sys/ofed/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 static int tavor_quirk = 0;
  59 module_param_named(tavor_quirk, tavor_quirk, int, 0644);
  60 MODULE_PARM_DESC(tavor_quirk, "Tavor performance quirk: limit MTU to 1K if > 0");
  61
  62 int unify_tcp_port_space = 1;
  63 module_param(unify_tcp_port_space, int, 0644);
  64 MODULE_PARM_DESC(unify_tcp_port_space, "Unify the host TCP and RDMA port "
  65                  "space allocation (default=1)");
  66
  67 #define CMA_CM_RESPONSE_TIMEOUT 20
  68 #define CMA_MAX_CM_RETRIES 15
  69 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  70 #define IBOE_PACKET_LIFETIME 18
  71
  72 static int cma_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
  73 module_param_named(cma_response_timeout, cma_response_timeout, int, 0644);
  74 MODULE_PARM_DESC(cma_response_timeout, "CMA_CM_RESPONSE_TIMEOUT default=20");
  75
  76 static int def_prec2sl = 3;
  77 module_param_named(def_prec2sl, def_prec2sl, int, 0644);
  78 MODULE_PARM_DESC(def_prec2sl, "Default value for SL priority with RoCE. Valid values 0 - 7");
  79
  80 static void cma_add_one(struct ib_device *device);
  81 static void cma_remove_one(struct ib_device *device);
  82
  83 static struct ib_client cma_client = {
  84         .name   = "cma",
  85         .add    = cma_add_one,
  86         .remove = cma_remove_one
  87 };
  88
  89 static struct ib_sa_client sa_client;
  90 static struct rdma_addr_client addr_client;
  91 static LIST_HEAD(dev_list);
  92 static LIST_HEAD(listen_any_list);
  93 static DEFINE_MUTEX(lock);
  94 static struct workqueue_struct *cma_wq;
  95 static DEFINE_IDR(sdp_ps);
  96 static DEFINE_IDR(tcp_ps);
  97 static DEFINE_IDR(udp_ps);
  98 static DEFINE_IDR(ipoib_ps);
  99 #if defined(INET)
 100 static int next_port;
 101 #endif
 102
 103 struct cma_device {
 104         struct list_head        list;
 105         struct ib_device        *device;
 106         struct completion       comp;
 107         atomic_t                refcount;
 108         struct list_head        id_list;
 109 };
 110
 111 enum cma_state {
 112         CMA_IDLE,
 113         CMA_ADDR_QUERY,
 114         CMA_ADDR_RESOLVED,
 115         CMA_ROUTE_QUERY,
 116         CMA_ROUTE_RESOLVED,
 117         CMA_CONNECT,
 118         CMA_DISCONNECT,
 119         CMA_ADDR_BOUND,
 120         CMA_LISTEN,
 121         CMA_DEVICE_REMOVAL,
 122         CMA_DESTROYING
 123 };
 124
 125 struct rdma_bind_list {
 126         struct idr              *ps;
 127         struct hlist_head       owners;
 128         unsigned short          port;
 129 };
 130
 131 /*
 132  * Device removal can occur at anytime, so we need extra handling to
 133  * serialize notifying the user of device removal with other callbacks.
 134  * We do this by disabling removal notification while a callback is in process,
 135  * and reporting it after the callback completes.
 136  */
 137 struct rdma_id_private {
 138         struct rdma_cm_id       id;
 139
 140         struct rdma_bind_list   *bind_list;
 141         struct socket           *sock;
 142         struct hlist_node       node;
 143         struct list_head        list; /* listen_any_list or cma_device.list */
 144         struct list_head        listen_list; /* per device listens */
 145         struct cma_device       *cma_dev;
 146         struct list_head        mc_list;
 147
 148         int                     internal_id;
 149         enum cma_state          state;
 150         spinlock_t              lock;
 151         struct mutex            qp_mutex;
 152
 153         struct completion       comp;
 154         atomic_t                refcount;
 155         struct mutex            handler_mutex;
 156
 157         int                     backlog;
 158         int                     timeout_ms;
 159         struct ib_sa_query      *query;
 160         int                     query_id;
 161         union {
 162                 struct ib_cm_id *ib;
 163                 struct iw_cm_id *iw;
 164         } cm_id;
 165
 166         u32                     seq_num;
 167         u32                     qkey;
 168         u32                     qp_num;
 169         u8                      srq;
 170         u8                      tos;
 171 };
 172
 173 struct cma_multicast {
 174         struct rdma_id_private *id_priv;
 175         union {
 176                 struct ib_sa_multicast *ib;
 177         } multicast;
 178         struct list_head        list;
 179         void                    *context;
 180         struct sockaddr_storage addr;
 181         struct kref             mcref;
 182 };
 183
 184 struct cma_work {
 185         struct work_struct      work;
 186         struct rdma_id_private  *id;
 187         enum cma_state          old_state;
 188         enum cma_state          new_state;
 189         struct rdma_cm_event    event;
 190 };
 191
 192 struct cma_ndev_work {
 193         struct work_struct      work;
 194         struct rdma_id_private  *id;
 195         struct rdma_cm_event    event;
 196 };
 197
 198 struct iboe_mcast_work {
 199         struct work_struct       work;
 200         struct rdma_id_private  *id;
 201         struct cma_multicast    *mc;
 202 };
 203
 204 union cma_ip_addr {
 205         struct in6_addr ip6;
 206         struct {
 207                 __be32 pad[3];
 208                 __be32 addr;
 209         } ip4;
 210 };
 211
 212 struct cma_hdr {
 213         u8 cma_version;
 214         u8 ip_version;  /* IP version: 7:4 */
 215         __be16 port;
 216         union cma_ip_addr src_addr;
 217         union cma_ip_addr dst_addr;
 218 };
 219
 220 struct sdp_hh {
 221         u8 bsdh[16];
 222         u8 sdp_version; /* Major version: 7:4 */
 223         u8 ip_version;  /* IP version: 7:4 */
 224         u8 sdp_specific1[10];
 225         __be16 port;
 226         __be16 sdp_specific2;
 227         union cma_ip_addr src_addr;
 228         union cma_ip_addr dst_addr;
 229 };
 230
 231 struct sdp_hah {
 232         u8 bsdh[16];
 233         u8 sdp_version;
 234 };
 235
 236 #define CMA_VERSION 0x00
 237 #define SDP_MAJ_VERSION 0x2
 238
 239 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 240 {
 241         unsigned long flags;
 242         int ret;
 243
 244         spin_lock_irqsave(&id_priv->lock, flags);
 245         ret = (id_priv->state == comp);
 246         spin_unlock_irqrestore(&id_priv->lock, flags);
 247         return ret;
 248 }
 249
 250 static int cma_comp_exch(struct rdma_id_private *id_priv,
 251                          enum cma_state comp, enum cma_state exch)
 252 {
 253         unsigned long flags;
 254         int ret;
 255
 256         spin_lock_irqsave(&id_priv->lock, flags);
 257         if ((ret = (id_priv->state == comp)))
 258                 id_priv->state = exch;
 259         spin_unlock_irqrestore(&id_priv->lock, flags);
 260         return ret;
 261 }
 262
 263 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 264                                enum cma_state exch)
 265 {
 266         unsigned long flags;
 267         enum cma_state old;
 268
 269         spin_lock_irqsave(&id_priv->lock, flags);
 270         old = id_priv->state;
 271         id_priv->state = exch;
 272         spin_unlock_irqrestore(&id_priv->lock, flags);
 273         return old;
 274 }
 275
 276 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 277 {
 278         return hdr->ip_version >> 4;
 279 }
 280
 281 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 282 {
 283         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 284 }
 285
 286 static inline u8 sdp_get_majv(u8 sdp_version)
 287 {
 288         return sdp_version >> 4;
 289 }
 290
 291 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 292 {
 293         return hh->ip_version >> 4;
 294 }
 295
 296 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 297 {
 298         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 299 }
 300
 301 static inline int cma_is_ud_ps(enum rdma_port_space ps)
 302 {
 303         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
 304 }
 305
 306 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 307                               struct cma_device *cma_dev)
 308 {
 309         atomic_inc(&cma_dev->refcount);
 310         id_priv->cma_dev = cma_dev;
 311         id_priv->id.device = cma_dev->device;
 312         id_priv->id.route.addr.dev_addr.transport =
 313                 rdma_node_get_transport(cma_dev->device->node_type);
 314         list_add_tail(&id_priv->list, &cma_dev->id_list);
 315 }
 316
 317 static inline void cma_deref_dev(struct cma_device *cma_dev)
 318 {
 319         if (atomic_dec_and_test(&cma_dev->refcount))
 320                 complete(&cma_dev->comp);
 321 }
 322
 323 static inline void release_mc(struct kref *kref)
 324 {
 325         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 326
 327         kfree(mc->multicast.ib);
 328         kfree(mc);
 329 }
 330
 331 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 332 {
 333         list_del(&id_priv->list);
 334         cma_deref_dev(id_priv->cma_dev);
 335         id_priv->cma_dev = NULL;
 336 }
 337
 338 static int cma_set_qkey(struct rdma_id_private *id_priv)
 339 {
 340         struct ib_sa_mcmember_rec rec;
 341         int ret = 0;
 342
 343         if (id_priv->qkey)
 344                 return 0;
 345
 346         switch (id_priv->id.ps) {
 347         case RDMA_PS_UDP:
 348                 id_priv->qkey = RDMA_UDP_QKEY;
 349                 break;
 350         case RDMA_PS_IPOIB:
 351                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 352                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 353                                              id_priv->id.port_num, &rec.mgid,
 354                                              &rec);
 355                 if (!ret)
 356                         id_priv->qkey = be32_to_cpu(rec.qkey);
 357                 break;
 358         default:
 359                 break;
 360         }
 361         return ret;
 362 }
 363
 364 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 365 {
 366         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 367         struct cma_device *cma_dev;
 368         union ib_gid gid;
 369         int ret = -ENODEV;
 370
 371         if (dev_addr->dev_type != ARPHRD_INFINIBAND) {
 372                 iboe_addr_get_sgid(dev_addr, &gid);
 373                 list_for_each_entry(cma_dev, &dev_list, list) {
 374                         ret = ib_find_cached_gid(cma_dev->device, &gid,
 375                                                  &id_priv->id.port_num, NULL);
 376                         if (!ret)
 377                                 goto out;
 378                 }
 379         }
 380
 381         memcpy(&gid, dev_addr->src_dev_addr +
 382                rdma_addr_gid_offset(dev_addr), sizeof gid);
 383         list_for_each_entry(cma_dev, &dev_list, list) {
 384                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 385                                          &id_priv->id.port_num, NULL);
 386                 if (!ret)
 387                         break;
 388         }
 389
 390 out:
 391         if (!ret)
 392                 cma_attach_to_dev(id_priv, cma_dev);
 393
 394         return ret;
 395 }
 396
 397 static void cma_deref_id(struct rdma_id_private *id_priv)
 398 {
 399         if (atomic_dec_and_test(&id_priv->refcount))
 400                 complete(&id_priv->comp);
 401 }
 402
 403 static int cma_disable_callback(struct rdma_id_private *id_priv,
 404                               enum cma_state state)
 405 {
 406         mutex_lock(&id_priv->handler_mutex);
 407         if (id_priv->state != state) {
 408                 mutex_unlock(&id_priv->handler_mutex);
 409                 return -EINVAL;
 410         }
 411         return 0;
 412 }
 413
 414 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
 415 {
 416         return (id_priv->id.device && id_priv->cm_id.ib);
 417 }
 418
 419 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 420                                   void *context, enum rdma_port_space ps)
 421 {
 422         struct rdma_id_private *id_priv;
 423
 424         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 425         if (!id_priv)
 426                 return ERR_PTR(-ENOMEM);
 427
 428         id_priv->state = CMA_IDLE;
 429         id_priv->id.context = context;
 430         id_priv->id.event_handler = event_handler;
 431         id_priv->id.ps = ps;
 432         spin_lock_init(&id_priv->lock);
 433         mutex_init(&id_priv->qp_mutex);
 434         init_completion(&id_priv->comp);
 435         atomic_set(&id_priv->refcount, 1);
 436         mutex_init(&id_priv->handler_mutex);
 437         INIT_LIST_HEAD(&id_priv->listen_list);
 438         INIT_LIST_HEAD(&id_priv->mc_list);
 439         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 440
 441         return &id_priv->id;
 442 }
 443 EXPORT_SYMBOL(rdma_create_id);
 444
 445 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 446 {
