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_exch(conn_id, CMA_DESTROYING);
 936                 cma_release_remove(conn_id);
 937                 rdma_destroy_id(&conn_id->id);
 938                 goto out;
 939         }
 940
 941         conn_id->cm_id.ib = cm_id;
 942         cm_id->context = conn_id;
 943         cm_id->cm_handler = cma_ib_handler;
 944
 945         offset = cma_user_data_offset(listen_id->id.ps);
 946         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
 947                               ib_event->private_data + offset,
 948                               IB_CM_REQ_PRIVATE_DATA_SIZE - offset);
 949         if (ret) {
 950                 /* Destroy the CM ID by returning a non-zero value. */
 951                 conn_id->cm_id.ib = NULL;
 952                 cma_exch(conn_id, CMA_DESTROYING);
 953                 cma_release_remove(conn_id);
 954                 rdma_destroy_id(&conn_id->id);
 955         }
 956 out:
 957         cma_release_remove(listen_id);
 958         return ret;
 959 }
 960
 961 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
 962 {
 963         return cpu_to_be64(((u64)ps << 16) +
 964                be16_to_cpu(((struct sockaddr_in *) addr)->sin_port));
 965 }
 966
 967 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
 968                                  struct ib_cm_compare_data *compare)
 969 {
 970         struct cma_hdr *cma_data, *cma_mask;
 971         struct sdp_hh *sdp_data, *sdp_mask;
 972         __u32 ip4_addr;
 973         struct in6_addr ip6_addr;
 974
 975         memset(compare, 0, sizeof *compare);
 976         cma_data = (void *) compare->data;
 977         cma_mask = (void *) compare->mask;
 978         sdp_data = (void *) compare->data;
 979         sdp_mask = (void *) compare->mask;
 980
 981         switch (addr->sa_family) {
 982         case AF_INET:
 983                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
 984                 if (ps == RDMA_PS_SDP) {
 985                         sdp_set_ip_ver(sdp_data, 4);
 986                         sdp_set_ip_ver(sdp_mask, 0xF);
 987                         sdp_data->dst_addr.ip4.addr = ip4_addr;
 988                         sdp_mask->dst_addr.ip4.addr = ~0;
 989                 } else {
 990                         cma_set_ip_ver(cma_data, 4);
 991                         cma_set_ip_ver(cma_mask, 0xF);
 992                         cma_data->dst_addr.ip4.addr = ip4_addr;
 993                         cma_mask->dst_addr.ip4.addr = ~0;
 994                 }
 995                 break;
 996         case AF_INET6:
 997                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
 998                 if (ps == RDMA_PS_SDP) {
 999                         sdp_set_ip_ver(sdp_data, 6);
1000                         sdp_set_ip_ver(sdp_mask, 0xF);
1001                         sdp_data->dst_addr.ip6 = ip6_addr;
1002                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1003                                sizeof sdp_mask->dst_addr.ip6);
1004                 } else {
1005                         cma_set_ip_ver(cma_data, 6);
1006                         cma_set_ip_ver(cma_mask, 0xF);
1007                         cma_data->dst_addr.ip6 = ip6_addr;
1008                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1009                                sizeof cma_mask->dst_addr.ip6);
1010                 }
1011                 break;
1012         default:
1013                 break;
1014         }
1015 }
1016
1017 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1018 {
1019         struct rdma_id_private *id_priv = iw_id->context;
1020         enum rdma_cm_event_type event = 0;
1021         struct sockaddr_in *sin;
1022         int ret = 0;
1023
1024         atomic_inc(&id_priv->dev_remove);
1025
1026         switch (iw_event->event) {
1027         case IW_CM_EVENT_CLOSE:
1028                 event = RDMA_CM_EVENT_DISCONNECTED;
1029                 break;
1030         case IW_CM_EVENT_CONNECT_REPLY:
1031                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1032                 *sin = iw_event->local_addr;
1033                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1034                 *sin = iw_event->remote_addr;
1035                 if (iw_event->status)
1036                         event = RDMA_CM_EVENT_REJECTED;
1037                 else
1038                         event = RDMA_CM_EVENT_ESTABLISHED;
1039                 break;
1040         case IW_CM_EVENT_ESTABLISHED:
1041                 event = RDMA_CM_EVENT_ESTABLISHED;
1042                 break;
1043         default:
1044                 BUG_ON(1);
1045         }
1046
1047         ret = cma_notify_user(id_priv, event, iw_event->status,
1048                               iw_event->private_data,
1049                               iw_event->private_data_len);
1050         if (ret) {
1051                 /* Destroy the CM ID by returning a non-zero value. */
1052                 id_priv->cm_id.iw = NULL;
1053                 cma_exch(id_priv, CMA_DESTROYING);
1054                 cma_release_remove(id_priv);
1055                 rdma_destroy_id(&id_priv->id);
1056                 return ret;
1057         }
1058
1059         cma_release_remove(id_priv);
1060         return ret;
1061 }
1062
1063 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1064                                struct iw_cm_event *iw_event)
1065 {
1066         struct rdma_cm_id *new_cm_id;
1067         struct rdma_id_private *listen_id, *conn_id;
1068         struct sockaddr_in *sin;
1069         struct net_device *dev = NULL;
1070         int ret;
1071
1072         listen_id = cm_id->context;
1073         atomic_inc(&listen_id->dev_remove);
1074         if (!cma_comp(listen_id, CMA_LISTEN)) {
1075                 ret = -ECONNABORTED;
