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 available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/idr.h>
  42 #include <linux/inetdevice.h>
  43
  44 #include <net/tcp.h>
  45 #include <net/ipv6.h>
  46
  47 #include <rdma/rdma_cm.h>
  48 #include <rdma/rdma_cm_ib.h>
  49 #include <rdma/ib_cache.h>
  50 #include <rdma/ib_cm.h>
  51 #include <rdma/ib_sa.h>
  52 #include <rdma/iw_cm.h>
  53
  54 MODULE_AUTHOR("Sean Hefty");
  55 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 #define CMA_CM_RESPONSE_TIMEOUT 20
  59 #define CMA_MAX_CM_RETRIES 15
  60 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  61
  62 static void cma_add_one(struct ib_device *device);
  63 static void cma_remove_one(struct ib_device *device);
  64
  65 static struct ib_client cma_client = {
  66         .name   = "cma",
  67         .add    = cma_add_one,
  68         .remove = cma_remove_one
  69 };
  70
  71 static struct ib_sa_client sa_client;
  72 static struct rdma_addr_client addr_client;
  73 static LIST_HEAD(dev_list);
  74 static LIST_HEAD(listen_any_list);
  75 static DEFINE_MUTEX(lock);
  76 static struct workqueue_struct *cma_wq;
  77 static DEFINE_IDR(sdp_ps);
  78 static DEFINE_IDR(tcp_ps);
  79 static DEFINE_IDR(udp_ps);
  80 static DEFINE_IDR(ipoib_ps);
  81 static int next_port;
  82
  83 struct cma_device {
  84         struct list_head        list;
  85         struct ib_device        *device;
  86         struct completion       comp;
  87         atomic_t                refcount;
  88         struct list_head        id_list;
  89 };
  90
  91 enum cma_state {
  92         CMA_IDLE,
  93         CMA_ADDR_QUERY,
  94         CMA_ADDR_RESOLVED,
  95         CMA_ROUTE_QUERY,
  96         CMA_ROUTE_RESOLVED,
  97         CMA_CONNECT,
  98         CMA_DISCONNECT,
  99         CMA_ADDR_BOUND,
 100         CMA_LISTEN,
 101         CMA_DEVICE_REMOVAL,
 102         CMA_DESTROYING
 103 };
 104
 105 struct rdma_bind_list {
 106         struct idr              *ps;
 107         struct hlist_head       owners;
 108         unsigned short          port;
 109 };
 110
 111 /*
 112  * Device removal can occur at anytime, so we need extra handling to
 113  * serialize notifying the user of device removal with other callbacks.
 114  * We do this by disabling removal notification while a callback is in process,
 115  * and reporting it after the callback completes.
 116  */
 117 struct rdma_id_private {
 118         struct rdma_cm_id       id;
 119
 120         struct rdma_bind_list   *bind_list;
 121         struct hlist_node       node;
 122         struct list_head        list; /* listen_any_list or cma_device.list */
 123         struct list_head        listen_list; /* per device listens */
 124         struct cma_device       *cma_dev;
 125         struct list_head        mc_list;
 126
 127         int                     internal_id;
 128         enum cma_state          state;
 129         spinlock_t              lock;
 130         struct mutex            qp_mutex;
 131
 132         struct completion       comp;
 133         atomic_t                refcount;
 134         struct mutex            handler_mutex;
 135
 136         int                     backlog;
 137         int                     timeout_ms;
 138         struct ib_sa_query      *query;
 139         int                     query_id;
 140         union {
 141                 struct ib_cm_id *ib;
 142                 struct iw_cm_id *iw;
 143         } cm_id;
 144
 145         u32                     seq_num;
 146         u32                     qkey;
 147         u32                     qp_num;
 148         u8                      srq;
 149         u8                      tos;
 150 };
 151
 152 struct cma_multicast {
 153         struct rdma_id_private *id_priv;
 154         union {
 155                 struct ib_sa_multicast *ib;
 156         } multicast;
 157         struct list_head        list;
 158         void                    *context;
 159         struct sockaddr_storage addr;
 160 };
 161
 162 struct cma_work {
 163         struct work_struct      work;
 164         struct rdma_id_private  *id;
 165         enum cma_state          old_state;
 166         enum cma_state          new_state;
 167         struct rdma_cm_event    event;
 168 };
 169
 170 struct cma_ndev_work {
 171         struct work_struct      work;
 172         struct rdma_id_private  *id;
 173         struct rdma_cm_event    event;
 174 };
 175
 176 union cma_ip_addr {
 177         struct in6_addr ip6;
 178         struct {
 179                 __be32 pad[3];
 180                 __be32 addr;
 181         } ip4;
 182 };
 183
 184 struct cma_hdr {
 185         u8 cma_version;
 186         u8 ip_version;  /* IP version: 7:4 */
 187         __be16 port;
 188         union cma_ip_addr src_addr;
 189         union cma_ip_addr dst_addr;
 190 };
 191
 192 struct sdp_hh {
 193         u8 bsdh[16];
 194         u8 sdp_version; /* Major version: 7:4 */
 195         u8 ip_version;  /* IP version: 7:4 */
 196         u8 sdp_specific1[10];
 197         __be16 port;
 198         __be16 sdp_specific2;
 199         union cma_ip_addr src_addr;
 200         union cma_ip_addr dst_addr;
 201 };
 202
 203 struct sdp_hah {
 204         u8 bsdh[16];
 205         u8 sdp_version;
 206 };
 207
 208 #define CMA_VERSION 0x00
 209 #define SDP_MAJ_VERSION 0x2
 210
 211 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 212 {
 213         unsigned long flags;
 214         int ret;
 215
 216         spin_lock_irqsave(&id_priv->lock, flags);
 217         ret = (id_priv->state == comp);
 218         spin_unlock_irqrestore(&id_priv->lock, flags);
 219         return ret;
 220 }
 221
 222 static int cma_comp_exch(struct rdma_id_private *id_priv,
 223                          enum cma_state comp, enum cma_state exch)
 224 {
 225         unsigned long flags;
 226         int ret;
 227
 228         spin_lock_irqsave(&id_priv->lock, flags);
 229         if ((ret = (id_priv->state == comp)))
 230                 id_priv->state = exch;
 231         spin_unlock_irqrestore(&id_priv->lock, flags);
 232         return ret;
 233 }
 234
 235 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 236                                enum cma_state exch)
 237 {
 238         unsigned long flags;
 239         enum cma_state old;
 240
 241         spin_lock_irqsave(&id_priv->lock, flags);
 242         old = id_priv->state;
 243         id_priv->state = exch;
 244         spin_unlock_irqrestore(&id_priv->lock, flags);
 245         return old;
 246 }
 247
 248 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 249 {
 250         return hdr->ip_version >> 4;
 251 }
 252
 253 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 254 {
 255         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 256 }
 257
 258 static inline u8 sdp_get_majv(u8 sdp_version)
 259 {
 260         return sdp_version >> 4;
 261 }
 262
 263 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 264 {
 265         return hh->ip_version >> 4;
 266 }
 267
 268 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 269 {
 270         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 271 }
 272
 273 static inline int cma_is_ud_ps(enum rdma_port_space ps)
 274 {
 275         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
 276 }
 277
 278 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 279                               struct cma_device *cma_dev)
 280 {
 281         atomic_inc(&cma_dev->refcount);
 282         id_priv->cma_dev = cma_dev;
 283         id_priv->id.device = cma_dev->device;
 284         list_add_tail(&id_priv->list, &cma_dev->id_list);
 285 }
 286
 287 static inline void cma_deref_dev(struct cma_device *cma_dev)
 288 {
 289         if (atomic_dec_and_test(&cma_dev->refcount))
 290                 complete(&cma_dev->comp);
 291 }
 292
 293 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 294 {
 295         list_del(&id_priv->list);
 296         cma_deref_dev(id_priv->cma_dev);
 297         id_priv->cma_dev = NULL;
 298 }
 299
 300 static int cma_set_qkey(struct rdma_id_private *id_priv)
 301 {
 302         struct ib_sa_mcmember_rec rec;
 303         int ret = 0;
 304
 305         if (id_priv->qkey)
 306                 return 0;
 307
 308         switch (id_priv->id.ps) {
 309         case RDMA_PS_UDP:
 310                 id_priv->qkey = RDMA_UDP_QKEY;
 311                 break;
 312         case RDMA_PS_IPOIB:
 313                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 314                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 315                                              id_priv->id.port_num, &rec.mgid,
 316                                              &rec);
 317                 if (!ret)
 318                         id_priv->qkey = be32_to_cpu(rec.qkey);
 319                 break;
 320         default:
 321                 break;
 322         }
 323         return ret;
 324 }
 325
 326 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 327 {
 328         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 329         struct cma_device *cma_dev;
 330         union ib_gid gid;
 331         int ret = -ENODEV;
 332
 333         switch (rdma_node_get_transport(dev_addr->dev_type)) {
 334         case RDMA_TRANSPORT_IB:
 335                 ib_addr_get_sgid(dev_addr, &gid);
 336                 break;
 337         case RDMA_TRANSPORT_IWARP:
 338                 iw_addr_get_sgid(dev_addr, &gid);
 339                 break;
 340         default:
 341                 return -ENODEV;
 342         }
 343
 344         list_for_each_entry(cma_dev, &dev_list, list) {
 345                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 346                                          &id_priv->id.port_num, NULL);
 347                 if (!ret) {
 348                         cma_attach_to_dev(id_priv, cma_dev);
 349                         break;
 350                 }
 351         }
 352         return ret;
 353 }
 354
 355 static void cma_deref_id(struct rdma_id_private *id_priv)
 356 {
 357         if (atomic_dec_and_test(&id_priv->refcount))
 358                 complete(&id_priv->comp);
 359 }
 360
 361 static int cma_disable_callback(struct rdma_id_private *id_priv,
 362                               enum cma_state state)
 363 {
 364         mutex_lock(&id_priv->handler_mutex);
 365         if (id_priv->state != state) {
 366                 mutex_unlock(&id_priv->handler_mutex);
 367                 return -EINVAL;
 368         }
 369         return 0;
 370 }
 371
 372 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
 373 {
 374         return (id_priv->id.device && id_priv->cm_id.ib);
 375 }
 376
 377 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 378                                   void *context, enum rdma_port_space ps)
 379 {
 380         struct rdma_id_private *id_priv;
 381
 382         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 383         if (!id_priv)
 384                 return ERR_PTR(-ENOMEM);
 385
 386         id_priv->state = CMA_IDLE;
 387         id_priv->id.context = context;
 388         id_priv->id.event_handler = event_handler;
 389         id_priv->id.ps = ps;
 390         spin_lock_init(&id_priv->lock);
 391         mutex_init(&id_priv->qp_mutex);
 392         init_completion(&id_priv->comp);
 393         atomic_set(&id_priv->refcount, 1);
 394         mutex_init(&id_priv->handler_mutex);
 395         INIT_LIST_HEAD(&id_priv->listen_list);
 396         INIT_LIST_HEAD(&id_priv->mc_list);
 397         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 398
 399         return &id_priv->id;
 400 }
 401 EXPORT_SYMBOL(rdma_create_id);
 402
 403 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 404 {
 405         struct ib_qp_attr qp_attr;
 406         int qp_attr_mask, ret;
 407
 408         qp_attr.qp_state = IB_QPS_INIT;
 409         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 410         if (ret)
 411                 return ret;
 412
 413         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 414         if (ret)
 415                 return ret;
 416
 417         qp_attr.qp_state = IB_QPS_RTR;
 418         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 419         if (ret)
 420                 return ret;
