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 struct rdma_addr_client addr_client;
  67 static LIST_HEAD(dev_list);
  68 static LIST_HEAD(listen_any_list);
  69 static DEFINE_MUTEX(lock);
  70 static struct workqueue_struct *cma_wq;
  71 static DEFINE_IDR(sdp_ps);
  72 static DEFINE_IDR(tcp_ps);
  73 static DEFINE_IDR(udp_ps);
  74
  75 struct cma_device {
  76         struct list_head        list;
  77         struct ib_device        *device;
  78         __be64                  node_guid;
  79         struct completion       comp;
  80         atomic_t                refcount;
  81         struct list_head        id_list;
  82 };
  83
  84 enum cma_state {
  85         CMA_IDLE,
  86         CMA_ADDR_QUERY,
  87         CMA_ADDR_RESOLVED,
  88         CMA_ROUTE_QUERY,
  89         CMA_ROUTE_RESOLVED,
  90         CMA_CONNECT,
  91         CMA_DISCONNECT,
  92         CMA_ADDR_BOUND,
  93         CMA_LISTEN,
  94         CMA_DEVICE_REMOVAL,
  95         CMA_DESTROYING
  96 };
  97
  98 struct rdma_bind_list {
  99         struct idr              *ps;
 100         struct hlist_head       owners;
 101         unsigned short          port;
 102 };
 103
 104 /*
 105  * Device removal can occur at anytime, so we need extra handling to
 106  * serialize notifying the user of device removal with other callbacks.
 107  * We do this by disabling removal notification while a callback is in process,
 108  * and reporting it after the callback completes.
 109  */
 110 struct rdma_id_private {
 111         struct rdma_cm_id       id;
 112
 113         struct rdma_bind_list   *bind_list;
 114         struct hlist_node       node;
 115         struct list_head        list;
 116         struct list_head        listen_list;
 117         struct cma_device       *cma_dev;
 118
 119         enum cma_state          state;
 120         spinlock_t              lock;
 121         struct completion       comp;
 122         atomic_t                refcount;
 123         wait_queue_head_t       wait_remove;
 124         atomic_t                dev_remove;
 125
 126         int                     backlog;
 127         int                     timeout_ms;
 128         struct ib_sa_query      *query;
 129         int                     query_id;
 130         union {
 131                 struct ib_cm_id *ib;
 132                 struct iw_cm_id *iw;
 133         } cm_id;
 134
 135         u32                     seq_num;
 136         u32                     qp_num;
 137         u8                      srq;
 138 };
 139
 140 struct cma_work {
 141         struct work_struct      work;
 142         struct rdma_id_private  *id;
 143         enum cma_state          old_state;
 144         enum cma_state          new_state;
 145         struct rdma_cm_event    event;
 146 };
 147
 148 union cma_ip_addr {
 149         struct in6_addr ip6;
 150         struct {
 151                 __u32 pad[3];
 152                 __u32 addr;
 153         } ip4;
 154 };
 155
 156 struct cma_hdr {
 157         u8 cma_version;
 158         u8 ip_version;  /* IP version: 7:4 */
 159         __u16 port;
 160         union cma_ip_addr src_addr;
 161         union cma_ip_addr dst_addr;
 162 };
 163
 164 struct sdp_hh {
 165         u8 bsdh[16];
 166         u8 sdp_version; /* Major version: 7:4 */
 167         u8 ip_version;  /* IP version: 7:4 */
 168         u8 sdp_specific1[10];
 169         __u16 port;
 170         __u16 sdp_specific2;
 171         union cma_ip_addr src_addr;
 172         union cma_ip_addr dst_addr;
 173 };
 174
 175 struct sdp_hah {
 176         u8 bsdh[16];
 177         u8 sdp_version;
 178 };
 179
 180 #define CMA_VERSION 0x00
 181 #define SDP_MAJ_VERSION 0x2
 182
 183 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 184 {
 185         unsigned long flags;
 186         int ret;
 187
 188         spin_lock_irqsave(&id_priv->lock, flags);
 189         ret = (id_priv->state == comp);
 190         spin_unlock_irqrestore(&id_priv->lock, flags);
 191         return ret;
 192 }
 193
 194 static int cma_comp_exch(struct rdma_id_private *id_priv,
 195                          enum cma_state comp, enum cma_state exch)
 196 {
 197         unsigned long flags;
 198         int ret;
 199
 200         spin_lock_irqsave(&id_priv->lock, flags);
 201         if ((ret = (id_priv->state == comp)))
 202                 id_priv->state = exch;
 203         spin_unlock_irqrestore(&id_priv->lock, flags);
 204         return ret;
 205 }
 206
 207 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 208                                enum cma_state exch)
 209 {
 210         unsigned long flags;
 211         enum cma_state old;
 212
 213         spin_lock_irqsave(&id_priv->lock, flags);
 214         old = id_priv->state;
 215         id_priv->state = exch;
 216         spin_unlock_irqrestore(&id_priv->lock, flags);
 217         return old;
 218 }
 219
 220 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 221 {
 222         return hdr->ip_version >> 4;
 223 }
 224
 225 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 226 {
 227         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 228 }
 229
 230 static inline u8 sdp_get_majv(u8 sdp_version)
 231 {
 232         return sdp_version >> 4;
 233 }
 234
 235 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 236 {
 237         return hh->ip_version >> 4;
 238 }
 239
 240 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 241 {
 242         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 243 }
 244
 245 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 246                               struct cma_device *cma_dev)
 247 {
 248         atomic_inc(&cma_dev->refcount);
 249         id_priv->cma_dev = cma_dev;
 250         id_priv->id.device = cma_dev->device;
 251         list_add_tail(&id_priv->list, &cma_dev->id_list);
 252 }
 253
 254 static inline void cma_deref_dev(struct cma_device *cma_dev)
 255 {
 256         if (atomic_dec_and_test(&cma_dev->refcount))
 257                 complete(&cma_dev->comp);
 258 }
 259
 260 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 261 {
 262         list_del(&id_priv->list);
 263         cma_deref_dev(id_priv->cma_dev);
 264         id_priv->cma_dev = NULL;
 265 }
 266
 267 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 268 {
 269         enum rdma_node_type dev_type = id_priv->id.route.addr.dev_addr.dev_type;
 270         struct cma_device *cma_dev;
 271         union ib_gid gid;
 272         int ret = -ENODEV;
 273
 274         switch (rdma_node_get_transport(dev_type)) {
 275         case RDMA_TRANSPORT_IB:
 276                 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
 277                 break;
 278         case RDMA_TRANSPORT_IWARP:
 279                 iw_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
 280                 break;
 281         default:
 282                 return -ENODEV;
 283         }
 284
 285         list_for_each_entry(cma_dev, &dev_list, list) {
 286                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 287                                          &id_priv->id.port_num, NULL);
 288                 if (!ret) {
 289                         cma_attach_to_dev(id_priv, cma_dev);
 290                         break;
 291                 }
 292         }
 293         return ret;
 294 }
 295
 296 static void cma_deref_id(struct rdma_id_private *id_priv)
 297 {
 298         if (atomic_dec_and_test(&id_priv->refcount))
 299                 complete(&id_priv->comp);
 300 }
 301
 302 static void cma_release_remove(struct rdma_id_private *id_priv)
 303 {
 304         if (atomic_dec_and_test(&id_priv->dev_remove))
 305                 wake_up(&id_priv->wait_remove);
 306 }
 307
 308 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 309                                   void *context, enum rdma_port_space ps)
 310 {
 311         struct rdma_id_private *id_priv;
 312
 313         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 314         if (!id_priv)
 315                 return ERR_PTR(-ENOMEM);
 316
 317         id_priv->state = CMA_IDLE;
 318         id_priv->id.context = context;
 319         id_priv->id.event_handler = event_handler;
 320         id_priv->id.ps = ps;
 321         spin_lock_init(&id_priv->lock);
 322         init_completion(&id_priv->comp);
 323         atomic_set(&id_priv->refcount, 1);
 324         init_waitqueue_head(&id_priv->wait_remove);
 325         atomic_set(&id_priv->dev_remove, 0);
 326         INIT_LIST_HEAD(&id_priv->listen_list);
 327         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 328
 329         return &id_priv->id;
 330 }
 331 EXPORT_SYMBOL(rdma_create_id);
 332
 333 static int cma_init_ib_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 334 {
 335         struct ib_qp_attr qp_attr;
 336         struct rdma_dev_addr *dev_addr;
 337         int ret;
 338
 339         dev_addr = &id_priv->id.route.addr.dev_addr;
 340         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 341                                   ib_addr_get_pkey(dev_addr),
 342                                   &qp_attr.pkey_index);
 343         if (ret)
 344                 return ret;
 345
 346         qp_attr.qp_state = IB_QPS_INIT;
 347         qp_attr.qp_access_flags = 0;
 348         qp_attr.port_num = id_priv->id.port_num;
 349         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS |
 350                                           IB_QP_PKEY_INDEX | IB_QP_PORT);
 351 }
 352
 353 static int cma_init_iw_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 354 {
 355         struct ib_qp_attr qp_attr;
 356
 357         qp_attr.qp_state = IB_QPS_INIT;
 358         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 359
 360         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS);
 361 }
 362
 363 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 364                    struct ib_qp_init_attr *qp_init_attr)
 365 {
 366         struct rdma_id_private *id_priv;
 367         struct ib_qp *qp;
 368         int ret;
 369
 370         id_priv = container_of(id, struct rdma_id_private, id);
 371         if (id->device != pd->device)
 372                 return -EINVAL;
 373