 447         struct ib_qp_attr qp_attr;
 448         int qp_attr_mask, ret;
 449
 450         qp_attr.qp_state = IB_QPS_INIT;
 451         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 452         if (ret)
 453                 return ret;
 454
 455         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 456         if (ret)
 457                 return ret;
 458
 459         qp_attr.qp_state = IB_QPS_RTR;
 460         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 461         if (ret)
 462                 return ret;
 463
 464         qp_attr.qp_state = IB_QPS_RTS;
 465         qp_attr.sq_psn = 0;
 466         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 467
 468         return ret;
 469 }
 470
 471 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 472 {
 473         struct ib_qp_attr qp_attr;
 474         int qp_attr_mask, ret;
 475
 476         qp_attr.qp_state = IB_QPS_INIT;
 477         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 478         if (ret)
 479                 return ret;
 480
 481         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 482 }
 483
 484 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 485                    struct ib_qp_init_attr *qp_init_attr)
 486 {
 487         struct rdma_id_private *id_priv;
 488         struct ib_qp *qp;
 489         int ret;
 490
 491         id_priv = container_of(id, struct rdma_id_private, id);
 492         if (id->device != pd->device)
 493                 return -EINVAL;
 494
 495         qp = ib_create_qp(pd, qp_init_attr);
 496         if (IS_ERR(qp))
 497                 return PTR_ERR(qp);
 498
 499         if (cma_is_ud_ps(id_priv->id.ps))
 500                 ret = cma_init_ud_qp(id_priv, qp);
 501         else
 502                 ret = cma_init_conn_qp(id_priv, qp);
 503         if (ret)
 504                 goto err;
 505
 506         id->qp = qp;
 507         id_priv->qp_num = qp->qp_num;
 508         id_priv->srq = (qp->srq != NULL);
 509         return 0;
 510 err:
 511         ib_destroy_qp(qp);
 512         return ret;
 513 }
 514 EXPORT_SYMBOL(rdma_create_qp);
 515
 516 void rdma_destroy_qp(struct rdma_cm_id *id)
 517 {
 518         struct rdma_id_private *id_priv;
 519
 520         id_priv = container_of(id, struct rdma_id_private, id);
 521         mutex_lock(&id_priv->qp_mutex);
 522         ib_destroy_qp(id_priv->id.qp);
 523         id_priv->id.qp = NULL;
 524         mutex_unlock(&id_priv->qp_mutex);
 525 }
 526 EXPORT_SYMBOL(rdma_destroy_qp);
 527
 528 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 529                              struct rdma_conn_param *conn_param)
 530 {
 531         struct ib_qp_attr qp_attr;
 532         int qp_attr_mask, ret;
 533
 534         mutex_lock(&id_priv->qp_mutex);
 535         if (!id_priv->id.qp) {
 536                 ret = 0;
 537                 goto out;
 538         }
 539
 540         /* Need to update QP attributes from default values. */
 541         qp_attr.qp_state = IB_QPS_INIT;
 542         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 543         if (ret)
 544                 goto out;
 545
 546         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 547         if (ret)
 548                 goto out;
 549
 550         qp_attr.qp_state = IB_QPS_RTR;
 551         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 552         if (ret)
 553                 goto out;
 554
 555         if (conn_param)
 556                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 557         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 558 out:
 559         mutex_unlock(&id_priv->qp_mutex);
 560         return ret;
 561 }
 562
 563 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 564                              struct rdma_conn_param *conn_param)
 565 {
 566         struct ib_qp_attr qp_attr;
 567         int qp_attr_mask, ret;
 568
 569         mutex_lock(&id_priv->qp_mutex);
 570         if (!id_priv->id.qp) {
 571                 ret = 0;
 572                 goto out;
 573         }
 574
 575         qp_attr.qp_state = IB_QPS_RTS;
 576         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 577         if (ret)
 578                 goto out;
 579
 580         if (conn_param)
 581                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 582         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 583 out:
 584         mutex_unlock(&id_priv->qp_mutex);
 585         return ret;
 586 }
 587
 588 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 589 {
 590         struct ib_qp_attr qp_attr;
 591         int ret;
 592
 593         mutex_lock(&id_priv->qp_mutex);
 594         if (!id_priv->id.qp) {
 595                 ret = 0;
 596                 goto out;
 597         }
 598
 599         qp_attr.qp_state = IB_QPS_ERR;
 600         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 601 out:
 602         mutex_unlock(&id_priv->qp_mutex);
 603         return ret;
 604 }
 605
 606 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 607                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 608 {
 609         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 610         int ret;
 611         u16 pkey;
 612
 613         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 614             IB_LINK_LAYER_INFINIBAND)
 615                 pkey = ib_addr_get_pkey(dev_addr);
 616         else
 617                 pkey = 0xffff;
 618
 619         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 620                                   pkey, &qp_attr->pkey_index);
 621         if (ret)
 622                 return ret;
 623
 624         qp_attr->port_num = id_priv->id.port_num;
 625         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 626
 627         if (cma_is_ud_ps(id_priv->id.ps)) {
 628                 ret = cma_set_qkey(id_priv);
 629                 if (ret)
 630                         return ret;
 631
 632                 qp_attr->qkey = id_priv->qkey;
 633                 *qp_attr_mask |= IB_QP_QKEY;
 634         } else {
 635                 qp_attr->qp_access_flags = 0;
 636                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 637         }
 638         return 0;
 639 }
 640
 641 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 642                        int *qp_attr_mask)
 643 {
 644         struct rdma_id_private *id_priv;
 645         int ret = 0;
 646
 647         id_priv = container_of(id, struct rdma_id_private, id);
 648         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 649         case RDMA_TRANSPORT_IB:
 650                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
 651                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 652                 else
 653                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 654                                                  qp_attr_mask);
 655                 if (qp_attr->qp_state == IB_QPS_RTR)
 656                         qp_attr->rq_psn = id_priv->seq_num;
 657                 break;
 658         case RDMA_TRANSPORT_IWARP:
 659                 if (!id_priv->cm_id.iw) {
 660                         qp_attr->qp_access_flags = 0;
 661                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 662                 } else
 663                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 664                                                  qp_attr_mask);
 665                 break;
 666         default:
 667                 ret = -ENOSYS;
 668                 break;
 669         }
 670
 671         return ret;
 672 }
 673 EXPORT_SYMBOL(rdma_init_qp_attr);
 674
 675 static inline int cma_zero_addr(struct sockaddr *addr)
 676 {
 677         struct in6_addr *ip6;
 678
 679         if (addr->sa_family == AF_INET)
 680                 return ipv4_is_zeronet(
 681                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 682         else {
 683                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 684                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 685                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 686         }
 687 }
 688
 689 static inline int cma_loopback_addr(struct sockaddr *addr)
 690 {
 691         if (addr->sa_family == AF_INET)
 692                 return ipv4_is_loopback(
 693                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 694         else
 695                 return ipv6_addr_loopback(
 696                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 697 }
 698
 699 static inline int cma_any_addr(struct sockaddr *addr)
 700 {
 701         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 702 }
 703
 704 static inline __be16 cma_port(struct sockaddr *addr)
 705 {
 706         if (addr->sa_family == AF_INET)
 707                 return ((struct sockaddr_in *) addr)->sin_port;
 708         else
 709                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 710 }
 711
 712 static inline int cma_any_port(struct sockaddr *addr)
 713 {
 714         return !cma_port(addr);
 715 }
 716
 717 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 718                             u8 *ip_ver, __be16 *port,
 719                             union cma_ip_addr **src, union cma_ip_addr **dst)
 720 {
 721         switch (ps) {
 722         case RDMA_PS_SDP:
 723                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 724                     SDP_MAJ_VERSION)
 725                         return -EINVAL;
 726
 727                 *ip_ver = sdp_get_ip_ver(hdr);
 728                 *port   = ((struct sdp_hh *) hdr)->port;
 729                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 730                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 731                 break;
 732         default:
 733                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 734                         return -EINVAL;
 735
 736                 *ip_ver = cma_get_ip_ver(hdr);
 737                 *port   = ((struct cma_hdr *) hdr)->port;
 738                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 739                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 740                 break;
 741         }
 742
 743         if (*ip_ver != 4 && *ip_ver != 6)
 744                 return -EINVAL;
 745         return 0;
 746 }
 747
 748 static void cma_save_net_info(struct rdma_addr *addr,
 749                               struct rdma_addr *listen_addr,
 750                               u8 ip_ver, __be16 port,
 751                               union cma_ip_addr *src, union cma_ip_addr *dst)
 752 {
 753         struct sockaddr_in *listen4, *ip4;
 754         struct sockaddr_in6 *listen6, *ip6;
 755
 756         switch (ip_ver) {
 757         case 4:
 758                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 759                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 760                 ip4->sin_family = listen4->sin_family;
 761                 ip4->sin_addr.s_addr = dst->ip4.addr;
 762                 ip4->sin_port = listen4->sin_port;
 763
 764                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 765                 ip4->sin_family = listen4->sin_family;
 766                 ip4->sin_addr.s_addr = src->ip4.addr;
 767                 ip4->sin_port = port;
 768                 break;
 769         case 6:
 770                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 771                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 772                 ip6->sin6_family = listen6->sin6_family;
 773                 ip6->sin6_addr = dst->ip6;
 774                 ip6->sin6_port = listen6->sin6_port;
 775
 776                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 777                 ip6->sin6_family = listen6->sin6_family;
 778                 ip6->sin6_addr = src->ip6;
 779                 ip6->sin6_port = port;
 780                 break;
 781         default:
 782                 break;
 783         }
 784 }
 785
 786 static inline int cma_user_data_offset(enum rdma_port_space ps)
 787 {
 788         switch (ps) {
 789         case RDMA_PS_SDP:
 790                 return 0;
 791         default:
 792                 return sizeof(struct cma_hdr);
 793         }
 794 }
 795
 796 static void cma_cancel_route(struct rdma_id_private *id_priv)
 797 {
 798         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 799         case IB_LINK_LAYER_INFINIBAND:
 800                 if (id_priv->query)
 801                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 802                 break;
 803         default:
 804                 break;
 805         }
 806 }
 807
 808 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 809 {
 810         struct rdma_id_private *dev_id_priv;
 811
 812         /*
 813          * Remove from listen_any_list to prevent added devices from spawning
 814          * additional listen requests.
