Pileus Git - ~andy/linux/blob - 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 #include <linux/module.h>
  45 #include <net/route.h>
  46
  47 #include <net/tcp.h>
  48 #include <net/ipv6.h>
  49
  50 #include <rdma/rdma_cm.h>
  51 #include <rdma/rdma_cm_ib.h>
  52 #include <rdma/rdma_netlink.h>
  53 #include <rdma/ib_cache.h>
  54 #include <rdma/ib_cm.h>
  55 #include <rdma/ib_sa.h>
  56 #include <rdma/iw_cm.h>
  57
  58 MODULE_AUTHOR("Sean Hefty");
  59 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  60 MODULE_LICENSE("Dual BSD/GPL");
  61
  62 #define CMA_CM_RESPONSE_TIMEOUT 20
  63 #define CMA_MAX_CM_RETRIES 15
  64 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  65 #define CMA_IBOE_PACKET_LIFETIME 18
  66
  67 static void cma_add_one(struct ib_device *device);
  68 static void cma_remove_one(struct ib_device *device);
  69
  70 static struct ib_client cma_client = {
  71         .name   = "cma",
  72         .add    = cma_add_one,
  73         .remove = cma_remove_one
  74 };
  75
  76 static struct ib_sa_client sa_client;
  77 static struct rdma_addr_client addr_client;
  78 static LIST_HEAD(dev_list);
  79 static LIST_HEAD(listen_any_list);
  80 static DEFINE_MUTEX(lock);
  81 static struct workqueue_struct *cma_wq;
  82 static DEFINE_IDR(sdp_ps);
  83 static DEFINE_IDR(tcp_ps);
  84 static DEFINE_IDR(udp_ps);
  85 static DEFINE_IDR(ipoib_ps);
  86 static DEFINE_IDR(ib_ps);
  87
  88 struct cma_device {
  89         struct list_head        list;
  90         struct ib_device        *device;
  91         struct completion       comp;
  92         atomic_t                refcount;
  93         struct list_head        id_list;
  94 };
  95
  96 struct rdma_bind_list {
  97         struct idr              *ps;
  98         struct hlist_head       owners;
  99         unsigned short          port;
 100 };
 101
 102 /*
 103  * Device removal can occur at anytime, so we need extra handling to
 104  * serialize notifying the user of device removal with other callbacks.
 105  * We do this by disabling removal notification while a callback is in process,
 106  * and reporting it after the callback completes.
 107  */
 108 struct rdma_id_private {
 109         struct rdma_cm_id       id;
 110
 111         struct rdma_bind_list   *bind_list;
 112         struct hlist_node       node;
 113         struct list_head        list; /* listen_any_list or cma_device.list */
 114         struct list_head        listen_list; /* per device listens */
 115         struct cma_device       *cma_dev;
 116         struct list_head        mc_list;
 117
 118         int                     internal_id;
 119         enum rdma_cm_state      state;
 120         spinlock_t              lock;
 121         struct mutex            qp_mutex;
 122
 123         struct completion       comp;
 124         atomic_t                refcount;
 125         struct mutex            handler_mutex;
 126
 127         int                     backlog;
 128         int                     timeout_ms;
 129         struct ib_sa_query      *query;
 130         int                     query_id;
 131         union {
 132                 struct ib_cm_id *ib;
 133                 struct iw_cm_id *iw;
 134         } cm_id;
 135
 136         u32                     seq_num;
 137         u32                     qkey;
 138         u32                     qp_num;
 139         pid_t                   owner;
 140         u8                      srq;
 141         u8                      tos;
 142         u8                      reuseaddr;
 143 };
 144
 145 struct cma_multicast {
 146         struct rdma_id_private *id_priv;
 147         union {
 148                 struct ib_sa_multicast *ib;
 149         } multicast;
 150         struct list_head        list;
 151         void                    *context;
 152         struct sockaddr_storage addr;
 153         struct kref             mcref;
 154 };
 155
 156 struct cma_work {
 157         struct work_struct      work;
 158         struct rdma_id_private  *id;
 159         enum rdma_cm_state      old_state;
 160         enum rdma_cm_state      new_state;
 161         struct rdma_cm_event    event;
 162 };
 163
 164 struct cma_ndev_work {
 165         struct work_struct      work;
 166         struct rdma_id_private  *id;
 167         struct rdma_cm_event    event;
 168 };
 169
 170 struct iboe_mcast_work {
 171         struct work_struct       work;
 172         struct rdma_id_private  *id;
 173         struct cma_multicast    *mc;
 174 };
 175
 176 union cma_ip_addr {
 177         struct in6_addr ip6;
 178         struct {
 179                 __be32 pad[3];
 180                 __be32 addr;
 181         } ip4;
 182 };
 183
 184 struct cma_hdr {
 185         u8 cma_version;
 186         u8 ip_version;  /* IP version: 7:4 */
 187         __be16 port;
 188         union cma_ip_addr src_addr;
 189         union cma_ip_addr dst_addr;
 190 };
 191
 192 struct sdp_hh {
 193         u8 bsdh[16];
 194         u8 sdp_version; /* Major version: 7:4 */
 195         u8 ip_version;  /* IP version: 7:4 */
 196         u8 sdp_specific1[10];
 197         __be16 port;
 198         __be16 sdp_specific2;
 199         union cma_ip_addr src_addr;
 200         union cma_ip_addr dst_addr;
 201 };
 202
 203 struct sdp_hah {
 204         u8 bsdh[16];
 205         u8 sdp_version;
 206 };
 207
 208 #define CMA_VERSION 0x00
 209 #define SDP_MAJ_VERSION 0x2
 210
 211 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 212 {
 213         unsigned long flags;
 214         int ret;
 215
 216         spin_lock_irqsave(&id_priv->lock, flags);
 217         ret = (id_priv->state == comp);
 218         spin_unlock_irqrestore(&id_priv->lock, flags);
 219         return ret;
 220 }
 221
 222 static int cma_comp_exch(struct rdma_id_private *id_priv,
 223                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 224 {
 225         unsigned long flags;
 226         int ret;
 227
 228         spin_lock_irqsave(&id_priv->lock, flags);
 229         if ((ret = (id_priv->state == comp)))
 230                 id_priv->state = exch;
 231         spin_unlock_irqrestore(&id_priv->lock, flags);
 232         return ret;
 233 }
 234
 235 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 236                                    enum rdma_cm_state exch)
 237 {
 238         unsigned long flags;
 239         enum rdma_cm_state old;
 240
 241         spin_lock_irqsave(&id_priv->lock, flags);
 242         old = id_priv->state;
 243         id_priv->state = exch;
 244         spin_unlock_irqrestore(&id_priv->lock, flags);
 245         return old;
 246 }
 247
 248 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 249 {
 250         return hdr->ip_version >> 4;
 251 }
 252
 253 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 254 {
 255         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 256 }
 257
 258 static inline u8 sdp_get_majv(u8 sdp_version)
 259 {
 260         return sdp_version >> 4;
 261 }
 262
 263 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 264 {
 265         return hh->ip_version >> 4;
 266 }
 267
 268 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 269 {
 270         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 271 }
 272
 273 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 274                               struct cma_device *cma_dev)
 275 {
 276         atomic_inc(&cma_dev->refcount);
 277         id_priv->cma_dev = cma_dev;
 278         id_priv->id.device = cma_dev->device;
 279         id_priv->id.route.addr.dev_addr.transport =
 280                 rdma_node_get_transport(cma_dev->device->node_type);
 281         list_add_tail(&id_priv->list, &cma_dev->id_list);
 282 }
 283
 284 static inline void cma_deref_dev(struct cma_device *cma_dev)
 285 {
 286         if (atomic_dec_and_test(&cma_dev->refcount))
 287                 complete(&cma_dev->comp);
 288 }
 289
 290 static inline void release_mc(struct kref *kref)
 291 {
 292         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 293
 294         kfree(mc->multicast.ib);
 295         kfree(mc);
 296 }
 297
 298 static void cma_release_dev(struct rdma_id_private *id_priv)
 299 {
 300         mutex_lock(&lock);
 301         list_del(&id_priv->list);
 302         cma_deref_dev(id_priv->cma_dev);
 303         id_priv->cma_dev = NULL;
 304         mutex_unlock(&lock);
 305 }
 306
 307 static int cma_set_qkey(struct rdma_id_private *id_priv)
 308 {
 309         struct ib_sa_mcmember_rec rec;
 310         int ret = 0;
 311
 312         if (id_priv->qkey)
 313                 return 0;
 314
 315         switch (id_priv->id.ps) {
 316         case RDMA_PS_UDP:
 317                 id_priv->qkey = RDMA_UDP_QKEY;
 318                 break;
 319         case RDMA_PS_IPOIB:
 320                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 321                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 322                                              id_priv->id.port_num, &rec.mgid,
 323                                              &rec);
 324                 if (!ret)
 325                         id_priv->qkey = be32_to_cpu(rec.qkey);
 326                 break;
 327         default:
 328                 break;
 329         }
 330         return ret;
 331 }
 332
 333 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 334 {
 335         int i;
 336         int err;
 337         struct ib_port_attr props;
 338         union ib_gid tmp;
 339
 340         err = ib_query_port(device, port_num, &props);
 341         if (err)
 342                 return 1;
 343
 344         for (i = 0; i < props.gid_tbl_len; ++i) {
 345                 err = ib_query_gid(device, port_num, i, &tmp);
 346                 if (err)
 347                         return 1;
 348                 if (!memcmp(&tmp, gid, sizeof tmp))
 349                         return 0;
 350         }
 351
 352         return -EAGAIN;
 353 }
 354
 355 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 356 {
 357         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 358         struct cma_device *cma_dev;
 359         union ib_gid gid, iboe_gid;
 360         int ret = -ENODEV;
 361         u8 port;
 362         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 363                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 364
 365         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 366             id_priv->id.ps == RDMA_PS_IPOIB)
 367                 return -EINVAL;
 368
 369         mutex_lock(&lock);
 370         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 371         memcpy(&gid, dev_addr->src_dev_addr +
 372                rdma_addr_gid_offset(dev_addr), sizeof gid);
 373         list_for_each_entry(cma_dev, &dev_list, list) {
 374                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 375                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 376                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 377                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 378                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 379                                 else
 380                                         ret = find_gid_port(cma_dev->device, &gid, port);
 381
 382                                 if (!ret) {
 383                                         id_priv->id.port_num = port;
 384                                         goto out;
 385                                 } else if (ret == 1)
 386                                         break;
 387                         }
 388                 }
 389         }
 390
 391 out:
 392         if (!ret)
 393                 cma_attach_to_dev(id_priv, cma_dev);
 394
 395         mutex_unlock(&lock);
 396         return ret;
 397 }
 398
 399 static void cma_deref_id(struct rdma_id_private *id_priv)
 400 {
 401         if (atomic_dec_and_test(&id_priv->refcount))
 402                 complete(&id_priv->comp);
 403 }
 404
 405 static int cma_disable_callback(struct rdma_id_private *id_priv,
 406                                 enum rdma_cm_state state)
 407 {
 408         mutex_lock(&id_priv->handler_mutex);
 409         if (id_priv->state != state) {
 410                 mutex_unlock(&id_priv->handler_mutex);
 411                 return -EINVAL;
 412         }
 413         return 0;
 414 }
 415
 416 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 417                                   void *context, enum rdma_port_space ps,
 418                                   enum ib_qp_type qp_type)
 419 {
 420         struct rdma_id_private *id_priv;
 421
 422         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 423         if (!id_priv)
 424                 return ERR_PTR(-ENOMEM);
 425
 426         id_priv->owner = task_pid_nr(current);
 427         id_priv->state = RDMA_CM_IDLE;
 428         id_priv->id.context = context;
 429         id_priv->id.event_handler = event_handler;
 430         id_priv->id.ps = ps;
 431         id_priv->id.qp_type = qp_type;
 432         spin_lock_init(&id_priv->lock);
 433         mutex_init(&id_priv->qp_mutex);
 434         init_completion(&id_priv->comp);
 435         atomic_set(&id_priv->refcount, 1);
 436         mutex_init(&id_priv->handler_mutex);
 437         INIT_LIST_HEAD(&id_priv->listen_list);
 438         INIT_LIST_HEAD(&id_priv->mc_list);
 439         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 440
 441         return &id_priv->id;
 442 }
 443 EXPORT_SYMBOL(rdma_create_id);
 444
 445 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 446 {
 447         struct ib_qp_attr qp_attr;
 448         int qp_attr_mask, ret;
 449
 450         qp_attr.qp_state = IB_QPS_INIT;
 451         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 452         if (ret)
 453                 return ret;
 454
 455         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 456         if (ret)
 457                 return ret;
 458
 459         qp_attr.qp_state = IB_QPS_RTR;
 460         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 461         if (ret)
 462                 return ret;
 463
 464         qp_attr.qp_state = IB_QPS_RTS;
 465         qp_attr.sq_psn = 0;
 466         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 467
 468         return ret;
 469 }
 470
 471 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 472 {
 473         struct ib_qp_attr qp_attr;
 474         int qp_attr_mask, ret;
 475
 476         qp_attr.qp_state = IB_QPS_INIT;
 477         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 478         if (ret)
 479                 return ret;
 480
 481         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 482 }
 483
 484 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 485                    struct ib_qp_init_attr *qp_init_attr)
 486 {
 487         struct rdma_id_private *id_priv;
 488         struct ib_qp *qp;
 489         int ret;
 490
 491         id_priv = container_of(id, struct rdma_id_private, id);
 492         if (id->device != pd->device)
 493                 return -EINVAL;
 494
 495         qp = ib_create_qp(pd, qp_init_attr);
 496         if (IS_ERR(qp))
 497                 return PTR_ERR(qp);
 498