 374         qp = ib_create_qp(pd, qp_init_attr);
 375         if (IS_ERR(qp))
 376                 return PTR_ERR(qp);
 377
 378         switch (rdma_node_get_transport(id->device->node_type)) {
 379         case RDMA_TRANSPORT_IB:
 380                 ret = cma_init_ib_qp(id_priv, qp);
 381                 break;
 382         case RDMA_TRANSPORT_IWARP:
 383                 ret = cma_init_iw_qp(id_priv, qp);
 384                 break;
 385         default:
 386                 ret = -ENOSYS;
 387                 break;
 388         }
 389
 390         if (ret)
 391                 goto err;
 392
 393         id->qp = qp;
 394         id_priv->qp_num = qp->qp_num;
 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 __be16 cma_port(struct sockaddr *addr)
 513 {
 514         if (addr->sa_family == AF_INET)
 515                 return ((struct sockaddr_in *) addr)->sin_port;
 516         else
 517                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 518 }
 519
 520 static inline int cma_any_port(struct sockaddr *addr)
 521 {
 522         return !cma_port(addr);
 523 }
 524
 525 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 526                             u8 *ip_ver, __u16 *port,
 527                             union cma_ip_addr **src, union cma_ip_addr **dst)
 528 {
 529         switch (ps) {
 530         case RDMA_PS_SDP:
 531                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 532                     SDP_MAJ_VERSION)
 533                         return -EINVAL;
 534
 535                 *ip_ver = sdp_get_ip_ver(hdr);
 536                 *port   = ((struct sdp_hh *) hdr)->port;
 537                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 538                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 539                 break;
 540         default:
 541                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 542                         return -EINVAL;
 543
 544                 *ip_ver = cma_get_ip_ver(hdr);
 545                 *port   = ((struct cma_hdr *) hdr)->port;
 546                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 547                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 548                 break;
 549         }
 550
 551         if (*ip_ver != 4 && *ip_ver != 6)
 552                 return -EINVAL;
 553         return 0;
 554 }
 555
 556 static void cma_save_net_info(struct rdma_addr *addr,
 557                               struct rdma_addr *listen_addr,
 558                               u8 ip_ver, __u16 port,
 559                               union cma_ip_addr *src, union cma_ip_addr *dst)
 560 {
 561         struct sockaddr_in *listen4, *ip4;
 562         struct sockaddr_in6 *listen6, *ip6;
 563
 564         switch (ip_ver) {
 565         case 4:
 566                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 567                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 568                 ip4->sin_family = listen4->sin_family;
 569                 ip4->sin_addr.s_addr = dst->ip4.addr;
 570                 ip4->sin_port = listen4->sin_port;
 571
 572                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 573                 ip4->sin_family = listen4->sin_family;
 574                 ip4->sin_addr.s_addr = src->ip4.addr;
 575                 ip4->sin_port = port;
 576                 break;
 577         case 6:
 578                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 579                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 580                 ip6->sin6_family = listen6->sin6_family;
 581                 ip6->sin6_addr = dst->ip6;
 582                 ip6->sin6_port = listen6->sin6_port;
 583
 584                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 585                 ip6->sin6_family = listen6->sin6_family;
 586                 ip6->sin6_addr = src->ip6;
 587                 ip6->sin6_port = port;
 588                 break;
 589         default:
 590                 break;
 591         }
 592 }
 593
 594 static inline int cma_user_data_offset(enum rdma_port_space ps)
 595 {
 596         switch (ps) {
 597         case RDMA_PS_SDP:
 598                 return 0;
 599         default:
 600                 return sizeof(struct cma_hdr);
 601         }
 602 }
 603
 604 static void cma_cancel_route(struct rdma_id_private *id_priv)
 605 {
 606         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 607         case RDMA_TRANSPORT_IB:
 608                 if (id_priv->query)
 609                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 610                 break;
 611         default:
 612                 break;
 613         }
 614 }
 615
 616 static inline int cma_internal_listen(struct rdma_id_private *id_priv)
 617 {
 618         return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
 619                cma_any_addr(&id_priv->id.route.addr.src_addr);
 620 }
 621
 622 static void cma_destroy_listen(struct rdma_id_private *id_priv)
 623 {
 624         cma_exch(id_priv, CMA_DESTROYING);
 625
 626         if (id_priv->cma_dev) {
 627                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 628                 case RDMA_TRANSPORT_IB:
 629                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 630                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 631                         break;
 632                 case RDMA_TRANSPORT_IWARP:
 633                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 634                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 635                         break;
 636                 default:
 637                         break;
 638                 }
 639                 cma_detach_from_dev(id_priv);
 640         }
 641         list_del(&id_priv->listen_list);
 642
 643         cma_deref_id(id_priv);
 644         wait_for_completion(&id_priv->comp);
 645
 646         kfree(id_priv);
 647 }
 648
 649 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 650 {
 651         struct rdma_id_private *dev_id_priv;
 652
 653         mutex_lock(&lock);
 654         list_del(&id_priv->list);
 655
 656         while (!list_empty(&id_priv->listen_list)) {
 657                 dev_id_priv = list_entry(id_priv->listen_list.next,
 658                                          struct rdma_id_private, listen_list);
 659                 cma_destroy_listen(dev_id_priv);
 660         }
 661         mutex_unlock(&lock);
 662 }
 663
 664 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 665                                  enum cma_state state)
 666 {
 667         switch (state) {
 668         case CMA_ADDR_QUERY:
 669                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 670                 break;
 671         case CMA_ROUTE_QUERY:
 672                 cma_cancel_route(id_priv);
 673                 break;
 674         case CMA_LISTEN:
 675                 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
 676                     !id_priv->cma_dev)
 677                         cma_cancel_listens(id_priv);
 678                 break;
 679         default:
 680                 break;
 681         }
 682 }
 683
 684 static void cma_release_port(struct rdma_id_private *id_priv)
 685 {
 686         struct rdma_bind_list *bind_list = id_priv->bind_list;
 687
 688         if (!bind_list)
 689                 return;
 690
 691         mutex_lock(&lock);
 692         hlist_del(&id_priv->node);
 693         if (hlist_empty(&bind_list->owners)) {
 694                 idr_remove(bind_list->ps, bind_list->port);
 695                 kfree(bind_list);
 696         }
 697         mutex_unlock(&lock);
 698 }
 699
 700 void rdma_destroy_id(struct rdma_cm_id *id)
 701 {
 702         struct rdma_id_private *id_priv;
 703         enum cma_state state;
 704
 705         id_priv = container_of(id, struct rdma_id_private, id);
 706         state = cma_exch(id_priv, CMA_DESTROYING);
 707         cma_cancel_operation(id_priv, state);
 708
 709         mutex_lock(&lock);
 710         if (id_priv->cma_dev) {
 711                 mutex_unlock(&lock);
 712                 switch (rdma_node_get_transport(id->device->node_type)) {
 713                 case RDMA_TRANSPORT_IB:
 714                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 715                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 716                         break;
 717                 case RDMA_TRANSPORT_IWARP:
 718                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 719                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 720                         break;
 721                 default:
 722                         break;
 723                 }
 724                 mutex_lock(&lock);
 725                 cma_detach_from_dev(id_priv);
 726         }
 727         mutex_unlock(&lock);
 728
 729         cma_release_port(id_priv);
 730         cma_deref_id(id_priv);
 731         wait_for_completion(&id_priv->comp);
 732
 733         kfree(id_priv->id.route.path_rec);
 734         kfree(id_priv);
 735 }
 736 EXPORT_SYMBOL(rdma_destroy_id);
 737
 738 static int cma_rep_recv(struct rdma_id_private *id_priv)
 739 {
 740         int ret;
 741
 742         ret = cma_modify_qp_rtr(&id_priv->id);
 743         if (ret)
 744                 goto reject;
 745
 746         ret = cma_modify_qp_rts(&id_priv->id);
 747         if (ret)
 748                 goto reject;
 749
 750         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 751         if (ret)
 752                 goto reject;
 753
 754         return 0;
 755 reject:
 756         cma_modify_qp_err(&id_priv->id);
 757         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 758                        NULL, 0, NULL, 0);
 759         return ret;
 760 }
 761
 762 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 763 {
 764         if (id_priv->id.ps == RDMA_PS_SDP &&
 765             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 766             SDP_MAJ_VERSION)
 767                 return -EINVAL;
 768
 769         return 0;
 770 }
 771