 421
 422         qp_attr.qp_state = IB_QPS_RTS;
 423         qp_attr.sq_psn = 0;
 424         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 425
 426         return ret;
 427 }
 428
 429 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 430 {
 431         struct ib_qp_attr qp_attr;
 432         int qp_attr_mask, ret;
 433
 434         qp_attr.qp_state = IB_QPS_INIT;
 435         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 436         if (ret)
 437                 return ret;
 438
 439         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 440 }
 441
 442 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 443                    struct ib_qp_init_attr *qp_init_attr)
 444 {
 445         struct rdma_id_private *id_priv;
 446         struct ib_qp *qp;
 447         int ret;
 448
 449         id_priv = container_of(id, struct rdma_id_private, id);
 450         if (id->device != pd->device)
 451                 return -EINVAL;
 452
 453         qp = ib_create_qp(pd, qp_init_attr);
 454         if (IS_ERR(qp))
 455                 return PTR_ERR(qp);
 456
 457         if (cma_is_ud_ps(id_priv->id.ps))
 458                 ret = cma_init_ud_qp(id_priv, qp);
 459         else
 460                 ret = cma_init_conn_qp(id_priv, qp);
 461         if (ret)
 462                 goto err;
 463
 464         id->qp = qp;
 465         id_priv->qp_num = qp->qp_num;
 466         id_priv->srq = (qp->srq != NULL);
 467         return 0;
 468 err:
 469         ib_destroy_qp(qp);
 470         return ret;
 471 }
 472 EXPORT_SYMBOL(rdma_create_qp);
 473
 474 void rdma_destroy_qp(struct rdma_cm_id *id)
 475 {
 476         struct rdma_id_private *id_priv;
 477
 478         id_priv = container_of(id, struct rdma_id_private, id);
 479         mutex_lock(&id_priv->qp_mutex);
 480         ib_destroy_qp(id_priv->id.qp);
 481         id_priv->id.qp = NULL;
 482         mutex_unlock(&id_priv->qp_mutex);
 483 }
 484 EXPORT_SYMBOL(rdma_destroy_qp);
 485
 486 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 487                              struct rdma_conn_param *conn_param)
 488 {
 489         struct ib_qp_attr qp_attr;
 490         int qp_attr_mask, ret;
 491
 492         mutex_lock(&id_priv->qp_mutex);
 493         if (!id_priv->id.qp) {
 494                 ret = 0;
 495                 goto out;
 496         }
 497
 498         /* Need to update QP attributes from default values. */
 499         qp_attr.qp_state = IB_QPS_INIT;
 500         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 501         if (ret)
 502                 goto out;
 503
 504         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 505         if (ret)
 506                 goto out;
 507
 508         qp_attr.qp_state = IB_QPS_RTR;
 509         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 510         if (ret)
 511                 goto out;
 512
 513         if (conn_param)
 514                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 515         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 516 out:
 517         mutex_unlock(&id_priv->qp_mutex);
 518         return ret;
 519 }
 520
 521 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 522                              struct rdma_conn_param *conn_param)
 523 {
 524         struct ib_qp_attr qp_attr;
 525         int qp_attr_mask, ret;
 526
 527         mutex_lock(&id_priv->qp_mutex);
 528         if (!id_priv->id.qp) {
 529                 ret = 0;
 530                 goto out;
 531         }
 532
 533         qp_attr.qp_state = IB_QPS_RTS;
 534         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 535         if (ret)
 536                 goto out;
 537
 538         if (conn_param)
 539                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 540         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 541 out:
 542         mutex_unlock(&id_priv->qp_mutex);
 543         return ret;
 544 }
 545
 546 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 547 {
 548         struct ib_qp_attr qp_attr;
 549         int ret;
 550
 551         mutex_lock(&id_priv->qp_mutex);
 552         if (!id_priv->id.qp) {
 553                 ret = 0;
 554                 goto out;
 555         }
 556
 557         qp_attr.qp_state = IB_QPS_ERR;
 558         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 559 out:
 560         mutex_unlock(&id_priv->qp_mutex);
 561         return ret;
 562 }
 563
 564 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 565                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 566 {
 567         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 568         int ret;
 569
 570         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 571                                   ib_addr_get_pkey(dev_addr),
 572                                   &qp_attr->pkey_index);
 573         if (ret)
 574                 return ret;
 575
 576         qp_attr->port_num = id_priv->id.port_num;
 577         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 578
 579         if (cma_is_ud_ps(id_priv->id.ps)) {
 580                 ret = cma_set_qkey(id_priv);
 581                 if (ret)
 582                         return ret;
 583
 584                 qp_attr->qkey = id_priv->qkey;
 585                 *qp_attr_mask |= IB_QP_QKEY;
 586         } else {
 587                 qp_attr->qp_access_flags = 0;
 588                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 589         }
 590         return 0;
 591 }
 592
 593 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 594                        int *qp_attr_mask)
 595 {
 596         struct rdma_id_private *id_priv;
 597         int ret = 0;
 598
 599         id_priv = container_of(id, struct rdma_id_private, id);
 600         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 601         case RDMA_TRANSPORT_IB:
 602                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
 603                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 604                 else
 605                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 606                                                  qp_attr_mask);
 607                 if (qp_attr->qp_state == IB_QPS_RTR)
 608                         qp_attr->rq_psn = id_priv->seq_num;
 609                 break;
 610         case RDMA_TRANSPORT_IWARP:
 611                 if (!id_priv->cm_id.iw) {
 612                         qp_attr->qp_access_flags = 0;
 613                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 614                 } else
 615                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 616                                                  qp_attr_mask);
 617                 break;
 618         default:
 619                 ret = -ENOSYS;
 620                 break;
 621         }
 622
 623         return ret;
 624 }
 625 EXPORT_SYMBOL(rdma_init_qp_attr);
 626
 627 static inline int cma_zero_addr(struct sockaddr *addr)
 628 {
 629         struct in6_addr *ip6;
 630
 631         if (addr->sa_family == AF_INET)
 632                 return ipv4_is_zeronet(
 633                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 634         else {
 635                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 636                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 637                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 638         }
 639 }
 640
 641 static inline int cma_loopback_addr(struct sockaddr *addr)
 642 {
 643         if (addr->sa_family == AF_INET)
 644                 return ipv4_is_loopback(
 645                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 646         else
 647                 return ipv6_addr_loopback(
 648                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 649 }
 650
 651 static inline int cma_any_addr(struct sockaddr *addr)
 652 {
 653         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 654 }
 655
 656 static inline __be16 cma_port(struct sockaddr *addr)
 657 {
 658         if (addr->sa_family == AF_INET)
 659                 return ((struct sockaddr_in *) addr)->sin_port;
 660         else
 661                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 662 }
 663
 664 static inline int cma_any_port(struct sockaddr *addr)
 665 {
 666         return !cma_port(addr);
 667 }
 668
 669 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 670                             u8 *ip_ver, __be16 *port,
 671                             union cma_ip_addr **src, union cma_ip_addr **dst)
 672 {
 673         switch (ps) {
 674         case RDMA_PS_SDP:
 675                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 676                     SDP_MAJ_VERSION)
 677                         return -EINVAL;
 678
 679                 *ip_ver = sdp_get_ip_ver(hdr);
 680                 *port   = ((struct sdp_hh *) hdr)->port;
 681                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 682                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 683                 break;
 684         default:
 685                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 686                         return -EINVAL;
 687
 688                 *ip_ver = cma_get_ip_ver(hdr);
 689                 *port   = ((struct cma_hdr *) hdr)->port;
 690                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 691                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 692                 break;
 693         }
 694
 695         if (*ip_ver != 4 && *ip_ver != 6)
 696                 return -EINVAL;
 697         return 0;
 698 }
 699
 700 static void cma_save_net_info(struct rdma_addr *addr,
 701                               struct rdma_addr *listen_addr,
 702                               u8 ip_ver, __be16 port,
 703                               union cma_ip_addr *src, union cma_ip_addr *dst)
 704 {
 705         struct sockaddr_in *listen4, *ip4;
 706         struct sockaddr_in6 *listen6, *ip6;
 707
 708         switch (ip_ver) {
 709         case 4:
 710                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 711                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 712                 ip4->sin_family = listen4->sin_family;
 713                 ip4->sin_addr.s_addr = dst->ip4.addr;
 714                 ip4->sin_port = listen4->sin_port;
 715
 716                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 717                 ip4->sin_family = listen4->sin_family;
 718                 ip4->sin_addr.s_addr = src->ip4.addr;
 719                 ip4->sin_port = port;
 720                 break;
 721         case 6:
 722                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 723                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 724                 ip6->sin6_family = listen6->sin6_family;
 725                 ip6->sin6_addr = dst->ip6;
 726                 ip6->sin6_port = listen6->sin6_port;
 727
 728                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 729                 ip6->sin6_family = listen6->sin6_family;
 730                 ip6->sin6_addr = src->ip6;
 731                 ip6->sin6_port = port;
 732                 break;
 733         default:
 734                 break;
 735         }
 736 }
 737
 738 static inline int cma_user_data_offset(enum rdma_port_space ps)
 739 {
 740         switch (ps) {
 741         case RDMA_PS_SDP:
 742                 return 0;
 743         default:
 744                 return sizeof(struct cma_hdr);
 745         }
 746 }
 747
 748 static void cma_cancel_route(struct rdma_id_private *id_priv)
 749 {
 750         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 751         case RDMA_TRANSPORT_IB:
 752                 if (id_priv->query)
 753                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 754                 break;
 755         default:
 756                 break;
 757         }
 758 }
 759
 760 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 761 {
 762         struct rdma_id_private *dev_id_priv;
 763
 764         /*
 765          * Remove from listen_any_list to prevent added devices from spawning
 766          * additional listen requests.
