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 #include <linux/slab.h>
  44
  45 #include <net/tcp.h>
  46 #include <net/ipv6.h>
  47
  48 #include <rdma/rdma_cm.h>
  49 #include <rdma/rdma_cm_ib.h>
  50 #include <rdma/rdma_netlink.h>
  51 #include <rdma/ib_cache.h>
  52 #include <rdma/ib_cm.h>
  53 #include <rdma/ib_sa.h>
  54 #include <rdma/iw_cm.h>
  55
  56 MODULE_AUTHOR("Sean Hefty");
  57 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59
  60 #define CMA_CM_RESPONSE_TIMEOUT 20
  61 #define CMA_MAX_CM_RETRIES 15
  62 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  63 #define CMA_IBOE_PACKET_LIFETIME 18
  64
  65 static void cma_add_one(struct ib_device *device);
  66 static void cma_remove_one(struct ib_device *device);
  67
  68 static struct ib_client cma_client = {
  69         .name   = "cma",
  70         .add    = cma_add_one,
  71         .remove = cma_remove_one
  72 };
  73
  74 static struct ib_sa_client sa_client;
  75 static struct rdma_addr_client addr_client;
  76 static LIST_HEAD(dev_list);
  77 static LIST_HEAD(listen_any_list);
  78 static DEFINE_MUTEX(lock);
  79 static struct workqueue_struct *cma_wq;
  80 static DEFINE_IDR(sdp_ps);
  81 static DEFINE_IDR(tcp_ps);
  82 static DEFINE_IDR(udp_ps);
  83 static DEFINE_IDR(ipoib_ps);
  84
  85 struct cma_device {
  86         struct list_head        list;
  87         struct ib_device        *device;
  88         struct completion       comp;
  89         atomic_t                refcount;
  90         struct list_head        id_list;
  91 };
  92
  93 struct rdma_bind_list {
  94         struct idr              *ps;
  95         struct hlist_head       owners;
  96         unsigned short          port;
  97 };
  98
  99 /*
 100  * Device removal can occur at anytime, so we need extra handling to
 101  * serialize notifying the user of device removal with other callbacks.
 102  * We do this by disabling removal notification while a callback is in process,
 103  * and reporting it after the callback completes.
 104  */
 105 struct rdma_id_private {
 106         struct rdma_cm_id       id;
 107
 108         struct rdma_bind_list   *bind_list;
 109         struct hlist_node       node;
 110         struct list_head        list; /* listen_any_list or cma_device.list */
 111         struct list_head        listen_list; /* per device listens */
 112         struct cma_device       *cma_dev;
 113         struct list_head        mc_list;
 114
 115         int                     internal_id;
 116         enum rdma_cm_state      state;
 117         spinlock_t              lock;
 118         struct mutex            qp_mutex;
 119
 120         struct completion       comp;
 121         atomic_t                refcount;
 122         struct mutex            handler_mutex;
 123
 124         int                     backlog;
 125         int                     timeout_ms;
 126         struct ib_sa_query      *query;
 127         int                     query_id;
 128         union {
 129                 struct ib_cm_id *ib;
 130                 struct iw_cm_id *iw;
 131         } cm_id;
 132
 133         u32                     seq_num;
 134         u32                     qkey;
 135         u32                     qp_num;
 136         pid_t                   owner;
 137         u8                      srq;
 138         u8                      tos;
 139         u8                      reuseaddr;
 140 };
 141
 142 struct cma_multicast {
 143         struct rdma_id_private *id_priv;
 144         union {
 145                 struct ib_sa_multicast *ib;
 146         } multicast;
 147         struct list_head        list;
 148         void                    *context;
 149         struct sockaddr_storage addr;
 150         struct kref             mcref;
 151 };
 152
 153 struct cma_work {
 154         struct work_struct      work;
 155         struct rdma_id_private  *id;
 156         enum rdma_cm_state      old_state;
 157         enum rdma_cm_state      new_state;
 158         struct rdma_cm_event    event;
 159 };
 160
 161 struct cma_ndev_work {
 162         struct work_struct      work;
 163         struct rdma_id_private  *id;
 164         struct rdma_cm_event    event;
 165 };
 166
 167 struct iboe_mcast_work {
 168         struct work_struct       work;
 169         struct rdma_id_private  *id;
 170         struct cma_multicast    *mc;
 171 };
 172
 173 union cma_ip_addr {
 174         struct in6_addr ip6;
 175         struct {
 176                 __be32 pad[3];
 177                 __be32 addr;
 178         } ip4;
 179 };
 180
 181 struct cma_hdr {
 182         u8 cma_version;
 183         u8 ip_version;  /* IP version: 7:4 */
 184         __be16 port;
 185         union cma_ip_addr src_addr;
 186         union cma_ip_addr dst_addr;
 187 };
 188
 189 struct sdp_hh {
 190         u8 bsdh[16];
 191         u8 sdp_version; /* Major version: 7:4 */
 192         u8 ip_version;  /* IP version: 7:4 */
 193         u8 sdp_specific1[10];
 194         __be16 port;
 195         __be16 sdp_specific2;
 196         union cma_ip_addr src_addr;
 197         union cma_ip_addr dst_addr;
 198 };
 199
 200 struct sdp_hah {
 201         u8 bsdh[16];
 202         u8 sdp_version;
 203 };
 204
 205 #define CMA_VERSION 0x00
 206 #define SDP_MAJ_VERSION 0x2
 207
 208 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 209 {
 210         unsigned long flags;
 211         int ret;
 212
 213         spin_lock_irqsave(&id_priv->lock, flags);
 214         ret = (id_priv->state == comp);
 215         spin_unlock_irqrestore(&id_priv->lock, flags);
 216         return ret;
 217 }
 218
 219 static int cma_comp_exch(struct rdma_id_private *id_priv,
 220                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 221 {
 222         unsigned long flags;
 223         int ret;
 224
 225         spin_lock_irqsave(&id_priv->lock, flags);
 226         if ((ret = (id_priv->state == comp)))
 227                 id_priv->state = exch;
 228         spin_unlock_irqrestore(&id_priv->lock, flags);
 229         return ret;
 230 }
 231
 232 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 233                                    enum rdma_cm_state exch)
 234 {
 235         unsigned long flags;
 236         enum rdma_cm_state old;
 237
 238         spin_lock_irqsave(&id_priv->lock, flags);
 239         old = id_priv->state;
 240         id_priv->state = exch;
 241         spin_unlock_irqrestore(&id_priv->lock, flags);
 242         return old;
 243 }
 244
 245 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 246 {
 247         return hdr->ip_version >> 4;
 248 }
 249
 250 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 251 {
 252         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 253 }
 254
 255 static inline u8 sdp_get_majv(u8 sdp_version)
 256 {
 257         return sdp_version >> 4;
 258 }
 259
 260 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 261 {
 262         return hh->ip_version >> 4;
 263 }
 264
 265 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 266 {
 267         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 268 }
 269
 270 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 271                               struct cma_device *cma_dev)
 272 {
 273         atomic_inc(&cma_dev->refcount);
 274         id_priv->cma_dev = cma_dev;
 275         id_priv->id.device = cma_dev->device;
 276         id_priv->id.route.addr.dev_addr.transport =
 277                 rdma_node_get_transport(cma_dev->device->node_type);
 278         list_add_tail(&id_priv->list, &cma_dev->id_list);
 279 }
 280
 281 static inline void cma_deref_dev(struct cma_device *cma_dev)
 282 {
 283         if (atomic_dec_and_test(&cma_dev->refcount))
 284                 complete(&cma_dev->comp);
 285 }
 286
 287 static inline void release_mc(struct kref *kref)
 288 {
 289         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 290
 291         kfree(mc->multicast.ib);
 292         kfree(mc);
 293 }
 294
 295 static void cma_release_dev(struct rdma_id_private *id_priv)
 296 {
 297         mutex_lock(&lock);
 298         list_del(&id_priv->list);
 299         cma_deref_dev(id_priv->cma_dev);
 300         id_priv->cma_dev = NULL;
 301         mutex_unlock(&lock);
 302 }
 303
 304 static int cma_set_qkey(struct rdma_id_private *id_priv)
 305 {
 306         struct ib_sa_mcmember_rec rec;
 307         int ret = 0;
 308
 309         if (id_priv->qkey)
 310                 return 0;
 311
 312         switch (id_priv->id.ps) {
 313         case RDMA_PS_UDP:
 314                 id_priv->qkey = RDMA_UDP_QKEY;
 315                 break;
 316         case RDMA_PS_IPOIB:
 317                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 318                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 319                                              id_priv->id.port_num, &rec.mgid,
 320                                              &rec);
 321                 if (!ret)
 322                         id_priv->qkey = be32_to_cpu(rec.qkey);
 323                 break;
 324         default:
 325                 break;
 326         }
 327         return ret;
 328 }
 329
 330 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 331 {
 332         int i;
 333         int err;
 334         struct ib_port_attr props;
 335         union ib_gid tmp;
 336
 337         err = ib_query_port(device, port_num, &props);
 338         if (err)
 339                 return 1;
 340
 341         for (i = 0; i < props.gid_tbl_len; ++i) {
 342                 err = ib_query_gid(device, port_num, i, &tmp);
 343                 if (err)
 344                         return 1;
 345                 if (!memcmp(&tmp, gid, sizeof tmp))
 346                         return 0;
 347         }
 348
 349         return -EAGAIN;
 350 }
 351
 352 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 353 {
 354         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 355         struct cma_device *cma_dev;
 356         union ib_gid gid, iboe_gid;
 357         int ret = -ENODEV;
 358         u8 port;
 359         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 360                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 361
 362         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 363             id_priv->id.ps == RDMA_PS_IPOIB)
 364                 return -EINVAL;
 365
 366         mutex_lock(&lock);
 367         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 368         memcpy(&gid, dev_addr->src_dev_addr +
 369                rdma_addr_gid_offset(dev_addr), sizeof gid);
 370         list_for_each_entry(cma_dev, &dev_list, list) {
 371                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 372                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 373                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 374                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 375                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 376                                 else
 377                                         ret = find_gid_port(cma_dev->device, &gid, port);
 378
 379                                 if (!ret) {
 380                                         id_priv->id.port_num = port;
 381                                         goto out;
 382                                 } else if (ret == 1)
 383                                         break;
 384                         }
 385                 }
 386         }
 387
 388 out:
 389         if (!ret)
 390                 cma_attach_to_dev(id_priv, cma_dev);
 391
 392         mutex_unlock(&lock);
 393         return ret;
 394 }
 395
 396 static void cma_deref_id(struct rdma_id_private *id_priv)
 397 {
 398         if (atomic_dec_and_test(&id_priv->refcount))
 399                 complete(&id_priv->comp);
 400 }
 401
 402 static int cma_disable_callback(struct rdma_id_private *id_priv,
 403                                 enum rdma_cm_state state)
 404 {
 405         mutex_lock(&id_priv->handler_mutex);
 406         if (id_priv->state != state) {
 407                 mutex_unlock(&id_priv->handler_mutex);
 408                 return -EINVAL;
 409         }
 410         return 0;
 411 }
 412
 413 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 414                                   void *context, enum rdma_port_space ps,
 415                                   enum ib_qp_type qp_type)
 416 {
 417         struct rdma_id_private *id_priv;
 418
 419         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 420         if (!id_priv)
 421                 return ERR_PTR(-ENOMEM);
 422
 423         id_priv->owner = task_pid_nr(current);
 424         id_priv->state = RDMA_CM_IDLE;
 425         id_priv->id.context = context;
 426         id_priv->id.event_handler = event_handler;
 427         id_priv->id.ps = ps;
 428         id_priv->id.qp_type = qp_type;
 429         spin_lock_init(&id_priv->lock);
 430         mutex_init(&id_priv->qp_mutex);
 431         init_completion(&id_priv->comp);
 432         atomic_set(&id_priv->refcount, 1);
 433         mutex_init(&id_priv->handler_mutex);
 434         INIT_LIST_HEAD(&id_priv->listen_list);
 435         INIT_LIST_HEAD(&id_priv->mc_list);
 436         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 437
 438         return &id_priv->id;
 439 }
 440 EXPORT_SYMBOL(rdma_create_id);
 441
 442 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 443 {
 444         struct ib_qp_attr qp_attr;
 445         int qp_attr_mask, ret;
 446
 447         qp_attr.qp_state = IB_QPS_INIT;
 448         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 449         if (ret)
 450                 return ret;
 451
 452         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 453         if (ret)
 454                 return ret;
 455
 456         qp_attr.qp_state = IB_QPS_RTR;
 457         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 458         if (ret)
 459                 return ret;
 460
 461         qp_attr.qp_state = IB_QPS_RTS;
 462         qp_attr.sq_psn = 0;
 463         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 464
 465         return ret;
 466 }
 467
 468 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 469 {
 470         struct ib_qp_attr qp_attr;
 471         int qp_attr_mask, ret;
 472
 473         qp_attr.qp_state = IB_QPS_INIT;
 474         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 475         if (ret)
 476                 return ret;
 477
 478         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 479 }
 480
 481 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 482                    struct ib_qp_init_attr *qp_init_attr)
 483 {
 484         struct rdma_id_private *id_priv;
 485         struct ib_qp *qp;
 486         int ret;
 487
 488         id_priv = container_of(id, struct rdma_id_private, id);
 489         if (id->device != pd->device)
 490                 return -EINVAL;
 491
 492         qp = ib_create_qp(pd, qp_init_attr);
 493         if (IS_ERR(qp))
 494                 return PTR_ERR(qp);
 495
 496         if (id->qp_type == IB_QPT_UD)
 497                 ret = cma_init_ud_qp(id_priv, qp);