 815          */
 816         mutex_lock(&lock);
 817         list_del(&id_priv->list);
 818
 819         while (!list_empty(&id_priv->listen_list)) {
 820                 dev_id_priv = list_entry(id_priv->listen_list.next,
 821                                          struct rdma_id_private, listen_list);
 822                 /* sync with device removal to avoid duplicate destruction */
 823                 list_del_init(&dev_id_priv->list);
 824                 list_del(&dev_id_priv->listen_list);
 825                 mutex_unlock(&lock);
 826
 827                 rdma_destroy_id(&dev_id_priv->id);
 828                 mutex_lock(&lock);
 829         }
 830         mutex_unlock(&lock);
 831 }
 832
 833 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 834                                  enum cma_state state)
 835 {
 836         switch (state) {
 837         case CMA_ADDR_QUERY:
 838                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 839                 break;
 840         case CMA_ROUTE_QUERY:
 841                 cma_cancel_route(id_priv);
 842                 break;
 843         case CMA_LISTEN:
 844                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 845                                 && !id_priv->cma_dev)
 846                         cma_cancel_listens(id_priv);
 847                 break;
 848         default:
 849                 break;
 850         }
 851 }
 852
 853 static void cma_release_port(struct rdma_id_private *id_priv)
 854 {
 855         struct rdma_bind_list *bind_list = id_priv->bind_list;
 856
 857         if (!bind_list)
 858                 return;
 859
 860         mutex_lock(&lock);
 861         hlist_del(&id_priv->node);
 862         if (hlist_empty(&bind_list->owners)) {
 863                 idr_remove(bind_list->ps, bind_list->port);
 864                 kfree(bind_list);
 865         }
 866         mutex_unlock(&lock);
 867         if (id_priv->sock)
 868                 sock_release(id_priv->sock);
 869 }
 870
 871 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 872 {
 873         struct cma_multicast *mc;
 874
 875         while (!list_empty(&id_priv->mc_list)) {
 876                 mc = container_of(id_priv->mc_list.next,
 877                                   struct cma_multicast, list);
 878                 list_del(&mc->list);
 879                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 880                 case IB_LINK_LAYER_INFINIBAND:
 881                         ib_sa_free_multicast(mc->multicast.ib);
 882                         kfree(mc);
 883                         break;
 884                 case IB_LINK_LAYER_ETHERNET:
 885                         kref_put(&mc->mcref, release_mc);
 886                         break;
 887                 default:
 888                         break;
 889                 }
 890         }
 891 }
 892
 893 void rdma_destroy_id(struct rdma_cm_id *id)
 894 {
 895         struct rdma_id_private *id_priv;
 896         enum cma_state state;
 897
 898         id_priv = container_of(id, struct rdma_id_private, id);
 899         state = cma_exch(id_priv, CMA_DESTROYING);
 900         cma_cancel_operation(id_priv, state);
 901
 902         mutex_lock(&lock);
 903         if (id_priv->cma_dev) {
 904                 mutex_unlock(&lock);
 905                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 906                 case RDMA_TRANSPORT_IB:
 907                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 908                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 909                         break;
 910                 case RDMA_TRANSPORT_IWARP:
 911                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 912                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 913                         break;
 914                 default:
 915                         break;
 916                 }
 917                 cma_leave_mc_groups(id_priv);
 918                 mutex_lock(&lock);
 919                 cma_detach_from_dev(id_priv);
 920         }
 921         mutex_unlock(&lock);
 922
 923         cma_release_port(id_priv);
 924         cma_deref_id(id_priv);
 925         wait_for_completion(&id_priv->comp);
 926
 927         if (id_priv->internal_id)
 928                 cma_deref_id(id_priv->id.context);
 929
 930         kfree(id_priv->id.route.path_rec);
 931         kfree(id_priv);
 932 }
 933 EXPORT_SYMBOL(rdma_destroy_id);
 934
 935 static int cma_rep_recv(struct rdma_id_private *id_priv)
 936 {
 937         int ret;
 938
 939         ret = cma_modify_qp_rtr(id_priv, NULL);
 940         if (ret)
 941                 goto reject;
 942
 943         ret = cma_modify_qp_rts(id_priv, NULL);
 944         if (ret)
 945                 goto reject;
 946
 947         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 948         if (ret)
 949                 goto reject;
 950
 951         return 0;
 952 reject:
 953         cma_modify_qp_err(id_priv);
 954         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 955                        NULL, 0, NULL, 0);
 956         return ret;
 957 }
 958
 959 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 960 {
 961         if (id_priv->id.ps == RDMA_PS_SDP &&
 962             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 963             SDP_MAJ_VERSION)
 964                 return -EINVAL;
 965
 966         return 0;
 967 }
 968
 969 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 970                                    struct ib_cm_rep_event_param *rep_data,
 971                                    void *private_data)
 972 {
 973         event->param.conn.private_data = private_data;
 974         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 975         event->param.conn.responder_resources = rep_data->responder_resources;
 976         event->param.conn.initiator_depth = rep_data->initiator_depth;
 977         event->param.conn.flow_control = rep_data->flow_control;
 978         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 979         event->param.conn.srq = rep_data->srq;
 980         event->param.conn.qp_num = rep_data->remote_qpn;
 981 }
 982
 983 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 984 {
 985         struct rdma_id_private *id_priv = cm_id->context;
 986         struct rdma_cm_event event;
 987         int ret = 0;
 988
 989         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
 990                 cma_disable_callback(id_priv, CMA_CONNECT)) ||
 991             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
 992                 cma_disable_callback(id_priv, CMA_DISCONNECT)))
 993                 return 0;
 994
 995         memset(&event, 0, sizeof event);
 996         switch (ib_event->event) {
 997         case IB_CM_REQ_ERROR:
 998         case IB_CM_REP_ERROR:
 999                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1000                 event.status = -ETIMEDOUT;
1001                 break;
1002         case IB_CM_REP_RECEIVED:
1003                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1004                 if (event.status)
1005                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1006                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1007                         event.status = cma_rep_recv(id_priv);
1008                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1009                                                      RDMA_CM_EVENT_ESTABLISHED;
1010                 } else
1011                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1012                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1013                                        ib_event->private_data);
1014                 break;
1015         case IB_CM_RTU_RECEIVED:
1016         case IB_CM_USER_ESTABLISHED:
1017                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1018                 break;
1019         case IB_CM_DREQ_ERROR:
1020                 event.status = -ETIMEDOUT; /* fall through */
1021         case IB_CM_DREQ_RECEIVED:
1022         case IB_CM_DREP_RECEIVED:
1023                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
1024                         goto out;
1025                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1026                 break;
1027         case IB_CM_TIMEWAIT_EXIT:
1028                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1029                 break;
1030         case IB_CM_MRA_RECEIVED:
1031                 /* ignore event */
1032                 goto out;
1033         case IB_CM_REJ_RECEIVED:
1034                 cma_modify_qp_err(id_priv);
1035                 event.status = ib_event->param.rej_rcvd.reason;
1036                 event.event = RDMA_CM_EVENT_REJECTED;
1037                 event.param.conn.private_data = ib_event->private_data;
1038                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1039                 break;
1040         default:
1041                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1042                        ib_event->event);
1043                 goto out;
1044         }
1045
1046         ret = id_priv->id.event_handler(&id_priv->id, &event);
1047         if (ret) {
1048                 /* Destroy the CM ID by returning a non-zero value. */
1049                 id_priv->cm_id.ib = NULL;
1050                 cma_exch(id_priv, CMA_DESTROYING);
1051                 mutex_unlock(&id_priv->handler_mutex);
1052                 rdma_destroy_id(&id_priv->id);
1053                 return ret;
1054         }
1055 out:
1056         mutex_unlock(&id_priv->handler_mutex);
1057         return ret;
1058 }
1059
1060 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1061                                                struct ib_cm_event *ib_event)
1062 {
1063         struct rdma_id_private *id_priv;
1064         struct rdma_cm_id *id;
1065         struct rdma_route *rt;
1066         union cma_ip_addr *src, *dst;
1067         __be16 port;
1068         u8 ip_ver;
1069         int ret;
1070
1071         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1072                              &ip_ver, &port, &src, &dst))
1073                 goto err;
1074
1075         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1076                             listen_id->ps);
1077         if (IS_ERR(id))
1078                 goto err;
1079
1080         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1081                           ip_ver, port, src, dst);
1082
1083         rt = &id->route;
1084         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1085         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1086                                GFP_KERNEL);
1087         if (!rt->path_rec)
1088                 goto destroy_id;
1089
1090         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1091         if (rt->num_paths == 2)
1092                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1093
1094         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1095                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1096                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1097                 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1098         } else {
1099                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1100                                         &rt->addr.dev_addr);
1101                 if (ret)
1102                         goto destroy_id;
1103         }
1104         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1105
1106         id_priv = container_of(id, struct rdma_id_private, id);
1107         id_priv->state = CMA_CONNECT;
1108         return id_priv;
1109
1110 destroy_id:
1111         rdma_destroy_id(id);
1112 err:
1113         return NULL;
1114 }
1115
1116 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1117                                               struct ib_cm_event *ib_event)
1118 {
1119         struct rdma_id_private *id_priv;
1120         struct rdma_cm_id *id;
1121         union cma_ip_addr *src, *dst;
1122         __be16 port;
1123         u8 ip_ver;
1124         int ret;
1125
1126         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1127                             listen_id->ps);
1128         if (IS_ERR(id))
1129                 return NULL;
1130
1131
1132         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1133                              &ip_ver, &port, &src, &dst))
1134                 goto err;
1135
1136         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1137                           ip_ver, port, src, dst);
1138
1139         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1140                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1141                                         &id->route.addr.dev_addr);
1142                 if (ret)
1143                         goto err;
1144         }
1145
1146         id_priv = container_of(id, struct rdma_id_private, id);
1147         id_priv->state = CMA_CONNECT;
1148         return id_priv;
1149 err:
1150         rdma_destroy_id(id);
1151         return NULL;
1152 }
1153
1154 static void cma_set_req_event_data(struct rdma_cm_event *event,
1155                                    struct ib_cm_req_event_param *req_data,
1156                                    void *private_data, int offset)
1157 {
1158         event->param.conn.private_data = private_data + offset;
1159         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1160         event->param.conn.responder_resources = req_data->responder_resources;
1161         event->param.conn.initiator_depth = req_data->initiator_depth;
1162         event->param.conn.flow_control = req_data->flow_control;
1163         event->param.conn.retry_count = req_data->retry_count;
1164         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1165         event->param.conn.srq = req_data->srq;
1166         event->param.conn.qp_num = req_data->remote_qpn;
1167 }
1168
1169 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1170 {
1171         struct rdma_id_private *listen_id, *conn_id;
1172         struct rdma_cm_event event;
1173         int offset, ret;
1174
1175         listen_id = cm_id->context;
1176         if (cma_disable_callback(listen_id, CMA_LISTEN))
1177                 return -ECONNABORTED;
1178
1179         memset(&event, 0, sizeof event);
1180         offset = cma_user_data_offset(listen_id->id.ps);
1181         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1182         if (cma_is_ud_ps(listen_id->id.ps)) {
1183                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1184                 event.param.ud.private_data = ib_event->private_data + offset;
1185                 event.param.ud.private_data_len =
1186                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1187         } else {
1188                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1189                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1190                                        ib_event->private_data, offset);
1191         }
1192         if (!conn_id) {
1193                 ret = -ENOMEM;
1194                 goto out;
1195         }
1196
1197         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1198         mutex_lock(&lock);
1199         ret = cma_acquire_dev(conn_id);
1200         mutex_unlock(&lock);
1201         if (ret)
1202                 goto release_conn_id;
1203
1204         conn_id->cm_id.ib = cm_id;
1205         cm_id->context = conn_id;
1206         cm_id->cm_handler = cma_ib_handler;
1207
1208         ret = conn_id->id.event_handler(&conn_id->id, &event);
1209         if (!ret) {
1210                 /*
1211                  * Acquire mutex to prevent user executing rdma_destroy_id()
1212                  * while we're accessing the cm_id.