 767          */
 768         mutex_lock(&lock);
 769         list_del(&id_priv->list);
 770
 771         while (!list_empty(&id_priv->listen_list)) {
 772                 dev_id_priv = list_entry(id_priv->listen_list.next,
 773                                          struct rdma_id_private, listen_list);
 774                 /* sync with device removal to avoid duplicate destruction */
 775                 list_del_init(&dev_id_priv->list);
 776                 list_del(&dev_id_priv->listen_list);
 777                 mutex_unlock(&lock);
 778
 779                 rdma_destroy_id(&dev_id_priv->id);
 780                 mutex_lock(&lock);
 781         }
 782         mutex_unlock(&lock);
 783 }
 784
 785 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 786                                  enum cma_state state)
 787 {
 788         switch (state) {
 789         case CMA_ADDR_QUERY:
 790                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 791                 break;
 792         case CMA_ROUTE_QUERY:
 793                 cma_cancel_route(id_priv);
 794                 break;
 795         case CMA_LISTEN:
 796                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 797                                 && !id_priv->cma_dev)
 798                         cma_cancel_listens(id_priv);
 799                 break;
 800         default:
 801                 break;
 802         }
 803 }
 804
 805 static void cma_release_port(struct rdma_id_private *id_priv)
 806 {
 807         struct rdma_bind_list *bind_list = id_priv->bind_list;
 808
 809         if (!bind_list)
 810                 return;
 811
 812         mutex_lock(&lock);
 813         hlist_del(&id_priv->node);
 814         if (hlist_empty(&bind_list->owners)) {
 815                 idr_remove(bind_list->ps, bind_list->port);
 816                 kfree(bind_list);
 817         }
 818         mutex_unlock(&lock);
 819 }
 820
 821 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 822 {
 823         struct cma_multicast *mc;
 824
 825         while (!list_empty(&id_priv->mc_list)) {
 826                 mc = container_of(id_priv->mc_list.next,
 827                                   struct cma_multicast, list);
 828                 list_del(&mc->list);
 829                 ib_sa_free_multicast(mc->multicast.ib);
 830                 kfree(mc);
 831         }
 832 }
 833
 834 void rdma_destroy_id(struct rdma_cm_id *id)
 835 {
 836         struct rdma_id_private *id_priv;
 837         enum cma_state state;
 838
 839         id_priv = container_of(id, struct rdma_id_private, id);
 840         state = cma_exch(id_priv, CMA_DESTROYING);
 841         cma_cancel_operation(id_priv, state);
 842
 843         mutex_lock(&lock);
 844         if (id_priv->cma_dev) {
 845                 mutex_unlock(&lock);
 846                 switch (rdma_node_get_transport(id->device->node_type)) {
 847                 case RDMA_TRANSPORT_IB:
 848                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 849                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 850                         break;
 851                 case RDMA_TRANSPORT_IWARP:
 852                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 853                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 854                         break;
 855                 default:
 856                         break;
 857                 }
 858                 cma_leave_mc_groups(id_priv);
 859                 mutex_lock(&lock);
 860                 cma_detach_from_dev(id_priv);
 861         }
 862         mutex_unlock(&lock);
 863
 864         cma_release_port(id_priv);
 865         cma_deref_id(id_priv);
 866         wait_for_completion(&id_priv->comp);
 867
 868         if (id_priv->internal_id)
 869                 cma_deref_id(id_priv->id.context);
 870
 871         kfree(id_priv->id.route.path_rec);
 872         kfree(id_priv);
 873 }
 874 EXPORT_SYMBOL(rdma_destroy_id);
 875
 876 static int cma_rep_recv(struct rdma_id_private *id_priv)
 877 {
 878         int ret;
 879
 880         ret = cma_modify_qp_rtr(id_priv, NULL);
 881         if (ret)
 882                 goto reject;
 883
 884         ret = cma_modify_qp_rts(id_priv, NULL);
 885         if (ret)
 886                 goto reject;
 887
 888         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 889         if (ret)
 890                 goto reject;
 891
 892         return 0;
 893 reject:
 894         cma_modify_qp_err(id_priv);
 895         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 896                        NULL, 0, NULL, 0);
 897         return ret;
 898 }
 899
 900 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 901 {
 902         if (id_priv->id.ps == RDMA_PS_SDP &&
 903             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 904             SDP_MAJ_VERSION)
 905                 return -EINVAL;
 906
 907         return 0;
 908 }
 909
 910 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 911                                    struct ib_cm_rep_event_param *rep_data,
 912                                    void *private_data)
 913 {
 914         event->param.conn.private_data = private_data;
 915         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 916         event->param.conn.responder_resources = rep_data->responder_resources;
 917         event->param.conn.initiator_depth = rep_data->initiator_depth;
 918         event->param.conn.flow_control = rep_data->flow_control;
 919         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 920         event->param.conn.srq = rep_data->srq;
 921         event->param.conn.qp_num = rep_data->remote_qpn;
 922 }
 923
 924 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 925 {
 926         struct rdma_id_private *id_priv = cm_id->context;
 927         struct rdma_cm_event event;
 928         int ret = 0;
 929
 930         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
 931                 cma_disable_callback(id_priv, CMA_CONNECT)) ||
 932             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
 933                 cma_disable_callback(id_priv, CMA_DISCONNECT)))
 934                 return 0;
 935
 936         memset(&event, 0, sizeof event);
 937         switch (ib_event->event) {
 938         case IB_CM_REQ_ERROR:
 939         case IB_CM_REP_ERROR:
 940                 event.event = RDMA_CM_EVENT_UNREACHABLE;
 941                 event.status = -ETIMEDOUT;
 942                 break;
 943         case IB_CM_REP_RECEIVED:
 944                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
 945                 if (event.status)
 946                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
 947                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 948                         event.status = cma_rep_recv(id_priv);
 949                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
 950                                                      RDMA_CM_EVENT_ESTABLISHED;
 951                 } else
 952                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 953                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
 954                                        ib_event->private_data);
 955                 break;
 956         case IB_CM_RTU_RECEIVED:
 957         case IB_CM_USER_ESTABLISHED:
 958                 event.event = RDMA_CM_EVENT_ESTABLISHED;
 959                 break;
 960         case IB_CM_DREQ_ERROR:
 961                 event.status = -ETIMEDOUT; /* fall through */
 962         case IB_CM_DREQ_RECEIVED:
 963         case IB_CM_DREP_RECEIVED:
 964                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 965                         goto out;
 966                 event.event = RDMA_CM_EVENT_DISCONNECTED;
 967                 break;
 968         case IB_CM_TIMEWAIT_EXIT:
 969                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
 970                 break;
 971         case IB_CM_MRA_RECEIVED:
 972                 /* ignore event */
 973                 goto out;
 974         case IB_CM_REJ_RECEIVED:
 975                 cma_modify_qp_err(id_priv);
 976                 event.status = ib_event->param.rej_rcvd.reason;
 977                 event.event = RDMA_CM_EVENT_REJECTED;
 978                 event.param.conn.private_data = ib_event->private_data;
 979                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 980                 break;
 981         default:
 982                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
 983                        ib_event->event);
 984                 goto out;
 985         }
 986
 987         ret = id_priv->id.event_handler(&id_priv->id, &event);
 988         if (ret) {
 989                 /* Destroy the CM ID by returning a non-zero value. */
 990                 id_priv->cm_id.ib = NULL;
 991                 cma_exch(id_priv, CMA_DESTROYING);
 992                 mutex_unlock(&id_priv->handler_mutex);
 993                 rdma_destroy_id(&id_priv->id);
 994                 return ret;
 995         }
 996 out:
 997         mutex_unlock(&id_priv->handler_mutex);
 998         return ret;
 999 }
1000
1001 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1002                                                struct ib_cm_event *ib_event)
1003 {
1004         struct rdma_id_private *id_priv;
1005         struct rdma_cm_id *id;
1006         struct rdma_route *rt;
1007         union cma_ip_addr *src, *dst;
1008         __be16 port;
1009         u8 ip_ver;
1010         int ret;
1011
1012         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1013                              &ip_ver, &port, &src, &dst))
1014                 goto err;
1015
1016         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1017                             listen_id->ps);
1018         if (IS_ERR(id))
1019                 goto err;
1020
1021         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1022                           ip_ver, port, src, dst);
1023
1024         rt = &id->route;
1025         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1026         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1027                                GFP_KERNEL);
1028         if (!rt->path_rec)
1029                 goto destroy_id;
1030
1031         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1032         if (rt->num_paths == 2)
1033                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1034
1035         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1036         ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1037                                 &id->route.addr.dev_addr);
1038         if (ret)
1039                 goto destroy_id;
1040
1041         id_priv = container_of(id, struct rdma_id_private, id);
1042         id_priv->state = CMA_CONNECT;
1043         return id_priv;
1044
1045 destroy_id:
1046         rdma_destroy_id(id);
1047 err:
1048         return NULL;
1049 }
1050
1051 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1052                                               struct ib_cm_event *ib_event)
1053 {
1054         struct rdma_id_private *id_priv;
1055         struct rdma_cm_id *id;
1056         union cma_ip_addr *src, *dst;
1057         __be16 port;
1058         u8 ip_ver;
1059         int ret;
1060
1061         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1062                             listen_id->ps);
1063         if (IS_ERR(id))
1064                 return NULL;
1065
1066
1067         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1068                              &ip_ver, &port, &src, &dst))
1069                 goto err;
1070
1071         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1072                           ip_ver, port, src, dst);
1073
1074         ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1075                                 &id->route.addr.dev_addr);
1076         if (ret)
1077                 goto err;
1078
1079         id_priv = container_of(id, struct rdma_id_private, id);
1080         id_priv->state = CMA_CONNECT;
1081         return id_priv;
1082 err:
1083         rdma_destroy_id(id);
1084         return NULL;
1085 }
1086
1087 static void cma_set_req_event_data(struct rdma_cm_event *event,
1088                                    struct ib_cm_req_event_param *req_data,
1089                                    void *private_data, int offset)
1090 {
1091         event->param.conn.private_data = private_data + offset;
1092         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1093         event->param.conn.responder_resources = req_data->responder_resources;
1094         event->param.conn.initiator_depth = req_data->initiator_depth;
1095         event->param.conn.flow_control = req_data->flow_control;
1096         event->param.conn.retry_count = req_data->retry_count;
1097         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1098         event->param.conn.srq = req_data->srq;
1099         event->param.conn.qp_num = req_data->remote_qpn;
1100 }
1101
1102 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1103 {
1104         struct rdma_id_private *listen_id, *conn_id;
1105         struct rdma_cm_event event;
1106         int offset, ret;
1107
1108         listen_id = cm_id->context;
1109         if (cma_disable_callback(listen_id, CMA_LISTEN))
1110                 return -ECONNABORTED;
1111
1112         memset(&event, 0, sizeof event);
1113         offset = cma_user_data_offset(listen_id->id.ps);
1114         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1115         if (cma_is_ud_ps(listen_id->id.ps)) {
1116                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1117                 event.param.ud.private_data = ib_event->private_data + offset;
1118                 event.param.ud.private_data_len =
1119                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1120         } else {
1121                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1122                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1123                                        ib_event->private_data, offset);
1124         }
1125         if (!conn_id) {
1126                 ret = -ENOMEM;
1127                 goto out;
1128         }
1129
1130         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1131         mutex_lock(&lock);
1132         ret = cma_acquire_dev(conn_id);
1133         mutex_unlock(&lock);
1134         if (ret)
1135                 goto release_conn_id;
1136
1137         conn_id->cm_id.ib = cm_id;
1138         cm_id->context = conn_id;
1139         cm_id->cm_handler = cma_ib_handler;
1140
1141         /*
1142          * Protect against the user destroying conn_id from another thread
1143          * until we're done accessing it.
1144          */
1145         atomic_inc(&conn_id->refcount);
1146         ret = conn_id->id.event_handler(&conn_id->id, &event);
1147         if (!ret) {
1148                 /*
1149                  * Acquire mutex to prevent user executing rdma_destroy_id()
1150                  * while we're accessing the cm_id.
