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 static DEFINE_IDR(ipoib_ps);
  75 static int next_port;
  76
  77 struct cma_device {
  78         struct list_head        list;
  79         struct ib_device        *device;
  80         struct completion       comp;
  81         atomic_t                refcount;
  82         struct list_head        id_list;
  83 };
  84
  85 enum cma_state {
  86         CMA_IDLE,
  87         CMA_ADDR_QUERY,
  88         CMA_ADDR_RESOLVED,
  89         CMA_ROUTE_QUERY,
  90         CMA_ROUTE_RESOLVED,
  91         CMA_CONNECT,
  92         CMA_DISCONNECT,
  93         CMA_ADDR_BOUND,
  94         CMA_LISTEN,
  95         CMA_DEVICE_REMOVAL,
  96         CMA_DESTROYING
  97 };
  98
  99 struct rdma_bind_list {
 100         struct idr              *ps;
 101         struct hlist_head       owners;
 102         unsigned short          port;
 103 };
 104
 105 /*
 106  * Device removal can occur at anytime, so we need extra handling to
 107  * serialize notifying the user of device removal with other callbacks.
 108  * We do this by disabling removal notification while a callback is in process,
 109  * and reporting it after the callback completes.
 110  */
 111 struct rdma_id_private {
 112         struct rdma_cm_id       id;
 113
 114         struct rdma_bind_list   *bind_list;
 115         struct hlist_node       node;
 116         struct list_head        list;
 117         struct list_head        listen_list;
 118         struct cma_device       *cma_dev;
 119         struct list_head        mc_list;
 120
 121         enum cma_state          state;
 122         spinlock_t              lock;
 123         struct completion       comp;
 124         atomic_t                refcount;
 125         wait_queue_head_t       wait_remove;
 126         atomic_t                dev_remove;
 127
 128         int                     backlog;
 129         int                     timeout_ms;
 130         struct ib_sa_query      *query;
 131         int                     query_id;
 132         union {
 133                 struct ib_cm_id *ib;
 134                 struct iw_cm_id *iw;
 135         } cm_id;
 136
 137         u32                     seq_num;
 138         u32                     qkey;
 139         u32                     qp_num;
 140         u8                      srq;
 141 };
 142
 143 struct cma_multicast {
 144         struct rdma_id_private *id_priv;
 145         union {
 146                 struct ib_sa_multicast *ib;
 147         } multicast;
 148         struct list_head        list;
 149         void                    *context;
 150         struct sockaddr         addr;
 151         u8                      pad[sizeof(struct sockaddr_in6) -
 152                                     sizeof(struct sockaddr)];
 153 };
 154
 155 struct cma_work {
 156         struct work_struct      work;
 157         struct rdma_id_private  *id;
 158         enum cma_state          old_state;
 159         enum cma_state          new_state;
 160         struct rdma_cm_event    event;
 161 };
 162
 163 union cma_ip_addr {
 164         struct in6_addr ip6;
 165         struct {
 166                 __u32 pad[3];
 167                 __u32 addr;
 168         } ip4;
 169 };
 170
 171 struct cma_hdr {
 172         u8 cma_version;
 173         u8 ip_version;  /* IP version: 7:4 */
 174         __u16 port;
 175         union cma_ip_addr src_addr;
 176         union cma_ip_addr dst_addr;
 177 };
 178
 179 struct sdp_hh {
 180         u8 bsdh[16];
 181         u8 sdp_version; /* Major version: 7:4 */
 182         u8 ip_version;  /* IP version: 7:4 */
 183         u8 sdp_specific1[10];
 184         __u16 port;
 185         __u16 sdp_specific2;
 186         union cma_ip_addr src_addr;
 187         union cma_ip_addr dst_addr;
 188 };
 189
 190 struct sdp_hah {
 191         u8 bsdh[16];
 192         u8 sdp_version;
 193 };
 194
 195 #define CMA_VERSION 0x00
 196 #define SDP_MAJ_VERSION 0x2
 197
 198 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 199 {
 200         unsigned long flags;
 201         int ret;
 202
 203         spin_lock_irqsave(&id_priv->lock, flags);
 204         ret = (id_priv->state == comp);
 205         spin_unlock_irqrestore(&id_priv->lock, flags);
 206         return ret;
 207 }
 208
 209 static int cma_comp_exch(struct rdma_id_private *id_priv,
 210                          enum cma_state comp, enum cma_state exch)
 211 {
 212         unsigned long flags;
 213         int ret;
 214
 215         spin_lock_irqsave(&id_priv->lock, flags);
 216         if ((ret = (id_priv->state == comp)))
 217                 id_priv->state = exch;
 218         spin_unlock_irqrestore(&id_priv->lock, flags);
 219         return ret;
 220 }
 221
 222 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 223                                enum cma_state exch)
 224 {
 225         unsigned long flags;
 226         enum cma_state old;
 227
 228         spin_lock_irqsave(&id_priv->lock, flags);
 229         old = id_priv->state;
 230         id_priv->state = exch;
 231         spin_unlock_irqrestore(&id_priv->lock, flags);
 232         return old;
 233 }
 234
 235 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 236 {
 237         return hdr->ip_version >> 4;
 238 }
 239
 240 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 241 {
 242         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 243 }
 244
 245 static inline u8 sdp_get_majv(u8 sdp_version)
 246 {
 247         return sdp_version >> 4;
 248 }
 249
 250 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 251 {
 252         return hh->ip_version >> 4;
 253 }
 254
 255 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 256 {
 257         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 258 }
 259
 260 static inline int cma_is_ud_ps(enum rdma_port_space ps)
 261 {
 262         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
 263 }
 264
 265 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 266                               struct cma_device *cma_dev)
 267 {
 268         atomic_inc(&cma_dev->refcount);
 269         id_priv->cma_dev = cma_dev;
 270         id_priv->id.device = cma_dev->device;
 271         list_add_tail(&id_priv->list, &cma_dev->id_list);
 272 }
 273
 274 static inline void cma_deref_dev(struct cma_device *cma_dev)
 275 {
 276         if (atomic_dec_and_test(&cma_dev->refcount))
 277                 complete(&cma_dev->comp);
 278 }
 279
 280 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 281 {
 282         list_del(&id_priv->list);
 283         cma_deref_dev(id_priv->cma_dev);
 284         id_priv->cma_dev = NULL;
 285 }
 286
 287 static int cma_set_qkey(struct ib_device *device, u8 port_num,
 288                         enum rdma_port_space ps,
 289                         struct rdma_dev_addr *dev_addr, u32 *qkey)
 290 {
 291         struct ib_sa_mcmember_rec rec;
 292         int ret = 0;
 293
 294         switch (ps) {
 295         case RDMA_PS_UDP:
 296                 *qkey = RDMA_UDP_QKEY;
 297                 break;
 298         case RDMA_PS_IPOIB:
 299                 ib_addr_get_mgid(dev_addr, &rec.mgid);
 300                 ret = ib_sa_get_mcmember_rec(device, port_num, &rec.mgid, &rec);
 301                 *qkey = be32_to_cpu(rec.qkey);
 302                 break;
 303         default:
 304                 break;
 305         }
 306         return ret;
 307 }
 308
 309 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 310 {
 311         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 312         struct cma_device *cma_dev;
 313         union ib_gid gid;
 314         int ret = -ENODEV;
 315
 316         switch (rdma_node_get_transport(dev_addr->dev_type)) {
 317         case RDMA_TRANSPORT_IB:
 318                 ib_addr_get_sgid(dev_addr, &gid);
 319                 break;
 320         case RDMA_TRANSPORT_IWARP:
 321                 iw_addr_get_sgid(dev_addr, &gid);
 322                 break;
 323         default:
 324                 return -ENODEV;
 325         }
 326
 327         list_for_each_entry(cma_dev, &dev_list, list) {
 328                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 329                                          &id_priv->id.port_num, NULL);
 330                 if (!ret) {
 331                         ret = cma_set_qkey(cma_dev->device,
 332                                            id_priv->id.port_num,
 333                                            id_priv->id.ps, dev_addr,
 334                                            &id_priv->qkey);
 335                         if (!ret)
 336                                 cma_attach_to_dev(id_priv, cma_dev);
 337                         break;
 338                 }
 339         }
 340         return ret;
 341 }
 342
 343 static void cma_deref_id(struct rdma_id_private *id_priv)
 344 {
 345         if (atomic_dec_and_test(&id_priv->refcount))
 346                 complete(&id_priv->comp);
 347 }
 348
 349 static int cma_disable_remove(struct rdma_id_private *id_priv,
 350                               enum cma_state state)
 351 {
 352         unsigned long flags;
 353         int ret;
 354
 355         spin_lock_irqsave(&id_priv->lock, flags);
 356         if (id_priv->state == state) {
 357                 atomic_inc(&id_priv->dev_remove);
 358                 ret = 0;
 359         } else
 360                 ret = -EINVAL;
 361         spin_unlock_irqrestore(&id_priv->lock, flags);
 362         return ret;
 363 }
 364
 365 static void cma_enable_remove(struct rdma_id_private *id_priv)
 366 {
 367         if (atomic_dec_and_test(&id_priv->dev_remove))
 368                 wake_up(&id_priv->wait_remove);
 369 }
 370
 371 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
 372 {
 373         return (id_priv->id.device && id_priv->cm_id.ib);
 374 }
 375
 376 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 377                                   void *context, enum rdma_port_space ps)
 378 {
 379         struct rdma_id_private *id_priv;
 380
 381         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 382         if (!id_priv)
 383                 return ERR_PTR(-ENOMEM);
 384
 385         id_priv->state = CMA_IDLE;
 386         id_priv->id.context = context;
 387         id_priv->id.event_handler = event_handler;
 388         id_priv->id.ps = ps;
 389         spin_lock_init(&id_priv->lock);
 390         init_completion(&id_priv->comp);
 391         atomic_set(&id_priv->refcount, 1);
 392         init_waitqueue_head(&id_priv->wait_remove);
 393         atomic_set(&id_priv->dev_remove, 0);
 394         INIT_LIST_HEAD(&id_priv->listen_list);
 395         INIT_LIST_HEAD(&id_priv->mc_list);
 396         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 397
 398         return &id_priv->id;
 399 }
 400 EXPORT_SYMBOL(rdma_create_id);
 401
 402 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 403 {
 404         struct ib_qp_attr qp_attr;
 405         int qp_attr_mask, ret;
 406
 407         qp_attr.qp_state = IB_QPS_INIT;
 408         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 409         if (ret)
 410                 return ret;
 411
 412         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 413         if (ret)
 414                 return ret;
 415
 416         qp_attr.qp_state = IB_QPS_RTR;
 417         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 418         if (ret)
 419                 return ret;
 420
 421         qp_attr.qp_state = IB_QPS_RTS;
 422         qp_attr.sq_psn = 0;
 423         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 424
 425         return ret;
 426 }
 427
 428 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 429 {
 430         struct ib_qp_attr qp_attr;
 431         int qp_attr_mask, ret;
 432