 498         else
 499                 ret = cma_init_conn_qp(id_priv, qp);
 500         if (ret)
 501                 goto err;
 502
 503         id->qp = qp;
 504         id_priv->qp_num = qp->qp_num;
 505         id_priv->srq = (qp->srq != NULL);
 506         return 0;
 507 err:
 508         ib_destroy_qp(qp);
 509         return ret;
 510 }
 511 EXPORT_SYMBOL(rdma_create_qp);
 512
 513 void rdma_destroy_qp(struct rdma_cm_id *id)
 514 {
 515         struct rdma_id_private *id_priv;
 516
 517         id_priv = container_of(id, struct rdma_id_private, id);
 518         mutex_lock(&id_priv->qp_mutex);
 519         ib_destroy_qp(id_priv->id.qp);
 520         id_priv->id.qp = NULL;
 521         mutex_unlock(&id_priv->qp_mutex);
 522 }
 523 EXPORT_SYMBOL(rdma_destroy_qp);
 524
 525 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 526                              struct rdma_conn_param *conn_param)
 527 {
 528         struct ib_qp_attr qp_attr;
 529         int qp_attr_mask, ret;
 530
 531         mutex_lock(&id_priv->qp_mutex);
 532         if (!id_priv->id.qp) {
 533                 ret = 0;
 534                 goto out;
 535         }
 536
 537         /* Need to update QP attributes from default values. */
 538         qp_attr.qp_state = IB_QPS_INIT;
 539         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 540         if (ret)
 541                 goto out;
 542
 543         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 544         if (ret)
 545                 goto out;
 546
 547         qp_attr.qp_state = IB_QPS_RTR;
 548         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 549         if (ret)
 550                 goto out;
 551
 552         if (conn_param)
 553                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 554         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 555 out:
 556         mutex_unlock(&id_priv->qp_mutex);
 557         return ret;
 558 }
 559
 560 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 561                              struct rdma_conn_param *conn_param)
 562 {
 563         struct ib_qp_attr qp_attr;
 564         int qp_attr_mask, ret;
 565
 566         mutex_lock(&id_priv->qp_mutex);
 567         if (!id_priv->id.qp) {
 568                 ret = 0;
 569                 goto out;
 570         }
 571
 572         qp_attr.qp_state = IB_QPS_RTS;
 573         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 574         if (ret)
 575                 goto out;
 576
 577         if (conn_param)
 578                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 579         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 580 out:
 581         mutex_unlock(&id_priv->qp_mutex);
 582         return ret;
 583 }
 584
 585 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 586 {
 587         struct ib_qp_attr qp_attr;
 588         int ret;
 589
 590         mutex_lock(&id_priv->qp_mutex);
 591         if (!id_priv->id.qp) {
 592                 ret = 0;
 593                 goto out;
 594         }
 595
 596         qp_attr.qp_state = IB_QPS_ERR;
 597         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 598 out:
 599         mutex_unlock(&id_priv->qp_mutex);
 600         return ret;
 601 }
 602
 603 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 604                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 605 {
 606         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 607         int ret;
 608         u16 pkey;
 609
 610         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 611             IB_LINK_LAYER_INFINIBAND)
 612                 pkey = ib_addr_get_pkey(dev_addr);
 613         else
 614                 pkey = 0xffff;
 615
 616         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 617                                   pkey, &qp_attr->pkey_index);
 618         if (ret)
 619                 return ret;
 620
 621         qp_attr->port_num = id_priv->id.port_num;
 622         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 623
 624         if (id_priv->id.qp_type == IB_QPT_UD) {
 625                 ret = cma_set_qkey(id_priv);
 626                 if (ret)
 627                         return ret;
 628
 629                 qp_attr->qkey = id_priv->qkey;
 630                 *qp_attr_mask |= IB_QP_QKEY;
 631         } else {
 632                 qp_attr->qp_access_flags = 0;
 633                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 634         }
 635         return 0;
 636 }
 637
 638 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 639                        int *qp_attr_mask)
 640 {
 641         struct rdma_id_private *id_priv;
 642         int ret = 0;
 643
 644         id_priv = container_of(id, struct rdma_id_private, id);
 645         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 646         case RDMA_TRANSPORT_IB:
 647                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 648                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 649                 else
 650                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 651                                                  qp_attr_mask);
 652                 if (qp_attr->qp_state == IB_QPS_RTR)
 653                         qp_attr->rq_psn = id_priv->seq_num;
 654                 break;
 655         case RDMA_TRANSPORT_IWARP:
 656                 if (!id_priv->cm_id.iw) {
 657                         qp_attr->qp_access_flags = 0;
 658                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 659                 } else
 660                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 661                                                  qp_attr_mask);
 662                 break;
 663         default:
 664                 ret = -ENOSYS;
 665                 break;
 666         }
 667
 668         return ret;
 669 }
 670 EXPORT_SYMBOL(rdma_init_qp_attr);
 671
 672 static inline int cma_zero_addr(struct sockaddr *addr)
 673 {
 674         struct in6_addr *ip6;
 675
 676         if (addr->sa_family == AF_INET)
 677                 return ipv4_is_zeronet(
 678                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 679         else {
 680                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 681                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 682                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 683         }
 684 }
 685
 686 static inline int cma_loopback_addr(struct sockaddr *addr)
 687 {
 688         if (addr->sa_family == AF_INET)
 689                 return ipv4_is_loopback(
 690                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 691         else
 692                 return ipv6_addr_loopback(
 693                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 694 }
 695
 696 static inline int cma_any_addr(struct sockaddr *addr)
 697 {
 698         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 699 }
 700
 701 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 702 {
 703         if (src->sa_family != dst->sa_family)
 704                 return -1;
 705
 706         switch (src->sa_family) {
 707         case AF_INET:
 708                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 709                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 710         default:
 711                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 712                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 713         }
 714 }
 715
 716 static inline __be16 cma_port(struct sockaddr *addr)
 717 {
 718         if (addr->sa_family == AF_INET)
 719                 return ((struct sockaddr_in *) addr)->sin_port;
 720         else
 721                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 722 }
 723
 724 static inline int cma_any_port(struct sockaddr *addr)
 725 {
 726         return !cma_port(addr);
 727 }
 728
 729 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 730                             u8 *ip_ver, __be16 *port,
 731                             union cma_ip_addr **src, union cma_ip_addr **dst)
 732 {
 733         switch (ps) {
 734         case RDMA_PS_SDP:
 735                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 736                     SDP_MAJ_VERSION)
 737                         return -EINVAL;
 738
 739                 *ip_ver = sdp_get_ip_ver(hdr);
 740                 *port   = ((struct sdp_hh *) hdr)->port;
 741                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 742                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 743                 break;
 744         default:
 745                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 746                         return -EINVAL;
 747
 748                 *ip_ver = cma_get_ip_ver(hdr);
 749                 *port   = ((struct cma_hdr *) hdr)->port;
 750                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 751                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 752                 break;
 753         }
 754
 755         if (*ip_ver != 4 && *ip_ver != 6)
 756                 return -EINVAL;
 757         return 0;
 758 }
 759
 760 static void cma_save_net_info(struct rdma_addr *addr,
 761                               struct rdma_addr *listen_addr,
 762                               u8 ip_ver, __be16 port,
 763                               union cma_ip_addr *src, union cma_ip_addr *dst)
 764 {
 765         struct sockaddr_in *listen4, *ip4;
 766         struct sockaddr_in6 *listen6, *ip6;
 767
 768         switch (ip_ver) {
 769         case 4:
 770                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 771                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 772                 ip4->sin_family = listen4->sin_family;
 773                 ip4->sin_addr.s_addr = dst->ip4.addr;
 774                 ip4->sin_port = listen4->sin_port;
 775
 776                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 777                 ip4->sin_family = listen4->sin_family;
 778                 ip4->sin_addr.s_addr = src->ip4.addr;
 779                 ip4->sin_port = port;
 780                 break;
 781         case 6:
 782                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 783                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 784                 ip6->sin6_family = listen6->sin6_family;
 785                 ip6->sin6_addr = dst->ip6;
 786                 ip6->sin6_port = listen6->sin6_port;
 787
 788                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 789                 ip6->sin6_family = listen6->sin6_family;
 790                 ip6->sin6_addr = src->ip6;
 791                 ip6->sin6_port = port;
 792                 break;
 793         default:
 794                 break;
 795         }
 796 }
 797
 798 static inline int cma_user_data_offset(enum rdma_port_space ps)
 799 {
 800         switch (ps) {
 801         case RDMA_PS_SDP:
 802                 return 0;
 803         default:
 804                 return sizeof(struct cma_hdr);
 805         }
 806 }
 807
 808 static void cma_cancel_route(struct rdma_id_private *id_priv)
 809 {
 810         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 811         case IB_LINK_LAYER_INFINIBAND:
 812                 if (id_priv->query)
 813                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 814                 break;
 815         default:
 816                 break;
 817         }
 818 }
 819
 820 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 821 {
 822         struct rdma_id_private *dev_id_priv;
 823
 824         /*
 825          * Remove from listen_any_list to prevent added devices from spawning
 826          * additional listen requests.
 827          */
 828         mutex_lock(&lock);
 829         list_del(&id_priv->list);
 830
 831         while (!list_empty(&id_priv->listen_list)) {
 832                 dev_id_priv = list_entry(id_priv->listen_list.next,
 833                                          struct rdma_id_private, listen_list);
 834                 /* sync with device removal to avoid duplicate destruction */
 835                 list_del_init(&dev_id_priv->list);
 836                 list_del(&dev_id_priv->listen_list);
 837                 mutex_unlock(&lock);
 838
 839                 rdma_destroy_id(&dev_id_priv->id);
 840                 mutex_lock(&lock);
 841         }
 842         mutex_unlock(&lock);
 843 }
 844
 845 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 846                                  enum rdma_cm_state state)
 847 {
 848         switch (state) {
 849         case RDMA_CM_ADDR_QUERY:
 850                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 851                 break;
 852         case RDMA_CM_ROUTE_QUERY:
 853                 cma_cancel_route(id_priv);
 854                 break;
 855         case RDMA_CM_LISTEN:
 856                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 857                                 && !id_priv->cma_dev)
 858                         cma_cancel_listens(id_priv);
 859                 break;
 860         default:
 861                 break;
 862         }
 863 }
 864
 865 static void cma_release_port(struct rdma_id_private *id_priv)
 866 {
 867         struct rdma_bind_list *bind_list = id_priv->bind_list;
 868
 869         if (!bind_list)
 870                 return;
 871
 872         mutex_lock(&lock);
 873         hlist_del(&id_priv->node);
 874         if (hlist_empty(&bind_list->owners)) {
 875                 idr_remove(bind_list->ps, bind_list->port);
 876                 kfree(bind_list);
 877         }
 878         mutex_unlock(&lock);
 879 }
 880
 881 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 882 {
 883         struct cma_multicast *mc;
 884
 885         while (!list_empty(&id_priv->mc_list)) {
 886                 mc = container_of(id_priv->mc_list.next,
 887                                   struct cma_multicast, list);
 888                 list_del(&mc->list);
 889                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 890                 case IB_LINK_LAYER_INFINIBAND:
 891                         ib_sa_free_multicast(mc->multicast.ib);
 892                         kfree(mc);
 893                         break;
 894                 case IB_LINK_LAYER_ETHERNET:
 895                         kref_put(&mc->mcref, release_mc);
 896                         break;
 897                 default:
 898                         break;
 899                 }
 900         }
 901 }
 902
 903 void rdma_destroy_id(struct rdma_cm_id *id)
 904 {
 905         struct rdma_id_private *id_priv;
 906         enum rdma_cm_state state;
 907
 908         id_priv = container_of(id, struct rdma_id_private, id);
 909         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 910         cma_cancel_operation(id_priv, state);
 911
 912         /*
 913          * Wait for any active callback to finish.  New callbacks will find
 914          * the id_priv state set to destroying and abort.