1076                 goto out;
1077         }
1078
1079         /* Create a new RDMA id for the new IW CM ID */
1080         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1081                                    listen_id->id.context,
1082                                    RDMA_PS_TCP);
1083         if (!new_cm_id) {
1084                 ret = -ENOMEM;
1085                 goto out;
1086         }
1087         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1088         atomic_inc(&conn_id->dev_remove);
1089         conn_id->state = CMA_CONNECT;
1090
1091         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1092         if (!dev) {
1093                 ret = -EADDRNOTAVAIL;
1094                 cma_release_remove(conn_id);
1095                 rdma_destroy_id(new_cm_id);
1096                 goto out;
1097         }
1098         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1099         if (ret) {
1100                 cma_release_remove(conn_id);
1101                 rdma_destroy_id(new_cm_id);
1102                 goto out;
1103         }
1104
1105         mutex_lock(&lock);
1106         ret = cma_acquire_dev(conn_id);
1107         mutex_unlock(&lock);
1108         if (ret) {
1109                 cma_release_remove(conn_id);
1110                 rdma_destroy_id(new_cm_id);
1111                 goto out;
1112         }
1113
1114         conn_id->cm_id.iw = cm_id;
1115         cm_id->context = conn_id;
1116         cm_id->cm_handler = cma_iw_handler;
1117
1118         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1119         *sin = iw_event->local_addr;
1120         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1121         *sin = iw_event->remote_addr;
1122
1123         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
1124                               iw_event->private_data,
1125                               iw_event->private_data_len);
1126         if (ret) {
1127                 /* User wants to destroy the CM ID */
1128                 conn_id->cm_id.iw = NULL;
1129                 cma_exch(conn_id, CMA_DESTROYING);
1130                 cma_release_remove(conn_id);
1131                 rdma_destroy_id(&conn_id->id);
1132         }
1133
1134 out:
1135         if (dev)
1136                 dev_put(dev);
1137         cma_release_remove(listen_id);
1138         return ret;
1139 }
1140
1141 static int cma_ib_listen(struct rdma_id_private *id_priv)
1142 {
1143         struct ib_cm_compare_data compare_data;
1144         struct sockaddr *addr;
1145         __be64 svc_id;
1146         int ret;
1147
1148         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1149                                             id_priv);
1150         if (IS_ERR(id_priv->cm_id.ib))
1151                 return PTR_ERR(id_priv->cm_id.ib);
1152
1153         addr = &id_priv->id.route.addr.src_addr;
1154         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1155         if (cma_any_addr(addr))
1156                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1157         else {
1158                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1159                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1160         }
1161
1162         if (ret) {
1163                 ib_destroy_cm_id(id_priv->cm_id.ib);
1164                 id_priv->cm_id.ib = NULL;
1165         }
1166
1167         return ret;
1168 }
1169
1170 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1171 {
1172         int ret;
1173         struct sockaddr_in *sin;
1174
1175         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1176                                             iw_conn_req_handler,
1177                                             id_priv);
1178         if (IS_ERR(id_priv->cm_id.iw))
1179                 return PTR_ERR(id_priv->cm_id.iw);
1180
1181         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1182         id_priv->cm_id.iw->local_addr = *sin;
1183
1184         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1185
1186         if (ret) {
1187                 iw_destroy_cm_id(id_priv->cm_id.iw);
1188                 id_priv->cm_id.iw = NULL;
1189         }
1190
1191         return ret;
1192 }
1193
1194 static int cma_listen_handler(struct rdma_cm_id *id,
1195                               struct rdma_cm_event *event)
1196 {
1197         struct rdma_id_private *id_priv = id->context;
1198
1199         id->context = id_priv->id.context;
1200         id->event_handler = id_priv->id.event_handler;
1201         return id_priv->id.event_handler(id, event);
1202 }
1203
1204 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1205                               struct cma_device *cma_dev)
1206 {
1207         struct rdma_id_private *dev_id_priv;
1208         struct rdma_cm_id *id;
1209         int ret;
1210
1211         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1212         if (IS_ERR(id))
1213                 return;
1214
1215         dev_id_priv = container_of(id, struct rdma_id_private, id);
1216
1217         dev_id_priv->state = CMA_ADDR_BOUND;
1218         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1219                ip_addr_size(&id_priv->id.route.addr.src_addr));
1220
1221         cma_attach_to_dev(dev_id_priv, cma_dev);
1222         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1223
1224         ret = rdma_listen(id, id_priv->backlog);
1225         if (ret)
1226                 goto err;
1227
1228         return;
1229 err:
1230         cma_destroy_listen(dev_id_priv);
1231 }
1232
1233 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1234 {
1235         struct cma_device *cma_dev;
1236
1237         mutex_lock(&lock);
1238         list_add_tail(&id_priv->list, &listen_any_list);
1239         list_for_each_entry(cma_dev, &dev_list, list)