 433         qp_attr.qp_state = IB_QPS_INIT;
 434         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 435         if (ret)
 436                 return ret;
 437
 438         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 439 }
 440
 441 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 442                    struct ib_qp_init_attr *qp_init_attr)
 443 {
 444         struct rdma_id_private *id_priv;
 445         struct ib_qp *qp;
 446         int ret;
 447
 448         id_priv = container_of(id, struct rdma_id_private, id);
 449         if (id->device != pd->device)
 450                 return -EINVAL;
 451
 452         qp = ib_create_qp(pd, qp_init_attr);
 453         if (IS_ERR(qp))
 454                 return PTR_ERR(qp);
 455
 456         if (cma_is_ud_ps(id_priv->id.ps))
 457                 ret = cma_init_ud_qp(id_priv, qp);
 458         else
 459                 ret = cma_init_conn_qp(id_priv, qp);
 460         if (ret)
 461                 goto err;
 462
 463         id->qp = qp;
 464         id_priv->qp_num = qp->qp_num;
 465         id_priv->srq = (qp->srq != NULL);
 466         return 0;
 467 err:
 468         ib_destroy_qp(qp);
 469         return ret;
 470 }
 471 EXPORT_SYMBOL(rdma_create_qp);
 472
 473 void rdma_destroy_qp(struct rdma_cm_id *id)
 474 {
 475         ib_destroy_qp(id->qp);
 476 }
 477 EXPORT_SYMBOL(rdma_destroy_qp);
 478
 479 static int cma_modify_qp_rtr(struct rdma_cm_id *id)
 480 {
 481         struct ib_qp_attr qp_attr;
 482         int qp_attr_mask, ret;
 483
 484         if (!id->qp)
 485                 return 0;
 486
 487         /* Need to update QP attributes from default values. */
 488         qp_attr.qp_state = IB_QPS_INIT;
 489         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 490         if (ret)
 491                 return ret;
 492
 493         ret = ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 494         if (ret)
 495                 return ret;
 496
 497         qp_attr.qp_state = IB_QPS_RTR;
 498         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 499         if (ret)
 500                 return ret;
 501
 502         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 503 }
 504
 505 static int cma_modify_qp_rts(struct rdma_cm_id *id)
 506 {
 507         struct ib_qp_attr qp_attr;
 508         int qp_attr_mask, ret;
 509
 510         if (!id->qp)
 511                 return 0;
 512
 513         qp_attr.qp_state = IB_QPS_RTS;
 514         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 515         if (ret)
 516                 return ret;
 517
 518         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 519 }
 520
 521 static int cma_modify_qp_err(struct rdma_cm_id *id)
 522 {
 523         struct ib_qp_attr qp_attr;
 524
 525         if (!id->qp)
 526                 return 0;
 527
 528         qp_attr.qp_state = IB_QPS_ERR;
 529         return ib_modify_qp(id->qp, &qp_attr, IB_QP_STATE);
 530 }
 531
 532 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 533                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 534 {
 535         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 536         int ret;
 537
 538         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 539                                   ib_addr_get_pkey(dev_addr),
 540                                   &qp_attr->pkey_index);
 541         if (ret)
 542                 return ret;
 543
 544         qp_attr->port_num = id_priv->id.port_num;
 545         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 546
 547         if (cma_is_ud_ps(id_priv->id.ps)) {
 548                 qp_attr->qkey = id_priv->qkey;
 549                 *qp_attr_mask |= IB_QP_QKEY;
 550         } else {
 551                 qp_attr->qp_access_flags = 0;
 552                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 553         }
 554         return 0;
 555 }
 556
 557 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 558                        int *qp_attr_mask)
 559 {
 560         struct rdma_id_private *id_priv;
 561         int ret = 0;
 562
 563         id_priv = container_of(id, struct rdma_id_private, id);
 564         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 565         case RDMA_TRANSPORT_IB:
 566                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
 567                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 568                 else
 569                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 570                                                  qp_attr_mask);
 571                 if (qp_attr->qp_state == IB_QPS_RTR)
 572                         qp_attr->rq_psn = id_priv->seq_num;
 573                 break;
 574         case RDMA_TRANSPORT_IWARP:
 575                 if (!id_priv->cm_id.iw) {
 576                         qp_attr->qp_access_flags = 0;
 577                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 578                 } else
 579                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 580                                                  qp_attr_mask);
 581                 break;
 582         default:
 583                 ret = -ENOSYS;
 584                 break;
 585         }
 586
 587         return ret;
 588 }
 589 EXPORT_SYMBOL(rdma_init_qp_attr);
 590
 591 static inline int cma_zero_addr(struct sockaddr *addr)
 592 {
 593         struct in6_addr *ip6;
 594
 595         if (addr->sa_family == AF_INET)
 596                 return ZERONET(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 597         else {
 598                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 599                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 600                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 601         }
 602 }
 603
 604 static inline int cma_loopback_addr(struct sockaddr *addr)
 605 {
 606         return LOOPBACK(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 607 }
 608
 609 static inline int cma_any_addr(struct sockaddr *addr)
 610 {
 611         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 612 }
 613
 614 static inline __be16 cma_port(struct sockaddr *addr)
 615 {
 616         if (addr->sa_family == AF_INET)
 617                 return ((struct sockaddr_in *) addr)->sin_port;
 618         else
 619                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 620 }
 621
 622 static inline int cma_any_port(struct sockaddr *addr)
 623 {
 624         return !cma_port(addr);
 625 }
 626
 627 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 628                             u8 *ip_ver, __u16 *port,
 629                             union cma_ip_addr **src, union cma_ip_addr **dst)
 630 {
 631         switch (ps) {
 632         case RDMA_PS_SDP:
 633                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 634                     SDP_MAJ_VERSION)
 635                         return -EINVAL;
 636
 637                 *ip_ver = sdp_get_ip_ver(hdr);
 638                 *port   = ((struct sdp_hh *) hdr)->port;
 639                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 640                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 641                 break;
 642         default:
 643                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 644                         return -EINVAL;
 645
 646                 *ip_ver = cma_get_ip_ver(hdr);
 647                 *port   = ((struct cma_hdr *) hdr)->port;
 648                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 649                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 650                 break;
 651         }
 652
 653         if (*ip_ver != 4 && *ip_ver != 6)
 654                 return -EINVAL;
 655         return 0;
 656 }
 657
 658 static void cma_save_net_info(struct rdma_addr *addr,
 659                               struct rdma_addr *listen_addr,
 660                               u8 ip_ver, __u16 port,
 661                               union cma_ip_addr *src, union cma_ip_addr *dst)
 662 {
 663         struct sockaddr_in *listen4, *ip4;
 664         struct sockaddr_in6 *listen6, *ip6;
 665
 666         switch (ip_ver) {
 667         case 4:
 668                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 669                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 670                 ip4->sin_family = listen4->sin_family;
 671                 ip4->sin_addr.s_addr = dst->ip4.addr;
 672                 ip4->sin_port = listen4->sin_port;
 673
 674                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 675                 ip4->sin_family = listen4->sin_family;
 676                 ip4->sin_addr.s_addr = src->ip4.addr;
 677                 ip4->sin_port = port;
 678                 break;
 679         case 6:
 680                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 681                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 682                 ip6->sin6_family = listen6->sin6_family;
 683                 ip6->sin6_addr = dst->ip6;
 684                 ip6->sin6_port = listen6->sin6_port;
 685
 686                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 687                 ip6->sin6_family = listen6->sin6_family;
 688                 ip6->sin6_addr = src->ip6;
 689                 ip6->sin6_port = port;
 690                 break;
 691         default:
 692                 break;
 693         }
 694 }
 695
 696 static inline int cma_user_data_offset(enum rdma_port_space ps)
 697 {
 698         switch (ps) {
 699         case RDMA_PS_SDP:
 700                 return 0;
 701         default:
 702                 return sizeof(struct cma_hdr);
 703         }
 704 }
 705
 706 static void cma_cancel_route(struct rdma_id_private *id_priv)
 707 {
 708         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 709         case RDMA_TRANSPORT_IB:
 710                 if (id_priv->query)
 711                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 712                 break;
 713         default:
 714                 break;
 715         }
 716 }
 717
 718 static inline int cma_internal_listen(struct rdma_id_private *id_priv)
 719 {
 720         return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
 721                cma_any_addr(&id_priv->id.route.addr.src_addr);
 722 }
 723
 724 static void cma_destroy_listen(struct rdma_id_private *id_priv)
 725 {
 726         cma_exch(id_priv, CMA_DESTROYING);
 727
 728         if (id_priv->cma_dev) {
 729                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 730                 case RDMA_TRANSPORT_IB:
 731                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 732                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 733                         break;
 734                 case RDMA_TRANSPORT_IWARP:
 735                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 736                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 737                         break;
 738                 default:
 739                         break;
 740                 }
 741                 cma_detach_from_dev(id_priv);
 742         }
 743         list_del(&id_priv->listen_list);
 744
 745         cma_deref_id(id_priv);
 746         wait_for_completion(&id_priv->comp);
 747
 748         kfree(id_priv);
 749 }
 750
 751 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 752 {
 753         struct rdma_id_private *dev_id_priv;
 754
 755         mutex_lock(&lock);
 756         list_del(&id_priv->list);
 757
 758         while (!list_empty(&id_priv->listen_list)) {
 759                 dev_id_priv = list_entry(id_priv->listen_list.next,