1151                  */
1152                 mutex_lock(&lock);
1153                 if (cma_comp(conn_id, CMA_CONNECT) &&
1154                     !cma_is_ud_ps(conn_id->id.ps))
1155                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1156                 mutex_unlock(&lock);
1157                 mutex_unlock(&conn_id->handler_mutex);
1158                 cma_deref_id(conn_id);
1159                 goto out;
1160         }
1161         cma_deref_id(conn_id);
1162
1163         /* Destroy the CM ID by returning a non-zero value. */
1164         conn_id->cm_id.ib = NULL;
1165
1166 release_conn_id:
1167         cma_exch(conn_id, CMA_DESTROYING);
1168         mutex_unlock(&conn_id->handler_mutex);
1169         rdma_destroy_id(&conn_id->id);
1170
1171 out:
1172         mutex_unlock(&listen_id->handler_mutex);
1173         return ret;
1174 }
1175
1176 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1177 {
1178         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1179 }
1180
1181 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1182                                  struct ib_cm_compare_data *compare)
1183 {
1184         struct cma_hdr *cma_data, *cma_mask;
1185         struct sdp_hh *sdp_data, *sdp_mask;
1186         __be32 ip4_addr;
1187         struct in6_addr ip6_addr;
1188
1189         memset(compare, 0, sizeof *compare);
1190         cma_data = (void *) compare->data;
1191         cma_mask = (void *) compare->mask;
1192         sdp_data = (void *) compare->data;
1193         sdp_mask = (void *) compare->mask;
1194
1195         switch (addr->sa_family) {
1196         case AF_INET:
1197                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1198                 if (ps == RDMA_PS_SDP) {
1199                         sdp_set_ip_ver(sdp_data, 4);
1200                         sdp_set_ip_ver(sdp_mask, 0xF);
1201                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1202                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1203                 } else {
1204                         cma_set_ip_ver(cma_data, 4);
1205                         cma_set_ip_ver(cma_mask, 0xF);
1206                         cma_data->dst_addr.ip4.addr = ip4_addr;
1207                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1208                 }
1209                 break;
1210         case AF_INET6:
1211                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1212                 if (ps == RDMA_PS_SDP) {
1213                         sdp_set_ip_ver(sdp_data, 6);
1214                         sdp_set_ip_ver(sdp_mask, 0xF);
1215                         sdp_data->dst_addr.ip6 = ip6_addr;
1216                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1217                                sizeof sdp_mask->dst_addr.ip6);
1218                 } else {
1219                         cma_set_ip_ver(cma_data, 6);
1220                         cma_set_ip_ver(cma_mask, 0xF);
1221                         cma_data->dst_addr.ip6 = ip6_addr;
1222                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1223                                sizeof cma_mask->dst_addr.ip6);
1224                 }
1225                 break;
1226         default:
1227                 break;
1228         }
1229 }
1230
1231 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1232 {
1233         struct rdma_id_private *id_priv = iw_id->context;
1234         struct rdma_cm_event event;
1235         struct sockaddr_in *sin;
1236         int ret = 0;
1237
1238         if (cma_disable_callback(id_priv, CMA_CONNECT))
1239                 return 0;
1240
1241         memset(&event, 0, sizeof event);
1242         switch (iw_event->event) {
1243         case IW_CM_EVENT_CLOSE:
1244                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1245                 break;
1246         case IW_CM_EVENT_CONNECT_REPLY:
1247                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1248                 *sin = iw_event->local_addr;
1249                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1250                 *sin = iw_event->remote_addr;
1251                 switch (iw_event->status) {
1252                 case 0:
1253                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1254                         break;
1255                 case -ECONNRESET:
1256                 case -ECONNREFUSED:
1257                         event.event = RDMA_CM_EVENT_REJECTED;
1258                         break;
1259                 case -ETIMEDOUT:
1260                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1261                         break;
1262                 default:
1263                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1264                         break;
1265                 }
1266                 break;
1267         case IW_CM_EVENT_ESTABLISHED:
1268                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1269                 break;
1270         default:
1271                 BUG_ON(1);
1272         }
1273
1274         event.status = iw_event->status;
1275         event.param.conn.private_data = iw_event->private_data;
1276         event.param.conn.private_data_len = iw_event->private_data_len;
1277         ret = id_priv->id.event_handler(&id_priv->id, &event);
1278         if (ret) {
1279                 /* Destroy the CM ID by returning a non-zero value. */
1280                 id_priv->cm_id.iw = NULL;
1281                 cma_exch(id_priv, CMA_DESTROYING);
1282                 mutex_unlock(&id_priv->handler_mutex);
1283                 rdma_destroy_id(&id_priv->id);
1284                 return ret;
1285         }
1286
1287         mutex_unlock(&id_priv->handler_mutex);
1288         return ret;
1289 }
1290
1291 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1292                                struct iw_cm_event *iw_event)
1293 {
1294         struct rdma_cm_id *new_cm_id;
1295         struct rdma_id_private *listen_id, *conn_id;
1296         struct sockaddr_in *sin;
1297         struct net_device *dev = NULL;
1298         struct rdma_cm_event event;
1299         int ret;
1300         struct ib_device_attr attr;
1301
1302         listen_id = cm_id->context;
1303         if (cma_disable_callback(listen_id, CMA_LISTEN))
1304                 return -ECONNABORTED;
1305
1306         /* Create a new RDMA id for the new IW CM ID */
1307         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1308                                    listen_id->id.context,
1309                                    RDMA_PS_TCP);
1310         if (IS_ERR(new_cm_id)) {
1311                 ret = -ENOMEM;
1312                 goto out;
1313         }
1314         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1315         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1316         conn_id->state = CMA_CONNECT;
1317
1318         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1319         if (!dev) {
1320                 ret = -EADDRNOTAVAIL;
1321                 mutex_unlock(&conn_id->handler_mutex);
1322                 rdma_destroy_id(new_cm_id);
1323                 goto out;
1324         }
1325         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1326         if (ret) {
1327                 mutex_unlock(&conn_id->handler_mutex);
1328                 rdma_destroy_id(new_cm_id);
1329                 goto out;
1330         }
1331
1332         mutex_lock(&lock);
1333         ret = cma_acquire_dev(conn_id);
1334         mutex_unlock(&lock);
1335         if (ret) {
1336                 mutex_unlock(&conn_id->handler_mutex);
1337                 rdma_destroy_id(new_cm_id);
1338                 goto out;
1339         }
1340
1341         conn_id->cm_id.iw = cm_id;
1342         cm_id->context = conn_id;
1343         cm_id->cm_handler = cma_iw_handler;
1344
1345         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1346         *sin = iw_event->local_addr;
1347         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1348         *sin = iw_event->remote_addr;
1349
1350         ret = ib_query_device(conn_id->id.device, &attr);
1351         if (ret) {
1352                 mutex_unlock(&conn_id->handler_mutex);
1353                 rdma_destroy_id(new_cm_id);
1354                 goto out;
1355         }
1356
1357         memset(&event, 0, sizeof event);
1358         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1359         event.param.conn.private_data = iw_event->private_data;
1360         event.param.conn.private_data_len = iw_event->private_data_len;
1361         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1362         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1363
1364         /*
1365          * Protect against the user destroying conn_id from another thread
1366          * until we're done accessing it.
1367          */
1368         atomic_inc(&conn_id->refcount);
1369         ret = conn_id->id.event_handler(&conn_id->id, &event);
1370         if (ret) {
1371                 /* User wants to destroy the CM ID */
1372                 conn_id->cm_id.iw = NULL;
1373                 cma_exch(conn_id, CMA_DESTROYING);
1374                 mutex_unlock(&conn_id->handler_mutex);
1375                 cma_deref_id(conn_id);
1376                 rdma_destroy_id(&conn_id->id);
1377                 goto out;
1378         }
1379
1380         mutex_unlock(&conn_id->handler_mutex);
1381         cma_deref_id(conn_id);
1382
1383 out:
1384         if (dev)
1385                 dev_put(dev);
1386         mutex_unlock(&listen_id->handler_mutex);
1387         return ret;
1388 }
1389
1390 static int cma_ib_listen(struct rdma_id_private *id_priv)
1391 {
1392         struct ib_cm_compare_data compare_data;
1393         struct sockaddr *addr;
1394         __be64 svc_id;
1395         int ret;
1396
1397         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1398                                             id_priv);
1399         if (IS_ERR(id_priv->cm_id.ib))
1400                 return PTR_ERR(id_priv->cm_id.ib);
1401
1402         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1403         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1404         if (cma_any_addr(addr))
1405                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1406         else {
1407                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1408                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1409         }
1410
1411         if (ret) {
1412                 ib_destroy_cm_id(id_priv->cm_id.ib);
1413                 id_priv->cm_id.ib = NULL;
1414         }
1415
1416         return ret;
1417 }
1418
1419 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1420 {
1421         int ret;
1422         struct sockaddr_in *sin;
1423
1424         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1425                                             iw_conn_req_handler,
1426                                             id_priv);
1427         if (IS_ERR(id_priv->cm_id.iw))
1428                 return PTR_ERR(id_priv->cm_id.iw);
1429
1430         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1431         id_priv->cm_id.iw->local_addr = *sin;
1432
1433         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1434
1435         if (ret) {
1436                 iw_destroy_cm_id(id_priv->cm_id.iw);
1437                 id_priv->cm_id.iw = NULL;
1438         }
1439
1440         return ret;
1441 }
1442
1443 static int cma_listen_handler(struct rdma_cm_id *id,
1444                               struct rdma_cm_event *event)
1445 {
1446         struct rdma_id_private *id_priv = id->context;
1447
1448         id->context = id_priv->id.context;
1449         id->event_handler = id_priv->id.event_handler;
1450         return id_priv->id.event_handler(id, event);
1451 }
1452
1453 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1454                               struct cma_device *cma_dev)
1455 {
1456         struct rdma_id_private *dev_id_priv;
1457         struct rdma_cm_id *id;
1458         int ret;
1459
1460         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1461         if (IS_ERR(id))
1462                 return;
1463
1464         dev_id_priv = container_of(id, struct rdma_id_private, id);
1465
1466         dev_id_priv->state = CMA_ADDR_BOUND;
1467         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1468                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1469