 499         if (id->qp_type == IB_QPT_UD)
 500                 ret = cma_init_ud_qp(id_priv, qp);
 501         else
 502                 ret = cma_init_conn_qp(id_priv, qp);
 503         if (ret)
 504                 goto err;
 505
 506         id->qp = qp;
 507         id_priv->qp_num = qp->qp_num;
 508         id_priv->srq = (qp->srq != NULL);
 509         return 0;
 510 err:
 511         ib_destroy_qp(qp);
 512         return ret;
 513 }
 514 EXPORT_SYMBOL(rdma_create_qp);
 515
 516 void rdma_destroy_qp(struct rdma_cm_id *id)
 517 {
 518         struct rdma_id_private *id_priv;
 519
 520         id_priv = container_of(id, struct rdma_id_private, id);
 521         mutex_lock(&id_priv->qp_mutex);
 522         ib_destroy_qp(id_priv->id.qp);
 523         id_priv->id.qp = NULL;
 524         mutex_unlock(&id_priv->qp_mutex);
 525 }
 526 EXPORT_SYMBOL(rdma_destroy_qp);
 527
 528 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 529                              struct rdma_conn_param *conn_param)
 530 {
 531         struct ib_qp_attr qp_attr;
 532         int qp_attr_mask, ret;
 533
 534         mutex_lock(&id_priv->qp_mutex);
 535         if (!id_priv->id.qp) {
 536                 ret = 0;
 537                 goto out;
 538         }
 539
 540         /* Need to update QP attributes from default values. */
 541         qp_attr.qp_state = IB_QPS_INIT;
 542         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 543         if (ret)
 544                 goto out;
 545
 546         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 547         if (ret)
 548                 goto out;
 549
 550         qp_attr.qp_state = IB_QPS_RTR;
 551         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 552         if (ret)
 553                 goto out;
 554
 555         if (conn_param)
 556                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 557         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 558 out:
 559         mutex_unlock(&id_priv->qp_mutex);
 560         return ret;
 561 }
 562
 563 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 564                              struct rdma_conn_param *conn_param)
 565 {
 566         struct ib_qp_attr qp_attr;
 567         int qp_attr_mask, ret;
 568
 569         mutex_lock(&id_priv->qp_mutex);
 570         if (!id_priv->id.qp) {
 571                 ret = 0;
 572                 goto out;
 573         }
 574
 575         qp_attr.qp_state = IB_QPS_RTS;
 576         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 577         if (ret)
 578                 goto out;
 579
 580         if (conn_param)
 581                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 582         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 583 out:
 584         mutex_unlock(&id_priv->qp_mutex);
 585         return ret;
 586 }
 587
 588 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 589 {
 590         struct ib_qp_attr qp_attr;
 591         int ret;
 592
 593         mutex_lock(&id_priv->qp_mutex);
 594         if (!id_priv->id.qp) {
 595                 ret = 0;
 596                 goto out;
 597         }
 598
 599         qp_attr.qp_state = IB_QPS_ERR;
 600         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 601 out:
 602         mutex_unlock(&id_priv->qp_mutex);
 603         return ret;
 604 }
 605
 606 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 607                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 608 {
 609         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 610         int ret;
 611         u16 pkey;
 612
 613         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 614             IB_LINK_LAYER_INFINIBAND)
 615                 pkey = ib_addr_get_pkey(dev_addr);
 616         else
 617                 pkey = 0xffff;
 618
 619         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 620                                   pkey, &qp_attr->pkey_index);
 621         if (ret)
 622                 return ret;
 623
 624         qp_attr->port_num = id_priv->id.port_num;
 625         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 626
 627         if (id_priv->id.qp_type == IB_QPT_UD) {
 628                 ret = cma_set_qkey(id_priv);
 629                 if (ret)
 630                         return ret;
 631
 632                 qp_attr->qkey = id_priv->qkey;
 633                 *qp_attr_mask |= IB_QP_QKEY;
 634         } else {
 635                 qp_attr->qp_access_flags = 0;
 636                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 637         }
 638         return 0;
 639 }
 640
 641 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 642                        int *qp_attr_mask)
 643 {
 644         struct rdma_id_private *id_priv;
 645         int ret = 0;
 646
 647         id_priv = container_of(id, struct rdma_id_private, id);
 648         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 649         case RDMA_TRANSPORT_IB:
 650                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 651                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 652                 else
 653                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 654                                                  qp_attr_mask);
 655                 if (qp_attr->qp_state == IB_QPS_RTR)
 656                         qp_attr->rq_psn = id_priv->seq_num;
 657                 break;
 658         case RDMA_TRANSPORT_IWARP:
 659                 if (!id_priv->cm_id.iw) {
 660                         qp_attr->qp_access_flags = 0;
 661                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 662                 } else
 663                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 664                                                  qp_attr_mask);
 665                 break;
 666         default:
 667                 ret = -ENOSYS;
 668                 break;
 669         }
 670
 671         return ret;
 672 }
 673 EXPORT_SYMBOL(rdma_init_qp_attr);
 674
 675 static inline int cma_zero_addr(struct sockaddr *addr)
 676 {
 677         struct in6_addr *ip6;
 678
 679         if (addr->sa_family == AF_INET)
 680                 return ipv4_is_zeronet(
 681                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 682         else {
 683                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 684                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 685                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 686         }
 687 }
 688
 689 static inline int cma_loopback_addr(struct sockaddr *addr)
 690 {
 691         if (addr->sa_family == AF_INET)
 692                 return ipv4_is_loopback(
 693                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 694         else
 695                 return ipv6_addr_loopback(
 696                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 697 }
 698
 699 static inline int cma_any_addr(struct sockaddr *addr)
 700 {
 701         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 702 }
 703
 704 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 705 {
 706         if (src->sa_family != dst->sa_family)
 707                 return -1;
 708
 709         switch (src->sa_family) {
 710         case AF_INET:
 711                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 712                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 713         default:
 714                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 715                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 716         }
 717 }
 718
 719 static inline __be16 cma_port(struct sockaddr *addr)
 720 {
 721         if (addr->sa_family == AF_INET)
 722                 return ((struct sockaddr_in *) addr)->sin_port;
 723         else
 724                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 725 }
 726
 727 static inline int cma_any_port(struct sockaddr *addr)
 728 {
 729         return !cma_port(addr);
 730 }
 731
 732 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 733                             u8 *ip_ver, __be16 *port,
 734                             union cma_ip_addr **src, union cma_ip_addr **dst)
 735 {
 736         switch (ps) {
 737         case RDMA_PS_SDP:
 738                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 739                     SDP_MAJ_VERSION)
 740                         return -EINVAL;
 741
 742                 *ip_ver = sdp_get_ip_ver(hdr);
 743                 *port   = ((struct sdp_hh *) hdr)->port;
 744                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 745                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 746                 break;
 747         default:
 748                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 749                         return -EINVAL;
 750
 751                 *ip_ver = cma_get_ip_ver(hdr);
 752                 *port   = ((struct cma_hdr *) hdr)->port;
 753                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 754                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 755                 break;
 756         }
 757
 758         if (*ip_ver != 4 && *ip_ver != 6)
 759                 return -EINVAL;
 760         return 0;
 761 }
 762
 763 static void cma_save_net_info(struct rdma_addr *addr,
 764                               struct rdma_addr *listen_addr,
 765                               u8 ip_ver, __be16 port,
 766                               union cma_ip_addr *src, union cma_ip_addr *dst)
 767 {
 768         struct sockaddr_in *listen4, *ip4;
 769         struct sockaddr_in6 *listen6, *ip6;
 770
 771         switch (ip_ver) {
 772         case 4:
 773                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 774                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 775                 ip4->sin_family = listen4->sin_family;
 776                 ip4->sin_addr.s_addr = dst->ip4.addr;
 777                 ip4->sin_port = listen4->sin_port;
 778
 779                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 780                 ip4->sin_family = listen4->sin_family;
 781                 ip4->sin_addr.s_addr = src->ip4.addr;
 782                 ip4->sin_port = port;
 783                 break;
 784         case 6:
 785                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 786                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 787                 ip6->sin6_family = listen6->sin6_family;
 788                 ip6->sin6_addr = dst->ip6;
 789                 ip6->sin6_port = listen6->sin6_port;
 790
 791                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 792                 ip6->sin6_family = listen6->sin6_family;
 793                 ip6->sin6_addr = src->ip6;
 794                 ip6->sin6_port = port;
 795                 break;
 796         default:
 797                 break;
 798         }
 799 }
 800
 801 static inline int cma_user_data_offset(enum rdma_port_space ps)
 802 {
 803         switch (ps) {
 804         case RDMA_PS_SDP:
 805                 return 0;
 806         default:
 807                 return sizeof(struct cma_hdr);
 808         }
 809 }
 810
 811 static void cma_cancel_route(struct rdma_id_private *id_priv)
 812 {
 813         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 814         case IB_LINK_LAYER_INFINIBAND:
 815                 if (id_priv->query)
 816                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 817                 break;
 818         default:
 819                 break;
 820         }
 821 }
 822
 823 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 824 {
 825         struct rdma_id_private *dev_id_priv;
 826
 827         /*
 828          * Remove from listen_any_list to prevent added devices from spawning
 829          * additional listen requests.
 830          */
 831         mutex_lock(&lock);
 832         list_del(&id_priv->list);
 833
 834         while (!list_empty(&id_priv->listen_list)) {
 835                 dev_id_priv = list_entry(id_priv->listen_list.next,
 836                                          struct rdma_id_private, listen_list);
 837                 /* sync with device removal to avoid duplicate destruction */
 838                 list_del_init(&dev_id_priv->list);
 839                 list_del(&dev_id_priv->listen_list);
 840                 mutex_unlock(&lock);
 841
 842                 rdma_destroy_id(&dev_id_priv->id);
 843                 mutex_lock(&lock);
 844         }
 845         mutex_unlock(&lock);
 846 }
 847
 848 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 849                                  enum rdma_cm_state state)
 850 {
 851         switch (state) {
 852         case RDMA_CM_ADDR_QUERY:
 853                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 854                 break;
 855         case RDMA_CM_ROUTE_QUERY:
 856                 cma_cancel_route(id_priv);
 857                 break;
 858         case RDMA_CM_LISTEN:
 859                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 860                                 && !id_priv->cma_dev)
 861                         cma_cancel_listens(id_priv);
 862                 break;
 863         default:
 864                 break;
 865         }
 866 }
 867
 868 static void cma_release_port(struct rdma_id_private *id_priv)
 869 {
 870         struct rdma_bind_list *bind_list = id_priv->bind_list;
 871
 872         if (!bind_list)
 873                 return;
 874
 875         mutex_lock(&lock);
 876         hlist_del(&id_priv->node);
 877         if (hlist_empty(&bind_list->owners)) {
 878                 idr_remove(bind_list->ps, bind_list->port);
 879                 kfree(bind_list);
 880         }
 881         mutex_unlock(&lock);
 882 }
 883
 884 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 885 {
 886         struct cma_multicast *mc;
 887
 888         while (!list_empty(&id_priv->mc_list)) {
 889                 mc = container_of(id_priv->mc_list.next,
 890                                   struct cma_multicast, list);
 891                 list_del(&mc->list);
 892                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 893                 case IB_LINK_LAYER_INFINIBAND:
 894                         ib_sa_free_multicast(mc->multicast.ib);
 895                         kfree(mc);
 896                         break;
 897                 case IB_LINK_LAYER_ETHERNET:
 898                         kref_put(&mc->mcref, release_mc);
 899                         break;
 900                 default:
 901                         break;
 902                 }
 903         }
 904 }
 905
 906 void rdma_destroy_id(struct rdma_cm_id *id)
 907 {
 908         struct rdma_id_private *id_priv;
 909         enum rdma_cm_state state;
 910
 911         id_priv = container_of(id, struct rdma_id_private, id);
 912         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 913         cma_cancel_operation(id_priv, state);
 914
 915         /*
 916          * Wait for any active callback to finish.  New callbacks will find
 917          * the id_priv state set to destroying and abort.
