drivers/infiniband/core/cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/idr.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/slab.h>
  44
  45 #include <net/tcp.h>
  46 #include <net/ipv6.h>
  47
  48 #include <rdma/rdma_cm.h>
  49 #include <rdma/rdma_cm_ib.h>
  50 #include <rdma/rdma_netlink.h>
  51 #include <rdma/ib_cache.h>
  52 #include <rdma/ib_cm.h>
  53 #include <rdma/ib_sa.h>
  54 #include <rdma/iw_cm.h>
  55
  56 MODULE_AUTHOR("Sean Hefty");
  57 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59
  60 #define CMA_CM_RESPONSE_TIMEOUT 20
  61 #define CMA_MAX_CM_RETRIES 15
  62 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  63 #define CMA_IBOE_PACKET_LIFETIME 18
  64
  65 static void cma_add_one(struct ib_device *device);
  66 static void cma_remove_one(struct ib_device *device);
  67
  68 static struct ib_client cma_client = {
  69         .name   = "cma",
  70         .add    = cma_add_one,
  71         .remove = cma_remove_one
  72 };
  73
  74 static struct ib_sa_client sa_client;
  75 static struct rdma_addr_client addr_client;
  76 static LIST_HEAD(dev_list);
  77 static LIST_HEAD(listen_any_list);
  78 static DEFINE_MUTEX(lock);
  79 static struct workqueue_struct *cma_wq;
  80 static DEFINE_IDR(sdp_ps);
  81 static DEFINE_IDR(tcp_ps);
  82 static DEFINE_IDR(udp_ps);
  83 static DEFINE_IDR(ipoib_ps);
  84 static DEFINE_IDR(ib_ps);
  85
  86 struct cma_device {
  87         struct list_head        list;
  88         struct ib_device        *device;
  89         struct completion       comp;
  90         atomic_t                refcount;
  91         struct list_head        id_list;
  92 };
  93
  94 struct rdma_bind_list {
  95         struct idr              *ps;
  96         struct hlist_head       owners;
  97         unsigned short          port;
  98 };
  99
 100 /*
 101  * Device removal can occur at anytime, so we need extra handling to
 102  * serialize notifying the user of device removal with other callbacks.
 103  * We do this by disabling removal notification while a callback is in process,
 104  * and reporting it after the callback completes.
 105  */
 106 struct rdma_id_private {
 107         struct rdma_cm_id       id;
 108
 109         struct rdma_bind_list   *bind_list;
 110         struct hlist_node       node;
 111         struct list_head        list; /* listen_any_list or cma_device.list */
 112         struct list_head        listen_list; /* per device listens */
 113         struct cma_device       *cma_dev;
 114         struct list_head        mc_list;
 115
 116         int                     internal_id;
 117         enum rdma_cm_state      state;
 118         spinlock_t              lock;
 119         struct mutex            qp_mutex;
 120
 121         struct completion       comp;
 122         atomic_t                refcount;
 123         struct mutex            handler_mutex;
 124
 125         int                     backlog;
 126         int                     timeout_ms;
 127         struct ib_sa_query      *query;
 128         int                     query_id;
 129         union {
 130                 struct ib_cm_id *ib;
 131                 struct iw_cm_id *iw;
 132         } cm_id;
 133
 134         u32                     seq_num;
 135         u32                     qkey;
 136         u32                     qp_num;
 137         pid_t                   owner;
 138         u8                      srq;
 139         u8                      tos;
 140         u8                      reuseaddr;
 141 };
 142
 143 struct cma_multicast {
 144         struct rdma_id_private *id_priv;
 145         union {
 146                 struct ib_sa_multicast *ib;
 147         } multicast;
 148         struct list_head        list;
 149         void                    *context;
 150         struct sockaddr_storage addr;
 151         struct kref             mcref;
 152 };
 153
 154 struct cma_work {
 155         struct work_struct      work;
 156         struct rdma_id_private  *id;
 157         enum rdma_cm_state      old_state;
 158         enum rdma_cm_state      new_state;
 159         struct rdma_cm_event    event;
 160 };
 161
 162 struct cma_ndev_work {
 163         struct work_struct      work;
 164         struct rdma_id_private  *id;
 165         struct rdma_cm_event    event;
 166 };
 167
 168 struct iboe_mcast_work {
 169         struct work_struct       work;
 170         struct rdma_id_private  *id;
 171         struct cma_multicast    *mc;
 172 };
 173
 174 union cma_ip_addr {
 175         struct in6_addr ip6;
 176         struct {
 177                 __be32 pad[3];
 178                 __be32 addr;
 179         } ip4;
 180 };
 181
 182 struct cma_hdr {
 183         u8 cma_version;
 184         u8 ip_version;  /* IP version: 7:4 */
 185         __be16 port;
 186         union cma_ip_addr src_addr;
 187         union cma_ip_addr dst_addr;
 188 };
 189
 190 struct sdp_hh {
 191         u8 bsdh[16];
 192         u8 sdp_version; /* Major version: 7:4 */
 193         u8 ip_version;  /* IP version: 7:4 */
 194         u8 sdp_specific1[10];
 195         __be16 port;
 196         __be16 sdp_specific2;
 197         union cma_ip_addr src_addr;
 198         union cma_ip_addr dst_addr;
 199 };
 200
 201 struct sdp_hah {
 202         u8 bsdh[16];
 203         u8 sdp_version;
 204 };
 205
 206 #define CMA_VERSION 0x00
 207 #define SDP_MAJ_VERSION 0x2
 208
 209 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 210 {
 211         unsigned long flags;
 212         int ret;
 213
 214         spin_lock_irqsave(&id_priv->lock, flags);
 215         ret = (id_priv->state == comp);
 216         spin_unlock_irqrestore(&id_priv->lock, flags);
 217         return ret;
 218 }
 219
 220 static int cma_comp_exch(struct rdma_id_private *id_priv,
 221                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 222 {
 223         unsigned long flags;
 224         int ret;
 225
 226         spin_lock_irqsave(&id_priv->lock, flags);
 227         if ((ret = (id_priv->state == comp)))
 228                 id_priv->state = exch;
 229         spin_unlock_irqrestore(&id_priv->lock, flags);
 230         return ret;
 231 }
 232
 233 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 234                                    enum rdma_cm_state exch)
 235 {
 236         unsigned long flags;
 237         enum rdma_cm_state old;
 238
 239         spin_lock_irqsave(&id_priv->lock, flags);
 240         old = id_priv->state;
 241         id_priv->state = exch;
 242         spin_unlock_irqrestore(&id_priv->lock, flags);
 243         return old;
 244 }
 245
 246 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 247 {
 248         return hdr->ip_version >> 4;
 249 }
 250
 251 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 252 {
 253         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 254 }
 255
 256 static inline u8 sdp_get_majv(u8 sdp_version)
 257 {
 258         return sdp_version >> 4;
 259 }
 260
 261 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 262 {
 263         return hh->ip_version >> 4;
 264 }
 265
 266 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 267 {
 268         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 269 }
 270
 271 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 272                               struct cma_device *cma_dev)
 273 {
 274         atomic_inc(&cma_dev->refcount);
 275         id_priv->cma_dev = cma_dev;
 276         id_priv->id.device = cma_dev->device;
 277         id_priv->id.route.addr.dev_addr.transport =
 278                 rdma_node_get_transport(cma_dev->device->node_type);
 279         list_add_tail(&id_priv->list, &cma_dev->id_list);
 280 }
 281
 282 static inline void cma_deref_dev(struct cma_device *cma_dev)
 283 {
 284         if (atomic_dec_and_test(&cma_dev->refcount))
 285                 complete(&cma_dev->comp);
 286 }
 287
 288 static inline void release_mc(struct kref *kref)
 289 {
 290         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 291
 292         kfree(mc->multicast.ib);
 293         kfree(mc);
 294 }
 295
 296 static void cma_release_dev(struct rdma_id_private *id_priv)
 297 {
 298         mutex_lock(&lock);
 299         list_del(&id_priv->list);
 300         cma_deref_dev(id_priv->cma_dev);
 301         id_priv->cma_dev = NULL;
 302         mutex_unlock(&lock);
 303 }
 304
 305 static int cma_set_qkey(struct rdma_id_private *id_priv)
 306 {
 307         struct ib_sa_mcmember_rec rec;
 308         int ret = 0;
 309
 310         if (id_priv->qkey)
 311                 return 0;
 312
 313         switch (id_priv->id.ps) {
 314         case RDMA_PS_UDP:
 315                 id_priv->qkey = RDMA_UDP_QKEY;
 316                 break;
 317         case RDMA_PS_IPOIB:
 318                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 319                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 320                                              id_priv->id.port_num, &rec.mgid,
 321                                              &rec);
 322                 if (!ret)
 323                         id_priv->qkey = be32_to_cpu(rec.qkey);
 324                 break;
 325         default:
 326                 break;
 327         }
 328         return ret;
 329 }
 330
 331 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 332 {
 333         int i;
 334         int err;
 335         struct ib_port_attr props;
 336         union ib_gid tmp;
 337
 338         err = ib_query_port(device, port_num, &props);
 339         if (err)
 340                 return 1;
 341
 342         for (i = 0; i < props.gid_tbl_len; ++i) {
 343                 err = ib_query_gid(device, port_num, i, &tmp);
 344                 if (err)
 345                         return 1;
 346                 if (!memcmp(&tmp, gid, sizeof tmp))
 347                         return 0;
 348         }
 349
 350         return -EAGAIN;
 351 }
 352
 353 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 354 {
 355         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 356         struct cma_device *cma_dev;
 357         union ib_gid gid, iboe_gid;
 358         int ret = -ENODEV;
 359         u8 port;
 360         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 361                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 362
 363         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 364             id_priv->id.ps == RDMA_PS_IPOIB)
 365                 return -EINVAL;
 366
 367         mutex_lock(&lock);
 368         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 369         memcpy(&gid, dev_addr->src_dev_addr +
 370                rdma_addr_gid_offset(dev_addr), sizeof gid);
 371         list_for_each_entry(cma_dev, &dev_list, list) {
 372                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 373                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 374                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 375                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 376                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 377                                 else
 378                                         ret = find_gid_port(cma_dev->device, &gid, port);
 379
 380                                 if (!ret) {
 381                                         id_priv->id.port_num = port;
 382                                         goto out;
 383                                 } else if (ret == 1)
 384                                         break;
 385                         }
 386                 }
 387         }
 388
 389 out:
 390         if (!ret)
 391                 cma_attach_to_dev(id_priv, cma_dev);
 392
 393         mutex_unlock(&lock);
 394         return ret;
 395 }
 396
 397 static void cma_deref_id(struct rdma_id_private *id_priv)
 398 {
 399         if (atomic_dec_and_test(&id_priv->refcount))
 400                 complete(&id_priv->comp);
 401 }
 402
 403 static int cma_disable_callback(struct rdma_id_private *id_priv,
 404                                 enum rdma_cm_state state)
 405 {
 406         mutex_lock(&id_priv->handler_mutex);
 407         if (id_priv->state != state) {
 408                 mutex_unlock(&id_priv->handler_mutex);
 409                 return -EINVAL;
 410         }
 411         return 0;
 412 }
 413
 414 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 415                                   void *context, enum rdma_port_space ps,
 416                                   enum ib_qp_type qp_type)
 417 {
 418         struct rdma_id_private *id_priv;
 419
 420         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 421         if (!id_priv)
 422                 return ERR_PTR(-ENOMEM);
 423
 424         id_priv->owner = task_pid_nr(current);
 425         id_priv->state = RDMA_CM_IDLE;
 426         id_priv->id.context = context;
 427         id_priv->id.event_handler = event_handler;
 428         id_priv->id.ps = ps;
 429         id_priv->id.qp_type = qp_type;
 430         spin_lock_init(&id_priv->lock);
 431         mutex_init(&id_priv->qp_mutex);
 432         init_completion(&id_priv->comp);
 433         atomic_set(&id_priv->refcount, 1);
 434         mutex_init(&id_priv->handler_mutex);
 435         INIT_LIST_HEAD(&id_priv->listen_list);
 436         INIT_LIST_HEAD(&id_priv->mc_list);
 437         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 438
 439         return &id_priv->id;
 440 }
 441 EXPORT_SYMBOL(rdma_create_id);
 442
 443 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 444 {
