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 = 0;
  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=0)");
  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 (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                                             iw_conn_req_handler,
1482                                             id_priv);
1483         if (IS_ERR(id_priv->cm_id.iw))
1484                 return PTR_ERR(id_priv->cm_id.iw);
1485
1486         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1487         id_priv->cm_id.iw->local_addr = *sin;
1488
1489         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1490
1491         if (ret) {
1492                 iw_destroy_cm_id(id_priv->cm_id.iw);
1493                 id_priv->cm_id.iw = NULL;
1494         }
1495
1496         return ret;
1497 }
1498
1499 static int cma_listen_handler(struct rdma_cm_id *id,
1500                               struct rdma_cm_event *event)
1501 {
1502         struct rdma_id_private *id_priv = id->context;
1503
1504         id->context = id_priv->id.context;
1505         id->event_handler = id_priv->id.event_handler;
1506         return id_priv->id.event_handler(id, event);
1507 }
1508
1509 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1510                               struct cma_device *cma_dev)
1511 {
1512         struct rdma_id_private *dev_id_priv;
1513         struct rdma_cm_id *id;
1514         int ret;
1515
1516         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1517         if (IS_ERR(id))
1518                 return;
1519
1520         dev_id_priv = container_of(id, struct rdma_id_private, id);
1521
1522         dev_id_priv->state = CMA_ADDR_BOUND;
1523         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1524                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1525
1526         cma_attach_to_dev(dev_id_priv, cma_dev);
1527         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1528         atomic_inc(&id_priv->refcount);
1529         dev_id_priv->internal_id = 1;
1530
1531         ret = rdma_listen(id, id_priv->backlog);
1532         if (ret)
1533                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1534                        "listening on device %s\n", ret, cma_dev->device->name);
1535 }
1536
1537 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1538 {
1539         struct cma_device *cma_dev;
1540
1541         mutex_lock(&lock);
1542         list_add_tail(&id_priv->list, &listen_any_list);
1543         list_for_each_entry(cma_dev, &dev_list, list)
1544                 cma_listen_on_dev(id_priv, cma_dev);
1545         mutex_unlock(&lock);
1546 }
1547
1548 int rdma_listen(struct rdma_cm_id *id, int backlog)
1549 {
1550         struct rdma_id_private *id_priv;
1551         int ret;
1552
1553         id_priv = container_of(id, struct rdma_id_private, id);
1554         if (id_priv->state == CMA_IDLE) {
1555                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
1556                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
1557                 if (ret)
1558                         return ret;
1559         }
1560
1561         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1562                 return -EINVAL;
1563
1564         id_priv->backlog = backlog;
1565         if (id->device) {
1566                 switch (rdma_node_get_transport(id->device->node_type)) {
1567                 case RDMA_TRANSPORT_IB:
1568                         ret = cma_ib_listen(id_priv);
1569                         if (ret)
1570                                 goto err;
1571                         break;
1572                 case RDMA_TRANSPORT_IWARP:
1573                         ret = cma_iw_listen(id_priv, backlog);
1574                         if (ret)
1575                                 goto err;
1576                         break;
1577                 default:
1578                         ret = -ENOSYS;
1579                         goto err;
1580                 }
1581         } else
1582                 cma_listen_on_all(id_priv);
1583
1584         return 0;
1585 err:
1586         id_priv->backlog = 0;
1587         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1588         return ret;
1589 }
1590 EXPORT_SYMBOL(rdma_listen);
1591
1592 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1593 {
1594         struct rdma_id_private *id_priv;
1595
1596         id_priv = container_of(id, struct rdma_id_private, id);
1597         id_priv->tos = (u8) tos;
1598 }
1599 EXPORT_SYMBOL(rdma_set_service_type);
1600
1601 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1602                               void *context)
1603 {
1604         struct cma_work *work = context;
1605         struct rdma_route *route;
1606
1607         route = &work->id->id.route;
1608
1609         if (!status) {
1610                 route->num_paths = 1;
1611                 *route->path_rec = *path_rec;
1612         } else {
1613                 work->old_state = CMA_ROUTE_QUERY;
1614                 work->new_state = CMA_ADDR_RESOLVED;
1615                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1616                 work->event.status = status;
1617         }
1618
1619         queue_work(cma_wq, &work->work);
1620 }
1621
1622 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1623                               struct cma_work *work)
1624 {
1625         struct rdma_addr *addr = &id_priv->id.route.addr;
1626         struct ib_sa_path_rec path_rec;
1627         ib_sa_comp_mask comp_mask;
1628         struct sockaddr_in6 *sin6;
1629
1630         memset(&path_rec, 0, sizeof path_rec);
1631         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1632         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1633         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1634         path_rec.numb_path = 1;
1635         path_rec.reversible = 1;
1636         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1637                                                         (struct sockaddr *) &addr->dst_addr);
1638
1639         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1640                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1641                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1642
1643         if (addr->src_addr.ss_family == AF_INET) {
1644                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1645                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1646         } else {
1647                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1648                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1649                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1650         }
1651
1652         if (tavor_quirk) {
1653                 path_rec.mtu_selector = IB_SA_LT;
1654                 path_rec.mtu = IB_MTU_2048;
1655         }
1656
1657         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1658                                                id_priv->id.port_num, &path_rec,
1659                                                comp_mask, timeout_ms,
1660                                                GFP_KERNEL, cma_query_handler,
1661                                                work, &id_priv->query);
1662
1663         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1664 }
1665
1666 static void cma_work_handler(struct work_struct *_work)
1667 {
1668         struct cma_work *work = container_of(_work, struct cma_work, work);
1669         struct rdma_id_private *id_priv = work->id;
1670         int destroy = 0;
1671
1672         mutex_lock(&id_priv->handler_mutex);
1673         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1674                 goto out;
1675
1676         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1677                 cma_exch(id_priv, CMA_DESTROYING);
1678                 destroy = 1;
1679         }
1680 out:
1681         mutex_unlock(&id_priv->handler_mutex);
1682         cma_deref_id(id_priv);
1683         if (destroy)
1684                 rdma_destroy_id(&id_priv->id);
1685         kfree(work);
1686 }
1687
1688 static void cma_ndev_work_handler(struct work_struct *_work)
1689 {
1690         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1691         struct rdma_id_private *id_priv = work->id;
1692         int destroy = 0;
1693
1694         mutex_lock(&id_priv->handler_mutex);
1695         if (id_priv->state == CMA_DESTROYING ||
1696             id_priv->state == CMA_DEVICE_REMOVAL)
1697                 goto out;
1698
1699         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1700                 cma_exch(id_priv, CMA_DESTROYING);
1701                 destroy = 1;
1702         }
1703
1704 out:
1705         mutex_unlock(&id_priv->handler_mutex);
1706         cma_deref_id(id_priv);
1707         if (destroy)
