 827                 break;
 828         case IB_CM_DREQ_ERROR:
 829                 status = -ETIMEDOUT; /* fall through */
 830         case IB_CM_DREQ_RECEIVED:
 831         case IB_CM_DREP_RECEIVED:
 832                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 833                         goto out;
 834                 event = RDMA_CM_EVENT_DISCONNECTED;
 835                 break;
 836         case IB_CM_TIMEWAIT_EXIT:
 837         case IB_CM_MRA_RECEIVED:
 838                 /* ignore event */
 839                 goto out;
 840         case IB_CM_REJ_RECEIVED:
 841                 cma_modify_qp_err(&id_priv->id);
 842                 status = ib_event->param.rej_rcvd.reason;
 843                 event = RDMA_CM_EVENT_REJECTED;
 844                 private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 845                 break;
 846         default:
 847                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
 848                        ib_event->event);
 849                 goto out;
 850         }
 851
 852         ret = cma_notify_user(id_priv, event, status, ib_event->private_data,
 853                               private_data_len);
 854         if (ret) {
 855                 /* Destroy the CM ID by returning a non-zero value. */
 856                 id_priv->cm_id.ib = NULL;
 857                 cma_exch(id_priv, CMA_DESTROYING);
 858                 cma_release_remove(id_priv);
 859                 rdma_destroy_id(&id_priv->id);
 860                 return ret;
 861         }
 862 out:
 863         cma_release_remove(id_priv);
 864         return ret;
 865 }
 866
 867 static struct rdma_id_private *cma_new_id(struct rdma_cm_id *listen_id,
 868                                           struct ib_cm_event *ib_event)
 869 {
 870         struct rdma_id_private *id_priv;
 871         struct rdma_cm_id *id;
 872         struct rdma_route *rt;
 873         union cma_ip_addr *src, *dst;
 874         __u16 port;
 875         u8 ip_ver;
 876
 877         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 878                             listen_id->ps);
 879         if (IS_ERR(id))
 880                 return NULL;
 881
 882         rt = &id->route;
 883         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
 884         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, GFP_KERNEL);
 885         if (!rt->path_rec)
 886                 goto err;
 887
 888         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 889                              &ip_ver, &port, &src, &dst))
 890                 goto err;
 891
 892         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 893                           ip_ver, port, src, dst);
 894         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
 895         if (rt->num_paths == 2)
 896                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
 897
 898         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
 899         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
 900         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
 901         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
 902
 903         id_priv = container_of(id, struct rdma_id_private, id);
 904         id_priv->state = CMA_CONNECT;
 905         return id_priv;
 906 err:
 907         rdma_destroy_id(id);
 908         return NULL;
 909 }
 910
 911 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 912 {
 913         struct rdma_id_private *listen_id, *conn_id;
 914         int offset, ret;
 915
 916         listen_id = cm_id->context;
 917         atomic_inc(&listen_id->dev_remove);
 918         if (!cma_comp(listen_id, CMA_LISTEN)) {
 919                 ret = -ECONNABORTED;
 920                 goto out;
 921         }
 922
 923         conn_id = cma_new_id(&listen_id->id, ib_event);
 924         if (!conn_id) {
 925                 ret = -ENOMEM;
 926                 goto out;
 927         }
 928
 929         atomic_inc(&conn_id->dev_remove);
 930         mutex_lock(&lock);
 931         ret = cma_acquire_dev(conn_id);
 932         mutex_unlock(&lock);
 933         if (ret) {
 934                 ret = -ENODEV;
 935                 cma_release_remove(conn_id);
 936                 rdma_destroy_id(&conn_id->id);
 937                 goto out;
 938         }
 939
 940         conn_id->cm_id.ib = cm_id;
 941         cm_id->context = conn_id;
 942         cm_id->cm_handler = cma_ib_handler;
 943
 944         offset = cma_user_data_offset(listen_id->id.ps);
 945         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
 946                               ib_event->private_data + offset,
 947                               IB_CM_REQ_PRIVATE_DATA_SIZE - offset);
 948         if (ret) {
 949                 /* Destroy the CM ID by returning a non-zero value. */
 950                 conn_id->cm_id.ib = NULL;
 951                 cma_exch(conn_id, CMA_DESTROYING);
 952                 cma_release_remove(conn_id);
 953                 rdma_destroy_id(&conn_id->id);
 954         }
 955 out:
 956         cma_release_remove(listen_id);
 957         return ret;
 958 }
 959
 960 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
 961 {
 962         return cpu_to_be64(((u64)ps << 16) +
 963                be16_to_cpu(((struct sockaddr_in *) addr)->sin_port));
 964 }
 965
 966 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
 967                                  struct ib_cm_compare_data *compare)
 968 {
 969         struct cma_hdr *cma_data, *cma_mask;
 970         struct sdp_hh *sdp_data, *sdp_mask;
 971         __u32 ip4_addr;
 972         struct in6_addr ip6_addr;
 973
 974         memset(compare, 0, sizeof *compare);
 975         cma_data = (void *) compare->data;
 976         cma_mask = (void *) compare->mask;
 977         sdp_data = (void *) compare->data;
 978         sdp_mask = (void *) compare->mask;
 979
 980         switch (addr->sa_family) {
 981         case AF_INET:
 982                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
 983                 if (ps == RDMA_PS_SDP) {
 984                         sdp_set_ip_ver(sdp_data, 4);
 985                         sdp_set_ip_ver(sdp_mask, 0xF);
 986                         sdp_data->dst_addr.ip4.addr = ip4_addr;
 987                         sdp_mask->dst_addr.ip4.addr = ~0;
 988                 } else {