1470         cma_attach_to_dev(dev_id_priv, cma_dev);
1471         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1472         atomic_inc(&id_priv->refcount);
1473         dev_id_priv->internal_id = 1;
1474
1475         ret = rdma_listen(id, id_priv->backlog);
1476         if (ret)
1477                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1478                        "listening on device %s\n", ret, cma_dev->device->name);
1479 }
1480
1481 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1482 {
1483         struct cma_device *cma_dev;
1484
1485         mutex_lock(&lock);
1486         list_add_tail(&id_priv->list, &listen_any_list);
1487         list_for_each_entry(cma_dev, &dev_list, list)
1488                 cma_listen_on_dev(id_priv, cma_dev);
1489         mutex_unlock(&lock);
1490 }
1491
1492 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1493 {
1494         struct sockaddr_storage addr_in;
1495
1496         memset(&addr_in, 0, sizeof addr_in);
1497         addr_in.ss_family = af;
1498         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1499 }
1500
1501 int rdma_listen(struct rdma_cm_id *id, int backlog)
1502 {
1503         struct rdma_id_private *id_priv;
1504         int ret;
1505
1506         id_priv = container_of(id, struct rdma_id_private, id);
1507         if (id_priv->state == CMA_IDLE) {
1508                 ret = cma_bind_any(id, AF_INET);
1509                 if (ret)
1510                         return ret;
1511         }
1512
1513         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1514                 return -EINVAL;
1515
1516         id_priv->backlog = backlog;
1517         if (id->device) {
1518                 switch (rdma_node_get_transport(id->device->node_type)) {
1519                 case RDMA_TRANSPORT_IB:
1520                         ret = cma_ib_listen(id_priv);
1521                         if (ret)
1522                                 goto err;
1523                         break;
1524                 case RDMA_TRANSPORT_IWARP:
1525                         ret = cma_iw_listen(id_priv, backlog);
1526                         if (ret)
1527                                 goto err;
1528                         break;
1529                 default:
1530                         ret = -ENOSYS;
1531                         goto err;
1532                 }
1533         } else
1534                 cma_listen_on_all(id_priv);
1535
1536         return 0;
1537 err:
1538         id_priv->backlog = 0;
1539         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1540         return ret;
1541 }
1542 EXPORT_SYMBOL(rdma_listen);
1543
1544 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1545 {
1546         struct rdma_id_private *id_priv;
1547
1548         id_priv = container_of(id, struct rdma_id_private, id);
1549         id_priv->tos = (u8) tos;
1550 }
1551 EXPORT_SYMBOL(rdma_set_service_type);
1552
1553 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1554                               void *context)
1555 {
1556         struct cma_work *work = context;
1557         struct rdma_route *route;
1558
1559         route = &work->id->id.route;
1560
1561         if (!status) {
1562                 route->num_paths = 1;
1563                 *route->path_rec = *path_rec;
1564         } else {
1565                 work->old_state = CMA_ROUTE_QUERY;
1566                 work->new_state = CMA_ADDR_RESOLVED;
1567                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1568                 work->event.status = status;
1569         }
1570
1571         queue_work(cma_wq, &work->work);
1572 }
1573
1574 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1575                               struct cma_work *work)
1576 {
1577         struct rdma_addr *addr = &id_priv->id.route.addr;
1578         struct ib_sa_path_rec path_rec;
1579         ib_sa_comp_mask comp_mask;
1580         struct sockaddr_in6 *sin6;
1581
1582         memset(&path_rec, 0, sizeof path_rec);
1583         ib_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1584         ib_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1585         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1586         path_rec.numb_path = 1;
1587         path_rec.reversible = 1;
1588         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1589                                                         (struct sockaddr *) &addr->dst_addr);
1590
1591         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1592                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1593                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1594
1595         if (addr->src_addr.ss_family == AF_INET) {
1596                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1597                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1598         } else {
1599                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1600                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1601                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1602         }
1603
1604         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1605                                                id_priv->id.port_num, &path_rec,
1606                                                comp_mask, timeout_ms,
1607                                                GFP_KERNEL, cma_query_handler,
1608                                                work, &id_priv->query);
1609
1610         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1611 }
1612
1613 static void cma_work_handler(struct work_struct *_work)
1614 {
1615         struct cma_work *work = container_of(_work, struct cma_work, work);
1616         struct rdma_id_private *id_priv = work->id;
1617         int destroy = 0;
1618
1619         mutex_lock(&id_priv->handler_mutex);
1620         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1621                 goto out;
1622
1623         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1624                 cma_exch(id_priv, CMA_DESTROYING);
1625                 destroy = 1;
1626         }
1627 out:
1628         mutex_unlock(&id_priv->handler_mutex);
1629         cma_deref_id(id_priv);
1630         if (destroy)
1631                 rdma_destroy_id(&id_priv->id);
1632         kfree(work);
1633 }
1634
1635 static void cma_ndev_work_handler(struct work_struct *_work)
1636 {
1637         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1638         struct rdma_id_private *id_priv = work->id;
1639         int destroy = 0;
1640
1641         mutex_lock(&id_priv->handler_mutex);
1642         if (id_priv->state == CMA_DESTROYING ||
1643             id_priv->state == CMA_DEVICE_REMOVAL)
1644                 goto out;
1645
1646         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1647                 cma_exch(id_priv, CMA_DESTROYING);
1648                 destroy = 1;
1649         }
1650
1651 out:
1652         mutex_unlock(&id_priv->handler_mutex);
1653         cma_deref_id(id_priv);
1654         if (destroy)
1655                 rdma_destroy_id(&id_priv->id);
1656         kfree(work);
1657 }
1658
1659 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1660 {
1661         struct rdma_route *route = &id_priv->id.route;
1662         struct cma_work *work;
1663         int ret;
1664
1665         work = kzalloc(sizeof *work, GFP_KERNEL);
1666         if (!work)
1667                 return -ENOMEM;
1668
1669         work->id = id_priv;
1670         INIT_WORK(&work->work, cma_work_handler);
1671         work->old_state = CMA_ROUTE_QUERY;
1672         work->new_state = CMA_ROUTE_RESOLVED;
1673         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1674
1675         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1676         if (!route->path_rec) {
1677                 ret = -ENOMEM;
1678                 goto err1;
1679         }
1680
1681         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1682         if (ret)
1683                 goto err2;
1684
1685         return 0;
1686 err2:
1687         kfree(route->path_rec);
1688         route->path_rec = NULL;
1689 err1:
1690         kfree(work);
1691         return ret;
1692 }
1693
1694 int rdma_set_ib_paths(struct rdma_cm_id *id,
1695                       struct ib_sa_path_rec *path_rec, int num_paths)
1696 {
1697         struct rdma_id_private *id_priv;
1698         int ret;
1699
1700         id_priv = container_of(id, struct rdma_id_private, id);
1701         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1702                 return -EINVAL;
1703
1704         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1705         if (!id->route.path_rec) {
1706                 ret = -ENOMEM;
1707                 goto err;
1708         }
1709
1710         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1711         return 0;
1712 err:
1713         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1714         return ret;
1715 }
1716 EXPORT_SYMBOL(rdma_set_ib_paths);
1717
1718 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1719 {
1720         struct cma_work *work;
1721
1722         work = kzalloc(sizeof *work, GFP_KERNEL);
1723         if (!work)
1724                 return -ENOMEM;
1725
1726         work->id = id_priv;
1727         INIT_WORK(&work->work, cma_work_handler);
1728         work->old_state = CMA_ROUTE_QUERY;
1729         work->new_state = CMA_ROUTE_RESOLVED;
1730         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1731         queue_work(cma_wq, &work->work);
1732         return 0;
1733 }
1734
1735 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1736 {
1737         struct rdma_id_private *id_priv;
1738         int ret;
1739
1740         id_priv = container_of(id, struct rdma_id_private, id);
1741         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1742                 return -EINVAL;
1743
1744         atomic_inc(&id_priv->refcount);
1745         switch (rdma_node_get_transport(id->device->node_type)) {
1746         case RDMA_TRANSPORT_IB:
1747                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1748                 break;
1749         case RDMA_TRANSPORT_IWARP:
1750                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1751                 break;
1752         default:
1753                 ret = -ENOSYS;
1754                 break;
1755         }
1756         if (ret)
1757                 goto err;
1758
1759         return 0;
1760 err:
1761         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1762         cma_deref_id(id_priv);
1763         return ret;
1764 }
1765 EXPORT_SYMBOL(rdma_resolve_route);
1766
1767 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1768 {
1769         struct cma_device *cma_dev;
1770         struct ib_port_attr port_attr;
1771         union ib_gid gid;
1772         u16 pkey;
1773         int ret;
1774         u8 p;
1775
1776         mutex_lock(&lock);
1777         if (list_empty(&dev_list)) {
1778                 ret = -ENODEV;
1779                 goto out;
1780         }
1781         list_for_each_entry(cma_dev, &dev_list, list)
1782                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1783                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1784                             port_attr.state == IB_PORT_ACTIVE)
1785                                 goto port_found;
1786
1787         p = 1;
1788         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1789
1790 port_found:
1791         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1792         if (ret)
1793                 goto out;
1794
1795         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1796         if (ret)
1797                 goto out;
1798
1799         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1800         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1801         id_priv->id.port_num = p;
1802         cma_attach_to_dev(id_priv, cma_dev);
1803 out:
1804         mutex_unlock(&lock);
1805         return ret;
1806 }
1807
1808 static void addr_handler(int status, struct sockaddr *src_addr,
1809                          struct rdma_dev_addr *dev_addr, void *context)
1810 {
1811         struct rdma_id_private *id_priv = context;
1812         struct rdma_cm_event event;
1813
1814         memset(&event, 0, sizeof event);
1815         mutex_lock(&id_priv->handler_mutex);
1816
1817         /*
1818          * Grab mutex to block rdma_destroy_id() from removing the device while
1819          * we're trying to acquire it.