 760                                          struct rdma_id_private, listen_list);
 761                 cma_destroy_listen(dev_id_priv);
 762         }
 763         mutex_unlock(&lock);
 764 }
 765
 766 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 767                                  enum cma_state state)
 768 {
 769         switch (state) {
 770         case CMA_ADDR_QUERY:
 771                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 772                 break;
 773         case CMA_ROUTE_QUERY:
 774                 cma_cancel_route(id_priv);
 775                 break;
 776         case CMA_LISTEN:
 777                 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
 778                     !id_priv->cma_dev)
 779                         cma_cancel_listens(id_priv);
 780                 break;
 781         default:
 782                 break;
 783         }
 784 }
 785
 786 static void cma_release_port(struct rdma_id_private *id_priv)
 787 {
 788         struct rdma_bind_list *bind_list = id_priv->bind_list;
 789
 790         if (!bind_list)
 791                 return;
 792
 793         mutex_lock(&lock);
 794         hlist_del(&id_priv->node);
 795         if (hlist_empty(&bind_list->owners)) {
 796                 idr_remove(bind_list->ps, bind_list->port);
 797                 kfree(bind_list);
 798         }
 799         mutex_unlock(&lock);
 800 }
 801
 802 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 803 {
 804         struct cma_multicast *mc;
 805
 806         while (!list_empty(&id_priv->mc_list)) {
 807                 mc = container_of(id_priv->mc_list.next,
 808                                   struct cma_multicast, list);
 809                 list_del(&mc->list);
 810                 ib_sa_free_multicast(mc->multicast.ib);
 811                 kfree(mc);
 812         }
 813 }
 814
 815 void rdma_destroy_id(struct rdma_cm_id *id)
 816 {
 817         struct rdma_id_private *id_priv;
 818         enum cma_state state;
 819
 820         id_priv = container_of(id, struct rdma_id_private, id);
 821         state = cma_exch(id_priv, CMA_DESTROYING);
 822         cma_cancel_operation(id_priv, state);
 823
 824         mutex_lock(&lock);
 825         if (id_priv->cma_dev) {
 826                 mutex_unlock(&lock);
 827                 switch (rdma_node_get_transport(id->device->node_type)) {
 828                 case RDMA_TRANSPORT_IB:
 829                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 830                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 831                         break;
 832                 case RDMA_TRANSPORT_IWARP:
 833                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 834                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 835                         break;
 836                 default:
 837                         break;
 838                 }
 839                 cma_leave_mc_groups(id_priv);
 840                 mutex_lock(&lock);
 841                 cma_detach_from_dev(id_priv);
 842         }
 843         mutex_unlock(&lock);
 844
 845         cma_release_port(id_priv);
 846         cma_deref_id(id_priv);
 847         wait_for_completion(&id_priv->comp);
 848
 849         kfree(id_priv->id.route.path_rec);
 850         kfree(id_priv);
 851 }
 852 EXPORT_SYMBOL(rdma_destroy_id);
 853
 854 static int cma_rep_recv(struct rdma_id_private *id_priv)
 855 {
 856         int ret;
 857
 858         ret = cma_modify_qp_rtr(&id_priv->id);
 859         if (ret)
 860                 goto reject;
 861
 862         ret = cma_modify_qp_rts(&id_priv->id);
 863         if (ret)
 864                 goto reject;
 865
 866         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 867         if (ret)
 868                 goto reject;
 869
 870         return 0;
 871 reject:
 872         cma_modify_qp_err(&id_priv->id);
 873         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 874                        NULL, 0, NULL, 0);
 875         return ret;
 876 }
 877
 878 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 879 {
 880         if (id_priv->id.ps == RDMA_PS_SDP &&
 881             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 882             SDP_MAJ_VERSION)
 883                 return -EINVAL;
 884
 885         return 0;
 886 }
 887
 888 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 889                                    struct ib_cm_rep_event_param *rep_data,
 890                                    void *private_data)
 891 {
 892         event->param.conn.private_data = private_data;
 893         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 894         event->param.conn.responder_resources = rep_data->responder_resources;
 895         event->param.conn.initiator_depth = rep_data->initiator_depth;
 896         event->param.conn.flow_control = rep_data->flow_control;
 897         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 898         event->param.conn.srq = rep_data->srq;
 899         event->param.conn.qp_num = rep_data->remote_qpn;
 900 }
 901
 902 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 903 {
 904         struct rdma_id_private *id_priv = cm_id->context;
 905         struct rdma_cm_event event;
 906         int ret = 0;
 907
 908         if (cma_disable_remove(id_priv, CMA_CONNECT))
 909                 return 0;
 910
 911         memset(&event, 0, sizeof event);
 912         switch (ib_event->event) {
 913         case IB_CM_REQ_ERROR:
 914         case IB_CM_REP_ERROR:
 915                 event.event = RDMA_CM_EVENT_UNREACHABLE;
 916                 event.status = -ETIMEDOUT;
 917                 break;
 918         case IB_CM_REP_RECEIVED:
 919                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
 920                 if (event.status)
 921                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
 922                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 923                         event.status = cma_rep_recv(id_priv);
 924                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
 925                                                      RDMA_CM_EVENT_ESTABLISHED;
 926                 } else
 927                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 928                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
 929                                        ib_event->private_data);
 930                 break;
 931         case IB_CM_RTU_RECEIVED:
 932         case IB_CM_USER_ESTABLISHED:
 933                 event.event = RDMA_CM_EVENT_ESTABLISHED;
 934                 break;
 935         case IB_CM_DREQ_ERROR:
 936                 event.status = -ETIMEDOUT; /* fall through */
 937         case IB_CM_DREQ_RECEIVED:
 938         case IB_CM_DREP_RECEIVED:
 939                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 940                         goto out;
 941                 event.event = RDMA_CM_EVENT_DISCONNECTED;
 942                 break;
 943         case IB_CM_TIMEWAIT_EXIT:
 944         case IB_CM_MRA_RECEIVED:
 945                 /* ignore event */
 946                 goto out;
 947         case IB_CM_REJ_RECEIVED:
 948                 cma_modify_qp_err(&id_priv->id);
 949                 event.status = ib_event->param.rej_rcvd.reason;
 950                 event.event = RDMA_CM_EVENT_REJECTED;
 951                 event.param.conn.private_data = ib_event->private_data;
 952                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 953                 break;
 954         default:
 955                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
 956                        ib_event->event);
 957                 goto out;
 958         }
 959
 960         ret = id_priv->id.event_handler(&id_priv->id, &event);
 961         if (ret) {
 962                 /* Destroy the CM ID by returning a non-zero value. */
 963                 id_priv->cm_id.ib = NULL;
 964                 cma_exch(id_priv, CMA_DESTROYING);
 965                 cma_enable_remove(id_priv);
 966                 rdma_destroy_id(&id_priv->id);
 967                 return ret;
 968         }
 969 out:
 970         cma_enable_remove(id_priv);
 971         return ret;
 972 }
 973
 974 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
 975                                                struct ib_cm_event *ib_event)
 976 {
 977         struct rdma_id_private *id_priv;
 978         struct rdma_cm_id *id;
 979         struct rdma_route *rt;
 980         union cma_ip_addr *src, *dst;
 981         __u16 port;
 982         u8 ip_ver;
 983
 984         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 985                              &ip_ver, &port, &src, &dst))
 986                 goto err;
 987
 988         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 989                             listen_id->ps);
 990         if (IS_ERR(id))
 991                 goto err;
 992
 993         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 994                           ip_ver, port, src, dst);
 995
 996         rt = &id->route;
 997         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
 998         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
 999                                GFP_KERNEL);
1000         if (!rt->path_rec)
1001                 goto destroy_id;
1002
1003         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1004         if (rt->num_paths == 2)
1005                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1006
1007         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1008         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1009         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1010         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
1011
1012         id_priv = container_of(id, struct rdma_id_private, id);
1013         id_priv->state = CMA_CONNECT;
1014         return id_priv;
1015
1016 destroy_id:
1017         rdma_destroy_id(id);
1018 err:
1019         return NULL;
1020 }
1021
1022 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1023                                               struct ib_cm_event *ib_event)
1024 {
1025         struct rdma_id_private *id_priv;
1026         struct rdma_cm_id *id;
1027         union cma_ip_addr *src, *dst;
1028         __u16 port;
1029         u8 ip_ver;
1030         int ret;
1031
1032         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1033                             listen_id->ps);
1034         if (IS_ERR(id))
1035                 return NULL;
1036
1037
1038         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1039                              &ip_ver, &port, &src, &dst))
1040                 goto err;
1041
1042         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1043                           ip_ver, port, src, dst);
1044
1045         ret = rdma_translate_ip(&id->route.addr.src_addr,
1046                                 &id->route.addr.dev_addr);
1047         if (ret)
1048                 goto err;
1049
1050         id_priv = container_of(id, struct rdma_id_private, id);
1051         id_priv->state = CMA_CONNECT;
1052         return id_priv;
1053 err:
1054         rdma_destroy_id(id);
1055         return NULL;
1056 }
1057
1058 static void cma_set_req_event_data(struct rdma_cm_event *event,
1059                                    struct ib_cm_req_event_param *req_data,
1060                                    void *private_data, int offset)
1061 {
1062         event->param.conn.private_data = private_data + offset;
1063         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1064         event->param.conn.responder_resources = req_data->responder_resources;
1065         event->param.conn.initiator_depth = req_data->initiator_depth;