 772 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 773                                    struct ib_cm_rep_event_param *rep_data,
 774                                    void *private_data)
 775 {
 776         event->param.conn.private_data = private_data;
 777         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 778         event->param.conn.responder_resources = rep_data->responder_resources;
 779         event->param.conn.initiator_depth = rep_data->initiator_depth;
 780         event->param.conn.flow_control = rep_data->flow_control;
 781         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 782         event->param.conn.srq = rep_data->srq;
 783         event->param.conn.qp_num = rep_data->remote_qpn;
 784 }
 785
 786 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 787 {
 788         struct rdma_id_private *id_priv = cm_id->context;
 789         struct rdma_cm_event event;
 790         int ret = 0;
 791
 792         atomic_inc(&id_priv->dev_remove);
 793         if (!cma_comp(id_priv, CMA_CONNECT))
 794                 goto out;
 795
 796         memset(&event, 0, sizeof event);
 797         switch (ib_event->event) {
 798         case IB_CM_REQ_ERROR:
 799         case IB_CM_REP_ERROR:
 800                 event.event = RDMA_CM_EVENT_UNREACHABLE;
 801                 event.status = -ETIMEDOUT;
 802                 break;
 803         case IB_CM_REP_RECEIVED:
 804                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
 805                 if (event.status)
 806                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
 807                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 808                         event.status = cma_rep_recv(id_priv);
 809                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
 810                                                      RDMA_CM_EVENT_ESTABLISHED;
 811                 } else
 812                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 813                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
 814                                        ib_event->private_data);
 815                 break;
 816         case IB_CM_RTU_RECEIVED:
 817         case IB_CM_USER_ESTABLISHED:
 818                 event.event = RDMA_CM_EVENT_ESTABLISHED;
 819                 break;
 820         case IB_CM_DREQ_ERROR:
 821                 event.status = -ETIMEDOUT; /* fall through */
 822         case IB_CM_DREQ_RECEIVED:
 823         case IB_CM_DREP_RECEIVED:
 824                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 825                         goto out;
 826                 event.event = RDMA_CM_EVENT_DISCONNECTED;
 827                 break;
 828         case IB_CM_TIMEWAIT_EXIT:
 829         case IB_CM_MRA_RECEIVED:
 830                 /* ignore event */
 831                 goto out;
 832         case IB_CM_REJ_RECEIVED:
 833                 cma_modify_qp_err(&id_priv->id);
 834                 event.status = ib_event->param.rej_rcvd.reason;
 835                 event.event = RDMA_CM_EVENT_REJECTED;
 836                 event.param.conn.private_data = ib_event->private_data;
 837                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 838                 break;
 839         default:
 840                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
 841                        ib_event->event);
 842                 goto out;
 843         }
 844
 845         ret = id_priv->id.event_handler(&id_priv->id, &event);
 846         if (ret) {
 847                 /* Destroy the CM ID by returning a non-zero value. */
 848                 id_priv->cm_id.ib = NULL;
 849                 cma_exch(id_priv, CMA_DESTROYING);
 850                 cma_release_remove(id_priv);
 851                 rdma_destroy_id(&id_priv->id);
 852                 return ret;
 853         }
 854 out:
 855         cma_release_remove(id_priv);
 856         return ret;
 857 }
 858
 859 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
 860                                                struct ib_cm_event *ib_event)
 861 {
 862         struct rdma_id_private *id_priv;
 863         struct rdma_cm_id *id;
 864         struct rdma_route *rt;
 865         union cma_ip_addr *src, *dst;
 866         __u16 port;
 867         u8 ip_ver;
 868
 869         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 870                              &ip_ver, &port, &src, &dst))
 871                 goto err;
 872
 873         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 874                             listen_id->ps);
 875         if (IS_ERR(id))
 876                 goto err;
 877
 878         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 879                           ip_ver, port, src, dst);
 880
 881         rt = &id->route;
 882         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
 883         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
 884                                GFP_KERNEL);
 885         if (!rt->path_rec)
 886                 goto destroy_id;
 887
 888         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
 889         if (rt->num_paths == 2)
 890                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
 891
 892         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
 893         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
 894         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
 895         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
 896
 897         id_priv = container_of(id, struct rdma_id_private, id);
 898         id_priv->state = CMA_CONNECT;
 899         return id_priv;
 900
 901 destroy_id:
 902         rdma_destroy_id(id);
 903 err:
 904         return NULL;
 905 }
 906
 907 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
 908                                               struct ib_cm_event *ib_event)
 909 {
 910         struct rdma_id_private *id_priv;
 911         struct rdma_cm_id *id;
 912         union cma_ip_addr *src, *dst;
 913         __u16 port;
 914         u8 ip_ver;
 915         int ret;
 916
 917         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 918                             listen_id->ps);
 919         if (IS_ERR(id))
 920                 return NULL;
 921
 922
 923         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 924                              &ip_ver, &port, &src, &dst))
 925                 goto err;
 926
 927         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 928                           ip_ver, port, src, dst);
 929
 930         ret = rdma_translate_ip(&id->route.addr.src_addr,
 931                                 &id->route.addr.dev_addr);
 932         if (ret)
 933                 goto err;
 934
 935         id_priv = container_of(id, struct rdma_id_private, id);
 936         id_priv->state = CMA_CONNECT;
 937         return id_priv;
 938 err:
 939         rdma_destroy_id(id);
 940         return NULL;
 941 }
 942
 943 static void cma_set_req_event_data(struct rdma_cm_event *event,
 944                                    struct ib_cm_req_event_param *req_data,
 945                                    void *private_data, int offset)
 946 {
 947         event->param.conn.private_data = private_data + offset;
 948         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
 949         event->param.conn.responder_resources = req_data->responder_resources;
 950         event->param.conn.initiator_depth = req_data->initiator_depth;
 951         event->param.conn.flow_control = req_data->flow_control;
 952         event->param.conn.retry_count = req_data->retry_count;
 953         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
 954         event->param.conn.srq = req_data->srq;
 955         event->param.conn.qp_num = req_data->remote_qpn;
 956 }
 957
 958 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 959 {
 960         struct rdma_id_private *listen_id, *conn_id;
 961         struct rdma_cm_event event;
 962         int offset, ret;
 963
 964         listen_id = cm_id->context;
 965         atomic_inc(&listen_id->dev_remove);
 966         if (!cma_comp(listen_id, CMA_LISTEN)) {
 967                 ret = -ECONNABORTED;
 968                 goto out;
 969         }
 970
 971         memset(&event, 0, sizeof event);
 972         offset = cma_user_data_offset(listen_id->id.ps);
 973         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
 974         if (listen_id->id.ps == RDMA_PS_UDP) {
 975                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
 976                 event.param.ud.private_data = ib_event->private_data + offset;
 977                 event.param.ud.private_data_len =
 978                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
 979         } else {
 980                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
 981                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
 982                                        ib_event->private_data, offset);
 983         }
 984         if (!conn_id) {
 985                 ret = -ENOMEM;
 986                 goto out;
 987         }
 988
 989         atomic_inc(&conn_id->dev_remove);
 990         mutex_lock(&lock);
 991         ret = cma_acquire_dev(conn_id);
 992         mutex_unlock(&lock);
 993         if (ret)
 994                 goto release_conn_id;
 995
 996         conn_id->cm_id.ib = cm_id;
 997         cm_id->context = conn_id;
 998         cm_id->cm_handler = cma_ib_handler;
 999
1000         ret = conn_id->id.event_handler(&conn_id->id, &event);
1001         if (!ret)
1002                 goto out;
1003
1004         /* Destroy the CM ID by returning a non-zero value. */
1005         conn_id->cm_id.ib = NULL;
1006
1007 release_conn_id:
1008         cma_exch(conn_id, CMA_DESTROYING);
1009         cma_release_remove(conn_id);
1010         rdma_destroy_id(&conn_id->id);
1011
1012 out:
1013         cma_release_remove(listen_id);
1014         return ret;
1015 }
1016
1017 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1018 {
1019         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1020 }
1021
1022 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1023                                  struct ib_cm_compare_data *compare)