 918          */
 919         mutex_lock(&id_priv->handler_mutex);
 920         mutex_unlock(&id_priv->handler_mutex);
 921
 922         if (id_priv->cma_dev) {
 923                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 924                 case RDMA_TRANSPORT_IB:
 925                         if (id_priv->cm_id.ib)
 926                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 927                         break;
 928                 case RDMA_TRANSPORT_IWARP:
 929                         if (id_priv->cm_id.iw)
 930                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 931                         break;
 932                 default:
 933                         break;
 934                 }
 935                 cma_leave_mc_groups(id_priv);
 936                 cma_release_dev(id_priv);
 937         }
 938
 939         cma_release_port(id_priv);
 940         cma_deref_id(id_priv);
 941         wait_for_completion(&id_priv->comp);
 942
 943         if (id_priv->internal_id)
 944                 cma_deref_id(id_priv->id.context);
 945
 946         kfree(id_priv->id.route.path_rec);
 947         kfree(id_priv);
 948 }
 949 EXPORT_SYMBOL(rdma_destroy_id);
 950
 951 static int cma_rep_recv(struct rdma_id_private *id_priv)
 952 {
 953         int ret;
 954
 955         ret = cma_modify_qp_rtr(id_priv, NULL);
 956         if (ret)
 957                 goto reject;
 958
 959         ret = cma_modify_qp_rts(id_priv, NULL);
 960         if (ret)
 961                 goto reject;
 962
 963         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 964         if (ret)
 965                 goto reject;
 966
 967         return 0;
 968 reject:
 969         cma_modify_qp_err(id_priv);
 970         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 971                        NULL, 0, NULL, 0);
 972         return ret;
 973 }
 974
 975 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 976 {
 977         if (id_priv->id.ps == RDMA_PS_SDP &&
 978             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 979             SDP_MAJ_VERSION)
 980                 return -EINVAL;
 981
 982         return 0;
 983 }
 984
 985 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 986                                    struct ib_cm_rep_event_param *rep_data,
 987                                    void *private_data)
 988 {
 989         event->param.conn.private_data = private_data;
 990         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 991         event->param.conn.responder_resources = rep_data->responder_resources;
 992         event->param.conn.initiator_depth = rep_data->initiator_depth;
 993         event->param.conn.flow_control = rep_data->flow_control;
 994         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 995         event->param.conn.srq = rep_data->srq;
 996         event->param.conn.qp_num = rep_data->remote_qpn;
 997 }
 998
 999 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1000 {
1001         struct rdma_id_private *id_priv = cm_id->context;
1002         struct rdma_cm_event event;
1003         int ret = 0;
1004
1005         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1006                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1007             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1008                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1009                 return 0;
1010
1011         memset(&event, 0, sizeof event);
1012         switch (ib_event->event) {
1013         case IB_CM_REQ_ERROR:
1014         case IB_CM_REP_ERROR:
1015                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1016                 event.status = -ETIMEDOUT;
1017                 break;
1018         case IB_CM_REP_RECEIVED:
1019                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1020                 if (event.status)
1021                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1022                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1023                         event.status = cma_rep_recv(id_priv);
1024                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1025                                                      RDMA_CM_EVENT_ESTABLISHED;
1026                 } else
1027                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1028                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1029                                        ib_event->private_data);
1030                 break;
1031         case IB_CM_RTU_RECEIVED:
1032         case IB_CM_USER_ESTABLISHED:
1033                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1034                 break;
1035         case IB_CM_DREQ_ERROR:
1036                 event.status = -ETIMEDOUT; /* fall through */
1037         case IB_CM_DREQ_RECEIVED:
1038         case IB_CM_DREP_RECEIVED:
1039                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1040                                    RDMA_CM_DISCONNECT))
1041                         goto out;
1042                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1043                 break;
1044         case IB_CM_TIMEWAIT_EXIT:
1045                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1046                 break;
1047         case IB_CM_MRA_RECEIVED:
1048                 /* ignore event */
1049                 goto out;
1050         case IB_CM_REJ_RECEIVED:
1051                 cma_modify_qp_err(id_priv);
1052                 event.status = ib_event->param.rej_rcvd.reason;
1053                 event.event = RDMA_CM_EVENT_REJECTED;
1054                 event.param.conn.private_data = ib_event->private_data;
1055                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1056                 break;
1057         default:
1058                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1059                        ib_event->event);
1060                 goto out;
1061         }
1062
1063         ret = id_priv->id.event_handler(&id_priv->id, &event);
1064         if (ret) {
1065                 /* Destroy the CM ID by returning a non-zero value. */
1066                 id_priv->cm_id.ib = NULL;
1067                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1068                 mutex_unlock(&id_priv->handler_mutex);
1069                 rdma_destroy_id(&id_priv->id);
1070                 return ret;
1071         }
1072 out:
1073         mutex_unlock(&id_priv->handler_mutex);
1074         return ret;
1075 }
1076
1077 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1078                                                struct ib_cm_event *ib_event)
1079 {
1080         struct rdma_id_private *id_priv;
1081         struct rdma_cm_id *id;
1082         struct rdma_route *rt;
1083         union cma_ip_addr *src, *dst;
1084         __be16 port;
1085         u8 ip_ver;
1086         int ret;
1087
1088         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1089                              &ip_ver, &port, &src, &dst))
1090                 return NULL;
1091
1092         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1093                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1094         if (IS_ERR(id))
1095                 return NULL;
1096
1097         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1098                           ip_ver, port, src, dst);
1099
1100         rt = &id->route;
1101         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1102         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1103                                GFP_KERNEL);
1104         if (!rt->path_rec)
1105                 goto err;
1106
1107         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1108         if (rt->num_paths == 2)
1109                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1110
1111         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1112                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1113                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1114                 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1115         } else {
1116                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1117                                         &rt->addr.dev_addr);
1118                 if (ret)
1119                         goto err;
1120         }
1121         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1122
1123         id_priv = container_of(id, struct rdma_id_private, id);
1124         id_priv->state = RDMA_CM_CONNECT;
1125         return id_priv;
1126
1127 err:
1128         rdma_destroy_id(id);
1129         return NULL;
1130 }
1131
1132 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1133                                               struct ib_cm_event *ib_event)
1134 {
1135         struct rdma_id_private *id_priv;
1136         struct rdma_cm_id *id;
1137         union cma_ip_addr *src, *dst;
1138         __be16 port;
1139         u8 ip_ver;
1140         int ret;
1141
1142         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1143                             listen_id->ps, IB_QPT_UD);
1144         if (IS_ERR(id))
1145                 return NULL;
1146
1147
1148         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1149                              &ip_ver, &port, &src, &dst))
1150                 goto err;
1151
1152         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1153                           ip_ver, port, src, dst);
1154
1155         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1156                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1157                                         &id->route.addr.dev_addr);
1158                 if (ret)
1159                         goto err;
1160         }
1161
1162         id_priv = container_of(id, struct rdma_id_private, id);
1163         id_priv->state = RDMA_CM_CONNECT;
1164         return id_priv;
1165 err:
1166         rdma_destroy_id(id);
1167         return NULL;
1168 }
1169
1170 static void cma_set_req_event_data(struct rdma_cm_event *event,
1171                                    struct ib_cm_req_event_param *req_data,
1172                                    void *private_data, int offset)
1173 {
1174         event->param.conn.private_data = private_data + offset;
1175         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1176         event->param.conn.responder_resources = req_data->responder_resources;
1177         event->param.conn.initiator_depth = req_data->initiator_depth;
1178         event->param.conn.flow_control = req_data->flow_control;
1179         event->param.conn.retry_count = req_data->retry_count;
1180         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1181         event->param.conn.srq = req_data->srq;
1182         event->param.conn.qp_num = req_data->remote_qpn;
1183 }
1184
1185 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1186 {
1187         return (((ib_event->event == IB_CM_REQ_RECEIVED) ||
1188                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1189                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1190                  (id->qp_type == IB_QPT_UD)) ||
1191                 (!id->qp_type));
1192 }
1193
1194 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1195 {
1196         struct rdma_id_private *listen_id, *conn_id;
1197         struct rdma_cm_event event;
1198         int offset, ret;
1199
1200         listen_id = cm_id->context;
1201         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1202                 return -EINVAL;
1203
1204         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1205                 return -ECONNABORTED;
1206
1207         memset(&event, 0, sizeof event);
1208         offset = cma_user_data_offset(listen_id->id.ps);
1209         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1210         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1211                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1212                 event.param.ud.private_data = ib_event->private_data + offset;
1213                 event.param.ud.private_data_len =
1214                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1215         } else {
1216                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1217                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1218                                        ib_event->private_data, offset);
1219         }
1220         if (!conn_id) {
1221                 ret = -ENOMEM;
1222                 goto err1;
1223         }
1224
1225         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1226         ret = cma_acquire_dev(conn_id);
1227         if (ret)
1228                 goto err2;
1229
1230         conn_id->cm_id.ib = cm_id;
1231         cm_id->context = conn_id;
1232         cm_id->cm_handler = cma_ib_handler;
1233
1234         /*
1235          * Protect against the user destroying conn_id from another thread
1236          * until we're done accessing it.
1237          */
1238         atomic_inc(&conn_id->refcount);
1239         ret = conn_id->id.event_handler(&conn_id->id, &event);
1240         if (ret)
1241                 goto err3;
1242
1243         /*
1244          * Acquire mutex to prevent user executing rdma_destroy_id()
1245          * while we're accessing the cm_id.