 915          */
 916         mutex_lock(&id_priv->handler_mutex);
 917         mutex_unlock(&id_priv->handler_mutex);
 918
 919         if (id_priv->cma_dev) {
 920                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 921                 case RDMA_TRANSPORT_IB:
 922                         if (id_priv->cm_id.ib)
 923                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 924                         break;
 925                 case RDMA_TRANSPORT_IWARP:
 926                         if (id_priv->cm_id.iw)
 927                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 928                         break;
 929                 default:
 930                         break;
 931                 }
 932                 cma_leave_mc_groups(id_priv);
 933                 cma_release_dev(id_priv);
 934         }
 935
 936         cma_release_port(id_priv);
 937         cma_deref_id(id_priv);
 938         wait_for_completion(&id_priv->comp);
 939
 940         if (id_priv->internal_id)
 941                 cma_deref_id(id_priv->id.context);
 942
 943         kfree(id_priv->id.route.path_rec);
 944         kfree(id_priv);
 945 }
 946 EXPORT_SYMBOL(rdma_destroy_id);
 947
 948 static int cma_rep_recv(struct rdma_id_private *id_priv)
 949 {
 950         int ret;
 951
 952         ret = cma_modify_qp_rtr(id_priv, NULL);
 953         if (ret)
 954                 goto reject;
 955
 956         ret = cma_modify_qp_rts(id_priv, NULL);
 957         if (ret)
 958                 goto reject;
 959
 960         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 961         if (ret)
 962                 goto reject;
 963
 964         return 0;
 965 reject:
 966         cma_modify_qp_err(id_priv);
 967         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 968                        NULL, 0, NULL, 0);
 969         return ret;
 970 }
 971
 972 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 973 {
 974         if (id_priv->id.ps == RDMA_PS_SDP &&
 975             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 976             SDP_MAJ_VERSION)
 977                 return -EINVAL;
 978
 979         return 0;
 980 }
 981
 982 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 983                                    struct ib_cm_rep_event_param *rep_data,
 984                                    void *private_data)
 985 {
 986         event->param.conn.private_data = private_data;
 987         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 988         event->param.conn.responder_resources = rep_data->responder_resources;
 989         event->param.conn.initiator_depth = rep_data->initiator_depth;
 990         event->param.conn.flow_control = rep_data->flow_control;
 991         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 992         event->param.conn.srq = rep_data->srq;
 993         event->param.conn.qp_num = rep_data->remote_qpn;
 994 }
 995
 996 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 997 {
 998         struct rdma_id_private *id_priv = cm_id->context;
 999         struct rdma_cm_event event;
1000         int ret = 0;
1001
1002         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1003                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1004             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1005                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1006                 return 0;
1007
1008         memset(&event, 0, sizeof event);
1009         switch (ib_event->event) {
1010         case IB_CM_REQ_ERROR:
1011         case IB_CM_REP_ERROR:
1012                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1013                 event.status = -ETIMEDOUT;
1014                 break;
1015         case IB_CM_REP_RECEIVED:
1016                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1017                 if (event.status)
1018                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1019                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1020                         event.status = cma_rep_recv(id_priv);
1021                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1022                                                      RDMA_CM_EVENT_ESTABLISHED;
1023                 } else
1024                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1025                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1026                                        ib_event->private_data);
1027                 break;
1028         case IB_CM_RTU_RECEIVED:
1029         case IB_CM_USER_ESTABLISHED:
1030                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1031                 break;
1032         case IB_CM_DREQ_ERROR:
1033                 event.status = -ETIMEDOUT; /* fall through */
1034         case IB_CM_DREQ_RECEIVED:
1035         case IB_CM_DREP_RECEIVED:
1036                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1037                                    RDMA_CM_DISCONNECT))
1038                         goto out;
1039                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1040                 break;
1041         case IB_CM_TIMEWAIT_EXIT:
1042                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1043                 break;
1044         case IB_CM_MRA_RECEIVED:
1045                 /* ignore event */
1046                 goto out;
1047         case IB_CM_REJ_RECEIVED:
1048                 cma_modify_qp_err(id_priv);
1049                 event.status = ib_event->param.rej_rcvd.reason;
1050                 event.event = RDMA_CM_EVENT_REJECTED;
1051                 event.param.conn.private_data = ib_event->private_data;
1052                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1053                 break;
1054         default:
1055                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1056                        ib_event->event);
1057                 goto out;
1058         }
1059
1060         ret = id_priv->id.event_handler(&id_priv->id, &event);
1061         if (ret) {
1062                 /* Destroy the CM ID by returning a non-zero value. */
1063                 id_priv->cm_id.ib = NULL;
1064                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1065                 mutex_unlock(&id_priv->handler_mutex);
1066                 rdma_destroy_id(&id_priv->id);
1067                 return ret;
1068         }
1069 out:
1070         mutex_unlock(&id_priv->handler_mutex);
1071         return ret;
1072 }
1073
1074 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1075                                                struct ib_cm_event *ib_event)
1076 {
1077         struct rdma_id_private *id_priv;
1078         struct rdma_cm_id *id;
1079         struct rdma_route *rt;
1080         union cma_ip_addr *src, *dst;
1081         __be16 port;
1082         u8 ip_ver;
1083         int ret;
1084
1085         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1086                              &ip_ver, &port, &src, &dst))
1087                 return NULL;
1088
1089         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1090                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1091         if (IS_ERR(id))
1092                 return NULL;
1093
1094         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1095                           ip_ver, port, src, dst);
1096
1097         rt = &id->route;
1098         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1099         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1100                                GFP_KERNEL);
1101         if (!rt->path_rec)
1102                 goto err;
1103
1104         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1105         if (rt->num_paths == 2)
1106                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1107
1108         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1109                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1110                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1111                 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1112         } else {
1113                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1114                                         &rt->addr.dev_addr);
1115                 if (ret)
1116                         goto err;
1117         }
1118         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1119
1120         id_priv = container_of(id, struct rdma_id_private, id);
1121         id_priv->state = RDMA_CM_CONNECT;
1122         return id_priv;
1123
1124 err:
1125         rdma_destroy_id(id);
1126         return NULL;
1127 }
1128
1129 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1130                                               struct ib_cm_event *ib_event)
1131 {
1132         struct rdma_id_private *id_priv;
1133         struct rdma_cm_id *id;
1134         union cma_ip_addr *src, *dst;
1135         __be16 port;
1136         u8 ip_ver;
1137         int ret;
1138
1139         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1140                             listen_id->ps, IB_QPT_UD);
1141         if (IS_ERR(id))
1142                 return NULL;
1143
1144
1145         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1146                              &ip_ver, &port, &src, &dst))
1147                 goto err;
1148
1149         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1150                           ip_ver, port, src, dst);
1151
1152         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1153                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1154                                         &id->route.addr.dev_addr);
1155                 if (ret)
1156                         goto err;
1157         }
1158
1159         id_priv = container_of(id, struct rdma_id_private, id);
1160         id_priv->state = RDMA_CM_CONNECT;
1161         return id_priv;
1162 err:
1163         rdma_destroy_id(id);
1164         return NULL;
1165 }
1166
1167 static void cma_set_req_event_data(struct rdma_cm_event *event,
1168                                    struct ib_cm_req_event_param *req_data,
1169                                    void *private_data, int offset)
1170 {
1171         event->param.conn.private_data = private_data + offset;
1172         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1173         event->param.conn.responder_resources = req_data->responder_resources;
1174         event->param.conn.initiator_depth = req_data->initiator_depth;
1175         event->param.conn.flow_control = req_data->flow_control;
1176         event->param.conn.retry_count = req_data->retry_count;
1177         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1178         event->param.conn.srq = req_data->srq;
1179         event->param.conn.qp_num = req_data->remote_qpn;
1180 }
1181
1182 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1183 {
1184         struct rdma_id_private *listen_id, *conn_id;
1185         struct rdma_cm_event event;
1186         int offset, ret;
1187
1188         listen_id = cm_id->context;
1189         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1190                 return -ECONNABORTED;
1191
1192         memset(&event, 0, sizeof event);
1193         offset = cma_user_data_offset(listen_id->id.ps);
1194         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1195         if (listen_id->id.qp_type == IB_QPT_UD) {
1196                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1197                 event.param.ud.private_data = ib_event->private_data + offset;
1198                 event.param.ud.private_data_len =
1199                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1200         } else {
1201                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1202                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1203                                        ib_event->private_data, offset);
1204         }
1205         if (!conn_id) {
1206                 ret = -ENOMEM;
1207                 goto out;
1208         }
1209
1210         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1211         ret = cma_acquire_dev(conn_id);
1212         if (ret)
1213                 goto release_conn_id;
1214
1215         conn_id->cm_id.ib = cm_id;
1216         cm_id->context = conn_id;
1217         cm_id->cm_handler = cma_ib_handler;
1218
1219         /*
1220          * Protect against the user destroying conn_id from another thread
1221          * until we're done accessing it.
1222          */
1223         atomic_inc(&conn_id->refcount);
1224         ret = conn_id->id.event_handler(&conn_id->id, &event);
1225         if (!ret) {
1226                 /*
1227                  * Acquire mutex to prevent user executing rdma_destroy_id()
1228                  * while we're accessing the cm_id.