 445         struct ib_qp_attr qp_attr;
 446         int qp_attr_mask, ret;
 447
 448         qp_attr.qp_state = IB_QPS_INIT;
 449         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 450         if (ret)
 451                 return ret;
 452
 453         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 454         if (ret)
 455                 return ret;
 456
 457         qp_attr.qp_state = IB_QPS_RTR;
 458         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 459         if (ret)
 460                 return ret;
 461
 462         qp_attr.qp_state = IB_QPS_RTS;
 463         qp_attr.sq_psn = 0;
 464         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 465
 466         return ret;
 467 }
 468
 469 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 470 {
 471         struct ib_qp_attr qp_attr;
 472         int qp_attr_mask, ret;
 473
 474         qp_attr.qp_state = IB_QPS_INIT;
 475         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 476         if (ret)
 477                 return ret;
 478
 479         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 480 }
 481
 482 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 483                    struct ib_qp_init_attr *qp_init_attr)
 484 {
 485         struct rdma_id_private *id_priv;
 486         struct ib_qp *qp;
 487         int ret;
 488
 489         id_priv = container_of(id, struct rdma_id_private, id);
 490         if (id->device != pd->device)
 491                 return -EINVAL;
 492
 493         qp = ib_create_qp(pd, qp_init_attr);
 494         if (IS_ERR(qp))
 495                 return PTR_ERR(qp);
 496
 497         if (id->qp_type == IB_QPT_UD)
 498                 ret = cma_init_ud_qp(id_priv, qp);
 499         else
 500                 ret = cma_init_conn_qp(id_priv, qp);
 501         if (ret)
 502                 goto err;
 503
 504         id->qp = qp;
 505         id_priv->qp_num = qp->qp_num;
 506         id_priv->srq = (qp->srq != NULL);
 507         return 0;
 508 err:
 509         ib_destroy_qp(qp);
 510         return ret;
 511 }
 512 EXPORT_SYMBOL(rdma_create_qp);
 513
 514 void rdma_destroy_qp(struct rdma_cm_id *id)
 515 {
 516         struct rdma_id_private *id_priv;
 517
 518         id_priv = container_of(id, struct rdma_id_private, id);
 519         mutex_lock(&id_priv->qp_mutex);
 520         ib_destroy_qp(id_priv->id.qp);
 521         id_priv->id.qp = NULL;
 522         mutex_unlock(&id_priv->qp_mutex);
 523 }
 524 EXPORT_SYMBOL(rdma_destroy_qp);
 525
 526 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 527                              struct rdma_conn_param *conn_param)
 528 {
 529         struct ib_qp_attr qp_attr;
 530         int qp_attr_mask, ret;
 531
 532         mutex_lock(&id_priv->qp_mutex);
 533         if (!id_priv->id.qp) {
 534                 ret = 0;
 535                 goto out;
 536         }
 537
 538         /* Need to update QP attributes from default values. */
 539         qp_attr.qp_state = IB_QPS_INIT;
 540         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 541         if (ret)
 542                 goto out;
 543
 544         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 545         if (ret)
 546                 goto out;
 547
 548         qp_attr.qp_state = IB_QPS_RTR;
 549         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 550         if (ret)
 551                 goto out;
 552
 553         if (conn_param)
 554                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 555         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 556 out:
 557         mutex_unlock(&id_priv->qp_mutex);
 558         return ret;
 559 }
 560
 561 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 562                              struct rdma_conn_param *conn_param)
 563 {
 564         struct ib_qp_attr qp_attr;
 565         int qp_attr_mask, ret;
 566
 567         mutex_lock(&id_priv->qp_mutex);
 568         if (!id_priv->id.qp) {
 569                 ret = 0;
 570                 goto out;
 571         }
 572
 573         qp_attr.qp_state = IB_QPS_RTS;
 574         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 575         if (ret)
 576                 goto out;
 577
 578         if (conn_param)
 579                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 580         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 581 out:
 582         mutex_unlock(&id_priv->qp_mutex);
 583         return ret;
 584 }
 585
 586 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 587 {
 588         struct ib_qp_attr qp_attr;
 589         int ret;
 590
 591         mutex_lock(&id_priv->qp_mutex);
 592         if (!id_priv->id.qp) {
 593                 ret = 0;
 594                 goto out;
 595         }
 596
 597         qp_attr.qp_state = IB_QPS_ERR;
 598         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 599 out:
 600         mutex_unlock(&id_priv->qp_mutex);
 601         return ret;
 602 }
 603
 604 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 605                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 606 {
 607         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 608         int ret;
 609         u16 pkey;
 610
 611         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 612             IB_LINK_LAYER_INFINIBAND)
 613                 pkey = ib_addr_get_pkey(dev_addr);
 614         else
 615                 pkey = 0xffff;
 616
 617         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 618                                   pkey, &qp_attr->pkey_index);
 619         if (ret)
 620                 return ret;
 621
 622         qp_attr->port_num = id_priv->id.port_num;
 623         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 624
 625         if (id_priv->id.qp_type == IB_QPT_UD) {
 626                 ret = cma_set_qkey(id_priv);
 627                 if (ret)
 628                         return ret;
 629
 630                 qp_attr->qkey = id_priv->qkey;
 631                 *qp_attr_mask |= IB_QP_QKEY;
 632         } else {
 633                 qp_attr->qp_access_flags = 0;
 634                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 635         }
 636         return 0;
 637 }
 638
 639 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 640                        int *qp_attr_mask)
 641 {
 642         struct rdma_id_private *id_priv;
 643         int ret = 0;
 644
 645         id_priv = container_of(id, struct rdma_id_private, id);
 646         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 647         case RDMA_TRANSPORT_IB:
 648                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 649                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 650                 else
 651                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 652                                                  qp_attr_mask);
 653                 if (qp_attr->qp_state == IB_QPS_RTR)
 654                         qp_attr->rq_psn = id_priv->seq_num;
 655                 break;
 656         case RDMA_TRANSPORT_IWARP:
 657                 if (!id_priv->cm_id.iw) {
 658                         qp_attr->qp_access_flags = 0;
 659                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 660                 } else
 661                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 662                                                  qp_attr_mask);
 663                 break;
 664         default:
 665                 ret = -ENOSYS;
 666                 break;
 667         }
 668
 669         return ret;
 670 }
 671 EXPORT_SYMBOL(rdma_init_qp_attr);
 672
 673 static inline int cma_zero_addr(struct sockaddr *addr)
 674 {
 675         struct in6_addr *ip6;
 676
 677         if (addr->sa_family == AF_INET)
 678                 return ipv4_is_zeronet(
 679                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 680         else {
 681                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 682                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 683                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 684         }
 685 }
 686
 687 static inline int cma_loopback_addr(struct sockaddr *addr)
 688 {
 689         if (addr->sa_family == AF_INET)
 690                 return ipv4_is_loopback(
 691                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 692         else
 693                 return ipv6_addr_loopback(
 694                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 695 }
 696
 697 static inline int cma_any_addr(struct sockaddr *addr)
 698 {
 699         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 700 }
 701
 702 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 703 {
 704         if (src->sa_family != dst->sa_family)
 705                 return -1;
 706
 707         switch (src->sa_family) {
 708         case AF_INET:
 709                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 710                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 711         default:
 712                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 713                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 714         }
 715 }
 716
 717 static inline __be16 cma_port(struct sockaddr *addr)
 718 {
 719         if (addr->sa_family == AF_INET)
 720                 return ((struct sockaddr_in *) addr)->sin_port;
 721         else
 722                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 723 }
 724
 725 static inline int cma_any_port(struct sockaddr *addr)
 726 {
 727         return !cma_port(addr);
 728 }
 729
 730 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 731                             u8 *ip_ver, __be16 *port,
 732                             union cma_ip_addr **src, union cma_ip_addr **dst)
 733 {
 734         switch (ps) {
 735         case RDMA_PS_SDP:
 736                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 737                     SDP_MAJ_VERSION)
 738                         return -EINVAL;
 739
 740                 *ip_ver = sdp_get_ip_ver(hdr);
 741                 *port   = ((struct sdp_hh *) hdr)->port;
 742                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 743                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 744                 break;
 745         default:
 746                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 747                         return -EINVAL;
 748
 749                 *ip_ver = cma_get_ip_ver(hdr);
 750                 *port   = ((struct cma_hdr *) hdr)->port;
 751                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 752                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 753                 break;
 754         }
 755
 756         if (*ip_ver != 4 && *ip_ver != 6)
 757                 return -EINVAL;
 758         return 0;
 759 }
 760
 761 static void cma_save_net_info(struct rdma_addr *addr,
 762                               struct rdma_addr *listen_addr,
 763                               u8 ip_ver, __be16 port,
 764                               union cma_ip_addr *src, union cma_ip_addr *dst)
 765 {
 766         struct sockaddr_in *listen4, *ip4;
 767         struct sockaddr_in6 *listen6, *ip6;
 768
 769         switch (ip_ver) {
 770         case 4:
 771                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 772                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 773                 ip4->sin_family = listen4->sin_family;
 774                 ip4->sin_addr.s_addr = dst->ip4.addr;
 775                 ip4->sin_port = listen4->sin_port;
 776
 777                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 778                 ip4->sin_family = listen4->sin_family;
 779                 ip4->sin_addr.s_addr = src->ip4.addr;
 780                 ip4->sin_port = port;
 781                 break;
 782         case 6:
 783                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 784                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 785                 ip6->sin6_family = listen6->sin6_family;
 786                 ip6->sin6_addr = dst->ip6;
 787                 ip6->sin6_port = listen6->sin6_port;
 788
 789                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 790                 ip6->sin6_family = listen6->sin6_family;
 791                 ip6->sin6_addr = src->ip6;
 792                 ip6->sin6_port = port;
 793                 break;
 794         default:
 795                 break;
 796         }
 797 }
 798
 799 static inline int cma_user_data_offset(enum rdma_port_space ps)
 800 {
 801         switch (ps) {
 802         case RDMA_PS_SDP:
 803                 return 0;
 804         default:
 805                 return sizeof(struct cma_hdr);
 806         }
 807 }
 808
 809 static void cma_cancel_route(struct rdma_id_private *id_priv)
 810 {
 811         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 812         case IB_LINK_LAYER_INFINIBAND:
 813                 if (id_priv->query)
 814                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 815                 break;
 816         default:
 817                 break;
 818         }
 819 }
 820
 821 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 822 {
 823         struct rdma_id_private *dev_id_priv;
 824
 825         /*
 826          * Remove from listen_any_list to prevent added devices from spawning
 827          * additional listen requests.