1820          */
1821         mutex_lock(&lock);
1822         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1823                 mutex_unlock(&lock);
1824                 goto out;
1825         }
1826
1827         if (!status && !id_priv->cma_dev)
1828                 status = cma_acquire_dev(id_priv);
1829         mutex_unlock(&lock);
1830
1831         if (status) {
1832                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1833                         goto out;
1834                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1835                 event.status = status;
1836         } else {
1837                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1838                        ip_addr_size(src_addr));
1839                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1840         }
1841
1842         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1843                 cma_exch(id_priv, CMA_DESTROYING);
1844                 mutex_unlock(&id_priv->handler_mutex);
1845                 cma_deref_id(id_priv);
1846                 rdma_destroy_id(&id_priv->id);
1847                 return;
1848         }
1849 out:
1850         mutex_unlock(&id_priv->handler_mutex);
1851         cma_deref_id(id_priv);
1852 }
1853
1854 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1855 {
1856         struct cma_work *work;
1857         struct sockaddr_in *src_in, *dst_in;
1858         union ib_gid gid;
1859         int ret;
1860
1861         work = kzalloc(sizeof *work, GFP_KERNEL);
1862         if (!work)
1863                 return -ENOMEM;
1864
1865         if (!id_priv->cma_dev) {
1866                 ret = cma_bind_loopback(id_priv);
1867                 if (ret)
1868                         goto err;
1869         }
1870
1871         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1872         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1873
1874         if (cma_zero_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)) {
1875                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1876                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1877                 src_in->sin_family = dst_in->sin_family;
1878                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1879         }
1880
1881         work->id = id_priv;
1882         INIT_WORK(&work->work, cma_work_handler);
1883         work->old_state = CMA_ADDR_QUERY;
1884         work->new_state = CMA_ADDR_RESOLVED;
1885         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1886         queue_work(cma_wq, &work->work);
1887         return 0;
1888 err:
1889         kfree(work);
1890         return ret;
1891 }
1892
1893 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1894                          struct sockaddr *dst_addr)
1895 {
1896         if (src_addr && src_addr->sa_family)
1897                 return rdma_bind_addr(id, src_addr);
1898         else
1899                 return cma_bind_any(id, dst_addr->sa_family);
1900 }
1901
1902 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1903                       struct sockaddr *dst_addr, int timeout_ms)
1904 {
1905         struct rdma_id_private *id_priv;
1906         int ret;
1907
1908         id_priv = container_of(id, struct rdma_id_private, id);
1909         if (id_priv->state == CMA_IDLE) {
1910                 ret = cma_bind_addr(id, src_addr, dst_addr);
1911                 if (ret)
1912                         return ret;
1913         }
1914
1915         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1916                 return -EINVAL;
1917
1918         atomic_inc(&id_priv->refcount);
1919         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1920         if (cma_any_addr(dst_addr))
1921                 ret = cma_resolve_loopback(id_priv);
1922         else
1923                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
1924                                       dst_addr, &id->route.addr.dev_addr,
1925                                       timeout_ms, addr_handler, id_priv);
1926         if (ret)
1927                 goto err;
1928
1929         return 0;
1930 err:
1931         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1932         cma_deref_id(id_priv);
1933         return ret;
1934 }
1935 EXPORT_SYMBOL(rdma_resolve_addr);
1936
1937 static void cma_bind_port(struct rdma_bind_list *bind_list,
1938                           struct rdma_id_private *id_priv)
1939 {
1940         struct sockaddr_in *sin;
1941
1942         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1943         sin->sin_port = htons(bind_list->port);
1944         id_priv->bind_list = bind_list;
1945         hlist_add_head(&id_priv->node, &bind_list->owners);
1946 }
1947
1948 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1949                           unsigned short snum)
1950 {
1951         struct rdma_bind_list *bind_list;
1952         int port, ret;
1953
1954         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1955         if (!bind_list)
1956                 return -ENOMEM;
1957
1958         do {
1959                 ret = idr_get_new_above(ps, bind_list, snum, &port);
1960         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1961
1962         if (ret)
1963                 goto err1;
1964
1965         if (port != snum) {
1966                 ret = -EADDRNOTAVAIL;
1967                 goto err2;
1968         }
1969
1970         bind_list->ps = ps;
1971         bind_list->port = (unsigned short) port;
1972         cma_bind_port(bind_list, id_priv);
1973         return 0;
1974 err2:
1975         idr_remove(ps, port);
1976 err1:
1977         kfree(bind_list);
1978         return ret;
1979 }
1980
1981 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1982 {
1983         struct rdma_bind_list *bind_list;
1984         int port, ret, low, high;
1985
1986         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1987         if (!bind_list)
1988                 return -ENOMEM;
1989
1990 retry:
1991         /* FIXME: add proper port randomization per like inet_csk_get_port */
1992         do {
1993                 ret = idr_get_new_above(ps, bind_list, next_port, &port);
1994         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1995
1996         if (ret)
1997                 goto err1;
1998
1999         inet_get_local_port_range(&low, &high);
2000         if (port > high) {
2001                 if (next_port != low) {
2002                         idr_remove(ps, port);
2003                         next_port = low;
2004                         goto retry;
2005                 }
2006                 ret = -EADDRNOTAVAIL;
2007                 goto err2;
2008         }
2009
2010         if (port == high)
2011                 next_port = low;
2012         else
2013                 next_port = port + 1;
2014
2015         bind_list->ps = ps;
2016         bind_list->port = (unsigned short) port;
2017         cma_bind_port(bind_list, id_priv);
2018         return 0;
2019 err2:
2020         idr_remove(ps, port);
2021 err1:
2022         kfree(bind_list);
2023         return ret;
2024 }
2025
2026 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2027 {
2028         struct rdma_id_private *cur_id;
2029         struct sockaddr_in *sin, *cur_sin;
2030         struct rdma_bind_list *bind_list;
2031         struct hlist_node *node;
2032         unsigned short snum;
2033
2034         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2035         snum = ntohs(sin->sin_port);
2036         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2037                 return -EACCES;
2038
2039         bind_list = idr_find(ps, snum);
2040         if (!bind_list)
2041                 return cma_alloc_port(ps, id_priv, snum);
2042
2043         /*
2044          * We don't support binding to any address if anyone is bound to
2045          * a specific address on the same port.
2046          */
2047         if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2048                 return -EADDRNOTAVAIL;
2049
2050         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2051                 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2052                         return -EADDRNOTAVAIL;
2053
2054                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2055                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2056                         return -EADDRINUSE;
2057         }
2058
2059         cma_bind_port(bind_list, id_priv);
2060         return 0;
2061 }
2062
2063 static int cma_get_port(struct rdma_id_private *id_priv)
2064 {
2065         struct idr *ps;
2066         int ret;
2067
2068         switch (id_priv->id.ps) {
2069         case RDMA_PS_SDP:
2070                 ps = &sdp_ps;
2071                 break;
2072         case RDMA_PS_TCP:
2073                 ps = &tcp_ps;
2074                 break;
2075         case RDMA_PS_UDP:
2076                 ps = &udp_ps;
2077                 break;
2078         case RDMA_PS_IPOIB:
2079                 ps = &ipoib_ps;
2080                 break;
2081         default:
2082                 return -EPROTONOSUPPORT;
2083         }
2084
2085         mutex_lock(&lock);
2086         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2087                 ret = cma_alloc_any_port(ps, id_priv);
2088         else
2089                 ret = cma_use_port(ps, id_priv);
2090         mutex_unlock(&lock);
2091
2092         return ret;
2093 }
2094
2095 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2096 {
2097         struct rdma_id_private *id_priv;
2098         int ret;
2099
2100         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2101                 return -EAFNOSUPPORT;
2102
2103         id_priv = container_of(id, struct rdma_id_private, id);
2104         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2105                 return -EINVAL;
2106
2107         if (!cma_any_addr(addr)) {
2108                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2109                 if (ret)
2110                         goto err1;
2111
2112                 mutex_lock(&lock);
2113                 ret = cma_acquire_dev(id_priv);
2114                 mutex_unlock(&lock);
2115                 if (ret)
2116                         goto err1;
2117         }
2118
2119         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2120         ret = cma_get_port(id_priv);
2121         if (ret)
2122                 goto err2;
2123
2124         return 0;
2125 err2:
2126         if (!cma_any_addr(addr)) {
2127                 mutex_lock(&lock);
2128                 cma_detach_from_dev(id_priv);
2129                 mutex_unlock(&lock);
2130         }
2131 err1:
2132         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2133         return ret;
2134 }
2135 EXPORT_SYMBOL(rdma_bind_addr);
2136
2137 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2138                           struct rdma_route *route)
2139 {
2140         struct cma_hdr *cma_hdr;
2141         struct sdp_hh *sdp_hdr;
2142
2143         if (route->addr.src_addr.ss_family == AF_INET) {
2144                 struct sockaddr_in *src4, *dst4;
2145
2146                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2147                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2148
2149                 switch (ps) {
2150                 case RDMA_PS_SDP:
2151                         sdp_hdr = hdr;
2152                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2153                                 return -EINVAL;
2154                         sdp_set_ip_ver(sdp_hdr, 4);
2155                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2156                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2157                         sdp_hdr->port = src4->sin_port;
2158                         break;
2159                 default:
2160                         cma_hdr = hdr;
2161                         cma_hdr->cma_version = CMA_VERSION;
2162                         cma_set_ip_ver(cma_hdr, 4);
2163                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2164                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2165                         cma_hdr->port = src4->sin_port;
2166                         break;
2167                 }
2168         } else {
2169                 struct sockaddr_in6 *src6, *dst6;
2170
2171                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2172                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2173
2174                 switch (ps) {
2175                 case RDMA_PS_SDP:
2176                         sdp_hdr = hdr;
2177                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2178                                 return -EINVAL;
2179                         sdp_set_ip_ver(sdp_hdr, 6);
2180                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2181                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2182                         sdp_hdr->port = src6->sin6_port;
2183                         break;
2184                 default:
2185                         cma_hdr = hdr;
2186                         cma_hdr->cma_version = CMA_VERSION;
2187                         cma_set_ip_ver(cma_hdr, 6);
2188                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2189                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2190                         cma_hdr->port = src6->sin6_port;
2191                         break;
2192                 }
2193         }
2194         return 0;
2195 }
2196
2197 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2198                                 struct ib_cm_event *ib_event)
2199 {
2200         struct rdma_id_private *id_priv = cm_id->context;
2201         struct rdma_cm_event event;
2202         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2203         int ret = 0;
2204
2205         if (cma_disable_callback(id_priv, CMA_CONNECT))
2206                 return 0;
2207
2208         memset(&event, 0, sizeof event);
2209         switch (ib_event->event) {
2210         case IB_CM_SIDR_REQ_ERROR:
2211                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2212                 event.status = -ETIMEDOUT;
2213                 break;
2214         case IB_CM_SIDR_REP_RECEIVED:
2215                 event.param.ud.private_data = ib_event->private_data;
2216                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2217                 if (rep->status != IB_SIDR_SUCCESS) {
2218                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2219                         event.status = ib_event->param.sidr_rep_rcvd.status;
2220                         break;
2221                 }
2222                 ret = cma_set_qkey(id_priv);
2223                 if (ret) {
2224                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2225                         event.status = -EINVAL;
2226                         break;
2227                 }