1240                 cma_listen_on_dev(id_priv, cma_dev);
1241         mutex_unlock(&lock);
1242 }
1243
1244 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1245 {
1246         struct sockaddr_in addr_in;
1247
1248         memset(&addr_in, 0, sizeof addr_in);
1249         addr_in.sin_family = af;
1250         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1251 }
1252
1253 int rdma_listen(struct rdma_cm_id *id, int backlog)
1254 {
1255         struct rdma_id_private *id_priv;
1256         int ret;
1257
1258         id_priv = container_of(id, struct rdma_id_private, id);
1259         if (id_priv->state == CMA_IDLE) {
1260                 ret = cma_bind_any(id, AF_INET);
1261                 if (ret)
1262                         return ret;
1263         }
1264
1265         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1266                 return -EINVAL;
1267
1268         id_priv->backlog = backlog;
1269         if (id->device) {
1270                 switch (rdma_node_get_transport(id->device->node_type)) {
1271                 case RDMA_TRANSPORT_IB:
1272                         ret = cma_ib_listen(id_priv);
1273                         if (ret)
1274                                 goto err;
1275                         break;
1276                 case RDMA_TRANSPORT_IWARP:
1277                         ret = cma_iw_listen(id_priv, backlog);
1278                         if (ret)
1279                                 goto err;
1280                         break;
1281                 default:
1282                         ret = -ENOSYS;
1283                         goto err;
1284                 }
1285         } else
1286                 cma_listen_on_all(id_priv);
1287
1288         return 0;
1289 err:
1290         id_priv->backlog = 0;
1291         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1292         return ret;
1293 }
1294 EXPORT_SYMBOL(rdma_listen);
1295
1296 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1297                               void *context)
1298 {
1299         struct cma_work *work = context;
1300         struct rdma_route *route;
1301
1302         route = &work->id->id.route;
1303
1304         if (!status) {
1305                 route->num_paths = 1;
1306                 *route->path_rec = *path_rec;
1307         } else {
1308                 work->old_state = CMA_ROUTE_QUERY;
1309                 work->new_state = CMA_ADDR_RESOLVED;
1310                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1311                 work->event.status = status;
1312         }
1313
1314         queue_work(cma_wq, &work->work);
1315 }
1316
1317 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1318                               struct cma_work *work)
1319 {
1320         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1321         struct ib_sa_path_rec path_rec;
1322
1323         memset(&path_rec, 0, sizeof path_rec);
1324         ib_addr_get_sgid(addr, &path_rec.sgid);
1325         ib_addr_get_dgid(addr, &path_rec.dgid);
1326         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1327         path_rec.numb_path = 1;
1328
1329         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1330                                 id_priv->id.port_num, &path_rec,
1331                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1332                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
1333                                 timeout_ms, GFP_KERNEL,
1334                                 cma_query_handler, work, &id_priv->query);
1335
1336         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1337 }
1338
1339 static void cma_work_handler(void *data)
1340 {
1341         struct cma_work *work = data;
1342         struct rdma_id_private *id_priv = work->id;
1343         int destroy = 0;
1344
1345         atomic_inc(&id_priv->dev_remove);
1346         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1347                 goto out;
1348
1349         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1350                 cma_exch(id_priv, CMA_DESTROYING);
1351                 destroy = 1;
1352         }
1353 out:
1354         cma_release_remove(id_priv);
1355         cma_deref_id(id_priv);
1356         if (destroy)
1357                 rdma_destroy_id(&id_priv->id);
1358         kfree(work);
1359 }
1360
1361 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1362 {
1363         struct rdma_route *route = &id_priv->id.route;
1364         struct cma_work *work;
1365         int ret;
1366
1367         work = kzalloc(sizeof *work, GFP_KERNEL);
1368         if (!work)
1369                 return -ENOMEM;
1370
1371         work->id = id_priv;
1372         INIT_WORK(&work->work, cma_work_handler, work);
1373         work->old_state = CMA_ROUTE_QUERY;
1374         work->new_state = CMA_ROUTE_RESOLVED;
1375         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1376
1377         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1378         if (!route->path_rec) {
1379                 ret = -ENOMEM;
1380                 goto err1;
1381         }
1382
1383         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1384         if (ret)
1385                 goto err2;
1386
1387         return 0;
1388 err2:
1389         kfree(route->path_rec);
1390         route->path_rec = NULL;
1391 err1:
1392         kfree(work);
1393         return ret;
1394 }
1395
1396 int rdma_set_ib_paths(struct rdma_cm_id *id,
1397                       struct ib_sa_path_rec *path_rec, int num_paths)
1398 {
1399         struct rdma_id_private *id_priv;
1400         int ret;
1401
1402         id_priv = container_of(id, struct rdma_id_private, id);
1403         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1404                 return -EINVAL;
1405