 828          */
 829         mutex_lock(&lock);
 830         list_del(&id_priv->list);
 831
 832         while (!list_empty(&id_priv->listen_list)) {
 833                 dev_id_priv = list_entry(id_priv->listen_list.next,
 834                                          struct rdma_id_private, listen_list);
 835                 /* sync with device removal to avoid duplicate destruction */
 836                 list_del_init(&dev_id_priv->list);
 837                 list_del(&dev_id_priv->listen_list);
 838                 mutex_unlock(&lock);
 839
 840                 rdma_destroy_id(&dev_id_priv->id);
 841                 mutex_lock(&lock);
 842         }
 843         mutex_unlock(&lock);
 844 }
 845
 846 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 847                                  enum rdma_cm_state state)
 848 {
 849         switch (state) {
 850         case RDMA_CM_ADDR_QUERY:
 851                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 852                 break;
 853         case RDMA_CM_ROUTE_QUERY:
 854                 cma_cancel_route(id_priv);
 855                 break;
 856         case RDMA_CM_LISTEN:
 857                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 858                                 && !id_priv->cma_dev)
 859                         cma_cancel_listens(id_priv);
 860                 break;
 861         default:
 862                 break;
 863         }
 864 }
 865
 866 static void cma_release_port(struct rdma_id_private *id_priv)
 867 {
 868         struct rdma_bind_list *bind_list = id_priv->bind_list;
 869
 870         if (!bind_list)
 871                 return;
 872
 873         mutex_lock(&lock);
 874         hlist_del(&id_priv->node);
 875         if (hlist_empty(&bind_list->owners)) {
 876                 idr_remove(bind_list->ps, bind_list->port);
 877                 kfree(bind_list);
 878         }
 879         mutex_unlock(&lock);
 880 }
 881
 882 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 883 {
 884         struct cma_multicast *mc;
 885
 886         while (!list_empty(&id_priv->mc_list)) {
 887                 mc = container_of(id_priv->mc_list.next,
 888                                   struct cma_multicast, list);
 889                 list_del(&mc->list);
 890                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 891                 case IB_LINK_LAYER_INFINIBAND:
 892                         ib_sa_free_multicast(mc->multicast.ib);
 893                         kfree(mc);
 894                         break;
 895                 case IB_LINK_LAYER_ETHERNET:
 896                         kref_put(&mc->mcref, release_mc);
 897                         break;
 898                 default:
 899                         break;
 900                 }
 901         }
 902 }
 903
 904 void rdma_destroy_id(struct rdma_cm_id *id)
 905 {
 906         struct rdma_id_private *id_priv;
 907         enum rdma_cm_state state;
 908
 909         id_priv = container_of(id, struct rdma_id_private, id);
 910         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 911         cma_cancel_operation(id_priv, state);
 912
 913         /*
 914          * Wait for any active callback to finish.  New callbacks will find
 915          * the id_priv state set to destroying and abort.
 916          */
 917         mutex_lock(&id_priv->handler_mutex);
 918         mutex_unlock(&id_priv->handler_mutex);
 919
 920         if (id_priv->cma_dev) {
 921                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 922                 case RDMA_TRANSPORT_IB:
 923                         if (id_priv->cm_id.ib)
 924                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 925                         break;
 926                 case RDMA_TRANSPORT_IWARP:
 927                         if (id_priv->cm_id.iw)
 928                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 929                         break;
 930                 default:
 931                         break;
 932                 }
 933                 cma_leave_mc_groups(id_priv);
 934                 cma_release_dev(id_priv);
 935         }
 936
 937         cma_release_port(id_priv);
 938         cma_deref_id(id_priv);
 939         wait_for_completion(&id_priv->comp);
 940
 941         if (id_priv->internal_id)
 942                 cma_deref_id(id_priv->id.context);
 943
 944         kfree(id_priv->id.route.path_rec);
 945         kfree(id_priv);
 946 }
 947 EXPORT_SYMBOL(rdma_destroy_id);
 948
 949 static int cma_rep_recv(struct rdma_id_private *id_priv)
 950 {
 951         int ret;
 952
 953         ret = cma_modify_qp_rtr(id_priv, NULL);
 954         if (ret)
 955                 goto reject;
 956
 957         ret = cma_modify_qp_rts(id_priv, NULL);
 958         if (ret)
 959                 goto reject;
 960
 961         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 962         if (ret)
 963                 goto reject;
 964
 965         return 0;
 966 reject:
 967         cma_modify_qp_err(id_priv);
 968         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 969                        NULL, 0, NULL, 0);
 970         return ret;
 971 }
 972
 973 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 974 {
 975         if (id_priv->id.ps == RDMA_PS_SDP &&
 976             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 977             SDP_MAJ_VERSION)
 978                 return -EINVAL;
 979
 980         return 0;
 981 }
 982
 983 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 984                                    struct ib_cm_rep_event_param *rep_data,
 985                                    void *private_data)
 986 {
 987         event->param.conn.private_data = private_data;
 988         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 989         event->param.conn.responder_resources = rep_data->responder_resources;
 990         event->param.conn.initiator_depth = rep_data->initiator_depth;
 991         event->param.conn.flow_control = rep_data->flow_control;
 992         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 993         event->param.conn.srq = rep_data->srq;
 994         event->param.conn.qp_num = rep_data->remote_qpn;
 995 }
 996
 997 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 998 {
 999         struct rdma_id_private *id_priv = cm_id->context;
1000         struct rdma_cm_event event;
1001         int ret = 0;
1002
1003         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1004                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1005             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1006                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1007                 return 0;
1008
1009         memset(&event, 0, sizeof event);
1010         switch (ib_event->event) {
1011         case IB_CM_REQ_ERROR:
1012         case IB_CM_REP_ERROR:
1013                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1014                 event.status = -ETIMEDOUT;
1015                 break;
1016         case IB_CM_REP_RECEIVED:
1017                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1018                 if (event.status)
1019                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1020                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1021                         event.status = cma_rep_recv(id_priv);
1022                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1023                                                      RDMA_CM_EVENT_ESTABLISHED;
1024                 } else
1025                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1026                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1027                                        ib_event->private_data);
1028                 break;
1029         case IB_CM_RTU_RECEIVED:
1030         case IB_CM_USER_ESTABLISHED:
1031                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1032                 break;
1033         case IB_CM_DREQ_ERROR:
1034                 event.status = -ETIMEDOUT; /* fall through */
1035         case IB_CM_DREQ_RECEIVED:
1036         case IB_CM_DREP_RECEIVED:
1037                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1038                                    RDMA_CM_DISCONNECT))
1039                         goto out;
1040                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1041                 break;
1042         case IB_CM_TIMEWAIT_EXIT:
1043                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1044                 break;
1045         case IB_CM_MRA_RECEIVED:
1046                 /* ignore event */
1047                 goto out;
1048         case IB_CM_REJ_RECEIVED:
1049                 cma_modify_qp_err(id_priv);
1050                 event.status = ib_event->param.rej_rcvd.reason;
1051                 event.event = RDMA_CM_EVENT_REJECTED;
1052                 event.param.conn.private_data = ib_event->private_data;
1053                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1054                 break;
1055         default:
1056                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1057                        ib_event->event);
1058                 goto out;
1059         }
1060
1061         ret = id_priv->id.event_handler(&id_priv->id, &event);
1062         if (ret) {
1063                 /* Destroy the CM ID by returning a non-zero value. */
1064                 id_priv->cm_id.ib = NULL;
1065                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1066                 mutex_unlock(&id_priv->handler_mutex);
1067                 rdma_destroy_id(&id_priv->id);
1068                 return ret;
1069         }
1070 out:
1071         mutex_unlock(&id_priv->handler_mutex);
1072         return ret;
1073 }
1074
1075 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1076                                                struct ib_cm_event *ib_event)
1077 {
1078         struct rdma_id_private *id_priv;
1079         struct rdma_cm_id *id;
1080         struct rdma_route *rt;
1081         union cma_ip_addr *src, *dst;
1082         __be16 port;
1083         u8 ip_ver;
1084         int ret;
1085
1086         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1087                              &ip_ver, &port, &src, &dst))
1088                 return NULL;
1089
1090         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1091                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1092         if (IS_ERR(id))
1093                 return NULL;
1094
1095         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1096                           ip_ver, port, src, dst);
1097
1098         rt = &id->route;
1099         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1100         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1101                                GFP_KERNEL);
1102         if (!rt->path_rec)
1103                 goto err;
1104
1105         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1106         if (rt->num_paths == 2)
1107                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1108
1109         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1110                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1111                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1112                 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1113         } else {
1114                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1115                                         &rt->addr.dev_addr);
1116                 if (ret)
1117                         goto err;
1118         }
1119         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1120
1121         id_priv = container_of(id, struct rdma_id_private, id);
1122         id_priv->state = RDMA_CM_CONNECT;
1123         return id_priv;
1124
1125 err:
1126         rdma_destroy_id(id);
1127         return NULL;
1128 }
1129
1130 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1131                                               struct ib_cm_event *ib_event)
1132 {
1133         struct rdma_id_private *id_priv;
1134         struct rdma_cm_id *id;
1135         union cma_ip_addr *src, *dst;
1136         __be16 port;
1137         u8 ip_ver;
1138         int ret;
1139
1140         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1141                             listen_id->ps, IB_QPT_UD);
1142         if (IS_ERR(id))
1143                 return NULL;
1144
1145
1146         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1147                              &ip_ver, &port, &src, &dst))
1148                 goto err;
1149
1150         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1151                           ip_ver, port, src, dst);
1152
1153         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1154                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1155                                         &id->route.addr.dev_addr);
1156                 if (ret)
1157                         goto err;
1158         }
1159
1160         id_priv = container_of(id, struct rdma_id_private, id);
1161         id_priv->state = RDMA_CM_CONNECT;
1162         return id_priv;
1163 err:
1164         rdma_destroy_id(id);
1165         return NULL;
1166 }
1167
1168 static void cma_set_req_event_data(struct rdma_cm_event *event,
1169                                    struct ib_cm_req_event_param *req_data,
1170                                    void *private_data, int offset)
1171 {
1172         event->param.conn.private_data = private_data + offset;
1173         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1174         event->param.conn.responder_resources = req_data->responder_resources;
1175         event->param.conn.initiator_depth = req_data->initiator_depth;
1176         event->param.conn.flow_control = req_data->flow_control;
1177         event->param.conn.retry_count = req_data->retry_count;
1178         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1179         event->param.conn.srq = req_data->srq;
1180         event->param.conn.qp_num = req_data->remote_qpn;
1181 }
1182
1183 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1184 {
1185         return (((ib_event->event == IB_CM_REQ_RECEIVED) ||
1186                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1187                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1188                  (id->qp_type == IB_QPT_UD)) ||
1189                 (!id->qp_type));
1190 }
1191
1192 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1193 {
1194         struct rdma_id_private *listen_id, *conn_id;
1195         struct rdma_cm_event event;
1196         int offset, ret;
1197
1198         listen_id = cm_id->context;
1199         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1200                 return -EINVAL;
1201
1202         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1203                 return -ECONNABORTED;
1204
1205         memset(&event, 0, sizeof event);
1206         offset = cma_user_data_offset(listen_id->id.ps);
1207         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1208         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1209                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1210                 event.param.ud.private_data = ib_event->private_data + offset;
1211                 event.param.ud.private_data_len =
1212                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1213         } else {
1214                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1215                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1216                                        ib_event->private_data, offset);
1217         }
1218         if (!conn_id) {
1219                 ret = -ENOMEM;
1220                 goto out;
1221         }
1222
1223         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1224         ret = cma_acquire_dev(conn_id);
1225         if (ret)
1226                 goto release_conn_id;
1227
1228         conn_id->cm_id.ib = cm_id;
1229         cm_id->context = conn_id;
1230         cm_id->cm_handler = cma_ib_handler;
1231
1232         /*
1233          * Protect against the user destroying conn_id from another thread
1234          * until we're done accessing it.