1246          */
1247         mutex_lock(&lock);
1248         if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1249                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1250         mutex_unlock(&lock);
1251         mutex_unlock(&conn_id->handler_mutex);
1252         mutex_unlock(&listen_id->handler_mutex);
1253         cma_deref_id(conn_id);
1254         return 0;
1255
1256 err3:
1257         cma_deref_id(conn_id);
1258         /* Destroy the CM ID by returning a non-zero value. */
1259         conn_id->cm_id.ib = NULL;
1260 err2:
1261         cma_exch(conn_id, RDMA_CM_DESTROYING);
1262         mutex_unlock(&conn_id->handler_mutex);
1263 err1:
1264         mutex_unlock(&listen_id->handler_mutex);
1265         if (conn_id)
1266                 rdma_destroy_id(&conn_id->id);
1267         return ret;
1268 }
1269
1270 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1271 {
1272         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1273 }
1274
1275 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1276                                  struct ib_cm_compare_data *compare)
1277 {
1278         struct cma_hdr *cma_data, *cma_mask;
1279         struct sdp_hh *sdp_data, *sdp_mask;
1280         __be32 ip4_addr;
1281         struct in6_addr ip6_addr;
1282
1283         memset(compare, 0, sizeof *compare);
1284         cma_data = (void *) compare->data;
1285         cma_mask = (void *) compare->mask;
1286         sdp_data = (void *) compare->data;
1287         sdp_mask = (void *) compare->mask;
1288
1289         switch (addr->sa_family) {
1290         case AF_INET:
1291                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1292                 if (ps == RDMA_PS_SDP) {
1293                         sdp_set_ip_ver(sdp_data, 4);
1294                         sdp_set_ip_ver(sdp_mask, 0xF);
1295                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1296                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1297                 } else {
1298                         cma_set_ip_ver(cma_data, 4);
1299                         cma_set_ip_ver(cma_mask, 0xF);
1300                         cma_data->dst_addr.ip4.addr = ip4_addr;
1301                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1302                 }
1303                 break;
1304         case AF_INET6:
1305                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1306                 if (ps == RDMA_PS_SDP) {
1307                         sdp_set_ip_ver(sdp_data, 6);
1308                         sdp_set_ip_ver(sdp_mask, 0xF);
1309                         sdp_data->dst_addr.ip6 = ip6_addr;
1310                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1311                                sizeof sdp_mask->dst_addr.ip6);
1312                 } else {
1313                         cma_set_ip_ver(cma_data, 6);
1314                         cma_set_ip_ver(cma_mask, 0xF);
1315                         cma_data->dst_addr.ip6 = ip6_addr;
1316                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1317                                sizeof cma_mask->dst_addr.ip6);
1318                 }
1319                 break;
1320         default:
1321                 break;
1322         }
1323 }
1324
1325 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1326 {
1327         struct rdma_id_private *id_priv = iw_id->context;
1328         struct rdma_cm_event event;
1329         struct sockaddr_in *sin;
1330         int ret = 0;
1331
1332         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1333                 return 0;
1334
1335         memset(&event, 0, sizeof event);
1336         switch (iw_event->event) {
1337         case IW_CM_EVENT_CLOSE:
1338                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1339                 break;
1340         case IW_CM_EVENT_CONNECT_REPLY:
1341                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1342                 *sin = iw_event->local_addr;
1343                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1344                 *sin = iw_event->remote_addr;
1345                 switch (iw_event->status) {
1346                 case 0:
1347                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1348                         event.param.conn.initiator_depth = iw_event->ird;
1349                         event.param.conn.responder_resources = iw_event->ord;
1350                         break;
1351                 case -ECONNRESET:
1352                 case -ECONNREFUSED:
1353                         event.event = RDMA_CM_EVENT_REJECTED;
1354                         break;
1355                 case -ETIMEDOUT:
1356                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1357                         break;
1358                 default:
1359                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1360                         break;
1361                 }
1362                 break;
1363         case IW_CM_EVENT_ESTABLISHED:
1364                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1365                 event.param.conn.initiator_depth = iw_event->ird;
1366                 event.param.conn.responder_resources = iw_event->ord;
1367                 break;
1368         default:
1369                 BUG_ON(1);
1370         }
1371
1372         event.status = iw_event->status;
1373         event.param.conn.private_data = iw_event->private_data;
1374         event.param.conn.private_data_len = iw_event->private_data_len;
1375         ret = id_priv->id.event_handler(&id_priv->id, &event);
1376         if (ret) {
1377                 /* Destroy the CM ID by returning a non-zero value. */
1378                 id_priv->cm_id.iw = NULL;
1379                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1380                 mutex_unlock(&id_priv->handler_mutex);
1381                 rdma_destroy_id(&id_priv->id);
1382                 return ret;
1383         }
1384
1385         mutex_unlock(&id_priv->handler_mutex);
1386         return ret;
1387 }
1388
1389 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1390                                struct iw_cm_event *iw_event)
1391 {
1392         struct rdma_cm_id *new_cm_id;
1393         struct rdma_id_private *listen_id, *conn_id;
1394         struct sockaddr_in *sin;
1395         struct net_device *dev = NULL;
1396         struct rdma_cm_event event;
1397         int ret;
1398         struct ib_device_attr attr;
1399
1400         listen_id = cm_id->context;
1401         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1402                 return -ECONNABORTED;
1403
1404         /* Create a new RDMA id for the new IW CM ID */
1405         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1406                                    listen_id->id.context,
1407                                    RDMA_PS_TCP, IB_QPT_RC);
1408         if (IS_ERR(new_cm_id)) {
1409                 ret = -ENOMEM;
1410                 goto out;
1411         }
1412         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1413         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1414         conn_id->state = RDMA_CM_CONNECT;
1415
1416         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1417         if (!dev) {
1418                 ret = -EADDRNOTAVAIL;
1419                 mutex_unlock(&conn_id->handler_mutex);
1420                 rdma_destroy_id(new_cm_id);
1421                 goto out;
1422         }
1423         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1424         if (ret) {
1425                 mutex_unlock(&conn_id->handler_mutex);
1426                 rdma_destroy_id(new_cm_id);
1427                 goto out;
1428         }
1429
1430         ret = cma_acquire_dev(conn_id);
1431         if (ret) {
1432                 mutex_unlock(&conn_id->handler_mutex);
1433                 rdma_destroy_id(new_cm_id);
1434                 goto out;
1435         }
1436
1437         conn_id->cm_id.iw = cm_id;
1438         cm_id->context = conn_id;
1439         cm_id->cm_handler = cma_iw_handler;
1440
1441         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1442         *sin = iw_event->local_addr;
1443         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1444         *sin = iw_event->remote_addr;
1445
1446         ret = ib_query_device(conn_id->id.device, &attr);
1447         if (ret) {
1448                 mutex_unlock(&conn_id->handler_mutex);
1449                 rdma_destroy_id(new_cm_id);
1450                 goto out;
1451         }
1452
1453         memset(&event, 0, sizeof event);
1454         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1455         event.param.conn.private_data = iw_event->private_data;
1456         event.param.conn.private_data_len = iw_event->private_data_len;
1457         event.param.conn.initiator_depth = iw_event->ird;
1458         event.param.conn.responder_resources = iw_event->ord;
1459
1460         /*
1461          * Protect against the user destroying conn_id from another thread
1462          * until we're done accessing it.
1463          */
1464         atomic_inc(&conn_id->refcount);
1465         ret = conn_id->id.event_handler(&conn_id->id, &event);
1466         if (ret) {
1467                 /* User wants to destroy the CM ID */
1468                 conn_id->cm_id.iw = NULL;
1469                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1470                 mutex_unlock(&conn_id->handler_mutex);
1471                 cma_deref_id(conn_id);
1472                 rdma_destroy_id(&conn_id->id);
1473                 goto out;
1474         }
1475
1476         mutex_unlock(&conn_id->handler_mutex);
1477         cma_deref_id(conn_id);
1478
1479 out:
1480         if (dev)
1481                 dev_put(dev);
1482         mutex_unlock(&listen_id->handler_mutex);
1483         return ret;
1484 }
1485
1486 static int cma_ib_listen(struct rdma_id_private *id_priv)
1487 {
1488         struct ib_cm_compare_data compare_data;
1489         struct sockaddr *addr;
1490         struct ib_cm_id *id;
1491         __be64 svc_id;
1492         int ret;
1493
1494         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1495         if (IS_ERR(id))
1496                 return PTR_ERR(id);
1497
1498         id_priv->cm_id.ib = id;
1499
1500         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1501         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1502         if (cma_any_addr(addr))
1503                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1504         else {
1505                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1506                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1507         }
1508
1509         if (ret) {
1510                 ib_destroy_cm_id(id_priv->cm_id.ib);
1511                 id_priv->cm_id.ib = NULL;
1512         }
1513
1514         return ret;
1515 }
1516
1517 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1518 {
1519         int ret;
1520         struct sockaddr_in *sin;
1521         struct iw_cm_id *id;
1522
1523         id = iw_create_cm_id(id_priv->id.device,
1524                              iw_conn_req_handler,
1525                              id_priv);
1526         if (IS_ERR(id))
1527                 return PTR_ERR(id);
1528
1529         id_priv->cm_id.iw = id;
1530
1531         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1532         id_priv->cm_id.iw->local_addr = *sin;
1533
1534         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1535
1536         if (ret) {
1537                 iw_destroy_cm_id(id_priv->cm_id.iw);
1538                 id_priv->cm_id.iw = NULL;
1539         }
1540
1541         return ret;
1542 }
1543
1544 static int cma_listen_handler(struct rdma_cm_id *id,
1545                               struct rdma_cm_event *event)
1546 {
1547         struct rdma_id_private *id_priv = id->context;
1548
1549         id->context = id_priv->id.context;
1550         id->event_handler = id_priv->id.event_handler;
1551         return id_priv->id.event_handler(id, event);
1552 }
1553
1554 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1555                               struct cma_device *cma_dev)
1556 {
1557         struct rdma_id_private *dev_id_priv;
1558         struct rdma_cm_id *id;
1559         int ret;
1560
1561         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1562                             id_priv->id.qp_type);
1563         if (IS_ERR(id))
1564                 return;
1565
1566         dev_id_priv = container_of(id, struct rdma_id_private, id);
1567
1568         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1569         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1570                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1571
1572         cma_attach_to_dev(dev_id_priv, cma_dev);
1573         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1574         atomic_inc(&id_priv->refcount);
1575         dev_id_priv->internal_id = 1;
1576
1577         ret = rdma_listen(id, id_priv->backlog);
1578         if (ret)
1579                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1580                        "listening on device %s\n", ret, cma_dev->device->name);
1581 }
1582
1583 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1584 {
1585         struct cma_device *cma_dev;
1586
1587         mutex_lock(&lock);
1588         list_add_tail(&id_priv->list, &listen_any_list);
1589         list_for_each_entry(cma_dev, &dev_list, list)
1590                 cma_listen_on_dev(id_priv, cma_dev);
1591         mutex_unlock(&lock);
1592 }
1593
1594 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1595 {
1596         struct rdma_id_private *id_priv;
1597
1598         id_priv = container_of(id, struct rdma_id_private, id);
1599         id_priv->tos = (u8) tos;
1600 }
1601 EXPORT_SYMBOL(rdma_set_service_type);
1602
1603 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1604                               void *context)
1605 {
1606         struct cma_work *work = context;
1607         struct rdma_route *route;
1608
1609         route = &work->id->id.route;
1610
1611         if (!status) {
1612                 route->num_paths = 1;
1613                 *route->path_rec = *path_rec;
1614         } else {
1615                 work->old_state = RDMA_CM_ROUTE_QUERY;
1616                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1617                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1618                 work->event.status = status;
1619         }
1620
1621         queue_work(cma_wq, &work->work);
1622 }
1623
1624 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1625                               struct cma_work *work)
1626 {
1627         struct rdma_addr *addr = &id_priv->id.route.addr;
1628         struct ib_sa_path_rec path_rec;
1629         ib_sa_comp_mask comp_mask;
1630         struct sockaddr_in6 *sin6;
1631
1632         memset(&path_rec, 0, sizeof path_rec);
1633         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1634         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1635         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1636         path_rec.numb_path = 1;
1637         path_rec.reversible = 1;
1638         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1639                                                         (struct sockaddr *) &addr->dst_addr);
1640
1641         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1642                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1643                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1644
1645         if (addr->src_addr.ss_family == AF_INET) {
1646                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1647                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1648         } else {
1649                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1650                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1651                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1652         }
1653
1654         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1655                                                id_priv->id.port_num, &path_rec,
1656                                                comp_mask, timeout_ms,
1657                                                GFP_KERNEL, cma_query_handler,
1658                                                work, &id_priv->query);
1659
1660         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1661 }
1662
1663 static void cma_work_handler(struct work_struct *_work)
1664 {
1665         struct cma_work *work = container_of(_work, struct cma_work, work);
1666         struct rdma_id_private *id_priv = work->id;
1667         int destroy = 0;
1668
1669         mutex_lock(&id_priv->handler_mutex);
1670         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1671                 goto out;
1672
1673         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1674                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1675                 destroy = 1;
1676         }
1677 out:
1678         mutex_unlock(&id_priv->handler_mutex);
1679         cma_deref_id(id_priv);
1680         if (destroy)
1681                 rdma_destroy_id(&id_priv->id);
1682         kfree(work);
1683 }
1684
1685 static void cma_ndev_work_handler(struct work_struct *_work)
1686 {
1687         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1688         struct rdma_id_private *id_priv = work->id;
1689         int destroy = 0;
1690
1691         mutex_lock(&id_priv->handler_mutex);
1692         if (id_priv->state == RDMA_CM_DESTROYING ||
1693             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1694                 goto out;
1695
1696         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1697                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1698                 destroy = 1;
1699         }
1700
1701 out:
1702         mutex_unlock(&id_priv->handler_mutex);
1703         cma_deref_id(id_priv);
1704         if (destroy)
1705                 rdma_destroy_id(&id_priv->id);
1706         kfree(work);
1707 }
1708
1709 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1710 {
1711         struct rdma_route *route = &id_priv->id.route;
1712         struct cma_work *work;
1713         int ret;
1714
1715         work = kzalloc(sizeof *work, GFP_KERNEL);
1716         if (!work)
1717                 return -ENOMEM;
1718
1719         work->id = id_priv;
1720         INIT_WORK(&work->work, cma_work_handler);
1721         work->old_state = RDMA_CM_ROUTE_QUERY;
1722         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1723         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1724
1725         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1726         if (!route->path_rec) {
1727                 ret = -ENOMEM;
1728                 goto err1;
1729         }
1730
1731         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1732         if (ret)
1733                 goto err2;
1734
1735         return 0;
1736 err2:
1737         kfree(route->path_rec);
1738         route->path_rec = NULL;
1739 err1:
1740         kfree(work);
1741         return ret;
1742 }
1743
1744 int rdma_set_ib_paths(struct rdma_cm_id *id,
1745                       struct ib_sa_path_rec *path_rec, int num_paths)
1746 {
1747         struct rdma_id_private *id_priv;
1748         int ret;
1749
1750         id_priv = container_of(id, struct rdma_id_private, id);
1751         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1752                            RDMA_CM_ROUTE_RESOLVED))
1753                 return -EINVAL;
1754
1755         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1756                                      GFP_KERNEL);
1757         if (!id->route.path_rec) {
1758                 ret = -ENOMEM;
1759                 goto err;
1760         }
1761
1762         id->route.num_paths = num_paths;
1763         return 0;
1764 err:
1765         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1766         return ret;
1767 }
1768 EXPORT_SYMBOL(rdma_set_ib_paths);
1769
1770 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1771 {
1772         struct cma_work *work;
1773
1774         work = kzalloc(sizeof *work, GFP_KERNEL);
1775         if (!work)
1776                 return -ENOMEM;
1777
1778         work->id = id_priv;
1779         INIT_WORK(&work->work, cma_work_handler);
1780         work->old_state = RDMA_CM_ROUTE_QUERY;
1781         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1782         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1783         queue_work(cma_wq, &work->work);
1784         return 0;
1785 }
1786
1787 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1788 {
1789         struct rdma_route *route = &id_priv->id.route;
1790         struct rdma_addr *addr = &route->addr;
1791         struct cma_work *work;
1792         int ret;
1793         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1794         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1795         struct net_device *ndev = NULL;
1796         u16 vid;
1797
1798         if (src_addr->sin_family != dst_addr->sin_family)
1799                 return -EINVAL;
1800
1801         work = kzalloc(sizeof *work, GFP_KERNEL);
1802         if (!work)
1803                 return -ENOMEM;
1804
1805         work->id = id_priv;
1806         INIT_WORK(&work->work, cma_work_handler);
1807
1808         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1809         if (!route->path_rec) {
1810                 ret = -ENOMEM;
1811                 goto err1;
1812         }
1813
1814         route->num_paths = 1;
1815
1816         if (addr->dev_addr.bound_dev_if)
1817                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1818         if (!ndev) {
1819                 ret = -ENODEV;
1820                 goto err2;
1821         }
1822
1823         vid = rdma_vlan_dev_vlan_id(ndev);
1824
1825         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1826         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1827
1828         route->path_rec->hop_limit = 1;
1829         route->path_rec->reversible = 1;
1830         route->path_rec->pkey = cpu_to_be16(0xffff);
1831         route->path_rec->mtu_selector = IB_SA_EQ;
1832         route->path_rec->sl = netdev_get_prio_tc_map(
1833                         ndev->priv_flags & IFF_802_1Q_VLAN ?