1406         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1407         if (!id->route.path_rec) {
1408                 ret = -ENOMEM;
1409                 goto err;
1410         }
1411
1412         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1413         return 0;
1414 err:
1415         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1416         return ret;
1417 }
1418 EXPORT_SYMBOL(rdma_set_ib_paths);
1419
1420 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1421 {
1422         struct cma_work *work;
1423
1424         work = kzalloc(sizeof *work, GFP_KERNEL);
1425         if (!work)
1426                 return -ENOMEM;
1427
1428         work->id = id_priv;
1429         INIT_WORK(&work->work, cma_work_handler, work);
1430         work->old_state = CMA_ROUTE_QUERY;
1431         work->new_state = CMA_ROUTE_RESOLVED;
1432         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1433         queue_work(cma_wq, &work->work);
1434         return 0;
1435 }
1436
1437 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1438 {
1439         struct rdma_id_private *id_priv;
1440         int ret;
1441
1442         id_priv = container_of(id, struct rdma_id_private, id);
1443         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1444                 return -EINVAL;
1445
1446         atomic_inc(&id_priv->refcount);
1447         switch (rdma_node_get_transport(id->device->node_type)) {
1448         case RDMA_TRANSPORT_IB:
1449                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1450                 break;
1451         case RDMA_TRANSPORT_IWARP:
1452                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1453                 break;
1454         default:
1455                 ret = -ENOSYS;
1456                 break;
1457         }
1458         if (ret)
1459                 goto err;
1460
1461         return 0;
1462 err:
1463         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1464         cma_deref_id(id_priv);
1465         return ret;
1466 }
1467 EXPORT_SYMBOL(rdma_resolve_route);
1468
1469 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1470 {
1471         struct cma_device *cma_dev;
1472         struct ib_port_attr port_attr;
1473         union ib_gid gid;
1474         u16 pkey;
1475         int ret;
1476         u8 p;
1477
1478         mutex_lock(&lock);
1479         list_for_each_entry(cma_dev, &dev_list, list)
1480                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1481                         if (!ib_query_port (cma_dev->device, p, &port_attr) &&
1482                             port_attr.state == IB_PORT_ACTIVE)
1483                                 goto port_found;
1484
1485         if (!list_empty(&dev_list)) {
1486                 p = 1;
1487                 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1488         } else {
1489                 ret = -ENODEV;
1490                 goto out;
1491         }
1492
1493 port_found:
1494         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1495         if (ret)
1496                 goto out;
1497
1498         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1499         if (ret)
1500                 goto out;
1501
1502         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1503         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1504         id_priv->id.port_num = p;
1505         cma_attach_to_dev(id_priv, cma_dev);
1506 out:
1507         mutex_unlock(&lock);
1508         return ret;
1509 }
1510
1511 static void addr_handler(int status, struct sockaddr *src_addr,
1512                          struct rdma_dev_addr *dev_addr, void *context)
1513 {
1514         struct rdma_id_private *id_priv = context;
1515         enum rdma_cm_event_type event;
1516
1517         atomic_inc(&id_priv->dev_remove);
1518
1519         /*
1520          * Grab mutex to block rdma_destroy_id() from removing the device while
1521          * we're trying to acquire it.
1522          */
1523         mutex_lock(&lock);
1524         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1525                 mutex_unlock(&lock);
1526                 goto out;
1527         }
1528
1529         if (!status && !id_priv->cma_dev)
1530                 status = cma_acquire_dev(id_priv);
1531         mutex_unlock(&lock);
1532
1533         if (status) {
1534                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1535                         goto out;
1536                 event = RDMA_CM_EVENT_ADDR_ERROR;
1537         } else {
1538                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1539                        ip_addr_size(src_addr));
1540                 event = RDMA_CM_EVENT_ADDR_RESOLVED;
1541         }
1542
1543         if (cma_notify_user(id_priv, event, status, NULL, 0)) {
1544                 cma_exch(id_priv, CMA_DESTROYING);
1545                 cma_release_remove(id_priv);
1546                 cma_deref_id(id_priv);
1547                 rdma_destroy_id(&id_priv->id);
1548                 return;
1549         }
1550 out:
1551         cma_release_remove(id_priv);
1552         cma_deref_id(id_priv);
1553 }
1554
1555 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1556 {
1557         struct cma_work *work;
1558         struct sockaddr_in *src_in, *dst_in;
1559         union ib_gid gid;
1560         int ret;
1561
1562         work = kzalloc(sizeof *work, GFP_KERNEL);
1563         if (!work)
1564                 return -ENOMEM;
1565
1566         if (!id_priv->cma_dev) {
1567                 ret = cma_bind_loopback(id_priv);
1568                 if (ret)
1569                         goto err;
1570         }
1571
1572         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1573         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1574