1213                  */
1214                 mutex_lock(&lock);
1215                 if (cma_comp(conn_id, CMA_CONNECT) &&
1216                     !cma_is_ud_ps(conn_id->id.ps))
1217                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1218                 mutex_unlock(&lock);
1219                 mutex_unlock(&conn_id->handler_mutex);
1220                 goto out;
1221         }
1222
1223         /* Destroy the CM ID by returning a non-zero value. */
1224         conn_id->cm_id.ib = NULL;
1225
1226 release_conn_id:
1227         cma_exch(conn_id, CMA_DESTROYING);
1228         mutex_unlock(&conn_id->handler_mutex);
1229         rdma_destroy_id(&conn_id->id);
1230
1231 out:
1232         mutex_unlock(&listen_id->handler_mutex);
1233         return ret;
1234 }
1235
1236 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1237 {
1238         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1239 }
1240
1241 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1242                                  struct ib_cm_compare_data *compare)
1243 {
1244         struct cma_hdr *cma_data, *cma_mask;
1245         struct sdp_hh *sdp_data, *sdp_mask;
1246         __be32 ip4_addr;
1247 #ifdef INET6
1248         struct in6_addr ip6_addr;
1249 #endif
1250
1251         memset(compare, 0, sizeof *compare);
1252         cma_data = (void *) compare->data;
1253         cma_mask = (void *) compare->mask;
1254         sdp_data = (void *) compare->data;
1255         sdp_mask = (void *) compare->mask;
1256
1257         switch (addr->sa_family) {
1258         case AF_INET:
1259                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1260                 if (ps == RDMA_PS_SDP) {
1261                         sdp_set_ip_ver(sdp_data, 4);
1262                         sdp_set_ip_ver(sdp_mask, 0xF);
1263                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1264                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1265                 } else {
1266                         cma_set_ip_ver(cma_data, 4);
1267                         cma_set_ip_ver(cma_mask, 0xF);
1268                         cma_data->dst_addr.ip4.addr = ip4_addr;
1269                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1270                 }
1271                 break;
1272 #ifdef INET6
1273         case AF_INET6:
1274                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1275                 if (ps == RDMA_PS_SDP) {
1276                         sdp_set_ip_ver(sdp_data, 6);
1277                         sdp_set_ip_ver(sdp_mask, 0xF);
1278                         sdp_data->dst_addr.ip6 = ip6_addr;
1279                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1280                                sizeof sdp_mask->dst_addr.ip6);
1281                 } else {
1282                         cma_set_ip_ver(cma_data, 6);
1283                         cma_set_ip_ver(cma_mask, 0xF);
1284                         cma_data->dst_addr.ip6 = ip6_addr;
1285                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1286                                sizeof cma_mask->dst_addr.ip6);
1287                 }
1288                 break;
1289 #endif
1290         default:
1291                 break;
1292         }
1293 }
1294
1295 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1296 {
1297         struct rdma_id_private *id_priv = iw_id->context;
1298         struct rdma_cm_event event;
1299         struct sockaddr_in *sin;
1300         int ret = 0;
1301
1302         if (cma_disable_callback(id_priv, CMA_CONNECT))
1303                 return 0;
1304
1305         memset(&event, 0, sizeof event);
1306         switch (iw_event->event) {
1307         case IW_CM_EVENT_CLOSE:
1308                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1309                 break;
1310         case IW_CM_EVENT_CONNECT_REPLY:
1311                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1312                 *sin = iw_event->local_addr;
1313                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1314                 *sin = iw_event->remote_addr;
1315                 switch ((int)iw_event->status) {
1316                 case 0:
1317                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1318                         break;
1319                 case -ECONNRESET:
1320                 case -ECONNREFUSED:
1321                         event.event = RDMA_CM_EVENT_REJECTED;
1322                         break;
1323                 case -ETIMEDOUT:
1324                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1325                         break;
1326                 default:
1327                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1328                         break;
1329                 }
1330                 break;
1331         case IW_CM_EVENT_ESTABLISHED:
1332                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1333                 break;
1334         default:
1335                 BUG_ON(1);
1336         }
1337
1338         event.status = iw_event->status;
1339         event.param.conn.private_data = iw_event->private_data;
1340         event.param.conn.private_data_len = iw_event->private_data_len;
1341         ret = id_priv->id.event_handler(&id_priv->id, &event);
1342         if (ret) {
1343                 /* Destroy the CM ID by returning a non-zero value. */
1344                 id_priv->cm_id.iw = NULL;
1345                 cma_exch(id_priv, CMA_DESTROYING);
1346                 mutex_unlock(&id_priv->handler_mutex);
1347                 rdma_destroy_id(&id_priv->id);
1348                 return ret;
1349         }
1350
1351         mutex_unlock(&id_priv->handler_mutex);
1352         return ret;
1353 }
1354
1355 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1356                                struct iw_cm_event *iw_event)
1357 {
1358         struct rdma_cm_id *new_cm_id;
1359         struct rdma_id_private *listen_id, *conn_id;
1360         struct sockaddr_in *sin;
1361         struct net_device *dev = NULL;
1362         struct rdma_cm_event event;
1363         int ret;
1364         struct ib_device_attr attr;
1365
1366         listen_id = cm_id->context;
1367         if (cma_disable_callback(listen_id, CMA_LISTEN))
1368                 return -ECONNABORTED;
1369
1370         /* Create a new RDMA id for the new IW CM ID */
1371         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1372                                    listen_id->id.context,
1373                                    RDMA_PS_TCP);
1374         if (IS_ERR(new_cm_id)) {
1375                 ret = -ENOMEM;
1376                 goto out;
1377         }
1378         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1379         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1380         conn_id->state = CMA_CONNECT;
1381
1382         dev = ip_dev_find(NULL, iw_event->local_addr.sin_addr.s_addr);
1383         if (!dev) {
1384                 ret = -EADDRNOTAVAIL;
1385                 mutex_unlock(&conn_id->handler_mutex);
1386                 rdma_destroy_id(new_cm_id);
1387                 goto out;
1388         }
1389         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1390         if (ret) {
1391                 mutex_unlock(&conn_id->handler_mutex);
1392                 rdma_destroy_id(new_cm_id);
1393                 goto out;
1394         }
1395
1396         mutex_lock(&lock);
1397         ret = cma_acquire_dev(conn_id);
1398         mutex_unlock(&lock);
1399         if (ret) {
1400                 mutex_unlock(&conn_id->handler_mutex);
1401                 rdma_destroy_id(new_cm_id);
1402                 goto out;
1403         }
1404
1405         conn_id->cm_id.iw = cm_id;
1406         cm_id->context = conn_id;
1407         cm_id->cm_handler = cma_iw_handler;
1408
1409         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1410         *sin = iw_event->local_addr;
1411         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1412         *sin = iw_event->remote_addr;
1413
1414         ret = ib_query_device(conn_id->id.device, &attr);
1415         if (ret) {
1416                 mutex_unlock(&conn_id->handler_mutex);
1417                 rdma_destroy_id(new_cm_id);
1418                 goto out;
1419         }
1420
1421         memset(&event, 0, sizeof event);
1422         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1423         event.param.conn.private_data = iw_event->private_data;
1424         event.param.conn.private_data_len = iw_event->private_data_len;
1425         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1426         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1427         ret = conn_id->id.event_handler(&conn_id->id, &event);
1428         if (ret) {
1429                 /* User wants to destroy the CM ID */
1430                 conn_id->cm_id.iw = NULL;
1431                 cma_exch(conn_id, CMA_DESTROYING);
1432                 mutex_unlock(&conn_id->handler_mutex);
1433                 rdma_destroy_id(&conn_id->id);
1434                 goto out;
1435         }
1436
1437         mutex_unlock(&conn_id->handler_mutex);
1438
1439 out:
1440         if (dev)
1441                 dev_put(dev);
1442         mutex_unlock(&listen_id->handler_mutex);
1443         return ret;
1444 }
1445
1446 static int cma_ib_listen(struct rdma_id_private *id_priv)
1447 {
1448         struct ib_cm_compare_data compare_data;
1449         struct sockaddr *addr;
1450         __be64 svc_id;
1451         int ret;
1452
1453         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1454                                             id_priv);
1455         if (IS_ERR(id_priv->cm_id.ib))
1456                 return PTR_ERR(id_priv->cm_id.ib);
1457
1458         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1459         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1460         if (cma_any_addr(addr))
1461                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1462         else {
1463                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1464                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1465         }
1466
1467         if (ret) {
1468                 ib_destroy_cm_id(id_priv->cm_id.ib);
1469                 id_priv->cm_id.ib = NULL;
1470         }
1471
1472         return ret;
1473 }
1474
1475 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1476 {
1477         int ret;
1478         struct sockaddr_in *sin;
1479
1480         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1481                                             id_priv->sock,
1482                                             iw_conn_req_handler,
1483                                             id_priv);
1484         if (IS_ERR(id_priv->cm_id.iw))
1485                 return PTR_ERR(id_priv->cm_id.iw);
1486
1487         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1488         id_priv->cm_id.iw->local_addr = *sin;
1489
1490         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1491
1492         if (ret) {
1493                 iw_destroy_cm_id(id_priv->cm_id.iw);
1494                 id_priv->cm_id.iw = NULL;
1495         }
1496
1497         return ret;
1498 }
1499
1500 static int cma_listen_handler(struct rdma_cm_id *id,
1501                               struct rdma_cm_event *event)
1502 {
1503         struct rdma_id_private *id_priv = id->context;
1504
1505         id->context = id_priv->id.context;
1506         id->event_handler = id_priv->id.event_handler;
1507         return id_priv->id.event_handler(id, event);
1508 }
1509
1510 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1511                               struct cma_device *cma_dev)
1512 {
1513         struct rdma_id_private *dev_id_priv;
1514         struct rdma_cm_id *id;
1515         int ret;
1516
1517         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1518         if (IS_ERR(id))
1519                 return;
1520
1521         dev_id_priv = container_of(id, struct rdma_id_private, id);
1522
1523         dev_id_priv->state = CMA_ADDR_BOUND;
1524         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1525                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1526
1527         cma_attach_to_dev(dev_id_priv, cma_dev);
1528         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1529         atomic_inc(&id_priv->refcount);
1530         dev_id_priv->internal_id = 1;
1531
1532         ret = rdma_listen(id, id_priv->backlog);
1533         if (ret)
1534                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1535                        "listening on device %s\n", ret, cma_dev->device->name);
1536 }
1537
1538 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1539 {
1540         struct cma_device *cma_dev;
1541
1542         mutex_lock(&lock);
1543         list_add_tail(&id_priv->list, &listen_any_list);
1544         list_for_each_entry(cma_dev, &dev_list, list)
1545                 cma_listen_on_dev(id_priv, cma_dev);
1546         mutex_unlock(&lock);
1547 }
1548
1549 int rdma_listen(struct rdma_cm_id *id, int backlog)
1550 {
1551         struct rdma_id_private *id_priv;
1552         int ret;
1553
1554         id_priv = container_of(id, struct rdma_id_private, id);
1555         if (id_priv->state == CMA_IDLE) {
1556                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
1557                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
1558                 if (ret)
1559                         return ret;
1560         }
1561
1562         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1563                 return -EINVAL;
1564
1565         id_priv->backlog = backlog;
1566         if (id->device) {
1567                 switch (rdma_node_get_transport(id->device->node_type)) {
1568                 case RDMA_TRANSPORT_IB:
1569                         ret = cma_ib_listen(id_priv);
1570                         if (ret)
1571                                 goto err;
1572                         break;
1573                 case RDMA_TRANSPORT_IWARP:
1574                         ret = cma_iw_listen(id_priv, backlog);
1575                         if (ret)
1576                                 goto err;
1577                         break;
1578                 default:
1579                         ret = -ENOSYS;
1580                         goto err;
1581                 }
1582         } else
1583                 cma_listen_on_all(id_priv);
1584
1585         return 0;
1586 err:
1587         id_priv->backlog = 0;
1588         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1589         return ret;
1590 }
1591 EXPORT_SYMBOL(rdma_listen);
1592
1593 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1594 {
1595         struct rdma_id_private *id_priv;
1596
1597         id_priv = container_of(id, struct rdma_id_private, id);
1598         id_priv->tos = (u8) tos;
1599 }
1600 EXPORT_SYMBOL(rdma_set_service_type);
1601
1602 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1603                               void *context)
1604 {
1605         struct cma_work *work = context;
1606         struct rdma_route *route;
1607
1608         route = &work->id->id.route;
1609
1610         if (!status) {
1611                 route->num_paths = 1;
1612                 *route->path_rec = *path_rec;
1613         } else {
1614                 work->old_state = CMA_ROUTE_QUERY;
1615                 work->new_state = CMA_ADDR_RESOLVED;
1616                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1617                 work->event.status = status;
1618         }
1619
1620         queue_work(cma_wq, &work->work);
1621 }
1622
1623 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1624                               struct cma_work *work)
1625 {
1626         struct rdma_addr *addr = &id_priv->id.route.addr;
1627         struct ib_sa_path_rec path_rec;
1628         ib_sa_comp_mask comp_mask;
1629         struct sockaddr_in6 *sin6;
1630
1631         memset(&path_rec, 0, sizeof path_rec);
1632         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1633         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1634         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1635         path_rec.numb_path = 1;
1636         path_rec.reversible = 1;
1637         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1638                                                         (struct sockaddr *) &addr->dst_addr);
1639
1640         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1641                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1642                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1643
1644         if (addr->src_addr.ss_family == AF_INET) {
1645                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1646                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1647         } else {
1648                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1649                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1650                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1651         }
1652
1653         if (tavor_quirk) {
1654                 path_rec.mtu_selector = IB_SA_LT;
1655                 path_rec.mtu = IB_MTU_2048;
1656         }
1657
1658         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1659                                                id_priv->id.port_num, &path_rec,
1660                                                comp_mask, timeout_ms,
1661                                                GFP_KERNEL, cma_query_handler,
1662                                                work, &id_priv->query);
1663
1664         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1665 }
1666
1667 static void cma_work_handler(struct work_struct *_work)
1668 {
1669         struct cma_work *work = container_of(_work, struct cma_work, work);
1670         struct rdma_id_private *id_priv = work->id;
1671         int destroy = 0;
1672
1673         mutex_lock(&id_priv->handler_mutex);
1674         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1675                 goto out;
1676
1677         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1678                 cma_exch(id_priv, CMA_DESTROYING);
1679                 destroy = 1;
1680         }
1681 out:
1682         mutex_unlock(&id_priv->handler_mutex);
1683         cma_deref_id(id_priv);
1684         if (destroy)
1685                 rdma_destroy_id(&id_priv->id);