1235          */
1236         atomic_inc(&conn_id->refcount);
1237         ret = conn_id->id.event_handler(&conn_id->id, &event);
1238         if (!ret) {
1239                 /*
1240                  * Acquire mutex to prevent user executing rdma_destroy_id()
1241                  * while we're accessing the cm_id.
1242                  */
1243                 mutex_lock(&lock);
1244                 if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1245                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1246                 mutex_unlock(&lock);
1247                 mutex_unlock(&conn_id->handler_mutex);
1248                 cma_deref_id(conn_id);
1249                 goto out;
1250         }
1251         cma_deref_id(conn_id);
1252
1253         /* Destroy the CM ID by returning a non-zero value. */
1254         conn_id->cm_id.ib = NULL;
1255
1256 release_conn_id:
1257         cma_exch(conn_id, RDMA_CM_DESTROYING);
1258         mutex_unlock(&conn_id->handler_mutex);
1259         rdma_destroy_id(&conn_id->id);
1260
1261 out:
1262         mutex_unlock(&listen_id->handler_mutex);
1263         return ret;
1264 }
1265
1266 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1267 {
1268         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1269 }
1270
1271 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1272                                  struct ib_cm_compare_data *compare)
1273 {
1274         struct cma_hdr *cma_data, *cma_mask;
1275         struct sdp_hh *sdp_data, *sdp_mask;
1276         __be32 ip4_addr;
1277         struct in6_addr ip6_addr;
1278
1279         memset(compare, 0, sizeof *compare);
1280         cma_data = (void *) compare->data;
1281         cma_mask = (void *) compare->mask;
1282         sdp_data = (void *) compare->data;
1283         sdp_mask = (void *) compare->mask;
1284
1285         switch (addr->sa_family) {
1286         case AF_INET:
1287                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1288                 if (ps == RDMA_PS_SDP) {
1289                         sdp_set_ip_ver(sdp_data, 4);
1290                         sdp_set_ip_ver(sdp_mask, 0xF);
1291                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1292                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1293                 } else {
1294                         cma_set_ip_ver(cma_data, 4);
1295                         cma_set_ip_ver(cma_mask, 0xF);
1296                         cma_data->dst_addr.ip4.addr = ip4_addr;
1297                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1298                 }
1299                 break;
1300         case AF_INET6:
1301                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1302                 if (ps == RDMA_PS_SDP) {
1303                         sdp_set_ip_ver(sdp_data, 6);
1304                         sdp_set_ip_ver(sdp_mask, 0xF);
1305                         sdp_data->dst_addr.ip6 = ip6_addr;
1306                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1307                                sizeof sdp_mask->dst_addr.ip6);
1308                 } else {
1309                         cma_set_ip_ver(cma_data, 6);
1310                         cma_set_ip_ver(cma_mask, 0xF);
1311                         cma_data->dst_addr.ip6 = ip6_addr;
1312                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1313                                sizeof cma_mask->dst_addr.ip6);
1314                 }
1315                 break;
1316         default:
1317                 break;
1318         }
1319 }
1320
1321 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1322 {
1323         struct rdma_id_private *id_priv = iw_id->context;
1324         struct rdma_cm_event event;
1325         struct sockaddr_in *sin;
1326         int ret = 0;
1327
1328         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1329                 return 0;
1330
1331         memset(&event, 0, sizeof event);
1332         switch (iw_event->event) {
1333         case IW_CM_EVENT_CLOSE:
1334                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1335                 break;
1336         case IW_CM_EVENT_CONNECT_REPLY:
1337                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1338                 *sin = iw_event->local_addr;
1339                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1340                 *sin = iw_event->remote_addr;
1341                 switch (iw_event->status) {
1342                 case 0:
1343                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1344                         event.param.conn.initiator_depth = iw_event->ird;
1345                         event.param.conn.responder_resources = iw_event->ord;
1346                         break;
1347                 case -ECONNRESET:
1348                 case -ECONNREFUSED:
1349                         event.event = RDMA_CM_EVENT_REJECTED;
1350                         break;
1351                 case -ETIMEDOUT:
1352                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1353                         break;
1354                 default:
1355                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1356                         break;
1357                 }
1358                 break;
1359         case IW_CM_EVENT_ESTABLISHED:
1360                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1361                 event.param.conn.initiator_depth = iw_event->ird;
1362                 event.param.conn.responder_resources = iw_event->ord;
1363                 break;
1364         default:
1365                 BUG_ON(1);
1366         }
1367
1368         event.status = iw_event->status;
1369         event.param.conn.private_data = iw_event->private_data;
1370         event.param.conn.private_data_len = iw_event->private_data_len;
1371         ret = id_priv->id.event_handler(&id_priv->id, &event);
1372         if (ret) {
1373                 /* Destroy the CM ID by returning a non-zero value. */
1374                 id_priv->cm_id.iw = NULL;
1375                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1376                 mutex_unlock(&id_priv->handler_mutex);
1377                 rdma_destroy_id(&id_priv->id);
1378                 return ret;
1379         }
1380
1381         mutex_unlock(&id_priv->handler_mutex);
1382         return ret;
1383 }
1384
1385 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1386                                struct iw_cm_event *iw_event)
1387 {
1388         struct rdma_cm_id *new_cm_id;
1389         struct rdma_id_private *listen_id, *conn_id;
1390         struct sockaddr_in *sin;
1391         struct net_device *dev = NULL;
1392         struct rdma_cm_event event;
1393         int ret;
1394         struct ib_device_attr attr;
1395
1396         listen_id = cm_id->context;
1397         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1398                 return -ECONNABORTED;
1399
1400         /* Create a new RDMA id for the new IW CM ID */
1401         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1402                                    listen_id->id.context,
1403                                    RDMA_PS_TCP, IB_QPT_RC);
1404         if (IS_ERR(new_cm_id)) {
1405                 ret = -ENOMEM;
1406                 goto out;
1407         }
1408         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1409         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1410         conn_id->state = RDMA_CM_CONNECT;
1411
1412         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1413         if (!dev) {
1414                 ret = -EADDRNOTAVAIL;
1415                 mutex_unlock(&conn_id->handler_mutex);
1416                 rdma_destroy_id(new_cm_id);
1417                 goto out;
1418         }
1419         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1420         if (ret) {
1421                 mutex_unlock(&conn_id->handler_mutex);
1422                 rdma_destroy_id(new_cm_id);
1423                 goto out;
1424         }
1425
1426         ret = cma_acquire_dev(conn_id);
1427         if (ret) {
1428                 mutex_unlock(&conn_id->handler_mutex);
1429                 rdma_destroy_id(new_cm_id);
1430                 goto out;
1431         }
1432
1433         conn_id->cm_id.iw = cm_id;
1434         cm_id->context = conn_id;
1435         cm_id->cm_handler = cma_iw_handler;
1436
1437         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1438         *sin = iw_event->local_addr;
1439         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1440         *sin = iw_event->remote_addr;
1441
1442         ret = ib_query_device(conn_id->id.device, &attr);
1443         if (ret) {
1444                 mutex_unlock(&conn_id->handler_mutex);
1445                 rdma_destroy_id(new_cm_id);
1446                 goto out;
1447         }
1448
1449         memset(&event, 0, sizeof event);
1450         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1451         event.param.conn.private_data = iw_event->private_data;
1452         event.param.conn.private_data_len = iw_event->private_data_len;
1453         event.param.conn.initiator_depth = iw_event->ird;
1454         event.param.conn.responder_resources = iw_event->ord;
1455
1456         /*
1457          * Protect against the user destroying conn_id from another thread
1458          * until we're done accessing it.