1066         event->param.conn.flow_control = req_data->flow_control;
1067         event->param.conn.retry_count = req_data->retry_count;
1068         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1069         event->param.conn.srq = req_data->srq;
1070         event->param.conn.qp_num = req_data->remote_qpn;
1071 }
1072
1073 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1074 {
1075         struct rdma_id_private *listen_id, *conn_id;
1076         struct rdma_cm_event event;
1077         int offset, ret;
1078
1079         listen_id = cm_id->context;
1080         if (cma_disable_remove(listen_id, CMA_LISTEN))
1081                 return -ECONNABORTED;
1082
1083         memset(&event, 0, sizeof event);
1084         offset = cma_user_data_offset(listen_id->id.ps);
1085         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1086         if (cma_is_ud_ps(listen_id->id.ps)) {
1087                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1088                 event.param.ud.private_data = ib_event->private_data + offset;
1089                 event.param.ud.private_data_len =
1090                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1091         } else {
1092                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1093                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1094                                        ib_event->private_data, offset);
1095         }
1096         if (!conn_id) {
1097                 ret = -ENOMEM;
1098                 goto out;
1099         }
1100
1101         atomic_inc(&conn_id->dev_remove);
1102         mutex_lock(&lock);
1103         ret = cma_acquire_dev(conn_id);
1104         mutex_unlock(&lock);
1105         if (ret)
1106                 goto release_conn_id;
1107
1108         conn_id->cm_id.ib = cm_id;
1109         cm_id->context = conn_id;
1110         cm_id->cm_handler = cma_ib_handler;
1111
1112         ret = conn_id->id.event_handler(&conn_id->id, &event);
1113         if (!ret)
1114                 goto out;
1115
1116         /* Destroy the CM ID by returning a non-zero value. */
1117         conn_id->cm_id.ib = NULL;
1118
1119 release_conn_id:
1120         cma_exch(conn_id, CMA_DESTROYING);
1121         cma_enable_remove(conn_id);
1122         rdma_destroy_id(&conn_id->id);
1123
1124 out:
1125         cma_enable_remove(listen_id);
1126         return ret;
1127 }
1128
1129 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1130 {
1131         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1132 }
1133
1134 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1135                                  struct ib_cm_compare_data *compare)
1136 {
1137         struct cma_hdr *cma_data, *cma_mask;
1138         struct sdp_hh *sdp_data, *sdp_mask;
1139         __u32 ip4_addr;
1140         struct in6_addr ip6_addr;
1141
1142         memset(compare, 0, sizeof *compare);
1143         cma_data = (void *) compare->data;
1144         cma_mask = (void *) compare->mask;
1145         sdp_data = (void *) compare->data;
1146         sdp_mask = (void *) compare->mask;
1147
1148         switch (addr->sa_family) {
1149         case AF_INET:
1150                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1151                 if (ps == RDMA_PS_SDP) {
1152                         sdp_set_ip_ver(sdp_data, 4);
1153                         sdp_set_ip_ver(sdp_mask, 0xF);
1154                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1155                         sdp_mask->dst_addr.ip4.addr = ~0;
1156                 } else {
1157                         cma_set_ip_ver(cma_data, 4);
1158                         cma_set_ip_ver(cma_mask, 0xF);
1159                         cma_data->dst_addr.ip4.addr = ip4_addr;
1160                         cma_mask->dst_addr.ip4.addr = ~0;
1161                 }
1162                 break;
1163         case AF_INET6:
1164                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1165                 if (ps == RDMA_PS_SDP) {
1166                         sdp_set_ip_ver(sdp_data, 6);
1167                         sdp_set_ip_ver(sdp_mask, 0xF);
1168                         sdp_data->dst_addr.ip6 = ip6_addr;
1169                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1170                                sizeof sdp_mask->dst_addr.ip6);
1171                 } else {
1172                         cma_set_ip_ver(cma_data, 6);
1173                         cma_set_ip_ver(cma_mask, 0xF);
1174                         cma_data->dst_addr.ip6 = ip6_addr;
1175                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1176                                sizeof cma_mask->dst_addr.ip6);
1177                 }
1178                 break;
1179         default:
1180                 break;
1181         }
1182 }
1183
1184 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1185 {
1186         struct rdma_id_private *id_priv = iw_id->context;
1187         struct rdma_cm_event event;
1188         struct sockaddr_in *sin;
1189         int ret = 0;
1190
1191         if (cma_disable_remove(id_priv, CMA_CONNECT))
1192                 return 0;
1193
1194         memset(&event, 0, sizeof event);
1195         switch (iw_event->event) {
1196         case IW_CM_EVENT_CLOSE:
1197                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1198                 break;
1199         case IW_CM_EVENT_CONNECT_REPLY:
1200                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1201                 *sin = iw_event->local_addr;
1202                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1203                 *sin = iw_event->remote_addr;
1204                 switch (iw_event->status) {
1205                 case 0:
1206                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1207                         break;
1208                 case -ECONNRESET:
1209                 case -ECONNREFUSED:
1210                         event.event = RDMA_CM_EVENT_REJECTED;
1211                         break;
1212                 case -ETIMEDOUT:
1213                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1214                         break;
1215                 default:
1216                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1217                         break;
1218                 }
1219                 break;
1220         case IW_CM_EVENT_ESTABLISHED:
1221                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1222                 break;
1223         default:
1224                 BUG_ON(1);
1225         }
1226
1227         event.status = iw_event->status;
1228         event.param.conn.private_data = iw_event->private_data;
1229         event.param.conn.private_data_len = iw_event->private_data_len;
1230         ret = id_priv->id.event_handler(&id_priv->id, &event);
1231         if (ret) {
1232                 /* Destroy the CM ID by returning a non-zero value. */
1233                 id_priv->cm_id.iw = NULL;
1234                 cma_exch(id_priv, CMA_DESTROYING);
1235                 cma_enable_remove(id_priv);
1236                 rdma_destroy_id(&id_priv->id);
1237                 return ret;
1238         }
1239
1240         cma_enable_remove(id_priv);
1241         return ret;
1242 }
1243
1244 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1245                                struct iw_cm_event *iw_event)
1246 {
1247         struct rdma_cm_id *new_cm_id;
1248         struct rdma_id_private *listen_id, *conn_id;
1249         struct sockaddr_in *sin;
1250         struct net_device *dev = NULL;
1251         struct rdma_cm_event event;
1252         int ret;
1253
1254         listen_id = cm_id->context;
1255         if (cma_disable_remove(listen_id, CMA_LISTEN))
1256                 return -ECONNABORTED;
1257
1258         /* Create a new RDMA id for the new IW CM ID */
1259         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1260                                    listen_id->id.context,
1261                                    RDMA_PS_TCP);
1262         if (!new_cm_id) {
1263                 ret = -ENOMEM;
1264                 goto out;
1265         }
1266         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1267         atomic_inc(&conn_id->dev_remove);
1268         conn_id->state = CMA_CONNECT;
1269
1270         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1271         if (!dev) {
1272                 ret = -EADDRNOTAVAIL;
1273                 cma_enable_remove(conn_id);
1274                 rdma_destroy_id(new_cm_id);
1275                 goto out;
1276         }
1277         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1278         if (ret) {
1279                 cma_enable_remove(conn_id);
1280                 rdma_destroy_id(new_cm_id);
1281                 goto out;
1282         }
1283
1284         mutex_lock(&lock);
1285         ret = cma_acquire_dev(conn_id);
1286         mutex_unlock(&lock);
1287         if (ret) {
1288                 cma_enable_remove(conn_id);
1289                 rdma_destroy_id(new_cm_id);
1290                 goto out;
1291         }
1292
1293         conn_id->cm_id.iw = cm_id;
1294         cm_id->context = conn_id;
1295         cm_id->cm_handler = cma_iw_handler;
1296
1297         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1298         *sin = iw_event->local_addr;
1299         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1300         *sin = iw_event->remote_addr;
1301
1302         memset(&event, 0, sizeof event);
1303         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1304         event.param.conn.private_data = iw_event->private_data;
1305         event.param.conn.private_data_len = iw_event->private_data_len;
1306         ret = conn_id->id.event_handler(&conn_id->id, &event);
1307         if (ret) {
1308                 /* User wants to destroy the CM ID */
1309                 conn_id->cm_id.iw = NULL;
1310                 cma_exch(conn_id, CMA_DESTROYING);
1311                 cma_enable_remove(conn_id);
1312                 rdma_destroy_id(&conn_id->id);
1313         }
1314
1315 out:
1316         if (dev)
1317                 dev_put(dev);
1318         cma_enable_remove(listen_id);
1319         return ret;
1320 }
1321
1322 static int cma_ib_listen(struct rdma_id_private *id_priv)
1323 {
1324         struct ib_cm_compare_data compare_data;
1325         struct sockaddr *addr;
1326         __be64 svc_id;
1327         int ret;
1328
1329         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1330                                             id_priv);
1331         if (IS_ERR(id_priv->cm_id.ib))
1332                 return PTR_ERR(id_priv->cm_id.ib);
1333
1334         addr = &id_priv->id.route.addr.src_addr;
1335         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1336         if (cma_any_addr(addr))
1337                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1338         else {
1339                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1340                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1341         }
1342
1343         if (ret) {
1344                 ib_destroy_cm_id(id_priv->cm_id.ib);
1345                 id_priv->cm_id.ib = NULL;
1346         }
1347
1348         return ret;
1349 }
1350
1351 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1352 {
1353         int ret;
1354         struct sockaddr_in *sin;
1355
1356         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1357                                             iw_conn_req_handler,
1358                                             id_priv);
1359         if (IS_ERR(id_priv->cm_id.iw))
1360                 return PTR_ERR(id_priv->cm_id.iw);
1361
1362         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1363         id_priv->cm_id.iw->local_addr = *sin;