1575         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1576                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1577                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1578                 src_in->sin_family = dst_in->sin_family;
1579                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1580         }
1581
1582         work->id = id_priv;
1583         INIT_WORK(&work->work, cma_work_handler, work);
1584         work->old_state = CMA_ADDR_QUERY;
1585         work->new_state = CMA_ADDR_RESOLVED;
1586         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1587         queue_work(cma_wq, &work->work);
1588         return 0;
1589 err:
1590         kfree(work);
1591         return ret;
1592 }
1593
1594 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1595                          struct sockaddr *dst_addr)
1596 {
1597         if (src_addr && src_addr->sa_family)
1598                 return rdma_bind_addr(id, src_addr);
1599         else
1600                 return cma_bind_any(id, dst_addr->sa_family);
1601 }
1602
1603 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1604                       struct sockaddr *dst_addr, int timeout_ms)
1605 {
1606         struct rdma_id_private *id_priv;
1607         int ret;
1608
1609         id_priv = container_of(id, struct rdma_id_private, id);
1610         if (id_priv->state == CMA_IDLE) {
1611                 ret = cma_bind_addr(id, src_addr, dst_addr);
1612                 if (ret)
1613                         return ret;
1614         }
1615
1616         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1617                 return -EINVAL;
1618
1619         atomic_inc(&id_priv->refcount);
1620         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1621         if (cma_any_addr(dst_addr))
1622                 ret = cma_resolve_loopback(id_priv);
1623         else
1624                 ret = rdma_resolve_ip(&id->route.addr.src_addr, dst_addr,
1625                                       &id->route.addr.dev_addr,
1626                                       timeout_ms, addr_handler, id_priv);
1627         if (ret)
1628                 goto err;
1629
1630         return 0;
1631 err:
1632         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1633         cma_deref_id(id_priv);
1634         return ret;
1635 }
1636 EXPORT_SYMBOL(rdma_resolve_addr);
1637
1638 static void cma_bind_port(struct rdma_bind_list *bind_list,
1639                           struct rdma_id_private *id_priv)
1640 {
1641         struct sockaddr_in *sin;
1642
1643         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1644         sin->sin_port = htons(bind_list->port);
1645         id_priv->bind_list = bind_list;
1646         hlist_add_head(&id_priv->node, &bind_list->owners);
1647 }
1648
1649 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1650                           unsigned short snum)
1651 {
1652         struct rdma_bind_list *bind_list;
1653         int port, start, ret;
1654
1655         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1656         if (!bind_list)
1657                 return -ENOMEM;
1658
1659         start = snum ? snum : sysctl_local_port_range[0];
1660
1661         do {
1662                 ret = idr_get_new_above(ps, bind_list, start, &port);
1663         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1664
1665         if (ret)
1666                 goto err;
1667
1668         if ((snum && port != snum) ||
1669             (!snum && port > sysctl_local_port_range[1])) {
1670                 idr_remove(ps, port);
1671                 ret = -EADDRNOTAVAIL;
1672                 goto err;
1673         }
1674
1675         bind_list->ps = ps;
1676         bind_list->port = (unsigned short) port;
1677         cma_bind_port(bind_list, id_priv);
1678         return 0;
1679 err:
1680         kfree(bind_list);
1681         return ret;
1682 }
1683
1684 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1685 {
1686         struct rdma_id_private *cur_id;
1687         struct sockaddr_in *sin, *cur_sin;
1688         struct rdma_bind_list *bind_list;
1689         struct hlist_node *node;
1690         unsigned short snum;
1691
1692         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1693         snum = ntohs(sin->sin_port);
1694         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1695                 return -EACCES;
1696
1697         bind_list = idr_find(ps, snum);
1698         if (!bind_list)
1699                 return cma_alloc_port(ps, id_priv, snum);
1700
1701         /*
1702          * We don't support binding to any address if anyone is bound to
1703          * a specific address on the same port.
1704          */
1705         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1706                 return -EADDRNOTAVAIL;
1707
1708         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1709                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1710                         return -EADDRNOTAVAIL;
1711
1712                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1713                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1714                         return -EADDRINUSE;
1715         }
1716
1717         cma_bind_port(bind_list, id_priv);
1718         return 0;
1719 }
1720
1721 static int cma_get_port(struct rdma_id_private *id_priv)
1722 {
1723         struct idr *ps;
1724         int ret;
1725
1726         switch (id_priv->id.ps) {
1727         case RDMA_PS_SDP:
1728                 ps = &sdp_ps;
1729                 break;
1730         case RDMA_PS_TCP:
1731                 ps = &tcp_ps;
1732                 break;
1733         default:
1734                 return -EPROTONOSUPPORT;
1735         }
1736
1737         mutex_lock(&lock);
1738         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1739                 ret = cma_alloc_port(ps, id_priv, 0);