1708                 rdma_destroy_id(&id_priv->id);
1709         kfree(work);
1710 }
1711
1712 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1713 {
1714         struct rdma_route *route = &id_priv->id.route;
1715         struct cma_work *work;
1716         int ret;
1717
1718         work = kzalloc(sizeof *work, GFP_KERNEL);
1719         if (!work)
1720                 return -ENOMEM;
1721
1722         work->id = id_priv;
1723         INIT_WORK(&work->work, cma_work_handler);
1724         work->old_state = CMA_ROUTE_QUERY;
1725         work->new_state = CMA_ROUTE_RESOLVED;
1726         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1727
1728         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1729         if (!route->path_rec) {
1730                 ret = -ENOMEM;
1731                 goto err1;
1732         }
1733
1734         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1735         if (ret)
1736                 goto err2;
1737
1738         return 0;
1739 err2:
1740         kfree(route->path_rec);
1741         route->path_rec = NULL;
1742 err1:
1743         kfree(work);
1744         return ret;
1745 }
1746
1747 int rdma_set_ib_paths(struct rdma_cm_id *id,
1748                       struct ib_sa_path_rec *path_rec, int num_paths)
1749 {
1750         struct rdma_id_private *id_priv;
1751         int ret;
1752
1753         id_priv = container_of(id, struct rdma_id_private, id);
1754         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1755                 return -EINVAL;
1756
1757         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1758         if (!id->route.path_rec) {
1759                 ret = -ENOMEM;
1760                 goto err;
1761         }
1762
1763         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1764         return 0;
1765 err:
1766         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1767         return ret;
1768 }
1769 EXPORT_SYMBOL(rdma_set_ib_paths);
1770
1771 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1772 {
1773         struct cma_work *work;
1774
1775         work = kzalloc(sizeof *work, GFP_KERNEL);
1776         if (!work)
1777                 return -ENOMEM;
1778
1779         work->id = id_priv;
1780         INIT_WORK(&work->work, cma_work_handler);
1781         work->old_state = CMA_ROUTE_QUERY;
1782         work->new_state = CMA_ROUTE_RESOLVED;
1783         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1784         queue_work(cma_wq, &work->work);
1785         return 0;
1786 }
1787
1788 static u8 tos_to_sl(u8 tos)
1789 {
1790         return def_prec2sl & 7;
1791 }
1792
1793 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1794 {
1795         struct rdma_route *route = &id_priv->id.route;
1796         struct rdma_addr *addr = &route->addr;
1797         struct cma_work *work;
1798         int ret;
1799         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1800         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1801         struct net_device *ndev = NULL;
1802         u16 vid;
1803
1804         if (src_addr->sin_family != dst_addr->sin_family)
1805                 return -EINVAL;
1806
1807         work = kzalloc(sizeof *work, GFP_KERNEL);
1808         if (!work)
1809                 return -ENOMEM;
1810
1811         work->id = id_priv;
1812         INIT_WORK(&work->work, cma_work_handler);
1813
1814         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1815         if (!route->path_rec) {
1816                 ret = -ENOMEM;
1817                 goto err1;
1818         }
1819
1820         route->num_paths = 1;
1821
1822         if (addr->dev_addr.bound_dev_if)
1823                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1824         if (!ndev) {
1825                 ret = -ENODEV;
1826                 goto err2;
1827         }
1828
1829         vid = rdma_vlan_dev_vlan_id(ndev);
1830
1831         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1832         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1833
1834         route->path_rec->hop_limit = 1;
1835         route->path_rec->reversible = 1;
1836         route->path_rec->pkey = cpu_to_be16(0xffff);
1837         route->path_rec->mtu_selector = IB_SA_EQ;
1838         route->path_rec->sl = tos_to_sl(id_priv->tos);
1839
1840 #ifdef __linux__
1841         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1842 #else
1843         route->path_rec->mtu = iboe_get_mtu(ndev->if_mtu);
1844 #endif
1845         route->path_rec->rate_selector = IB_SA_EQ;
1846         route->path_rec->rate = iboe_get_rate(ndev);
1847         dev_put(ndev);
1848         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1849         route->path_rec->packet_life_time = IBOE_PACKET_LIFETIME;
1850         if (!route->path_rec->mtu) {
1851                 ret = -EINVAL;
1852                 goto err2;
1853         }
1854
1855         work->old_state = CMA_ROUTE_QUERY;
1856         work->new_state = CMA_ROUTE_RESOLVED;
1857         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1858         work->event.status = 0;
1859
1860         queue_work(cma_wq, &work->work);
1861
1862         return 0;
1863
1864 err2:
1865         kfree(route->path_rec);
1866         route->path_rec = NULL;
1867 err1:
1868         kfree(work);
1869         return ret;
1870 }
1871
1872 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1873 {
1874         struct rdma_id_private *id_priv;
1875         int ret;
1876
1877         id_priv = container_of(id, struct rdma_id_private, id);
1878         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1879                 return -EINVAL;
1880
1881         atomic_inc(&id_priv->refcount);
1882         switch (rdma_node_get_transport(id->device->node_type)) {
1883         case RDMA_TRANSPORT_IB:
1884                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1885                 case IB_LINK_LAYER_INFINIBAND:
1886                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1887                         break;
1888                 case IB_LINK_LAYER_ETHERNET:
1889                         ret = cma_resolve_iboe_route(id_priv);
1890                         break;
1891                 default:
1892                         ret = -ENOSYS;
1893                 }
1894                 break;
1895         case RDMA_TRANSPORT_IWARP:
1896                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1897                 break;
1898         default:
1899                 ret = -ENOSYS;
1900                 break;
1901         }
1902         if (ret)
1903                 goto err;
1904
1905         return 0;
1906 err:
1907         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1908         cma_deref_id(id_priv);
1909         return ret;
1910 }
1911 EXPORT_SYMBOL(rdma_resolve_route);
1912
1913 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1914 {
1915         struct cma_device *cma_dev;
1916         struct ib_port_attr port_attr;
1917         union ib_gid gid;
1918         u16 pkey;
1919         int ret;
1920         u8 p;
1921
1922         mutex_lock(&lock);
1923         if (list_empty(&dev_list)) {
1924                 ret = -ENODEV;
1925                 goto out;
1926         }
1927         list_for_each_entry(cma_dev, &dev_list, list)
1928                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1929                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1930                             port_attr.state == IB_PORT_ACTIVE)
1931                                 goto port_found;
1932
1933         p = 1;
1934         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1935
1936 port_found:
1937         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1938         if (ret)
1939                 goto out;
1940
1941         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1942         if (ret)
1943                 goto out;
1944
1945         id_priv->id.route.addr.dev_addr.dev_type =
1946                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1947                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1948
1949         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1950         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1951         id_priv->id.port_num = p;
1952         cma_attach_to_dev(id_priv, cma_dev);
1953 out:
1954         mutex_unlock(&lock);
1955         return ret;
1956 }
1957
1958 static void addr_handler(int status, struct sockaddr *src_addr,
1959                          struct rdma_dev_addr *dev_addr, void *context)
1960 {
1961         struct rdma_id_private *id_priv = context;
1962         struct rdma_cm_event event;
1963
1964         memset(&event, 0, sizeof event);
1965         mutex_lock(&id_priv->handler_mutex);
1966
1967         /*
1968          * Grab mutex to block rdma_destroy_id() from removing the device while
1969          * we're trying to acquire it.