1364
1365         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1366
1367         if (ret) {
1368                 iw_destroy_cm_id(id_priv->cm_id.iw);
1369                 id_priv->cm_id.iw = NULL;
1370         }
1371
1372         return ret;
1373 }
1374
1375 static int cma_listen_handler(struct rdma_cm_id *id,
1376                               struct rdma_cm_event *event)
1377 {
1378         struct rdma_id_private *id_priv = id->context;
1379
1380         id->context = id_priv->id.context;
1381         id->event_handler = id_priv->id.event_handler;
1382         return id_priv->id.event_handler(id, event);
1383 }
1384
1385 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1386                               struct cma_device *cma_dev)
1387 {
1388         struct rdma_id_private *dev_id_priv;
1389         struct rdma_cm_id *id;
1390         int ret;
1391
1392         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1393         if (IS_ERR(id))
1394                 return;
1395
1396         dev_id_priv = container_of(id, struct rdma_id_private, id);
1397
1398         dev_id_priv->state = CMA_ADDR_BOUND;
1399         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1400                ip_addr_size(&id_priv->id.route.addr.src_addr));
1401
1402         cma_attach_to_dev(dev_id_priv, cma_dev);
1403         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1404
1405         ret = rdma_listen(id, id_priv->backlog);
1406         if (ret)
1407                 goto err;
1408
1409         return;
1410 err:
1411         cma_destroy_listen(dev_id_priv);
1412 }
1413
1414 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1415 {
1416         struct cma_device *cma_dev;
1417
1418         mutex_lock(&lock);
1419         list_add_tail(&id_priv->list, &listen_any_list);
1420         list_for_each_entry(cma_dev, &dev_list, list)
1421                 cma_listen_on_dev(id_priv, cma_dev);
1422         mutex_unlock(&lock);
1423 }
1424
1425 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1426 {
1427         struct sockaddr_in addr_in;
1428
1429         memset(&addr_in, 0, sizeof addr_in);
1430         addr_in.sin_family = af;
1431         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1432 }
1433
1434 int rdma_listen(struct rdma_cm_id *id, int backlog)
1435 {
1436         struct rdma_id_private *id_priv;
1437         int ret;
1438
1439         id_priv = container_of(id, struct rdma_id_private, id);
1440         if (id_priv->state == CMA_IDLE) {
1441                 ret = cma_bind_any(id, AF_INET);
1442                 if (ret)
1443                         return ret;
1444         }
1445
1446         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1447                 return -EINVAL;
1448
1449         id_priv->backlog = backlog;
1450         if (id->device) {
1451                 switch (rdma_node_get_transport(id->device->node_type)) {
1452                 case RDMA_TRANSPORT_IB:
1453                         ret = cma_ib_listen(id_priv);
1454                         if (ret)
1455                                 goto err;
1456                         break;
1457                 case RDMA_TRANSPORT_IWARP:
1458                         ret = cma_iw_listen(id_priv, backlog);
1459                         if (ret)
1460                                 goto err;
1461                         break;
1462                 default:
1463                         ret = -ENOSYS;
1464                         goto err;
1465                 }
1466         } else
1467                 cma_listen_on_all(id_priv);
1468
1469         return 0;
1470 err:
1471         id_priv->backlog = 0;
1472         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1473         return ret;
1474 }
1475 EXPORT_SYMBOL(rdma_listen);
1476
1477 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1478                               void *context)
1479 {
1480         struct cma_work *work = context;
1481         struct rdma_route *route;
1482
1483         route = &work->id->id.route;
1484
1485         if (!status) {
1486                 route->num_paths = 1;
1487                 *route->path_rec = *path_rec;
1488         } else {
1489                 work->old_state = CMA_ROUTE_QUERY;
1490                 work->new_state = CMA_ADDR_RESOLVED;
1491                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1492                 work->event.status = status;
1493         }
1494
1495         queue_work(cma_wq, &work->work);
1496 }
1497
1498 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1499                               struct cma_work *work)
1500 {
1501         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1502         struct ib_sa_path_rec path_rec;
1503
1504         memset(&path_rec, 0, sizeof path_rec);
1505         ib_addr_get_sgid(addr, &path_rec.sgid);
1506         ib_addr_get_dgid(addr, &path_rec.dgid);
1507         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1508         path_rec.numb_path = 1;
1509         path_rec.reversible = 1;
1510
1511         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1512                                 id_priv->id.port_num, &path_rec,
1513                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1514                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1515                                 IB_SA_PATH_REC_REVERSIBLE,
1516                                 timeout_ms, GFP_KERNEL,
1517                                 cma_query_handler, work, &id_priv->query);
1518
1519         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1520 }
1521
1522 static void cma_work_handler(struct work_struct *_work)
1523 {
1524         struct cma_work *work = container_of(_work, struct cma_work, work);
1525         struct rdma_id_private *id_priv = work->id;
1526         int destroy = 0;
1527
1528         atomic_inc(&id_priv->dev_remove);
1529         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1530                 goto out;
1531
1532         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1533                 cma_exch(id_priv, CMA_DESTROYING);
1534                 destroy = 1;
1535         }
1536 out:
1537         cma_enable_remove(id_priv);
1538         cma_deref_id(id_priv);
1539         if (destroy)
1540                 rdma_destroy_id(&id_priv->id);
1541         kfree(work);
1542 }
1543
1544 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1545 {
1546         struct rdma_route *route = &id_priv->id.route;
1547         struct cma_work *work;
1548         int ret;
1549
1550         work = kzalloc(sizeof *work, GFP_KERNEL);
1551         if (!work)
1552                 return -ENOMEM;
1553
1554         work->id = id_priv;
1555         INIT_WORK(&work->work, cma_work_handler);
1556         work->old_state = CMA_ROUTE_QUERY;
1557         work->new_state = CMA_ROUTE_RESOLVED;
1558         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1559
1560         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1561         if (!route->path_rec) {
1562                 ret = -ENOMEM;
1563                 goto err1;
1564         }
1565
1566         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1567         if (ret)
1568                 goto err2;
1569
1570         return 0;
1571 err2:
1572         kfree(route->path_rec);
1573         route->path_rec = NULL;
1574 err1:
1575         kfree(work);
1576         return ret;
1577 }
1578
1579 int rdma_set_ib_paths(struct rdma_cm_id *id,
1580                       struct ib_sa_path_rec *path_rec, int num_paths)
1581 {
1582         struct rdma_id_private *id_priv;
1583         int ret;
1584
1585         id_priv = container_of(id, struct rdma_id_private, id);
1586         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1587                 return -EINVAL;
1588
1589         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1590         if (!id->route.path_rec) {
1591                 ret = -ENOMEM;
1592                 goto err;
1593         }
1594
1595         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1596         return 0;
1597 err:
1598         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1599         return ret;
1600 }
1601 EXPORT_SYMBOL(rdma_set_ib_paths);
1602
1603 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1604 {
1605         struct cma_work *work;
1606
1607         work = kzalloc(sizeof *work, GFP_KERNEL);
1608         if (!work)
1609                 return -ENOMEM;
1610
1611         work->id = id_priv;
1612         INIT_WORK(&work->work, cma_work_handler);
1613         work->old_state = CMA_ROUTE_QUERY;
1614         work->new_state = CMA_ROUTE_RESOLVED;
1615         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1616         queue_work(cma_wq, &work->work);
1617         return 0;
1618 }
1619
1620 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1621 {
1622         struct rdma_id_private *id_priv;
1623         int ret;
1624
1625         id_priv = container_of(id, struct rdma_id_private, id);
1626         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1627                 return -EINVAL;
1628
1629         atomic_inc(&id_priv->refcount);
1630         switch (rdma_node_get_transport(id->device->node_type)) {
1631         case RDMA_TRANSPORT_IB:
1632                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1633                 break;
1634         case RDMA_TRANSPORT_IWARP:
1635                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1636                 break;
1637         default:
1638                 ret = -ENOSYS;
1639                 break;
1640         }
1641         if (ret)
1642                 goto err;
1643
1644         return 0;
1645 err:
1646         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1647         cma_deref_id(id_priv);
1648         return ret;
1649 }
1650 EXPORT_SYMBOL(rdma_resolve_route);
1651
1652 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1653 {
1654         struct cma_device *cma_dev;
1655         struct ib_port_attr port_attr;
1656         union ib_gid gid;
1657         u16 pkey;
1658         int ret;
1659         u8 p;
1660
1661         mutex_lock(&lock);
1662         if (list_empty(&dev_list)) {
1663                 ret = -ENODEV;
1664                 goto out;
1665         }
1666         list_for_each_entry(cma_dev, &dev_list, list)
1667                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1668                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1669                             port_attr.state == IB_PORT_ACTIVE)
1670                                 goto port_found;
1671
1672         p = 1;
1673         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1674
1675 port_found:
1676         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1677         if (ret)
1678                 goto out;
1679
1680         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1681         if (ret)
1682                 goto out;
1683
1684         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1685         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1686         id_priv->id.port_num = p;
1687         cma_attach_to_dev(id_priv, cma_dev);
1688 out:
1689         mutex_unlock(&lock);
1690         return ret;
1691 }
1692
1693 static void addr_handler(int status, struct sockaddr *src_addr,
1694                          struct rdma_dev_addr *dev_addr, void *context)
1695 {
1696         struct rdma_id_private *id_priv = context;
1697         struct rdma_cm_event event;
1698
1699         memset(&event, 0, sizeof event);
1700         atomic_inc(&id_priv->dev_remove);
1701
1702         /*
1703          * Grab mutex to block rdma_destroy_id() from removing the device while
1704          * we're trying to acquire it.