 989                         cma_set_ip_ver(cma_data, 4);
 990                         cma_set_ip_ver(cma_mask, 0xF);
 991                         cma_data->dst_addr.ip4.addr = ip4_addr;
 992                         cma_mask->dst_addr.ip4.addr = ~0;
 993                 }
 994                 break;
 995         case AF_INET6:
 996                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
 997                 if (ps == RDMA_PS_SDP) {
 998                         sdp_set_ip_ver(sdp_data, 6);
 999                         sdp_set_ip_ver(sdp_mask, 0xF);
1000                         sdp_data->dst_addr.ip6 = ip6_addr;
1001                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1002                                sizeof sdp_mask->dst_addr.ip6);
1003                 } else {
1004                         cma_set_ip_ver(cma_data, 6);
1005                         cma_set_ip_ver(cma_mask, 0xF);
1006                         cma_data->dst_addr.ip6 = ip6_addr;
1007                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1008                                sizeof cma_mask->dst_addr.ip6);
1009                 }
1010                 break;
1011         default:
1012                 break;
1013         }
1014 }
1015
1016 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1017 {
1018         struct rdma_id_private *id_priv = iw_id->context;
1019         enum rdma_cm_event_type event = 0;
1020         struct sockaddr_in *sin;
1021         int ret = 0;
1022
1023         atomic_inc(&id_priv->dev_remove);
1024
1025         switch (iw_event->event) {
1026         case IW_CM_EVENT_CLOSE:
1027                 event = RDMA_CM_EVENT_DISCONNECTED;
1028                 break;
1029         case IW_CM_EVENT_CONNECT_REPLY:
1030                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1031                 *sin = iw_event->local_addr;
1032                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1033                 *sin = iw_event->remote_addr;
1034                 if (iw_event->status)
1035                         event = RDMA_CM_EVENT_REJECTED;
1036                 else
1037                         event = RDMA_CM_EVENT_ESTABLISHED;
1038                 break;
1039         case IW_CM_EVENT_ESTABLISHED:
1040                 event = RDMA_CM_EVENT_ESTABLISHED;
1041                 break;
1042         default:
1043                 BUG_ON(1);
1044         }
1045
1046         ret = cma_notify_user(id_priv, event, iw_event->status,
1047                               iw_event->private_data,
1048                               iw_event->private_data_len);
1049         if (ret) {
1050                 /* Destroy the CM ID by returning a non-zero value. */
1051                 id_priv->cm_id.iw = NULL;
1052                 cma_exch(id_priv, CMA_DESTROYING);
1053                 cma_release_remove(id_priv);
1054                 rdma_destroy_id(&id_priv->id);
1055                 return ret;
1056         }
1057
1058         cma_release_remove(id_priv);
1059         return ret;
1060 }
1061
1062 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1063                                struct iw_cm_event *iw_event)
1064 {
1065         struct rdma_cm_id *new_cm_id;
1066         struct rdma_id_private *listen_id, *conn_id;
1067         struct sockaddr_in *sin;
1068         struct net_device *dev = NULL;
1069         int ret;
1070
1071         listen_id = cm_id->context;
1072         atomic_inc(&listen_id->dev_remove);
1073         if (!cma_comp(listen_id, CMA_LISTEN)) {
1074                 ret = -ECONNABORTED;
1075                 goto out;
1076         }
1077
1078         /* Create a new RDMA id for the new IW CM ID */
1079         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1080                                    listen_id->id.context,
1081                                    RDMA_PS_TCP);
1082         if (!new_cm_id) {
1083                 ret = -ENOMEM;
1084                 goto out;
1085         }
1086         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1087         atomic_inc(&conn_id->dev_remove);
1088         conn_id->state = CMA_CONNECT;
1089
1090         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1091         if (!dev) {
1092                 ret = -EADDRNOTAVAIL;
1093                 cma_release_remove(conn_id);
1094                 rdma_destroy_id(new_cm_id);
1095                 goto out;
1096         }
1097         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1098         if (ret) {
1099                 cma_release_remove(conn_id);
1100                 rdma_destroy_id(new_cm_id);
1101                 goto out;
1102         }
1103
1104         mutex_lock(&lock);
1105         ret = cma_acquire_dev(conn_id);
1106         mutex_unlock(&lock);
1107         if (ret) {
1108                 cma_release_remove(conn_id);
1109                 rdma_destroy_id(new_cm_id);
1110                 goto out;
1111         }
1112
1113         conn_id->cm_id.iw = cm_id;
1114         cm_id->context = conn_id;
1115         cm_id->cm_handler = cma_iw_handler;
1116
1117         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1118         *sin = iw_event->local_addr;
1119         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1120         *sin = iw_event->remote_addr;
1121
1122         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
1123                               iw_event->private_data,
1124                               iw_event->private_data_len);
1125         if (ret) {
1126                 /* User wants to destroy the CM ID */
1127                 conn_id->cm_id.iw = NULL;
1128                 cma_exch(conn_id, CMA_DESTROYING);
1129                 cma_release_remove(conn_id);
1130                 rdma_destroy_id(&conn_id->id);
1131         }
1132
1133 out:
1134         if (dev)
1135                 dev_put(dev);
1136         cma_release_remove(listen_id);
1137         return ret;
1138 }
1139
1140 static int cma_ib_listen(struct rdma_id_private *id_priv)
1141 {
1142         struct ib_cm_compare_data compare_data;
1143         struct sockaddr *addr;
1144         __be64 svc_id;
1145         int ret;
1146
1147         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1148                                             id_priv);
1149         if (IS_ERR(id_priv->cm_id.ib))
1150                 return PTR_ERR(id_priv->cm_id.ib);