1970          */
1971         mutex_lock(&lock);
1972         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1973                 mutex_unlock(&lock);
1974                 goto out;
1975         }
1976
1977         if (!status && !id_priv->cma_dev)
1978                 status = cma_acquire_dev(id_priv);
1979         mutex_unlock(&lock);
1980
1981         if (status) {
1982                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1983                         goto out;
1984                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1985                 event.status = status;
1986         } else {
1987                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1988                        ip_addr_size(src_addr));
1989                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1990         }
1991
1992         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1993                 cma_exch(id_priv, CMA_DESTROYING);
1994                 mutex_unlock(&id_priv->handler_mutex);
1995                 cma_deref_id(id_priv);
1996                 rdma_destroy_id(&id_priv->id);
1997                 return;
1998         }
1999 out:
2000         mutex_unlock(&id_priv->handler_mutex);
2001         cma_deref_id(id_priv);
2002 }
2003
2004 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2005 {
2006         struct cma_work *work;
2007         struct sockaddr *src, *dst;
2008         union ib_gid gid;
2009         int ret;
2010
2011         work = kzalloc(sizeof *work, GFP_KERNEL);
2012         if (!work)
2013                 return -ENOMEM;
2014
2015         if (!id_priv->cma_dev) {
2016                 ret = cma_bind_loopback(id_priv);
2017                 if (ret)
2018                         goto err;
2019         }
2020
2021         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2022         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2023
2024         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2025         if (cma_zero_addr(src)) {
2026                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2027                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2028                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
2029                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
2030                 } else {
2031                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
2032                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
2033                 }
2034         }
2035
2036         work->id = id_priv;
2037         INIT_WORK(&work->work, cma_work_handler);
2038         work->old_state = CMA_ADDR_QUERY;
2039         work->new_state = CMA_ADDR_RESOLVED;
2040         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2041         queue_work(cma_wq, &work->work);
2042         return 0;
2043 err:
2044         kfree(work);
2045         return ret;
2046 }
2047
2048 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2049                          struct sockaddr *dst_addr)
2050 {
2051         if (!src_addr || !src_addr->sa_family) {
2052                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2053                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2054                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2055                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2056                 }
2057         }
2058         return rdma_bind_addr(id, src_addr);
2059 }
2060
2061 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2062                       struct sockaddr *dst_addr, int timeout_ms)
2063 {
2064         struct rdma_id_private *id_priv;
2065         int ret;
2066
2067         id_priv = container_of(id, struct rdma_id_private, id);
2068         if (id_priv->state == CMA_IDLE) {
2069                 ret = cma_bind_addr(id, src_addr, dst_addr);
2070                 if (ret)
2071                         return ret;
2072         }
2073
2074         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
2075                 return -EINVAL;
2076
2077         atomic_inc(&id_priv->refcount);
2078         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2079         if (cma_any_addr(dst_addr))
2080                 ret = cma_resolve_loopback(id_priv);
2081         else
2082                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2083                                       dst_addr, &id->route.addr.dev_addr,
2084                                       timeout_ms, addr_handler, id_priv);
2085         if (ret)
2086                 goto err;
2087
2088         return 0;
2089 err:
2090         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
2091         cma_deref_id(id_priv);
2092         return ret;
2093 }
2094 EXPORT_SYMBOL(rdma_resolve_addr);
2095
2096 static void cma_bind_port(struct rdma_bind_list *bind_list,
2097                           struct rdma_id_private *id_priv)
2098 {
2099         struct sockaddr_in *sin;
2100
2101         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2102         sin->sin_port = htons(bind_list->port);
2103         id_priv->bind_list = bind_list;
2104         hlist_add_head(&id_priv->node, &bind_list->owners);
2105 }
2106
2107 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2108                           unsigned short snum)
2109 {
2110         struct rdma_bind_list *bind_list;
2111         int port, ret;
2112
2113         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2114         if (!bind_list)
2115                 return -ENOMEM;
2116
2117         do {
2118                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2119         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2120
2121         if (ret)
2122                 goto err1;
2123
2124         if (port != snum) {
2125                 ret = -EADDRNOTAVAIL;
2126                 goto err2;
2127         }
2128
2129         bind_list->ps = ps;
2130         bind_list->port = (unsigned short) port;
2131         cma_bind_port(bind_list, id_priv);
2132         return 0;
2133 err2:
2134         idr_remove(ps, port);
2135 err1:
2136         kfree(bind_list);
2137         return ret;
2138 }
2139
2140 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2141 {
2142 #if defined(INET)
2143         struct rdma_bind_list *bind_list;
2144         int port, ret, low, high;
2145
2146         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2147         if (!bind_list)
2148                 return -ENOMEM;
2149
2150 retry:
2151         /* FIXME: add proper port randomization per like inet_csk_get_port */
2152         do {
2153                 ret = idr_get_new_above(ps, bind_list, next_port, &port);
2154         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2155
2156         if (ret)
2157                 goto err1;
2158
2159         inet_get_local_port_range(&low, &high);
2160         if (port > high) {
2161                 if (next_port != low) {
2162                         idr_remove(ps, port);
2163                         next_port = low;
2164                         goto retry;
2165                 }
2166                 ret = -EADDRNOTAVAIL;
2167                 goto err2;
2168         }
2169
2170         if (port == high)
2171                 next_port = low;
2172         else
2173                 next_port = port + 1;
2174
2175         bind_list->ps = ps;
2176         bind_list->port = (unsigned short) port;
2177         cma_bind_port(bind_list, id_priv);
2178         return 0;
2179 err2:
2180         idr_remove(ps, port);
2181 err1:
2182         kfree(bind_list);
2183         return ret;
2184 #else
2185         return -ENOSPC;
2186 #endif
2187 }
2188
2189 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2190 {
2191         struct rdma_id_private *cur_id;
2192         struct sockaddr_in *sin, *cur_sin;
2193         struct rdma_bind_list *bind_list;
2194         struct hlist_node *node;
2195         unsigned short snum;
2196
2197         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2198         snum = ntohs(sin->sin_port);
2199 #ifdef __linux__
2200         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2201                 return -EACCES;
2202 #endif
2203
2204         bind_list = idr_find(ps, snum);
2205         if (!bind_list)
2206                 return cma_alloc_port(ps, id_priv, snum);
2207
2208         /*
2209          * We don't support binding to any address if anyone is bound to
2210          * a specific address on the same port.