1229                  */
1230                 mutex_lock(&lock);
1231                 if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1232                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1233                 mutex_unlock(&lock);
1234                 mutex_unlock(&conn_id->handler_mutex);
1235                 cma_deref_id(conn_id);
1236                 goto out;
1237         }
1238         cma_deref_id(conn_id);
1239
1240         /* Destroy the CM ID by returning a non-zero value. */
1241         conn_id->cm_id.ib = NULL;
1242
1243 release_conn_id:
1244         cma_exch(conn_id, RDMA_CM_DESTROYING);
1245         mutex_unlock(&conn_id->handler_mutex);
1246         rdma_destroy_id(&conn_id->id);
1247
1248 out:
1249         mutex_unlock(&listen_id->handler_mutex);
1250         return ret;
1251 }
1252
1253 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1254 {
1255         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1256 }
1257
1258 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1259                                  struct ib_cm_compare_data *compare)
1260 {
1261         struct cma_hdr *cma_data, *cma_mask;
1262         struct sdp_hh *sdp_data, *sdp_mask;
1263         __be32 ip4_addr;
1264         struct in6_addr ip6_addr;
1265
1266         memset(compare, 0, sizeof *compare);
1267         cma_data = (void *) compare->data;
1268         cma_mask = (void *) compare->mask;
1269         sdp_data = (void *) compare->data;
1270         sdp_mask = (void *) compare->mask;
1271
1272         switch (addr->sa_family) {
1273         case AF_INET:
1274                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1275                 if (ps == RDMA_PS_SDP) {
1276                         sdp_set_ip_ver(sdp_data, 4);
1277                         sdp_set_ip_ver(sdp_mask, 0xF);
1278                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1279                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1280                 } else {
1281                         cma_set_ip_ver(cma_data, 4);
1282                         cma_set_ip_ver(cma_mask, 0xF);
1283                         cma_data->dst_addr.ip4.addr = ip4_addr;
1284                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1285                 }
1286                 break;
1287         case AF_INET6:
1288                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1289                 if (ps == RDMA_PS_SDP) {
1290                         sdp_set_ip_ver(sdp_data, 6);
1291                         sdp_set_ip_ver(sdp_mask, 0xF);
1292                         sdp_data->dst_addr.ip6 = ip6_addr;
1293                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1294                                sizeof sdp_mask->dst_addr.ip6);
1295                 } else {
1296                         cma_set_ip_ver(cma_data, 6);
1297                         cma_set_ip_ver(cma_mask, 0xF);
1298                         cma_data->dst_addr.ip6 = ip6_addr;
1299                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1300                                sizeof cma_mask->dst_addr.ip6);
1301                 }
1302                 break;
1303         default:
1304                 break;
1305         }
1306 }
1307
1308 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1309 {
1310         struct rdma_id_private *id_priv = iw_id->context;
1311         struct rdma_cm_event event;
1312         struct sockaddr_in *sin;
1313         int ret = 0;
1314
1315         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1316                 return 0;
1317
1318         memset(&event, 0, sizeof event);
1319         switch (iw_event->event) {
1320         case IW_CM_EVENT_CLOSE:
1321                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1322                 break;
1323         case IW_CM_EVENT_CONNECT_REPLY:
1324                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1325                 *sin = iw_event->local_addr;
1326                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1327                 *sin = iw_event->remote_addr;
1328                 switch (iw_event->status) {
1329                 case 0:
1330                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1331                         break;
1332                 case -ECONNRESET:
1333                 case -ECONNREFUSED:
1334                         event.event = RDMA_CM_EVENT_REJECTED;
1335                         break;
1336                 case -ETIMEDOUT:
1337                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1338                         break;
1339                 default:
1340                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1341                         break;
1342                 }
1343                 break;
1344         case IW_CM_EVENT_ESTABLISHED:
1345                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1346                 break;
1347         default:
1348                 BUG_ON(1);
1349         }
1350
1351         event.status = iw_event->status;
1352         event.param.conn.private_data = iw_event->private_data;
1353         event.param.conn.private_data_len = iw_event->private_data_len;
1354         ret = id_priv->id.event_handler(&id_priv->id, &event);
1355         if (ret) {
1356                 /* Destroy the CM ID by returning a non-zero value. */
1357                 id_priv->cm_id.iw = NULL;
1358                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1359                 mutex_unlock(&id_priv->handler_mutex);
1360                 rdma_destroy_id(&id_priv->id);
1361                 return ret;
1362         }
1363
1364         mutex_unlock(&id_priv->handler_mutex);
1365         return ret;
1366 }
1367
1368 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1369                                struct iw_cm_event *iw_event)
1370 {
1371         struct rdma_cm_id *new_cm_id;
1372         struct rdma_id_private *listen_id, *conn_id;
1373         struct sockaddr_in *sin;
1374         struct net_device *dev = NULL;
1375         struct rdma_cm_event event;
1376         int ret;
1377         struct ib_device_attr attr;
1378
1379         listen_id = cm_id->context;
1380         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1381                 return -ECONNABORTED;
1382
1383         /* Create a new RDMA id for the new IW CM ID */
1384         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1385                                    listen_id->id.context,
1386                                    RDMA_PS_TCP, IB_QPT_RC);
1387         if (IS_ERR(new_cm_id)) {
1388                 ret = -ENOMEM;
1389                 goto out;
1390         }
1391         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1392         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1393         conn_id->state = RDMA_CM_CONNECT;
1394
1395         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1396         if (!dev) {
1397                 ret = -EADDRNOTAVAIL;
1398                 mutex_unlock(&conn_id->handler_mutex);
1399                 rdma_destroy_id(new_cm_id);
1400                 goto out;
1401         }
1402         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1403         if (ret) {
1404                 mutex_unlock(&conn_id->handler_mutex);
1405                 rdma_destroy_id(new_cm_id);
1406                 goto out;
1407         }
1408
1409         ret = cma_acquire_dev(conn_id);
1410         if (ret) {
1411                 mutex_unlock(&conn_id->handler_mutex);
1412                 rdma_destroy_id(new_cm_id);
1413                 goto out;
1414         }
1415
1416         conn_id->cm_id.iw = cm_id;
1417         cm_id->context = conn_id;
1418         cm_id->cm_handler = cma_iw_handler;
1419
1420         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1421         *sin = iw_event->local_addr;
1422         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1423         *sin = iw_event->remote_addr;
1424
1425         ret = ib_query_device(conn_id->id.device, &attr);
1426         if (ret) {
1427                 mutex_unlock(&conn_id->handler_mutex);
1428                 rdma_destroy_id(new_cm_id);
1429                 goto out;
1430         }
1431
1432         memset(&event, 0, sizeof event);
1433         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1434         event.param.conn.private_data = iw_event->private_data;
1435         event.param.conn.private_data_len = iw_event->private_data_len;
1436         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1437         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1438
1439         /*
1440          * Protect against the user destroying conn_id from another thread
1441          * until we're done accessing it.
1442          */
1443         atomic_inc(&conn_id->refcount);
1444         ret = conn_id->id.event_handler(&conn_id->id, &event);
1445         if (ret) {
1446                 /* User wants to destroy the CM ID */
1447                 conn_id->cm_id.iw = NULL;
1448                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1449                 mutex_unlock(&conn_id->handler_mutex);
1450                 cma_deref_id(conn_id);
1451                 rdma_destroy_id(&conn_id->id);
1452                 goto out;
1453         }
1454
1455         mutex_unlock(&conn_id->handler_mutex);
1456         cma_deref_id(conn_id);
1457
1458 out:
1459         if (dev)
1460                 dev_put(dev);
1461         mutex_unlock(&listen_id->handler_mutex);
1462         return ret;
1463 }
1464
1465 static int cma_ib_listen(struct rdma_id_private *id_priv)
1466 {
1467         struct ib_cm_compare_data compare_data;
1468         struct sockaddr *addr;
1469         struct ib_cm_id *id;
1470         __be64 svc_id;
1471         int ret;
1472
1473         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1474         if (IS_ERR(id))
1475                 return PTR_ERR(id);
1476
1477         id_priv->cm_id.ib = id;
1478
1479         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1480         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1481         if (cma_any_addr(addr))
1482                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1483         else {
1484                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1485                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1486         }
1487
1488         if (ret) {
1489                 ib_destroy_cm_id(id_priv->cm_id.ib);
1490                 id_priv->cm_id.ib = NULL;
1491         }
1492
1493         return ret;
1494 }
1495
1496 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1497 {
1498         int ret;
1499         struct sockaddr_in *sin;
1500         struct iw_cm_id *id;
1501
1502         id = iw_create_cm_id(id_priv->id.device,
1503                              iw_conn_req_handler,
1504                              id_priv);
1505         if (IS_ERR(id))
1506                 return PTR_ERR(id);
1507
1508         id_priv->cm_id.iw = id;
1509
1510         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1511         id_priv->cm_id.iw->local_addr = *sin;
1512
1513         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1514
1515         if (ret) {
1516                 iw_destroy_cm_id(id_priv->cm_id.iw);
1517                 id_priv->cm_id.iw = NULL;
1518         }
1519
1520         return ret;
1521 }
1522
1523 static int cma_listen_handler(struct rdma_cm_id *id,
1524                               struct rdma_cm_event *event)
1525 {
1526         struct rdma_id_private *id_priv = id->context;
1527
1528         id->context = id_priv->id.context;
1529         id->event_handler = id_priv->id.event_handler;
1530         return id_priv->id.event_handler(id, event);
1531 }
1532
1533 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1534                               struct cma_device *cma_dev)
1535 {
1536         struct rdma_id_private *dev_id_priv;
1537         struct rdma_cm_id *id;
1538         int ret;
1539
1540         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1541                             id_priv->id.qp_type);
1542         if (IS_ERR(id))
1543                 return;
1544
1545         dev_id_priv = container_of(id, struct rdma_id_private, id);
1546
1547         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1548         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1549                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1550
1551         cma_attach_to_dev(dev_id_priv, cma_dev);
1552         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1553         atomic_inc(&id_priv->refcount);
1554         dev_id_priv->internal_id = 1;
1555
1556         ret = rdma_listen(id, id_priv->backlog);
1557         if (ret)
1558                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1559                        "listening on device %s\n", ret, cma_dev->device->name);
1560 }
1561
1562 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1563 {
1564         struct cma_device *cma_dev;
1565
1566         mutex_lock(&lock);
1567         list_add_tail(&id_priv->list, &listen_any_list);
1568         list_for_each_entry(cma_dev, &dev_list, list)
1569                 cma_listen_on_dev(id_priv, cma_dev);
1570         mutex_unlock(&lock);
1571 }
1572
1573 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1574 {
1575         struct rdma_id_private *id_priv;
1576
1577         id_priv = container_of(id, struct rdma_id_private, id);
1578         id_priv->tos = (u8) tos;
1579 }
1580 EXPORT_SYMBOL(rdma_set_service_type);
1581
1582 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1583                               void *context)
1584 {
1585         struct cma_work *work = context;
1586         struct rdma_route *route;
1587
1588         route = &work->id->id.route;
1589
1590         if (!status) {
1591                 route->num_paths = 1;
1592                 *route->path_rec = *path_rec;
1593         } else {
1594                 work->old_state = RDMA_CM_ROUTE_QUERY;
1595                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1596                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1597                 work->event.status = status;
1598         }
1599
1600         queue_work(cma_wq, &work->work);
1601 }
1602
1603 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1604                               struct cma_work *work)
1605 {
1606         struct rdma_addr *addr = &id_priv->id.route.addr;
1607         struct ib_sa_path_rec path_rec;
1608         ib_sa_comp_mask comp_mask;
1609         struct sockaddr_in6 *sin6;
1610
1611         memset(&path_rec, 0, sizeof path_rec);
1612         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1613         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1614         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1615         path_rec.numb_path = 1;
1616         path_rec.reversible = 1;
1617         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1618                                                         (struct sockaddr *) &addr->dst_addr);
1619
1620         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1621                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1622                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1623
1624         if (addr->src_addr.ss_family == AF_INET) {
1625                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1626                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1627         } else {
1628                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1629                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1630                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1631         }
1632
1633         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1634                                                id_priv->id.port_num, &path_rec,
1635                                                comp_mask, timeout_ms,
1636                                                GFP_KERNEL, cma_query_handler,
1637                                                work, &id_priv->query);
1638
1639         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1640 }
1641
1642 static void cma_work_handler(struct work_struct *_work)
1643 {
1644         struct cma_work *work = container_of(_work, struct cma_work, work);
1645         struct rdma_id_private *id_priv = work->id;
1646         int destroy = 0;
1647
1648         mutex_lock(&id_priv->handler_mutex);
1649         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1650                 goto out;
1651
1652         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1653                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1654                 destroy = 1;
1655         }
1656 out:
1657         mutex_unlock(&id_priv->handler_mutex);
1658         cma_deref_id(id_priv);
1659         if (destroy)
1660                 rdma_destroy_id(&id_priv->id);
1661         kfree(work);
1662 }
1663
1664 static void cma_ndev_work_handler(struct work_struct *_work)
1665 {
1666         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1667         struct rdma_id_private *id_priv = work->id;
1668         int destroy = 0;
1669