1151
1152         addr = &id_priv->id.route.addr.src_addr;
1153         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1154         if (cma_any_addr(addr))
1155                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1156         else {
1157                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1158                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1159         }
1160
1161         if (ret) {
1162                 ib_destroy_cm_id(id_priv->cm_id.ib);
1163                 id_priv->cm_id.ib = NULL;
1164         }
1165
1166         return ret;
1167 }
1168
1169 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1170 {
1171         int ret;
1172         struct sockaddr_in *sin;
1173
1174         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1175                                             iw_conn_req_handler,
1176                                             id_priv);
1177         if (IS_ERR(id_priv->cm_id.iw))
1178                 return PTR_ERR(id_priv->cm_id.iw);
1179
1180         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1181         id_priv->cm_id.iw->local_addr = *sin;
1182
1183         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1184
1185         if (ret) {
1186                 iw_destroy_cm_id(id_priv->cm_id.iw);
1187                 id_priv->cm_id.iw = NULL;
1188         }
1189
1190         return ret;
1191 }
1192
1193 static int cma_listen_handler(struct rdma_cm_id *id,
1194                               struct rdma_cm_event *event)
1195 {
1196         struct rdma_id_private *id_priv = id->context;
1197
1198         id->context = id_priv->id.context;
1199         id->event_handler = id_priv->id.event_handler;
1200         return id_priv->id.event_handler(id, event);
1201 }
1202
1203 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1204                               struct cma_device *cma_dev)
1205 {
1206         struct rdma_id_private *dev_id_priv;
1207         struct rdma_cm_id *id;
1208         int ret;
1209
1210         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1211         if (IS_ERR(id))
1212                 return;
1213
1214         dev_id_priv = container_of(id, struct rdma_id_private, id);
1215
1216         dev_id_priv->state = CMA_ADDR_BOUND;
1217         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1218                ip_addr_size(&id_priv->id.route.addr.src_addr));
1219
1220         cma_attach_to_dev(dev_id_priv, cma_dev);
1221         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1222
1223         ret = rdma_listen(id, id_priv->backlog);
1224         if (ret)
1225                 goto err;
1226
1227         return;
1228 err:
1229         cma_destroy_listen(dev_id_priv);
1230 }
1231
1232 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1233 {
1234         struct cma_device *cma_dev;
1235
1236         mutex_lock(&lock);
1237         list_add_tail(&id_priv->list, &listen_any_list);
1238         list_for_each_entry(cma_dev, &dev_list, list)
1239                 cma_listen_on_dev(id_priv, cma_dev);
1240         mutex_unlock(&lock);
1241 }
1242
1243 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1244 {
1245         struct sockaddr_in addr_in;
1246
1247         memset(&addr_in, 0, sizeof addr_in);
1248         addr_in.sin_family = af;
1249         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1250 }
1251
1252 int rdma_listen(struct rdma_cm_id *id, int backlog)
1253 {
1254         struct rdma_id_private *id_priv;
1255         int ret;
1256
1257         id_priv = container_of(id, struct rdma_id_private, id);
1258         if (id_priv->state == CMA_IDLE) {
1259                 ret = cma_bind_any(id, AF_INET);
1260                 if (ret)
1261                         return ret;
1262         }
1263
1264         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1265                 return -EINVAL;
1266
1267         id_priv->backlog = backlog;
1268         if (id->device) {
1269                 switch (rdma_node_get_transport(id->device->node_type)) {
1270                 case RDMA_TRANSPORT_IB:
1271                         ret = cma_ib_listen(id_priv);
1272                         if (ret)
1273                                 goto err;
1274                         break;
1275                 case RDMA_TRANSPORT_IWARP:
1276                         ret = cma_iw_listen(id_priv, backlog);
1277                         if (ret)
1278                                 goto err;
1279                         break;
1280                 default:
1281                         ret = -ENOSYS;
1282                         goto err;
1283                 }
1284         } else
1285                 cma_listen_on_all(id_priv);
1286
1287         return 0;
1288 err:
1289         id_priv->backlog = 0;
1290         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1291         return ret;
1292 }
1293 EXPORT_SYMBOL(rdma_listen);
1294
1295 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1296                               void *context)
1297 {
1298         struct cma_work *work = context;
1299         struct rdma_route *route;
1300
1301         route = &work->id->id.route;
1302
1303         if (!status) {
1304                 route->num_paths = 1;
1305                 *route->path_rec = *path_rec;
1306         } else {
1307                 work->old_state = CMA_ROUTE_QUERY;
1308                 work->new_state = CMA_ADDR_RESOLVED;
1309                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1310         }
1311
1312         queue_work(cma_wq, &work->work);
1313 }
1314
1315 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1316                               struct cma_work *work)
1317 {
1318         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1319         struct ib_sa_path_rec path_rec;
1320
1321         memset(&path_rec, 0, sizeof path_rec);
1322         ib_addr_get_sgid(addr, &path_rec.sgid);
1323         ib_addr_get_dgid(addr, &path_rec.dgid);
1324         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1325         path_rec.numb_path = 1;
1326
1327         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1328                                 id_priv->id.port_num, &path_rec,
1329                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1330                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