1740         else
1741                 ret = cma_use_port(ps, id_priv);
1742         mutex_unlock(&lock);
1743
1744         return ret;
1745 }
1746
1747 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1748 {
1749         struct rdma_id_private *id_priv;
1750         int ret;
1751
1752         if (addr->sa_family != AF_INET)
1753                 return -EAFNOSUPPORT;
1754
1755         id_priv = container_of(id, struct rdma_id_private, id);
1756         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1757                 return -EINVAL;
1758
1759         if (!cma_any_addr(addr)) {
1760                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1761                 if (!ret) {
1762                         mutex_lock(&lock);
1763                         ret = cma_acquire_dev(id_priv);
1764                         mutex_unlock(&lock);
1765                 }
1766                 if (ret)
1767                         goto err;
1768         }
1769
1770         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
1771         ret = cma_get_port(id_priv);
1772         if (ret)
1773                 goto err;
1774
1775         return 0;
1776 err:
1777         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
1778         return ret;
1779 }
1780 EXPORT_SYMBOL(rdma_bind_addr);
1781
1782 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
1783                           struct rdma_route *route)
1784 {
1785         struct sockaddr_in *src4, *dst4;
1786         struct cma_hdr *cma_hdr;
1787         struct sdp_hh *sdp_hdr;
1788
1789         src4 = (struct sockaddr_in *) &route->addr.src_addr;
1790         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
1791
1792         switch (ps) {
1793         case RDMA_PS_SDP:
1794                 sdp_hdr = hdr;
1795                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
1796                         return -EINVAL;
1797                 sdp_set_ip_ver(sdp_hdr, 4);
1798                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1799                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1800                 sdp_hdr->port = src4->sin_port;
1801                 break;
1802         default:
1803                 cma_hdr = hdr;
1804                 cma_hdr->cma_version = CMA_VERSION;
1805                 cma_set_ip_ver(cma_hdr, 4);
1806                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1807                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1808                 cma_hdr->port = src4->sin_port;
1809                 break;
1810         }
1811         return 0;
1812 }
1813
1814 static int cma_connect_ib(struct rdma_id_private *id_priv,
1815                           struct rdma_conn_param *conn_param)
1816 {
1817         struct ib_cm_req_param req;
1818         struct rdma_route *route;
1819         void *private_data;
1820         int offset, ret;
1821
1822         memset(&req, 0, sizeof req);
1823         offset = cma_user_data_offset(id_priv->id.ps);
1824         req.private_data_len = offset + conn_param->private_data_len;
1825         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
1826         if (!private_data)
1827                 return -ENOMEM;
1828
1829         if (conn_param->private_data && conn_param->private_data_len)
1830                 memcpy(private_data + offset, conn_param->private_data,
1831                        conn_param->private_data_len);
1832
1833         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
1834                                             id_priv);
1835         if (IS_ERR(id_priv->cm_id.ib)) {
1836                 ret = PTR_ERR(id_priv->cm_id.ib);
1837                 goto out;
1838         }
1839
1840         route = &id_priv->id.route;
1841         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
1842         if (ret)
1843                 goto out;
1844         req.private_data = private_data;
1845
1846         req.primary_path = &route->path_rec[0];
1847         if (route->num_paths == 2)
1848                 req.alternate_path = &route->path_rec[1];
1849
1850         req.service_id = cma_get_service_id(id_priv->id.ps,
1851                                             &route->addr.dst_addr);
1852         req.qp_num = id_priv->qp_num;
1853         req.qp_type = id_priv->qp_type;
1854         req.starting_psn = id_priv->seq_num;
1855         req.responder_resources = conn_param->responder_resources;
1856         req.initiator_depth = conn_param->initiator_depth;
1857         req.flow_control = conn_param->flow_control;
1858         req.retry_count = conn_param->retry_count;
1859         req.rnr_retry_count = conn_param->rnr_retry_count;
1860         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1861         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1862         req.max_cm_retries = CMA_MAX_CM_RETRIES;
1863         req.srq = id_priv->srq ? 1 : 0;
1864
1865         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
1866 out:
1867         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
1868                 ib_destroy_cm_id(id_priv->cm_id.ib);
1869                 id_priv->cm_id.ib = NULL;
1870         }
1871
1872         kfree(private_data);
1873         return ret;
1874 }
1875
1876 static int cma_connect_iw(struct rdma_id_private *id_priv,
1877                           struct rdma_conn_param *conn_param)
1878 {
1879         struct iw_cm_id *cm_id;
1880         struct sockaddr_in* sin;
1881         int ret;
1882         struct iw_cm_conn_param iw_param;
1883
1884         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
1885         if (IS_ERR(cm_id)) {
1886                 ret = PTR_ERR(cm_id);
1887                 goto out;
1888         }
1889
1890         id_priv->cm_id.iw = cm_id;
1891