2228                 if (id_priv->qkey != rep->qkey) {
2229                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2230                         event.status = -EINVAL;
2231                         break;
2232                 }
2233                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2234                                      id_priv->id.route.path_rec,
2235                                      &event.param.ud.ah_attr);
2236                 event.param.ud.qp_num = rep->qpn;
2237                 event.param.ud.qkey = rep->qkey;
2238                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2239                 event.status = 0;
2240                 break;
2241         default:
2242                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2243                        ib_event->event);
2244                 goto out;
2245         }
2246
2247         ret = id_priv->id.event_handler(&id_priv->id, &event);
2248         if (ret) {
2249                 /* Destroy the CM ID by returning a non-zero value. */
2250                 id_priv->cm_id.ib = NULL;
2251                 cma_exch(id_priv, CMA_DESTROYING);
2252                 mutex_unlock(&id_priv->handler_mutex);
2253                 rdma_destroy_id(&id_priv->id);
2254                 return ret;
2255         }
2256 out:
2257         mutex_unlock(&id_priv->handler_mutex);
2258         return ret;
2259 }
2260
2261 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2262                               struct rdma_conn_param *conn_param)
2263 {
2264         struct ib_cm_sidr_req_param req;
2265         struct rdma_route *route;
2266         int ret;
2267
2268         req.private_data_len = sizeof(struct cma_hdr) +
2269                                conn_param->private_data_len;
2270         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2271         if (!req.private_data)
2272                 return -ENOMEM;
2273
2274         if (conn_param->private_data && conn_param->private_data_len)
2275                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2276                        conn_param->private_data, conn_param->private_data_len);
2277
2278         route = &id_priv->id.route;
2279         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2280         if (ret)
2281                 goto out;
2282
2283         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2284                                             cma_sidr_rep_handler, id_priv);
2285         if (IS_ERR(id_priv->cm_id.ib)) {
2286                 ret = PTR_ERR(id_priv->cm_id.ib);
2287                 goto out;
2288         }
2289
2290         req.path = route->path_rec;
2291         req.service_id = cma_get_service_id(id_priv->id.ps,
2292                                             (struct sockaddr *) &route->addr.dst_addr);
2293         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2294         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2295
2296         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2297         if (ret) {
2298                 ib_destroy_cm_id(id_priv->cm_id.ib);
2299                 id_priv->cm_id.ib = NULL;
2300         }
2301 out:
2302         kfree(req.private_data);
2303         return ret;
2304 }
2305
2306 static int cma_connect_ib(struct rdma_id_private *id_priv,
2307                           struct rdma_conn_param *conn_param)
2308 {
2309         struct ib_cm_req_param req;
2310         struct rdma_route *route;
2311         void *private_data;
2312         int offset, ret;
2313
2314         memset(&req, 0, sizeof req);
2315         offset = cma_user_data_offset(id_priv->id.ps);
2316         req.private_data_len = offset + conn_param->private_data_len;
2317         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2318         if (!private_data)
2319                 return -ENOMEM;
2320
2321         if (conn_param->private_data && conn_param->private_data_len)
2322                 memcpy(private_data + offset, conn_param->private_data,
2323                        conn_param->private_data_len);
2324
2325         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2326                                             id_priv);
2327         if (IS_ERR(id_priv->cm_id.ib)) {
2328                 ret = PTR_ERR(id_priv->cm_id.ib);
2329                 goto out;
2330         }
2331
2332         route = &id_priv->id.route;
2333         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2334         if (ret)
2335                 goto out;
2336         req.private_data = private_data;
2337
2338         req.primary_path = &route->path_rec[0];
2339         if (route->num_paths == 2)
2340                 req.alternate_path = &route->path_rec[1];
2341
2342         req.service_id = cma_get_service_id(id_priv->id.ps,
2343                                             (struct sockaddr *) &route->addr.dst_addr);
2344         req.qp_num = id_priv->qp_num;
2345         req.qp_type = IB_QPT_RC;
2346         req.starting_psn = id_priv->seq_num;
2347         req.responder_resources = conn_param->responder_resources;
2348         req.initiator_depth = conn_param->initiator_depth;
2349         req.flow_control = conn_param->flow_control;
2350         req.retry_count = conn_param->retry_count;
2351         req.rnr_retry_count = conn_param->rnr_retry_count;
2352         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2353         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2354         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2355         req.srq = id_priv->srq ? 1 : 0;
2356
2357         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2358 out:
2359         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2360                 ib_destroy_cm_id(id_priv->cm_id.ib);
2361                 id_priv->cm_id.ib = NULL;
2362         }
2363
2364         kfree(private_data);
2365         return ret;
2366 }
2367
2368 static int cma_connect_iw(struct rdma_id_private *id_priv,
2369                           struct rdma_conn_param *conn_param)
2370 {
2371         struct iw_cm_id *cm_id;
2372         struct sockaddr_in* sin;
2373         int ret;
2374         struct iw_cm_conn_param iw_param;
2375
2376         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2377         if (IS_ERR(cm_id)) {
2378                 ret = PTR_ERR(cm_id);
2379                 goto out;
2380         }
2381
2382         id_priv->cm_id.iw = cm_id;
2383
2384         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2385         cm_id->local_addr = *sin;
2386
2387         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2388         cm_id->remote_addr = *sin;
2389
2390         ret = cma_modify_qp_rtr(id_priv, conn_param);
2391         if (ret)
2392                 goto out;
2393
2394         iw_param.ord = conn_param->initiator_depth;
2395         iw_param.ird = conn_param->responder_resources;
2396         iw_param.private_data = conn_param->private_data;
2397         iw_param.private_data_len = conn_param->private_data_len;
2398         if (id_priv->id.qp)
2399                 iw_param.qpn = id_priv->qp_num;
2400         else
2401                 iw_param.qpn = conn_param->qp_num;
2402         ret = iw_cm_connect(cm_id, &iw_param);
2403 out:
2404         if (ret && !IS_ERR(cm_id)) {
2405                 iw_destroy_cm_id(cm_id);
2406                 id_priv->cm_id.iw = NULL;
2407         }
2408         return ret;
2409 }
2410
2411 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2412 {
2413         struct rdma_id_private *id_priv;
2414         int ret;
2415
2416         id_priv = container_of(id, struct rdma_id_private, id);
2417         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2418                 return -EINVAL;
2419
2420         if (!id->qp) {
2421                 id_priv->qp_num = conn_param->qp_num;
2422                 id_priv->srq = conn_param->srq;
2423         }
2424
2425         switch (rdma_node_get_transport(id->device->node_type)) {
2426         case RDMA_TRANSPORT_IB:
2427                 if (cma_is_ud_ps(id->ps))
2428                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2429                 else
2430                         ret = cma_connect_ib(id_priv, conn_param);
2431                 break;
2432         case RDMA_TRANSPORT_IWARP:
2433                 ret = cma_connect_iw(id_priv, conn_param);
2434                 break;
2435         default:
2436                 ret = -ENOSYS;
2437                 break;
2438         }
2439         if (ret)
2440                 goto err;
2441
2442         return 0;
2443 err:
2444         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2445         return ret;
2446 }
2447 EXPORT_SYMBOL(rdma_connect);
2448
2449 static int cma_accept_ib(struct rdma_id_private *id_priv,
2450                          struct rdma_conn_param *conn_param)
2451 {
2452         struct ib_cm_rep_param rep;
2453         int ret;
2454
2455         ret = cma_modify_qp_rtr(id_priv, conn_param);
2456         if (ret)
2457                 goto out;
2458
2459         ret = cma_modify_qp_rts(id_priv, conn_param);
2460         if (ret)
2461                 goto out;
2462
2463         memset(&rep, 0, sizeof rep);
2464         rep.qp_num = id_priv->qp_num;
2465         rep.starting_psn = id_priv->seq_num;
2466         rep.private_data = conn_param->private_data;
2467         rep.private_data_len = conn_param->private_data_len;
2468         rep.responder_resources = conn_param->responder_resources;
2469         rep.initiator_depth = conn_param->initiator_depth;
2470         rep.failover_accepted = 0;
2471         rep.flow_control = conn_param->flow_control;
2472         rep.rnr_retry_count = conn_param->rnr_retry_count;
2473         rep.srq = id_priv->srq ? 1 : 0;
2474
2475         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2476 out:
2477         return ret;
2478 }
2479
2480 static int cma_accept_iw(struct rdma_id_private *id_priv,
2481                   struct rdma_conn_param *conn_param)
2482 {
2483         struct iw_cm_conn_param iw_param;
2484         int ret;
2485
2486         ret = cma_modify_qp_rtr(id_priv, conn_param);
2487         if (ret)
2488                 return ret;
2489
2490         iw_param.ord = conn_param->initiator_depth;
2491         iw_param.ird = conn_param->responder_resources;
2492         iw_param.private_data = conn_param->private_data;
2493         iw_param.private_data_len = conn_param->private_data_len;
2494         if (id_priv->id.qp) {
2495                 iw_param.qpn = id_priv->qp_num;
2496         } else
2497                 iw_param.qpn = conn_param->qp_num;
2498
2499         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2500 }
2501
2502 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2503                              enum ib_cm_sidr_status status,
2504                              const void *private_data, int private_data_len)
2505 {
2506         struct ib_cm_sidr_rep_param rep;
2507         int ret;
2508
2509         memset(&rep, 0, sizeof rep);
2510         rep.status = status;
2511         if (status == IB_SIDR_SUCCESS) {
2512                 ret = cma_set_qkey(id_priv);
2513                 if (ret)
2514                         return ret;
2515                 rep.qp_num = id_priv->qp_num;
2516                 rep.qkey = id_priv->qkey;
2517         }
2518         rep.private_data = private_data;
2519         rep.private_data_len = private_data_len;
2520
2521         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2522 }
2523
2524 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2525 {
2526         struct rdma_id_private *id_priv;
2527         int ret;
2528
2529         id_priv = container_of(id, struct rdma_id_private, id);
2530         if (!cma_comp(id_priv, CMA_CONNECT))
2531                 return -EINVAL;
2532
2533         if (!id->qp && conn_param) {
2534                 id_priv->qp_num = conn_param->qp_num;
2535                 id_priv->srq = conn_param->srq;
2536         }
2537
2538         switch (rdma_node_get_transport(id->device->node_type)) {
2539         case RDMA_TRANSPORT_IB:
2540                 if (cma_is_ud_ps(id->ps))
2541                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2542                                                 conn_param->private_data,
2543                                                 conn_param->private_data_len);
2544                 else if (conn_param)
2545                         ret = cma_accept_ib(id_priv, conn_param);
2546                 else
2547                         ret = cma_rep_recv(id_priv);
2548                 break;
2549         case RDMA_TRANSPORT_IWARP:
2550                 ret = cma_accept_iw(id_priv, conn_param);
2551                 break;
2552         default:
2553                 ret = -ENOSYS;
2554                 break;
2555         }
2556
2557         if (ret)
2558                 goto reject;
2559
2560         return 0;
2561 reject:
2562         cma_modify_qp_err(id_priv);
2563         rdma_reject(id, NULL, 0);
2564         return ret;
2565 }
2566 EXPORT_SYMBOL(rdma_accept);
2567
2568 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2569 {
2570         struct rdma_id_private *id_priv;
2571         int ret;
2572
2573         id_priv = container_of(id, struct rdma_id_private, id);
2574         if (!cma_has_cm_dev(id_priv))
2575                 return -EINVAL;
2576
2577         switch (id->device->node_type) {
2578         case RDMA_NODE_IB_CA:
2579                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2580                 break;
2581         default:
2582                 ret = 0;
2583                 break;
2584         }
2585         return ret;
2586 }
2587 EXPORT_SYMBOL(rdma_notify);
2588
2589 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2590                 u8 private_data_len)
2591 {
2592         struct rdma_id_private *id_priv;
2593         int ret;
2594
2595         id_priv = container_of(id, struct rdma_id_private, id);
2596         if (!cma_has_cm_dev(id_priv))
2597                 return -EINVAL;
2598
2599         switch (rdma_node_get_transport(id->device->node_type)) {
2600         case RDMA_TRANSPORT_IB:
2601                 if (cma_is_ud_ps(id->ps))
2602                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2603                                                 private_data, private_data_len);
2604                 else
2605                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2606                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2607                                              0, private_data, private_data_len);
2608                 break;
2609         case RDMA_TRANSPORT_IWARP:
2610                 ret = iw_cm_reject(id_priv->cm_id.iw,
2611                                    private_data, private_data_len);
2612                 break;
2613         default:
2614                 ret = -ENOSYS;
2615                 break;
2616         }
2617         return ret;
2618 }
2619 EXPORT_SYMBOL(rdma_reject);
2620
2621 int rdma_disconnect(struct rdma_cm_id *id)
2622 {
2623         struct rdma_id_private *id_priv;