2211          */
2212         if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2213                 return -EADDRNOTAVAIL;
2214
2215         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2216                 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2217                         return -EADDRNOTAVAIL;
2218
2219                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2220                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2221                         return -EADDRINUSE;
2222         }
2223
2224         cma_bind_port(bind_list, id_priv);
2225         return 0;
2226 }
2227
2228 static int cma_get_tcp_port(struct rdma_id_private *id_priv)
2229 {
2230         int ret;
2231         int size;
2232         struct socket *sock;
2233
2234         ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2235         if (ret)
2236                 return ret;
2237 #ifdef __linux__
2238         ret = sock->ops->bind(sock,
2239                         (struct sockaddr *) &id_priv->id.route.addr.src_addr,
2240                         ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2241 #else
2242         ret = -sobind(sock,
2243                         (struct sockaddr *)&id_priv->id.route.addr.src_addr,
2244                         curthread);
2245 #endif
2246         if (ret) {
2247                 sock_release(sock);
2248                 return ret;
2249         }
2250         size = ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr);
2251         ret = sock_getname(sock,
2252                         (struct sockaddr *) &id_priv->id.route.addr.src_addr,
2253                         &size, 0);
2254         if (ret) {
2255                 sock_release(sock);
2256                 return ret;
2257         }
2258         id_priv->sock = sock;
2259         return 0;
2260 }
2261
2262 static int cma_get_port(struct rdma_id_private *id_priv)
2263 {
2264         struct idr *ps;
2265         int ret;
2266
2267         switch (id_priv->id.ps) {
2268         case RDMA_PS_SDP:
2269                 ps = &sdp_ps;
2270                 break;
2271         case RDMA_PS_TCP:
2272                 ps = &tcp_ps;
2273                 if (unify_tcp_port_space) {
2274                         ret = cma_get_tcp_port(id_priv);
2275                         if (ret)
2276                                 goto out;
2277                 }
2278                 break;
2279         case RDMA_PS_UDP:
2280                 ps = &udp_ps;
2281                 break;
2282         case RDMA_PS_IPOIB:
2283                 ps = &ipoib_ps;
2284                 break;
2285         default:
2286                 return -EPROTONOSUPPORT;
2287         }
2288
2289         mutex_lock(&lock);
2290         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2291                 ret = cma_alloc_any_port(ps, id_priv);
2292         else
2293                 ret = cma_use_port(ps, id_priv);
2294         mutex_unlock(&lock);
2295 out:
2296         return ret;
2297 }
2298
2299 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2300                                struct sockaddr *addr)
2301 {
2302 #if defined(INET6)
2303         struct sockaddr_in6 *sin6;
2304
2305         if (addr->sa_family != AF_INET6)
2306                 return 0;
2307
2308         sin6 = (struct sockaddr_in6 *) addr;
2309 #ifdef __linux__
2310         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2311 #else
2312         if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr) &&
2313 #endif
2314             !sin6->sin6_scope_id)
2315                         return -EINVAL;
2316
2317         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2318 #endif
2319         return 0;
2320 }
2321
2322 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2323 {
2324         struct rdma_id_private *id_priv;
2325         int ret;
2326
2327         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2328                 return -EAFNOSUPPORT;
2329
2330         id_priv = container_of(id, struct rdma_id_private, id);
2331         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2332                 return -EINVAL;
2333
2334         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2335         if (ret)
2336                 goto err1;
2337
2338         if (!cma_any_addr(addr)) {
2339                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2340                 if (ret)
2341                         goto err1;
2342
2343                 mutex_lock(&lock);
2344                 ret = cma_acquire_dev(id_priv);
2345                 mutex_unlock(&lock);
2346                 if (ret)
2347                         goto err1;
2348         }
2349
2350         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2351         ret = cma_get_port(id_priv);
2352         if (ret)
2353                 goto err2;
2354
2355         return 0;
2356 err2:
2357         if (id_priv->cma_dev) {
2358                 mutex_lock(&lock);
2359                 cma_detach_from_dev(id_priv);
2360                 mutex_unlock(&lock);
2361         }
2362 err1:
2363         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2364         return ret;
2365 }
2366 EXPORT_SYMBOL(rdma_bind_addr);
2367
2368 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2369                           struct rdma_route *route)
2370 {
2371         struct cma_hdr *cma_hdr;
2372         struct sdp_hh *sdp_hdr;
2373
2374         if (route->addr.src_addr.ss_family == AF_INET) {
2375                 struct sockaddr_in *src4, *dst4;
2376
2377                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2378                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2379
2380                 switch (ps) {
2381                 case RDMA_PS_SDP:
2382                         sdp_hdr = hdr;
2383                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2384                                 return -EINVAL;
2385                         sdp_set_ip_ver(sdp_hdr, 4);
2386                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2387                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2388                         sdp_hdr->port = src4->sin_port;
2389                         break;
2390                 default:
2391                         cma_hdr = hdr;
2392                         cma_hdr->cma_version = CMA_VERSION;
2393                         cma_set_ip_ver(cma_hdr, 4);
2394                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2395                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2396                         cma_hdr->port = src4->sin_port;
2397                         break;
2398                 }
2399         } else {
2400                 struct sockaddr_in6 *src6, *dst6;
2401
2402                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2403                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2404
2405                 switch (ps) {
2406                 case RDMA_PS_SDP:
2407                         sdp_hdr = hdr;
2408                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2409                                 return -EINVAL;
2410                         sdp_set_ip_ver(sdp_hdr, 6);
2411                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2412                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2413                         sdp_hdr->port = src6->sin6_port;
2414                         break;
2415                 default:
2416                         cma_hdr = hdr;
2417                         cma_hdr->cma_version = CMA_VERSION;
2418                         cma_set_ip_ver(cma_hdr, 6);
2419                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2420                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2421                         cma_hdr->port = src6->sin6_port;
2422                         break;
2423                 }
2424         }
2425         return 0;
2426 }
2427
2428 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2429                                 struct ib_cm_event *ib_event)
2430 {
2431         struct rdma_id_private *id_priv = cm_id->context;
2432         struct rdma_cm_event event;
2433         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2434         int ret = 0;
2435
2436         if (cma_disable_callback(id_priv, CMA_CONNECT))
2437                 return 0;
2438
2439         memset(&event, 0, sizeof event);
2440         switch (ib_event->event) {
2441         case IB_CM_SIDR_REQ_ERROR:
2442                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2443                 event.status = -ETIMEDOUT;
2444                 break;
2445         case IB_CM_SIDR_REP_RECEIVED:
2446                 event.param.ud.private_data = ib_event->private_data;
2447                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2448                 if (rep->status != IB_SIDR_SUCCESS) {
2449                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2450                         event.status = ib_event->param.sidr_rep_rcvd.status;