1459          */
1460         atomic_inc(&conn_id->refcount);
1461         ret = conn_id->id.event_handler(&conn_id->id, &event);
1462         if (ret) {
1463                 /* User wants to destroy the CM ID */
1464                 conn_id->cm_id.iw = NULL;
1465                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1466                 mutex_unlock(&conn_id->handler_mutex);
1467                 cma_deref_id(conn_id);
1468                 rdma_destroy_id(&conn_id->id);
1469                 goto out;
1470         }
1471
1472         mutex_unlock(&conn_id->handler_mutex);
1473         cma_deref_id(conn_id);
1474
1475 out:
1476         if (dev)
1477                 dev_put(dev);
1478         mutex_unlock(&listen_id->handler_mutex);
1479         return ret;
1480 }
1481
1482 static int cma_ib_listen(struct rdma_id_private *id_priv)
1483 {
1484         struct ib_cm_compare_data compare_data;
1485         struct sockaddr *addr;
1486         struct ib_cm_id *id;
1487         __be64 svc_id;
1488         int ret;
1489
1490         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1491         if (IS_ERR(id))
1492                 return PTR_ERR(id);
1493
1494         id_priv->cm_id.ib = id;
1495
1496         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1497         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1498         if (cma_any_addr(addr))
1499                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1500         else {
1501                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1502                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1503         }
1504
1505         if (ret) {
1506                 ib_destroy_cm_id(id_priv->cm_id.ib);
1507                 id_priv->cm_id.ib = NULL;
1508         }
1509
1510         return ret;
1511 }
1512
1513 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1514 {
1515         int ret;
1516         struct sockaddr_in *sin;
1517         struct iw_cm_id *id;
1518
1519         id = iw_create_cm_id(id_priv->id.device,
1520                              iw_conn_req_handler,
1521                              id_priv);
1522         if (IS_ERR(id))
1523                 return PTR_ERR(id);
1524
1525         id_priv->cm_id.iw = id;
1526
1527         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1528         id_priv->cm_id.iw->local_addr = *sin;
1529
1530         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1531
1532         if (ret) {
1533                 iw_destroy_cm_id(id_priv->cm_id.iw);
1534                 id_priv->cm_id.iw = NULL;
1535         }
1536
1537         return ret;
1538 }
1539
1540 static int cma_listen_handler(struct rdma_cm_id *id,
1541                               struct rdma_cm_event *event)
1542 {
1543         struct rdma_id_private *id_priv = id->context;
1544
1545         id->context = id_priv->id.context;
1546         id->event_handler = id_priv->id.event_handler;
1547         return id_priv->id.event_handler(id, event);
1548 }
1549
1550 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1551                               struct cma_device *cma_dev)
1552 {
1553         struct rdma_id_private *dev_id_priv;
1554         struct rdma_cm_id *id;
1555         int ret;
1556
1557         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1558                             id_priv->id.qp_type);
1559         if (IS_ERR(id))
1560                 return;
1561
1562         dev_id_priv = container_of(id, struct rdma_id_private, id);
1563
1564         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1565         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1566                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1567
1568         cma_attach_to_dev(dev_id_priv, cma_dev);
1569         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1570         atomic_inc(&id_priv->refcount);
1571         dev_id_priv->internal_id = 1;
1572
1573         ret = rdma_listen(id, id_priv->backlog);
1574         if (ret)
1575                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1576                        "listening on device %s\n", ret, cma_dev->device->name);
1577 }
1578
1579 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1580 {
1581         struct cma_device *cma_dev;
1582
1583         mutex_lock(&lock);
1584         list_add_tail(&id_priv->list, &listen_any_list);
1585         list_for_each_entry(cma_dev, &dev_list, list)
1586                 cma_listen_on_dev(id_priv, cma_dev);
1587         mutex_unlock(&lock);
1588 }
1589
1590 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1591 {
1592         struct rdma_id_private *id_priv;
1593
1594         id_priv = container_of(id, struct rdma_id_private, id);
1595         id_priv->tos = (u8) tos;
1596 }
1597 EXPORT_SYMBOL(rdma_set_service_type);
1598
1599 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1600                               void *context)
1601 {
1602         struct cma_work *work = context;
1603         struct rdma_route *route;
1604
1605         route = &work->id->id.route;
1606
1607         if (!status) {
1608                 route->num_paths = 1;
1609                 *route->path_rec = *path_rec;
1610         } else {
1611                 work->old_state = RDMA_CM_ROUTE_QUERY;
1612                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1613                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1614                 work->event.status = status;
1615         }
1616
1617         queue_work(cma_wq, &work->work);
1618 }
1619
1620 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1621                               struct cma_work *work)
1622 {
1623         struct rdma_addr *addr = &id_priv->id.route.addr;
1624         struct ib_sa_path_rec path_rec;
1625         ib_sa_comp_mask comp_mask;
1626         struct sockaddr_in6 *sin6;
1627
1628         memset(&path_rec, 0, sizeof path_rec);
1629         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1630         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1631         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1632         path_rec.numb_path = 1;
1633         path_rec.reversible = 1;
1634         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1635                                                         (struct sockaddr *) &addr->dst_addr);
1636
1637         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1638                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1639                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1640
1641         if (addr->src_addr.ss_family == AF_INET) {
1642                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1643                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1644         } else {
1645                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1646                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1647                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1648         }
1649
1650         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1651                                                id_priv->id.port_num, &path_rec,
1652                                                comp_mask, timeout_ms,
1653                                                GFP_KERNEL, cma_query_handler,
1654                                                work, &id_priv->query);
1655
1656         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1657 }
1658
1659 static void cma_work_handler(struct work_struct *_work)
1660 {
1661         struct cma_work *work = container_of(_work, struct cma_work, work);
1662         struct rdma_id_private *id_priv = work->id;
1663         int destroy = 0;
1664
1665         mutex_lock(&id_priv->handler_mutex);
1666         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1667                 goto out;
1668
1669         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1670                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1671                 destroy = 1;
1672         }
1673 out:
1674         mutex_unlock(&id_priv->handler_mutex);
1675         cma_deref_id(id_priv);
1676         if (destroy)
1677                 rdma_destroy_id(&id_priv->id);
1678         kfree(work);
1679 }
1680
1681 static void cma_ndev_work_handler(struct work_struct *_work)
1682 {
1683         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1684         struct rdma_id_private *id_priv = work->id;
1685         int destroy = 0;
1686
1687         mutex_lock(&id_priv->handler_mutex);
1688         if (id_priv->state == RDMA_CM_DESTROYING ||
1689             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1690                 goto out;
1691
1692         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1693                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1694                 destroy = 1;
1695         }
1696
1697 out:
1698         mutex_unlock(&id_priv->handler_mutex);
1699         cma_deref_id(id_priv);
1700         if (destroy)
1701                 rdma_destroy_id(&id_priv->id);
1702         kfree(work);
1703 }
1704
1705 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1706 {
1707         struct rdma_route *route = &id_priv->id.route;
1708         struct cma_work *work;
1709         int ret;
1710
1711         work = kzalloc(sizeof *work, GFP_KERNEL);
1712         if (!work)
1713                 return -ENOMEM;
1714
1715         work->id = id_priv;
1716         INIT_WORK(&work->work, cma_work_handler);
1717         work->old_state = RDMA_CM_ROUTE_QUERY;
1718         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1719         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1720
1721         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1722         if (!route->path_rec) {
1723                 ret = -ENOMEM;
1724                 goto err1;
1725         }
1726
1727         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1728         if (ret)
1729                 goto err2;
1730
1731         return 0;
1732 err2:
1733         kfree(route->path_rec);
1734         route->path_rec = NULL;
1735 err1:
1736         kfree(work);
1737         return ret;
1738 }
1739
1740 int rdma_set_ib_paths(struct rdma_cm_id *id,
1741                       struct ib_sa_path_rec *path_rec, int num_paths)
1742 {
1743         struct rdma_id_private *id_priv;
1744         int ret;
1745
1746         id_priv = container_of(id, struct rdma_id_private, id);
1747         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1748                            RDMA_CM_ROUTE_RESOLVED))
1749                 return -EINVAL;
1750
1751         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1752                                      GFP_KERNEL);
1753         if (!id->route.path_rec) {
1754                 ret = -ENOMEM;
1755                 goto err;
1756         }
1757
1758         id->route.num_paths = num_paths;
1759         return 0;
1760 err:
1761         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1762         return ret;
1763 }
1764 EXPORT_SYMBOL(rdma_set_ib_paths);
1765
1766 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1767 {
1768         struct cma_work *work;
1769
1770         work = kzalloc(sizeof *work, GFP_KERNEL);
1771         if (!work)
1772                 return -ENOMEM;
1773
1774         work->id = id_priv;
1775         INIT_WORK(&work->work, cma_work_handler);
1776         work->old_state = RDMA_CM_ROUTE_QUERY;
1777         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1778         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1779         queue_work(cma_wq, &work->work);
1780         return 0;
1781 }
1782
1783 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1784 {
1785         struct rdma_route *route = &id_priv->id.route;
1786         struct rdma_addr *addr = &route->addr;
1787         struct cma_work *work;
1788         int ret;
1789         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1790         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1791         struct net_device *ndev = NULL;
1792         u16 vid;
1793
1794         if (src_addr->sin_family != dst_addr->sin_family)
1795                 return -EINVAL;
1796
1797         work = kzalloc(sizeof *work, GFP_KERNEL);
1798         if (!work)
1799                 return -ENOMEM;
1800
1801         work->id = id_priv;
1802         INIT_WORK(&work->work, cma_work_handler);
1803
1804         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1805         if (!route->path_rec) {
1806                 ret = -ENOMEM;
1807                 goto err1;
1808         }
1809
1810         route->num_paths = 1;
1811
1812         if (addr->dev_addr.bound_dev_if)
1813                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1814         if (!ndev) {
1815                 ret = -ENODEV;
1816                 goto err2;
1817         }
1818
1819         vid = rdma_vlan_dev_vlan_id(ndev);
1820
1821         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1822         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1823
1824         route->path_rec->hop_limit = 1;
1825         route->path_rec->reversible = 1;
1826         route->path_rec->pkey = cpu_to_be16(0xffff);
1827         route->path_rec->mtu_selector = IB_SA_EQ;
1828         route->path_rec->sl = id_priv->tos >> 5;
1829
1830         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1831         route->path_rec->rate_selector = IB_SA_EQ;
1832         route->path_rec->rate = iboe_get_rate(ndev);
1833         dev_put(ndev);
1834         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1835         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1836         if (!route->path_rec->mtu) {
1837                 ret = -EINVAL;
1838                 goto err2;
1839         }
1840
1841         work->old_state = RDMA_CM_ROUTE_QUERY;
1842         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1843         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1844         work->event.status = 0;
1845
1846         queue_work(cma_wq, &work->work);
1847
1848         return 0;
1849
1850 err2:
1851         kfree(route->path_rec);
1852         route->path_rec = NULL;
1853 err1:
1854         kfree(work);
1855         return ret;
1856 }
1857
1858 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1859 {
1860         struct rdma_id_private *id_priv;
1861         int ret;
1862
1863         id_priv = container_of(id, struct rdma_id_private, id);
1864         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1865                 return -EINVAL;
1866
1867         atomic_inc(&id_priv->refcount);
1868         switch (rdma_node_get_transport(id->device->node_type)) {
1869         case RDMA_TRANSPORT_IB:
1870                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1871                 case IB_LINK_LAYER_INFINIBAND:
1872                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1873                         break;
1874                 case IB_LINK_LAYER_ETHERNET:
1875                         ret = cma_resolve_iboe_route(id_priv);
1876                         break;
1877                 default:
1878                         ret = -ENOSYS;
1879                 }
1880                 break;
1881         case RDMA_TRANSPORT_IWARP:
1882                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1883                 break;