1024 {
1025         struct cma_hdr *cma_data, *cma_mask;
1026         struct sdp_hh *sdp_data, *sdp_mask;
1027         __u32 ip4_addr;
1028         struct in6_addr ip6_addr;
1029
1030         memset(compare, 0, sizeof *compare);
1031         cma_data = (void *) compare->data;
1032         cma_mask = (void *) compare->mask;
1033         sdp_data = (void *) compare->data;
1034         sdp_mask = (void *) compare->mask;
1035
1036         switch (addr->sa_family) {
1037         case AF_INET:
1038                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1039                 if (ps == RDMA_PS_SDP) {
1040                         sdp_set_ip_ver(sdp_data, 4);
1041                         sdp_set_ip_ver(sdp_mask, 0xF);
1042                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1043                         sdp_mask->dst_addr.ip4.addr = ~0;
1044                 } else {
1045                         cma_set_ip_ver(cma_data, 4);
1046                         cma_set_ip_ver(cma_mask, 0xF);
1047                         cma_data->dst_addr.ip4.addr = ip4_addr;
1048                         cma_mask->dst_addr.ip4.addr = ~0;
1049                 }
1050                 break;
1051         case AF_INET6:
1052                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1053                 if (ps == RDMA_PS_SDP) {
1054                         sdp_set_ip_ver(sdp_data, 6);
1055                         sdp_set_ip_ver(sdp_mask, 0xF);
1056                         sdp_data->dst_addr.ip6 = ip6_addr;
1057                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1058                                sizeof sdp_mask->dst_addr.ip6);
1059                 } else {
1060                         cma_set_ip_ver(cma_data, 6);
1061                         cma_set_ip_ver(cma_mask, 0xF);
1062                         cma_data->dst_addr.ip6 = ip6_addr;
1063                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1064                                sizeof cma_mask->dst_addr.ip6);
1065                 }
1066                 break;
1067         default:
1068                 break;
1069         }
1070 }
1071
1072 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1073 {
1074         struct rdma_id_private *id_priv = iw_id->context;
1075         struct rdma_cm_event event;
1076         struct sockaddr_in *sin;
1077         int ret = 0;
1078
1079         memset(&event, 0, sizeof event);
1080         atomic_inc(&id_priv->dev_remove);
1081
1082         switch (iw_event->event) {
1083         case IW_CM_EVENT_CLOSE:
1084                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1085                 break;
1086         case IW_CM_EVENT_CONNECT_REPLY:
1087                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1088                 *sin = iw_event->local_addr;
1089                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1090                 *sin = iw_event->remote_addr;
1091                 if (iw_event->status)
1092                         event.event = RDMA_CM_EVENT_REJECTED;
1093                 else
1094                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1095                 break;
1096         case IW_CM_EVENT_ESTABLISHED:
1097                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1098                 break;
1099         default:
1100                 BUG_ON(1);
1101         }
1102
1103         event.status = iw_event->status;
1104         event.param.conn.private_data = iw_event->private_data;
1105         event.param.conn.private_data_len = iw_event->private_data_len;
1106         ret = id_priv->id.event_handler(&id_priv->id, &event);
1107         if (ret) {
1108                 /* Destroy the CM ID by returning a non-zero value. */
1109                 id_priv->cm_id.iw = NULL;
1110                 cma_exch(id_priv, CMA_DESTROYING);
1111                 cma_release_remove(id_priv);
1112                 rdma_destroy_id(&id_priv->id);
1113                 return ret;
1114         }
1115
1116         cma_release_remove(id_priv);
1117         return ret;
1118 }
1119
1120 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1121                                struct iw_cm_event *iw_event)
1122 {
1123         struct rdma_cm_id *new_cm_id;
1124         struct rdma_id_private *listen_id, *conn_id;
1125         struct sockaddr_in *sin;
1126         struct net_device *dev = NULL;
1127         struct rdma_cm_event event;
1128         int ret;
1129
1130         listen_id = cm_id->context;
1131         atomic_inc(&listen_id->dev_remove);
1132         if (!cma_comp(listen_id, CMA_LISTEN)) {
1133                 ret = -ECONNABORTED;
1134                 goto out;
1135         }
1136
1137         /* Create a new RDMA id for the new IW CM ID */
1138         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1139                                    listen_id->id.context,
1140                                    RDMA_PS_TCP);
1141         if (!new_cm_id) {
1142                 ret = -ENOMEM;
1143                 goto out;
1144         }
1145         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1146         atomic_inc(&conn_id->dev_remove);
1147         conn_id->state = CMA_CONNECT;
1148
1149         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1150         if (!dev) {
1151                 ret = -EADDRNOTAVAIL;
1152                 cma_release_remove(conn_id);
1153                 rdma_destroy_id(new_cm_id);
1154                 goto out;
1155         }
1156         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1157         if (ret) {
1158                 cma_release_remove(conn_id);
1159                 rdma_destroy_id(new_cm_id);
1160                 goto out;
1161         }
1162
1163         mutex_lock(&lock);
1164         ret = cma_acquire_dev(conn_id);
1165         mutex_unlock(&lock);
1166         if (ret) {
1167                 cma_release_remove(conn_id);
1168                 rdma_destroy_id(new_cm_id);
1169                 goto out;
1170         }
1171
1172         conn_id->cm_id.iw = cm_id;
1173         cm_id->context = conn_id;
1174         cm_id->cm_handler = cma_iw_handler;
1175
1176         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1177         *sin = iw_event->local_addr;
1178         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1179         *sin = iw_event->remote_addr;
1180
1181         memset(&event, 0, sizeof event);
1182         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1183         event.param.conn.private_data = iw_event->private_data;
1184         event.param.conn.private_data_len = iw_event->private_data_len;
1185         ret = conn_id->id.event_handler(&conn_id->id, &event);
1186         if (ret) {
1187                 /* User wants to destroy the CM ID */
1188                 conn_id->cm_id.iw = NULL;
1189                 cma_exch(conn_id, CMA_DESTROYING);
1190                 cma_release_remove(conn_id);
1191                 rdma_destroy_id(&conn_id->id);
1192         }
1193
1194 out:
1195         if (dev)
1196                 dev_put(dev);
1197         cma_release_remove(listen_id);
1198         return ret;
1199 }
1200
1201 static int cma_ib_listen(struct rdma_id_private *id_priv)
1202 {
1203         struct ib_cm_compare_data compare_data;
1204         struct sockaddr *addr;
1205         __be64 svc_id;
1206         int ret;
1207
1208         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1209                                             id_priv);
1210         if (IS_ERR(id_priv->cm_id.ib))
1211                 return PTR_ERR(id_priv->cm_id.ib);
1212
1213         addr = &id_priv->id.route.addr.src_addr;
1214         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1215         if (cma_any_addr(addr))
1216                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1217         else {
1218                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1219                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1220         }
1221
1222         if (ret) {
1223                 ib_destroy_cm_id(id_priv->cm_id.ib);
1224                 id_priv->cm_id.ib = NULL;
1225         }
1226
1227         return ret;
1228 }
1229
1230 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1231 {
1232         int ret;
1233         struct sockaddr_in *sin;
1234
1235         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1236                                             iw_conn_req_handler,
1237                                             id_priv);
1238         if (IS_ERR(id_priv->cm_id.iw))
1239                 return PTR_ERR(id_priv->cm_id.iw);
1240
1241         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1242         id_priv->cm_id.iw->local_addr = *sin;
1243
1244         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1245
1246         if (ret) {
1247                 iw_destroy_cm_id(id_priv->cm_id.iw);
1248                 id_priv->cm_id.iw = NULL;
1249         }
1250
1251         return ret;
1252 }
1253
1254 static int cma_listen_handler(struct rdma_cm_id *id,
1255                               struct rdma_cm_event *event)
1256 {
1257         struct rdma_id_private *id_priv = id->context;
1258
1259         id->context = id_priv->id.context;
1260         id->event_handler = id_priv->id.event_handler;
1261         return id_priv->id.event_handler(id, event);
1262 }
1263
1264 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1265                               struct cma_device *cma_dev)
1266 {
1267         struct rdma_id_private *dev_id_priv;
1268         struct rdma_cm_id *id;
1269         int ret;
1270
1271         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1272         if (IS_ERR(id))
1273                 return;
1274
1275         dev_id_priv = container_of(id, struct rdma_id_private, id);
1276
1277         dev_id_priv->state = CMA_ADDR_BOUND;
1278         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1279                ip_addr_size(&id_priv->id.route.addr.src_addr));
1280
1281         cma_attach_to_dev(dev_id_priv, cma_dev);
1282         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1283
1284         ret = rdma_listen(id, id_priv->backlog);