1331                                 timeout_ms, GFP_KERNEL,
1332                                 cma_query_handler, work, &id_priv->query);
1333
1334         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1335 }
1336
1337 static void cma_work_handler(void *data)
1338 {
1339         struct cma_work *work = data;
1340         struct rdma_id_private *id_priv = work->id;
1341         int destroy = 0;
1342
1343         atomic_inc(&id_priv->dev_remove);
1344         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1345                 goto out;
1346
1347         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1348                 cma_exch(id_priv, CMA_DESTROYING);
1349                 destroy = 1;
1350         }
1351 out:
1352         cma_release_remove(id_priv);
1353         cma_deref_id(id_priv);
1354         if (destroy)
1355                 rdma_destroy_id(&id_priv->id);
1356         kfree(work);
1357 }
1358
1359 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1360 {
1361         struct rdma_route *route = &id_priv->id.route;
1362         struct cma_work *work;
1363         int ret;
1364
1365         work = kzalloc(sizeof *work, GFP_KERNEL);
1366         if (!work)
1367                 return -ENOMEM;
1368
1369         work->id = id_priv;
1370         INIT_WORK(&work->work, cma_work_handler, work);
1371         work->old_state = CMA_ROUTE_QUERY;
1372         work->new_state = CMA_ROUTE_RESOLVED;
1373         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1374
1375         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1376         if (!route->path_rec) {
1377                 ret = -ENOMEM;
1378                 goto err1;
1379         }
1380
1381         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1382         if (ret)
1383                 goto err2;
1384
1385         return 0;
1386 err2:
1387         kfree(route->path_rec);
1388         route->path_rec = NULL;
1389 err1:
1390         kfree(work);
1391         return ret;
1392 }
1393
1394 int rdma_set_ib_paths(struct rdma_cm_id *id,
1395                       struct ib_sa_path_rec *path_rec, int num_paths)
1396 {
1397         struct rdma_id_private *id_priv;
1398         int ret;
1399
1400         id_priv = container_of(id, struct rdma_id_private, id);
1401         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1402                 return -EINVAL;
1403
1404         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1405         if (!id->route.path_rec) {
1406                 ret = -ENOMEM;
1407                 goto err;
1408         }
1409
1410         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1411         return 0;
1412 err:
1413         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1414         return ret;
1415 }
1416 EXPORT_SYMBOL(rdma_set_ib_paths);
1417
1418 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1419 {
1420         struct cma_work *work;
1421
1422         work = kzalloc(sizeof *work, GFP_KERNEL);
1423         if (!work)
1424                 return -ENOMEM;
1425
1426         work->id = id_priv;
1427         INIT_WORK(&work->work, cma_work_handler, work);
1428         work->old_state = CMA_ROUTE_QUERY;
1429         work->new_state = CMA_ROUTE_RESOLVED;
1430         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1431         queue_work(cma_wq, &work->work);
1432         return 0;
1433 }
1434
1435 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1436 {
1437         struct rdma_id_private *id_priv;
1438         int ret;
1439
1440         id_priv = container_of(id, struct rdma_id_private, id);
1441         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1442                 return -EINVAL;
1443
1444         atomic_inc(&id_priv->refcount);
1445         switch (rdma_node_get_transport(id->device->node_type)) {
1446         case RDMA_TRANSPORT_IB:
1447                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1448                 break;
1449         case RDMA_TRANSPORT_IWARP:
1450                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1451                 break;
1452         default:
1453                 ret = -ENOSYS;
1454                 break;
1455         }
1456         if (ret)
1457                 goto err;
1458
1459         return 0;
1460 err:
1461         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1462         cma_deref_id(id_priv);
1463         return ret;
1464 }
1465 EXPORT_SYMBOL(rdma_resolve_route);
1466
1467 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1468 {
1469         struct cma_device *cma_dev;
1470         struct ib_port_attr port_attr;
1471         union ib_gid gid;
1472         u16 pkey;
1473         int ret;
1474         u8 p;
1475
1476         mutex_lock(&lock);
1477         list_for_each_entry(cma_dev, &dev_list, list)
1478                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1479                         if (!ib_query_port (cma_dev->device, p, &port_attr) &&
1480                             port_attr.state == IB_PORT_ACTIVE)
1481                                 goto port_found;
1482
1483         if (!list_empty(&dev_list)) {
1484                 p = 1;
1485                 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1486         } else {
1487                 ret = -ENODEV;
1488                 goto out;
1489         }
1490
1491 port_found:
1492         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1493         if (ret)
1494                 goto out;
1495
1496         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1497         if (ret)
1498                 goto out;
1499
1500         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1501         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1502         id_priv->id.port_num = p;
1503         cma_attach_to_dev(id_priv, cma_dev);
1504 out:
1505         mutex_unlock(&lock);
1506         return ret;
1507 }
1508
1509 static void addr_handler(int status, struct sockaddr *src_addr,
1510                          struct rdma_dev_addr *dev_addr, void *context)
1511 {
1512         struct rdma_id_private *id_priv = context;
1513         enum rdma_cm_event_type event;
1514
1515         atomic_inc(&id_priv->dev_remove);
1516
1517         /*
1518          * Grab mutex to block rdma_destroy_id() from removing the device while
1519          * we're trying to acquire it.