1884         default:
1885                 ret = -ENOSYS;
1886                 break;
1887         }
1888         if (ret)
1889                 goto err;
1890
1891         return 0;
1892 err:
1893         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1894         cma_deref_id(id_priv);
1895         return ret;
1896 }
1897 EXPORT_SYMBOL(rdma_resolve_route);
1898
1899 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1900 {
1901         struct cma_device *cma_dev;
1902         struct ib_port_attr port_attr;
1903         union ib_gid gid;
1904         u16 pkey;
1905         int ret;
1906         u8 p;
1907
1908         mutex_lock(&lock);
1909         if (list_empty(&dev_list)) {
1910                 ret = -ENODEV;
1911                 goto out;
1912         }
1913         list_for_each_entry(cma_dev, &dev_list, list)
1914                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1915                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1916                             port_attr.state == IB_PORT_ACTIVE)
1917                                 goto port_found;
1918
1919         p = 1;
1920         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1921
1922 port_found:
1923         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1924         if (ret)
1925                 goto out;
1926
1927         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1928         if (ret)
1929                 goto out;
1930
1931         id_priv->id.route.addr.dev_addr.dev_type =
1932                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1933                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1934
1935         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1936         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1937         id_priv->id.port_num = p;
1938         cma_attach_to_dev(id_priv, cma_dev);
1939 out:
1940         mutex_unlock(&lock);
1941         return ret;
1942 }
1943
1944 static void addr_handler(int status, struct sockaddr *src_addr,
1945                          struct rdma_dev_addr *dev_addr, void *context)
1946 {
1947         struct rdma_id_private *id_priv = context;
1948         struct rdma_cm_event event;
1949
1950         memset(&event, 0, sizeof event);
1951         mutex_lock(&id_priv->handler_mutex);
1952         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1953                            RDMA_CM_ADDR_RESOLVED))
1954                 goto out;
1955
1956         if (!status && !id_priv->cma_dev)
1957                 status = cma_acquire_dev(id_priv);
1958
1959         if (status) {
1960                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1961                                    RDMA_CM_ADDR_BOUND))
1962                         goto out;
1963                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1964                 event.status = status;
1965         } else {
1966                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1967                        ip_addr_size(src_addr));
1968                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1969         }
1970
1971         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1972                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1973                 mutex_unlock(&id_priv->handler_mutex);
1974                 cma_deref_id(id_priv);
1975                 rdma_destroy_id(&id_priv->id);
1976                 return;
1977         }
1978 out:
1979         mutex_unlock(&id_priv->handler_mutex);
1980         cma_deref_id(id_priv);
1981 }
1982
1983 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1984 {
1985         struct cma_work *work;
1986         struct sockaddr *src, *dst;
1987         union ib_gid gid;
1988         int ret;
1989
1990         work = kzalloc(sizeof *work, GFP_KERNEL);
1991         if (!work)
1992                 return -ENOMEM;
1993
1994         if (!id_priv->cma_dev) {
1995                 ret = cma_bind_loopback(id_priv);
1996                 if (ret)
1997                         goto err;
1998         }
1999
2000         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2001         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2002
2003         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2004         if (cma_zero_addr(src)) {
2005                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2006                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2007                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
2008                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
2009                 } else {
2010                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
2011                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
2012                 }
2013         }
2014
2015         work->id = id_priv;
2016         INIT_WORK(&work->work, cma_work_handler);
2017         work->old_state = RDMA_CM_ADDR_QUERY;
2018         work->new_state = RDMA_CM_ADDR_RESOLVED;
2019         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2020         queue_work(cma_wq, &work->work);
2021         return 0;
2022 err:
2023         kfree(work);
2024         return ret;
2025 }
2026
2027 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2028                          struct sockaddr *dst_addr)
2029 {
2030         if (!src_addr || !src_addr->sa_family) {
2031                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2032                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2033                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2034                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2035                 }
2036         }
2037         return rdma_bind_addr(id, src_addr);
2038 }
2039
2040 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2041                       struct sockaddr *dst_addr, int timeout_ms)
2042 {
2043         struct rdma_id_private *id_priv;
2044         int ret;
2045
2046         id_priv = container_of(id, struct rdma_id_private, id);
2047         if (id_priv->state == RDMA_CM_IDLE) {
2048                 ret = cma_bind_addr(id, src_addr, dst_addr);
2049                 if (ret)
2050                         return ret;
2051         }
2052
2053         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2054                 return -EINVAL;
2055
2056         atomic_inc(&id_priv->refcount);
2057         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2058         if (cma_any_addr(dst_addr))
2059                 ret = cma_resolve_loopback(id_priv);
2060         else
2061                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2062                                       dst_addr, &id->route.addr.dev_addr,
2063                                       timeout_ms, addr_handler, id_priv);
2064         if (ret)
2065                 goto err;
2066
2067         return 0;
2068 err:
2069         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2070         cma_deref_id(id_priv);
2071         return ret;
2072 }
2073 EXPORT_SYMBOL(rdma_resolve_addr);
2074
2075 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2076 {
2077         struct rdma_id_private *id_priv;
2078         unsigned long flags;
2079         int ret;
2080
2081         id_priv = container_of(id, struct rdma_id_private, id);
2082         spin_lock_irqsave(&id_priv->lock, flags);
2083         if (id_priv->state == RDMA_CM_IDLE) {
2084                 id_priv->reuseaddr = reuse;
2085                 ret = 0;
2086         } else {
2087                 ret = -EINVAL;
2088         }
2089         spin_unlock_irqrestore(&id_priv->lock, flags);
2090         return ret;
2091 }
2092 EXPORT_SYMBOL(rdma_set_reuseaddr);
2093
2094 static void cma_bind_port(struct rdma_bind_list *bind_list,
2095                           struct rdma_id_private *id_priv)
2096 {
2097         struct sockaddr_in *sin;
2098
2099         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2100         sin->sin_port = htons(bind_list->port);
2101         id_priv->bind_list = bind_list;
2102         hlist_add_head(&id_priv->node, &bind_list->owners);
2103 }
2104
2105 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2106                           unsigned short snum)
2107 {
2108         struct rdma_bind_list *bind_list;
2109         int port, ret;
2110
2111         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2112         if (!bind_list)
2113                 return -ENOMEM;
2114
2115         do {
2116                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2117         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2118
2119         if (ret)
2120                 goto err1;
2121
2122         if (port != snum) {
2123                 ret = -EADDRNOTAVAIL;
2124                 goto err2;
2125         }
2126
2127         bind_list->ps = ps;
2128         bind_list->port = (unsigned short) port;
2129         cma_bind_port(bind_list, id_priv);
2130         return 0;
2131 err2:
2132         idr_remove(ps, port);
2133 err1:
2134         kfree(bind_list);
2135         return ret;
2136 }
2137
2138 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2139 {
2140         static unsigned int last_used_port;
2141         int low, high, remaining;
2142         unsigned int rover;
2143
2144         inet_get_local_port_range(&low, &high);
2145         remaining = (high - low) + 1;
2146         rover = net_random() % remaining + low;
2147 retry:
2148         if (last_used_port != rover &&
2149             !idr_find(ps, (unsigned short) rover)) {
2150                 int ret = cma_alloc_port(ps, id_priv, rover);
2151                 /*
2152                  * Remember previously used port number in order to avoid
2153                  * re-using same port immediately after it is closed.
2154                  */
2155                 if (!ret)
2156                         last_used_port = rover;
2157                 if (ret != -EADDRNOTAVAIL)
2158                         return ret;
2159         }
2160         if (--remaining) {
2161                 rover++;
2162                 if ((rover < low) || (rover > high))
2163                         rover = low;
2164                 goto retry;
2165         }
2166         return -EADDRNOTAVAIL;
2167 }
2168
2169 /*
2170  * Check that the requested port is available.  This is called when trying to
2171  * bind to a specific port, or when trying to listen on a bound port.  In
2172  * the latter case, the provided id_priv may already be on the bind_list, but
2173  * we still need to check that it's okay to start listening.
2174  */
2175 static int cma_check_port(struct rdma_bind_list *bind_list,
2176                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2177 {
2178         struct rdma_id_private *cur_id;
2179         struct sockaddr *addr, *cur_addr;
2180         struct hlist_node *node;
2181
2182         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2183         if (cma_any_addr(addr) && !reuseaddr)
2184                 return -EADDRNOTAVAIL;
2185
2186         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2187                 if (id_priv == cur_id)
2188                         continue;
2189
2190                 if ((cur_id->state == RDMA_CM_LISTEN) ||
2191                     !reuseaddr || !cur_id->reuseaddr) {
2192                         cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2193                         if (cma_any_addr(cur_addr))
2194                                 return -EADDRNOTAVAIL;
2195
2196                         if (!cma_addr_cmp(addr, cur_addr))
2197                                 return -EADDRINUSE;
2198                 }
2199         }
2200         return 0;
2201 }
2202
2203 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2204 {
2205         struct rdma_bind_list *bind_list;
2206         unsigned short snum;
2207         int ret;
2208
2209         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2210         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2211                 return -EACCES;
2212
2213         bind_list = idr_find(ps, snum);
2214         if (!bind_list) {
2215                 ret = cma_alloc_port(ps, id_priv, snum);
2216         } else {
2217                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2218                 if (!ret)
2219                         cma_bind_port(bind_list, id_priv);
2220         }
2221         return ret;
2222 }
2223
2224 static int cma_bind_listen(struct rdma_id_private *id_priv)
2225 {
2226         struct rdma_bind_list *bind_list = id_priv->bind_list;
2227         int ret = 0;
2228
2229         mutex_lock(&lock);
2230         if (bind_list->owners.first->next)
2231                 ret = cma_check_port(bind_list, id_priv, 0);
2232         mutex_unlock(&lock);
2233         return ret;
2234 }
2235
2236 static int cma_get_port(struct rdma_id_private *id_priv)
2237 {
2238         struct idr *ps;
2239         int ret;
2240
2241         switch (id_priv->id.ps) {
2242         case RDMA_PS_SDP:
2243                 ps = &sdp_ps;
2244                 break;
2245         case RDMA_PS_TCP:
2246                 ps = &tcp_ps;
2247                 break;
2248         case RDMA_PS_UDP:
2249                 ps = &udp_ps;
2250                 break;
2251         case RDMA_PS_IPOIB:
2252                 ps = &ipoib_ps;
2253                 break;
2254         case RDMA_PS_IB:
2255                 ps = &ib_ps;
2256                 break;
2257         default:
2258                 return -EPROTONOSUPPORT;
2259         }
2260
2261         mutex_lock(&lock);
2262         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2263                 ret = cma_alloc_any_port(ps, id_priv);
2264         else
2265                 ret = cma_use_port(ps, id_priv);
2266         mutex_unlock(&lock);
2267
2268         return ret;
2269 }
2270
2271 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2272                                struct sockaddr *addr)
2273 {
2274 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2275         struct sockaddr_in6 *sin6;
2276
2277         if (addr->sa_family != AF_INET6)
2278                 return 0;
2279
2280         sin6 = (struct sockaddr_in6 *) addr;
2281         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2282             !sin6->sin6_scope_id)
2283                         return -EINVAL;
2284
2285         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2286 #endif
2287         return 0;
2288 }
2289
2290 int rdma_listen(struct rdma_cm_id *id, int backlog)
2291 {
2292         struct rdma_id_private *id_priv;
2293         int ret;
2294
2295         id_priv = container_of(id, struct rdma_id_private, id);
2296         if (id_priv->state == RDMA_CM_IDLE) {
2297                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2298                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2299                 if (ret)
2300                         return ret;
2301         }
2302
2303         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2304                 return -EINVAL;
2305
2306         if (id_priv->reuseaddr) {
2307                 ret = cma_bind_listen(id_priv);
2308                 if (ret)
2309                         goto err;
2310         }
2311
2312         id_priv->backlog = backlog;
2313         if (id->device) {