1285         if (ret)
1286                 goto err;
1287
1288         return;
1289 err:
1290         cma_destroy_listen(dev_id_priv);
1291 }
1292
1293 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1294 {
1295         struct cma_device *cma_dev;
1296
1297         mutex_lock(&lock);
1298         list_add_tail(&id_priv->list, &listen_any_list);
1299         list_for_each_entry(cma_dev, &dev_list, list)
1300                 cma_listen_on_dev(id_priv, cma_dev);
1301         mutex_unlock(&lock);
1302 }
1303
1304 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1305 {
1306         struct sockaddr_in addr_in;
1307
1308         memset(&addr_in, 0, sizeof addr_in);
1309         addr_in.sin_family = af;
1310         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1311 }
1312
1313 int rdma_listen(struct rdma_cm_id *id, int backlog)
1314 {
1315         struct rdma_id_private *id_priv;
1316         int ret;
1317
1318         id_priv = container_of(id, struct rdma_id_private, id);
1319         if (id_priv->state == CMA_IDLE) {
1320                 ret = cma_bind_any(id, AF_INET);
1321                 if (ret)
1322                         return ret;
1323         }
1324
1325         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1326                 return -EINVAL;
1327
1328         id_priv->backlog = backlog;
1329         if (id->device) {
1330                 switch (rdma_node_get_transport(id->device->node_type)) {
1331                 case RDMA_TRANSPORT_IB:
1332                         ret = cma_ib_listen(id_priv);
1333                         if (ret)
1334                                 goto err;
1335                         break;
1336                 case RDMA_TRANSPORT_IWARP:
1337                         ret = cma_iw_listen(id_priv, backlog);
1338                         if (ret)
1339                                 goto err;
1340                         break;
1341                 default:
1342                         ret = -ENOSYS;
1343                         goto err;
1344                 }
1345         } else
1346                 cma_listen_on_all(id_priv);
1347
1348         return 0;
1349 err:
1350         id_priv->backlog = 0;
1351         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1352         return ret;
1353 }
1354 EXPORT_SYMBOL(rdma_listen);
1355
1356 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1357                               void *context)
1358 {
1359         struct cma_work *work = context;
1360         struct rdma_route *route;
1361
1362         route = &work->id->id.route;
1363
1364         if (!status) {
1365                 route->num_paths = 1;
1366                 *route->path_rec = *path_rec;
1367         } else {
1368                 work->old_state = CMA_ROUTE_QUERY;
1369                 work->new_state = CMA_ADDR_RESOLVED;
1370                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1371                 work->event.status = status;
1372         }
1373
1374         queue_work(cma_wq, &work->work);
1375 }
1376
1377 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1378                               struct cma_work *work)
1379 {
1380         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1381         struct ib_sa_path_rec path_rec;
1382
1383         memset(&path_rec, 0, sizeof path_rec);
1384         ib_addr_get_sgid(addr, &path_rec.sgid);
1385         ib_addr_get_dgid(addr, &path_rec.dgid);
1386         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1387         path_rec.numb_path = 1;
1388
1389         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1390                                 id_priv->id.port_num, &path_rec,
1391                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1392                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
1393                                 timeout_ms, GFP_KERNEL,
1394                                 cma_query_handler, work, &id_priv->query);
1395
1396         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1397 }
1398
1399 static void cma_work_handler(struct work_struct *_work)
1400 {
1401         struct cma_work *work = container_of(_work, struct cma_work, work);
1402         struct rdma_id_private *id_priv = work->id;
1403         int destroy = 0;
1404
1405         atomic_inc(&id_priv->dev_remove);
1406         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1407                 goto out;
1408
1409         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1410                 cma_exch(id_priv, CMA_DESTROYING);
1411                 destroy = 1;
1412         }
1413 out:
1414         cma_release_remove(id_priv);
1415         cma_deref_id(id_priv);
1416         if (destroy)
1417                 rdma_destroy_id(&id_priv->id);
1418         kfree(work);
1419 }
1420
1421 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1422 {
1423         struct rdma_route *route = &id_priv->id.route;
1424         struct cma_work *work;
1425         int ret;
1426
1427         work = kzalloc(sizeof *work, GFP_KERNEL);
1428         if (!work)
1429                 return -ENOMEM;
1430
1431         work->id = id_priv;
1432         INIT_WORK(&work->work, cma_work_handler);
1433         work->old_state = CMA_ROUTE_QUERY;
1434         work->new_state = CMA_ROUTE_RESOLVED;
1435         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1436
1437         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1438         if (!route->path_rec) {
1439                 ret = -ENOMEM;
1440                 goto err1;
1441         }
1442
1443         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1444         if (ret)
1445                 goto err2;
1446
1447         return 0;
1448 err2:
1449         kfree(route->path_rec);
1450         route->path_rec = NULL;
1451 err1:
1452         kfree(work);
1453         return ret;
1454 }
1455
1456 int rdma_set_ib_paths(struct rdma_cm_id *id,
1457                       struct ib_sa_path_rec *path_rec, int num_paths)
1458 {
1459         struct rdma_id_private *id_priv;
1460         int ret;
1461
1462         id_priv = container_of(id, struct rdma_id_private, id);
1463         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1464                 return -EINVAL;
1465
1466         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1467         if (!id->route.path_rec) {
1468                 ret = -ENOMEM;
1469                 goto err;
1470         }
1471
1472         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1473         return 0;
1474 err:
1475         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1476         return ret;
1477 }
1478 EXPORT_SYMBOL(rdma_set_ib_paths);
1479
1480 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1481 {
1482         struct cma_work *work;
1483
1484         work = kzalloc(sizeof *work, GFP_KERNEL);
1485         if (!work)
1486                 return -ENOMEM;
1487
1488         work->id = id_priv;
1489         INIT_WORK(&work->work, cma_work_handler);
1490         work->old_state = CMA_ROUTE_QUERY;
1491         work->new_state = CMA_ROUTE_RESOLVED;
1492         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1493         queue_work(cma_wq, &work->work);
1494         return 0;
1495 }
1496
1497 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1498 {
1499         struct rdma_id_private *id_priv;
1500         int ret;
1501
1502         id_priv = container_of(id, struct rdma_id_private, id);
1503         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1504                 return -EINVAL;
1505
1506         atomic_inc(&id_priv->refcount);
1507         switch (rdma_node_get_transport(id->device->node_type)) {
1508         case RDMA_TRANSPORT_IB:
1509                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1510                 break;
1511         case RDMA_TRANSPORT_IWARP:
1512                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1513                 break;
1514         default:
1515                 ret = -ENOSYS;
1516                 break;
1517         }
1518         if (ret)
1519                 goto err;
1520
1521         return 0;
1522 err:
1523         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1524         cma_deref_id(id_priv);
1525         return ret;
1526 }
1527 EXPORT_SYMBOL(rdma_resolve_route);
1528
1529 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1530 {
1531         struct cma_device *cma_dev;
1532         struct ib_port_attr port_attr;
1533         union ib_gid gid;
1534         u16 pkey;
1535         int ret;
1536         u8 p;
1537
1538         mutex_lock(&lock);
1539         if (list_empty(&dev_list)) {
1540                 ret = -ENODEV;
1541                 goto out;
1542         }
1543         list_for_each_entry(cma_dev, &dev_list, list)
1544                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1545                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1546                             port_attr.state == IB_PORT_ACTIVE)
1547                                 goto port_found;
1548
1549         p = 1;
1550         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1551
1552 port_found:
1553         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1554         if (ret)
1555                 goto out;
1556
1557         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1558         if (ret)
1559                 goto out;
1560
1561         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1562         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1563         id_priv->id.port_num = p;
1564         cma_attach_to_dev(id_priv, cma_dev);
1565 out:
1566         mutex_unlock(&lock);
1567         return ret;
1568 }
1569
1570 static void addr_handler(int status, struct sockaddr *src_addr,
1571                          struct rdma_dev_addr *dev_addr, void *context)
1572 {
1573         struct rdma_id_private *id_priv = context;
1574         struct rdma_cm_event event;
1575
1576         memset(&event, 0, sizeof event);
1577         atomic_inc(&id_priv->dev_remove);
1578
1579         /*
1580          * Grab mutex to block rdma_destroy_id() from removing the device while
1581          * we're trying to acquire it.