1520          */
1521         mutex_lock(&lock);
1522         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1523                 mutex_unlock(&lock);
1524                 goto out;
1525         }
1526
1527         if (!status && !id_priv->cma_dev)
1528                 status = cma_acquire_dev(id_priv);
1529         mutex_unlock(&lock);
1530
1531         if (status) {
1532                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1533                         goto out;
1534                 event = RDMA_CM_EVENT_ADDR_ERROR;
1535         } else {
1536                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1537                        ip_addr_size(src_addr));
1538                 event = RDMA_CM_EVENT_ADDR_RESOLVED;
1539         }
1540
1541         if (cma_notify_user(id_priv, event, status, NULL, 0)) {
1542                 cma_exch(id_priv, CMA_DESTROYING);
1543                 cma_release_remove(id_priv);
1544                 cma_deref_id(id_priv);
1545                 rdma_destroy_id(&id_priv->id);
1546                 return;
1547         }
1548 out:
1549         cma_release_remove(id_priv);
1550         cma_deref_id(id_priv);
1551 }
1552
1553 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1554 {
1555         struct cma_work *work;
1556         struct sockaddr_in *src_in, *dst_in;
1557         union ib_gid gid;
1558         int ret;
1559
1560         work = kzalloc(sizeof *work, GFP_KERNEL);
1561         if (!work)
1562                 return -ENOMEM;
1563
1564         if (!id_priv->cma_dev) {
1565                 ret = cma_bind_loopback(id_priv);
1566                 if (ret)
1567                         goto err;
1568         }
1569
1570         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1571         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1572
1573         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1574                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1575                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1576                 src_in->sin_family = dst_in->sin_family;
1577                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1578         }
1579
1580         work->id = id_priv;
1581         INIT_WORK(&work->work, cma_work_handler, work);
1582         work->old_state = CMA_ADDR_QUERY;
1583         work->new_state = CMA_ADDR_RESOLVED;
1584         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1585         queue_work(cma_wq, &work->work);
1586         return 0;
1587 err:
1588         kfree(work);
1589         return ret;
1590 }
1591
1592 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1593                          struct sockaddr *dst_addr)
1594 {
1595         if (src_addr && src_addr->sa_family)
1596                 return rdma_bind_addr(id, src_addr);
1597         else
1598                 return cma_bind_any(id, dst_addr->sa_family);
1599 }
1600
1601 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1602                       struct sockaddr *dst_addr, int timeout_ms)
1603 {
1604         struct rdma_id_private *id_priv;
1605         int ret;
1606
1607         id_priv = container_of(id, struct rdma_id_private, id);
1608         if (id_priv->state == CMA_IDLE) {
1609                 ret = cma_bind_addr(id, src_addr, dst_addr);
1610                 if (ret)
1611                         return ret;
1612         }
1613
1614         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1615                 return -EINVAL;
1616
1617         atomic_inc(&id_priv->refcount);
1618         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1619         if (cma_any_addr(dst_addr))
1620                 ret = cma_resolve_loopback(id_priv);
1621         else
1622                 ret = rdma_resolve_ip(&id->route.addr.src_addr, dst_addr,
1623                                       &id->route.addr.dev_addr,
1624                                       timeout_ms, addr_handler, id_priv);
1625         if (ret)
1626                 goto err;
1627
1628         return 0;
1629 err:
1630         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1631         cma_deref_id(id_priv);
1632         return ret;
1633 }
1634 EXPORT_SYMBOL(rdma_resolve_addr);
1635
1636 static void cma_bind_port(struct rdma_bind_list *bind_list,
1637                           struct rdma_id_private *id_priv)
1638 {
1639         struct sockaddr_in *sin;
1640
1641         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1642         sin->sin_port = htons(bind_list->port);
1643         id_priv->bind_list = bind_list;
1644         hlist_add_head(&id_priv->node, &bind_list->owners);
1645 }
1646
1647 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1648                           unsigned short snum)
1649 {
1650         struct rdma_bind_list *bind_list;
1651         int port, start, ret;
1652
1653         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1654         if (!bind_list)
1655                 return -ENOMEM;
1656
1657         start = snum ? snum : sysctl_local_port_range[0];
1658
1659         do {
1660                 ret = idr_get_new_above(ps, bind_list, start, &port);
1661         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1662
1663         if (ret)
1664                 goto err;
1665
1666         if ((snum && port != snum) ||
1667             (!snum && port > sysctl_local_port_range[1])) {
1668                 idr_remove(ps, port);
1669                 ret = -EADDRNOTAVAIL;
1670                 goto err;
1671         }
1672
1673         bind_list->ps = ps;
1674         bind_list->port = (unsigned short) port;
1675         cma_bind_port(bind_list, id_priv);
1676         return 0;
1677 err:
1678         kfree(bind_list);
1679         return ret;
1680 }
1681
1682 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1683 {
1684         struct rdma_id_private *cur_id;
1685         struct sockaddr_in *sin, *cur_sin;
1686         struct rdma_bind_list *bind_list;
1687         struct hlist_node *node;
1688         unsigned short snum;
1689
1690         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1691         snum = ntohs(sin->sin_port);
1692         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1693                 return -EACCES;
1694
1695         bind_list = idr_find(ps, snum);
1696         if (!bind_list)
1697                 return cma_alloc_port(ps, id_priv, snum);
1698
1699         /*
1700          * We don't support binding to any address if anyone is bound to
1701          * a specific address on the same port.