1705          */
1706         mutex_lock(&lock);
1707         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1708                 mutex_unlock(&lock);
1709                 goto out;
1710         }
1711
1712         if (!status && !id_priv->cma_dev)
1713                 status = cma_acquire_dev(id_priv);
1714         mutex_unlock(&lock);
1715
1716         if (status) {
1717                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1718                         goto out;
1719                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1720                 event.status = status;
1721         } else {
1722                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1723                        ip_addr_size(src_addr));
1724                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1725         }
1726
1727         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1728                 cma_exch(id_priv, CMA_DESTROYING);
1729                 cma_enable_remove(id_priv);
1730                 cma_deref_id(id_priv);
1731                 rdma_destroy_id(&id_priv->id);
1732                 return;
1733         }
1734 out:
1735         cma_enable_remove(id_priv);
1736         cma_deref_id(id_priv);
1737 }
1738
1739 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1740 {
1741         struct cma_work *work;
1742         struct sockaddr_in *src_in, *dst_in;
1743         union ib_gid gid;
1744         int ret;
1745
1746         work = kzalloc(sizeof *work, GFP_KERNEL);
1747         if (!work)
1748                 return -ENOMEM;
1749
1750         if (!id_priv->cma_dev) {
1751                 ret = cma_bind_loopback(id_priv);
1752                 if (ret)
1753                         goto err;
1754         }
1755
1756         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1757         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1758
1759         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1760                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1761                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1762                 src_in->sin_family = dst_in->sin_family;
1763                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1764         }
1765
1766         work->id = id_priv;
1767         INIT_WORK(&work->work, cma_work_handler);
1768         work->old_state = CMA_ADDR_QUERY;
1769         work->new_state = CMA_ADDR_RESOLVED;
1770         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1771         queue_work(cma_wq, &work->work);
1772         return 0;
1773 err:
1774         kfree(work);
1775         return ret;
1776 }
1777
1778 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1779                          struct sockaddr *dst_addr)
1780 {
1781         if (src_addr && src_addr->sa_family)
1782                 return rdma_bind_addr(id, src_addr);
1783         else
1784                 return cma_bind_any(id, dst_addr->sa_family);
1785 }
1786
1787 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1788                       struct sockaddr *dst_addr, int timeout_ms)
1789 {
1790         struct rdma_id_private *id_priv;
1791         int ret;
1792
1793         id_priv = container_of(id, struct rdma_id_private, id);
1794         if (id_priv->state == CMA_IDLE) {
1795                 ret = cma_bind_addr(id, src_addr, dst_addr);
1796                 if (ret)
1797                         return ret;
1798         }
1799
1800         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1801                 return -EINVAL;
1802
1803         atomic_inc(&id_priv->refcount);
1804         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1805         if (cma_any_addr(dst_addr))
1806                 ret = cma_resolve_loopback(id_priv);
1807         else
1808                 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr,
1809                                       dst_addr, &id->route.addr.dev_addr,
1810                                       timeout_ms, addr_handler, id_priv);
1811         if (ret)
1812                 goto err;
1813
1814         return 0;
1815 err:
1816         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1817         cma_deref_id(id_priv);
1818         return ret;
1819 }
1820 EXPORT_SYMBOL(rdma_resolve_addr);
1821
1822 static void cma_bind_port(struct rdma_bind_list *bind_list,
1823                           struct rdma_id_private *id_priv)
1824 {
1825         struct sockaddr_in *sin;
1826
1827         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1828         sin->sin_port = htons(bind_list->port);
1829         id_priv->bind_list = bind_list;
1830         hlist_add_head(&id_priv->node, &bind_list->owners);
1831 }
1832
1833 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1834                           unsigned short snum)
1835 {
1836         struct rdma_bind_list *bind_list;
1837         int port, ret;
1838
1839         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1840         if (!bind_list)
1841                 return -ENOMEM;
1842
1843         do {
1844                 ret = idr_get_new_above(ps, bind_list, snum, &port);
1845         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1846
1847         if (ret)
1848                 goto err1;
1849
1850         if (port != snum) {
1851                 ret = -EADDRNOTAVAIL;
1852                 goto err2;
1853         }
1854
1855         bind_list->ps = ps;
1856         bind_list->port = (unsigned short) port;
1857         cma_bind_port(bind_list, id_priv);
1858         return 0;
1859 err2:
1860         idr_remove(ps, port);
1861 err1:
1862         kfree(bind_list);
1863         return ret;
1864 }
1865
1866 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1867 {
1868         struct rdma_bind_list *bind_list;
1869         int port, ret;
1870
1871         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1872         if (!bind_list)
1873                 return -ENOMEM;
1874
1875 retry:
1876         do {
1877                 ret = idr_get_new_above(ps, bind_list, next_port, &port);
1878         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1879
1880         if (ret)
1881                 goto err1;
1882
1883         if (port > sysctl_local_port_range[1]) {
1884                 if (next_port != sysctl_local_port_range[0]) {
1885                         idr_remove(ps, port);
1886                         next_port = sysctl_local_port_range[0];
1887                         goto retry;
1888                 }
1889                 ret = -EADDRNOTAVAIL;
1890                 goto err2;
1891         }
1892
1893         if (port == sysctl_local_port_range[1])
1894                 next_port = sysctl_local_port_range[0];
1895         else
1896                 next_port = port + 1;
1897
1898         bind_list->ps = ps;
1899         bind_list->port = (unsigned short) port;
1900         cma_bind_port(bind_list, id_priv);
1901         return 0;
1902 err2:
1903         idr_remove(ps, port);
1904 err1:
1905         kfree(bind_list);
1906         return ret;
1907 }
1908
1909 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1910 {
1911         struct rdma_id_private *cur_id;
1912         struct sockaddr_in *sin, *cur_sin;
1913         struct rdma_bind_list *bind_list;
1914         struct hlist_node *node;
1915         unsigned short snum;
1916
1917         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1918         snum = ntohs(sin->sin_port);
1919         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1920                 return -EACCES;
1921
1922         bind_list = idr_find(ps, snum);
1923         if (!bind_list)
1924                 return cma_alloc_port(ps, id_priv, snum);
1925
1926         /*
1927          * We don't support binding to any address if anyone is bound to
1928          * a specific address on the same port.
1929          */
1930         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1931                 return -EADDRNOTAVAIL;
1932
1933         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1934                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1935                         return -EADDRNOTAVAIL;
1936
1937                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1938                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1939                         return -EADDRINUSE;
1940         }
1941
1942         cma_bind_port(bind_list, id_priv);
1943         return 0;
1944 }
1945
1946 static int cma_get_port(struct rdma_id_private *id_priv)
1947 {
1948         struct idr *ps;
1949         int ret;
1950
1951         switch (id_priv->id.ps) {
1952         case RDMA_PS_SDP:
1953                 ps = &sdp_ps;
1954                 break;
1955         case RDMA_PS_TCP:
1956                 ps = &tcp_ps;
1957                 break;
1958         case RDMA_PS_UDP:
1959                 ps = &udp_ps;
1960                 break;
1961         case RDMA_PS_IPOIB:
1962                 ps = &ipoib_ps;
1963                 break;
1964         default:
1965                 return -EPROTONOSUPPORT;
1966         }
1967
1968         mutex_lock(&lock);
1969         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1970                 ret = cma_alloc_any_port(ps, id_priv);
1971         else
1972                 ret = cma_use_port(ps, id_priv);
1973         mutex_unlock(&lock);
1974
1975         return ret;
1976 }
1977
1978 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1979 {
1980         struct rdma_id_private *id_priv;
1981         int ret;
1982
1983         if (addr->sa_family != AF_INET)
1984                 return -EAFNOSUPPORT;
1985
1986         id_priv = container_of(id, struct rdma_id_private, id);
1987         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1988                 return -EINVAL;
1989
1990         if (!cma_any_addr(addr)) {
1991                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1992                 if (ret)
1993                         goto err1;
1994
1995                 mutex_lock(&lock);
1996                 ret = cma_acquire_dev(id_priv);
1997                 mutex_unlock(&lock);
1998                 if (ret)
1999                         goto err1;
2000         }
2001
2002         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2003         ret = cma_get_port(id_priv);
2004         if (ret)
2005                 goto err2;
2006
2007         return 0;
2008 err2:
2009         if (!cma_any_addr(addr)) {
2010                 mutex_lock(&lock);
2011                 cma_detach_from_dev(id_priv);
2012                 mutex_unlock(&lock);
2013         }
2014 err1:
2015         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2016         return ret;
2017 }
2018 EXPORT_SYMBOL(rdma_bind_addr);
2019
2020 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2021                           struct rdma_route *route)
2022 {
2023         struct sockaddr_in *src4, *dst4;
2024         struct cma_hdr *cma_hdr;
2025         struct sdp_hh *sdp_hdr;
2026
2027         src4 = (struct sockaddr_in *) &route->addr.src_addr;
2028         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2029
2030         switch (ps) {
2031         case RDMA_PS_SDP:
2032                 sdp_hdr = hdr;
2033                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2034                         return -EINVAL;
2035                 sdp_set_ip_ver(sdp_hdr, 4);
2036                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2037                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2038                 sdp_hdr->port = src4->sin_port;
2039                 break;
2040         default:
2041                 cma_hdr = hdr;
2042                 cma_hdr->cma_version = CMA_VERSION;
2043                 cma_set_ip_ver(cma_hdr, 4);
2044                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2045                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2046                 cma_hdr->port = src4->sin_port;
2047                 break;
2048         }
2049         return 0;
2050 }
2051
2052 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2053                                 struct ib_cm_event *ib_event)
2054 {
2055         struct rdma_id_private *id_priv = cm_id->context;
2056         struct rdma_cm_event event;
2057         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2058         int ret = 0;
2059
2060         if (cma_disable_remove(id_priv, CMA_CONNECT))
2061                 return 0;
2062
2063         memset(&event, 0, sizeof event);
2064         switch (ib_event->event) {
2065         case IB_CM_SIDR_REQ_ERROR:
2066                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2067                 event.status = -ETIMEDOUT;
2068                 break;
2069         case IB_CM_SIDR_REP_RECEIVED:
2070                 event.param.ud.private_data = ib_event->private_data;
2071                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2072                 if (rep->status != IB_SIDR_SUCCESS) {
2073                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2074                         event.status = ib_event->param.sidr_rep_rcvd.status;
2075                         break;
2076                 }
2077                 if (id_priv->qkey != rep->qkey) {
2078                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2079                         event.status = -EINVAL;
2080                         break;
2081                 }
2082                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2083                                      id_priv->id.route.path_rec,
2084                                      &event.param.ud.ah_attr);
2085                 event.param.ud.qp_num = rep->qpn;