1582          */
1583         mutex_lock(&lock);
1584         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1585                 mutex_unlock(&lock);
1586                 goto out;
1587         }
1588
1589         if (!status && !id_priv->cma_dev)
1590                 status = cma_acquire_dev(id_priv);
1591         mutex_unlock(&lock);
1592
1593         if (status) {
1594                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1595                         goto out;
1596                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1597                 event.status = status;
1598         } else {
1599                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1600                        ip_addr_size(src_addr));
1601                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1602         }
1603
1604         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1605                 cma_exch(id_priv, CMA_DESTROYING);
1606                 cma_release_remove(id_priv);
1607                 cma_deref_id(id_priv);
1608                 rdma_destroy_id(&id_priv->id);
1609                 return;
1610         }
1611 out:
1612         cma_release_remove(id_priv);
1613         cma_deref_id(id_priv);
1614 }
1615
1616 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1617 {
1618         struct cma_work *work;
1619         struct sockaddr_in *src_in, *dst_in;
1620         union ib_gid gid;
1621         int ret;
1622
1623         work = kzalloc(sizeof *work, GFP_KERNEL);
1624         if (!work)
1625                 return -ENOMEM;
1626
1627         if (!id_priv->cma_dev) {
1628                 ret = cma_bind_loopback(id_priv);
1629                 if (ret)
1630                         goto err;
1631         }
1632
1633         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1634         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1635
1636         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1637                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1638                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1639                 src_in->sin_family = dst_in->sin_family;
1640                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1641         }
1642
1643         work->id = id_priv;
1644         INIT_WORK(&work->work, cma_work_handler);
1645         work->old_state = CMA_ADDR_QUERY;
1646         work->new_state = CMA_ADDR_RESOLVED;
1647         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1648         queue_work(cma_wq, &work->work);
1649         return 0;
1650 err:
1651         kfree(work);
1652         return ret;
1653 }
1654
1655 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1656                          struct sockaddr *dst_addr)
1657 {
1658         if (src_addr && src_addr->sa_family)
1659                 return rdma_bind_addr(id, src_addr);
1660         else
1661                 return cma_bind_any(id, dst_addr->sa_family);
1662 }
1663
1664 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1665                       struct sockaddr *dst_addr, int timeout_ms)
1666 {
1667         struct rdma_id_private *id_priv;
1668         int ret;
1669
1670         id_priv = container_of(id, struct rdma_id_private, id);
1671         if (id_priv->state == CMA_IDLE) {
1672                 ret = cma_bind_addr(id, src_addr, dst_addr);
1673                 if (ret)
1674                         return ret;
1675         }
1676
1677         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1678                 return -EINVAL;
1679
1680         atomic_inc(&id_priv->refcount);
1681         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1682         if (cma_any_addr(dst_addr))
1683                 ret = cma_resolve_loopback(id_priv);
1684         else
1685                 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr,
1686                                       dst_addr, &id->route.addr.dev_addr,
1687                                       timeout_ms, addr_handler, id_priv);
1688         if (ret)
1689                 goto err;
1690
1691         return 0;
1692 err:
1693         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1694         cma_deref_id(id_priv);
1695         return ret;
1696 }
1697 EXPORT_SYMBOL(rdma_resolve_addr);
1698
1699 static void cma_bind_port(struct rdma_bind_list *bind_list,
1700                           struct rdma_id_private *id_priv)
1701 {
1702         struct sockaddr_in *sin;
1703
1704         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1705         sin->sin_port = htons(bind_list->port);
1706         id_priv->bind_list = bind_list;
1707         hlist_add_head(&id_priv->node, &bind_list->owners);
1708 }
1709
1710 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1711                           unsigned short snum)
1712 {
1713         struct rdma_bind_list *bind_list;
1714         int port, start, ret;
1715
1716         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1717         if (!bind_list)
1718                 return -ENOMEM;
1719
1720         start = snum ? snum : sysctl_local_port_range[0];
1721
1722         do {
1723                 ret = idr_get_new_above(ps, bind_list, start, &port);
1724         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1725
1726         if (ret)
1727                 goto err;
1728
1729         if ((snum && port != snum) ||
1730             (!snum && port > sysctl_local_port_range[1])) {
1731                 idr_remove(ps, port);
1732                 ret = -EADDRNOTAVAIL;
1733                 goto err;
1734         }
1735
1736         bind_list->ps = ps;
1737         bind_list->port = (unsigned short) port;
1738         cma_bind_port(bind_list, id_priv);
1739         return 0;
1740 err:
1741         kfree(bind_list);
1742         return ret;
1743 }
1744
1745 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1746 {
1747         struct rdma_id_private *cur_id;
1748         struct sockaddr_in *sin, *cur_sin;
1749         struct rdma_bind_list *bind_list;
1750         struct hlist_node *node;
1751         unsigned short snum;
1752
1753         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1754         snum = ntohs(sin->sin_port);
1755         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1756                 return -EACCES;
1757
1758         bind_list = idr_find(ps, snum);
1759         if (!bind_list)
1760                 return cma_alloc_port(ps, id_priv, snum);
1761
1762         /*
1763          * We don't support binding to any address if anyone is bound to
1764          * a specific address on the same port.
1765          */
1766         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1767                 return -EADDRNOTAVAIL;
1768
1769         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1770                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1771                         return -EADDRNOTAVAIL;
1772
1773                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1774                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1775                         return -EADDRINUSE;
1776         }
1777
1778         cma_bind_port(bind_list, id_priv);
1779         return 0;
1780 }
1781
1782 static int cma_get_port(struct rdma_id_private *id_priv)
1783 {
1784         struct idr *ps;
1785         int ret;
1786
1787         switch (id_priv->id.ps) {
1788         case RDMA_PS_SDP:
1789                 ps = &sdp_ps;
1790                 break;
1791         case RDMA_PS_TCP:
1792                 ps = &tcp_ps;
1793                 break;
1794         case RDMA_PS_UDP:
1795                 ps = &udp_ps;
1796                 break;
1797         default:
1798                 return -EPROTONOSUPPORT;
1799         }
1800
1801         mutex_lock(&lock);
1802         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1803                 ret = cma_alloc_port(ps, id_priv, 0);
1804         else
1805                 ret = cma_use_port(ps, id_priv);
1806         mutex_unlock(&lock);
1807
1808         return ret;
1809 }
1810
1811 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1812 {
1813         struct rdma_id_private *id_priv;
1814         int ret;
1815
1816         if (addr->sa_family != AF_INET)
1817                 return -EAFNOSUPPORT;
1818
1819         id_priv = container_of(id, struct rdma_id_private, id);
1820         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1821                 return -EINVAL;
1822
1823         if (!cma_any_addr(addr)) {
1824                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1825                 if (ret)
1826                         goto err1;
1827
1828                 mutex_lock(&lock);
1829                 ret = cma_acquire_dev(id_priv);
1830                 mutex_unlock(&lock);
1831                 if (ret)
1832                         goto err1;
1833         }
1834
1835         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
1836         ret = cma_get_port(id_priv);
1837         if (ret)
1838                 goto err2;
1839
1840         return 0;
1841 err2:
1842         if (!cma_any_addr(addr)) {
1843                 mutex_lock(&lock);
1844                 cma_detach_from_dev(id_priv);
1845                 mutex_unlock(&lock);
1846         }
1847 err1:
1848         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
1849         return ret;
1850 }
1851 EXPORT_SYMBOL(rdma_bind_addr);
1852
1853 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
1854                           struct rdma_route *route)
1855 {
1856         struct sockaddr_in *src4, *dst4;
1857         struct cma_hdr *cma_hdr;
1858         struct sdp_hh *sdp_hdr;
1859
1860         src4 = (struct sockaddr_in *) &route->addr.src_addr;
1861         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
1862
1863         switch (ps) {
1864         case RDMA_PS_SDP:
1865                 sdp_hdr = hdr;
1866                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
1867                         return -EINVAL;
1868                 sdp_set_ip_ver(sdp_hdr, 4);
1869                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1870                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1871                 sdp_hdr->port = src4->sin_port;
1872                 break;
1873         default:
1874                 cma_hdr = hdr;
1875                 cma_hdr->cma_version = CMA_VERSION;
1876                 cma_set_ip_ver(cma_hdr, 4);
1877                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1878                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1879                 cma_hdr->port = src4->sin_port;
1880                 break;
1881         }
1882         return 0;
1883 }
1884
1885 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
1886                                 struct ib_cm_event *ib_event)
1887 {