1702          */
1703         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1704                 return -EADDRNOTAVAIL;
1705
1706         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1707                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1708                         return -EADDRNOTAVAIL;
1709
1710                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1711                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1712                         return -EADDRINUSE;
1713         }
1714
1715         cma_bind_port(bind_list, id_priv);
1716         return 0;
1717 }
1718
1719 static int cma_get_port(struct rdma_id_private *id_priv)
1720 {
1721         struct idr *ps;
1722         int ret;
1723
1724         switch (id_priv->id.ps) {
1725         case RDMA_PS_SDP:
1726                 ps = &sdp_ps;
1727                 break;
1728         case RDMA_PS_TCP:
1729                 ps = &tcp_ps;
1730                 break;
1731         default:
1732                 return -EPROTONOSUPPORT;
1733         }
1734
1735         mutex_lock(&lock);
1736         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1737                 ret = cma_alloc_port(ps, id_priv, 0);
1738         else
1739                 ret = cma_use_port(ps, id_priv);
1740         mutex_unlock(&lock);
1741
1742         return ret;
1743 }
1744
1745 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1746 {
1747         struct rdma_id_private *id_priv;
1748         int ret;
1749
1750         if (addr->sa_family != AF_INET)
1751                 return -EAFNOSUPPORT;
1752
1753         id_priv = container_of(id, struct rdma_id_private, id);
1754         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1755                 return -EINVAL;
1756
1757         if (!cma_any_addr(addr)) {
1758                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1759                 if (!ret) {
1760                         mutex_lock(&lock);
1761                         ret = cma_acquire_dev(id_priv);
1762                         mutex_unlock(&lock);
1763                 }
1764                 if (ret)
1765                         goto err;
1766         }
1767
1768         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
1769         ret = cma_get_port(id_priv);
1770         if (ret)
1771                 goto err;
1772
1773         return 0;
1774 err:
1775         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
1776         return ret;
1777 }
1778 EXPORT_SYMBOL(rdma_bind_addr);
1779
1780 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
1781                           struct rdma_route *route)
1782 {
1783         struct sockaddr_in *src4, *dst4;
1784         struct cma_hdr *cma_hdr;
1785         struct sdp_hh *sdp_hdr;
1786
1787         src4 = (struct sockaddr_in *) &route->addr.src_addr;
1788         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
1789
1790         switch (ps) {
1791         case RDMA_PS_SDP:
1792                 sdp_hdr = hdr;
1793                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
1794                         return -EINVAL;
1795                 sdp_set_ip_ver(sdp_hdr, 4);
1796                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1797                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1798                 sdp_hdr->port = src4->sin_port;
1799                 break;
1800         default:
1801                 cma_hdr = hdr;
1802                 cma_hdr->cma_version = CMA_VERSION;
1803                 cma_set_ip_ver(cma_hdr, 4);
1804                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1805                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1806                 cma_hdr->port = src4->sin_port;
1807                 break;
1808         }
1809         return 0;
1810 }
1811
1812 static int cma_connect_ib(struct rdma_id_private *id_priv,
1813                           struct rdma_conn_param *conn_param)
1814 {
1815         struct ib_cm_req_param req;
1816         struct rdma_route *route;
1817         void *private_data;
1818         int offset, ret;
1819
1820         memset(&req, 0, sizeof req);
1821         offset = cma_user_data_offset(id_priv->id.ps);
1822         req.private_data_len = offset + conn_param->private_data_len;
1823         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
1824         if (!private_data)
1825                 return -ENOMEM;
1826
1827         if (conn_param->private_data && conn_param->private_data_len)
1828                 memcpy(private_data + offset, conn_param->private_data,
1829                        conn_param->private_data_len);
1830
1831         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
1832                                             id_priv);
1833         if (IS_ERR(id_priv->cm_id.ib)) {
1834                 ret = PTR_ERR(id_priv->cm_id.ib);
1835                 goto out;
1836         }
1837
1838         route = &id_priv->id.route;
1839         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
1840         if (ret)
1841                 goto out;
1842         req.private_data = private_data;
1843
1844         req.primary_path = &route->path_rec[0];
1845         if (route->num_paths == 2)
1846                 req.alternate_path = &route->path_rec[1];
1847
1848         req.service_id = cma_get_service_id(id_priv->id.ps,
1849                                             &route->addr.dst_addr);
1850         req.qp_num = id_priv->qp_num;
1851         req.qp_type = id_priv->qp_type;
1852         req.starting_psn = id_priv->seq_num;
1853         req.responder_resources = conn_param->responder_resources;
1854         req.initiator_depth = conn_param->initiator_depth;
1855         req.flow_control = conn_param->flow_control;
1856         req.retry_count = conn_param->retry_count;
1857         req.rnr_retry_count = conn_param->rnr_retry_count;
1858         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1859         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1860         req.max_cm_retries = CMA_MAX_CM_RETRIES;
1861         req.srq = id_priv->srq ? 1 : 0;
1862
1863         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