1834                                 vlan_dev_real_dev(ndev) : ndev,
1835                         rt_tos2priority(id_priv->tos));
1836
1837         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1838         route->path_rec->rate_selector = IB_SA_EQ;
1839         route->path_rec->rate = iboe_get_rate(ndev);
1840         dev_put(ndev);
1841         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1842         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1843         if (!route->path_rec->mtu) {
1844                 ret = -EINVAL;
1845                 goto err2;
1846         }
1847
1848         work->old_state = RDMA_CM_ROUTE_QUERY;
1849         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1850         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1851         work->event.status = 0;
1852
1853         queue_work(cma_wq, &work->work);
1854
1855         return 0;
1856
1857 err2:
1858         kfree(route->path_rec);
1859         route->path_rec = NULL;
1860 err1:
1861         kfree(work);
1862         return ret;
1863 }
1864
1865 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1866 {
1867         struct rdma_id_private *id_priv;
1868         int ret;
1869
1870         id_priv = container_of(id, struct rdma_id_private, id);
1871         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1872                 return -EINVAL;
1873
1874         atomic_inc(&id_priv->refcount);
1875         switch (rdma_node_get_transport(id->device->node_type)) {
1876         case RDMA_TRANSPORT_IB:
1877                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1878                 case IB_LINK_LAYER_INFINIBAND:
1879                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1880                         break;
1881                 case IB_LINK_LAYER_ETHERNET:
1882                         ret = cma_resolve_iboe_route(id_priv);
1883                         break;
1884                 default:
1885                         ret = -ENOSYS;
1886                 }
1887                 break;
1888         case RDMA_TRANSPORT_IWARP:
1889                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1890                 break;
1891         default:
1892                 ret = -ENOSYS;
1893                 break;
1894         }
1895         if (ret)
1896                 goto err;
1897
1898         return 0;
1899 err:
1900         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1901         cma_deref_id(id_priv);
1902         return ret;
1903 }
1904 EXPORT_SYMBOL(rdma_resolve_route);
1905
1906 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1907 {
1908         struct cma_device *cma_dev;
1909         struct ib_port_attr port_attr;
1910         union ib_gid gid;
1911         u16 pkey;
1912         int ret;
1913         u8 p;
1914
1915         mutex_lock(&lock);
1916         if (list_empty(&dev_list)) {
1917                 ret = -ENODEV;
1918                 goto out;
1919         }
1920         list_for_each_entry(cma_dev, &dev_list, list)
1921                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1922                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1923                             port_attr.state == IB_PORT_ACTIVE)
1924                                 goto port_found;
1925
1926         p = 1;
1927         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1928
1929 port_found:
1930         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1931         if (ret)
1932                 goto out;
1933
1934         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1935         if (ret)
1936                 goto out;
1937
1938         id_priv->id.route.addr.dev_addr.dev_type =
1939                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1940                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1941
1942         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1943         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1944         id_priv->id.port_num = p;
1945         cma_attach_to_dev(id_priv, cma_dev);
1946 out:
1947         mutex_unlock(&lock);
1948         return ret;
1949 }
1950
1951 static void addr_handler(int status, struct sockaddr *src_addr,
1952                          struct rdma_dev_addr *dev_addr, void *context)
1953 {
1954         struct rdma_id_private *id_priv = context;
1955         struct rdma_cm_event event;
1956
1957         memset(&event, 0, sizeof event);
1958         mutex_lock(&id_priv->handler_mutex);
1959         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1960                            RDMA_CM_ADDR_RESOLVED))
1961                 goto out;
1962
1963         if (!status && !id_priv->cma_dev)
1964                 status = cma_acquire_dev(id_priv);
1965
1966         if (status) {
1967                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1968                                    RDMA_CM_ADDR_BOUND))
1969                         goto out;
1970                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1971                 event.status = status;
1972         } else {
1973                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1974                        ip_addr_size(src_addr));
1975                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1976         }
1977
1978         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1979                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1980                 mutex_unlock(&id_priv->handler_mutex);
1981                 cma_deref_id(id_priv);
1982                 rdma_destroy_id(&id_priv->id);
1983                 return;
1984         }
1985 out:
1986         mutex_unlock(&id_priv->handler_mutex);
1987         cma_deref_id(id_priv);
1988 }
1989
1990 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1991 {
1992         struct cma_work *work;
1993         struct sockaddr *src, *dst;
1994         union ib_gid gid;
1995         int ret;
1996
1997         work = kzalloc(sizeof *work, GFP_KERNEL);
1998         if (!work)
1999                 return -ENOMEM;
2000
2001         if (!id_priv->cma_dev) {
2002                 ret = cma_bind_loopback(id_priv);
2003                 if (ret)
2004                         goto err;
2005         }
2006
2007         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2008         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2009
2010         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2011         if (cma_zero_addr(src)) {
2012                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2013                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2014                         ((struct sockaddr_in *)src)->sin_addr =
2015                                 ((struct sockaddr_in *)dst)->sin_addr;
2016                 } else {
2017                         ((struct sockaddr_in6 *)src)->sin6_addr =
2018                                 ((struct sockaddr_in6 *)dst)->sin6_addr;
2019                 }
2020         }
2021
2022         work->id = id_priv;
2023         INIT_WORK(&work->work, cma_work_handler);
2024         work->old_state = RDMA_CM_ADDR_QUERY;
2025         work->new_state = RDMA_CM_ADDR_RESOLVED;
2026         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2027         queue_work(cma_wq, &work->work);
2028         return 0;
2029 err:
2030         kfree(work);
2031         return ret;
2032 }
2033
2034 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2035                          struct sockaddr *dst_addr)
2036 {
2037         if (!src_addr || !src_addr->sa_family) {
2038                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2039                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2040                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2041                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2042                 }
2043         }
2044         return rdma_bind_addr(id, src_addr);
2045 }
2046
2047 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2048                       struct sockaddr *dst_addr, int timeout_ms)
2049 {
2050         struct rdma_id_private *id_priv;
2051         int ret;
2052
2053         id_priv = container_of(id, struct rdma_id_private, id);
2054         if (id_priv->state == RDMA_CM_IDLE) {
2055                 ret = cma_bind_addr(id, src_addr, dst_addr);
2056                 if (ret)
2057                         return ret;
2058         }
2059
2060         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2061                 return -EINVAL;
2062
2063         atomic_inc(&id_priv->refcount);
2064         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2065         if (cma_any_addr(dst_addr))
2066                 ret = cma_resolve_loopback(id_priv);
2067         else
2068                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2069                                       dst_addr, &id->route.addr.dev_addr,
2070                                       timeout_ms, addr_handler, id_priv);
2071         if (ret)
2072                 goto err;
2073
2074         return 0;
2075 err:
2076         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2077         cma_deref_id(id_priv);
2078         return ret;
2079 }
2080 EXPORT_SYMBOL(rdma_resolve_addr);
2081
2082 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2083 {
2084         struct rdma_id_private *id_priv;
2085         unsigned long flags;
2086         int ret;
2087
2088         id_priv = container_of(id, struct rdma_id_private, id);
2089         spin_lock_irqsave(&id_priv->lock, flags);
2090         if (id_priv->state == RDMA_CM_IDLE) {
2091                 id_priv->reuseaddr = reuse;
2092                 ret = 0;
2093         } else {
2094                 ret = -EINVAL;
2095         }
2096         spin_unlock_irqrestore(&id_priv->lock, flags);
2097         return ret;
2098 }
2099 EXPORT_SYMBOL(rdma_set_reuseaddr);
2100
2101 static void cma_bind_port(struct rdma_bind_list *bind_list,
2102                           struct rdma_id_private *id_priv)
2103 {
2104         struct sockaddr_in *sin;
2105
2106         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2107         sin->sin_port = htons(bind_list->port);
2108         id_priv->bind_list = bind_list;
2109         hlist_add_head(&id_priv->node, &bind_list->owners);
2110 }
2111
2112 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2113                           unsigned short snum)
2114 {
2115         struct rdma_bind_list *bind_list;
2116         int port, ret;
2117
2118         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2119         if (!bind_list)
2120                 return -ENOMEM;
2121
2122         do {
2123                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2124         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2125
2126         if (ret)
2127                 goto err1;
2128
2129         if (port != snum) {
2130                 ret = -EADDRNOTAVAIL;
2131                 goto err2;
2132         }
2133
2134         bind_list->ps = ps;
2135         bind_list->port = (unsigned short) port;
2136         cma_bind_port(bind_list, id_priv);
2137         return 0;
2138 err2:
2139         idr_remove(ps, port);
2140 err1:
2141         kfree(bind_list);
2142         return ret;
2143 }
2144
2145 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2146 {
2147         static unsigned int last_used_port;
2148         int low, high, remaining;
2149         unsigned int rover;
2150
2151         inet_get_local_port_range(&low, &high);
2152         remaining = (high - low) + 1;
2153         rover = net_random() % remaining + low;
2154 retry:
2155         if (last_used_port != rover &&
2156             !idr_find(ps, (unsigned short) rover)) {
2157                 int ret = cma_alloc_port(ps, id_priv, rover);
2158                 /*
2159                  * Remember previously used port number in order to avoid
2160                  * re-using same port immediately after it is closed.
2161                  */
2162                 if (!ret)
2163                         last_used_port = rover;
2164                 if (ret != -EADDRNOTAVAIL)
2165                         return ret;
2166         }
2167         if (--remaining) {
2168                 rover++;
2169                 if ((rover < low) || (rover > high))
2170                         rover = low;
2171                 goto retry;
2172         }
2173         return -EADDRNOTAVAIL;
2174 }
2175
2176 /*
2177  * Check that the requested port is available.  This is called when trying to
2178  * bind to a specific port, or when trying to listen on a bound port.  In
2179  * the latter case, the provided id_priv may already be on the bind_list, but
2180  * we still need to check that it's okay to start listening.