1686         kfree(work);
1687 }
1688
1689 static void cma_ndev_work_handler(struct work_struct *_work)
1690 {
1691         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1692         struct rdma_id_private *id_priv = work->id;
1693         int destroy = 0;
1694
1695         mutex_lock(&id_priv->handler_mutex);
1696         if (id_priv->state == CMA_DESTROYING ||
1697             id_priv->state == CMA_DEVICE_REMOVAL)
1698                 goto out;
1699
1700         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1701                 cma_exch(id_priv, CMA_DESTROYING);
1702                 destroy = 1;
1703         }
1704
1705 out:
1706         mutex_unlock(&id_priv->handler_mutex);
1707         cma_deref_id(id_priv);
1708         if (destroy)
1709                 rdma_destroy_id(&id_priv->id);
1710         kfree(work);
1711 }
1712
1713 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1714 {
1715         struct rdma_route *route = &id_priv->id.route;
1716         struct cma_work *work;
1717         int ret;
1718
1719         work = kzalloc(sizeof *work, GFP_KERNEL);
1720         if (!work)
1721                 return -ENOMEM;
1722
1723         work->id = id_priv;
1724         INIT_WORK(&work->work, cma_work_handler);
1725         work->old_state = CMA_ROUTE_QUERY;
1726         work->new_state = CMA_ROUTE_RESOLVED;
1727         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1728
1729         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1730         if (!route->path_rec) {
1731                 ret = -ENOMEM;
1732                 goto err1;
1733         }
1734
1735         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1736         if (ret)
1737                 goto err2;
1738
1739         return 0;
1740 err2:
1741         kfree(route->path_rec);
1742         route->path_rec = NULL;
1743 err1:
1744         kfree(work);
1745         return ret;
1746 }
1747
1748 int rdma_set_ib_paths(struct rdma_cm_id *id,
1749                       struct ib_sa_path_rec *path_rec, int num_paths)
1750 {
1751         struct rdma_id_private *id_priv;
1752         int ret;
1753
1754         id_priv = container_of(id, struct rdma_id_private, id);
1755         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1756                 return -EINVAL;
1757
1758         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1759         if (!id->route.path_rec) {
1760                 ret = -ENOMEM;
1761                 goto err;
1762         }
1763
1764         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1765         return 0;
1766 err:
1767         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1768         return ret;
1769 }
1770 EXPORT_SYMBOL(rdma_set_ib_paths);
1771
1772 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1773 {
1774         struct cma_work *work;
1775
1776         work = kzalloc(sizeof *work, GFP_KERNEL);
1777         if (!work)
1778                 return -ENOMEM;
1779
1780         work->id = id_priv;
1781         INIT_WORK(&work->work, cma_work_handler);
1782         work->old_state = CMA_ROUTE_QUERY;
1783         work->new_state = CMA_ROUTE_RESOLVED;
1784         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1785         queue_work(cma_wq, &work->work);
1786         return 0;
1787 }
1788
1789 static u8 tos_to_sl(u8 tos)
1790 {
1791         return def_prec2sl & 7;
1792 }
1793
1794 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1795 {
1796         struct rdma_route *route = &id_priv->id.route;
1797         struct rdma_addr *addr = &route->addr;
1798         struct cma_work *work;
1799         int ret;
1800         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1801         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1802         struct net_device *ndev = NULL;
1803         u16 vid;
1804
1805         if (src_addr->sin_family != dst_addr->sin_family)
1806                 return -EINVAL;
1807
1808         work = kzalloc(sizeof *work, GFP_KERNEL);
1809         if (!work)
1810                 return -ENOMEM;
1811
1812         work->id = id_priv;
1813         INIT_WORK(&work->work, cma_work_handler);
1814
1815         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1816         if (!route->path_rec) {
1817                 ret = -ENOMEM;
1818                 goto err1;
1819         }
1820
1821         route->num_paths = 1;
1822
1823         if (addr->dev_addr.bound_dev_if)
1824                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1825         if (!ndev) {
1826                 ret = -ENODEV;
1827                 goto err2;
1828         }
1829
1830         vid = rdma_vlan_dev_vlan_id(ndev);
1831
1832         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1833         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1834
1835         route->path_rec->hop_limit = 1;
1836         route->path_rec->reversible = 1;
1837         route->path_rec->pkey = cpu_to_be16(0xffff);
1838         route->path_rec->mtu_selector = IB_SA_EQ;
1839         route->path_rec->sl = tos_to_sl(id_priv->tos);
1840
1841 #ifdef __linux__
1842         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1843 #else
1844         route->path_rec->mtu = iboe_get_mtu(ndev->if_mtu);
1845 #endif
1846         route->path_rec->rate_selector = IB_SA_EQ;
1847         route->path_rec->rate = iboe_get_rate(ndev);
1848         dev_put(ndev);
1849         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1850         route->path_rec->packet_life_time = IBOE_PACKET_LIFETIME;
1851         if (!route->path_rec->mtu) {
1852                 ret = -EINVAL;
1853                 goto err2;
1854         }
1855
1856         work->old_state = CMA_ROUTE_QUERY;
1857         work->new_state = CMA_ROUTE_RESOLVED;
1858         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1859         work->event.status = 0;
1860
1861         queue_work(cma_wq, &work->work);
1862
1863         return 0;
1864
1865 err2:
1866         kfree(route->path_rec);
1867         route->path_rec = NULL;
1868 err1:
1869         kfree(work);
1870         return ret;
1871 }
1872
1873 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1874 {
1875         struct rdma_id_private *id_priv;
1876         int ret;
1877
1878         id_priv = container_of(id, struct rdma_id_private, id);
1879         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1880                 return -EINVAL;
1881
1882         atomic_inc(&id_priv->refcount);
1883         switch (rdma_node_get_transport(id->device->node_type)) {
1884         case RDMA_TRANSPORT_IB:
1885                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1886                 case IB_LINK_LAYER_INFINIBAND:
1887                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1888                         break;
1889                 case IB_LINK_LAYER_ETHERNET:
1890                         ret = cma_resolve_iboe_route(id_priv);
1891                         break;
1892                 default:
1893                         ret = -ENOSYS;
1894                 }
1895                 break;
1896         case RDMA_TRANSPORT_IWARP:
1897                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1898                 break;
1899         default:
1900                 ret = -ENOSYS;
1901                 break;
1902         }
1903         if (ret)
1904                 goto err;
1905
1906         return 0;
1907 err:
1908         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1909         cma_deref_id(id_priv);
1910         return ret;
1911 }
1912 EXPORT_SYMBOL(rdma_resolve_route);
1913
1914 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1915 {
1916         struct cma_device *cma_dev;
1917         struct ib_port_attr port_attr;
1918         union ib_gid gid;
1919         u16 pkey;
1920         int ret;
1921         u8 p;
1922
1923         mutex_lock(&lock);
1924         if (list_empty(&dev_list)) {
1925                 ret = -ENODEV;
1926                 goto out;
1927         }
1928         list_for_each_entry(cma_dev, &dev_list, list)
1929                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1930                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1931                             port_attr.state == IB_PORT_ACTIVE)
1932                                 goto port_found;
1933
1934         p = 1;
1935         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1936
1937 port_found:
1938         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1939         if (ret)
1940                 goto out;
1941
1942         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1943         if (ret)
1944                 goto out;
1945
1946         id_priv->id.route.addr.dev_addr.dev_type =
1947                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1948                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1949
1950         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1951         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1952         id_priv->id.port_num = p;
1953         cma_attach_to_dev(id_priv, cma_dev);
1954 out:
1955         mutex_unlock(&lock);
1956         return ret;
1957 }
1958
1959 static void addr_handler(int status, struct sockaddr *src_addr,
1960                          struct rdma_dev_addr *dev_addr, void *context)
1961 {
1962         struct rdma_id_private *id_priv = context;
1963         struct rdma_cm_event event;
1964
1965         memset(&event, 0, sizeof event);
1966         mutex_lock(&id_priv->handler_mutex);
1967
1968         /*
1969          * Grab mutex to block rdma_destroy_id() from removing the device while
1970          * we're trying to acquire it.
1971          */
1972         mutex_lock(&lock);
1973         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1974                 mutex_unlock(&lock);
1975                 goto out;
1976         }
1977
1978         if (!status && !id_priv->cma_dev)
1979                 status = cma_acquire_dev(id_priv);
1980         mutex_unlock(&lock);
1981
1982         if (status) {
1983                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1984                         goto out;
1985                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1986                 event.status = status;
1987         } else {
1988                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1989                        ip_addr_size(src_addr));
1990                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1991         }
1992
1993         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1994                 cma_exch(id_priv, CMA_DESTROYING);
1995                 mutex_unlock(&id_priv->handler_mutex);
1996                 cma_deref_id(id_priv);
1997                 rdma_destroy_id(&id_priv->id);
1998                 return;
1999         }
2000 out:
2001         mutex_unlock(&id_priv->handler_mutex);
2002         cma_deref_id(id_priv);
2003 }
2004
2005 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2006 {
2007         struct cma_work *work;
2008         struct sockaddr *src, *dst;
2009         union ib_gid gid;
2010         int ret;
2011
2012         work = kzalloc(sizeof *work, GFP_KERNEL);
2013         if (!work)
2014                 return -ENOMEM;
2015
2016         if (!id_priv->cma_dev) {
2017                 ret = cma_bind_loopback(id_priv);
2018                 if (ret)
2019                         goto err;
2020         }
2021
2022         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2023         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2024
2025         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2026         if (cma_zero_addr(src)) {
2027                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2028                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2029                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
2030                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
2031                 } else {
2032                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
2033                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
2034                 }
2035         }
2036
2037         work->id = id_priv;
2038         INIT_WORK(&work->work, cma_work_handler);
2039         work->old_state = CMA_ADDR_QUERY;
2040         work->new_state = CMA_ADDR_RESOLVED;
2041         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2042         queue_work(cma_wq, &work->work);
2043         return 0;
2044 err:
2045         kfree(work);
2046         return ret;
2047 }
2048
2049 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2050                          struct sockaddr *dst_addr)
2051 {
2052         if (!src_addr || !src_addr->sa_family) {
2053                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2054                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2055                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2056                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2057                 }
2058         }
2059         if (!cma_any_addr(src_addr))
2060                 return rdma_bind_addr(id, src_addr);
2061         else {
2062                 struct sockaddr_in addr_in;
2063
2064                 memset(&addr_in, 0, sizeof addr_in);
2065                 addr_in.sin_family = dst_addr->sa_family;
2066                 addr_in.sin_len = sizeof addr_in;
2067                 return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
2068         }
2069 }
2070
2071 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2072                       struct sockaddr *dst_addr, int timeout_ms)
2073 {
2074         struct rdma_id_private *id_priv;
2075         int ret;
2076
2077         id_priv = container_of(id, struct rdma_id_private, id);
2078         if (id_priv->state == CMA_IDLE) {
2079                 ret = cma_bind_addr(id, src_addr, dst_addr);
2080                 if (ret)
2081                         return ret;
2082         }
2083
2084         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
2085                 return -EINVAL;
2086
2087         atomic_inc(&id_priv->refcount);
2088         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2089         if (cma_any_addr(dst_addr))
2090                 ret = cma_resolve_loopback(id_priv);
2091         else
2092                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2093                                       dst_addr, &id->route.addr.dev_addr,
2094                                       timeout_ms, addr_handler, id_priv);
2095         if (ret)
2096                 goto err;
2097
2098         return 0;
2099 err:
2100         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
2101         cma_deref_id(id_priv);
2102         return ret;
2103 }
2104 EXPORT_SYMBOL(rdma_resolve_addr);
2105
2106 static void cma_bind_port(struct rdma_bind_list *bind_list,
2107                           struct rdma_id_private *id_priv)
2108 {
2109         struct sockaddr_in *sin;
2110
2111         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2112         sin->sin_port = htons(bind_list->port);
2113         id_priv->bind_list = bind_list;
2114         hlist_add_head(&id_priv->node, &bind_list->owners);
2115 }
2116
2117 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2118                           unsigned short snum)
2119 {
2120         struct rdma_bind_list *bind_list;
2121         int port, ret;
2122
2123         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2124         if (!bind_list)
2125                 return -ENOMEM;
2126
2127         do {
2128                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2129         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2130
2131         if (ret)
2132                 goto err1;
2133
2134         if (port != snum) {
2135                 ret = -EADDRNOTAVAIL;
2136                 goto err2;
2137         }
2138
2139         bind_list->ps = ps;
2140         bind_list->port = (unsigned short) port;
2141         cma_bind_port(bind_list, id_priv);
2142         return 0;
2143 err2:
2144         idr_remove(ps, port);
2145 err1:
2146         kfree(bind_list);
2147         return ret;
2148 }
2149
2150 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2151 {
2152 #if defined(INET)
2153         struct rdma_bind_list *bind_list;
2154         int port, ret, low, high;
2155
2156         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2157         if (!bind_list)
2158                 return -ENOMEM;
2159
2160 retry:
2161         /* FIXME: add proper port randomization per like inet_csk_get_port */
2162         do {
2163                 ret = idr_get_new_above(ps, bind_list, next_port, &port);
2164         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2165
2166         if (ret)
2167                 goto err1;
2168
2169         inet_get_local_port_range(&low, &high);
2170         if (port > high) {
2171                 if (next_port != low) {
2172                         idr_remove(ps, port);
2173                         next_port = low;
2174                         goto retry;
2175                 }
2176                 ret = -EADDRNOTAVAIL;
2177                 goto err2;
2178         }
2179
2180         if (port == high)
2181                 next_port = low;
2182         else
2183                 next_port = port + 1;
2184
2185         bind_list->ps = ps;
2186         bind_list->port = (unsigned short) port;
2187         cma_bind_port(bind_list, id_priv);
2188         return 0;
2189 err2:
2190         idr_remove(ps, port);
2191 err1:
2192         kfree(bind_list);
2193         return ret;
2194 #else
2195         return -ENOSPC;
2196 #endif
2197 }
2198
2199 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2200 {
2201         struct rdma_id_private *cur_id;
2202         struct sockaddr_in *sin, *cur_sin;
2203         struct rdma_bind_list *bind_list;
2204         struct hlist_node *node;
2205         unsigned short snum;
2206
2207         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2208         snum = ntohs(sin->sin_port);
2209 #ifdef __linux__
2210         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2211                 return -EACCES;
2212 #endif
2213
2214         bind_list = idr_find(ps, snum);
2215         if (!bind_list)
2216                 return cma_alloc_port(ps, id_priv, snum);
2217
2218         /*
2219          * We don't support binding to any address if anyone is bound to
2220          * a specific address on the same port.