1864 out:
1865         kfree(private_data);
1866         return ret;
1867 }
1868
1869 static int cma_connect_iw(struct rdma_id_private *id_priv,
1870                           struct rdma_conn_param *conn_param)
1871 {
1872         struct iw_cm_id *cm_id;
1873         struct sockaddr_in* sin;
1874         int ret;
1875         struct iw_cm_conn_param iw_param;
1876
1877         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
1878         if (IS_ERR(cm_id)) {
1879                 ret = PTR_ERR(cm_id);
1880                 goto out;
1881         }
1882
1883         id_priv->cm_id.iw = cm_id;
1884
1885         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
1886         cm_id->local_addr = *sin;
1887
1888         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
1889         cm_id->remote_addr = *sin;
1890
1891         ret = cma_modify_qp_rtr(&id_priv->id);
1892         if (ret) {
1893                 iw_destroy_cm_id(cm_id);
1894                 return ret;
1895         }
1896
1897         iw_param.ord = conn_param->initiator_depth;
1898         iw_param.ird = conn_param->responder_resources;
1899         iw_param.private_data = conn_param->private_data;
1900         iw_param.private_data_len = conn_param->private_data_len;
1901         if (id_priv->id.qp)
1902                 iw_param.qpn = id_priv->qp_num;
1903         else
1904                 iw_param.qpn = conn_param->qp_num;
1905         ret = iw_cm_connect(cm_id, &iw_param);
1906 out:
1907         return ret;
1908 }
1909
1910 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1911 {
1912         struct rdma_id_private *id_priv;
1913         int ret;
1914
1915         id_priv = container_of(id, struct rdma_id_private, id);
1916         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
1917                 return -EINVAL;
1918
1919         if (!id->qp) {
1920                 id_priv->qp_num = conn_param->qp_num;
1921                 id_priv->qp_type = conn_param->qp_type;
1922                 id_priv->srq = conn_param->srq;
1923         }
1924
1925         switch (rdma_node_get_transport(id->device->node_type)) {
1926         case RDMA_TRANSPORT_IB:
1927                 ret = cma_connect_ib(id_priv, conn_param);
1928                 break;
1929         case RDMA_TRANSPORT_IWARP:
1930                 ret = cma_connect_iw(id_priv, conn_param);
1931                 break;
1932         default:
1933                 ret = -ENOSYS;
1934                 break;
1935         }
1936         if (ret)
1937                 goto err;
1938
1939         return 0;
1940 err:
1941         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
1942         return ret;
1943 }
1944 EXPORT_SYMBOL(rdma_connect);
1945
1946 static int cma_accept_ib(struct rdma_id_private *id_priv,
1947                          struct rdma_conn_param *conn_param)
1948 {
1949         struct ib_cm_rep_param rep;
1950         int ret;
1951
1952         ret = cma_modify_qp_rtr(&id_priv->id);
1953         if (ret)
1954                 return ret;
1955
1956         memset(&rep, 0, sizeof rep);
1957         rep.qp_num = id_priv->qp_num;
1958         rep.starting_psn = id_priv->seq_num;
1959         rep.private_data = conn_param->private_data;
1960         rep.private_data_len = conn_param->private_data_len;
1961         rep.responder_resources = conn_param->responder_resources;
1962         rep.initiator_depth = conn_param->initiator_depth;
1963         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
1964         rep.failover_accepted = 0;
1965         rep.flow_control = conn_param->flow_control;
1966         rep.rnr_retry_count = conn_param->rnr_retry_count;
1967         rep.srq = id_priv->srq ? 1 : 0;
1968
1969         return ib_send_cm_rep(id_priv->cm_id.ib, &rep);
1970 }
1971
1972 static int cma_accept_iw(struct rdma_id_private *id_priv,
1973                   struct rdma_conn_param *conn_param)
1974 {
1975         struct iw_cm_conn_param iw_param;
1976         int ret;
1977
1978         ret = cma_modify_qp_rtr(&id_priv->id);
1979         if (ret)
1980                 return ret;
1981
1982         iw_param.ord = conn_param->initiator_depth;
1983         iw_param.ird = conn_param->responder_resources;
1984         iw_param.private_data = conn_param->private_data;
1985         iw_param.private_data_len = conn_param->private_data_len;
1986         if (id_priv->id.qp) {
1987                 iw_param.qpn = id_priv->qp_num;
1988         } else
1989                 iw_param.qpn = conn_param->qp_num;
1990
1991         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
1992 }
1993
1994 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1995 {
1996         struct rdma_id_private *id_priv;
1997         int ret;
1998
1999         id_priv = container_of(id, struct rdma_id_private, id);
2000         if (!cma_comp(id_priv, CMA_CONNECT))
2001                 return -EINVAL;
2002
2003         if (!id->qp && conn_param) {
2004                 id_priv->qp_num = conn_param->qp_num;
2005                 id_priv->qp_type = conn_param->qp_type;
2006                 id_priv->srq = conn_param->srq;
2007         }
2008
2009         switch (rdma_node_get_transport(id->device->node_type)) {
2010         case RDMA_TRANSPORT_IB:
2011                 if (conn_param)
2012                         ret = cma_accept_ib(id_priv, conn_param);
2013                 else
2014                         ret = cma_rep_recv(id_priv);
2015                 break;
2016         case RDMA_TRANSPORT_IWARP:
2017                 ret = cma_accept_iw(id_priv, conn_param);
2018                 break;
2019         default:
2020                 ret = -ENOSYS;
2021                 break;
2022         }
2023
2024         if (ret)
2025                 goto reject;
2026
2027         return 0;
2028 reject:
2029         cma_modify_qp_err(id);
2030         rdma_reject(id, NULL, 0);
2031         return ret;
2032 }
2033 EXPORT_SYMBOL(rdma_accept);
2034
2035 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2036                 u8 private_data_len)
2037 {
2038         struct rdma_id_private *id_priv;
2039         int ret;
2040
2041         id_priv = container_of(id, struct rdma_id_private, id);