2314                 switch (rdma_node_get_transport(id->device->node_type)) {
2315                 case RDMA_TRANSPORT_IB:
2316                         ret = cma_ib_listen(id_priv);
2317                         if (ret)
2318                                 goto err;
2319                         break;
2320                 case RDMA_TRANSPORT_IWARP:
2321                         ret = cma_iw_listen(id_priv, backlog);
2322                         if (ret)
2323                                 goto err;
2324                         break;
2325                 default:
2326                         ret = -ENOSYS;
2327                         goto err;
2328                 }
2329         } else
2330                 cma_listen_on_all(id_priv);
2331
2332         return 0;
2333 err:
2334         id_priv->backlog = 0;
2335         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2336         return ret;
2337 }
2338 EXPORT_SYMBOL(rdma_listen);
2339
2340 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2341 {
2342         struct rdma_id_private *id_priv;
2343         int ret;
2344
2345         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2346                 return -EAFNOSUPPORT;
2347
2348         id_priv = container_of(id, struct rdma_id_private, id);
2349         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2350                 return -EINVAL;
2351
2352         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2353         if (ret)
2354                 goto err1;
2355
2356         if (!cma_any_addr(addr)) {
2357                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2358                 if (ret)
2359                         goto err1;
2360
2361                 ret = cma_acquire_dev(id_priv);
2362                 if (ret)
2363                         goto err1;
2364         }
2365
2366         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2367         ret = cma_get_port(id_priv);
2368         if (ret)
2369                 goto err2;
2370
2371         return 0;
2372 err2:
2373         if (id_priv->cma_dev)
2374                 cma_release_dev(id_priv);
2375 err1:
2376         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2377         return ret;
2378 }
2379 EXPORT_SYMBOL(rdma_bind_addr);
2380
2381 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2382                           struct rdma_route *route)
2383 {
2384         struct cma_hdr *cma_hdr;
2385         struct sdp_hh *sdp_hdr;
2386
2387         if (route->addr.src_addr.ss_family == AF_INET) {
2388                 struct sockaddr_in *src4, *dst4;
2389
2390                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2391                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2392
2393                 switch (ps) {
2394                 case RDMA_PS_SDP:
2395                         sdp_hdr = hdr;
2396                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2397                                 return -EINVAL;
2398                         sdp_set_ip_ver(sdp_hdr, 4);
2399                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2400                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2401                         sdp_hdr->port = src4->sin_port;
2402                         break;
2403                 default:
2404                         cma_hdr = hdr;
2405                         cma_hdr->cma_version = CMA_VERSION;
2406                         cma_set_ip_ver(cma_hdr, 4);
2407                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2408                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2409                         cma_hdr->port = src4->sin_port;
2410                         break;
2411                 }
2412         } else {
2413                 struct sockaddr_in6 *src6, *dst6;
2414
2415                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2416                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2417
2418                 switch (ps) {
2419                 case RDMA_PS_SDP:
2420                         sdp_hdr = hdr;
2421                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2422                                 return -EINVAL;
2423                         sdp_set_ip_ver(sdp_hdr, 6);
2424                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2425                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2426                         sdp_hdr->port = src6->sin6_port;
2427                         break;
2428                 default:
2429                         cma_hdr = hdr;
2430                         cma_hdr->cma_version = CMA_VERSION;
2431                         cma_set_ip_ver(cma_hdr, 6);
2432                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2433                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2434                         cma_hdr->port = src6->sin6_port;
2435                         break;
2436                 }
2437         }
2438         return 0;
2439 }
2440
2441 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2442                                 struct ib_cm_event *ib_event)
2443 {
2444         struct rdma_id_private *id_priv = cm_id->context;
2445         struct rdma_cm_event event;
2446         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2447         int ret = 0;
2448
2449         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2450                 return 0;
2451
2452         memset(&event, 0, sizeof event);
2453         switch (ib_event->event) {
2454         case IB_CM_SIDR_REQ_ERROR:
2455                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2456                 event.status = -ETIMEDOUT;
2457                 break;
2458         case IB_CM_SIDR_REP_RECEIVED:
2459                 event.param.ud.private_data = ib_event->private_data;
2460                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2461                 if (rep->status != IB_SIDR_SUCCESS) {
2462                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2463                         event.status = ib_event->param.sidr_rep_rcvd.status;
2464                         break;
2465                 }
2466                 ret = cma_set_qkey(id_priv);
2467                 if (ret) {
2468                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2469                         event.status = -EINVAL;
2470                         break;
2471                 }
2472                 if (id_priv->qkey != rep->qkey) {
2473                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2474                         event.status = -EINVAL;
2475                         break;
2476                 }
2477                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2478                                      id_priv->id.route.path_rec,
2479                                      &event.param.ud.ah_attr);
2480                 event.param.ud.qp_num = rep->qpn;
2481                 event.param.ud.qkey = rep->qkey;
2482                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2483                 event.status = 0;
2484                 break;
2485         default:
2486                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2487                        ib_event->event);
2488                 goto out;
2489         }
2490
2491         ret = id_priv->id.event_handler(&id_priv->id, &event);
2492         if (ret) {
2493                 /* Destroy the CM ID by returning a non-zero value. */
2494                 id_priv->cm_id.ib = NULL;
2495                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2496                 mutex_unlock(&id_priv->handler_mutex);
2497                 rdma_destroy_id(&id_priv->id);
2498                 return ret;
2499         }
2500 out:
2501         mutex_unlock(&id_priv->handler_mutex);
2502         return ret;
2503 }
2504
2505 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2506                               struct rdma_conn_param *conn_param)
2507 {
2508         struct ib_cm_sidr_req_param req;
2509         struct rdma_route *route;
2510         struct ib_cm_id *id;
2511         int ret;
2512
2513         req.private_data_len = sizeof(struct cma_hdr) +
2514                                conn_param->private_data_len;
2515         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2516         if (!req.private_data)
2517                 return -ENOMEM;
2518
2519         if (conn_param->private_data && conn_param->private_data_len)
2520                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2521                        conn_param->private_data, conn_param->private_data_len);
2522
2523         route = &id_priv->id.route;
2524         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2525         if (ret)
2526                 goto out;
2527
2528         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2529                              id_priv);
2530         if (IS_ERR(id)) {
2531                 ret = PTR_ERR(id);
2532                 goto out;
2533         }
2534         id_priv->cm_id.ib = id;
2535
2536         req.path = route->path_rec;
2537         req.service_id = cma_get_service_id(id_priv->id.ps,
2538                                             (struct sockaddr *) &route->addr.dst_addr);
2539         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2540         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2541
2542         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2543         if (ret) {
2544                 ib_destroy_cm_id(id_priv->cm_id.ib);
2545                 id_priv->cm_id.ib = NULL;
2546         }
2547 out:
2548         kfree(req.private_data);
2549         return ret;
2550 }
2551
2552 static int cma_connect_ib(struct rdma_id_private *id_priv,
2553                           struct rdma_conn_param *conn_param)
2554 {
2555         struct ib_cm_req_param req;
2556         struct rdma_route *route;
2557         void *private_data;
2558         struct ib_cm_id *id;
2559         int offset, ret;
2560
2561         memset(&req, 0, sizeof req);
2562         offset = cma_user_data_offset(id_priv->id.ps);
2563         req.private_data_len = offset + conn_param->private_data_len;
2564         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2565         if (!private_data)
2566                 return -ENOMEM;
2567
2568         if (conn_param->private_data && conn_param->private_data_len)
2569                 memcpy(private_data + offset, conn_param->private_data,
2570                        conn_param->private_data_len);
2571
2572         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2573         if (IS_ERR(id)) {
2574                 ret = PTR_ERR(id);
2575                 goto out;
2576         }
2577         id_priv->cm_id.ib = id;
2578
2579         route = &id_priv->id.route;
2580         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2581         if (ret)
2582                 goto out;
2583         req.private_data = private_data;
2584
2585         req.primary_path = &route->path_rec[0];
2586         if (route->num_paths == 2)
2587                 req.alternate_path = &route->path_rec[1];
2588
2589         req.service_id = cma_get_service_id(id_priv->id.ps,
2590                                             (struct sockaddr *) &route->addr.dst_addr);
2591         req.qp_num = id_priv->qp_num;
2592         req.qp_type = id_priv->id.qp_type;
2593         req.starting_psn = id_priv->seq_num;
2594         req.responder_resources = conn_param->responder_resources;
2595         req.initiator_depth = conn_param->initiator_depth;
2596         req.flow_control = conn_param->flow_control;
2597         req.retry_count = conn_param->retry_count;
2598         req.rnr_retry_count = conn_param->rnr_retry_count;
2599         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2600         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2601         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2602         req.srq = id_priv->srq ? 1 : 0;
2603
2604         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2605 out:
2606         if (ret && !IS_ERR(id)) {
2607                 ib_destroy_cm_id(id);
2608                 id_priv->cm_id.ib = NULL;
2609         }
2610
2611         kfree(private_data);
2612         return ret;
2613 }
2614
2615 static int cma_connect_iw(struct rdma_id_private *id_priv,
2616                           struct rdma_conn_param *conn_param)
2617 {
2618         struct iw_cm_id *cm_id;
2619         struct sockaddr_in* sin;
2620         int ret;
2621         struct iw_cm_conn_param iw_param;
2622
2623         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2624         if (IS_ERR(cm_id))
2625                 return PTR_ERR(cm_id);
2626
2627         id_priv->cm_id.iw = cm_id;
2628
2629         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2630         cm_id->local_addr = *sin;
2631
2632         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2633         cm_id->remote_addr = *sin;
2634
2635         ret = cma_modify_qp_rtr(id_priv, conn_param);
2636         if (ret)
2637                 goto out;
2638
2639         if (conn_param) {
2640                 iw_param.ord = conn_param->initiator_depth;
2641                 iw_param.ird = conn_param->responder_resources;
2642                 iw_param.private_data = conn_param->private_data;
2643                 iw_param.private_data_len = conn_param->private_data_len;
2644                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2645         } else {
2646                 memset(&iw_param, 0, sizeof iw_param);
2647                 iw_param.qpn = id_priv->qp_num;
2648         }
2649         ret = iw_cm_connect(cm_id, &iw_param);
2650 out:
2651         if (ret) {
2652                 iw_destroy_cm_id(cm_id);
2653                 id_priv->cm_id.iw = NULL;
2654         }
2655         return ret;
2656 }
2657
2658 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2659 {
2660         struct rdma_id_private *id_priv;
2661         int ret;
2662
2663         id_priv = container_of(id, struct rdma_id_private, id);
2664         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2665                 return -EINVAL;
2666
2667         if (!id->qp) {
2668                 id_priv->qp_num = conn_param->qp_num;
2669                 id_priv->srq = conn_param->srq;
2670         }
2671
2672         switch (rdma_node_get_transport(id->device->node_type)) {
2673         case RDMA_TRANSPORT_IB:
2674                 if (id->qp_type == IB_QPT_UD)
2675                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2676                 else
2677                         ret = cma_connect_ib(id_priv, conn_param);
2678                 break;
2679         case RDMA_TRANSPORT_IWARP:
2680                 ret = cma_connect_iw(id_priv, conn_param);
2681                 break;
2682         default:
2683                 ret = -ENOSYS;
2684                 break;
2685         }
2686         if (ret)
2687                 goto err;
2688
2689         return 0;
2690 err:
2691         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2692         return ret;
2693 }
2694 EXPORT_SYMBOL(rdma_connect);
2695
2696 static int cma_accept_ib(struct rdma_id_private *id_priv,
2697                          struct rdma_conn_param *conn_param)
2698 {
2699         struct ib_cm_rep_param rep;