1670         mutex_lock(&id_priv->handler_mutex);
1671         if (id_priv->state == RDMA_CM_DESTROYING ||
1672             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1673                 goto out;
1674
1675         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1676                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1677                 destroy = 1;
1678         }
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 int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1689 {
1690         struct rdma_route *route = &id_priv->id.route;
1691         struct cma_work *work;
1692         int ret;
1693
1694         work = kzalloc(sizeof *work, GFP_KERNEL);
1695         if (!work)
1696                 return -ENOMEM;
1697
1698         work->id = id_priv;
1699         INIT_WORK(&work->work, cma_work_handler);
1700         work->old_state = RDMA_CM_ROUTE_QUERY;
1701         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1702         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1703
1704         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1705         if (!route->path_rec) {
1706                 ret = -ENOMEM;
1707                 goto err1;
1708         }
1709
1710         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1711         if (ret)
1712                 goto err2;
1713
1714         return 0;
1715 err2:
1716         kfree(route->path_rec);
1717         route->path_rec = NULL;
1718 err1:
1719         kfree(work);
1720         return ret;
1721 }
1722
1723 int rdma_set_ib_paths(struct rdma_cm_id *id,
1724                       struct ib_sa_path_rec *path_rec, int num_paths)
1725 {
1726         struct rdma_id_private *id_priv;
1727         int ret;
1728
1729         id_priv = container_of(id, struct rdma_id_private, id);
1730         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1731                            RDMA_CM_ROUTE_RESOLVED))
1732                 return -EINVAL;
1733
1734         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1735                                      GFP_KERNEL);
1736         if (!id->route.path_rec) {
1737                 ret = -ENOMEM;
1738                 goto err;
1739         }
1740
1741         id->route.num_paths = num_paths;
1742         return 0;
1743 err:
1744         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1745         return ret;
1746 }
1747 EXPORT_SYMBOL(rdma_set_ib_paths);
1748
1749 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1750 {
1751         struct cma_work *work;
1752
1753         work = kzalloc(sizeof *work, GFP_KERNEL);
1754         if (!work)
1755                 return -ENOMEM;
1756
1757         work->id = id_priv;
1758         INIT_WORK(&work->work, cma_work_handler);
1759         work->old_state = RDMA_CM_ROUTE_QUERY;
1760         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1761         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1762         queue_work(cma_wq, &work->work);
1763         return 0;
1764 }
1765
1766 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1767 {
1768         struct rdma_route *route = &id_priv->id.route;
1769         struct rdma_addr *addr = &route->addr;
1770         struct cma_work *work;
1771         int ret;
1772         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1773         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1774         struct net_device *ndev = NULL;
1775         u16 vid;
1776
1777         if (src_addr->sin_family != dst_addr->sin_family)
1778                 return -EINVAL;
1779
1780         work = kzalloc(sizeof *work, GFP_KERNEL);
1781         if (!work)
1782                 return -ENOMEM;
1783
1784         work->id = id_priv;
1785         INIT_WORK(&work->work, cma_work_handler);
1786
1787         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1788         if (!route->path_rec) {
1789                 ret = -ENOMEM;
1790                 goto err1;
1791         }
1792
1793         route->num_paths = 1;
1794
1795         if (addr->dev_addr.bound_dev_if)
1796                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1797         if (!ndev) {
1798                 ret = -ENODEV;
1799                 goto err2;
1800         }
1801
1802         vid = rdma_vlan_dev_vlan_id(ndev);
1803
1804         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1805         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1806
1807         route->path_rec->hop_limit = 1;
1808         route->path_rec->reversible = 1;
1809         route->path_rec->pkey = cpu_to_be16(0xffff);
1810         route->path_rec->mtu_selector = IB_SA_EQ;
1811         route->path_rec->sl = id_priv->tos >> 5;
1812
1813         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1814         route->path_rec->rate_selector = IB_SA_EQ;
1815         route->path_rec->rate = iboe_get_rate(ndev);
1816         dev_put(ndev);
1817         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1818         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1819         if (!route->path_rec->mtu) {
1820                 ret = -EINVAL;
1821                 goto err2;
1822         }
1823
1824         work->old_state = RDMA_CM_ROUTE_QUERY;
1825         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1826         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1827         work->event.status = 0;
1828
1829         queue_work(cma_wq, &work->work);
1830
1831         return 0;
1832
1833 err2:
1834         kfree(route->path_rec);
1835         route->path_rec = NULL;
1836 err1:
1837         kfree(work);
1838         return ret;
1839 }
1840
1841 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1842 {
1843         struct rdma_id_private *id_priv;
1844         int ret;
1845
1846         id_priv = container_of(id, struct rdma_id_private, id);
1847         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1848                 return -EINVAL;
1849
1850         atomic_inc(&id_priv->refcount);
1851         switch (rdma_node_get_transport(id->device->node_type)) {
1852         case RDMA_TRANSPORT_IB:
1853                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1854                 case IB_LINK_LAYER_INFINIBAND:
1855                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1856                         break;
1857                 case IB_LINK_LAYER_ETHERNET:
1858                         ret = cma_resolve_iboe_route(id_priv);
1859                         break;
1860                 default:
1861                         ret = -ENOSYS;
1862                 }
1863                 break;
1864         case RDMA_TRANSPORT_IWARP:
1865                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1866                 break;
1867         default:
1868                 ret = -ENOSYS;
1869                 break;
1870         }
1871         if (ret)
1872                 goto err;
1873
1874         return 0;
1875 err:
1876         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1877         cma_deref_id(id_priv);
1878         return ret;
1879 }
1880 EXPORT_SYMBOL(rdma_resolve_route);
1881
1882 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1883 {
1884         struct cma_device *cma_dev;
1885         struct ib_port_attr port_attr;
1886         union ib_gid gid;
1887         u16 pkey;
1888         int ret;
1889         u8 p;
1890
1891         mutex_lock(&lock);
1892         if (list_empty(&dev_list)) {
1893                 ret = -ENODEV;
1894                 goto out;
1895         }
1896         list_for_each_entry(cma_dev, &dev_list, list)
1897                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1898                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1899                             port_attr.state == IB_PORT_ACTIVE)
1900                                 goto port_found;
1901
1902         p = 1;
1903         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1904
1905 port_found:
1906         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1907         if (ret)
1908                 goto out;
1909
1910         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1911         if (ret)
1912                 goto out;
1913
1914         id_priv->id.route.addr.dev_addr.dev_type =
1915                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1916                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1917
1918         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1919         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1920         id_priv->id.port_num = p;
1921         cma_attach_to_dev(id_priv, cma_dev);
1922 out:
1923         mutex_unlock(&lock);
1924         return ret;
1925 }
1926
1927 static void addr_handler(int status, struct sockaddr *src_addr,
1928                          struct rdma_dev_addr *dev_addr, void *context)
1929 {
1930         struct rdma_id_private *id_priv = context;
1931         struct rdma_cm_event event;
1932
1933         memset(&event, 0, sizeof event);
1934         mutex_lock(&id_priv->handler_mutex);
1935         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1936                            RDMA_CM_ADDR_RESOLVED))
1937                 goto out;
1938
1939         if (!status && !id_priv->cma_dev)
1940                 status = cma_acquire_dev(id_priv);
1941
1942         if (status) {
1943                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1944                                    RDMA_CM_ADDR_BOUND))
1945                         goto out;
1946                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1947                 event.status = status;
1948         } else {
1949                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1950                        ip_addr_size(src_addr));
1951                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1952         }
1953
1954         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1955                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1956                 mutex_unlock(&id_priv->handler_mutex);
1957                 cma_deref_id(id_priv);
1958                 rdma_destroy_id(&id_priv->id);
1959                 return;
1960         }
1961 out:
1962         mutex_unlock(&id_priv->handler_mutex);
1963         cma_deref_id(id_priv);
1964 }
1965
1966 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1967 {
1968         struct cma_work *work;
1969         struct sockaddr *src, *dst;
1970         union ib_gid gid;
1971         int ret;
1972
1973         work = kzalloc(sizeof *work, GFP_KERNEL);
1974         if (!work)
1975                 return -ENOMEM;
1976
1977         if (!id_priv->cma_dev) {
1978                 ret = cma_bind_loopback(id_priv);
1979                 if (ret)
1980                         goto err;
1981         }
1982
1983         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1984         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1985
1986         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1987         if (cma_zero_addr(src)) {
1988                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
1989                 if ((src->sa_family = dst->sa_family) == AF_INET) {
1990                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
1991                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1992                 } else {
1993                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
1994                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
1995                 }
1996         }
1997
1998         work->id = id_priv;
1999         INIT_WORK(&work->work, cma_work_handler);
2000         work->old_state = RDMA_CM_ADDR_QUERY;
2001         work->new_state = RDMA_CM_ADDR_RESOLVED;
2002         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2003         queue_work(cma_wq, &work->work);
2004         return 0;
2005 err:
2006         kfree(work);
2007         return ret;
2008 }
2009
2010 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2011                          struct sockaddr *dst_addr)
2012 {
2013         if (!src_addr || !src_addr->sa_family) {
2014                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2015                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2016                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2017                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2018                 }
2019         }
2020         return rdma_bind_addr(id, src_addr);
2021 }
2022
2023 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2024                       struct sockaddr *dst_addr, int timeout_ms)
2025 {
2026         struct rdma_id_private *id_priv;
2027         int ret;
2028
2029         id_priv = container_of(id, struct rdma_id_private, id);
2030         if (id_priv->state == RDMA_CM_IDLE) {
2031                 ret = cma_bind_addr(id, src_addr, dst_addr);
2032                 if (ret)
2033                         return ret;
2034         }
2035
2036         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2037                 return -EINVAL;
2038
2039         atomic_inc(&id_priv->refcount);
2040         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2041         if (cma_any_addr(dst_addr))
2042                 ret = cma_resolve_loopback(id_priv);
2043         else
2044                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2045                                       dst_addr, &id->route.addr.dev_addr,
2046                                       timeout_ms, addr_handler, id_priv);
2047         if (ret)
2048                 goto err;
2049
2050         return 0;
2051 err:
2052         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2053         cma_deref_id(id_priv);
2054         return ret;
2055 }
2056 EXPORT_SYMBOL(rdma_resolve_addr);
2057
2058 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2059 {
2060         struct rdma_id_private *id_priv;
2061         unsigned long flags;
2062         int ret;
2063
2064         id_priv = container_of(id, struct rdma_id_private, id);
2065         spin_lock_irqsave(&id_priv->lock, flags);
2066         if (id_priv->state == RDMA_CM_IDLE) {
2067                 id_priv->reuseaddr = reuse;
2068                 ret = 0;
2069         } else {
2070                 ret = -EINVAL;
2071         }
2072         spin_unlock_irqrestore(&id_priv->lock, flags);
2073         return ret;
2074 }
2075 EXPORT_SYMBOL(rdma_set_reuseaddr);
2076
2077 static void cma_bind_port(struct rdma_bind_list *bind_list,
2078                           struct rdma_id_private *id_priv)
2079 {
2080         struct sockaddr_in *sin;
2081
2082         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2083         sin->sin_port = htons(bind_list->port);
2084         id_priv->bind_list = bind_list;
2085         hlist_add_head(&id_priv->node, &bind_list->owners);
2086 }
2087
2088 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2089                           unsigned short snum)
2090 {
2091         struct rdma_bind_list *bind_list;
2092         int port, ret;
2093
2094         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2095         if (!bind_list)
2096                 return -ENOMEM;
2097
2098         do {
2099                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2100         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2101
2102         if (ret)
2103                 goto err1;
2104
2105         if (port != snum) {
2106                 ret = -EADDRNOTAVAIL;
2107                 goto err2;
2108         }
2109
2110         bind_list->ps = ps;
2111         bind_list->port = (unsigned short) port;
2112         cma_bind_port(bind_list, id_priv);
2113         return 0;
2114 err2:
2115         idr_remove(ps, port);
2116 err1:
2117         kfree(bind_list);
2118         return ret;
2119 }
2120
2121 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2122 {
2123         static unsigned int last_used_port;
2124         int low, high, remaining;
2125         unsigned int rover;
2126
2127         inet_get_local_port_range(&low, &high);
2128         remaining = (high - low) + 1;
2129         rover = net_random() % remaining + low;
2130 retry:
2131         if (last_used_port != rover &&
2132             !idr_find(ps, (unsigned short) rover)) {
2133                 int ret = cma_alloc_port(ps, id_priv, rover);
2134                 /*
2135                  * Remember previously used port number in order to avoid
2136                  * re-using same port immediately after it is closed.
2137                  */
2138                 if (!ret)
2139                         last_used_port = rover;
2140                 if (ret != -EADDRNOTAVAIL)
2141                         return ret;
2142         }
2143         if (--remaining) {
2144                 rover++;
2145                 if ((rover < low) || (rover > high))
2146                         rover = low;
2147                 goto retry;
2148         }
2149         return -EADDRNOTAVAIL;
2150 }
2151
2152 /*
2153  * Check that the requested port is available.  This is called when trying to
2154  * bind to a specific port, or when trying to listen on a bound port.  In
2155  * the latter case, the provided id_priv may already be on the bind_list, but
2156  * we still need to check that it's okay to start listening.
