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         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 878                              &ip_ver, &port, &src, &dst))
 879                 goto err;
 880
 881         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 882                             listen_id->ps);
 883         if (IS_ERR(id))
 884                 goto err;
 885
 886         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 887                           ip_ver, port, src, dst);
 888
 889         rt = &id->route;
 890         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
 891         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
 892                                GFP_KERNEL);
 893         if (!rt->path_rec)
 894                 goto destroy_id;
 895
 896         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
 897         if (rt->num_paths == 2)
 898                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
 899
 900         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
 901         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
 902         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
 903         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
 904
 905         id_priv = container_of(id, struct rdma_id_private, id);
 906         id_priv->state = CMA_CONNECT;
 907         return id_priv;
 908
 909 destroy_id:
 910         rdma_destroy_id(id);
 911 err:
 912         return NULL;
 913 }
 914
 915 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 916 {
 917         struct rdma_id_private *listen_id, *conn_id;
 918         int offset, ret;
 919
 920         listen_id = cm_id->context;
 921         atomic_inc(&listen_id->dev_remove);
 922         if (!cma_comp(listen_id, CMA_LISTEN)) {
 923                 ret = -ECONNABORTED;
 924                 goto out;
 925         }
 926
 927         conn_id = cma_new_id(&listen_id->id, ib_event);
 928         if (!conn_id) {
 929                 ret = -ENOMEM;
 930                 goto out;
 931         }
 932
 933         atomic_inc(&conn_id->dev_remove);
 934         mutex_lock(&lock);
 935         ret = cma_acquire_dev(conn_id);
 936         mutex_unlock(&lock);
 937         if (ret) {
 938                 ret = -ENODEV;
 939                 cma_exch(conn_id, CMA_DESTROYING);
 940                 cma_release_remove(conn_id);
 941                 rdma_destroy_id(&conn_id->id);
 942                 goto out;
 943         }
 944
 945         conn_id->cm_id.ib = cm_id;
 946         cm_id->context = conn_id;
 947         cm_id->cm_handler = cma_ib_handler;
 948
 949         offset = cma_user_data_offset(listen_id->id.ps);
 950         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
 951                               ib_event->private_data + offset,
 952                               IB_CM_REQ_PRIVATE_DATA_SIZE - offset);
 953         if (ret) {
 954                 /* Destroy the CM ID by returning a non-zero value. */
 955                 conn_id->cm_id.ib = NULL;
 956                 cma_exch(conn_id, CMA_DESTROYING);
 957                 cma_release_remove(conn_id);
 958                 rdma_destroy_id(&conn_id->id);
 959         }
 960 out:
 961         cma_release_remove(listen_id);
 962         return ret;
 963 }
 964
 965 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
 966 {
 967         return cpu_to_be64(((u64)ps << 16) +
 968                be16_to_cpu(((struct sockaddr_in *) addr)->sin_port));
 969 }
 970
 971 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
 972                                  struct ib_cm_compare_data *compare)
 973 {
 974         struct cma_hdr *cma_data, *cma_mask;
 975         struct sdp_hh *sdp_data, *sdp_mask;
 976         __u32 ip4_addr;
 977         struct in6_addr ip6_addr;
 978
 979         memset(compare, 0, sizeof *compare);
 980         cma_data = (void *) compare->data;
 981         cma_mask = (void *) compare->mask;
 982         sdp_data = (void *) compare->data;
 983         sdp_mask = (void *) compare->mask;
 984
 985         switch (addr->sa_family) {
 986         case AF_INET:
 987                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
 988                 if (ps == RDMA_PS_SDP) {
 989                         sdp_set_ip_ver(sdp_data, 4);
 990                         sdp_set_ip_ver(sdp_mask, 0xF);
 991                         sdp_data->dst_addr.ip4.addr = ip4_addr;
 992                         sdp_mask->dst_addr.ip4.addr = ~0;
 993                 } else {
 994                         cma_set_ip_ver(cma_data, 4);
 995                         cma_set_ip_ver(cma_mask, 0xF);
 996                         cma_data->dst_addr.ip4.addr = ip4_addr;
 997                         cma_mask->dst_addr.ip4.addr = ~0;
 998                 }
 999                 break;
1000         case AF_INET6:
1001                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1002                 if (ps == RDMA_PS_SDP) {
1003                         sdp_set_ip_ver(sdp_data, 6);
1004                         sdp_set_ip_ver(sdp_mask, 0xF);
1005                         sdp_data->dst_addr.ip6 = ip6_addr;
1006                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1007                                sizeof sdp_mask->dst_addr.ip6);
1008                 } else {
1009                         cma_set_ip_ver(cma_data, 6);
1010                         cma_set_ip_ver(cma_mask, 0xF);
1011                         cma_data->dst_addr.ip6 = ip6_addr;
1012                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1013                                sizeof cma_mask->dst_addr.ip6);
1014                 }
1015                 break;
1016         default:
1017                 break;
1018         }
1019 }
1020
1021 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1022 {
1023         struct rdma_id_private *id_priv = iw_id->context;