2181  */
2182 static int cma_check_port(struct rdma_bind_list *bind_list,
2183                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2184 {
2185         struct rdma_id_private *cur_id;
2186         struct sockaddr *addr, *cur_addr;
2187         struct hlist_node *node;
2188
2189         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2190         if (cma_any_addr(addr) && !reuseaddr)
2191                 return -EADDRNOTAVAIL;
2192
2193         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2194                 if (id_priv == cur_id)
2195                         continue;
2196
2197                 if ((cur_id->state == RDMA_CM_LISTEN) ||
2198                     !reuseaddr || !cur_id->reuseaddr) {
2199                         cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2200                         if (cma_any_addr(cur_addr))
2201                                 return -EADDRNOTAVAIL;
2202
2203                         if (!cma_addr_cmp(addr, cur_addr))
2204                                 return -EADDRINUSE;
2205                 }
2206         }
2207         return 0;
2208 }
2209
2210 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2211 {
2212         struct rdma_bind_list *bind_list;
2213         unsigned short snum;
2214         int ret;
2215
2216         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2217         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2218                 return -EACCES;
2219
2220         bind_list = idr_find(ps, snum);
2221         if (!bind_list) {
2222                 ret = cma_alloc_port(ps, id_priv, snum);
2223         } else {
2224                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2225                 if (!ret)
2226                         cma_bind_port(bind_list, id_priv);
2227         }
2228         return ret;
2229 }
2230
2231 static int cma_bind_listen(struct rdma_id_private *id_priv)
2232 {
2233         struct rdma_bind_list *bind_list = id_priv->bind_list;
2234         int ret = 0;
2235
2236         mutex_lock(&lock);
2237         if (bind_list->owners.first->next)
2238                 ret = cma_check_port(bind_list, id_priv, 0);
2239         mutex_unlock(&lock);
2240         return ret;
2241 }
2242
2243 static int cma_get_port(struct rdma_id_private *id_priv)
2244 {
2245         struct idr *ps;
2246         int ret;
2247
2248         switch (id_priv->id.ps) {
2249         case RDMA_PS_SDP:
2250                 ps = &sdp_ps;
2251                 break;
2252         case RDMA_PS_TCP:
2253                 ps = &tcp_ps;
2254                 break;
2255         case RDMA_PS_UDP:
2256                 ps = &udp_ps;
2257                 break;
2258         case RDMA_PS_IPOIB:
2259                 ps = &ipoib_ps;
2260                 break;
2261         case RDMA_PS_IB:
2262                 ps = &ib_ps;
2263                 break;
2264         default:
2265                 return -EPROTONOSUPPORT;
2266         }
2267
2268         mutex_lock(&lock);
2269         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2270                 ret = cma_alloc_any_port(ps, id_priv);
2271         else
2272                 ret = cma_use_port(ps, id_priv);
2273         mutex_unlock(&lock);
2274
2275         return ret;
2276 }
2277
2278 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2279                                struct sockaddr *addr)
2280 {
2281 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2282         struct sockaddr_in6 *sin6;
2283
2284         if (addr->sa_family != AF_INET6)
2285                 return 0;
2286
2287         sin6 = (struct sockaddr_in6 *) addr;
2288         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2289             !sin6->sin6_scope_id)
2290                         return -EINVAL;
2291
2292         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2293 #endif
2294         return 0;
2295 }
2296
2297 int rdma_listen(struct rdma_cm_id *id, int backlog)
2298 {
2299         struct rdma_id_private *id_priv;
2300         int ret;
2301
2302         id_priv = container_of(id, struct rdma_id_private, id);
2303         if (id_priv->state == RDMA_CM_IDLE) {
2304                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2305                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2306                 if (ret)
2307                         return ret;
2308         }
2309
2310         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2311                 return -EINVAL;
2312
2313         if (id_priv->reuseaddr) {
2314                 ret = cma_bind_listen(id_priv);
2315                 if (ret)
2316                         goto err;
2317         }
2318
2319         id_priv->backlog = backlog;
2320         if (id->device) {
2321                 switch (rdma_node_get_transport(id->device->node_type)) {
2322                 case RDMA_TRANSPORT_IB:
2323                         ret = cma_ib_listen(id_priv);
2324                         if (ret)
2325                                 goto err;
2326                         break;
2327                 case RDMA_TRANSPORT_IWARP:
2328                         ret = cma_iw_listen(id_priv, backlog);
2329                         if (ret)
2330                                 goto err;
2331                         break;
2332                 default:
2333                         ret = -ENOSYS;
2334                         goto err;
2335                 }
2336         } else
2337                 cma_listen_on_all(id_priv);
2338
2339         return 0;
2340 err:
2341         id_priv->backlog = 0;
2342         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2343         return ret;
2344 }
2345 EXPORT_SYMBOL(rdma_listen);
2346
2347 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2348 {
2349         struct rdma_id_private *id_priv;
2350         int ret;
2351
2352         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2353                 return -EAFNOSUPPORT;
2354
2355         id_priv = container_of(id, struct rdma_id_private, id);
2356         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2357                 return -EINVAL;
2358
2359         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2360         if (ret)
2361                 goto err1;
2362
2363         if (!cma_any_addr(addr)) {
2364                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2365                 if (ret)
2366                         goto err1;
2367
2368                 ret = cma_acquire_dev(id_priv);
2369                 if (ret)
2370                         goto err1;
2371         }
2372
2373         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2374         ret = cma_get_port(id_priv);
2375         if (ret)
2376                 goto err2;
2377
2378         return 0;
2379 err2:
2380         if (id_priv->cma_dev)
2381                 cma_release_dev(id_priv);
2382 err1:
2383         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2384         return ret;
2385 }
2386 EXPORT_SYMBOL(rdma_bind_addr);
2387
2388 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2389                           struct rdma_route *route)
2390 {
2391         struct cma_hdr *cma_hdr;
2392         struct sdp_hh *sdp_hdr;
2393
2394         if (route->addr.src_addr.ss_family == AF_INET) {
2395                 struct sockaddr_in *src4, *dst4;
2396
2397                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2398                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2399
2400                 switch (ps) {
2401                 case RDMA_PS_SDP:
2402                         sdp_hdr = hdr;
2403                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2404                                 return -EINVAL;
2405                         sdp_set_ip_ver(sdp_hdr, 4);
2406                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2407                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2408                         sdp_hdr->port = src4->sin_port;
2409                         break;
2410                 default:
2411                         cma_hdr = hdr;
2412                         cma_hdr->cma_version = CMA_VERSION;
2413                         cma_set_ip_ver(cma_hdr, 4);
2414                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2415                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2416                         cma_hdr->port = src4->sin_port;
2417                         break;
2418                 }
2419         } else {
2420                 struct sockaddr_in6 *src6, *dst6;
2421
2422                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2423                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2424
2425                 switch (ps) {
2426                 case RDMA_PS_SDP:
2427                         sdp_hdr = hdr;
2428                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2429                                 return -EINVAL;
2430                         sdp_set_ip_ver(sdp_hdr, 6);
2431                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2432                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2433                         sdp_hdr->port = src6->sin6_port;
2434                         break;
2435                 default:
2436                         cma_hdr = hdr;
2437                         cma_hdr->cma_version = CMA_VERSION;
2438                         cma_set_ip_ver(cma_hdr, 6);
2439                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2440                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2441                         cma_hdr->port = src6->sin6_port;
2442                         break;
2443                 }
2444         }
2445         return 0;
2446 }
2447
2448 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2449                                 struct ib_cm_event *ib_event)
2450 {
2451         struct rdma_id_private *id_priv = cm_id->context;
2452         struct rdma_cm_event event;
2453         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2454         int ret = 0;
2455
2456         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2457                 return 0;
2458
2459         memset(&event, 0, sizeof event);
2460         switch (ib_event->event) {
2461         case IB_CM_SIDR_REQ_ERROR:
2462                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2463                 event.status = -ETIMEDOUT;
2464                 break;
2465         case IB_CM_SIDR_REP_RECEIVED:
2466                 event.param.ud.private_data = ib_event->private_data;
2467                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2468                 if (rep->status != IB_SIDR_SUCCESS) {
2469                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2470                         event.status = ib_event->param.sidr_rep_rcvd.status;
2471                         break;
2472                 }
2473                 ret = cma_set_qkey(id_priv);
2474                 if (ret) {
2475                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2476                         event.status = -EINVAL;
2477                         break;
2478                 }
2479                 if (id_priv->qkey != rep->qkey) {
2480                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2481                         event.status = -EINVAL;
2482                         break;
2483                 }
2484                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2485                                      id_priv->id.route.path_rec,
2486                                      &event.param.ud.ah_attr);
2487                 event.param.ud.qp_num = rep->qpn;
2488                 event.param.ud.qkey = rep->qkey;
2489                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2490                 event.status = 0;
2491                 break;
2492         default:
2493                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2494                        ib_event->event);
2495                 goto out;
2496         }
2497
2498         ret = id_priv->id.event_handler(&id_priv->id, &event);
2499         if (ret) {
2500                 /* Destroy the CM ID by returning a non-zero value. */
2501                 id_priv->cm_id.ib = NULL;
2502                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2503                 mutex_unlock(&id_priv->handler_mutex);
2504                 rdma_destroy_id(&id_priv->id);
2505                 return ret;
2506         }
2507 out:
2508         mutex_unlock(&id_priv->handler_mutex);
2509         return ret;
2510 }
2511
2512 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2513                               struct rdma_conn_param *conn_param)
2514 {
2515         struct ib_cm_sidr_req_param req;
2516         struct rdma_route *route;
2517         struct ib_cm_id *id;
2518         int ret;
2519
2520         req.private_data_len = sizeof(struct cma_hdr) +
2521                                conn_param->private_data_len;
2522         if (req.private_data_len < conn_param->private_data_len)
2523                 return -EINVAL;
2524
2525         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2526         if (!req.private_data)
2527                 return -ENOMEM;
2528
2529         if (conn_param->private_data && conn_param->private_data_len)
2530                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2531                        conn_param->private_data, conn_param->private_data_len);
2532
2533         route = &id_priv->id.route;
2534         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2535         if (ret)
2536                 goto out;
2537
2538         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2539                              id_priv);
2540         if (IS_ERR(id)) {
2541                 ret = PTR_ERR(id);
2542                 goto out;
2543         }
2544         id_priv->cm_id.ib = id;
2545
2546         req.path = route->path_rec;
2547         req.service_id = cma_get_service_id(id_priv->id.ps,
2548                                             (struct sockaddr *) &route->addr.dst_addr);
2549         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2550         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2551
2552         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2553         if (ret) {
2554                 ib_destroy_cm_id(id_priv->cm_id.ib);
2555                 id_priv->cm_id.ib = NULL;
2556         }
2557 out:
2558         kfree(req.private_data);
2559         return ret;
2560 }
2561
2562 static int cma_connect_ib(struct rdma_id_private *id_priv,
2563                           struct rdma_conn_param *conn_param)
2564 {
2565         struct ib_cm_req_param req;
2566         struct rdma_route *route;
2567         void *private_data;
2568         struct ib_cm_id *id;
2569         int offset, ret;
2570
2571         memset(&req, 0, sizeof req);
2572         offset = cma_user_data_offset(id_priv->id.ps);
2573         req.private_data_len = offset + conn_param->private_data_len;
2574         if (req.private_data_len < conn_param->private_data_len)
2575                 return -EINVAL;
2576
2577         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2578         if (!private_data)
2579                 return -ENOMEM;
2580
2581         if (conn_param->private_data && conn_param->private_data_len)
2582                 memcpy(private_data + offset, conn_param->private_data,
2583                        conn_param->private_data_len);
2584
2585         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2586         if (IS_ERR(id)) {
2587                 ret = PTR_ERR(id);
2588                 goto out;
2589         }
2590         id_priv->cm_id.ib = id;
2591
2592         route = &id_priv->id.route;
2593         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2594         if (ret)
2595                 goto out;
2596         req.private_data = private_data;
2597
2598         req.primary_path = &route->path_rec[0];
2599         if (route->num_paths == 2)
2600                 req.alternate_path = &route->path_rec[1];
2601
2602         req.service_id = cma_get_service_id(id_priv->id.ps,
2603                                             (struct sockaddr *) &route->addr.dst_addr);
2604         req.qp_num = id_priv->qp_num;
2605         req.qp_type = id_priv->id.qp_type;
2606         req.starting_psn = id_priv->seq_num;
2607         req.responder_resources = conn_param->responder_resources;
2608         req.initiator_depth = conn_param->initiator_depth;
2609         req.flow_control = conn_param->flow_control;
2610         req.retry_count = conn_param->retry_count;
2611         req.rnr_retry_count = conn_param->rnr_retry_count;
2612         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2613         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2614         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2615         req.srq = id_priv->srq ? 1 : 0;
2616
2617         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2618 out:
2619         if (ret && !IS_ERR(id)) {
2620                 ib_destroy_cm_id(id);
2621                 id_priv->cm_id.ib = NULL;
2622         }
2623
2624         kfree(private_data);
2625         return ret;
2626 }
2627
2628 static int cma_connect_iw(struct rdma_id_private *id_priv,
2629                           struct rdma_conn_param *conn_param)
2630 {
2631         struct iw_cm_id *cm_id;
2632         struct sockaddr_in* sin;
2633         int ret;
2634         struct iw_cm_conn_param iw_param;
2635
2636         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2637         if (IS_ERR(cm_id))
2638                 return PTR_ERR(cm_id);
2639
2640         id_priv->cm_id.iw = cm_id;
2641
2642         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2643         cm_id->local_addr = *sin;
2644
2645         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2646         cm_id->remote_addr = *sin;
2647
2648         ret = cma_modify_qp_rtr(id_priv, conn_param);
2649         if (ret)
2650                 goto out;
2651
2652         if (conn_param) {
2653                 iw_param.ord = conn_param->initiator_depth;
2654                 iw_param.ird = conn_param->responder_resources;
2655                 iw_param.private_data = conn_param->private_data;
2656                 iw_param.private_data_len = conn_param->private_data_len;
2657                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2658         } else {
2659                 memset(&iw_param, 0, sizeof iw_param);
2660                 iw_param.qpn = id_priv->qp_num;
2661         }
2662         ret = iw_cm_connect(cm_id, &iw_param);
2663 out:
2664         if (ret) {
2665                 iw_destroy_cm_id(cm_id);
2666                 id_priv->cm_id.iw = NULL;
2667         }
2668         return ret;
2669 }
2670
2671 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2672 {
2673         struct rdma_id_private *id_priv;
2674         int ret;
2675
2676         id_priv = container_of(id, struct rdma_id_private, id);
2677         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2678                 return -EINVAL;
2679
2680         if (!id->qp) {
2681                 id_priv->qp_num = conn_param->qp_num;
2682                 id_priv->srq = conn_param->srq;
2683         }
2684
2685         switch (rdma_node_get_transport(id->device->node_type)) {
2686         case RDMA_TRANSPORT_IB:
2687                 if (id->qp_type == IB_QPT_UD)
2688                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2689                 else
2690                         ret = cma_connect_ib(id_priv, conn_param);
2691                 break;
2692         case RDMA_TRANSPORT_IWARP:
2693                 ret = cma_connect_iw(id_priv, conn_param);
2694                 break;
2695         default:
2696                 ret = -ENOSYS;
2697                 break;
2698         }
2699         if (ret)
2700                 goto err;
2701
2702         return 0;
2703 err:
2704         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2705         return ret;
2706 }
2707 EXPORT_SYMBOL(rdma_connect);
2708
2709 static int cma_accept_ib(struct rdma_id_private *id_priv,
2710                          struct rdma_conn_param *conn_param)
2711 {
2712         struct ib_cm_rep_param rep;
2713         int ret;
2714
2715         ret = cma_modify_qp_rtr(id_priv, conn_param);
2716         if (ret)
2717                 goto out;
2718
2719         ret = cma_modify_qp_rts(id_priv, conn_param);
2720         if (ret)
2721                 goto out;
2722
2723         memset(&rep, 0, sizeof rep);
2724         rep.qp_num = id_priv->qp_num;
2725         rep.starting_psn = id_priv->seq_num;
2726         rep.private_data = conn_param->private_data;
2727         rep.private_data_len = conn_param->private_data_len;
2728         rep.responder_resources = conn_param->responder_resources;
2729         rep.initiator_depth = conn_param->initiator_depth;
2730         rep.failover_accepted = 0;
2731         rep.flow_control = conn_param->flow_control;
2732         rep.rnr_retry_count = conn_param->rnr_retry_count;
2733         rep.srq = id_priv->srq ? 1 : 0;
2734
2735         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2736 out:
2737         return ret;
2738 }
2739
2740 static int cma_accept_iw(struct rdma_id_private *id_priv,
2741                   struct rdma_conn_param *conn_param)
2742 {
2743         struct iw_cm_conn_param iw_param;
2744         int ret;
2745
2746         ret = cma_modify_qp_rtr(id_priv, conn_param);
2747         if (ret)
2748                 return ret;
2749
2750         iw_param.ord = conn_param->initiator_depth;
2751         iw_param.ird = conn_param->responder_resources;
2752         iw_param.private_data = conn_param->private_data;
2753         iw_param.private_data_len = conn_param->private_data_len;
2754         if (id_priv->id.qp) {
2755                 iw_param.qpn = id_priv->qp_num;
2756         } else
2757                 iw_param.qpn = conn_param->qp_num;
2758
2759         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2760 }
2761
2762 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2763                              enum ib_cm_sidr_status status,
2764                              const void *private_data, int private_data_len)
2765 {
2766         struct ib_cm_sidr_rep_param rep;
2767         int ret;
2768
2769         memset(&rep, 0, sizeof rep);
2770         rep.status = status;
2771         if (status == IB_SIDR_SUCCESS) {
2772                 ret = cma_set_qkey(id_priv);
2773                 if (ret)
2774                         return ret;
2775                 rep.qp_num = id_priv->qp_num;
2776                 rep.qkey = id_priv->qkey;
2777         }
2778         rep.private_data = private_data;
2779         rep.private_data_len = private_data_len;
2780
2781         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2782 }
2783
2784 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2785 {
2786         struct rdma_id_private *id_priv;
2787         int ret;
2788
2789         id_priv = container_of(id, struct rdma_id_private, id);
2790
2791         id_priv->owner = task_pid_nr(current);
2792
2793         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2794                 return -EINVAL;
2795
2796         if (!id->qp && conn_param) {
2797                 id_priv->qp_num = conn_param->qp_num;
2798                 id_priv->srq = conn_param->srq;
2799         }
2800
2801         switch (rdma_node_get_transport(id->device->node_type)) {
2802         case RDMA_TRANSPORT_IB:
2803                 if (id->qp_type == IB_QPT_UD) {
2804                         if (conn_param)
2805                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2806                                                         conn_param->private_data,
2807                                                         conn_param->private_data_len);
2808                         else
2809                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2810                                                         NULL, 0);
2811                 } else {
2812                         if (conn_param)
2813                                 ret = cma_accept_ib(id_priv, conn_param);
2814                         else
2815                                 ret = cma_rep_recv(id_priv);
2816                 }
2817                 break;
2818         case RDMA_TRANSPORT_IWARP:
2819                 ret = cma_accept_iw(id_priv, conn_param);
2820                 break;
2821         default:
2822                 ret = -ENOSYS;
2823                 break;
2824         }
2825
2826         if (ret)
2827                 goto reject;
2828
2829         return 0;
2830 reject:
2831         cma_modify_qp_err(id_priv);
2832         rdma_reject(id, NULL, 0);
2833         return ret;
2834 }
2835 EXPORT_SYMBOL(rdma_accept);
2836
2837 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2838 {
2839         struct rdma_id_private *id_priv;
2840         int ret;
2841
2842         id_priv = container_of(id, struct rdma_id_private, id);
2843         if (!id_priv->cm_id.ib)
2844                 return -EINVAL;
2845
2846         switch (id->device->node_type) {
2847         case RDMA_NODE_IB_CA:
2848                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2849                 break;
2850         default:
2851                 ret = 0;
2852                 break;
2853         }
2854         return ret;
2855 }
2856 EXPORT_SYMBOL(rdma_notify);
2857
2858 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2859                 u8 private_data_len)
2860 {
2861         struct rdma_id_private *id_priv;
2862         int ret;
2863
2864         id_priv = container_of(id, struct rdma_id_private, id);
2865         if (!id_priv->cm_id.ib)
2866                 return -EINVAL;
2867
2868         switch (rdma_node_get_transport(id->device->node_type)) {
2869         case RDMA_TRANSPORT_IB:
2870                 if (id->qp_type == IB_QPT_UD)
2871                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2872                                                 private_data, private_data_len);
2873                 else
2874                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2875                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2876                                              0, private_data, private_data_len);
2877                 break;
2878         case RDMA_TRANSPORT_IWARP:
2879                 ret = iw_cm_reject(id_priv->cm_id.iw,
2880                                    private_data, private_data_len);
2881                 break;
2882         default:
2883                 ret = -ENOSYS;
2884                 break;
2885         }
2886         return ret;
2887 }
2888 EXPORT_SYMBOL(rdma_reject);
2889
2890 int rdma_disconnect(struct rdma_cm_id *id)
2891 {
2892         struct rdma_id_private *id_priv;
2893         int ret;
2894
2895         id_priv = container_of(id, struct rdma_id_private, id);
2896         if (!id_priv->cm_id.ib)
2897                 return -EINVAL;
2898
2899         switch (rdma_node_get_transport(id->device->node_type)) {
2900         case RDMA_TRANSPORT_IB:
2901                 ret = cma_modify_qp_err(id_priv);
2902                 if (ret)
2903                         goto out;
2904                 /* Initiate or respond to a disconnect. */
2905                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2906                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2907                 break;
2908         case RDMA_TRANSPORT_IWARP:
2909                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2910                 break;
2911         default:
2912                 ret = -EINVAL;
2913                 break;
2914         }
2915 out:
2916         return ret;
2917 }
2918 EXPORT_SYMBOL(rdma_disconnect);
2919
2920 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2921 {
2922         struct rdma_id_private *id_priv;
2923         struct cma_multicast *mc = multicast->context;
2924         struct rdma_cm_event event;
2925         int ret;
2926
2927         id_priv = mc->id_priv;
2928         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2929             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2930                 return 0;
2931
2932         mutex_lock(&id_priv->qp_mutex);
2933         if (!status && id_priv->id.qp)
2934                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2935                                          be16_to_cpu(multicast->rec.mlid));
2936         mutex_unlock(&id_priv->qp_mutex);
2937
2938         memset(&event, 0, sizeof event);
2939         event.status = status;
2940         event.param.ud.private_data = mc->context;
2941         if (!status) {
2942                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2943                 ib_init_ah_from_mcmember(id_priv->id.device,
2944                                          id_priv->id.port_num, &multicast->rec,
2945                                          &event.param.ud.ah_attr);
2946                 event.param.ud.qp_num = 0xFFFFFF;
2947                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2948         } else
2949                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2950
2951         ret = id_priv->id.event_handler(&id_priv->id, &event);
2952         if (ret) {
2953                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2954                 mutex_unlock(&id_priv->handler_mutex);
2955                 rdma_destroy_id(&id_priv->id);
2956                 return 0;
2957         }
2958
2959         mutex_unlock(&id_priv->handler_mutex);
2960         return 0;
2961 }
2962
2963 static void cma_set_mgid(struct rdma_id_private *id_priv,
2964                          struct sockaddr *addr, union ib_gid *mgid)
2965 {
2966         unsigned char mc_map[MAX_ADDR_LEN];
2967         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2968         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2969         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2970
2971         if (cma_any_addr(addr)) {
2972                 memset(mgid, 0, sizeof *mgid);
2973         } else if ((addr->sa_family == AF_INET6) &&
2974                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2975                                                                  0xFF10A01B)) {
2976                 /* IPv6 address is an SA assigned MGID. */
2977                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2978         } else if ((addr->sa_family == AF_INET6)) {
2979                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2980                 if (id_priv->id.ps == RDMA_PS_UDP)
2981                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2982                 *mgid = *(union ib_gid *) (mc_map + 4);
2983         } else {
2984                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2985                 if (id_priv->id.ps == RDMA_PS_UDP)
2986                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2987                 *mgid = *(union ib_gid *) (mc_map + 4);
2988         }
2989 }
2990
2991 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2992                                  struct cma_multicast *mc)
2993 {
2994         struct ib_sa_mcmember_rec rec;
2995         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2996         ib_sa_comp_mask comp_mask;
2997         int ret;
2998
2999         ib_addr_get_mgid(dev_addr, &rec.mgid);
3000         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3001                                      &rec.mgid, &rec);
3002         if (ret)
3003                 return ret;
3004
3005         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3006         if (id_priv->id.ps == RDMA_PS_UDP)
3007                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3008         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3009         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3010         rec.join_state = 1;
3011
3012         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3013                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3014                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3015                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3016                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3017
3018         if (id_priv->id.ps == RDMA_PS_IPOIB)
3019                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3020                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
3021
3022         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3023                                                 id_priv->id.port_num, &rec,
3024                                                 comp_mask, GFP_KERNEL,
3025                                                 cma_ib_mc_handler, mc);
3026         if (IS_ERR(mc->multicast.ib))
3027                 return PTR_ERR(mc->multicast.ib);
3028
3029         return 0;
3030 }
3031
3032 static void iboe_mcast_work_handler(struct work_struct *work)
3033 {
3034         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3035         struct cma_multicast *mc = mw->mc;
3036         struct ib_sa_multicast *m = mc->multicast.ib;
3037
3038         mc->multicast.ib->context = mc;
3039         cma_ib_mc_handler(0, m);
3040         kref_put(&mc->mcref, release_mc);
3041         kfree(mw);
3042 }
3043
3044 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3045 {
3046         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3047         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3048
3049         if (cma_any_addr(addr)) {
3050                 memset(mgid, 0, sizeof *mgid);
3051         } else if (addr->sa_family == AF_INET6) {
3052                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3053         } else {
3054                 mgid->raw[0] = 0xff;
3055                 mgid->raw[1] = 0x0e;
3056                 mgid->raw[2] = 0;
3057                 mgid->raw[3] = 0;
3058                 mgid->raw[4] = 0;