2157  */
2158 static int cma_check_port(struct rdma_bind_list *bind_list,
2159                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2160 {
2161         struct rdma_id_private *cur_id;
2162         struct sockaddr *addr, *cur_addr;
2163         struct hlist_node *node;
2164
2165         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2166         if (cma_any_addr(addr) && !reuseaddr)
2167                 return -EADDRNOTAVAIL;
2168
2169         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2170                 if (id_priv == cur_id)
2171                         continue;
2172
2173                 if ((cur_id->state == RDMA_CM_LISTEN) ||
2174                     !reuseaddr || !cur_id->reuseaddr) {
2175                         cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2176                         if (cma_any_addr(cur_addr))
2177                                 return -EADDRNOTAVAIL;
2178
2179                         if (!cma_addr_cmp(addr, cur_addr))
2180                                 return -EADDRINUSE;
2181                 }
2182         }
2183         return 0;
2184 }
2185
2186 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2187 {
2188         struct rdma_bind_list *bind_list;
2189         unsigned short snum;
2190         int ret;
2191
2192         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2193         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2194                 return -EACCES;
2195
2196         bind_list = idr_find(ps, snum);
2197         if (!bind_list) {
2198                 ret = cma_alloc_port(ps, id_priv, snum);
2199         } else {
2200                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2201                 if (!ret)
2202                         cma_bind_port(bind_list, id_priv);
2203         }
2204         return ret;
2205 }
2206
2207 static int cma_bind_listen(struct rdma_id_private *id_priv)
2208 {
2209         struct rdma_bind_list *bind_list = id_priv->bind_list;
2210         int ret = 0;
2211
2212         mutex_lock(&lock);
2213         if (bind_list->owners.first->next)
2214                 ret = cma_check_port(bind_list, id_priv, 0);
2215         mutex_unlock(&lock);
2216         return ret;
2217 }
2218
2219 static int cma_get_port(struct rdma_id_private *id_priv)
2220 {
2221         struct idr *ps;
2222         int ret;
2223
2224         switch (id_priv->id.ps) {
2225         case RDMA_PS_SDP:
2226                 ps = &sdp_ps;
2227                 break;
2228         case RDMA_PS_TCP:
2229                 ps = &tcp_ps;
2230                 break;
2231         case RDMA_PS_UDP:
2232                 ps = &udp_ps;
2233                 break;
2234         case RDMA_PS_IPOIB:
2235                 ps = &ipoib_ps;
2236                 break;
2237         default:
2238                 return -EPROTONOSUPPORT;
2239         }
2240
2241         mutex_lock(&lock);
2242         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2243                 ret = cma_alloc_any_port(ps, id_priv);
2244         else
2245                 ret = cma_use_port(ps, id_priv);
2246         mutex_unlock(&lock);
2247
2248         return ret;
2249 }
2250
2251 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2252                                struct sockaddr *addr)
2253 {
2254 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2255         struct sockaddr_in6 *sin6;
2256
2257         if (addr->sa_family != AF_INET6)
2258                 return 0;
2259
2260         sin6 = (struct sockaddr_in6 *) addr;
2261         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2262             !sin6->sin6_scope_id)
2263                         return -EINVAL;
2264
2265         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2266 #endif
2267         return 0;
2268 }
2269
2270 int rdma_listen(struct rdma_cm_id *id, int backlog)
2271 {
2272         struct rdma_id_private *id_priv;
2273         int ret;
2274
2275         id_priv = container_of(id, struct rdma_id_private, id);
2276         if (id_priv->state == RDMA_CM_IDLE) {
2277                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2278                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2279                 if (ret)
2280                         return ret;
2281         }
2282
2283         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2284                 return -EINVAL;
2285
2286         if (id_priv->reuseaddr) {
2287                 ret = cma_bind_listen(id_priv);
2288                 if (ret)
2289                         goto err;
2290         }
2291
2292         id_priv->backlog = backlog;
2293         if (id->device) {
2294                 switch (rdma_node_get_transport(id->device->node_type)) {
2295                 case RDMA_TRANSPORT_IB:
2296                         ret = cma_ib_listen(id_priv);
2297                         if (ret)
2298                                 goto err;
2299                         break;
2300                 case RDMA_TRANSPORT_IWARP:
2301                         ret = cma_iw_listen(id_priv, backlog);
2302                         if (ret)
2303                                 goto err;
2304                         break;
2305                 default:
2306                         ret = -ENOSYS;
2307                         goto err;
2308                 }
2309         } else
2310                 cma_listen_on_all(id_priv);
2311
2312         return 0;
2313 err:
2314         id_priv->backlog = 0;
2315         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2316         return ret;
2317 }
2318 EXPORT_SYMBOL(rdma_listen);
2319
2320 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2321 {
2322         struct rdma_id_private *id_priv;
2323         int ret;
2324
2325         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2326                 return -EAFNOSUPPORT;
2327
2328         id_priv = container_of(id, struct rdma_id_private, id);
2329         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2330                 return -EINVAL;
2331
2332         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2333         if (ret)
2334                 goto err1;
2335
2336         if (!cma_any_addr(addr)) {
2337                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2338                 if (ret)
2339                         goto err1;
2340
2341                 ret = cma_acquire_dev(id_priv);
2342                 if (ret)
2343                         goto err1;
2344         }
2345
2346         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2347         ret = cma_get_port(id_priv);
2348         if (ret)
2349                 goto err2;
2350
2351         return 0;
2352 err2:
2353         if (id_priv->cma_dev)
2354                 cma_release_dev(id_priv);
2355 err1:
2356         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2357         return ret;
2358 }
2359 EXPORT_SYMBOL(rdma_bind_addr);
2360
2361 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2362                           struct rdma_route *route)
2363 {
2364         struct cma_hdr *cma_hdr;
2365         struct sdp_hh *sdp_hdr;
2366
2367         if (route->addr.src_addr.ss_family == AF_INET) {
2368                 struct sockaddr_in *src4, *dst4;
2369
2370                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2371                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2372
2373                 switch (ps) {
2374                 case RDMA_PS_SDP:
2375                         sdp_hdr = hdr;
2376                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2377                                 return -EINVAL;
2378                         sdp_set_ip_ver(sdp_hdr, 4);
2379                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2380                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2381                         sdp_hdr->port = src4->sin_port;
2382                         break;
2383                 default:
2384                         cma_hdr = hdr;
2385                         cma_hdr->cma_version = CMA_VERSION;
2386                         cma_set_ip_ver(cma_hdr, 4);
2387                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2388                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2389                         cma_hdr->port = src4->sin_port;
2390                         break;
2391                 }
2392         } else {
2393                 struct sockaddr_in6 *src6, *dst6;
2394
2395                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2396                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2397
2398                 switch (ps) {
2399                 case RDMA_PS_SDP:
2400                         sdp_hdr = hdr;
2401                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2402                                 return -EINVAL;
2403                         sdp_set_ip_ver(sdp_hdr, 6);
2404                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2405                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2406                         sdp_hdr->port = src6->sin6_port;
2407                         break;
2408                 default:
2409                         cma_hdr = hdr;
2410                         cma_hdr->cma_version = CMA_VERSION;
2411                         cma_set_ip_ver(cma_hdr, 6);
2412                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2413                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2414                         cma_hdr->port = src6->sin6_port;
2415                         break;
2416                 }
2417         }
2418         return 0;
2419 }
2420
2421 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2422                                 struct ib_cm_event *ib_event)
2423 {
2424         struct rdma_id_private *id_priv = cm_id->context;
2425         struct rdma_cm_event event;
2426         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2427         int ret = 0;
2428
2429         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2430                 return 0;
2431
2432         memset(&event, 0, sizeof event);
2433         switch (ib_event->event) {
2434         case IB_CM_SIDR_REQ_ERROR:
2435                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2436                 event.status = -ETIMEDOUT;
2437                 break;
2438         case IB_CM_SIDR_REP_RECEIVED:
2439                 event.param.ud.private_data = ib_event->private_data;
2440                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2441                 if (rep->status != IB_SIDR_SUCCESS) {
2442                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2443                         event.status = ib_event->param.sidr_rep_rcvd.status;
2444                         break;
2445                 }
2446                 ret = cma_set_qkey(id_priv);
2447                 if (ret) {
2448                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2449                         event.status = -EINVAL;
2450                         break;
2451                 }
2452                 if (id_priv->qkey != rep->qkey) {
2453                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2454                         event.status = -EINVAL;
2455                         break;
2456                 }
2457                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2458                                      id_priv->id.route.path_rec,
2459                                      &event.param.ud.ah_attr);
2460                 event.param.ud.qp_num = rep->qpn;
2461                 event.param.ud.qkey = rep->qkey;
2462                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2463                 event.status = 0;
2464                 break;
2465         default:
2466                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2467                        ib_event->event);
2468                 goto out;
2469         }
2470
2471         ret = id_priv->id.event_handler(&id_priv->id, &event);
2472         if (ret) {
2473                 /* Destroy the CM ID by returning a non-zero value. */
2474                 id_priv->cm_id.ib = NULL;
2475                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2476                 mutex_unlock(&id_priv->handler_mutex);
2477                 rdma_destroy_id(&id_priv->id);
2478                 return ret;
2479         }
2480 out:
2481         mutex_unlock(&id_priv->handler_mutex);
2482         return ret;
2483 }
2484
2485 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2486                               struct rdma_conn_param *conn_param)
2487 {
2488         struct ib_cm_sidr_req_param req;
2489         struct rdma_route *route;
2490         struct ib_cm_id *id;
2491         int ret;
2492
2493         req.private_data_len = sizeof(struct cma_hdr) +
2494                                conn_param->private_data_len;
2495         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2496         if (!req.private_data)
2497                 return -ENOMEM;
2498
2499         if (conn_param->private_data && conn_param->private_data_len)
2500                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2501                        conn_param->private_data, conn_param->private_data_len);
2502
2503         route = &id_priv->id.route;
2504         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2505         if (ret)
2506                 goto out;
2507
2508         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2509                              id_priv);
2510         if (IS_ERR(id)) {
2511                 ret = PTR_ERR(id);
2512                 goto out;
2513         }
2514         id_priv->cm_id.ib = id;
2515
2516         req.path = route->path_rec;
2517         req.service_id = cma_get_service_id(id_priv->id.ps,
2518                                             (struct sockaddr *) &route->addr.dst_addr);
2519         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2520         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2521
2522         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2523         if (ret) {
2524                 ib_destroy_cm_id(id_priv->cm_id.ib);
2525                 id_priv->cm_id.ib = NULL;
2526         }
2527 out:
2528         kfree(req.private_data);
2529         return ret;
2530 }
2531
2532 static int cma_connect_ib(struct rdma_id_private *id_priv,
2533                           struct rdma_conn_param *conn_param)
2534 {
2535         struct ib_cm_req_param req;
2536         struct rdma_route *route;
2537         void *private_data;
2538         struct ib_cm_id *id;
2539         int offset, ret;
2540
2541         memset(&req, 0, sizeof req);
2542         offset = cma_user_data_offset(id_priv->id.ps);
2543         req.private_data_len = offset + conn_param->private_data_len;
2544         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2545         if (!private_data)
2546                 return -ENOMEM;
2547
2548         if (conn_param->private_data && conn_param->private_data_len)
2549                 memcpy(private_data + offset, conn_param->private_data,
2550                        conn_param->private_data_len);
2551
2552         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2553         if (IS_ERR(id)) {
2554                 ret = PTR_ERR(id);
2555                 goto out;
2556         }
2557         id_priv->cm_id.ib = id;
2558
2559         route = &id_priv->id.route;
2560         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2561         if (ret)
2562                 goto out;
2563         req.private_data = private_data;
2564
2565         req.primary_path = &route->path_rec[0];
2566         if (route->num_paths == 2)
2567                 req.alternate_path = &route->path_rec[1];
2568
2569         req.service_id = cma_get_service_id(id_priv->id.ps,
2570                                             (struct sockaddr *) &route->addr.dst_addr);
2571         req.qp_num = id_priv->qp_num;
2572         req.qp_type = IB_QPT_RC;
2573         req.starting_psn = id_priv->seq_num;
2574         req.responder_resources = conn_param->responder_resources;
2575         req.initiator_depth = conn_param->initiator_depth;
2576         req.flow_control = conn_param->flow_control;
2577         req.retry_count = conn_param->retry_count;
2578         req.rnr_retry_count = conn_param->rnr_retry_count;
2579         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2580         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2581         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2582         req.srq = id_priv->srq ? 1 : 0;
2583
2584         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2585 out:
2586         if (ret && !IS_ERR(id)) {
2587                 ib_destroy_cm_id(id);