2451                         break;
2452                 }
2453                 ret = cma_set_qkey(id_priv);
2454                 if (ret) {
2455                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2456                         event.status = -EINVAL;
2457                         break;
2458                 }
2459                 if (id_priv->qkey != rep->qkey) {
2460                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2461                         event.status = -EINVAL;
2462                         break;
2463                 }
2464                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2465                                      id_priv->id.route.path_rec,
2466                                      &event.param.ud.ah_attr);
2467                 event.param.ud.qp_num = rep->qpn;
2468                 event.param.ud.qkey = rep->qkey;
2469                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2470                 event.status = 0;
2471                 break;
2472         default:
2473                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2474                        ib_event->event);
2475                 goto out;
2476         }
2477
2478         ret = id_priv->id.event_handler(&id_priv->id, &event);
2479         if (ret) {
2480                 /* Destroy the CM ID by returning a non-zero value. */
2481                 id_priv->cm_id.ib = NULL;
2482                 cma_exch(id_priv, CMA_DESTROYING);
2483                 mutex_unlock(&id_priv->handler_mutex);
2484                 rdma_destroy_id(&id_priv->id);
2485                 return ret;
2486         }
2487 out:
2488         mutex_unlock(&id_priv->handler_mutex);
2489         return ret;
2490 }
2491
2492 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2493                               struct rdma_conn_param *conn_param)
2494 {
2495         struct ib_cm_sidr_req_param req;
2496         struct rdma_route *route;
2497         int ret;
2498
2499         req.private_data_len = sizeof(struct cma_hdr) +
2500                                conn_param->private_data_len;
2501         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2502         if (!req.private_data)
2503                 return -ENOMEM;
2504
2505         if (conn_param->private_data && conn_param->private_data_len)
2506                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2507                        conn_param->private_data, conn_param->private_data_len);
2508
2509         route = &id_priv->id.route;
2510         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2511         if (ret)
2512                 goto out;
2513
2514         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2515                                             cma_sidr_rep_handler, id_priv);
2516         if (IS_ERR(id_priv->cm_id.ib)) {
2517                 ret = PTR_ERR(id_priv->cm_id.ib);
2518                 goto out;
2519         }
2520
2521         req.path = route->path_rec;
2522         req.service_id = cma_get_service_id(id_priv->id.ps,
2523                                             (struct sockaddr *) &route->addr.dst_addr);
2524         req.timeout_ms = 1 << (cma_response_timeout - 8);
2525         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2526
2527         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2528         if (ret) {
2529                 ib_destroy_cm_id(id_priv->cm_id.ib);
2530                 id_priv->cm_id.ib = NULL;
2531         }
2532 out:
2533         kfree(req.private_data);
2534         return ret;
2535 }
2536
2537 static int cma_connect_ib(struct rdma_id_private *id_priv,
2538                           struct rdma_conn_param *conn_param)
2539 {
2540         struct ib_cm_req_param req;
2541         struct rdma_route *route;
2542         void *private_data;
2543         int offset, ret;
2544
2545         memset(&req, 0, sizeof req);
2546         offset = cma_user_data_offset(id_priv->id.ps);
2547         req.private_data_len = offset + conn_param->private_data_len;
2548         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2549         if (!private_data)
2550                 return -ENOMEM;
2551
2552         if (conn_param->private_data && conn_param->private_data_len)
2553                 memcpy(private_data + offset, conn_param->private_data,
2554                        conn_param->private_data_len);
2555
2556         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2557                                             id_priv);
2558         if (IS_ERR(id_priv->cm_id.ib)) {
2559                 ret = PTR_ERR(id_priv->cm_id.ib);
2560                 goto out;
2561         }
2562
2563         route = &id_priv->id.route;
2564         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2565         if (ret)
2566                 goto out;
2567         req.private_data = private_data;
2568
2569         req.primary_path = &route->path_rec[0];
2570         if (route->num_paths == 2)
2571                 req.alternate_path = &route->path_rec[1];
2572
2573         req.service_id = cma_get_service_id(id_priv->id.ps,
2574                                             (struct sockaddr *) &route->addr.dst_addr);
2575         req.qp_num = id_priv->qp_num;
2576         req.qp_type = IB_QPT_RC;
2577         req.starting_psn = id_priv->seq_num;
2578         req.responder_resources = conn_param->responder_resources;
2579         req.initiator_depth = conn_param->initiator_depth;
2580         req.flow_control = conn_param->flow_control;
2581         req.retry_count = conn_param->retry_count;
2582         req.rnr_retry_count = conn_param->rnr_retry_count;
2583        req.remote_cm_response_timeout = cma_response_timeout;
2584        req.local_cm_response_timeout = cma_response_timeout;
2585         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2586         req.srq = id_priv->srq ? 1 : 0;
2587
2588         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2589 out:
2590         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2591                 ib_destroy_cm_id(id_priv->cm_id.ib);
2592                 id_priv->cm_id.ib = NULL;
2593         }
2594
2595         kfree(private_data);
2596         return ret;
2597 }
2598
2599 static int cma_connect_iw(struct rdma_id_private *id_priv,
2600                           struct rdma_conn_param *conn_param)
2601 {
2602         struct iw_cm_id *cm_id;
2603         struct sockaddr_in* sin;
2604         int ret;
2605         struct iw_cm_conn_param iw_param;
2606
2607         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2608         if (IS_ERR(cm_id)) {
2609                 ret = PTR_ERR(cm_id);
2610                 goto out;
2611         }
2612
2613         id_priv->cm_id.iw = cm_id;
2614
2615         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2616         cm_id->local_addr = *sin;
2617
2618         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2619         cm_id->remote_addr = *sin;
2620
2621         ret = cma_modify_qp_rtr(id_priv, conn_param);
2622         if (ret)
2623                 goto out;
2624
2625         iw_param.ord = conn_param->initiator_depth;
2626         iw_param.ird = conn_param->responder_resources;
2627         iw_param.private_data = conn_param->private_data;
2628         iw_param.private_data_len = conn_param->private_data_len;
2629         if (id_priv->id.qp)
2630                 iw_param.qpn = id_priv->qp_num;
2631         else
2632                 iw_param.qpn = conn_param->qp_num;
2633         ret = iw_cm_connect(cm_id, &iw_param);
2634 out:
2635         if (ret && !IS_ERR(cm_id)) {
2636                 iw_destroy_cm_id(cm_id);
2637                 id_priv->cm_id.iw = NULL;
2638         }
2639         return ret;
2640 }
2641
2642 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2643 {
2644         struct rdma_id_private *id_priv;
2645         int ret;
2646
2647         id_priv = container_of(id, struct rdma_id_private, id);
2648         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2649                 return -EINVAL;
2650
2651         if (!id->qp) {
2652                 id_priv->qp_num = conn_param->qp_num;
2653                 id_priv->srq = conn_param->srq;
2654         }
2655
2656         switch (rdma_node_get_transport(id->device->node_type)) {
2657         case RDMA_TRANSPORT_IB:
2658                 if (cma_is_ud_ps(id->ps))
2659                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2660                 else
2661                         ret = cma_connect_ib(id_priv, conn_param);
2662                 break;
2663         case RDMA_TRANSPORT_IWARP:
2664                 ret = cma_connect_iw(id_priv, conn_param);
2665                 break;
2666         default:
2667                 ret = -ENOSYS;
2668                 break;
2669         }
2670         if (ret)
2671                 goto err;
2672
2673         return 0;
2674 err:
2675         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2676         return ret;
2677 }
2678 EXPORT_SYMBOL(rdma_connect);
2679
2680 static int cma_accept_ib(struct rdma_id_private *id_priv,