2086                 event.param.ud.qkey = rep->qkey;
2087                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2088                 event.status = 0;
2089                 break;
2090         default:
2091                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
2092                        ib_event->event);
2093                 goto out;
2094         }
2095
2096         ret = id_priv->id.event_handler(&id_priv->id, &event);
2097         if (ret) {
2098                 /* Destroy the CM ID by returning a non-zero value. */
2099                 id_priv->cm_id.ib = NULL;
2100                 cma_exch(id_priv, CMA_DESTROYING);
2101                 cma_enable_remove(id_priv);
2102                 rdma_destroy_id(&id_priv->id);
2103                 return ret;
2104         }
2105 out:
2106         cma_enable_remove(id_priv);
2107         return ret;
2108 }
2109
2110 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2111                               struct rdma_conn_param *conn_param)
2112 {
2113         struct ib_cm_sidr_req_param req;
2114         struct rdma_route *route;
2115         int ret;
2116
2117         req.private_data_len = sizeof(struct cma_hdr) +
2118                                conn_param->private_data_len;
2119         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2120         if (!req.private_data)
2121                 return -ENOMEM;
2122
2123         if (conn_param->private_data && conn_param->private_data_len)
2124                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2125                        conn_param->private_data, conn_param->private_data_len);
2126
2127         route = &id_priv->id.route;
2128         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2129         if (ret)
2130                 goto out;
2131
2132         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2133                                             cma_sidr_rep_handler, id_priv);
2134         if (IS_ERR(id_priv->cm_id.ib)) {
2135                 ret = PTR_ERR(id_priv->cm_id.ib);
2136                 goto out;
2137         }
2138
2139         req.path = route->path_rec;
2140         req.service_id = cma_get_service_id(id_priv->id.ps,
2141                                             &route->addr.dst_addr);
2142         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2143         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2144
2145         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2146         if (ret) {
2147                 ib_destroy_cm_id(id_priv->cm_id.ib);
2148                 id_priv->cm_id.ib = NULL;
2149         }
2150 out:
2151         kfree(req.private_data);
2152         return ret;
2153 }
2154
2155 static int cma_connect_ib(struct rdma_id_private *id_priv,
2156                           struct rdma_conn_param *conn_param)
2157 {
2158         struct ib_cm_req_param req;
2159         struct rdma_route *route;
2160         void *private_data;
2161         int offset, ret;
2162
2163         memset(&req, 0, sizeof req);
2164         offset = cma_user_data_offset(id_priv->id.ps);
2165         req.private_data_len = offset + conn_param->private_data_len;
2166         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2167         if (!private_data)
2168                 return -ENOMEM;
2169
2170         if (conn_param->private_data && conn_param->private_data_len)
2171                 memcpy(private_data + offset, conn_param->private_data,
2172                        conn_param->private_data_len);
2173
2174         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2175                                             id_priv);
2176         if (IS_ERR(id_priv->cm_id.ib)) {
2177                 ret = PTR_ERR(id_priv->cm_id.ib);
2178                 goto out;
2179         }
2180
2181         route = &id_priv->id.route;
2182         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2183         if (ret)
2184                 goto out;
2185         req.private_data = private_data;
2186
2187         req.primary_path = &route->path_rec[0];
2188         if (route->num_paths == 2)
2189                 req.alternate_path = &route->path_rec[1];
2190
2191         req.service_id = cma_get_service_id(id_priv->id.ps,
2192                                             &route->addr.dst_addr);
2193         req.qp_num = id_priv->qp_num;
2194         req.qp_type = IB_QPT_RC;
2195         req.starting_psn = id_priv->seq_num;
2196         req.responder_resources = conn_param->responder_resources;
2197         req.initiator_depth = conn_param->initiator_depth;
2198         req.flow_control = conn_param->flow_control;
2199         req.retry_count = conn_param->retry_count;
2200         req.rnr_retry_count = conn_param->rnr_retry_count;
2201         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2202         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2203         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2204         req.srq = id_priv->srq ? 1 : 0;
2205
2206         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2207 out:
2208         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2209                 ib_destroy_cm_id(id_priv->cm_id.ib);
2210                 id_priv->cm_id.ib = NULL;
2211         }
2212
2213         kfree(private_data);
2214         return ret;
2215 }
2216
2217 static int cma_connect_iw(struct rdma_id_private *id_priv,
2218                           struct rdma_conn_param *conn_param)
2219 {
2220         struct iw_cm_id *cm_id;
2221         struct sockaddr_in* sin;
2222         int ret;
2223         struct iw_cm_conn_param iw_param;
2224
2225         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2226         if (IS_ERR(cm_id)) {
2227                 ret = PTR_ERR(cm_id);
2228                 goto out;
2229         }
2230
2231         id_priv->cm_id.iw = cm_id;
2232
2233         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2234         cm_id->local_addr = *sin;
2235
2236         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2237         cm_id->remote_addr = *sin;
2238
2239         ret = cma_modify_qp_rtr(&id_priv->id);
2240         if (ret)
2241                 goto out;
2242
2243         iw_param.ord = conn_param->initiator_depth;
2244         iw_param.ird = conn_param->responder_resources;
2245         iw_param.private_data = conn_param->private_data;
2246         iw_param.private_data_len = conn_param->private_data_len;
2247         if (id_priv->id.qp)
2248                 iw_param.qpn = id_priv->qp_num;
2249         else
2250                 iw_param.qpn = conn_param->qp_num;
2251         ret = iw_cm_connect(cm_id, &iw_param);
2252 out:
2253         if (ret && !IS_ERR(cm_id)) {
2254                 iw_destroy_cm_id(cm_id);
2255                 id_priv->cm_id.iw = NULL;
2256         }
2257         return ret;
2258 }
2259
2260 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2261 {
2262         struct rdma_id_private *id_priv;
2263         int ret;
2264
2265         id_priv = container_of(id, struct rdma_id_private, id);
2266         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2267                 return -EINVAL;
2268
2269         if (!id->qp) {
2270                 id_priv->qp_num = conn_param->qp_num;
2271                 id_priv->srq = conn_param->srq;
2272         }
2273
2274         switch (rdma_node_get_transport(id->device->node_type)) {
2275         case RDMA_TRANSPORT_IB:
2276                 if (cma_is_ud_ps(id->ps))
2277                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2278                 else
2279                         ret = cma_connect_ib(id_priv, conn_param);
2280                 break;
2281         case RDMA_TRANSPORT_IWARP:
2282                 ret = cma_connect_iw(id_priv, conn_param);
2283                 break;
2284         default:
2285                 ret = -ENOSYS;
2286                 break;
2287         }
2288         if (ret)
2289                 goto err;
2290
2291         return 0;
2292 err:
2293         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2294         return ret;
2295 }
2296 EXPORT_SYMBOL(rdma_connect);
2297
2298 static int cma_accept_ib(struct rdma_id_private *id_priv,
2299                          struct rdma_conn_param *conn_param)
2300 {
2301         struct ib_cm_rep_param rep;
2302         struct ib_qp_attr qp_attr;
2303         int qp_attr_mask, ret;
2304
2305         if (id_priv->id.qp) {
2306                 ret = cma_modify_qp_rtr(&id_priv->id);
2307                 if (ret)
2308                         goto out;
2309
2310                 qp_attr.qp_state = IB_QPS_RTS;
2311                 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, &qp_attr,
2312                                          &qp_attr_mask);
2313                 if (ret)
2314                         goto out;
2315
2316                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
2317                 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
2318                 if (ret)
2319                         goto out;
2320         }
2321
2322         memset(&rep, 0, sizeof rep);
2323         rep.qp_num = id_priv->qp_num;
2324         rep.starting_psn = id_priv->seq_num;
2325         rep.private_data = conn_param->private_data;
2326         rep.private_data_len = conn_param->private_data_len;
2327         rep.responder_resources = conn_param->responder_resources;
2328         rep.initiator_depth = conn_param->initiator_depth;
2329         rep.failover_accepted = 0;
2330         rep.flow_control = conn_param->flow_control;
2331         rep.rnr_retry_count = conn_param->rnr_retry_count;
2332         rep.srq = id_priv->srq ? 1 : 0;
2333
2334         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2335 out:
2336         return ret;
2337 }
2338
2339 static int cma_accept_iw(struct rdma_id_private *id_priv,
2340                   struct rdma_conn_param *conn_param)
2341 {
2342         struct iw_cm_conn_param iw_param;
2343         int ret;
2344
2345         ret = cma_modify_qp_rtr(&id_priv->id);
2346         if (ret)
2347                 return ret;
2348
2349         iw_param.ord = conn_param->initiator_depth;
2350         iw_param.ird = conn_param->responder_resources;
2351         iw_param.private_data = conn_param->private_data;
2352         iw_param.private_data_len = conn_param->private_data_len;
2353         if (id_priv->id.qp) {
2354                 iw_param.qpn = id_priv->qp_num;
2355         } else
2356                 iw_param.qpn = conn_param->qp_num;
2357
2358         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2359 }
2360
2361 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2362                              enum ib_cm_sidr_status status,
2363                              const void *private_data, int private_data_len)
2364 {
2365         struct ib_cm_sidr_rep_param rep;
2366
2367         memset(&rep, 0, sizeof rep);
2368         rep.status = status;
2369         if (status == IB_SIDR_SUCCESS) {
2370                 rep.qp_num = id_priv->qp_num;
2371                 rep.qkey = id_priv->qkey;
2372         }
2373         rep.private_data = private_data;
2374         rep.private_data_len = private_data_len;
2375
2376         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2377 }
2378
2379 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2380 {
2381         struct rdma_id_private *id_priv;
2382         int ret;
2383
2384         id_priv = container_of(id, struct rdma_id_private, id);
2385         if (!cma_comp(id_priv, CMA_CONNECT))
2386                 return -EINVAL;
2387
2388         if (!id->qp && conn_param) {
2389                 id_priv->qp_num = conn_param->qp_num;
2390                 id_priv->srq = conn_param->srq;
2391         }
2392
2393         switch (rdma_node_get_transport(id->device->node_type)) {
2394         case RDMA_TRANSPORT_IB:
2395                 if (cma_is_ud_ps(id->ps))
2396                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2397                                                 conn_param->private_data,
2398                                                 conn_param->private_data_len);
2399                 else if (conn_param)
2400                         ret = cma_accept_ib(id_priv, conn_param);
2401                 else
2402                         ret = cma_rep_recv(id_priv);
2403                 break;
2404         case RDMA_TRANSPORT_IWARP:
2405                 ret = cma_accept_iw(id_priv, conn_param);
2406                 break;
2407         default:
2408                 ret = -ENOSYS;
2409                 break;
2410         }
2411
2412         if (ret)
2413                 goto reject;
2414
2415         return 0;
2416 reject:
2417         cma_modify_qp_err(id);
2418         rdma_reject(id, NULL, 0);
2419         return ret;
2420 }
2421 EXPORT_SYMBOL(rdma_accept);
2422
2423 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2424 {
2425         struct rdma_id_private *id_priv;
2426         int ret;
2427
2428         id_priv = container_of(id, struct rdma_id_private, id);
2429         if (!cma_has_cm_dev(id_priv))
2430                 return -EINVAL;
2431
2432         switch (id->device->node_type) {
2433         case RDMA_NODE_IB_CA:
2434                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2435                 break;
2436         default:
2437                 ret = 0;
2438                 break;
2439         }
2440         return ret;
2441 }
2442 EXPORT_SYMBOL(rdma_notify);
2443
2444 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2445                 u8 private_data_len)
2446 {
2447         struct rdma_id_private *id_priv;
2448         int ret;
2449
2450         id_priv = container_of(id, struct rdma_id_private, id);