2221          */
2222         if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2223                 return -EADDRNOTAVAIL;
2224
2225         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2226                 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2227                         return -EADDRNOTAVAIL;
2228
2229                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2230                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2231                         return -EADDRINUSE;
2232         }
2233
2234         cma_bind_port(bind_list, id_priv);
2235         return 0;
2236 }
2237
2238 static int cma_get_tcp_port(struct rdma_id_private *id_priv)
2239 {
2240         int ret;
2241         int size;
2242         struct socket *sock;
2243
2244         ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2245         if (ret)
2246                 return ret;
2247 #ifdef __linux__
2248         ret = sock->ops->bind(sock,
2249                         (struct sockaddr *) &id_priv->id.route.addr.src_addr,
2250                         ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2251 #else
2252         ret = -sobind(sock,
2253                         (struct sockaddr *)&id_priv->id.route.addr.src_addr,
2254                         curthread);
2255 #endif
2256         if (ret) {
2257                 sock_release(sock);
2258                 return ret;
2259         }
2260
2261         size = ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr);
2262         ret = sock_getname(sock,
2263                         (struct sockaddr *) &id_priv->id.route.addr.src_addr,
2264                         &size, 0);
2265         if (ret) {
2266                 sock_release(sock);
2267                 return ret;
2268         }
2269
2270         id_priv->sock = sock;
2271         return 0;
2272 }
2273
2274 static int cma_get_port(struct rdma_id_private *id_priv)
2275 {
2276         struct idr *ps;
2277         int ret;
2278
2279         switch (id_priv->id.ps) {
2280         case RDMA_PS_SDP:
2281                 ps = &sdp_ps;
2282                 break;
2283         case RDMA_PS_TCP:
2284                 ps = &tcp_ps;
2285                 if (unify_tcp_port_space) {
2286                         ret = cma_get_tcp_port(id_priv);
2287                         if (ret)
2288                                 goto out;
2289                 }
2290                 break;
2291         case RDMA_PS_UDP:
2292                 ps = &udp_ps;
2293                 break;
2294         case RDMA_PS_IPOIB:
2295                 ps = &ipoib_ps;
2296                 break;
2297         default:
2298                 return -EPROTONOSUPPORT;
2299         }
2300
2301         mutex_lock(&lock);
2302         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2303                 ret = cma_alloc_any_port(ps, id_priv);
2304         else
2305                 ret = cma_use_port(ps, id_priv);
2306         mutex_unlock(&lock);
2307 out:
2308         return ret;
2309 }
2310
2311 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2312                                struct sockaddr *addr)
2313 {
2314 #if defined(INET6)
2315         struct sockaddr_in6 *sin6;
2316
2317         if (addr->sa_family != AF_INET6)
2318                 return 0;
2319
2320         sin6 = (struct sockaddr_in6 *) addr;
2321 #ifdef __linux__
2322         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2323 #else
2324         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr) &&
2325 #endif
2326             !sin6->sin6_scope_id)
2327                         return -EINVAL;
2328
2329         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2330 #endif
2331         return 0;
2332 }
2333
2334 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2335 {
2336         struct rdma_id_private *id_priv;
2337         int ret;
2338
2339         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2340                 return -EAFNOSUPPORT;
2341
2342         id_priv = container_of(id, struct rdma_id_private, id);
2343         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2344                 return -EINVAL;
2345
2346         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2347         if (ret)
2348                 goto err1;
2349
2350         if (!cma_any_addr(addr)) {
2351                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2352                 if (ret)
2353                         goto err1;
2354
2355                 mutex_lock(&lock);
2356                 ret = cma_acquire_dev(id_priv);
2357                 mutex_unlock(&lock);
2358                 if (ret)
2359                         goto err1;
2360         }
2361
2362         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2363         ret = cma_get_port(id_priv);
2364         if (ret)
2365                 goto err2;
2366
2367         return 0;
2368 err2:
2369         if (id_priv->cma_dev) {
2370                 mutex_lock(&lock);
2371                 cma_detach_from_dev(id_priv);
2372                 mutex_unlock(&lock);
2373         }
2374 err1:
2375         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2376         return ret;
2377 }
2378 EXPORT_SYMBOL(rdma_bind_addr);
2379
2380 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2381                           struct rdma_route *route)
2382 {
2383         struct cma_hdr *cma_hdr;
2384         struct sdp_hh *sdp_hdr;
2385
2386         if (route->addr.src_addr.ss_family == AF_INET) {
2387                 struct sockaddr_in *src4, *dst4;
2388
2389                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2390                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2391
2392                 switch (ps) {
2393                 case RDMA_PS_SDP:
2394                         sdp_hdr = hdr;
2395                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2396                                 return -EINVAL;
2397                         sdp_set_ip_ver(sdp_hdr, 4);
2398                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2399                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2400                         sdp_hdr->port = src4->sin_port;
2401                         break;
2402                 default:
2403                         cma_hdr = hdr;
2404                         cma_hdr->cma_version = CMA_VERSION;
2405                         cma_set_ip_ver(cma_hdr, 4);
2406                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2407                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2408                         cma_hdr->port = src4->sin_port;
2409                         break;
2410                 }
2411         } else {
2412                 struct sockaddr_in6 *src6, *dst6;
2413
2414                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2415                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2416
2417                 switch (ps) {
2418                 case RDMA_PS_SDP:
2419                         sdp_hdr = hdr;
2420                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2421                                 return -EINVAL;
2422                         sdp_set_ip_ver(sdp_hdr, 6);
2423                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2424                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2425                         sdp_hdr->port = src6->sin6_port;
2426                         break;
2427                 default:
2428                         cma_hdr = hdr;
2429                         cma_hdr->cma_version = CMA_VERSION;
2430                         cma_set_ip_ver(cma_hdr, 6);
2431                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2432                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2433                         cma_hdr->port = src6->sin6_port;
2434                         break;
2435                 }
2436         }
2437         return 0;
2438 }
2439
2440 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2441                                 struct ib_cm_event *ib_event)
2442 {
2443         struct rdma_id_private *id_priv = cm_id->context;
2444         struct rdma_cm_event event;
2445         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2446         int ret = 0;
2447
2448         if (cma_disable_callback(id_priv, CMA_CONNECT))
2449                 return 0;
2450
2451         memset(&event, 0, sizeof event);
2452         switch (ib_event->event) {
2453         case IB_CM_SIDR_REQ_ERROR:
2454                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2455                 event.status = -ETIMEDOUT;
2456                 break;
2457         case IB_CM_SIDR_REP_RECEIVED:
2458                 event.param.ud.private_data = ib_event->private_data;
2459                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2460                 if (rep->status != IB_SIDR_SUCCESS) {
2461                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2462                         event.status = ib_event->param.sidr_rep_rcvd.status;
2463                         break;
2464                 }
2465                 ret = cma_set_qkey(id_priv);
2466                 if (ret) {
2467                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2468                         event.status = -EINVAL;
2469                         break;
2470                 }
2471                 if (id_priv->qkey != rep->qkey) {
2472                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2473                         event.status = -EINVAL;
2474                         break;
2475                 }
2476                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2477                                      id_priv->id.route.path_rec,
2478                                      &event.param.ud.ah_attr);
2479                 event.param.ud.qp_num = rep->qpn;
2480                 event.param.ud.qkey = rep->qkey;
2481                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2482                 event.status = 0;
2483                 break;
2484         default:
2485                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2486                        ib_event->event);
2487                 goto out;
2488         }
2489
2490         ret = id_priv->id.event_handler(&id_priv->id, &event);
2491         if (ret) {
2492                 /* Destroy the CM ID by returning a non-zero value. */
2493                 id_priv->cm_id.ib = NULL;
2494                 cma_exch(id_priv, CMA_DESTROYING);
2495                 mutex_unlock(&id_priv->handler_mutex);
2496                 rdma_destroy_id(&id_priv->id);
2497                 return ret;
2498         }
2499 out:
2500         mutex_unlock(&id_priv->handler_mutex);
2501         return ret;
2502 }
2503
2504 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2505                               struct rdma_conn_param *conn_param)
2506 {
2507         struct ib_cm_sidr_req_param req;
2508         struct rdma_route *route;
2509         int ret;
2510
2511         req.private_data_len = sizeof(struct cma_hdr) +
2512                                conn_param->private_data_len;
2513         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2514         if (!req.private_data)
2515                 return -ENOMEM;
2516
2517         if (conn_param->private_data && conn_param->private_data_len)
2518                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2519                        conn_param->private_data, conn_param->private_data_len);
2520
2521         route = &id_priv->id.route;
2522         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2523         if (ret)
2524                 goto out;
2525
2526         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2527                                             cma_sidr_rep_handler, id_priv);
2528         if (IS_ERR(id_priv->cm_id.ib)) {
2529                 ret = PTR_ERR(id_priv->cm_id.ib);
2530                 goto out;
2531         }
2532
2533         req.path = route->path_rec;
2534         req.service_id = cma_get_service_id(id_priv->id.ps,
2535                                             (struct sockaddr *) &route->addr.dst_addr);
2536         req.timeout_ms = 1 << (cma_response_timeout - 8);
2537         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2538
2539         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2540         if (ret) {
2541                 ib_destroy_cm_id(id_priv->cm_id.ib);
2542                 id_priv->cm_id.ib = NULL;
2543         }
2544 out:
2545         kfree(req.private_data);
2546         return ret;
2547 }
2548
2549 static int cma_connect_ib(struct rdma_id_private *id_priv,
2550                           struct rdma_conn_param *conn_param)
2551 {
2552         struct ib_cm_req_param req;
2553         struct rdma_route *route;
2554         void *private_data;
2555         int offset, ret;
2556
2557         memset(&req, 0, sizeof req);
2558         offset = cma_user_data_offset(id_priv->id.ps);
2559         req.private_data_len = offset + conn_param->private_data_len;
2560         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2561         if (!private_data)
2562                 return -ENOMEM;
2563
2564         if (conn_param->private_data && conn_param->private_data_len)
2565                 memcpy(private_data + offset, conn_param->private_data,
2566                        conn_param->private_data_len);
2567
2568         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2569                                             id_priv);
2570         if (IS_ERR(id_priv->cm_id.ib)) {
2571                 ret = PTR_ERR(id_priv->cm_id.ib);
2572                 goto out;
2573         }
2574
2575         route = &id_priv->id.route;
2576         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2577         if (ret)
2578                 goto out;
2579         req.private_data = private_data;
2580
2581         req.primary_path = &route->path_rec[0];
2582         if (route->num_paths == 2)
2583                 req.alternate_path = &route->path_rec[1];
2584
2585         req.service_id = cma_get_service_id(id_priv->id.ps,
2586                                             (struct sockaddr *) &route->addr.dst_addr);
2587         req.qp_num = id_priv->qp_num;
2588         req.qp_type = IB_QPT_RC;
2589         req.starting_psn = id_priv->seq_num;
2590         req.responder_resources = conn_param->responder_resources;
2591         req.initiator_depth = conn_param->initiator_depth;
2592         req.flow_control = conn_param->flow_control;
2593         req.retry_count = conn_param->retry_count;
2594         req.rnr_retry_count = conn_param->rnr_retry_count;
2595        req.remote_cm_response_timeout = cma_response_timeout;
2596        req.local_cm_response_timeout = cma_response_timeout;
2597         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2598         req.srq = id_priv->srq ? 1 : 0;
2599
2600         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2601 out:
2602         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2603                 ib_destroy_cm_id(id_priv->cm_id.ib);