2681                          struct rdma_conn_param *conn_param)
2682 {
2683         struct ib_cm_rep_param rep;
2684         int ret;
2685
2686         ret = cma_modify_qp_rtr(id_priv, conn_param);
2687         if (ret)
2688                 goto out;
2689
2690         ret = cma_modify_qp_rts(id_priv, conn_param);
2691         if (ret)
2692                 goto out;
2693
2694         memset(&rep, 0, sizeof rep);
2695         rep.qp_num = id_priv->qp_num;
2696         rep.starting_psn = id_priv->seq_num;
2697         rep.private_data = conn_param->private_data;
2698         rep.private_data_len = conn_param->private_data_len;
2699         rep.responder_resources = conn_param->responder_resources;
2700         rep.initiator_depth = conn_param->initiator_depth;
2701         rep.failover_accepted = 0;
2702         rep.flow_control = conn_param->flow_control;
2703         rep.rnr_retry_count = conn_param->rnr_retry_count;
2704         rep.srq = id_priv->srq ? 1 : 0;
2705
2706         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2707 out:
2708         return ret;
2709 }
2710
2711 static int cma_accept_iw(struct rdma_id_private *id_priv,
2712                   struct rdma_conn_param *conn_param)
2713 {
2714         struct iw_cm_conn_param iw_param;
2715         int ret;
2716
2717         ret = cma_modify_qp_rtr(id_priv, conn_param);
2718         if (ret)
2719                 return ret;
2720
2721         iw_param.ord = conn_param->initiator_depth;
2722         iw_param.ird = conn_param->responder_resources;
2723         iw_param.private_data = conn_param->private_data;
2724         iw_param.private_data_len = conn_param->private_data_len;
2725         if (id_priv->id.qp) {
2726                 iw_param.qpn = id_priv->qp_num;
2727         } else
2728                 iw_param.qpn = conn_param->qp_num;
2729
2730         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2731 }
2732
2733 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2734                              enum ib_cm_sidr_status status,
2735                              const void *private_data, int private_data_len)
2736 {
2737         struct ib_cm_sidr_rep_param rep;
2738         int ret;
2739
2740         memset(&rep, 0, sizeof rep);
2741         rep.status = status;
2742         if (status == IB_SIDR_SUCCESS) {
2743                 ret = cma_set_qkey(id_priv);
2744                 if (ret)
2745                         return ret;
2746                 rep.qp_num = id_priv->qp_num;
2747                 rep.qkey = id_priv->qkey;
2748         }
2749         rep.private_data = private_data;
2750         rep.private_data_len = private_data_len;
2751
2752         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2753 }
2754
2755 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2756 {
2757         struct rdma_id_private *id_priv;
2758         int ret;
2759
2760         id_priv = container_of(id, struct rdma_id_private, id);
2761         if (!cma_comp(id_priv, CMA_CONNECT))
2762                 return -EINVAL;
2763
2764         if (!id->qp && conn_param) {
2765                 id_priv->qp_num = conn_param->qp_num;
2766                 id_priv->srq = conn_param->srq;
2767         }
2768
2769         switch (rdma_node_get_transport(id->device->node_type)) {
2770         case RDMA_TRANSPORT_IB:
2771                 if (cma_is_ud_ps(id->ps))
2772                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2773                                                 conn_param->private_data,
2774                                                 conn_param->private_data_len);
2775                 else if (conn_param)
2776                         ret = cma_accept_ib(id_priv, conn_param);
2777                 else
2778                         ret = cma_rep_recv(id_priv);
2779                 break;
2780         case RDMA_TRANSPORT_IWARP:
2781                 ret = cma_accept_iw(id_priv, conn_param);
2782                 break;
2783         default:
2784                 ret = -ENOSYS;
2785                 break;
2786         }
2787
2788         if (ret)
2789                 goto reject;
2790
2791         return 0;
2792 reject:
2793         cma_modify_qp_err(id_priv);
2794         rdma_reject(id, NULL, 0);
2795         return ret;
2796 }
2797 EXPORT_SYMBOL(rdma_accept);
2798
2799 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2800 {
2801         struct rdma_id_private *id_priv;
2802         int ret;
2803
2804         id_priv = container_of(id, struct rdma_id_private, id);
2805         if (!cma_has_cm_dev(id_priv))
2806                 return -EINVAL;
2807
2808         switch (id->device->node_type) {
2809         case RDMA_NODE_IB_CA:
2810                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2811                 break;
2812         default:
2813                 ret = 0;
2814                 break;
2815         }
2816         return ret;
2817 }
2818 EXPORT_SYMBOL(rdma_notify);
2819
2820 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2821                 u8 private_data_len)
2822 {
2823         struct rdma_id_private *id_priv;
2824         int ret;
2825
2826         id_priv = container_of(id, struct rdma_id_private, id);
2827         if (!cma_has_cm_dev(id_priv))
2828                 return -EINVAL;
2829
2830         switch (rdma_node_get_transport(id->device->node_type)) {
2831         case RDMA_TRANSPORT_IB:
2832                 if (cma_is_ud_ps(id->ps))
2833                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2834                                                 private_data, private_data_len);
2835                 else
2836                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2837                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2838                                              0, private_data, private_data_len);
2839                 break;
2840         case RDMA_TRANSPORT_IWARP:
2841                 ret = iw_cm_reject(id_priv->cm_id.iw,
2842                                    private_data, private_data_len);
2843                 break;
2844         default:
2845                 ret = -ENOSYS;
2846                 break;
2847         }
2848         return ret;
2849 }
2850 EXPORT_SYMBOL(rdma_reject);
2851
2852 int rdma_disconnect(struct rdma_cm_id *id)
2853 {
2854         struct rdma_id_private *id_priv;
2855         int ret;
2856
2857         id_priv = container_of(id, struct rdma_id_private, id);
2858         if (!cma_has_cm_dev(id_priv))
2859                 return -EINVAL;
2860
2861         switch (rdma_node_get_transport(id->device->node_type)) {
2862         case RDMA_TRANSPORT_IB:
2863                 ret = cma_modify_qp_err(id_priv);
2864                 if (ret)
2865                         goto out;
2866                 /* Initiate or respond to a disconnect. */
2867                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2868                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2869                 break;
2870         case RDMA_TRANSPORT_IWARP:
2871                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2872                 break;
2873         default:
2874                 ret = -EINVAL;
2875                 break;
2876         }
2877 out:
2878         return ret;
2879 }
2880 EXPORT_SYMBOL(rdma_disconnect);
2881
2882 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2883 {
2884         struct rdma_id_private *id_priv;
2885         struct cma_multicast *mc = multicast->context;
2886         struct rdma_cm_event event;
2887         int ret;
2888
2889         id_priv = mc->id_priv;
2890         if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2891             cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2892                 return 0;
2893
2894         mutex_lock(&id_priv->qp_mutex);
2895         if (!status && id_priv->id.qp)
2896                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2897                                          multicast->rec.mlid);
2898         mutex_unlock(&id_priv->qp_mutex);
2899
2900         memset(&event, 0, sizeof event);
2901         event.status = status;
2902         event.param.ud.private_data = mc->context;
2903         if (!status) {
2904                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2905                 ib_init_ah_from_mcmember(id_priv->id.device,
2906                                          id_priv->id.port_num, &multicast->rec,
2907                                          &event.param.ud.ah_attr);
2908                 event.param.ud.qp_num = 0xFFFFFF;
2909                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2910         } else
2911                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2912
2913         ret = id_priv->id.event_handler(&id_priv->id, &event);
2914         if (ret) {
2915                 cma_exch(id_priv, CMA_DESTROYING);
2916                 mutex_unlock(&id_priv->handler_mutex);
2917                 rdma_destroy_id(&id_priv->id);
2918                 return 0;
2919         }
2920