1024         enum rdma_cm_event_type event = 0;
1025         struct sockaddr_in *sin;
1026         int ret = 0;
1027
1028         atomic_inc(&id_priv->dev_remove);
1029
1030         switch (iw_event->event) {
1031         case IW_CM_EVENT_CLOSE:
1032                 event = RDMA_CM_EVENT_DISCONNECTED;
1033                 break;
1034         case IW_CM_EVENT_CONNECT_REPLY:
1035                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1036                 *sin = iw_event->local_addr;
1037                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1038                 *sin = iw_event->remote_addr;
1039                 if (iw_event->status)
1040                         event = RDMA_CM_EVENT_REJECTED;
1041                 else
1042                         event = RDMA_CM_EVENT_ESTABLISHED;
1043                 break;
1044         case IW_CM_EVENT_ESTABLISHED:
1045                 event = RDMA_CM_EVENT_ESTABLISHED;
1046                 break;
1047         default:
1048                 BUG_ON(1);
1049         }
1050
1051         ret = cma_notify_user(id_priv, event, iw_event->status,
1052                               iw_event->private_data,
1053                               iw_event->private_data_len);
1054         if (ret) {
1055                 /* Destroy the CM ID by returning a non-zero value. */
1056                 id_priv->cm_id.iw = NULL;
1057                 cma_exch(id_priv, CMA_DESTROYING);
1058                 cma_release_remove(id_priv);
1059                 rdma_destroy_id(&id_priv->id);
1060                 return ret;
1061         }
1062
1063         cma_release_remove(id_priv);
1064         return ret;
1065 }
1066
1067 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1068                                struct iw_cm_event *iw_event)
1069 {
1070         struct rdma_cm_id *new_cm_id;
1071         struct rdma_id_private *listen_id, *conn_id;
1072         struct sockaddr_in *sin;
1073         struct net_device *dev = NULL;
1074         int ret;
1075
1076         listen_id = cm_id->context;
1077         atomic_inc(&listen_id->dev_remove);
1078         if (!cma_comp(listen_id, CMA_LISTEN)) {
1079                 ret = -ECONNABORTED;
1080                 goto out;
1081         }
1082
1083         /* Create a new RDMA id for the new IW CM ID */
1084         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1085                                    listen_id->id.context,
1086                                    RDMA_PS_TCP);
1087         if (!new_cm_id) {
1088                 ret = -ENOMEM;
1089                 goto out;
1090         }
1091         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1092         atomic_inc(&conn_id->dev_remove);
1093         conn_id->state = CMA_CONNECT;
1094
1095         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1096         if (!dev) {
1097                 ret = -EADDRNOTAVAIL;
1098                 cma_release_remove(conn_id);
1099                 rdma_destroy_id(new_cm_id);
1100                 goto out;
1101         }
1102         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1103         if (ret) {
1104                 cma_release_remove(conn_id);
1105                 rdma_destroy_id(new_cm_id);
1106                 goto out;
1107         }
1108
1109         mutex_lock(&lock);
1110         ret = cma_acquire_dev(conn_id);
1111         mutex_unlock(&lock);
1112         if (ret) {
1113                 cma_release_remove(conn_id);
1114                 rdma_destroy_id(new_cm_id);
1115                 goto out;
1116         }
1117
1118         conn_id->cm_id.iw = cm_id;
1119         cm_id->context = conn_id;
1120         cm_id->cm_handler = cma_iw_handler;
1121
1122         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1123         *sin = iw_event->local_addr;
1124         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1125         *sin = iw_event->remote_addr;
1126
1127         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
1128                               iw_event->private_data,
1129                               iw_event->private_data_len);
1130         if (ret) {
1131                 /* User wants to destroy the CM ID */
1132                 conn_id->cm_id.iw = NULL;
1133                 cma_exch(conn_id, CMA_DESTROYING);
1134                 cma_release_remove(conn_id);
1135                 rdma_destroy_id(&conn_id->id);
1136         }
1137
1138 out:
1139         if (dev)
1140                 dev_put(dev);
1141         cma_release_remove(listen_id);
1142         return ret;
1143 }
1144
1145 static int cma_ib_listen(struct rdma_id_private *id_priv)
1146 {
1147         struct ib_cm_compare_data compare_data;
1148         struct sockaddr *addr;
1149         __be64 svc_id;
1150         int ret;
1151
1152         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1153                                             id_priv);
1154         if (IS_ERR(id_priv->cm_id.ib))
1155                 return PTR_ERR(id_priv->cm_id.ib);
1156
1157         addr = &id_priv->id.route.addr.src_addr;
1158         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1159         if (cma_any_addr(addr))
1160                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1161         else {
1162                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1163                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1164         }
1165
1166         if (ret) {
1167                 ib_destroy_cm_id(id_priv->cm_id.ib);
1168                 id_priv->cm_id.ib = NULL;
1169         }
1170
1171         return ret;
1172 }
1173
1174 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1175 {
1176         int ret;
1177         struct sockaddr_in *sin;
1178
1179         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1180                                             iw_conn_req_handler,
1181                                             id_priv);