2588                 id_priv->cm_id.ib = NULL;
2589         }
2590
2591         kfree(private_data);
2592         return ret;
2593 }
2594
2595 static int cma_connect_iw(struct rdma_id_private *id_priv,
2596                           struct rdma_conn_param *conn_param)
2597 {
2598         struct iw_cm_id *cm_id;
2599         struct sockaddr_in* sin;
2600         int ret;
2601         struct iw_cm_conn_param iw_param;
2602
2603         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2604         if (IS_ERR(cm_id))
2605                 return PTR_ERR(cm_id);
2606
2607         id_priv->cm_id.iw = cm_id;
2608
2609         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2610         cm_id->local_addr = *sin;
2611
2612         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2613         cm_id->remote_addr = *sin;
2614
2615         ret = cma_modify_qp_rtr(id_priv, conn_param);
2616         if (ret)
2617                 goto out;
2618
2619         iw_param.ord = conn_param->initiator_depth;
2620         iw_param.ird = conn_param->responder_resources;
2621         iw_param.private_data = conn_param->private_data;
2622         iw_param.private_data_len = conn_param->private_data_len;
2623         if (id_priv->id.qp)
2624                 iw_param.qpn = id_priv->qp_num;
2625         else
2626                 iw_param.qpn = conn_param->qp_num;
2627         ret = iw_cm_connect(cm_id, &iw_param);
2628 out:
2629         if (ret) {
2630                 iw_destroy_cm_id(cm_id);
2631                 id_priv->cm_id.iw = NULL;
2632         }
2633         return ret;
2634 }
2635
2636 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2637 {
2638         struct rdma_id_private *id_priv;
2639         int ret;
2640
2641         id_priv = container_of(id, struct rdma_id_private, id);
2642         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2643                 return -EINVAL;
2644
2645         if (!id->qp) {
2646                 id_priv->qp_num = conn_param->qp_num;
2647                 id_priv->srq = conn_param->srq;
2648         }
2649
2650         switch (rdma_node_get_transport(id->device->node_type)) {
2651         case RDMA_TRANSPORT_IB:
2652                 if (id->qp_type == IB_QPT_UD)
2653                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2654                 else
2655                         ret = cma_connect_ib(id_priv, conn_param);
2656                 break;
2657         case RDMA_TRANSPORT_IWARP:
2658                 ret = cma_connect_iw(id_priv, conn_param);
2659                 break;
2660         default:
2661                 ret = -ENOSYS;
2662                 break;
2663         }
2664         if (ret)
2665                 goto err;
2666
2667         return 0;
2668 err:
2669         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2670         return ret;
2671 }
2672 EXPORT_SYMBOL(rdma_connect);
2673
2674 static int cma_accept_ib(struct rdma_id_private *id_priv,
2675                          struct rdma_conn_param *conn_param)
2676 {
2677         struct ib_cm_rep_param rep;
2678         int ret;
2679
2680         ret = cma_modify_qp_rtr(id_priv, conn_param);
2681         if (ret)
2682                 goto out;
2683
2684         ret = cma_modify_qp_rts(id_priv, conn_param);
2685         if (ret)
2686                 goto out;
2687
2688         memset(&rep, 0, sizeof rep);
2689         rep.qp_num = id_priv->qp_num;
2690         rep.starting_psn = id_priv->seq_num;
2691         rep.private_data = conn_param->private_data;
2692         rep.private_data_len = conn_param->private_data_len;
2693         rep.responder_resources = conn_param->responder_resources;
2694         rep.initiator_depth = conn_param->initiator_depth;
2695         rep.failover_accepted = 0;
2696         rep.flow_control = conn_param->flow_control;
2697         rep.rnr_retry_count = conn_param->rnr_retry_count;
2698         rep.srq = id_priv->srq ? 1 : 0;
2699
2700         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2701 out:
2702         return ret;
2703 }
2704
2705 static int cma_accept_iw(struct rdma_id_private *id_priv,
2706                   struct rdma_conn_param *conn_param)
2707 {
2708         struct iw_cm_conn_param iw_param;
2709         int ret;
2710
2711         ret = cma_modify_qp_rtr(id_priv, conn_param);
2712         if (ret)
2713                 return ret;
2714
2715         iw_param.ord = conn_param->initiator_depth;
2716         iw_param.ird = conn_param->responder_resources;
2717         iw_param.private_data = conn_param->private_data;
2718         iw_param.private_data_len = conn_param->private_data_len;
2719         if (id_priv->id.qp) {
2720                 iw_param.qpn = id_priv->qp_num;
2721         } else
2722                 iw_param.qpn = conn_param->qp_num;
2723
2724         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2725 }
2726
2727 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2728                              enum ib_cm_sidr_status status,
2729                              const void *private_data, int private_data_len)
2730 {
2731         struct ib_cm_sidr_rep_param rep;
2732         int ret;
2733
2734         memset(&rep, 0, sizeof rep);
2735         rep.status = status;
2736         if (status == IB_SIDR_SUCCESS) {
2737                 ret = cma_set_qkey(id_priv);
2738                 if (ret)
2739                         return ret;
2740                 rep.qp_num = id_priv->qp_num;
2741                 rep.qkey = id_priv->qkey;
2742         }
2743         rep.private_data = private_data;
2744         rep.private_data_len = private_data_len;
2745
2746         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2747 }
2748
2749 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2750 {
2751         struct rdma_id_private *id_priv;
2752         int ret;
2753
2754         id_priv = container_of(id, struct rdma_id_private, id);
2755
2756         id_priv->owner = task_pid_nr(current);
2757
2758         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2759                 return -EINVAL;
2760
2761         if (!id->qp && conn_param) {
2762                 id_priv->qp_num = conn_param->qp_num;
2763                 id_priv->srq = conn_param->srq;
2764         }
2765
2766         switch (rdma_node_get_transport(id->device->node_type)) {
2767         case RDMA_TRANSPORT_IB:
2768                 if (id->qp_type == IB_QPT_UD)
2769                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2770                                                 conn_param->private_data,
2771                                                 conn_param->private_data_len);
2772                 else if (conn_param)
2773                         ret = cma_accept_ib(id_priv, conn_param);
2774                 else
2775                         ret = cma_rep_recv(id_priv);
2776                 break;
2777         case RDMA_TRANSPORT_IWARP:
2778                 ret = cma_accept_iw(id_priv, conn_param);
2779                 break;
2780         default:
2781                 ret = -ENOSYS;
2782                 break;
2783         }
2784
2785         if (ret)
2786                 goto reject;
2787
2788         return 0;
2789 reject:
2790         cma_modify_qp_err(id_priv);
2791         rdma_reject(id, NULL, 0);
2792         return ret;
2793 }
2794 EXPORT_SYMBOL(rdma_accept);
2795
2796 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2797 {
2798         struct rdma_id_private *id_priv;
2799         int ret;
2800
2801         id_priv = container_of(id, struct rdma_id_private, id);
2802         if (!id_priv->cm_id.ib)
2803                 return -EINVAL;
2804
2805         switch (id->device->node_type) {
2806         case RDMA_NODE_IB_CA:
2807                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2808                 break;
2809         default:
2810                 ret = 0;
2811                 break;
2812         }
2813         return ret;
2814 }
2815 EXPORT_SYMBOL(rdma_notify);
2816
2817 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2818                 u8 private_data_len)
2819 {
2820         struct rdma_id_private *id_priv;
2821         int ret;
2822
2823         id_priv = container_of(id, struct rdma_id_private, id);
2824         if (!id_priv->cm_id.ib)
2825                 return -EINVAL;
2826
2827         switch (rdma_node_get_transport(id->device->node_type)) {
2828         case RDMA_TRANSPORT_IB:
2829                 if (id->qp_type == IB_QPT_UD)
2830                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2831                                                 private_data, private_data_len);
2832                 else
2833                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2834                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2835                                              0, private_data, private_data_len);
2836                 break;
2837         case RDMA_TRANSPORT_IWARP:
2838                 ret = iw_cm_reject(id_priv->cm_id.iw,
2839                                    private_data, private_data_len);
2840                 break;
2841         default:
2842                 ret = -ENOSYS;
2843                 break;
2844         }
2845         return ret;
2846 }
2847 EXPORT_SYMBOL(rdma_reject);
2848
2849 int rdma_disconnect(struct rdma_cm_id *id)
2850 {
2851         struct rdma_id_private *id_priv;
2852         int ret;
2853
2854         id_priv = container_of(id, struct rdma_id_private, id);
2855         if (!id_priv->cm_id.ib)
2856                 return -EINVAL;
2857
2858         switch (rdma_node_get_transport(id->device->node_type)) {
2859         case RDMA_TRANSPORT_IB:
2860                 ret = cma_modify_qp_err(id_priv);
2861                 if (ret)
2862                         goto out;
2863                 /* Initiate or respond to a disconnect. */
2864                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2865                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2866                 break;
2867         case RDMA_TRANSPORT_IWARP:
2868                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2869                 break;
2870         default:
2871                 ret = -EINVAL;
2872                 break;
2873         }
2874 out:
2875         return ret;
2876 }
2877 EXPORT_SYMBOL(rdma_disconnect);
2878
2879 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2880 {
2881         struct rdma_id_private *id_priv;
2882         struct cma_multicast *mc = multicast->context;
2883         struct rdma_cm_event event;
2884         int ret;
2885
2886         id_priv = mc->id_priv;
2887         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2888             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2889                 return 0;
2890
2891         mutex_lock(&id_priv->qp_mutex);
2892         if (!status && id_priv->id.qp)
2893                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2894                                          multicast->rec.mlid);
2895         mutex_unlock(&id_priv->qp_mutex);
2896
2897         memset(&event, 0, sizeof event);
2898         event.status = status;
2899         event.param.ud.private_data = mc->context;
2900         if (!status) {
2901                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2902                 ib_init_ah_from_mcmember(id_priv->id.device,
2903                                          id_priv->id.port_num, &multicast->rec,
2904                                          &event.param.ud.ah_attr);
2905                 event.param.ud.qp_num = 0xFFFFFF;
2906                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2907         } else
2908                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2909
2910         ret = id_priv->id.event_handler(&id_priv->id, &event);
2911         if (ret) {
2912                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2913                 mutex_unlock(&id_priv->handler_mutex);
2914                 rdma_destroy_id(&id_priv->id);
2915                 return 0;
2916         }
2917
2918         mutex_unlock(&id_priv->handler_mutex);
2919         return 0;
2920 }
2921
2922 static void cma_set_mgid(struct rdma_id_private *id_priv,
2923                          struct sockaddr *addr, union ib_gid *mgid)
2924 {
2925         unsigned char mc_map[MAX_ADDR_LEN];
2926         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2927         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2928         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2929
2930         if (cma_any_addr(addr)) {
2931                 memset(mgid, 0, sizeof *mgid);
2932         } else if ((addr->sa_family == AF_INET6) &&
2933                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2934                                                                  0xFF10A01B)) {
2935                 /* IPv6 address is an SA assigned MGID. */
2936                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2937         } else if ((addr->sa_family == AF_INET6)) {
2938                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2939                 if (id_priv->id.ps == RDMA_PS_UDP)
2940                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2941                 *mgid = *(union ib_gid *) (mc_map + 4);
2942         } else {
2943                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2944                 if (id_priv->id.ps == RDMA_PS_UDP)
2945                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2946                 *mgid = *(union ib_gid *) (mc_map + 4);
2947         }
2948 }
2949
2950 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2951                                  struct cma_multicast *mc)
2952 {
2953         struct ib_sa_mcmember_rec rec;
2954         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2955         ib_sa_comp_mask comp_mask;
2956         int ret;
2957
2958         ib_addr_get_mgid(dev_addr, &rec.mgid);
2959         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2960                                      &rec.mgid, &rec);
2961         if (ret)
2962                 return ret;
2963
2964         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2965         if (id_priv->id.ps == RDMA_PS_UDP)
2966                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2967         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2968         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2969         rec.join_state = 1;
2970
2971         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2972                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2973                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2974                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2975                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2976
2977         if (id_priv->id.ps == RDMA_PS_IPOIB)
2978                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2979                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2980
2981         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2982                                                 id_priv->id.port_num, &rec,
2983                                                 comp_mask, GFP_KERNEL,
2984                                                 cma_ib_mc_handler, mc);
2985         if (IS_ERR(mc->multicast.ib))
2986                 return PTR_ERR(mc->multicast.ib);
2987
2988         return 0;
2989 }
2990
2991 static void iboe_mcast_work_handler(struct work_struct *work)
2992 {
2993         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
2994         struct cma_multicast *mc = mw->mc;
2995         struct ib_sa_multicast *m = mc->multicast.ib;
2996
2997         mc->multicast.ib->context = mc;
2998         cma_ib_mc_handler(0, m);
2999         kref_put(&mc->mcref, release_mc);
3000         kfree(mw);
3001 }
3002
3003 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3004 {
3005         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3006         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3007
3008         if (cma_any_addr(addr)) {
3009                 memset(mgid, 0, sizeof *mgid);
3010         } else if (addr->sa_family == AF_INET6) {
3011                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3012         } else {
3013                 mgid->raw[0] = 0xff;
3014                 mgid->raw[1] = 0x0e;
3015                 mgid->raw[2] = 0;
3016                 mgid->raw[3] = 0;
3017                 mgid->raw[4] = 0;
3018                 mgid->raw[5] = 0;
3019                 mgid->raw[6] = 0;
3020                 mgid->raw[7] = 0;
3021                 mgid->raw[8] = 0;
3022                 mgid->raw[9] = 0;
3023                 mgid->raw[10] = 0xff;