2624         int ret;
2625
2626         id_priv = container_of(id, struct rdma_id_private, id);
2627         if (!cma_has_cm_dev(id_priv))
2628                 return -EINVAL;
2629
2630         switch (rdma_node_get_transport(id->device->node_type)) {
2631         case RDMA_TRANSPORT_IB:
2632                 ret = cma_modify_qp_err(id_priv);
2633                 if (ret)
2634                         goto out;
2635                 /* Initiate or respond to a disconnect. */
2636                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2637                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2638                 break;
2639         case RDMA_TRANSPORT_IWARP:
2640                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2641                 break;
2642         default:
2643                 ret = -EINVAL;
2644                 break;
2645         }
2646 out:
2647         return ret;
2648 }
2649 EXPORT_SYMBOL(rdma_disconnect);
2650
2651 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2652 {
2653         struct rdma_id_private *id_priv;
2654         struct cma_multicast *mc = multicast->context;
2655         struct rdma_cm_event event;
2656         int ret;
2657
2658         id_priv = mc->id_priv;
2659         if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2660             cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2661                 return 0;
2662
2663         mutex_lock(&id_priv->qp_mutex);
2664         if (!status && id_priv->id.qp)
2665                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2666                                          multicast->rec.mlid);
2667         mutex_unlock(&id_priv->qp_mutex);
2668
2669         memset(&event, 0, sizeof event);
2670         event.status = status;
2671         event.param.ud.private_data = mc->context;
2672         if (!status) {
2673                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2674                 ib_init_ah_from_mcmember(id_priv->id.device,
2675                                          id_priv->id.port_num, &multicast->rec,
2676                                          &event.param.ud.ah_attr);
2677                 event.param.ud.qp_num = 0xFFFFFF;
2678                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2679         } else
2680                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2681
2682         ret = id_priv->id.event_handler(&id_priv->id, &event);
2683         if (ret) {
2684                 cma_exch(id_priv, CMA_DESTROYING);
2685                 mutex_unlock(&id_priv->handler_mutex);
2686                 rdma_destroy_id(&id_priv->id);
2687                 return 0;
2688         }
2689
2690         mutex_unlock(&id_priv->handler_mutex);
2691         return 0;
2692 }
2693
2694 static void cma_set_mgid(struct rdma_id_private *id_priv,
2695                          struct sockaddr *addr, union ib_gid *mgid)
2696 {
2697         unsigned char mc_map[MAX_ADDR_LEN];
2698         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2699         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2700         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2701
2702         if (cma_any_addr(addr)) {
2703                 memset(mgid, 0, sizeof *mgid);
2704         } else if ((addr->sa_family == AF_INET6) &&
2705                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) ==
2706                                                                  0xFF10A01B)) {
2707                 /* IPv6 address is an SA assigned MGID. */
2708                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2709         } else {
2710                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2711                 if (id_priv->id.ps == RDMA_PS_UDP)
2712                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2713                 *mgid = *(union ib_gid *) (mc_map + 4);
2714         }
2715 }
2716
2717 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2718                                  struct cma_multicast *mc)
2719 {
2720         struct ib_sa_mcmember_rec rec;
2721         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2722         ib_sa_comp_mask comp_mask;
2723         int ret;
2724
2725         ib_addr_get_mgid(dev_addr, &rec.mgid);
2726         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2727                                      &rec.mgid, &rec);
2728         if (ret)
2729                 return ret;
2730
2731         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2732         if (id_priv->id.ps == RDMA_PS_UDP)
2733                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2734         ib_addr_get_sgid(dev_addr, &rec.port_gid);
2735         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2736         rec.join_state = 1;
2737
2738         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2739                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2740                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2741                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2742                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2743
2744         if (id_priv->id.ps == RDMA_PS_IPOIB)
2745                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2746                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2747
2748         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2749                                                 id_priv->id.port_num, &rec,
2750                                                 comp_mask, GFP_KERNEL,
2751                                                 cma_ib_mc_handler, mc);
2752         if (IS_ERR(mc->multicast.ib))
2753                 return PTR_ERR(mc->multicast.ib);
2754
2755         return 0;
2756 }
2757
2758 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2759                         void *context)
2760 {
2761         struct rdma_id_private *id_priv;
2762         struct cma_multicast *mc;
2763         int ret;
2764
2765         id_priv = container_of(id, struct rdma_id_private, id);
2766         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2767             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2768                 return -EINVAL;
2769
2770         mc = kmalloc(sizeof *mc, GFP_KERNEL);
2771         if (!mc)
2772                 return -ENOMEM;
2773
2774         memcpy(&mc->addr, addr, ip_addr_size(addr));
2775         mc->context = context;
2776         mc->id_priv = id_priv;
2777
2778         spin_lock(&id_priv->lock);
2779         list_add(&mc->list, &id_priv->mc_list);
2780         spin_unlock(&id_priv->lock);
2781
2782         switch (rdma_node_get_transport(id->device->node_type)) {
2783         case RDMA_TRANSPORT_IB:
2784                 ret = cma_join_ib_multicast(id_priv, mc);
2785                 break;
2786         default:
2787                 ret = -ENOSYS;
2788                 break;
2789         }
2790
2791         if (ret) {
2792                 spin_lock_irq(&id_priv->lock);
2793                 list_del(&mc->list);
2794                 spin_unlock_irq(&id_priv->lock);
2795                 kfree(mc);
2796         }
2797         return ret;
2798 }
2799 EXPORT_SYMBOL(rdma_join_multicast);
2800
2801 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2802 {
2803         struct rdma_id_private *id_priv;
2804         struct cma_multicast *mc;
2805
2806         id_priv = container_of(id, struct rdma_id_private, id);
2807         spin_lock_irq(&id_priv->lock);
2808         list_for_each_entry(mc, &id_priv->mc_list, list) {
2809                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2810                         list_del(&mc->list);
2811                         spin_unlock_irq(&id_priv->lock);
2812
2813                         if (id->qp)
2814                                 ib_detach_mcast(id->qp,
2815                                                 &mc->multicast.ib->rec.mgid,
2816                                                 mc->multicast.ib->rec.mlid);
2817                         ib_sa_free_multicast(mc->multicast.ib);
2818                         kfree(mc);
2819                         return;
2820                 }
2821         }
2822         spin_unlock_irq(&id_priv->lock);
2823 }
2824 EXPORT_SYMBOL(rdma_leave_multicast);
2825
2826 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
2827 {
2828         struct rdma_dev_addr *dev_addr;
2829         struct cma_ndev_work *work;
2830
2831         dev_addr = &id_priv->id.route.addr.dev_addr;
2832
2833         if ((dev_addr->src_dev == ndev) &&
2834             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
2835                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
2836                        ndev->name, &id_priv->id);
2837                 work = kzalloc(sizeof *work, GFP_KERNEL);
2838                 if (!work)
2839                         return -ENOMEM;
2840
2841                 INIT_WORK(&work->work, cma_ndev_work_handler);
2842                 work->id = id_priv;
2843                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
2844                 atomic_inc(&id_priv->refcount);
2845                 queue_work(cma_wq, &work->work);
2846         }
2847
2848         return 0;
2849 }
2850
2851 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
2852                                void *ctx)
2853 {
2854         struct net_device *ndev = (struct net_device *)ctx;
2855         struct cma_device *cma_dev;
2856         struct rdma_id_private *id_priv;
2857         int ret = NOTIFY_DONE;
2858
2859         if (dev_net(ndev) != &init_net)
2860                 return NOTIFY_DONE;
2861
2862         if (event != NETDEV_BONDING_FAILOVER)
2863                 return NOTIFY_DONE;
2864
2865         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
2866                 return NOTIFY_DONE;
2867
2868         mutex_lock(&lock);
2869         list_for_each_entry(cma_dev, &dev_list, list)
2870                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
2871                         ret = cma_netdev_change(ndev, id_priv);
2872                         if (ret)
2873                                 goto out;
2874                 }
2875
2876 out:
2877         mutex_unlock(&lock);
2878         return ret;
2879 }
2880
2881 static struct notifier_block cma_nb = {
2882         .notifier_call = cma_netdev_callback
2883 };
2884
2885 static void cma_add_one(struct ib_device *device)
2886 {
2887         struct cma_device *cma_dev;
2888         struct rdma_id_private *id_priv;
2889
2890         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2891         if (!cma_dev)
2892                 return;
2893
2894         cma_dev->device = device;
2895
2896         init_completion(&cma_dev->comp);
2897         atomic_set(&cma_dev->refcount, 1);
2898         INIT_LIST_HEAD(&cma_dev->id_list);
2899         ib_set_client_data(device, &cma_client, cma_dev);
2900
2901         mutex_lock(&lock);
2902         list_add_tail(&cma_dev->list, &dev_list);
2903         list_for_each_entry(id_priv, &listen_any_list, list)
2904                 cma_listen_on_dev(id_priv, cma_dev);
2905         mutex_unlock(&lock);
2906 }
2907
2908 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2909 {
2910         struct rdma_cm_event event;
2911         enum cma_state state;
2912         int ret = 0;
2913
2914         /* Record that we want to remove the device */
2915         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2916         if (state == CMA_DESTROYING)
2917                 return 0;
2918
2919         cma_cancel_operation(id_priv, state);
2920         mutex_lock(&id_priv->handler_mutex);
2921
2922         /* Check for destruction from another callback. */
2923         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2924                 goto out;
2925
2926         memset(&event, 0, sizeof event);
2927         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2928         ret = id_priv->id.event_handler(&id_priv->id, &event);
2929 out:
2930         mutex_unlock(&id_priv->handler_mutex);
2931         return ret;
2932 }
2933
2934 static void cma_process_remove(struct cma_device *cma_dev)
2935 {
2936         struct rdma_id_private *id_priv;
2937         int ret;
2938
2939         mutex_lock(&lock);
2940         while (!list_empty(&cma_dev->id_list)) {
2941                 id_priv = list_entry(cma_dev->id_list.next,
2942                                      struct rdma_id_private, list);
2943
2944                 list_del(&id_priv->listen_list);
2945                 list_del_init(&id_priv->list);
2946                 atomic_inc(&id_priv->refcount);
2947                 mutex_unlock(&lock);
2948
2949                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
2950                 cma_deref_id(id_priv);
2951                 if (ret)
2952                         rdma_destroy_id(&id_priv->id);
2953
2954                 mutex_lock(&lock);
2955         }
2956         mutex_unlock(&lock);
2957
2958         cma_deref_dev(cma_dev);
2959         wait_for_completion(&cma_dev->comp);
2960 }
2961
2962 static void cma_remove_one(struct ib_device *device)
2963 {
2964         struct cma_device *cma_dev;
2965
2966         cma_dev = ib_get_client_data(device, &cma_client);
2967         if (!cma_dev)
2968                 return;
2969
2970         mutex_lock(&lock);
2971         list_del(&cma_dev->list);
2972         mutex_unlock(&lock);
2973
2974         cma_process_remove(cma_dev);
2975         kfree(cma_dev);
2976 }
2977
2978 static int __init cma_init(void)
2979 {
2980         int ret, low, high, remaining;
2981
2982         get_random_bytes(&next_port, sizeof next_port);
2983         inet_get_local_port_range(&low, &high);
2984         remaining = (high - low) + 1;
2985         next_port = ((unsigned int) next_port % remaining) + low;
2986
2987         cma_wq = create_singlethread_workqueue("rdma_cm");
2988         if (!cma_wq)
2989                 return -ENOMEM;
2990
2991         ib_sa_register_client(&sa_client);
2992         rdma_addr_register_client(&addr_client);
2993         register_netdevice_notifier(&cma_nb);
2994
2995         ret = ib_register_client(&cma_client);
2996         if (ret)
2997                 goto err;
2998         return 0;
2999
3000 err:
3001         unregister_netdevice_notifier(&cma_nb);
3002         rdma_addr_unregister_client(&addr_client);
3003         ib_sa_unregister_client(&sa_client);
3004         destroy_workqueue(cma_wq);
3005         return ret;
3006 }
3007
3008 static void __exit cma_cleanup(void)
3009 {
3010         ib_unregister_client(&cma_client);
3011         unregister_netdevice_notifier(&cma_nb);
3012         rdma_addr_unregister_client(&addr_client);
3013         ib_sa_unregister_client(&sa_client);
3014         destroy_workqueue(cma_wq);
3015         idr_destroy(&sdp_ps);
3016         idr_destroy(&tcp_ps);
3017         idr_destroy(&udp_ps);
3018         idr_destroy(&ipoib_ps);
3019 }
3020
3021 module_init(cma_init);
3022 module_exit(cma_cleanup);