1182         if (IS_ERR(id_priv->cm_id.iw))
1183                 return PTR_ERR(id_priv->cm_id.iw);
1184
1185         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1186         id_priv->cm_id.iw->local_addr = *sin;
1187
1188         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1189
1190         if (ret) {
1191                 iw_destroy_cm_id(id_priv->cm_id.iw);
1192                 id_priv->cm_id.iw = NULL;
1193         }
1194
1195         return ret;
1196 }
1197
1198 static int cma_listen_handler(struct rdma_cm_id *id,
1199                               struct rdma_cm_event *event)
1200 {
1201         struct rdma_id_private *id_priv = id->context;
1202
1203         id->context = id_priv->id.context;
1204         id->event_handler = id_priv->id.event_handler;
1205         return id_priv->id.event_handler(id, event);
1206 }
1207
1208 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1209                               struct cma_device *cma_dev)
1210 {
1211         struct rdma_id_private *dev_id_priv;
1212         struct rdma_cm_id *id;
1213         int ret;
1214
1215         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1216         if (IS_ERR(id))
1217                 return;
1218
1219         dev_id_priv = container_of(id, struct rdma_id_private, id);
1220
1221         dev_id_priv->state = CMA_ADDR_BOUND;
1222         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1223                ip_addr_size(&id_priv->id.route.addr.src_addr));
1224
1225         cma_attach_to_dev(dev_id_priv, cma_dev);
1226         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1227
1228         ret = rdma_listen(id, id_priv->backlog);
1229         if (ret)
1230                 goto err;
1231
1232         return;
1233 err:
1234         cma_destroy_listen(dev_id_priv);
1235 }
1236
1237 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1238 {
1239         struct cma_device *cma_dev;
1240
1241         mutex_lock(&lock);
1242         list_add_tail(&id_priv->list, &listen_any_list);
1243         list_for_each_entry(cma_dev, &dev_list, list)
1244                 cma_listen_on_dev(id_priv, cma_dev);
1245         mutex_unlock(&lock);
1246 }
1247
1248 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1249 {
1250         struct sockaddr_in addr_in;
1251
1252         memset(&addr_in, 0, sizeof addr_in);
1253         addr_in.sin_family = af;
1254         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1255 }
1256
1257 int rdma_listen(struct rdma_cm_id *id, int backlog)
1258 {
1259         struct rdma_id_private *id_priv;
1260         int ret;
1261
1262         id_priv = container_of(id, struct rdma_id_private, id);
1263         if (id_priv->state == CMA_IDLE) {
1264                 ret = cma_bind_any(id, AF_INET);
1265                 if (ret)
1266                         return ret;
1267         }
1268
1269         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1270                 return -EINVAL;
1271
1272         id_priv->backlog = backlog;
1273         if (id->device) {
1274                 switch (rdma_node_get_transport(id->device->node_type)) {
1275                 case RDMA_TRANSPORT_IB:
1276                         ret = cma_ib_listen(id_priv);
1277                         if (ret)
1278                                 goto err;
1279                         break;
1280                 case RDMA_TRANSPORT_IWARP:
1281                         ret = cma_iw_listen(id_priv, backlog);
1282                         if (ret)
1283                                 goto err;
1284                         break;
1285                 default:
1286                         ret = -ENOSYS;
1287                         goto err;
1288                 }
1289         } else
1290                 cma_listen_on_all(id_priv);
1291
1292         return 0;
1293 err:
1294         id_priv->backlog = 0;
1295         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1296         return ret;
1297 }
1298 EXPORT_SYMBOL(rdma_listen);
1299
1300 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1301                               void *context)
1302 {
1303         struct cma_work *work = context;
1304         struct rdma_route *route;
1305
1306         route = &work->id->id.route;
1307
1308         if (!status) {
1309                 route->num_paths = 1;
1310                 *route->path_rec = *path_rec;
1311         } else {
1312                 work->old_state = CMA_ROUTE_QUERY;
1313                 work->new_state = CMA_ADDR_RESOLVED;
1314                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1315                 work->event.status = status;
1316         }
1317
1318         queue_work(cma_wq, &work->work);
1319 }
1320
1321 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1322                               struct cma_work *work)
1323 {
1324         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1325         struct ib_sa_path_rec path_rec;
1326
1327         memset(&path_rec, 0, sizeof path_rec);
1328         ib_addr_get_sgid(addr, &path_rec.sgid);
1329         ib_addr_get_dgid(addr, &path_rec.dgid);
1330         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1331         path_rec.numb_path = 1;
1332
1333         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1334                                 id_priv->id.port_num, &path_rec,
1335                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1336                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
1337                                 timeout_ms, GFP_KERNEL,
1338                                 cma_query_handler, work, &id_priv->query);
1339
1340         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1341 }
1342
1343 static void cma_work_handler(void *data)
1344 {
1345         struct cma_work *work = data;
1346         struct rdma_id_private *id_priv = work->id;
1347         int destroy = 0;
1348