1888         struct rdma_id_private *id_priv = cm_id->context;
1889         struct rdma_cm_event event;
1890         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
1891         int ret = 0;
1892
1893         memset(&event, 0, sizeof event);
1894         atomic_inc(&id_priv->dev_remove);
1895         if (!cma_comp(id_priv, CMA_CONNECT))
1896                 goto out;
1897
1898         switch (ib_event->event) {
1899         case IB_CM_SIDR_REQ_ERROR:
1900                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1901                 event.status = -ETIMEDOUT;
1902                 break;
1903         case IB_CM_SIDR_REP_RECEIVED:
1904                 event.param.ud.private_data = ib_event->private_data;
1905                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
1906                 if (rep->status != IB_SIDR_SUCCESS) {
1907                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1908                         event.status = ib_event->param.sidr_rep_rcvd.status;
1909                         break;
1910                 }
1911                 if (rep->qkey != RDMA_UD_QKEY) {
1912                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1913                         event.status = -EINVAL;
1914                         break;
1915                 }
1916                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
1917                                      id_priv->id.route.path_rec,
1918                                      &event.param.ud.ah_attr);
1919                 event.param.ud.qp_num = rep->qpn;
1920                 event.param.ud.qkey = rep->qkey;
1921                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1922                 event.status = 0;
1923                 break;
1924         default:
1925                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
1926                        ib_event->event);
1927                 goto out;
1928         }
1929
1930         ret = id_priv->id.event_handler(&id_priv->id, &event);
1931         if (ret) {
1932                 /* Destroy the CM ID by returning a non-zero value. */
1933                 id_priv->cm_id.ib = NULL;
1934                 cma_exch(id_priv, CMA_DESTROYING);
1935                 cma_release_remove(id_priv);
1936                 rdma_destroy_id(&id_priv->id);
1937                 return ret;
1938         }
1939 out:
1940         cma_release_remove(id_priv);
1941         return ret;
1942 }
1943
1944 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
1945                               struct rdma_conn_param *conn_param)
1946 {
1947         struct ib_cm_sidr_req_param req;
1948         struct rdma_route *route;
1949         int ret;
1950
1951         req.private_data_len = sizeof(struct cma_hdr) +
1952                                conn_param->private_data_len;
1953         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
1954         if (!req.private_data)
1955                 return -ENOMEM;
1956
1957         if (conn_param->private_data && conn_param->private_data_len)
1958                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
1959                        conn_param->private_data, conn_param->private_data_len);
1960
1961         route = &id_priv->id.route;
1962         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
1963         if (ret)
1964                 goto out;
1965
1966         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
1967                                             cma_sidr_rep_handler, id_priv);
1968         if (IS_ERR(id_priv->cm_id.ib)) {
1969                 ret = PTR_ERR(id_priv->cm_id.ib);
1970                 goto out;
1971         }
1972
1973         req.path = route->path_rec;
1974         req.service_id = cma_get_service_id(id_priv->id.ps,
1975                                             &route->addr.dst_addr);
1976         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
1977         req.max_cm_retries = CMA_MAX_CM_RETRIES;
1978
1979         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
1980         if (ret) {
1981                 ib_destroy_cm_id(id_priv->cm_id.ib);
1982                 id_priv->cm_id.ib = NULL;
1983         }
1984 out:
1985         kfree(req.private_data);
1986         return ret;
1987 }
1988
1989 static int cma_connect_ib(struct rdma_id_private *id_priv,
1990                           struct rdma_conn_param *conn_param)
1991 {
1992         struct ib_cm_req_param req;
1993         struct rdma_route *route;
1994         void *private_data;
1995         int offset, ret;
1996
1997         memset(&req, 0, sizeof req);
1998         offset = cma_user_data_offset(id_priv->id.ps);
1999         req.private_data_len = offset + conn_param->private_data_len;
2000         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2001         if (!private_data)
2002                 return -ENOMEM;
2003
2004         if (conn_param->private_data && conn_param->private_data_len)
2005                 memcpy(private_data + offset, conn_param->private_data,
2006                        conn_param->private_data_len);
2007
2008         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2009                                             id_priv);
2010         if (IS_ERR(id_priv->cm_id.ib)) {
2011                 ret = PTR_ERR(id_priv->cm_id.ib);
2012                 goto out;
2013         }
2014
2015         route = &id_priv->id.route;
2016         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2017         if (ret)
2018                 goto out;
2019         req.private_data = private_data;
2020
2021         req.primary_path = &route->path_rec[0];
2022         if (route->num_paths == 2)
2023                 req.alternate_path = &route->path_rec[1];
2024
2025         req.service_id = cma_get_service_id(id_priv->id.ps,
2026                                             &route->addr.dst_addr);
2027         req.qp_num = id_priv->qp_num;
2028         req.qp_type = IB_QPT_RC;
2029         req.starting_psn = id_priv->seq_num;
2030         req.responder_resources = conn_param->responder_resources;
2031         req.initiator_depth = conn_param->initiator_depth;
2032         req.flow_control = conn_param->flow_control;
2033         req.retry_count = conn_param->retry_count;
2034         req.rnr_retry_count = conn_param->rnr_retry_count;
2035         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2036         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2037         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2038         req.srq = id_priv->srq ? 1 : 0;
2039
2040         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2041 out:
2042         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2043                 ib_destroy_cm_id(id_priv->cm_id.ib);
2044                 id_priv->cm_id.ib = NULL;
2045         }
2046
2047         kfree(private_data);
2048         return ret;
2049 }
2050
2051 static int cma_connect_iw(struct rdma_id_private *id_priv,
2052                           struct rdma_conn_param *conn_param)
2053 {
2054         struct iw_cm_id *cm_id;
2055         struct sockaddr_in* sin;
2056         int ret;
2057         struct iw_cm_conn_param iw_param;
2058
2059         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2060         if (IS_ERR(cm_id)) {
2061                 ret = PTR_ERR(cm_id);
2062                 goto out;
2063         }
2064
2065         id_priv->cm_id.iw = cm_id;
2066
2067         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2068         cm_id->local_addr = *sin;
2069
2070         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2071         cm_id->remote_addr = *sin;
2072
2073         ret = cma_modify_qp_rtr(&id_priv->id);
2074         if (ret)
2075                 goto out;
2076
2077         iw_param.ord = conn_param->initiator_depth;
2078         iw_param.ird = conn_param->responder_resources;
2079         iw_param.private_data = conn_param->private_data;
2080         iw_param.private_data_len = conn_param->private_data_len;
2081         if (id_priv->id.qp)
2082                 iw_param.qpn = id_priv->qp_num;
2083         else
2084                 iw_param.qpn = conn_param->qp_num;
2085         ret = iw_cm_connect(cm_id, &iw_param);
2086 out:
2087         if (ret && !IS_ERR(cm_id)) {
2088                 iw_destroy_cm_id(cm_id);
2089                 id_priv->cm_id.iw = NULL;
2090         }
2091         return ret;
2092 }
2093
2094 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2095 {
2096         struct rdma_id_private *id_priv;
2097         int ret;
2098
2099         id_priv = container_of(id, struct rdma_id_private, id);
2100         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2101                 return -EINVAL;
2102
2103         if (!id->qp) {
2104                 id_priv->qp_num = conn_param->qp_num;
2105                 id_priv->srq = conn_param->srq;
2106         }
2107
2108         switch (rdma_node_get_transport(id->device->node_type)) {
2109         case RDMA_TRANSPORT_IB:
2110                 if (id->ps == RDMA_PS_UDP)
2111                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2112                 else
2113                         ret = cma_connect_ib(id_priv, conn_param);
2114                 break;
2115         case RDMA_TRANSPORT_IWARP:
2116                 ret = cma_connect_iw(id_priv, conn_param);
2117                 break;
2118         default:
2119                 ret = -ENOSYS;
2120                 break;
2121         }
2122         if (ret)
2123                 goto err;
2124
2125         return 0;
2126 err:
2127         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2128         return ret;
2129 }
2130 EXPORT_SYMBOL(rdma_connect);
2131
2132 static int cma_accept_ib(struct rdma_id_private *id_priv,
2133                          struct rdma_conn_param *conn_param)
2134 {
2135         struct ib_cm_rep_param rep;
2136         struct ib_qp_attr qp_attr;
2137         int qp_attr_mask, ret;
2138
2139         if (id_priv->id.qp) {
2140                 ret = cma_modify_qp_rtr(&id_priv->id);
2141                 if (ret)
2142                         goto out;
2143
2144                 qp_attr.qp_state = IB_QPS_RTS;
2145                 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, &qp_attr,
2146                                          &qp_attr_mask);
2147                 if (ret)
2148                         goto out;
2149
2150                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
2151                 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
2152                 if (ret)
2153                         goto out;
2154         }
2155
2156         memset(&rep, 0, sizeof rep);
2157         rep.qp_num = id_priv->qp_num;
2158         rep.starting_psn = id_priv->seq_num;
2159         rep.private_data = conn_param->private_data;
2160         rep.private_data_len = conn_param->private_data_len;