2604                 id_priv->cm_id.ib = NULL;
2605         }
2606
2607         kfree(private_data);
2608         return ret;
2609 }
2610
2611 static int cma_connect_iw(struct rdma_id_private *id_priv,
2612                           struct rdma_conn_param *conn_param)
2613 {
2614         struct iw_cm_id *cm_id;
2615         struct sockaddr_in* sin;
2616         int ret;
2617         struct iw_cm_conn_param iw_param;
2618
2619         cm_id = iw_create_cm_id(id_priv->id.device, id_priv->sock,
2620                                 cma_iw_handler, id_priv);
2621         if (IS_ERR(cm_id)) {
2622                 ret = PTR_ERR(cm_id);
2623                 goto out;
2624         }
2625
2626         id_priv->cm_id.iw = cm_id;
2627
2628         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2629         cm_id->local_addr = *sin;
2630
2631         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2632         cm_id->remote_addr = *sin;
2633
2634         ret = cma_modify_qp_rtr(id_priv, conn_param);
2635         if (ret)
2636                 goto out;
2637
2638         iw_param.ord = conn_param->initiator_depth;
2639         iw_param.ird = conn_param->responder_resources;
2640         iw_param.private_data = conn_param->private_data;
2641         iw_param.private_data_len = conn_param->private_data_len;
2642         if (id_priv->id.qp)
2643                 iw_param.qpn = id_priv->qp_num;
2644         else
2645                 iw_param.qpn = conn_param->qp_num;
2646         ret = iw_cm_connect(cm_id, &iw_param);
2647 out:
2648         if (ret && !IS_ERR(cm_id)) {
2649                 iw_destroy_cm_id(cm_id);
2650                 id_priv->cm_id.iw = NULL;
2651         }
2652         return ret;
2653 }
2654
2655 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2656 {
2657         struct rdma_id_private *id_priv;
2658         int ret;
2659
2660         id_priv = container_of(id, struct rdma_id_private, id);
2661         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2662                 return -EINVAL;
2663
2664         if (!id->qp) {
2665                 id_priv->qp_num = conn_param->qp_num;
2666                 id_priv->srq = conn_param->srq;
2667         }
2668
2669         switch (rdma_node_get_transport(id->device->node_type)) {
2670         case RDMA_TRANSPORT_IB:
2671                 if (cma_is_ud_ps(id->ps))
2672                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2673                 else
2674                         ret = cma_connect_ib(id_priv, conn_param);
2675                 break;
2676         case RDMA_TRANSPORT_IWARP:
2677                 ret = cma_connect_iw(id_priv, conn_param);
2678                 break;
2679         default:
2680                 ret = -ENOSYS;
2681                 break;
2682         }
2683         if (ret)
2684                 goto err;
2685
2686         return 0;
2687 err:
2688         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2689         return ret;
2690 }
2691 EXPORT_SYMBOL(rdma_connect);
2692
2693 static int cma_accept_ib(struct rdma_id_private *id_priv,
2694                          struct rdma_conn_param *conn_param)
2695 {
2696         struct ib_cm_rep_param rep;
2697         int ret;
2698
2699         ret = cma_modify_qp_rtr(id_priv, conn_param);
2700         if (ret)
2701                 goto out;
2702
2703         ret = cma_modify_qp_rts(id_priv, conn_param);
2704         if (ret)
2705                 goto out;
2706
2707         memset(&rep, 0, sizeof rep);
2708         rep.qp_num = id_priv->qp_num;
2709         rep.starting_psn = id_priv->seq_num;
2710         rep.private_data = conn_param->private_data;
2711         rep.private_data_len = conn_param->private_data_len;
2712         rep.responder_resources = conn_param->responder_resources;
2713         rep.initiator_depth = conn_param->initiator_depth;
2714         rep.failover_accepted = 0;
2715         rep.flow_control = conn_param->flow_control;
2716         rep.rnr_retry_count = conn_param->rnr_retry_count;
2717         rep.srq = id_priv->srq ? 1 : 0;
2718
2719         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2720 out:
2721         return ret;
2722 }
2723
2724 static int cma_accept_iw(struct rdma_id_private *id_priv,
2725                   struct rdma_conn_param *conn_param)
2726 {
2727         struct iw_cm_conn_param iw_param;
2728         int ret;
2729
2730         ret = cma_modify_qp_rtr(id_priv, conn_param);
2731         if (ret)
2732                 return ret;
2733
2734         iw_param.ord = conn_param->initiator_depth;
2735         iw_param.ird = conn_param->responder_resources;
2736         iw_param.private_data = conn_param->private_data;
2737         iw_param.private_data_len = conn_param->private_data_len;
2738         if (id_priv->id.qp) {
2739                 iw_param.qpn = id_priv->qp_num;
2740         } else
2741                 iw_param.qpn = conn_param->qp_num;
2742
2743         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2744 }
2745
2746 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2747                              enum ib_cm_sidr_status status,
2748                              const void *private_data, int private_data_len)
2749 {
2750         struct ib_cm_sidr_rep_param rep;
2751         int ret;
2752
2753         memset(&rep, 0, sizeof rep);
2754         rep.status = status;
2755         if (status == IB_SIDR_SUCCESS) {
2756                 ret = cma_set_qkey(id_priv);
2757                 if (ret)
2758                         return ret;
2759                 rep.qp_num = id_priv->qp_num;
2760                 rep.qkey = id_priv->qkey;
2761         }
2762         rep.private_data = private_data;
2763         rep.private_data_len = private_data_len;
2764
2765         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2766 }
2767
2768 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2769 {
2770         struct rdma_id_private *id_priv;
2771         int ret;
2772
2773         id_priv = container_of(id, struct rdma_id_private, id);
2774         if (!cma_comp(id_priv, CMA_CONNECT))
2775                 return -EINVAL;
2776
2777         if (!id->qp && conn_param) {
2778                 id_priv->qp_num = conn_param->qp_num;
2779                 id_priv->srq = conn_param->srq;
2780         }
2781
2782         switch (rdma_node_get_transport(id->device->node_type)) {
2783         case RDMA_TRANSPORT_IB:
2784                 if (cma_is_ud_ps(id->ps))
2785                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2786                                                 conn_param->private_data,
2787                                                 conn_param->private_data_len);
2788                 else if (conn_param)
2789                         ret = cma_accept_ib(id_priv, conn_param);
2790                 else
2791                         ret = cma_rep_recv(id_priv);
2792                 break;
2793         case RDMA_TRANSPORT_IWARP:
2794                 ret = cma_accept_iw(id_priv, conn_param);
2795                 break;
2796         default:
2797                 ret = -ENOSYS;
2798                 break;
2799         }
2800
2801         if (ret)
2802                 goto reject;
2803
2804         return 0;
2805 reject:
2806         cma_modify_qp_err(id_priv);
2807         rdma_reject(id, NULL, 0);
2808         return ret;
2809 }
2810 EXPORT_SYMBOL(rdma_accept);
2811
2812 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2813 {
2814         struct rdma_id_private *id_priv;
2815         int ret;
2816
2817         id_priv = container_of(id, struct rdma_id_private, id);
2818         if (!cma_has_cm_dev(id_priv))
2819                 return -EINVAL;
2820
2821         switch (id->device->node_type) {
2822         case RDMA_NODE_IB_CA:
2823                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2824                 break;
2825         default:
2826                 ret = 0;
2827                 break;
2828         }
2829         return ret;
2830 }
2831 EXPORT_SYMBOL(rdma_notify);
2832
2833 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2834                 u8 private_data_len)
2835 {
2836         struct rdma_id_private *id_priv;
2837         int ret;
2838
2839         id_priv = container_of(id, struct rdma_id_private, id);
2840         if (!cma_has_cm_dev(id_priv))
2841                 return -EINVAL;
2842
2843         switch (rdma_node_get_transport(id->device->node_type)) {
2844         case RDMA_TRANSPORT_IB:
2845                 if (cma_is_ud_ps(id->ps))
2846                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2847                                                 private_data, private_data_len);
2848                 else
2849                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2850                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2851                                              0, private_data, private_data_len);
2852                 break;
2853         case RDMA_TRANSPORT_IWARP:
2854                 ret = iw_cm_reject(id_priv->cm_id.iw,
2855                                    private_data, private_data_len);
2856                 break;
2857         default:
2858                 ret = -ENOSYS;
2859                 break;
2860         }
2861         return ret;
2862 }
2863 EXPORT_SYMBOL(rdma_reject);
2864
2865 int rdma_disconnect(struct rdma_cm_id *id)
2866 {
2867         struct rdma_id_private *id_priv;
2868         int ret;
2869
2870         id_priv = container_of(id, struct rdma_id_private, id);
2871         if (!cma_has_cm_dev(id_priv))
2872                 return -EINVAL;
2873
2874         switch (rdma_node_get_transport(id->device->node_type)) {
2875         case RDMA_TRANSPORT_IB:
2876                 ret = cma_modify_qp_err(id_priv);
2877                 if (ret)
2878                         goto out;
2879                 /* Initiate or respond to a disconnect. */
2880                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2881                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2882                 break;
2883         case RDMA_TRANSPORT_IWARP:
2884                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2885                 break;
2886         default:
2887                 ret = -EINVAL;
2888                 break;
2889         }
2890 out:
2891         return ret;
2892 }
2893 EXPORT_SYMBOL(rdma_disconnect);
2894
2895 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2896 {
2897         struct rdma_id_private *id_priv;
2898         struct cma_multicast *mc = multicast->context;
2899         struct rdma_cm_event event;
2900         int ret;
2901
2902         id_priv = mc->id_priv;
2903         if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2904             cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2905                 return 0;
2906
2907         mutex_lock(&id_priv->qp_mutex);
2908         if (!status && id_priv->id.qp)
2909                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2910                                          multicast->rec.mlid);
2911         mutex_unlock(&id_priv->qp_mutex);
2912
2913         memset(&event, 0, sizeof event);
2914         event.status = status;
2915         event.param.ud.private_data = mc->context;
2916         if (!status) {
2917                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2918                 ib_init_ah_from_mcmember(id_priv->id.device,
2919                                          id_priv->id.port_num, &multicast->rec,
2920                                          &event.param.ud.ah_attr);
2921                 event.param.ud.qp_num = 0xFFFFFF;
2922                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2923         } else
2924                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2925
2926         ret = id_priv->id.event_handler(&id_priv->id, &event);
2927         if (ret) {
2928                 cma_exch(id_priv, CMA_DESTROYING);
2929                 mutex_unlock(&id_priv->handler_mutex);
2930                 rdma_destroy_id(&id_priv->id);
2931                 return 0;
2932         }
2933
2934         mutex_unlock(&id_priv->handler_mutex);
2935         return 0;
2936 }
2937
2938 static void cma_set_mgid(struct rdma_id_private *id_priv,
2939                          struct sockaddr *addr, union ib_gid *mgid)
2940 {
2941 #if defined(INET) || defined(INET6)
2942         unsigned char mc_map[MAX_ADDR_LEN];
2943         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2944 #endif
2945 #ifdef INET
2946         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2947 #endif
2948 #ifdef INET6
2949         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2950 #endif
2951
2952         if (cma_any_addr(addr)) {
2953                 memset(mgid, 0, sizeof *mgid);
2954 #ifdef INET6
2955         } else if ((addr->sa_family == AF_INET6) &&
2956                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2957                                                                  0xFF10A01B)) {
2958                 /* IPv6 address is an SA assigned MGID. */
2959                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2960         } else if (addr->sa_family == AF_INET6) {
2961                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2962                 if (id_priv->id.ps == RDMA_PS_UDP)
2963                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2964                 *mgid = *(union ib_gid *) (mc_map + 4);
2965 #endif
2966 #ifdef INET
2967         } else {
2968                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2969                 if (id_priv->id.ps == RDMA_PS_UDP)
2970                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2971                 *mgid = *(union ib_gid *) (mc_map + 4);
2972 #endif
2973         }
2974 }
2975
2976 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2977                                  struct cma_multicast *mc)
2978 {
2979         struct ib_sa_mcmember_rec rec;
2980         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2981         ib_sa_comp_mask comp_mask;
2982         int ret;
2983
2984         ib_addr_get_mgid(dev_addr, &rec.mgid);
2985         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2986                                      &rec.mgid, &rec);
2987         if (ret)
2988                 return ret;
2989
2990         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2991         if (id_priv->id.ps == RDMA_PS_UDP)
2992                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2993         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2994         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2995         rec.join_state = 1;
2996
2997         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2998                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2999                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3000                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3001                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3002
3003         if (id_priv->id.ps == RDMA_PS_IPOIB)