1349         atomic_inc(&id_priv->dev_remove);
1350         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1351                 goto out;
1352
1353         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1354                 cma_exch(id_priv, CMA_DESTROYING);
1355                 destroy = 1;
1356         }
1357 out:
1358         cma_release_remove(id_priv);
1359         cma_deref_id(id_priv);
1360         if (destroy)
1361                 rdma_destroy_id(&id_priv->id);
1362         kfree(work);
1363 }
1364
1365 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1366 {
1367         struct rdma_route *route = &id_priv->id.route;
1368         struct cma_work *work;
1369         int ret;
1370
1371         work = kzalloc(sizeof *work, GFP_KERNEL);
1372         if (!work)
1373                 return -ENOMEM;
1374
1375         work->id = id_priv;
1376         INIT_WORK(&work->work, cma_work_handler, work);
1377         work->old_state = CMA_ROUTE_QUERY;
1378         work->new_state = CMA_ROUTE_RESOLVED;
1379         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1380
1381         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1382         if (!route->path_rec) {
1383                 ret = -ENOMEM;
1384                 goto err1;
1385         }
1386
1387         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1388         if (ret)
1389                 goto err2;
1390
1391         return 0;
1392 err2:
1393         kfree(route->path_rec);
1394         route->path_rec = NULL;
1395 err1:
1396         kfree(work);
1397         return ret;
1398 }
1399
1400 int rdma_set_ib_paths(struct rdma_cm_id *id,
1401                       struct ib_sa_path_rec *path_rec, int num_paths)
1402 {
1403         struct rdma_id_private *id_priv;
1404         int ret;
1405
1406         id_priv = container_of(id, struct rdma_id_private, id);
1407         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1408                 return -EINVAL;
1409
1410         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1411         if (!id->route.path_rec) {
1412                 ret = -ENOMEM;
1413                 goto err;
1414         }
1415
1416         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1417         return 0;
1418 err:
1419         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1420         return ret;
1421 }
1422 EXPORT_SYMBOL(rdma_set_ib_paths);
1423
1424 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1425 {
1426         struct cma_work *work;
1427
1428         work = kzalloc(sizeof *work, GFP_KERNEL);
1429         if (!work)
1430                 return -ENOMEM;
1431
1432         work->id = id_priv;
1433         INIT_WORK(&work->work, cma_work_handler, work);
1434         work->old_state = CMA_ROUTE_QUERY;
1435         work->new_state = CMA_ROUTE_RESOLVED;
1436         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1437         queue_work(cma_wq, &work->work);
1438         return 0;
1439 }
1440
1441 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1442 {
1443         struct rdma_id_private *id_priv;
1444         int ret;
1445
1446         id_priv = container_of(id, struct rdma_id_private, id);
1447         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1448                 return -EINVAL;
1449
1450         atomic_inc(&id_priv->refcount);
1451         switch (rdma_node_get_transport(id->device->node_type)) {
1452         case RDMA_TRANSPORT_IB:
1453                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1454                 break;
1455         case RDMA_TRANSPORT_IWARP:
1456                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1457                 break;
1458         default:
1459                 ret = -ENOSYS;
1460                 break;
1461         }
1462         if (ret)
1463                 goto err;
1464
1465         return 0;
1466 err:
1467         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1468         cma_deref_id(id_priv);
1469         return ret;
1470 }
1471 EXPORT_SYMBOL(rdma_resolve_route);
1472
1473 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1474 {
1475         struct cma_device *cma_dev;
1476         struct ib_port_attr port_attr;
1477         union ib_gid gid;
1478         u16 pkey;
1479         int ret;
1480         u8 p;
1481
1482         mutex_lock(&lock);
1483         list_for_each_entry(cma_dev, &dev_list, list)
1484                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1485                         if (!ib_query_port (cma_dev->device, p, &port_attr) &&
1486                             port_attr.state == IB_PORT_ACTIVE)
1487                                 goto port_found;
1488
1489         if (!list_empty(&dev_list)) {
1490                 p = 1;
1491                 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1492         } else {
1493                 ret = -ENODEV;
1494                 goto out;
1495         }
1496
1497 port_found:
1498         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1499         if (ret)
1500                 goto out;
1501
1502         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1503         if (ret)
1504                 goto out;
1505
1506         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1507         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1508         id_priv->id.port_num = p;
1509         cma_attach_to_dev(id_priv, cma_dev);
1510 out:
1511         mutex_unlock(&lock);
1512         return ret;
1513 }
1514
1515 static void addr_handler(int status, struct sockaddr *src_addr,
1516                          struct rdma_dev_addr *dev_addr, void *context)
1517 {
1518         struct rdma_id_private *id_priv = context;
1519         enum rdma_cm_event_type event;
1520
1521         atomic_inc(&id_priv->dev_remove);
1522
1523         /*
1524          * Grab mutex to block rdma_destroy_id() from removing the device while
1525          * we're trying to acquire it.