1892         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
1893         cm_id->local_addr = *sin;
1894
1895         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
1896         cm_id->remote_addr = *sin;
1897
1898         ret = cma_modify_qp_rtr(&id_priv->id);
1899         if (ret)
1900                 goto out;
1901
1902         iw_param.ord = conn_param->initiator_depth;
1903         iw_param.ird = conn_param->responder_resources;
1904         iw_param.private_data = conn_param->private_data;
1905         iw_param.private_data_len = conn_param->private_data_len;
1906         if (id_priv->id.qp)
1907                 iw_param.qpn = id_priv->qp_num;
1908         else
1909                 iw_param.qpn = conn_param->qp_num;
1910         ret = iw_cm_connect(cm_id, &iw_param);
1911 out:
1912         if (ret && !IS_ERR(cm_id)) {
1913                 iw_destroy_cm_id(cm_id);
1914                 id_priv->cm_id.iw = NULL;
1915         }
1916         return ret;
1917 }
1918
1919 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1920 {
1921         struct rdma_id_private *id_priv;
1922         int ret;
1923
1924         id_priv = container_of(id, struct rdma_id_private, id);
1925         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
1926                 return -EINVAL;
1927
1928         if (!id->qp) {
1929                 id_priv->qp_num = conn_param->qp_num;
1930                 id_priv->qp_type = conn_param->qp_type;
1931                 id_priv->srq = conn_param->srq;
1932         }
1933
1934         switch (rdma_node_get_transport(id->device->node_type)) {
1935         case RDMA_TRANSPORT_IB:
1936                 ret = cma_connect_ib(id_priv, conn_param);
1937                 break;
1938         case RDMA_TRANSPORT_IWARP:
1939                 ret = cma_connect_iw(id_priv, conn_param);
1940                 break;
1941         default:
1942                 ret = -ENOSYS;
1943                 break;
1944         }
1945         if (ret)
1946                 goto err;
1947
1948         return 0;
1949 err:
1950         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
1951         return ret;
1952 }
1953 EXPORT_SYMBOL(rdma_connect);
1954
1955 static int cma_accept_ib(struct rdma_id_private *id_priv,
1956                          struct rdma_conn_param *conn_param)
1957 {
1958         struct ib_cm_rep_param rep;
1959         int ret;
1960
1961         ret = cma_modify_qp_rtr(&id_priv->id);
1962         if (ret)
1963                 return ret;
1964
1965         memset(&rep, 0, sizeof rep);
1966         rep.qp_num = id_priv->qp_num;
1967         rep.starting_psn = id_priv->seq_num;
1968         rep.private_data = conn_param->private_data;
1969         rep.private_data_len = conn_param->private_data_len;
1970         rep.responder_resources = conn_param->responder_resources;
1971         rep.initiator_depth = conn_param->initiator_depth;
1972         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
1973         rep.failover_accepted = 0;
1974         rep.flow_control = conn_param->flow_control;
1975         rep.rnr_retry_count = conn_param->rnr_retry_count;
1976         rep.srq = id_priv->srq ? 1 : 0;
1977
1978         return ib_send_cm_rep(id_priv->cm_id.ib, &rep);
1979 }
1980
1981 static int cma_accept_iw(struct rdma_id_private *id_priv,
1982                   struct rdma_conn_param *conn_param)
1983 {
1984         struct iw_cm_conn_param iw_param;
1985         int ret;
1986
1987         ret = cma_modify_qp_rtr(&id_priv->id);
1988         if (ret)
1989                 return ret;
1990
1991         iw_param.ord = conn_param->initiator_depth;
1992         iw_param.ird = conn_param->responder_resources;
1993         iw_param.private_data = conn_param->private_data;
1994         iw_param.private_data_len = conn_param->private_data_len;
1995         if (id_priv->id.qp) {
1996                 iw_param.qpn = id_priv->qp_num;
1997         } else
1998                 iw_param.qpn = conn_param->qp_num;
1999
2000         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2001 }
2002
2003 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2004 {
2005         struct rdma_id_private *id_priv;
2006         int ret;
2007
2008         id_priv = container_of(id, struct rdma_id_private, id);
2009         if (!cma_comp(id_priv, CMA_CONNECT))
2010                 return -EINVAL;
2011
2012         if (!id->qp && conn_param) {
2013                 id_priv->qp_num = conn_param->qp_num;
2014                 id_priv->qp_type = conn_param->qp_type;
2015                 id_priv->srq = conn_param->srq;
2016         }
2017
2018         switch (rdma_node_get_transport(id->device->node_type)) {
2019         case RDMA_TRANSPORT_IB:
2020                 if (conn_param)
2021                         ret = cma_accept_ib(id_priv, conn_param);
2022                 else
2023                         ret = cma_rep_recv(id_priv);
2024                 break;
2025         case RDMA_TRANSPORT_IWARP:
2026                 ret = cma_accept_iw(id_priv, conn_param);
2027                 break;
2028         default:
2029                 ret = -ENOSYS;
2030                 break;
2031         }
2032
2033         if (ret)
2034                 goto reject;
2035
2036         return 0;
2037 reject:
2038         cma_modify_qp_err(id);
2039         rdma_reject(id, NULL, 0);
2040         return ret;
2041 }
2042 EXPORT_SYMBOL(rdma_accept);
2043
2044 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2045                 u8 private_data_len)
2046 {
2047         struct rdma_id_private *id_priv;
2048         int ret;
2049
2050         id_priv = container_of(id, struct rdma_id_private, id);
2051         if (!cma_comp(id_priv, CMA_CONNECT))
2052                 return -EINVAL;
2053
2054         switch (rdma_node_get_transport(id->device->node_type)) {
2055         case RDMA_TRANSPORT_IB:
2056                 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2057                                      IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2058                                      private_data, private_data_len);
2059                 break;
2060         case RDMA_TRANSPORT_IWARP:
2061                 ret = iw_cm_reject(id_priv->cm_id.iw,
2062                                    private_data, private_data_len);
2063                 break;
2064         default:
2065                 ret = -ENOSYS;
2066                 break;
2067         }
2068         return ret;
2069 }
2070 EXPORT_SYMBOL(rdma_reject);
2071
2072 int rdma_disconnect(struct rdma_cm_id *id)
2073 {
2074         struct rdma_id_private *id_priv;
2075         int ret;
2076
2077         id_priv = container_of(id, struct rdma_id_private, id);
2078         if (!cma_comp(id_priv, CMA_CONNECT) &&
2079             !cma_comp(id_priv, CMA_DISCONNECT))
2080                 return -EINVAL;
2081
2082         switch (rdma_node_get_transport(id->device->node_type)) {
2083         case RDMA_TRANSPORT_IB:
2084                 ret = cma_modify_qp_err(id);
2085                 if (ret)
2086                         goto out;
2087                 /* Initiate or respond to a disconnect. */
2088                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2089                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2090                 break;
2091         case RDMA_TRANSPORT_IWARP:
2092                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2093                 break;
2094         default:
2095                 ret = -EINVAL;
2096                 break;
2097         }
2098 out:
2099         return ret;
2100 }
2101 EXPORT_SYMBOL(rdma_disconnect);
2102
2103 static void cma_add_one(struct ib_device *device)
2104 {
2105         struct cma_device *cma_dev;
2106         struct rdma_id_private *id_priv;
2107
2108         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2109         if (!cma_dev)
2110                 return;
2111
2112         cma_dev->device = device;
2113         cma_dev->node_guid = device->node_guid;
2114         if (!cma_dev->node_guid)
2115                 goto err;
2116
2117         init_completion(&cma_dev->comp);
2118         atomic_set(&cma_dev->refcount, 1);
2119         INIT_LIST_HEAD(&cma_dev->id_list);
2120         ib_set_client_data(device, &cma_client, cma_dev);
2121
2122         mutex_lock(&lock);
2123         list_add_tail(&cma_dev->list, &dev_list);
2124         list_for_each_entry(id_priv, &listen_any_list, list)
2125                 cma_listen_on_dev(id_priv, cma_dev);
2126         mutex_unlock(&lock);
2127         return;
2128 err:
2129         kfree(cma_dev);
2130 }
2131
2132 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2133 {
2134         enum cma_state state;
2135
2136         /* Record that we want to remove the device */
2137         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2138         if (state == CMA_DESTROYING)
2139                 return 0;
2140
2141         cma_cancel_operation(id_priv, state);
2142         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2143
2144         /* Check for destruction from another callback. */
2145         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2146                 return 0;
2147
2148         return cma_notify_user(id_priv, RDMA_CM_EVENT_DEVICE_REMOVAL,
2149                                0, NULL, 0);
2150 }
2151
2152 static void cma_process_remove(struct cma_device *cma_dev)
2153 {
2154         struct rdma_id_private *id_priv;
2155         int ret;
2156
2157         mutex_lock(&lock);
2158         while (!list_empty(&cma_dev->id_list)) {
2159                 id_priv = list_entry(cma_dev->id_list.next,
2160                                      struct rdma_id_private, list);
2161
2162                 if (cma_internal_listen(id_priv)) {
2163                         cma_destroy_listen(id_priv);
2164                         continue;
2165                 }
2166
2167                 list_del_init(&id_priv->list);
2168                 atomic_inc(&id_priv->refcount);
2169                 mutex_unlock(&lock);
2170
2171                 ret = cma_remove_id_dev(id_priv);
2172                 cma_deref_id(id_priv);
2173                 if (ret)
2174                         rdma_destroy_id(&id_priv->id);
2175
2176                 mutex_lock(&lock);
2177         }
2178         mutex_unlock(&lock);
2179
2180         cma_deref_dev(cma_dev);
2181         wait_for_completion(&cma_dev->comp);
2182 }
2183
2184 static void cma_remove_one(struct ib_device *device)
2185 {
2186         struct cma_device *cma_dev;
2187
2188         cma_dev = ib_get_client_data(device, &cma_client);
2189         if (!cma_dev)
2190                 return;
2191
2192         mutex_lock(&lock);
2193         list_del(&cma_dev->list);
2194         mutex_unlock(&lock);
2195
2196         cma_process_remove(cma_dev);
2197         kfree(cma_dev);
2198 }
2199
2200 static int cma_init(void)
2201 {
2202         int ret;
2203
2204         cma_wq = create_singlethread_workqueue("rdma_cm_wq");
2205         if (!cma_wq)
2206                 return -ENOMEM;
2207
2208         ib_sa_register_client(&sa_client);
2209
2210         ret = ib_register_client(&cma_client);
2211         if (ret)
2212                 goto err;
2213         return 0;
2214
2215 err:
2216         ib_sa_unregister_client(&sa_client);
2217         destroy_workqueue(cma_wq);
2218         return ret;
2219 }
2220
2221 static void cma_cleanup(void)
2222 {
2223         ib_unregister_client(&cma_client);
2224         ib_sa_unregister_client(&sa_client);
2225         destroy_workqueue(cma_wq);
2226         idr_destroy(&sdp_ps);
2227         idr_destroy(&tcp_ps);
2228 }
2229
2230 module_init(cma_init);
2231 module_exit(cma_cleanup);