2921         mutex_unlock(&id_priv->handler_mutex);
2922         return 0;
2923 }
2924
2925 static void cma_set_mgid(struct rdma_id_private *id_priv,
2926                          struct sockaddr *addr, union ib_gid *mgid)
2927 {
2928 #if defined(INET) || defined(INET6)
2929         unsigned char mc_map[MAX_ADDR_LEN];
2930         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2931 #endif
2932 #ifdef INET
2933         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2934 #endif
2935 #ifdef INET6
2936         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2937 #endif
2938
2939         if (cma_any_addr(addr)) {
2940                 memset(mgid, 0, sizeof *mgid);
2941 #ifdef INET6
2942         } else if ((addr->sa_family == AF_INET6) &&
2943                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2944                                                                  0xFF10A01B)) {
2945                 /* IPv6 address is an SA assigned MGID. */
2946                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2947         } else if ((addr->sa_family == AF_INET6)) {
2948                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2949                 if (id_priv->id.ps == RDMA_PS_UDP)
2950                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2951                 *mgid = *(union ib_gid *) (mc_map + 4);
2952 #endif
2953 #ifdef INET
2954         } else {
2955                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2956                 if (id_priv->id.ps == RDMA_PS_UDP)
2957                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2958                 *mgid = *(union ib_gid *) (mc_map + 4);
2959 #endif
2960         }
2961 }
2962
2963 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2964                                  struct cma_multicast *mc)
2965 {
2966         struct ib_sa_mcmember_rec rec;
2967         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2968         ib_sa_comp_mask comp_mask;
2969         int ret;
2970
2971         ib_addr_get_mgid(dev_addr, &rec.mgid);
2972         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2973                                      &rec.mgid, &rec);
2974         if (ret)
2975                 return ret;
2976
2977         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2978         if (id_priv->id.ps == RDMA_PS_UDP)
2979                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2980         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2981         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2982         rec.join_state = 1;
2983
2984         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2985                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2986                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2987                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2988                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2989
2990         if (id_priv->id.ps == RDMA_PS_IPOIB)
2991                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2992                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2993
2994         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2995                                                 id_priv->id.port_num, &rec,
2996                                                 comp_mask, GFP_KERNEL,
2997                                                 cma_ib_mc_handler, mc);
2998         if (IS_ERR(mc->multicast.ib))
2999                 return PTR_ERR(mc->multicast.ib);
3000
3001         return 0;
3002 }
3003
3004
3005 static void iboe_mcast_work_handler(struct work_struct *work)
3006 {
3007         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3008         struct cma_multicast *mc = mw->mc;
3009         struct ib_sa_multicast *m = mc->multicast.ib;
3010
3011         mc->multicast.ib->context = mc;
3012         cma_ib_mc_handler(0, m);
3013         kref_put(&mc->mcref, release_mc);
3014         kfree(mw);
3015 }
3016
3017 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3018 {
3019         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3020         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3021
3022         if (cma_any_addr(addr)) {
3023                 memset(mgid, 0, sizeof *mgid);
3024         } else if (addr->sa_family == AF_INET6)
3025                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3026         else {
3027                 mgid->raw[0] = 0xff;
3028                 mgid->raw[1] = 0x0e;
3029                 mgid->raw[2] = 0;
3030                 mgid->raw[3] = 0;
3031                 mgid->raw[4] = 0;
3032                 mgid->raw[5] = 0;
3033                 mgid->raw[6] = 0;
3034                 mgid->raw[7] = 0;
3035                 mgid->raw[8] = 0;
3036                 mgid->raw[9] = 0;
3037                 mgid->raw[10] = 0xff;
3038                 mgid->raw[11] = 0xff;
3039                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3040         }
3041 }
3042
3043 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3044                                    struct cma_multicast *mc)
3045 {
3046         struct iboe_mcast_work *work;
3047         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3048         int err;
3049         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3050         struct net_device *ndev = NULL;
3051
3052         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3053                 return -EINVAL;
3054
3055         work = kzalloc(sizeof *work, GFP_KERNEL);
3056         if (!work)
3057                 return -ENOMEM;
3058
3059         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3060         if (!mc->multicast.ib) {
3061                 err = -ENOMEM;
3062                 goto out1;
3063         }
3064
3065         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3066
3067         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3068         if (id_priv->id.ps == RDMA_PS_UDP)
3069                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3070
3071         if (dev_addr->bound_dev_if)
3072                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3073         if (!ndev) {
3074                 err = -ENODEV;
3075                 goto out2;
3076         }
3077
3078         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3079         mc->multicast.ib->rec.hop_limit = 1;
3080 #ifdef __linux__
3081         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3082 #else
3083         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->if_mtu);
3084 #endif
3085         dev_put(ndev);
3086         if (!mc->multicast.ib->rec.mtu) {
3087                 err = -EINVAL;
3088                 goto out2;
3089         }
3090         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3091         work->id = id_priv;
3092         work->mc = mc;
3093         INIT_WORK(&work->work, iboe_mcast_work_handler);
3094         kref_get(&mc->mcref);
3095         queue_work(cma_wq, &work->work);
3096
3097         return 0;
3098
3099 out2:
3100         kfree(mc->multicast.ib);
3101 out1:
3102         kfree(work);
3103         return err;
3104 }
3105
3106 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3107                         void *context)
3108 {
3109         struct rdma_id_private *id_priv;
3110         struct cma_multicast *mc;
3111         int ret;
3112
3113         id_priv = container_of(id, struct rdma_id_private, id);
3114         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
3115             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
3116                 return -EINVAL;
3117
3118         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3119         if (!mc)
3120                 return -ENOMEM;
3121
3122         memcpy(&mc->addr, addr, ip_addr_size(addr));
3123         mc->context = context;
3124         mc->id_priv = id_priv;
3125
3126         spin_lock(&id_priv->lock);
3127         list_add(&mc->list, &id_priv->mc_list);
3128         spin_unlock(&id_priv->lock);
3129
3130         switch (rdma_node_get_transport(id->device->node_type)) {
3131         case RDMA_TRANSPORT_IB:
3132                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3133                 case IB_LINK_LAYER_INFINIBAND:
3134                         ret = cma_join_ib_multicast(id_priv, mc);
3135                         break;
3136                 case IB_LINK_LAYER_ETHERNET:
3137                         kref_init(&mc->mcref);
3138                         ret = cma_iboe_join_multicast(id_priv, mc);
3139                         break;
3140                 default:
3141                         ret = -EINVAL;
3142                 }
3143                 break;
3144         default:
3145                 ret = -ENOSYS;
3146                 break;
3147         }
3148
3149         if (ret) {
3150                 spin_lock_irq(&id_priv->lock);
3151                 list_del(&mc->list);
3152                 spin_unlock_irq(&id_priv->lock);
3153                 kfree(mc);