1526          */
1527         mutex_lock(&lock);
1528         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1529                 mutex_unlock(&lock);
1530                 goto out;
1531         }
1532
1533         if (!status && !id_priv->cma_dev)
1534                 status = cma_acquire_dev(id_priv);
1535         mutex_unlock(&lock);
1536
1537         if (status) {
1538                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1539                         goto out;
1540                 event = RDMA_CM_EVENT_ADDR_ERROR;
1541         } else {
1542                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1543                        ip_addr_size(src_addr));
1544                 event = RDMA_CM_EVENT_ADDR_RESOLVED;
1545         }
1546
1547         if (cma_notify_user(id_priv, event, status, NULL, 0)) {
1548                 cma_exch(id_priv, CMA_DESTROYING);
1549                 cma_release_remove(id_priv);
1550                 cma_deref_id(id_priv);
1551                 rdma_destroy_id(&id_priv->id);
1552                 return;
1553         }
1554 out:
1555         cma_release_remove(id_priv);
1556         cma_deref_id(id_priv);
1557 }
1558
1559 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1560 {
1561         struct cma_work *work;
1562         struct sockaddr_in *src_in, *dst_in;
1563         union ib_gid gid;
1564         int ret;
1565
1566         work = kzalloc(sizeof *work, GFP_KERNEL);
1567         if (!work)
1568                 return -ENOMEM;
1569
1570         if (!id_priv->cma_dev) {
1571                 ret = cma_bind_loopback(id_priv);
1572                 if (ret)
1573                         goto err;
1574         }
1575
1576         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1577         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1578
1579         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1580                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1581                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1582                 src_in->sin_family = dst_in->sin_family;
1583                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1584         }
1585
1586         work->id = id_priv;
1587         INIT_WORK(&work->work, cma_work_handler, work);
1588         work->old_state = CMA_ADDR_QUERY;
1589         work->new_state = CMA_ADDR_RESOLVED;
1590         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1591         queue_work(cma_wq, &work->work);
1592         return 0;
1593 err:
1594         kfree(work);
1595         return ret;
1596 }
1597
1598 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1599                          struct sockaddr *dst_addr)
1600 {
1601         if (src_addr && src_addr->sa_family)
1602                 return rdma_bind_addr(id, src_addr);
1603         else
1604                 return cma_bind_any(id, dst_addr->sa_family);
1605 }
1606
1607 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1608                       struct sockaddr *dst_addr, int timeout_ms)
1609 {
1610         struct rdma_id_private *id_priv;
1611         int ret;
1612
1613         id_priv = container_of(id, struct rdma_id_private, id);
1614         if (id_priv->state == CMA_IDLE) {
1615                 ret = cma_bind_addr(id, src_addr, dst_addr);
1616                 if (ret)
1617                         return ret;
1618         }
1619
1620         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1621                 return -EINVAL;
1622
1623         atomic_inc(&id_priv->refcount);
1624         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1625         if (cma_any_addr(dst_addr))
1626                 ret = cma_resolve_loopback(id_priv);
1627         else
1628                 ret = rdma_resolve_ip(&id->route.addr.src_addr, dst_addr,
1629                                       &id->route.addr.dev_addr,
1630                                       timeout_ms, addr_handler, id_priv);
1631         if (ret)
1632                 goto err;
1633
1634         return 0;
1635 err:
1636         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1637         cma_deref_id(id_priv);
1638         return ret;
1639 }
1640 EXPORT_SYMBOL(rdma_resolve_addr);
1641
1642 static void cma_bind_port(struct rdma_bind_list *bind_list,
1643                           struct rdma_id_private *id_priv)
1644 {
1645         struct sockaddr_in *sin;
1646
1647         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1648         sin->sin_port = htons(bind_list->port);
1649         id_priv->bind_list = bind_list;
1650         hlist_add_head(&id_priv->node, &bind_list->owners);
1651 }
1652
1653 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1654                           unsigned short snum)
1655 {
1656         struct rdma_bind_list *bind_list;
1657         int port, start, ret;
1658
1659         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1660         if (!bind_list)
1661                 return -ENOMEM;
1662
1663         start = snum ? snum : sysctl_local_port_range[0];
1664
1665         do {
1666                 ret = idr_get_new_above(ps, bind_list, start, &port);
1667         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1668
1669         if (ret)
1670                 goto err;
1671
1672         if ((snum && port != snum) ||
1673             (!snum && port > sysctl_local_port_range[1])) {
1674                 idr_remove(ps, port);
1675                 ret = -EADDRNOTAVAIL;
1676                 goto err;
1677         }
1678
1679         bind_list->ps = ps;
1680         bind_list->port = (unsigned short) port;
1681         cma_bind_port(bind_list, id_priv);
1682         return 0;
1683 err:
1684         kfree(bind_list);
1685         return ret;
1686 }
1687
1688 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1689 {
1690         struct rdma_id_private *cur_id;
1691         struct sockaddr_in *sin, *cur_sin;
1692         struct rdma_bind_list *bind_list;
1693         struct hlist_node *node;
1694         unsigned short snum;
1695
1696         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1697         snum = ntohs(sin->sin_port);
1698         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1699                 return -EACCES;
1700
1701         bind_list = idr_find(ps, snum);
1702         if (!bind_list)
1703                 return cma_alloc_port(ps, id_priv, snum);
1704
1705         /*
1706          * We don't support binding to any address if anyone is bound to
1707          * a specific address on the same port.