3024                 mgid->raw[11] = 0xff;
3025                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3026         }
3027 }
3028
3029 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3030                                    struct cma_multicast *mc)
3031 {
3032         struct iboe_mcast_work *work;
3033         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3034         int err;
3035         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3036         struct net_device *ndev = NULL;
3037
3038         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3039                 return -EINVAL;
3040
3041         work = kzalloc(sizeof *work, GFP_KERNEL);
3042         if (!work)
3043                 return -ENOMEM;
3044
3045         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3046         if (!mc->multicast.ib) {
3047                 err = -ENOMEM;
3048                 goto out1;
3049         }
3050
3051         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3052
3053         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3054         if (id_priv->id.ps == RDMA_PS_UDP)
3055                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3056
3057         if (dev_addr->bound_dev_if)
3058                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3059         if (!ndev) {
3060                 err = -ENODEV;
3061                 goto out2;
3062         }
3063         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3064         mc->multicast.ib->rec.hop_limit = 1;
3065         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3066         dev_put(ndev);
3067         if (!mc->multicast.ib->rec.mtu) {
3068                 err = -EINVAL;
3069                 goto out2;
3070         }
3071         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3072         work->id = id_priv;
3073         work->mc = mc;
3074         INIT_WORK(&work->work, iboe_mcast_work_handler);
3075         kref_get(&mc->mcref);
3076         queue_work(cma_wq, &work->work);
3077
3078         return 0;
3079
3080 out2:
3081         kfree(mc->multicast.ib);
3082 out1:
3083         kfree(work);
3084         return err;
3085 }
3086
3087 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3088                         void *context)
3089 {
3090         struct rdma_id_private *id_priv;
3091         struct cma_multicast *mc;
3092         int ret;
3093
3094         id_priv = container_of(id, struct rdma_id_private, id);
3095         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3096             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3097                 return -EINVAL;
3098
3099         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3100         if (!mc)
3101                 return -ENOMEM;
3102
3103         memcpy(&mc->addr, addr, ip_addr_size(addr));
3104         mc->context = context;
3105         mc->id_priv = id_priv;
3106
3107         spin_lock(&id_priv->lock);
3108         list_add(&mc->list, &id_priv->mc_list);
3109         spin_unlock(&id_priv->lock);
3110
3111         switch (rdma_node_get_transport(id->device->node_type)) {
3112         case RDMA_TRANSPORT_IB:
3113                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3114                 case IB_LINK_LAYER_INFINIBAND:
3115                         ret = cma_join_ib_multicast(id_priv, mc);
3116                         break;
3117                 case IB_LINK_LAYER_ETHERNET:
3118                         kref_init(&mc->mcref);
3119                         ret = cma_iboe_join_multicast(id_priv, mc);
3120                         break;
3121                 default:
3122                         ret = -EINVAL;
3123                 }
3124                 break;
3125         default:
3126                 ret = -ENOSYS;
3127                 break;
3128         }
3129
3130         if (ret) {
3131                 spin_lock_irq(&id_priv->lock);
3132                 list_del(&mc->list);
3133                 spin_unlock_irq(&id_priv->lock);
3134                 kfree(mc);
3135         }
3136         return ret;
3137 }
3138 EXPORT_SYMBOL(rdma_join_multicast);
3139
3140 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3141 {
3142         struct rdma_id_private *id_priv;
3143         struct cma_multicast *mc;
3144
3145         id_priv = container_of(id, struct rdma_id_private, id);
3146         spin_lock_irq(&id_priv->lock);
3147         list_for_each_entry(mc, &id_priv->mc_list, list) {
3148                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3149                         list_del(&mc->list);
3150                         spin_unlock_irq(&id_priv->lock);
3151
3152                         if (id->qp)
3153                                 ib_detach_mcast(id->qp,
3154                                                 &mc->multicast.ib->rec.mgid,
3155                                                 mc->multicast.ib->rec.mlid);
3156                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3157                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3158                                 case IB_LINK_LAYER_INFINIBAND:
3159                                         ib_sa_free_multicast(mc->multicast.ib);
3160                                         kfree(mc);
3161                                         break;
3162                                 case IB_LINK_LAYER_ETHERNET:
3163                                         kref_put(&mc->mcref, release_mc);
3164                                         break;
3165                                 default:
3166                                         break;
3167                                 }
3168                         }
3169                         return;
3170                 }
3171         }
3172         spin_unlock_irq(&id_priv->lock);
3173 }
3174 EXPORT_SYMBOL(rdma_leave_multicast);
3175
3176 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3177 {
3178         struct rdma_dev_addr *dev_addr;
3179         struct cma_ndev_work *work;
3180
3181         dev_addr = &id_priv->id.route.addr.dev_addr;
3182
3183         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3184             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3185                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3186                        ndev->name, &id_priv->id);
3187                 work = kzalloc(sizeof *work, GFP_KERNEL);
3188                 if (!work)
3189                         return -ENOMEM;
3190
3191                 INIT_WORK(&work->work, cma_ndev_work_handler);
3192                 work->id = id_priv;
3193                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3194                 atomic_inc(&id_priv->refcount);
3195                 queue_work(cma_wq, &work->work);
3196         }
3197
3198         return 0;
3199 }
3200
3201 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3202                                void *ctx)
3203 {
3204         struct net_device *ndev = (struct net_device *)ctx;
3205         struct cma_device *cma_dev;
3206         struct rdma_id_private *id_priv;
3207         int ret = NOTIFY_DONE;
3208
3209         if (dev_net(ndev) != &init_net)
3210                 return NOTIFY_DONE;
3211
3212         if (event != NETDEV_BONDING_FAILOVER)
3213                 return NOTIFY_DONE;
3214
3215         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3216                 return NOTIFY_DONE;
3217
3218         mutex_lock(&lock);
3219         list_for_each_entry(cma_dev, &dev_list, list)
3220                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3221                         ret = cma_netdev_change(ndev, id_priv);
3222                         if (ret)
3223                                 goto out;
3224                 }
3225
3226 out:
3227         mutex_unlock(&lock);
3228         return ret;
3229 }
3230
3231 static struct notifier_block cma_nb = {
3232         .notifier_call = cma_netdev_callback
3233 };
3234
3235 static void cma_add_one(struct ib_device *device)
3236 {
3237         struct cma_device *cma_dev;
3238         struct rdma_id_private *id_priv;
3239
3240         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3241         if (!cma_dev)
3242                 return;
3243
3244         cma_dev->device = device;
3245
3246         init_completion(&cma_dev->comp);
3247         atomic_set(&cma_dev->refcount, 1);
3248         INIT_LIST_HEAD(&cma_dev->id_list);
3249         ib_set_client_data(device, &cma_client, cma_dev);
3250
3251         mutex_lock(&lock);
3252         list_add_tail(&cma_dev->list, &dev_list);
3253         list_for_each_entry(id_priv, &listen_any_list, list)
3254                 cma_listen_on_dev(id_priv, cma_dev);
3255         mutex_unlock(&lock);
3256 }
3257
3258 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3259 {
3260         struct rdma_cm_event event;
3261         enum rdma_cm_state state;
3262         int ret = 0;
3263
3264         /* Record that we want to remove the device */
3265         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3266         if (state == RDMA_CM_DESTROYING)
3267                 return 0;
3268
3269         cma_cancel_operation(id_priv, state);
3270         mutex_lock(&id_priv->handler_mutex);
3271
3272         /* Check for destruction from another callback. */
3273         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3274                 goto out;
3275
3276         memset(&event, 0, sizeof event);
3277         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3278         ret = id_priv->id.event_handler(&id_priv->id, &event);
3279 out:
3280         mutex_unlock(&id_priv->handler_mutex);
3281         return ret;
3282 }
3283
3284 static void cma_process_remove(struct cma_device *cma_dev)
3285 {
3286         struct rdma_id_private *id_priv;
3287         int ret;
3288
3289         mutex_lock(&lock);
3290         while (!list_empty(&cma_dev->id_list)) {
3291                 id_priv = list_entry(cma_dev->id_list.next,
3292                                      struct rdma_id_private, list);
3293
3294                 list_del(&id_priv->listen_list);
3295                 list_del_init(&id_priv->list);
3296                 atomic_inc(&id_priv->refcount);
3297                 mutex_unlock(&lock);
3298
3299                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3300                 cma_deref_id(id_priv);
3301                 if (ret)
3302                         rdma_destroy_id(&id_priv->id);
3303
3304                 mutex_lock(&lock);
3305         }
3306         mutex_unlock(&lock);
3307
3308         cma_deref_dev(cma_dev);
3309         wait_for_completion(&cma_dev->comp);
3310 }
3311
3312 static void cma_remove_one(struct ib_device *device)
3313 {
3314         struct cma_device *cma_dev;
3315
3316         cma_dev = ib_get_client_data(device, &cma_client);
3317         if (!cma_dev)
3318                 return;
3319
3320         mutex_lock(&lock);
3321         list_del(&cma_dev->list);
3322         mutex_unlock(&lock);
3323
3324         cma_process_remove(cma_dev);
3325         kfree(cma_dev);
3326 }
3327
3328 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3329 {
3330         struct nlmsghdr *nlh;
3331         struct rdma_cm_id_stats *id_stats;
3332         struct rdma_id_private *id_priv;
3333         struct rdma_cm_id *id = NULL;
3334         struct cma_device *cma_dev;
3335         int i_dev = 0, i_id = 0;
3336
3337         /*
3338          * We export all of the IDs as a sequence of messages.  Each
3339          * ID gets its own netlink message.
3340          */
3341         mutex_lock(&lock);
3342
3343         list_for_each_entry(cma_dev, &dev_list, list) {
3344                 if (i_dev < cb->args[0]) {
3345                         i_dev++;
3346                         continue;
3347                 }
3348
3349                 i_id = 0;
3350                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3351                         if (i_id < cb->args[1]) {
3352                                 i_id++;
3353                                 continue;
3354                         }
3355
3356                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3357                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3358                                                 RDMA_NL_RDMA_CM_ID_STATS);
3359                         if (!id_stats)
3360                                 goto out;
3361
3362                         memset(id_stats, 0, sizeof *id_stats);
3363                         id = &id_priv->id;
3364                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3365                         id_stats->port_num = id->port_num;
3366                         id_stats->bound_dev_if =
3367                                 id->route.addr.dev_addr.bound_dev_if;
3368
3369                         if (id->route.addr.src_addr.ss_family == AF_INET) {
3370                                 if (ibnl_put_attr(skb, nlh,
3371                                                   sizeof(struct sockaddr_in),
3372                                                   &id->route.addr.src_addr,
3373                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3374                                         goto out;
3375                                 }
3376                                 if (ibnl_put_attr(skb, nlh,
3377                                                   sizeof(struct sockaddr_in),
3378                                                   &id->route.addr.dst_addr,
3379                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3380                                         goto out;
3381                                 }
3382                         } else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3383                                 if (ibnl_put_attr(skb, nlh,
3384                                                   sizeof(struct sockaddr_in6),
3385                                                   &id->route.addr.src_addr,
3386                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3387                                         goto out;
3388                                 }
3389                                 if (ibnl_put_attr(skb, nlh,
3390                                                   sizeof(struct sockaddr_in6),
3391                                                   &id->route.addr.dst_addr,
3392                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3393                                         goto out;
3394                                 }
3395                         }
3396
3397                         id_stats->pid           = id_priv->owner;
3398                         id_stats->port_space    = id->ps;
3399                         id_stats->cm_state      = id_priv->state;
3400                         id_stats->qp_num        = id_priv->qp_num;
3401                         id_stats->qp_type       = id->qp_type;
3402
3403                         i_id++;
3404                 }
3405
3406                 cb->args[1] = 0;
3407                 i_dev++;
3408         }
3409
3410 out:
3411         mutex_unlock(&lock);
3412         cb->args[0] = i_dev;
3413         cb->args[1] = i_id;
3414
3415         return skb->len;
3416 }
3417
3418 static const struct ibnl_client_cbs cma_cb_table[] = {
3419         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats },
3420 };
3421
3422 static int __init cma_init(void)
3423 {
3424         int ret;
3425
3426         cma_wq = create_singlethread_workqueue("rdma_cm");
3427         if (!cma_wq)
3428                 return -ENOMEM;
3429
3430         ib_sa_register_client(&sa_client);
3431         rdma_addr_register_client(&addr_client);
3432         register_netdevice_notifier(&cma_nb);
3433
3434         ret = ib_register_client(&cma_client);
3435         if (ret)
3436                 goto err;
3437
3438         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3439                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3440
3441         return 0;
3442
3443 err:
3444         unregister_netdevice_notifier(&cma_nb);
3445         rdma_addr_unregister_client(&addr_client);
3446         ib_sa_unregister_client(&sa_client);
3447         destroy_workqueue(cma_wq);
3448         return ret;
3449 }
3450
3451 static void __exit cma_cleanup(void)
3452 {
3453         ibnl_remove_client(RDMA_NL_RDMA_CM);
3454         ib_unregister_client(&cma_client);
3455         unregister_netdevice_notifier(&cma_nb);
3456         rdma_addr_unregister_client(&addr_client);
3457         ib_sa_unregister_client(&sa_client);
3458         destroy_workqueue(cma_wq);
3459         idr_destroy(&sdp_ps);
3460         idr_destroy(&tcp_ps);
3461         idr_destroy(&udp_ps);
3462         idr_destroy(&ipoib_ps);
3463 }
3464
3465 module_init(cma_init);
3466 module_exit(cma_cleanup);