3004                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3005                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
3006
3007         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3008                                                 id_priv->id.port_num, &rec,
3009                                                 comp_mask, GFP_KERNEL,
3010                                                 cma_ib_mc_handler, mc);
3011         if (IS_ERR(mc->multicast.ib))
3012                 return PTR_ERR(mc->multicast.ib);
3013
3014         return 0;
3015 }
3016
3017
3018 static void iboe_mcast_work_handler(struct work_struct *work)
3019 {
3020         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3021         struct cma_multicast *mc = mw->mc;
3022         struct ib_sa_multicast *m = mc->multicast.ib;
3023
3024         mc->multicast.ib->context = mc;
3025         cma_ib_mc_handler(0, m);
3026         kref_put(&mc->mcref, release_mc);
3027         kfree(mw);
3028 }
3029
3030 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3031 {
3032         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3033         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3034
3035         if (cma_any_addr(addr)) {
3036                 memset(mgid, 0, sizeof *mgid);
3037         } else if (addr->sa_family == AF_INET6)
3038                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3039         else {
3040                 mgid->raw[0] = 0xff;
3041                 mgid->raw[1] = 0x0e;
3042                 mgid->raw[2] = 0;
3043                 mgid->raw[3] = 0;
3044                 mgid->raw[4] = 0;
3045                 mgid->raw[5] = 0;
3046                 mgid->raw[6] = 0;
3047                 mgid->raw[7] = 0;
3048                 mgid->raw[8] = 0;
3049                 mgid->raw[9] = 0;
3050                 mgid->raw[10] = 0xff;
3051                 mgid->raw[11] = 0xff;
3052                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3053         }
3054 }
3055
3056 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3057                                    struct cma_multicast *mc)
3058 {
3059         struct iboe_mcast_work *work;
3060         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3061         int err;
3062         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3063         struct net_device *ndev = NULL;
3064
3065         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3066                 return -EINVAL;
3067
3068         work = kzalloc(sizeof *work, GFP_KERNEL);
3069         if (!work)
3070                 return -ENOMEM;
3071
3072         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3073         if (!mc->multicast.ib) {
3074                 err = -ENOMEM;
3075                 goto out1;
3076         }
3077
3078         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3079
3080         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3081         if (id_priv->id.ps == RDMA_PS_UDP)
3082                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3083
3084         if (dev_addr->bound_dev_if)
3085                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3086         if (!ndev) {
3087                 err = -ENODEV;
3088                 goto out2;
3089         }
3090
3091         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3092         mc->multicast.ib->rec.hop_limit = 1;
3093 #ifdef __linux__
3094         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3095 #else
3096         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->if_mtu);
3097 #endif
3098         dev_put(ndev);
3099         if (!mc->multicast.ib->rec.mtu) {
3100                 err = -EINVAL;
3101                 goto out2;
3102         }
3103         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3104         work->id = id_priv;
3105         work->mc = mc;
3106         INIT_WORK(&work->work, iboe_mcast_work_handler);
3107         kref_get(&mc->mcref);
3108         queue_work(cma_wq, &work->work);
3109
3110         return 0;
3111
3112 out2:
3113         kfree(mc->multicast.ib);
3114 out1:
3115         kfree(work);
3116         return err;
3117 }
3118
3119 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3120                         void *context)
3121 {
3122         struct rdma_id_private *id_priv;
3123         struct cma_multicast *mc;
3124         int ret;
3125
3126         id_priv = container_of(id, struct rdma_id_private, id);
3127         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
3128             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
3129                 return -EINVAL;
3130
3131         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3132         if (!mc)
3133                 return -ENOMEM;
3134
3135         memcpy(&mc->addr, addr, ip_addr_size(addr));
3136         mc->context = context;
3137         mc->id_priv = id_priv;
3138
3139         spin_lock(&id_priv->lock);
3140         list_add(&mc->list, &id_priv->mc_list);
3141         spin_unlock(&id_priv->lock);
3142
3143         switch (rdma_node_get_transport(id->device->node_type)) {
3144         case RDMA_TRANSPORT_IB:
3145                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3146                 case IB_LINK_LAYER_INFINIBAND:
3147                         ret = cma_join_ib_multicast(id_priv, mc);
3148                         break;
3149                 case IB_LINK_LAYER_ETHERNET:
3150                         kref_init(&mc->mcref);
3151                         ret = cma_iboe_join_multicast(id_priv, mc);
3152                         break;
3153                 default:
3154                         ret = -EINVAL;
3155                 }
3156                 break;
3157         default:
3158                 ret = -ENOSYS;
3159                 break;
3160         }
3161
3162         if (ret) {
3163                 spin_lock_irq(&id_priv->lock);
3164                 list_del(&mc->list);
3165                 spin_unlock_irq(&id_priv->lock);
3166                 kfree(mc);
3167         }
3168
3169         return ret;
3170 }
3171 EXPORT_SYMBOL(rdma_join_multicast);
3172
3173 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3174 {
3175         struct rdma_id_private *id_priv;
3176         struct cma_multicast *mc;
3177
3178         id_priv = container_of(id, struct rdma_id_private, id);
3179         spin_lock_irq(&id_priv->lock);
3180         list_for_each_entry(mc, &id_priv->mc_list, list) {
3181                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3182                         list_del(&mc->list);
3183                         spin_unlock_irq(&id_priv->lock);
3184
3185                         if (id->qp)
3186                                 ib_detach_mcast(id->qp,
3187                                                 &mc->multicast.ib->rec.mgid,
3188                                                 mc->multicast.ib->rec.mlid);
3189                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3190                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3191                                 case IB_LINK_LAYER_INFINIBAND:
3192                                         ib_sa_free_multicast(mc->multicast.ib);
3193                                         kfree(mc);
3194                                         break;
3195                                 case IB_LINK_LAYER_ETHERNET:
3196                                         kref_put(&mc->mcref, release_mc);
3197                                         break;
3198                                 default:
3199                                         break;
3200                                 }
3201                         }
3202                         return;
3203                 }
3204         }
3205         spin_unlock_irq(&id_priv->lock);
3206 }
3207 EXPORT_SYMBOL(rdma_leave_multicast);
3208
3209 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3210 {
3211         struct rdma_dev_addr *dev_addr;
3212         struct cma_ndev_work *work;
3213
3214         dev_addr = &id_priv->id.route.addr.dev_addr;
3215
3216 #ifdef __linux__
3217         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3218             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3219                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3220                        ndev->name, &id_priv->id);
3221 #else
3222         if ((dev_addr->bound_dev_if == ndev->if_index) &&
3223             memcmp(dev_addr->src_dev_addr, IF_LLADDR(ndev), ndev->if_addrlen)) {
3224                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3225                        ndev->if_xname, &id_priv->id);
3226 #endif
3227                 work = kzalloc(sizeof *work, GFP_KERNEL);
3228                 if (!work)
3229                         return -ENOMEM;
3230
3231                 INIT_WORK(&work->work, cma_ndev_work_handler);
3232                 work->id = id_priv;
3233                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3234                 atomic_inc(&id_priv->refcount);
3235                 queue_work(cma_wq, &work->work);
3236         }
3237
3238         return 0;
3239 }
3240
3241 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3242                                void *ctx)
3243 {
3244         struct net_device *ndev = (struct net_device *)ctx;
3245         struct cma_device *cma_dev;
3246         struct rdma_id_private *id_priv;
3247         int ret = NOTIFY_DONE;
3248
3249 #ifdef __linux__
3250         if (dev_net(ndev) != &init_net)
3251                 return NOTIFY_DONE;
3252
3253         if (event != NETDEV_BONDING_FAILOVER)
3254                 return NOTIFY_DONE;
3255
3256         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3257                 return NOTIFY_DONE;
3258 #else
3259         if (event != NETDEV_DOWN && event != NETDEV_UNREGISTER)
3260                 return NOTIFY_DONE;
3261 #endif
3262
3263         mutex_lock(&lock);
3264         list_for_each_entry(cma_dev, &dev_list, list)
3265                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3266                         ret = cma_netdev_change(ndev, id_priv);
3267                         if (ret)
3268                                 goto out;
3269                 }
3270
3271 out:
3272         mutex_unlock(&lock);
3273         return ret;
3274 }
3275
3276 static struct notifier_block cma_nb = {
3277         .notifier_call = cma_netdev_callback
3278 };
3279
3280 static void cma_add_one(struct ib_device *device)
3281 {
3282         struct cma_device *cma_dev;
3283         struct rdma_id_private *id_priv;
3284
3285         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3286         if (!cma_dev)
3287                 return;
3288
3289         cma_dev->device = device;
3290
3291         init_completion(&cma_dev->comp);
3292         atomic_set(&cma_dev->refcount, 1);
3293         INIT_LIST_HEAD(&cma_dev->id_list);
3294         ib_set_client_data(device, &cma_client, cma_dev);
3295
3296         mutex_lock(&lock);
3297         list_add_tail(&cma_dev->list, &dev_list);
3298         list_for_each_entry(id_priv, &listen_any_list, list)
3299                 cma_listen_on_dev(id_priv, cma_dev);
3300         mutex_unlock(&lock);
3301 }
3302
3303 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3304 {
3305         struct rdma_cm_event event;
3306         enum cma_state state;
3307         int ret = 0;
3308
3309         /* Record that we want to remove the device */
3310         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
3311         if (state == CMA_DESTROYING)
3312                 return 0;
3313
3314         cma_cancel_operation(id_priv, state);
3315         mutex_lock(&id_priv->handler_mutex);
3316
3317         /* Check for destruction from another callback. */
3318         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
3319                 goto out;
3320
3321         memset(&event, 0, sizeof event);
3322         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3323         ret = id_priv->id.event_handler(&id_priv->id, &event);
3324 out:
3325         mutex_unlock(&id_priv->handler_mutex);
3326         return ret;
3327 }
3328
3329 static void cma_process_remove(struct cma_device *cma_dev)
3330 {
3331         struct rdma_id_private *id_priv;
3332         int ret;
3333
3334         mutex_lock(&lock);
3335         while (!list_empty(&cma_dev->id_list)) {
3336                 id_priv = list_entry(cma_dev->id_list.next,
3337                                      struct rdma_id_private, list);
3338
3339                 list_del(&id_priv->listen_list);
3340                 list_del_init(&id_priv->list);
3341                 atomic_inc(&id_priv->refcount);
3342                 mutex_unlock(&lock);
3343
3344                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3345                 cma_deref_id(id_priv);
3346                 if (ret)
3347                         rdma_destroy_id(&id_priv->id);
3348
3349                 mutex_lock(&lock);
3350         }
3351         mutex_unlock(&lock);
3352
3353         cma_deref_dev(cma_dev);
3354         wait_for_completion(&cma_dev->comp);
3355 }
3356
3357 static void cma_remove_one(struct ib_device *device)
3358 {
3359         struct cma_device *cma_dev;
3360
3361         cma_dev = ib_get_client_data(device, &cma_client);
3362         if (!cma_dev)
3363                 return;
3364
3365         mutex_lock(&lock);
3366         list_del(&cma_dev->list);
3367         mutex_unlock(&lock);
3368
3369         cma_process_remove(cma_dev);
3370         kfree(cma_dev);
3371 }
3372
3373 static int cma_init(void)
3374 {
3375         int ret;
3376 #if defined(INET)
3377         int low, high, remaining;
3378
3379         get_random_bytes(&next_port, sizeof next_port);
3380         inet_get_local_port_range(&low, &high);
3381         remaining = (high - low) + 1;
3382         next_port = ((unsigned int) next_port % remaining) + low;
3383 #endif
3384
3385         cma_wq = create_singlethread_workqueue("rdma_cm");
3386         if (!cma_wq)
3387                 return -ENOMEM;
3388
3389         ib_sa_register_client(&sa_client);
3390         rdma_addr_register_client(&addr_client);
3391         register_netdevice_notifier(&cma_nb);
3392
3393         ret = ib_register_client(&cma_client);
3394         if (ret)
3395                 goto err;
3396         return 0;
3397
3398 err:
3399         unregister_netdevice_notifier(&cma_nb);
3400         rdma_addr_unregister_client(&addr_client);
3401         ib_sa_unregister_client(&sa_client);
3402         destroy_workqueue(cma_wq);
3403         return ret;
3404 }
3405
3406 static void cma_cleanup(void)
3407 {
3408         ib_unregister_client(&cma_client);
3409         unregister_netdevice_notifier(&cma_nb);
3410         rdma_addr_unregister_client(&addr_client);
3411         ib_sa_unregister_client(&sa_client);
3412         destroy_workqueue(cma_wq);
3413         idr_destroy(&sdp_ps);
3414         idr_destroy(&tcp_ps);
3415         idr_destroy(&udp_ps);
3416         idr_destroy(&ipoib_ps);
3417 }
3418
3419 module_init(cma_init);
3420 module_exit(cma_cleanup);