1708          */
1709         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1710                 return -EADDRNOTAVAIL;
1711
1712         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1713                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1714                         return -EADDRNOTAVAIL;
1715
1716                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1717                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1718                         return -EADDRINUSE;
1719         }
1720
1721         cma_bind_port(bind_list, id_priv);
1722         return 0;
1723 }
1724
1725 static int cma_get_port(struct rdma_id_private *id_priv)
1726 {
1727         struct idr *ps;
1728         int ret;
1729
1730         switch (id_priv->id.ps) {
1731         case RDMA_PS_SDP:
1732                 ps = &sdp_ps;
1733                 break;
1734         case RDMA_PS_TCP:
1735                 ps = &tcp_ps;
1736                 break;
1737         default:
1738                 return -EPROTONOSUPPORT;
1739         }
1740
1741         mutex_lock(&lock);
1742         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1743                 ret = cma_alloc_port(ps, id_priv, 0);
1744         else
1745                 ret = cma_use_port(ps, id_priv);
1746         mutex_unlock(&lock);
1747
1748         return ret;
1749 }
1750
1751 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1752 {
1753         struct rdma_id_private *id_priv;
1754         int ret;
1755
1756         if (addr->sa_family != AF_INET)
1757                 return -EAFNOSUPPORT;
1758
1759         id_priv = container_of(id, struct rdma_id_private, id);
1760         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1761                 return -EINVAL;
1762
1763         if (!cma_any_addr(addr)) {
1764                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1765                 if (!ret) {
1766                         mutex_lock(&lock);
1767                         ret = cma_acquire_dev(id_priv);
1768                         mutex_unlock(&lock);
1769                 }
1770                 if (ret)
1771                         goto err;
1772         }
1773
1774         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
1775         ret = cma_get_port(id_priv);
1776         if (ret)
1777                 goto err;
1778
1779         return 0;
1780 err:
1781         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
1782         return ret;
1783 }
1784 EXPORT_SYMBOL(rdma_bind_addr);
1785
1786 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
1787                           struct rdma_route *route)
1788 {
1789         struct sockaddr_in *src4, *dst4;
1790         struct cma_hdr *cma_hdr;
1791         struct sdp_hh *sdp_hdr;
1792
1793         src4 = (struct sockaddr_in *) &route->addr.src_addr;
1794         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
1795
1796         switch (ps) {
1797         case RDMA_PS_SDP:
1798                 sdp_hdr = hdr;
1799                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
1800                         return -EINVAL;
1801                 sdp_set_ip_ver(sdp_hdr, 4);
1802                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1803                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1804                 sdp_hdr->port = src4->sin_port;
1805                 break;
1806         default:
1807                 cma_hdr = hdr;
1808                 cma_hdr->cma_version = CMA_VERSION;
1809                 cma_set_ip_ver(cma_hdr, 4);
1810                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1811                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1812                 cma_hdr->port = src4->sin_port;
1813                 break;
1814         }
1815         return 0;
1816 }
1817
1818 static int cma_connect_ib(struct rdma_id_private *id_priv,
1819                           struct rdma_conn_param *conn_param)
1820 {
1821         struct ib_cm_req_param req;
1822         struct rdma_route *route;
1823         void *private_data;
1824         int offset, ret;
1825
1826         memset(&req, 0, sizeof req);
1827         offset = cma_user_data_offset(id_priv->id.ps);
1828         req.private_data_len = offset + conn_param->private_data_len;
1829         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
1830         if (!private_data)
1831                 return -ENOMEM;
1832
1833         if (conn_param->private_data && conn_param->private_data_len)
1834                 memcpy(private_data + offset, conn_param->private_data,
1835                        conn_param->private_data_len);
1836
1837         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
1838                                             id_priv);
1839         if (IS_ERR(id_priv->cm_id.ib)) {
1840                 ret = PTR_ERR(id_priv->cm_id.ib);
1841                 goto out;
1842         }
1843
1844         route = &id_priv->id.route;
1845         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
1846         if (ret)
1847                 goto out;
1848         req.private_data = private_data;
1849
1850         req.primary_path = &route->path_rec[0];
1851         if (route->num_paths == 2)
1852                 req.alternate_path = &route->path_rec[1];
1853
1854         req.service_id = cma_get_service_id(id_priv->id.ps,
1855                                             &route->addr.dst_addr);
1856         req.qp_num = id_priv->qp_num;
1857         req.qp_type = id_priv->qp_type;
1858         req.starting_psn = id_priv->seq_num;
1859         req.responder_resources = conn_param->responder_resources;
1860         req.initiator_depth = conn_param->initiator_depth;
1861         req.flow_control = conn_param->flow_control;
1862         req.retry_count = conn_param->retry_count;
1863         req.rnr_retry_count = conn_param->rnr_retry_count;
1864         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1865         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1866         req.max_cm_retries = CMA_MAX_CM_RETRIES;
1867         req.srq = id_priv->srq ? 1 : 0;
1868
1869         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
1870 out:
1871         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
1872                 ib_destroy_cm_id(id_priv->cm_id.ib);
1873                 id_priv->cm_id.ib = NULL;
1874         }
1875
1876         kfree(private_data);
1877         return ret;
1878 }
1879
1880 static int cma_connect_iw(struct rdma_id_private *id_priv,
1881                           struct rdma_conn_param *conn_param)
1882 {
1883         struct iw_cm_id *cm_id;
1884         struct sockaddr_in* sin;
1885         int ret;
1886         struct iw_cm_conn_param iw_param;
1887
1888         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
1889         if (IS_ERR(cm_id)) {
1890                 ret = PTR_ERR(cm_id);
1891                 goto out;
1892         }
1893
1894         id_priv->cm_id.iw = cm_id;
1895
1896         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
1897         cm_id->local_addr = *sin;
1898
1899         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
1900         cm_id->remote_addr = *sin;
1901
1902         ret = cma_modify_qp_rtr(&id_priv->id);
1903         if (ret)
1904                 goto out;
1905
1906         iw_param.ord = conn_param->initiator_depth;
1907         iw_param.ird = conn_param->responder_resources;
1908         iw_param.private_data = conn_param->private_data;
1909         iw_param.private_data_len = conn_param->private_data_len;
1910         if (id_priv->id.qp)
1911                 iw_param.qpn = id_priv->qp_num;
1912         else
1913                 iw_param.qpn = conn_param->qp_num;
1914         ret = iw_cm_connect(cm_id, &iw_param);
1915 out:
1916         if (ret && !IS_ERR(cm_id)) {
1917                 iw_destroy_cm_id(cm_id);
1918                 id_priv->cm_id.iw = NULL;
1919         }
1920         return ret;
1921 }
1922
1923 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1924 {
1925         struct rdma_id_private *id_priv;
1926         int ret;
1927
1928         id_priv = container_of(id, struct rdma_id_private, id);
1929         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
1930                 return -EINVAL;
1931
1932         if (!id->qp) {
1933                 id_priv->qp_num = conn_param->qp_num;
1934                 id_priv->qp_type = conn_param->qp_type;
1935                 id_priv->srq = conn_param->srq;
1936         }
1937
1938         switch (rdma_node_get_transport(id->device->node_type)) {
1939         case RDMA_TRANSPORT_IB:
1940                 ret = cma_connect_ib(id_priv, conn_param);
1941                 break;
1942         case RDMA_TRANSPORT_IWARP:
1943                 ret = cma_connect_iw(id_priv, conn_param);
1944                 break;
1945         default:
1946                 ret = -ENOSYS;
1947                 break;
1948         }
1949         if (ret)
1950                 goto err;
1951
1952         return 0;
1953 err:
1954         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
1955         return ret;
1956 }
1957 EXPORT_SYMBOL(rdma_connect);
1958
1959 static int cma_accept_ib(struct rdma_id_private *id_priv,
1960                          struct rdma_conn_param *conn_param)
1961 {
1962         struct ib_cm_rep_param rep;
1963         int ret;
1964
1965         ret = cma_modify_qp_rtr(&id_priv->id);
1966         if (ret)
1967                 return ret;
1968
1969         memset(&rep, 0, sizeof rep);
1970         rep.qp_num = id_priv->qp_num;
1971         rep.starting_psn = id_priv->seq_num;
1972         rep.private_data = conn_param->private_data;
1973         rep.private_data_len = conn_param->private_data_len;
1974         rep.responder_resources = conn_param->responder_resources;
1975         rep.initiator_depth = conn_param->initiator_depth;
1976         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
1977         rep.failover_accepted = 0;
1978         rep.flow_control = conn_param->flow_control;
1979         rep.rnr_retry_count = conn_param->rnr_retry_count;
1980         rep.srq = id_priv->srq ? 1 : 0;
1981
1982         return ib_send_cm_rep(id_priv->cm_id.ib, &rep);
1983 }
1984
1985 static int cma_accept_iw(struct rdma_id_private *id_priv,
1986                   struct rdma_conn_param *conn_param)
1987 {
1988         struct iw_cm_conn_param iw_param;
1989         int ret;
1990
1991         ret = cma_modify_qp_rtr(&id_priv->id);
1992         if (ret)
1993                 return ret;
1994
1995         iw_param.ord = conn_param->initiator_depth;
1996         iw_param.ird = conn_param->responder_resources;
1997         iw_param.private_data = conn_param->private_data;
1998         iw_param.private_data_len = conn_param->private_data_len;
1999         if (id_priv->id.qp) {
2000                 iw_param.qpn = id_priv->qp_num;
2001         } else
2002                 iw_param.qpn = conn_param->qp_num;
2003
2004         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2005 }
2006
2007 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2008 {
2009         struct rdma_id_private *id_priv;
2010         int ret;
2011
2012         id_priv = container_of(id, struct rdma_id_private, id);
2013         if (!cma_comp(id_priv, CMA_CONNECT))
2014                 return -EINVAL;
2015
2016         if (!id->qp && conn_param) {
2017                 id_priv->qp_num = conn_param->qp_num;
2018                 id_priv->qp_type = conn_param->qp_type;
2019                 id_priv->srq = conn_param->srq;
2020         }
2021
2022         switch (rdma_node_get_transport(id->device->node_type)) {
2023         case RDMA_TRANSPORT_IB:
2024                 if (conn_param)
2025                         ret = cma_accept_ib(id_priv, conn_param);
2026                 else
2027                         ret = cma_rep_recv(id_priv);
2028                 break;
2029         case RDMA_TRANSPORT_IWARP:
2030                 ret = cma_accept_iw(id_priv, conn_param);
2031                 break;
2032         default:
2033                 ret = -ENOSYS;
2034                 break;
2035         }
2036
2037         if (ret)
2038                 goto reject;
2039
2040         return 0;
2041 reject:
2042         cma_modify_qp_err(id);
2043         rdma_reject(id, NULL, 0);
2044         return ret;
2045 }
2046 EXPORT_SYMBOL(rdma_accept);
2047
2048 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2049                 u8 private_data_len)