2451         if (!cma_has_cm_dev(id_priv))
2452                 return -EINVAL;
2453
2454         switch (rdma_node_get_transport(id->device->node_type)) {
2455         case RDMA_TRANSPORT_IB:
2456                 if (cma_is_ud_ps(id->ps))
2457                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2458                                                 private_data, private_data_len);
2459                 else
2460                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2461                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2462                                              0, private_data, private_data_len);
2463                 break;
2464         case RDMA_TRANSPORT_IWARP:
2465                 ret = iw_cm_reject(id_priv->cm_id.iw,
2466                                    private_data, private_data_len);
2467                 break;
2468         default:
2469                 ret = -ENOSYS;
2470                 break;
2471         }
2472         return ret;
2473 }
2474 EXPORT_SYMBOL(rdma_reject);
2475
2476 int rdma_disconnect(struct rdma_cm_id *id)
2477 {
2478         struct rdma_id_private *id_priv;
2479         int ret;
2480
2481         id_priv = container_of(id, struct rdma_id_private, id);
2482         if (!cma_has_cm_dev(id_priv))
2483                 return -EINVAL;
2484
2485         switch (rdma_node_get_transport(id->device->node_type)) {
2486         case RDMA_TRANSPORT_IB:
2487                 ret = cma_modify_qp_err(id);
2488                 if (ret)
2489                         goto out;
2490                 /* Initiate or respond to a disconnect. */
2491                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2492                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2493                 break;
2494         case RDMA_TRANSPORT_IWARP:
2495                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2496                 break;
2497         default:
2498                 ret = -EINVAL;
2499                 break;
2500         }
2501 out:
2502         return ret;
2503 }
2504 EXPORT_SYMBOL(rdma_disconnect);
2505
2506 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2507 {
2508         struct rdma_id_private *id_priv;
2509         struct cma_multicast *mc = multicast->context;
2510         struct rdma_cm_event event;
2511         int ret;
2512
2513         id_priv = mc->id_priv;
2514         if (cma_disable_remove(id_priv, CMA_ADDR_BOUND) &&
2515             cma_disable_remove(id_priv, CMA_ADDR_RESOLVED))
2516                 return 0;
2517
2518         if (!status && id_priv->id.qp)
2519                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2520                                          multicast->rec.mlid);
2521
2522         memset(&event, 0, sizeof event);
2523         event.status = status;
2524         event.param.ud.private_data = mc->context;
2525         if (!status) {
2526                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2527                 ib_init_ah_from_mcmember(id_priv->id.device,
2528                                          id_priv->id.port_num, &multicast->rec,
2529                                          &event.param.ud.ah_attr);
2530                 event.param.ud.qp_num = 0xFFFFFF;
2531                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2532         } else
2533                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2534
2535         ret = id_priv->id.event_handler(&id_priv->id, &event);
2536         if (ret) {
2537                 cma_exch(id_priv, CMA_DESTROYING);
2538                 cma_enable_remove(id_priv);
2539                 rdma_destroy_id(&id_priv->id);
2540                 return 0;
2541         }
2542
2543         cma_enable_remove(id_priv);
2544         return 0;
2545 }
2546
2547 static void cma_set_mgid(struct rdma_id_private *id_priv,
2548                          struct sockaddr *addr, union ib_gid *mgid)
2549 {
2550         unsigned char mc_map[MAX_ADDR_LEN];
2551         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2552         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2553         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2554
2555         if (cma_any_addr(addr)) {
2556                 memset(mgid, 0, sizeof *mgid);
2557         } else if ((addr->sa_family == AF_INET6) &&
2558                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) ==
2559                                                                  0xFF10A01B)) {
2560                 /* IPv6 address is an SA assigned MGID. */
2561                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2562         } else {
2563                 ip_ib_mc_map(sin->sin_addr.s_addr, mc_map);
2564                 if (id_priv->id.ps == RDMA_PS_UDP)
2565                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2566                 mc_map[8] = ib_addr_get_pkey(dev_addr) >> 8;
2567                 mc_map[9] = (unsigned char) ib_addr_get_pkey(dev_addr);
2568                 *mgid = *(union ib_gid *) (mc_map + 4);
2569         }
2570 }
2571
2572 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2573                                  struct cma_multicast *mc)
2574 {
2575         struct ib_sa_mcmember_rec rec;
2576         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2577         ib_sa_comp_mask comp_mask;
2578         int ret;
2579
2580         ib_addr_get_mgid(dev_addr, &rec.mgid);
2581         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2582                                      &rec.mgid, &rec);
2583         if (ret)
2584                 return ret;
2585
2586         cma_set_mgid(id_priv, &mc->addr, &rec.mgid);
2587         if (id_priv->id.ps == RDMA_PS_UDP)
2588                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2589         ib_addr_get_sgid(dev_addr, &rec.port_gid);
2590         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2591         rec.join_state = 1;
2592
2593         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2594                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2595                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2596                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2597                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2598
2599         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2600                                                 id_priv->id.port_num, &rec,
2601                                                 comp_mask, GFP_KERNEL,
2602                                                 cma_ib_mc_handler, mc);
2603         if (IS_ERR(mc->multicast.ib))
2604                 return PTR_ERR(mc->multicast.ib);
2605
2606         return 0;
2607 }
2608
2609 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2610                         void *context)
2611 {
2612         struct rdma_id_private *id_priv;
2613         struct cma_multicast *mc;
2614         int ret;
2615
2616         id_priv = container_of(id, struct rdma_id_private, id);
2617         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2618             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2619                 return -EINVAL;
2620
2621         mc = kmalloc(sizeof *mc, GFP_KERNEL);
2622         if (!mc)
2623                 return -ENOMEM;
2624
2625         memcpy(&mc->addr, addr, ip_addr_size(addr));
2626         mc->context = context;
2627         mc->id_priv = id_priv;
2628
2629         spin_lock(&id_priv->lock);
2630         list_add(&mc->list, &id_priv->mc_list);
2631         spin_unlock(&id_priv->lock);
2632
2633         switch (rdma_node_get_transport(id->device->node_type)) {
2634         case RDMA_TRANSPORT_IB:
2635                 ret = cma_join_ib_multicast(id_priv, mc);
2636                 break;
2637         default:
2638                 ret = -ENOSYS;
2639                 break;
2640         }
2641
2642         if (ret) {
2643                 spin_lock_irq(&id_priv->lock);
2644                 list_del(&mc->list);
2645                 spin_unlock_irq(&id_priv->lock);
2646                 kfree(mc);
2647         }
2648         return ret;
2649 }
2650 EXPORT_SYMBOL(rdma_join_multicast);
2651
2652 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2653 {
2654         struct rdma_id_private *id_priv;
2655         struct cma_multicast *mc;
2656
2657         id_priv = container_of(id, struct rdma_id_private, id);
2658         spin_lock_irq(&id_priv->lock);
2659         list_for_each_entry(mc, &id_priv->mc_list, list) {
2660                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2661                         list_del(&mc->list);
2662                         spin_unlock_irq(&id_priv->lock);
2663
2664                         if (id->qp)
2665                                 ib_detach_mcast(id->qp,
2666                                                 &mc->multicast.ib->rec.mgid,
2667                                                 mc->multicast.ib->rec.mlid);
2668                         ib_sa_free_multicast(mc->multicast.ib);
2669                         kfree(mc);
2670                         return;
2671                 }
2672         }
2673         spin_unlock_irq(&id_priv->lock);
2674 }
2675 EXPORT_SYMBOL(rdma_leave_multicast);
2676
2677 static void cma_add_one(struct ib_device *device)
2678 {
2679         struct cma_device *cma_dev;
2680         struct rdma_id_private *id_priv;
2681
2682         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2683         if (!cma_dev)
2684                 return;
2685
2686         cma_dev->device = device;
2687
2688         init_completion(&cma_dev->comp);
2689         atomic_set(&cma_dev->refcount, 1);
2690         INIT_LIST_HEAD(&cma_dev->id_list);
2691         ib_set_client_data(device, &cma_client, cma_dev);
2692
2693         mutex_lock(&lock);
2694         list_add_tail(&cma_dev->list, &dev_list);
2695         list_for_each_entry(id_priv, &listen_any_list, list)
2696                 cma_listen_on_dev(id_priv, cma_dev);
2697         mutex_unlock(&lock);
2698 }
2699
2700 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2701 {
2702         struct rdma_cm_event event;
2703         enum cma_state state;
2704
2705         /* Record that we want to remove the device */
2706         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2707         if (state == CMA_DESTROYING)
2708                 return 0;
2709
2710         cma_cancel_operation(id_priv, state);
2711         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2712
2713         /* Check for destruction from another callback. */
2714         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2715                 return 0;
2716
2717         memset(&event, 0, sizeof event);
2718         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2719         return id_priv->id.event_handler(&id_priv->id, &event);
2720 }
2721
2722 static void cma_process_remove(struct cma_device *cma_dev)
2723 {
2724         struct rdma_id_private *id_priv;
2725         int ret;
2726
2727         mutex_lock(&lock);
2728         while (!list_empty(&cma_dev->id_list)) {
2729                 id_priv = list_entry(cma_dev->id_list.next,
2730                                      struct rdma_id_private, list);
2731
2732                 if (cma_internal_listen(id_priv)) {
2733                         cma_destroy_listen(id_priv);
2734                         continue;
2735                 }
2736
2737                 list_del_init(&id_priv->list);
2738                 atomic_inc(&id_priv->refcount);
2739                 mutex_unlock(&lock);
2740
2741                 ret = cma_remove_id_dev(id_priv);
2742                 cma_deref_id(id_priv);
2743                 if (ret)
2744                         rdma_destroy_id(&id_priv->id);
2745
2746                 mutex_lock(&lock);
2747         }
2748         mutex_unlock(&lock);
2749
2750         cma_deref_dev(cma_dev);
2751         wait_for_completion(&cma_dev->comp);
2752 }
2753
2754 static void cma_remove_one(struct ib_device *device)
2755 {
2756         struct cma_device *cma_dev;
2757
2758         cma_dev = ib_get_client_data(device, &cma_client);
2759         if (!cma_dev)
2760                 return;
2761
2762         mutex_lock(&lock);
2763         list_del(&cma_dev->list);
2764         mutex_unlock(&lock);
2765
2766         cma_process_remove(cma_dev);
2767         kfree(cma_dev);
2768 }
2769
2770 static int cma_init(void)
2771 {
2772         int ret;
2773
2774         get_random_bytes(&next_port, sizeof next_port);
2775         next_port = ((unsigned int) next_port %
2776                     (sysctl_local_port_range[1] - sysctl_local_port_range[0])) +
2777                     sysctl_local_port_range[0];
2778         cma_wq = create_singlethread_workqueue("rdma_cm");
2779         if (!cma_wq)
2780                 return -ENOMEM;
2781
2782         ib_sa_register_client(&sa_client);
2783         rdma_addr_register_client(&addr_client);
2784
2785         ret = ib_register_client(&cma_client);
2786         if (ret)
2787                 goto err;
2788         return 0;
2789
2790 err:
2791         rdma_addr_unregister_client(&addr_client);
2792         ib_sa_unregister_client(&sa_client);
2793         destroy_workqueue(cma_wq);
2794         return ret;
2795 }
2796
2797 static void cma_cleanup(void)
2798 {
2799         ib_unregister_client(&cma_client);
2800         rdma_addr_unregister_client(&addr_client);
2801         ib_sa_unregister_client(&sa_client);
2802         destroy_workqueue(cma_wq);
2803         idr_destroy(&sdp_ps);
2804         idr_destroy(&tcp_ps);
2805         idr_destroy(&udp_ps);
2806         idr_destroy(&ipoib_ps);
2807 }
2808
2809 module_init(cma_init);
2810 module_exit(cma_cleanup);