3154         }
3155
3156         return ret;
3157 }
3158 EXPORT_SYMBOL(rdma_join_multicast);
3159
3160 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3161 {
3162         struct rdma_id_private *id_priv;
3163         struct cma_multicast *mc;
3164
3165         id_priv = container_of(id, struct rdma_id_private, id);
3166         spin_lock_irq(&id_priv->lock);
3167         list_for_each_entry(mc, &id_priv->mc_list, list) {
3168                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3169                         list_del(&mc->list);
3170                         spin_unlock_irq(&id_priv->lock);
3171
3172                         if (id->qp)
3173                                 ib_detach_mcast(id->qp,
3174                                                 &mc->multicast.ib->rec.mgid,
3175                                                 mc->multicast.ib->rec.mlid);
3176                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3177                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3178                                 case IB_LINK_LAYER_INFINIBAND:
3179                                         ib_sa_free_multicast(mc->multicast.ib);
3180                                         kfree(mc);
3181                                         break;
3182                                 case IB_LINK_LAYER_ETHERNET:
3183                                         kref_put(&mc->mcref, release_mc);
3184                                         break;
3185                                 default:
3186                                         break;
3187                                 }
3188                         }
3189                         return;
3190                 }
3191         }
3192         spin_unlock_irq(&id_priv->lock);
3193 }
3194 EXPORT_SYMBOL(rdma_leave_multicast);
3195
3196 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3197 {
3198         struct rdma_dev_addr *dev_addr;
3199         struct cma_ndev_work *work;
3200
3201         dev_addr = &id_priv->id.route.addr.dev_addr;
3202
3203 #ifdef __linux__
3204         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3205             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3206                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3207                        ndev->name, &id_priv->id);
3208 #else
3209         if ((dev_addr->bound_dev_if == ndev->if_index) &&
3210             memcmp(dev_addr->src_dev_addr, IF_LLADDR(ndev), ndev->if_addrlen)) {
3211                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3212                        ndev->if_xname, &id_priv->id);
3213 #endif
3214                 work = kzalloc(sizeof *work, GFP_KERNEL);
3215                 if (!work)
3216                         return -ENOMEM;
3217
3218                 INIT_WORK(&work->work, cma_ndev_work_handler);
3219                 work->id = id_priv;
3220                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3221                 atomic_inc(&id_priv->refcount);
3222                 queue_work(cma_wq, &work->work);
3223         }
3224
3225         return 0;
3226 }
3227
3228 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3229                                void *ctx)
3230 {
3231         struct net_device *ndev = (struct net_device *)ctx;
3232         struct cma_device *cma_dev;
3233         struct rdma_id_private *id_priv;
3234         int ret = NOTIFY_DONE;
3235
3236 #ifdef __linux__
3237         if (dev_net(ndev) != &init_net)
3238                 return NOTIFY_DONE;
3239
3240         if (event != NETDEV_BONDING_FAILOVER)
3241                 return NOTIFY_DONE;
3242
3243         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3244                 return NOTIFY_DONE;
3245 #else
3246         if (event != NETDEV_DOWN && event != NETDEV_UNREGISTER)
3247                 return NOTIFY_DONE;
3248 #endif
3249
3250         mutex_lock(&lock);
3251         list_for_each_entry(cma_dev, &dev_list, list)
3252                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3253                         ret = cma_netdev_change(ndev, id_priv);
3254                         if (ret)
3255                                 goto out;
3256                 }
3257
3258 out:
3259         mutex_unlock(&lock);
3260         return ret;
3261 }
3262
3263 static struct notifier_block cma_nb = {
3264         .notifier_call = cma_netdev_callback
3265 };
3266
3267 static void cma_add_one(struct ib_device *device)
3268 {
3269         struct cma_device *cma_dev;
3270         struct rdma_id_private *id_priv;
3271
3272         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3273         if (!cma_dev)
3274                 return;
3275
3276         cma_dev->device = device;
3277
3278         init_completion(&cma_dev->comp);
3279         atomic_set(&cma_dev->refcount, 1);
3280         INIT_LIST_HEAD(&cma_dev->id_list);
3281         ib_set_client_data(device, &cma_client, cma_dev);
3282
3283         mutex_lock(&lock);
3284         list_add_tail(&cma_dev->list, &dev_list);
3285         list_for_each_entry(id_priv, &listen_any_list, list)
3286                 cma_listen_on_dev(id_priv, cma_dev);
3287         mutex_unlock(&lock);
3288 }
3289
3290 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3291 {
3292         struct rdma_cm_event event;
3293         enum cma_state state;
3294         int ret = 0;
3295
3296         /* Record that we want to remove the device */
3297         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
3298         if (state == CMA_DESTROYING)
3299                 return 0;
3300
3301         cma_cancel_operation(id_priv, state);
3302         mutex_lock(&id_priv->handler_mutex);
3303
3304         /* Check for destruction from another callback. */
3305         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
3306                 goto out;
3307
3308         memset(&event, 0, sizeof event);
3309         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3310         ret = id_priv->id.event_handler(&id_priv->id, &event);
3311 out:
3312         mutex_unlock(&id_priv->handler_mutex);
3313         return ret;
3314 }
3315
3316 static void cma_process_remove(struct cma_device *cma_dev)
3317 {
3318         struct rdma_id_private *id_priv;
3319         int ret;
3320
3321         mutex_lock(&lock);
3322         while (!list_empty(&cma_dev->id_list)) {
3323                 id_priv = list_entry(cma_dev->id_list.next,
3324                                      struct rdma_id_private, list);
3325
3326                 list_del(&id_priv->listen_list);
3327                 list_del_init(&id_priv->list);
3328                 atomic_inc(&id_priv->refcount);
3329                 mutex_unlock(&lock);
3330
3331                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3332                 cma_deref_id(id_priv);
3333                 if (ret)
3334                         rdma_destroy_id(&id_priv->id);
3335
3336                 mutex_lock(&lock);
3337         }
3338         mutex_unlock(&lock);
3339
3340         cma_deref_dev(cma_dev);
3341         wait_for_completion(&cma_dev->comp);
3342 }
3343
3344 static void cma_remove_one(struct ib_device *device)
3345 {
3346         struct cma_device *cma_dev;
3347
3348         cma_dev = ib_get_client_data(device, &cma_client);
3349         if (!cma_dev)
3350                 return;
3351
3352         mutex_lock(&lock);
3353         list_del(&cma_dev->list);
3354         mutex_unlock(&lock);
3355
3356         cma_process_remove(cma_dev);
3357         kfree(cma_dev);
3358 }
3359
3360 static int cma_init(void)
3361 {
3362         int ret;
3363 #if defined(INET)
3364         int low, high, remaining;
3365
3366         get_random_bytes(&next_port, sizeof next_port);
3367         inet_get_local_port_range(&low, &high);
3368         remaining = (high - low) + 1;
3369         next_port = ((unsigned int) next_port % remaining) + low;
3370 #endif
3371
3372         cma_wq = create_singlethread_workqueue("rdma_cm");
3373         if (!cma_wq)
3374                 return -ENOMEM;
3375
3376         ib_sa_register_client(&sa_client);
3377         rdma_addr_register_client(&addr_client);
3378         register_netdevice_notifier(&cma_nb);
3379
3380         ret = ib_register_client(&cma_client);
3381         if (ret)
3382                 goto err;
3383         return 0;
3384
3385 err:
3386         unregister_netdevice_notifier(&cma_nb);
3387         rdma_addr_unregister_client(&addr_client);
3388         ib_sa_unregister_client(&sa_client);
3389         destroy_workqueue(cma_wq);
3390         return ret;
3391 }
3392
3393 static void cma_cleanup(void)
3394 {
3395         ib_unregister_client(&cma_client);
3396         unregister_netdevice_notifier(&cma_nb);
3397         rdma_addr_unregister_client(&addr_client);
3398         ib_sa_unregister_client(&sa_client);
3399         destroy_workqueue(cma_wq);
3400         idr_destroy(&sdp_ps);
3401         idr_destroy(&tcp_ps);
3402         idr_destroy(&udp_ps);
3403         idr_destroy(&ipoib_ps);
3404 }
3405
3406 module_init(cma_init);
3407 module_exit(cma_cleanup);