3059                 mgid->raw[5] = 0;
3060                 mgid->raw[6] = 0;
3061                 mgid->raw[7] = 0;
3062                 mgid->raw[8] = 0;
3063                 mgid->raw[9] = 0;
3064                 mgid->raw[10] = 0xff;
3065                 mgid->raw[11] = 0xff;
3066                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3067         }
3068 }
3069
3070 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3071                                    struct cma_multicast *mc)
3072 {
3073         struct iboe_mcast_work *work;
3074         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3075         int err;
3076         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3077         struct net_device *ndev = NULL;
3078
3079         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3080                 return -EINVAL;
3081
3082         work = kzalloc(sizeof *work, GFP_KERNEL);
3083         if (!work)
3084                 return -ENOMEM;
3085
3086         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3087         if (!mc->multicast.ib) {
3088                 err = -ENOMEM;
3089                 goto out1;
3090         }
3091
3092         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3093
3094         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3095         if (id_priv->id.ps == RDMA_PS_UDP)
3096                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3097
3098         if (dev_addr->bound_dev_if)
3099                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3100         if (!ndev) {
3101                 err = -ENODEV;
3102                 goto out2;
3103         }
3104         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3105         mc->multicast.ib->rec.hop_limit = 1;
3106         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3107         dev_put(ndev);
3108         if (!mc->multicast.ib->rec.mtu) {
3109                 err = -EINVAL;
3110                 goto out2;
3111         }
3112         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3113         work->id = id_priv;
3114         work->mc = mc;
3115         INIT_WORK(&work->work, iboe_mcast_work_handler);
3116         kref_get(&mc->mcref);
3117         queue_work(cma_wq, &work->work);
3118
3119         return 0;
3120
3121 out2:
3122         kfree(mc->multicast.ib);
3123 out1:
3124         kfree(work);
3125         return err;
3126 }
3127
3128 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3129                         void *context)
3130 {
3131         struct rdma_id_private *id_priv;
3132         struct cma_multicast *mc;
3133         int ret;
3134
3135         id_priv = container_of(id, struct rdma_id_private, id);
3136         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3137             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3138                 return -EINVAL;
3139
3140         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3141         if (!mc)
3142                 return -ENOMEM;
3143
3144         memcpy(&mc->addr, addr, ip_addr_size(addr));
3145         mc->context = context;
3146         mc->id_priv = id_priv;
3147
3148         spin_lock(&id_priv->lock);
3149         list_add(&mc->list, &id_priv->mc_list);
3150         spin_unlock(&id_priv->lock);
3151
3152         switch (rdma_node_get_transport(id->device->node_type)) {
3153         case RDMA_TRANSPORT_IB:
3154                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3155                 case IB_LINK_LAYER_INFINIBAND:
3156                         ret = cma_join_ib_multicast(id_priv, mc);
3157                         break;
3158                 case IB_LINK_LAYER_ETHERNET:
3159                         kref_init(&mc->mcref);
3160                         ret = cma_iboe_join_multicast(id_priv, mc);
3161                         break;
3162                 default:
3163                         ret = -EINVAL;
3164                 }
3165                 break;
3166         default:
3167                 ret = -ENOSYS;
3168                 break;
3169         }
3170
3171         if (ret) {
3172                 spin_lock_irq(&id_priv->lock);
3173                 list_del(&mc->list);
3174                 spin_unlock_irq(&id_priv->lock);
3175                 kfree(mc);
3176         }
3177         return ret;
3178 }
3179 EXPORT_SYMBOL(rdma_join_multicast);
3180
3181 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3182 {
3183         struct rdma_id_private *id_priv;
3184         struct cma_multicast *mc;
3185
3186         id_priv = container_of(id, struct rdma_id_private, id);
3187         spin_lock_irq(&id_priv->lock);
3188         list_for_each_entry(mc, &id_priv->mc_list, list) {
3189                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3190                         list_del(&mc->list);
3191                         spin_unlock_irq(&id_priv->lock);
3192
3193                         if (id->qp)
3194                                 ib_detach_mcast(id->qp,
3195                                                 &mc->multicast.ib->rec.mgid,
3196                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3197                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3198                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3199                                 case IB_LINK_LAYER_INFINIBAND:
3200                                         ib_sa_free_multicast(mc->multicast.ib);
3201                                         kfree(mc);
3202                                         break;
3203                                 case IB_LINK_LAYER_ETHERNET:
3204                                         kref_put(&mc->mcref, release_mc);
3205                                         break;
3206                                 default:
3207                                         break;
3208                                 }
3209                         }
3210                         return;
3211                 }
3212         }
3213         spin_unlock_irq(&id_priv->lock);
3214 }
3215 EXPORT_SYMBOL(rdma_leave_multicast);
3216
3217 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3218 {
3219         struct rdma_dev_addr *dev_addr;
3220         struct cma_ndev_work *work;
3221
3222         dev_addr = &id_priv->id.route.addr.dev_addr;
3223
3224         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3225             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3226                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3227                        ndev->name, &id_priv->id);
3228                 work = kzalloc(sizeof *work, GFP_KERNEL);
3229                 if (!work)
3230                         return -ENOMEM;
3231
3232                 INIT_WORK(&work->work, cma_ndev_work_handler);
3233                 work->id = id_priv;
3234                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3235                 atomic_inc(&id_priv->refcount);
3236                 queue_work(cma_wq, &work->work);
3237         }
3238
3239         return 0;
3240 }
3241
3242 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3243                                void *ctx)
3244 {
3245         struct net_device *ndev = (struct net_device *)ctx;
3246         struct cma_device *cma_dev;
3247         struct rdma_id_private *id_priv;
3248         int ret = NOTIFY_DONE;
3249
3250         if (dev_net(ndev) != &init_net)
3251                 return NOTIFY_DONE;
3252
3253         if (event != NETDEV_BONDING_FAILOVER)
3254                 return NOTIFY_DONE;
3255
3256         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3257                 return NOTIFY_DONE;
3258
3259         mutex_lock(&lock);
3260         list_for_each_entry(cma_dev, &dev_list, list)
3261                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3262                         ret = cma_netdev_change(ndev, id_priv);
3263                         if (ret)
3264                                 goto out;
3265                 }
3266
3267 out:
3268         mutex_unlock(&lock);
3269         return ret;
3270 }
3271
3272 static struct notifier_block cma_nb = {
3273         .notifier_call = cma_netdev_callback
3274 };
3275
3276 static void cma_add_one(struct ib_device *device)
3277 {
3278         struct cma_device *cma_dev;
3279         struct rdma_id_private *id_priv;
3280
3281         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3282         if (!cma_dev)
3283                 return;
3284
3285         cma_dev->device = device;
3286
3287         init_completion(&cma_dev->comp);
3288         atomic_set(&cma_dev->refcount, 1);
3289         INIT_LIST_HEAD(&cma_dev->id_list);
3290         ib_set_client_data(device, &cma_client, cma_dev);
3291
3292         mutex_lock(&lock);
3293         list_add_tail(&cma_dev->list, &dev_list);
3294         list_for_each_entry(id_priv, &listen_any_list, list)
3295                 cma_listen_on_dev(id_priv, cma_dev);
3296         mutex_unlock(&lock);
3297 }
3298
3299 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3300 {
3301         struct rdma_cm_event event;
3302         enum rdma_cm_state state;
3303         int ret = 0;
3304
3305         /* Record that we want to remove the device */
3306         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3307         if (state == RDMA_CM_DESTROYING)
3308                 return 0;
3309
3310         cma_cancel_operation(id_priv, state);
3311         mutex_lock(&id_priv->handler_mutex);
3312
3313         /* Check for destruction from another callback. */
3314         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3315                 goto out;
3316
3317         memset(&event, 0, sizeof event);
3318         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3319         ret = id_priv->id.event_handler(&id_priv->id, &event);
3320 out:
3321         mutex_unlock(&id_priv->handler_mutex);
3322         return ret;
3323 }
3324
3325 static void cma_process_remove(struct cma_device *cma_dev)
3326 {
3327         struct rdma_id_private *id_priv;
3328         int ret;
3329
3330         mutex_lock(&lock);
3331         while (!list_empty(&cma_dev->id_list)) {
3332                 id_priv = list_entry(cma_dev->id_list.next,
3333                                      struct rdma_id_private, list);
3334
3335                 list_del(&id_priv->listen_list);
3336                 list_del_init(&id_priv->list);
3337                 atomic_inc(&id_priv->refcount);
3338                 mutex_unlock(&lock);
3339
3340                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3341                 cma_deref_id(id_priv);
3342                 if (ret)
3343                         rdma_destroy_id(&id_priv->id);
3344
3345                 mutex_lock(&lock);
3346         }
3347         mutex_unlock(&lock);
3348
3349         cma_deref_dev(cma_dev);
3350         wait_for_completion(&cma_dev->comp);
3351 }
3352
3353 static void cma_remove_one(struct ib_device *device)
3354 {
3355         struct cma_device *cma_dev;
3356
3357         cma_dev = ib_get_client_data(device, &cma_client);
3358         if (!cma_dev)
3359                 return;
3360
3361         mutex_lock(&lock);
3362         list_del(&cma_dev->list);
3363         mutex_unlock(&lock);
3364
3365         cma_process_remove(cma_dev);
3366         kfree(cma_dev);
3367 }
3368
3369 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3370 {
3371         struct nlmsghdr *nlh;
3372         struct rdma_cm_id_stats *id_stats;
3373         struct rdma_id_private *id_priv;
3374         struct rdma_cm_id *id = NULL;
3375         struct cma_device *cma_dev;
3376         int i_dev = 0, i_id = 0;
3377
3378         /*
3379          * We export all of the IDs as a sequence of messages.  Each
3380          * ID gets its own netlink message.
3381          */
3382         mutex_lock(&lock);
3383
3384         list_for_each_entry(cma_dev, &dev_list, list) {
3385                 if (i_dev < cb->args[0]) {
3386                         i_dev++;
3387                         continue;
3388                 }
3389
3390                 i_id = 0;
3391                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3392                         if (i_id < cb->args[1]) {
3393                                 i_id++;
3394                                 continue;
3395                         }
3396
3397                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3398                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3399                                                 RDMA_NL_RDMA_CM_ID_STATS);
3400                         if (!id_stats)
3401                                 goto out;
3402
3403                         memset(id_stats, 0, sizeof *id_stats);
3404                         id = &id_priv->id;
3405                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3406                         id_stats->port_num = id->port_num;
3407                         id_stats->bound_dev_if =
3408                                 id->route.addr.dev_addr.bound_dev_if;
3409
3410                         if (id->route.addr.src_addr.ss_family == AF_INET) {
3411                                 if (ibnl_put_attr(skb, nlh,
3412                                                   sizeof(struct sockaddr_in),
3413                                                   &id->route.addr.src_addr,
3414                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3415                                         goto out;
3416                                 }
3417                                 if (ibnl_put_attr(skb, nlh,
3418                                                   sizeof(struct sockaddr_in),
3419                                                   &id->route.addr.dst_addr,
3420                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3421                                         goto out;
3422                                 }
3423                         } else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3424                                 if (ibnl_put_attr(skb, nlh,
3425                                                   sizeof(struct sockaddr_in6),
3426                                                   &id->route.addr.src_addr,
3427                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3428                                         goto out;
3429                                 }
3430                                 if (ibnl_put_attr(skb, nlh,
3431                                                   sizeof(struct sockaddr_in6),
3432                                                   &id->route.addr.dst_addr,
3433                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3434                                         goto out;
3435                                 }
3436                         }
3437
3438                         id_stats->pid           = id_priv->owner;
3439                         id_stats->port_space    = id->ps;
3440                         id_stats->cm_state      = id_priv->state;
3441                         id_stats->qp_num        = id_priv->qp_num;
3442                         id_stats->qp_type       = id->qp_type;
3443
3444                         i_id++;
3445                 }
3446
3447                 cb->args[1] = 0;
3448                 i_dev++;
3449         }
3450
3451 out:
3452         mutex_unlock(&lock);
3453         cb->args[0] = i_dev;
3454         cb->args[1] = i_id;
3455
3456         return skb->len;
3457 }
3458
3459 static const struct ibnl_client_cbs cma_cb_table[] = {
3460         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats },
3461 };
3462
3463 static int __init cma_init(void)
3464 {
3465         int ret;
3466
3467         cma_wq = create_singlethread_workqueue("rdma_cm");
3468         if (!cma_wq)
3469                 return -ENOMEM;
3470
3471         ib_sa_register_client(&sa_client);
3472         rdma_addr_register_client(&addr_client);
3473         register_netdevice_notifier(&cma_nb);
3474
3475         ret = ib_register_client(&cma_client);
3476         if (ret)
3477                 goto err;
3478
3479         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3480                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3481
3482         return 0;
3483
3484 err:
3485         unregister_netdevice_notifier(&cma_nb);
3486         rdma_addr_unregister_client(&addr_client);
3487         ib_sa_unregister_client(&sa_client);
3488         destroy_workqueue(cma_wq);
3489         return ret;
3490 }
3491
3492 static void __exit cma_cleanup(void)
3493 {
3494         ibnl_remove_client(RDMA_NL_RDMA_CM);
3495         ib_unregister_client(&cma_client);
3496         unregister_netdevice_notifier(&cma_nb);
3497         rdma_addr_unregister_client(&addr_client);
3498         ib_sa_unregister_client(&sa_client);
3499         destroy_workqueue(cma_wq);
3500         idr_destroy(&sdp_ps);
3501         idr_destroy(&tcp_ps);
3502         idr_destroy(&udp_ps);
3503         idr_destroy(&ipoib_ps);
3504         idr_destroy(&ib_ps);
3505 }
3506
3507 module_init(cma_init);
3508 module_exit(cma_cleanup);