2161         rep.responder_resources = conn_param->responder_resources;
2162         rep.initiator_depth = conn_param->initiator_depth;
2163         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
2164         rep.failover_accepted = 0;
2165         rep.flow_control = conn_param->flow_control;
2166         rep.rnr_retry_count = conn_param->rnr_retry_count;
2167         rep.srq = id_priv->srq ? 1 : 0;
2168
2169         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2170 out:
2171         return ret;
2172 }
2173
2174 static int cma_accept_iw(struct rdma_id_private *id_priv,
2175                   struct rdma_conn_param *conn_param)
2176 {
2177         struct iw_cm_conn_param iw_param;
2178         int ret;
2179
2180         ret = cma_modify_qp_rtr(&id_priv->id);
2181         if (ret)
2182                 return ret;
2183
2184         iw_param.ord = conn_param->initiator_depth;
2185         iw_param.ird = conn_param->responder_resources;
2186         iw_param.private_data = conn_param->private_data;
2187         iw_param.private_data_len = conn_param->private_data_len;
2188         if (id_priv->id.qp) {
2189                 iw_param.qpn = id_priv->qp_num;
2190         } else
2191                 iw_param.qpn = conn_param->qp_num;
2192
2193         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2194 }
2195
2196 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2197                              enum ib_cm_sidr_status status,
2198                              const void *private_data, int private_data_len)
2199 {
2200         struct ib_cm_sidr_rep_param rep;
2201
2202         memset(&rep, 0, sizeof rep);
2203         rep.status = status;
2204         if (status == IB_SIDR_SUCCESS) {
2205                 rep.qp_num = id_priv->qp_num;
2206                 rep.qkey = RDMA_UD_QKEY;
2207         }
2208         rep.private_data = private_data;
2209         rep.private_data_len = private_data_len;
2210
2211         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2212 }
2213
2214 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2215 {
2216         struct rdma_id_private *id_priv;
2217         int ret;
2218
2219         id_priv = container_of(id, struct rdma_id_private, id);
2220         if (!cma_comp(id_priv, CMA_CONNECT))
2221                 return -EINVAL;
2222
2223         if (!id->qp && conn_param) {
2224                 id_priv->qp_num = conn_param->qp_num;
2225                 id_priv->srq = conn_param->srq;
2226         }
2227
2228         switch (rdma_node_get_transport(id->device->node_type)) {
2229         case RDMA_TRANSPORT_IB:
2230                 if (id->ps == RDMA_PS_UDP)
2231                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2232                                                 conn_param->private_data,
2233                                                 conn_param->private_data_len);
2234                 else if (conn_param)
2235                         ret = cma_accept_ib(id_priv, conn_param);
2236                 else
2237                         ret = cma_rep_recv(id_priv);
2238                 break;
2239         case RDMA_TRANSPORT_IWARP:
2240                 ret = cma_accept_iw(id_priv, conn_param);
2241                 break;
2242         default:
2243                 ret = -ENOSYS;
2244                 break;
2245         }
2246
2247         if (ret)
2248                 goto reject;
2249
2250         return 0;
2251 reject:
2252         cma_modify_qp_err(id);
2253         rdma_reject(id, NULL, 0);
2254         return ret;
2255 }
2256 EXPORT_SYMBOL(rdma_accept);
2257
2258 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2259 {
2260         struct rdma_id_private *id_priv;
2261         int ret;
2262
2263         id_priv = container_of(id, struct rdma_id_private, id);
2264         if (!cma_comp(id_priv, CMA_CONNECT))
2265                 return -EINVAL;
2266
2267         switch (id->device->node_type) {
2268         case RDMA_NODE_IB_CA:
2269                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2270                 break;
2271         default:
2272                 ret = 0;
2273                 break;
2274         }
2275         return ret;
2276 }
2277 EXPORT_SYMBOL(rdma_notify);
2278
2279 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2280                 u8 private_data_len)
2281 {
2282         struct rdma_id_private *id_priv;
2283         int ret;
2284
2285         id_priv = container_of(id, struct rdma_id_private, id);
2286         if (!cma_comp(id_priv, CMA_CONNECT))
2287                 return -EINVAL;
2288
2289         switch (rdma_node_get_transport(id->device->node_type)) {
2290         case RDMA_TRANSPORT_IB:
2291                 if (id->ps == RDMA_PS_UDP)
2292                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2293                                                 private_data, private_data_len);
2294                 else
2295                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2296                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2297                                              0, private_data, private_data_len);
2298                 break;
2299         case RDMA_TRANSPORT_IWARP:
2300                 ret = iw_cm_reject(id_priv->cm_id.iw,
2301                                    private_data, private_data_len);
2302                 break;
2303         default:
2304                 ret = -ENOSYS;
2305                 break;
2306         }
2307         return ret;
2308 }
2309 EXPORT_SYMBOL(rdma_reject);
2310
2311 int rdma_disconnect(struct rdma_cm_id *id)
2312 {
2313         struct rdma_id_private *id_priv;
2314         int ret;
2315
2316         id_priv = container_of(id, struct rdma_id_private, id);
2317         if (!cma_comp(id_priv, CMA_CONNECT) &&
2318             !cma_comp(id_priv, CMA_DISCONNECT))
2319                 return -EINVAL;
2320
2321         switch (rdma_node_get_transport(id->device->node_type)) {
2322         case RDMA_TRANSPORT_IB:
2323                 ret = cma_modify_qp_err(id);
2324                 if (ret)
2325                         goto out;
2326                 /* Initiate or respond to a disconnect. */
2327                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2328                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2329                 break;
2330         case RDMA_TRANSPORT_IWARP:
2331                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2332                 break;
2333         default:
2334                 ret = -EINVAL;
2335                 break;
2336         }
2337 out:
2338         return ret;
2339 }
2340 EXPORT_SYMBOL(rdma_disconnect);
2341
2342 static void cma_add_one(struct ib_device *device)
2343 {
2344         struct cma_device *cma_dev;
2345         struct rdma_id_private *id_priv;
2346
2347         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2348         if (!cma_dev)
2349                 return;
2350
2351         cma_dev->device = device;
2352         cma_dev->node_guid = device->node_guid;
2353
2354         init_completion(&cma_dev->comp);
2355         atomic_set(&cma_dev->refcount, 1);
2356         INIT_LIST_HEAD(&cma_dev->id_list);
2357         ib_set_client_data(device, &cma_client, cma_dev);
2358
2359         mutex_lock(&lock);
2360         list_add_tail(&cma_dev->list, &dev_list);
2361         list_for_each_entry(id_priv, &listen_any_list, list)
2362                 cma_listen_on_dev(id_priv, cma_dev);
2363         mutex_unlock(&lock);
2364 }
2365
2366 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2367 {
2368         struct rdma_cm_event event;
2369         enum cma_state state;
2370
2371         /* Record that we want to remove the device */
2372         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2373         if (state == CMA_DESTROYING)
2374                 return 0;
2375
2376         cma_cancel_operation(id_priv, state);
2377         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2378
2379         /* Check for destruction from another callback. */
2380         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2381                 return 0;
2382
2383         memset(&event, 0, sizeof event);
2384         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2385         return id_priv->id.event_handler(&id_priv->id, &event);
2386 }
2387
2388 static void cma_process_remove(struct cma_device *cma_dev)
2389 {
2390         struct rdma_id_private *id_priv;
2391         int ret;
2392
2393         mutex_lock(&lock);
2394         while (!list_empty(&cma_dev->id_list)) {
2395                 id_priv = list_entry(cma_dev->id_list.next,
2396                                      struct rdma_id_private, list);
2397
2398                 if (cma_internal_listen(id_priv)) {
2399                         cma_destroy_listen(id_priv);
2400                         continue;
2401                 }
2402
2403                 list_del_init(&id_priv->list);
2404                 atomic_inc(&id_priv->refcount);
2405                 mutex_unlock(&lock);
2406
2407                 ret = cma_remove_id_dev(id_priv);
2408                 cma_deref_id(id_priv);
2409                 if (ret)
2410                         rdma_destroy_id(&id_priv->id);
2411
2412                 mutex_lock(&lock);
2413         }
2414         mutex_unlock(&lock);
2415
2416         cma_deref_dev(cma_dev);
2417         wait_for_completion(&cma_dev->comp);
2418 }
2419
2420 static void cma_remove_one(struct ib_device *device)
2421 {
2422         struct cma_device *cma_dev;
2423
2424         cma_dev = ib_get_client_data(device, &cma_client);
2425         if (!cma_dev)
2426                 return;
2427
2428         mutex_lock(&lock);
2429         list_del(&cma_dev->list);
2430         mutex_unlock(&lock);
2431
2432         cma_process_remove(cma_dev);
2433         kfree(cma_dev);
2434 }
2435
2436 static int cma_init(void)
2437 {
2438         int ret;
2439
2440         cma_wq = create_singlethread_workqueue("rdma_cm_wq");
2441         if (!cma_wq)
2442                 return -ENOMEM;
2443
2444         ib_sa_register_client(&sa_client);
2445         rdma_addr_register_client(&addr_client);
2446
2447         ret = ib_register_client(&cma_client);
2448         if (ret)
2449                 goto err;
2450         return 0;
2451
2452 err:
2453         rdma_addr_unregister_client(&addr_client);
2454         ib_sa_unregister_client(&sa_client);
2455         destroy_workqueue(cma_wq);
2456         return ret;
2457 }
2458
2459 static void cma_cleanup(void)
2460 {
2461         ib_unregister_client(&cma_client);
2462         rdma_addr_unregister_client(&addr_client);
2463         ib_sa_unregister_client(&sa_client);
2464         destroy_workqueue(cma_wq);
2465         idr_destroy(&sdp_ps);
2466         idr_destroy(&tcp_ps);
2467         idr_destroy(&udp_ps);
2468 }
2469
2470 module_init(cma_init);
2471 module_exit(cma_cleanup);