2042         if (!cma_comp(id_priv, CMA_CONNECT))
2043                 return -EINVAL;
2044
2045         switch (rdma_node_get_transport(id->device->node_type)) {
2046         case RDMA_TRANSPORT_IB:
2047                 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2048                                      IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2049                                      private_data, private_data_len);
2050                 break;
2051         case RDMA_TRANSPORT_IWARP:
2052                 ret = iw_cm_reject(id_priv->cm_id.iw,
2053                                    private_data, private_data_len);
2054                 break;
2055         default:
2056                 ret = -ENOSYS;
2057                 break;
2058         }
2059         return ret;
2060 }
2061 EXPORT_SYMBOL(rdma_reject);
2062
2063 int rdma_disconnect(struct rdma_cm_id *id)
2064 {
2065         struct rdma_id_private *id_priv;
2066         int ret;
2067
2068         id_priv = container_of(id, struct rdma_id_private, id);
2069         if (!cma_comp(id_priv, CMA_CONNECT) &&
2070             !cma_comp(id_priv, CMA_DISCONNECT))
2071                 return -EINVAL;
2072
2073         switch (rdma_node_get_transport(id->device->node_type)) {
2074         case RDMA_TRANSPORT_IB:
2075                 ret = cma_modify_qp_err(id);
2076                 if (ret)
2077                         goto out;
2078                 /* Initiate or respond to a disconnect. */
2079                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2080                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2081                 break;
2082         case RDMA_TRANSPORT_IWARP:
2083                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2084                 break;
2085         default:
2086                 ret = -EINVAL;
2087                 break;
2088         }
2089 out:
2090         return ret;
2091 }
2092 EXPORT_SYMBOL(rdma_disconnect);
2093
2094 static void cma_add_one(struct ib_device *device)
2095 {
2096         struct cma_device *cma_dev;
2097         struct rdma_id_private *id_priv;
2098
2099         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2100         if (!cma_dev)
2101                 return;
2102
2103         cma_dev->device = device;
2104         cma_dev->node_guid = device->node_guid;
2105         if (!cma_dev->node_guid)
2106                 goto err;
2107
2108         init_completion(&cma_dev->comp);
2109         atomic_set(&cma_dev->refcount, 1);
2110         INIT_LIST_HEAD(&cma_dev->id_list);
2111         ib_set_client_data(device, &cma_client, cma_dev);
2112
2113         mutex_lock(&lock);
2114         list_add_tail(&cma_dev->list, &dev_list);
2115         list_for_each_entry(id_priv, &listen_any_list, list)
2116                 cma_listen_on_dev(id_priv, cma_dev);
2117         mutex_unlock(&lock);
2118         return;
2119 err:
2120         kfree(cma_dev);
2121 }
2122
2123 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2124 {
2125         enum cma_state state;
2126
2127         /* Record that we want to remove the device */
2128         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2129         if (state == CMA_DESTROYING)
2130                 return 0;
2131
2132         cma_cancel_operation(id_priv, state);
2133         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2134
2135         /* Check for destruction from another callback. */
2136         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2137                 return 0;
2138
2139         return cma_notify_user(id_priv, RDMA_CM_EVENT_DEVICE_REMOVAL,
2140                                0, NULL, 0);
2141 }
2142
2143 static void cma_process_remove(struct cma_device *cma_dev)
2144 {
2145         struct list_head remove_list;
2146         struct rdma_id_private *id_priv;
2147         int ret;
2148
2149         INIT_LIST_HEAD(&remove_list);
2150
2151         mutex_lock(&lock);
2152         while (!list_empty(&cma_dev->id_list)) {
2153                 id_priv = list_entry(cma_dev->id_list.next,
2154                                      struct rdma_id_private, list);
2155
2156                 if (cma_internal_listen(id_priv)) {
2157                         cma_destroy_listen(id_priv);
2158                         continue;
2159                 }
2160
2161                 list_del(&id_priv->list);
2162                 list_add_tail(&id_priv->list, &remove_list);
2163                 atomic_inc(&id_priv->refcount);
2164                 mutex_unlock(&lock);
2165
2166                 ret = cma_remove_id_dev(id_priv);
2167                 cma_deref_id(id_priv);
2168                 if (ret)
2169                         rdma_destroy_id(&id_priv->id);
2170
2171                 mutex_lock(&lock);
2172         }
2173         mutex_unlock(&lock);
2174
2175         cma_deref_dev(cma_dev);
2176         wait_for_completion(&cma_dev->comp);
2177 }
2178
2179 static void cma_remove_one(struct ib_device *device)
2180 {
2181         struct cma_device *cma_dev;
2182
2183         cma_dev = ib_get_client_data(device, &cma_client);
2184         if (!cma_dev)
2185                 return;
2186
2187         mutex_lock(&lock);
2188         list_del(&cma_dev->list);
2189         mutex_unlock(&lock);
2190
2191         cma_process_remove(cma_dev);
2192         kfree(cma_dev);
2193 }
2194
2195 static int cma_init(void)
2196 {
2197         int ret;
2198
2199         cma_wq = create_singlethread_workqueue("rdma_cm_wq");
2200         if (!cma_wq)
2201                 return -ENOMEM;
2202
2203         ib_sa_register_client(&sa_client);
2204
2205         ret = ib_register_client(&cma_client);
2206         if (ret)
2207                 goto err;
2208         return 0;
2209
2210 err:
2211         ib_sa_unregister_client(&sa_client);
2212         destroy_workqueue(cma_wq);
2213         return ret;
2214 }
2215
2216 static void cma_cleanup(void)
2217 {
2218         ib_unregister_client(&cma_client);
2219         ib_sa_unregister_client(&sa_client);
2220         destroy_workqueue(cma_wq);
2221         idr_destroy(&sdp_ps);
2222         idr_destroy(&tcp_ps);
2223 }
2224
2225 module_init(cma_init);
2226 module_exit(cma_cleanup);