2050 {
2051         struct rdma_id_private *id_priv;
2052         int ret;
2053
2054         id_priv = container_of(id, struct rdma_id_private, id);
2055         if (!cma_comp(id_priv, CMA_CONNECT))
2056                 return -EINVAL;
2057
2058         switch (rdma_node_get_transport(id->device->node_type)) {
2059         case RDMA_TRANSPORT_IB:
2060                 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2061                                      IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2062                                      private_data, private_data_len);
2063                 break;
2064         case RDMA_TRANSPORT_IWARP:
2065                 ret = iw_cm_reject(id_priv->cm_id.iw,
2066                                    private_data, private_data_len);
2067                 break;
2068         default:
2069                 ret = -ENOSYS;
2070                 break;
2071         }
2072         return ret;
2073 }
2074 EXPORT_SYMBOL(rdma_reject);
2075
2076 int rdma_disconnect(struct rdma_cm_id *id)
2077 {
2078         struct rdma_id_private *id_priv;
2079         int ret;
2080
2081         id_priv = container_of(id, struct rdma_id_private, id);
2082         if (!cma_comp(id_priv, CMA_CONNECT) &&
2083             !cma_comp(id_priv, CMA_DISCONNECT))
2084                 return -EINVAL;
2085
2086         switch (rdma_node_get_transport(id->device->node_type)) {
2087         case RDMA_TRANSPORT_IB:
2088                 ret = cma_modify_qp_err(id);
2089                 if (ret)
2090                         goto out;
2091                 /* Initiate or respond to a disconnect. */
2092                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2093                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2094                 break;
2095         case RDMA_TRANSPORT_IWARP:
2096                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2097                 break;
2098         default:
2099                 ret = -EINVAL;
2100                 break;
2101         }
2102 out:
2103         return ret;
2104 }
2105 EXPORT_SYMBOL(rdma_disconnect);
2106
2107 static void cma_add_one(struct ib_device *device)
2108 {
2109         struct cma_device *cma_dev;
2110         struct rdma_id_private *id_priv;
2111
2112         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2113         if (!cma_dev)
2114                 return;
2115
2116         cma_dev->device = device;
2117         cma_dev->node_guid = device->node_guid;
2118         if (!cma_dev->node_guid)
2119                 goto err;
2120
2121         init_completion(&cma_dev->comp);
2122         atomic_set(&cma_dev->refcount, 1);
2123         INIT_LIST_HEAD(&cma_dev->id_list);
2124         ib_set_client_data(device, &cma_client, cma_dev);
2125
2126         mutex_lock(&lock);
2127         list_add_tail(&cma_dev->list, &dev_list);
2128         list_for_each_entry(id_priv, &listen_any_list, list)
2129                 cma_listen_on_dev(id_priv, cma_dev);
2130         mutex_unlock(&lock);
2131         return;
2132 err:
2133         kfree(cma_dev);
2134 }
2135
2136 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2137 {
2138         enum cma_state state;
2139
2140         /* Record that we want to remove the device */
2141         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2142         if (state == CMA_DESTROYING)
2143                 return 0;
2144
2145         cma_cancel_operation(id_priv, state);
2146         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2147
2148         /* Check for destruction from another callback. */
2149         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2150                 return 0;
2151
2152         return cma_notify_user(id_priv, RDMA_CM_EVENT_DEVICE_REMOVAL,
2153                                0, NULL, 0);
2154 }
2155
2156 static void cma_process_remove(struct cma_device *cma_dev)
2157 {
2158         struct rdma_id_private *id_priv;
2159         int ret;
2160
2161         mutex_lock(&lock);
2162         while (!list_empty(&cma_dev->id_list)) {
2163                 id_priv = list_entry(cma_dev->id_list.next,
2164                                      struct rdma_id_private, list);
2165
2166                 if (cma_internal_listen(id_priv)) {
2167                         cma_destroy_listen(id_priv);
2168                         continue;
2169                 }
2170
2171                 list_del_init(&id_priv->list);
2172                 atomic_inc(&id_priv->refcount);
2173                 mutex_unlock(&lock);
2174
2175                 ret = cma_remove_id_dev(id_priv);
2176                 cma_deref_id(id_priv);
2177                 if (ret)
2178                         rdma_destroy_id(&id_priv->id);
2179
2180                 mutex_lock(&lock);
2181         }
2182         mutex_unlock(&lock);
2183
2184         cma_deref_dev(cma_dev);
2185         wait_for_completion(&cma_dev->comp);
2186 }
2187
2188 static void cma_remove_one(struct ib_device *device)
2189 {
2190         struct cma_device *cma_dev;
2191
2192         cma_dev = ib_get_client_data(device, &cma_client);
2193         if (!cma_dev)
2194                 return;
2195
2196         mutex_lock(&lock);
2197         list_del(&cma_dev->list);
2198         mutex_unlock(&lock);
2199
2200         cma_process_remove(cma_dev);
2201         kfree(cma_dev);
2202 }
2203
2204 static int cma_init(void)
2205 {
2206         int ret;
2207
2208         cma_wq = create_singlethread_workqueue("rdma_cm_wq");
2209         if (!cma_wq)
2210                 return -ENOMEM;
2211
2212         ib_sa_register_client(&sa_client);
2213
2214         ret = ib_register_client(&cma_client);
2215         if (ret)
2216                 goto err;
2217         return 0;
2218
2219 err:
2220         ib_sa_unregister_client(&sa_client);
2221         destroy_workqueue(cma_wq);
2222         return ret;
2223 }
2224
2225 static void cma_cleanup(void)
2226 {
2227         ib_unregister_client(&cma_client);
2228         ib_sa_unregister_client(&sa_client);
2229         destroy_workqueue(cma_wq);
2230         idr_destroy(&sdp_ps);
2231         idr_destroy(&tcp_ps);
2232 }
2233
2234 module_init(cma_init);
2235 module_exit(cma_cleanup);