2700         int ret;
2701
2702         ret = cma_modify_qp_rtr(id_priv, conn_param);
2703         if (ret)
2704                 goto out;
2705
2706         ret = cma_modify_qp_rts(id_priv, conn_param);
2707         if (ret)
2708                 goto out;
2709
2710         memset(&rep, 0, sizeof rep);
2711         rep.qp_num = id_priv->qp_num;
2712         rep.starting_psn = id_priv->seq_num;
2713         rep.private_data = conn_param->private_data;
2714         rep.private_data_len = conn_param->private_data_len;
2715         rep.responder_resources = conn_param->responder_resources;
2716         rep.initiator_depth = conn_param->initiator_depth;
2717         rep.failover_accepted = 0;
2718         rep.flow_control = conn_param->flow_control;
2719         rep.rnr_retry_count = conn_param->rnr_retry_count;
2720         rep.srq = id_priv->srq ? 1 : 0;
2721
2722         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2723 out:
2724         return ret;
2725 }
2726
2727 static int cma_accept_iw(struct rdma_id_private *id_priv,
2728                   struct rdma_conn_param *conn_param)
2729 {
2730         struct iw_cm_conn_param iw_param;
2731         int ret;
2732
2733         ret = cma_modify_qp_rtr(id_priv, conn_param);
2734         if (ret)
2735                 return ret;
2736
2737         iw_param.ord = conn_param->initiator_depth;
2738         iw_param.ird = conn_param->responder_resources;
2739         iw_param.private_data = conn_param->private_data;
2740         iw_param.private_data_len = conn_param->private_data_len;
2741         if (id_priv->id.qp) {
2742                 iw_param.qpn = id_priv->qp_num;
2743         } else
2744                 iw_param.qpn = conn_param->qp_num;
2745
2746         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2747 }
2748
2749 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2750                              enum ib_cm_sidr_status status,
2751                              const void *private_data, int private_data_len)
2752 {
2753         struct ib_cm_sidr_rep_param rep;
2754         int ret;
2755
2756         memset(&rep, 0, sizeof rep);
2757         rep.status = status;
2758         if (status == IB_SIDR_SUCCESS) {
2759                 ret = cma_set_qkey(id_priv);
2760                 if (ret)
2761                         return ret;
2762                 rep.qp_num = id_priv->qp_num;
2763                 rep.qkey = id_priv->qkey;
2764         }
2765         rep.private_data = private_data;
2766         rep.private_data_len = private_data_len;
2767
2768         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2769 }
2770
2771 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2772 {
2773         struct rdma_id_private *id_priv;
2774         int ret;
2775
2776         id_priv = container_of(id, struct rdma_id_private, id);
2777
2778         id_priv->owner = task_pid_nr(current);
2779
2780         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2781                 return -EINVAL;
2782
2783         if (!id->qp && conn_param) {
2784                 id_priv->qp_num = conn_param->qp_num;
2785                 id_priv->srq = conn_param->srq;
2786         }
2787
2788         switch (rdma_node_get_transport(id->device->node_type)) {
2789         case RDMA_TRANSPORT_IB:
2790                 if (id->qp_type == IB_QPT_UD) {
2791                         if (conn_param)
2792                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2793                                                         conn_param->private_data,
2794                                                         conn_param->private_data_len);
2795                         else
2796                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2797                                                         NULL, 0);
2798                 } else {
2799                         if (conn_param)
2800                                 ret = cma_accept_ib(id_priv, conn_param);
2801                         else
2802                                 ret = cma_rep_recv(id_priv);
2803                 }
2804                 break;
2805         case RDMA_TRANSPORT_IWARP:
2806                 ret = cma_accept_iw(id_priv, conn_param);
2807                 break;
2808         default:
2809                 ret = -ENOSYS;
2810                 break;
2811         }
2812
2813         if (ret)
2814                 goto reject;
2815
2816         return 0;
2817 reject:
2818         cma_modify_qp_err(id_priv);
2819         rdma_reject(id, NULL, 0);
2820         return ret;
2821 }
2822 EXPORT_SYMBOL(rdma_accept);
2823
2824 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2825 {
2826         struct rdma_id_private *id_priv;
2827         int ret;
2828
2829         id_priv = container_of(id, struct rdma_id_private, id);
2830         if (!id_priv->cm_id.ib)
2831                 return -EINVAL;
2832
2833         switch (id->device->node_type) {
2834         case RDMA_NODE_IB_CA:
2835                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2836                 break;
2837         default:
2838                 ret = 0;
2839                 break;
2840         }
2841         return ret;
2842 }
2843 EXPORT_SYMBOL(rdma_notify);
2844
2845 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2846                 u8 private_data_len)
2847 {
2848         struct rdma_id_private *id_priv;
2849         int ret;
2850
2851         id_priv = container_of(id, struct rdma_id_private, id);
2852         if (!id_priv->cm_id.ib)
2853                 return -EINVAL;
2854
2855         switch (rdma_node_get_transport(id->device->node_type)) {
2856         case RDMA_TRANSPORT_IB:
2857                 if (id->qp_type == IB_QPT_UD)
2858                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2859                                                 private_data, private_data_len);
2860                 else
2861                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2862                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2863                                              0, private_data, private_data_len);
2864                 break;
2865         case RDMA_TRANSPORT_IWARP:
2866                 ret = iw_cm_reject(id_priv->cm_id.iw,
2867                                    private_data, private_data_len);
2868                 break;
2869         default:
2870                 ret = -ENOSYS;
2871                 break;
2872         }
2873         return ret;
2874 }
2875 EXPORT_SYMBOL(rdma_reject);
2876
2877 int rdma_disconnect(struct rdma_cm_id *id)
2878 {
2879         struct rdma_id_private *id_priv;
2880         int ret;
2881
2882         id_priv = container_of(id, struct rdma_id_private, id);
2883         if (!id_priv->cm_id.ib)
2884                 return -EINVAL;
2885
2886         switch (rdma_node_get_transport(id->device->node_type)) {
2887         case RDMA_TRANSPORT_IB:
2888                 ret = cma_modify_qp_err(id_priv);
2889                 if (ret)
2890                         goto out;
2891                 /* Initiate or respond to a disconnect. */
2892                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2893                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2894                 break;
2895         case RDMA_TRANSPORT_IWARP:
2896                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2897                 break;
2898         default:
2899                 ret = -EINVAL;
2900                 break;
2901         }
2902 out:
2903         return ret;
2904 }
2905 EXPORT_SYMBOL(rdma_disconnect);
2906
2907 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2908 {
2909         struct rdma_id_private *id_priv;
2910         struct cma_multicast *mc = multicast->context;
2911         struct rdma_cm_event event;
2912         int ret;
2913
2914         id_priv = mc->id_priv;
2915         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2916             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2917                 return 0;
2918
2919         mutex_lock(&id_priv->qp_mutex);
2920         if (!status && id_priv->id.qp)
2921                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2922                                          multicast->rec.mlid);
2923         mutex_unlock(&id_priv->qp_mutex);
2924
2925         memset(&event, 0, sizeof event);
2926         event.status = status;
2927         event.param.ud.private_data = mc->context;
2928         if (!status) {
2929                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2930                 ib_init_ah_from_mcmember(id_priv->id.device,
2931                                          id_priv->id.port_num, &multicast->rec,
2932                                          &event.param.ud.ah_attr);
2933                 event.param.ud.qp_num = 0xFFFFFF;
2934                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2935         } else
2936                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2937
2938         ret = id_priv->id.event_handler(&id_priv->id, &event);
2939         if (ret) {
2940                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2941                 mutex_unlock(&id_priv->handler_mutex);
2942                 rdma_destroy_id(&id_priv->id);
2943                 return 0;
2944         }
2945
2946         mutex_unlock(&id_priv->handler_mutex);
2947         return 0;
2948 }
2949
2950 static void cma_set_mgid(struct rdma_id_private *id_priv,
2951                          struct sockaddr *addr, union ib_gid *mgid)
2952 {
2953         unsigned char mc_map[MAX_ADDR_LEN];
2954         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2955         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2956         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2957
2958         if (cma_any_addr(addr)) {
2959                 memset(mgid, 0, sizeof *mgid);
2960         } else if ((addr->sa_family == AF_INET6) &&
2961                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2962                                                                  0xFF10A01B)) {
2963                 /* IPv6 address is an SA assigned MGID. */
2964                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2965         } else if ((addr->sa_family == AF_INET6)) {
2966                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2967                 if (id_priv->id.ps == RDMA_PS_UDP)
2968                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2969                 *mgid = *(union ib_gid *) (mc_map + 4);
2970         } else {
2971                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2972                 if (id_priv->id.ps == RDMA_PS_UDP)
2973                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2974                 *mgid = *(union ib_gid *) (mc_map + 4);
2975         }
2976 }
2977
2978 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2979                                  struct cma_multicast *mc)
2980 {
2981         struct ib_sa_mcmember_rec rec;
2982         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2983         ib_sa_comp_mask comp_mask;
2984         int ret;
2985
2986         ib_addr_get_mgid(dev_addr, &rec.mgid);
2987         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2988                                      &rec.mgid, &rec);
2989         if (ret)
2990                 return ret;
2991
2992         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2993         if (id_priv->id.ps == RDMA_PS_UDP)
2994                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2995         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2996         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2997         rec.join_state = 1;
2998
2999         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3000                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3001                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3002                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3003                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3004
3005         if (id_priv->id.ps == RDMA_PS_IPOIB)
3006                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3007                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
3008
3009         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3010                                                 id_priv->id.port_num, &rec,
3011                                                 comp_mask, GFP_KERNEL,
3012                                                 cma_ib_mc_handler, mc);
3013         if (IS_ERR(mc->multicast.ib))
3014                 return PTR_ERR(mc->multicast.ib);
3015
3016         return 0;
3017 }
3018
3019 static void iboe_mcast_work_handler(struct work_struct *work)
3020 {
3021         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3022         struct cma_multicast *mc = mw->mc;
3023         struct ib_sa_multicast *m = mc->multicast.ib;
3024
3025         mc->multicast.ib->context = mc;
3026         cma_ib_mc_handler(0, m);
3027         kref_put(&mc->mcref, release_mc);
3028         kfree(mw);
3029 }
3030
3031 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3032 {
3033         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3034         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3035
3036         if (cma_any_addr(addr)) {
3037                 memset(mgid, 0, sizeof *mgid);
3038         } else if (addr->sa_family == AF_INET6) {
3039                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3040         } else {
3041                 mgid->raw[0] = 0xff;
3042                 mgid->raw[1] = 0x0e;
3043                 mgid->raw[2] = 0;
3044                 mgid->raw[3] = 0;
3045                 mgid->raw[4] = 0;
3046                 mgid->raw[5] = 0;
3047                 mgid->raw[6] = 0;
3048                 mgid->raw[7] = 0;
3049                 mgid->raw[8] = 0;
3050                 mgid->raw[9] = 0;
3051                 mgid->raw[10] = 0xff;
3052                 mgid->raw[11] = 0xff;
3053                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3054         }
3055 }
3056
3057 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3058                                    struct cma_multicast *mc)
3059 {
3060         struct iboe_mcast_work *work;
3061         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3062         int err;
3063         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3064         struct net_device *ndev = NULL;
3065
3066         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3067                 return -EINVAL;
3068
3069         work = kzalloc(sizeof *work, GFP_KERNEL);
3070         if (!work)
3071                 return -ENOMEM;
3072
3073         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3074         if (!mc->multicast.ib) {
3075                 err = -ENOMEM;
3076                 goto out1;
3077         }
3078
3079         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3080
3081         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3082         if (id_priv->id.ps == RDMA_PS_UDP)
3083                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3084
3085         if (dev_addr->bound_dev_if)
3086                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3087         if (!ndev) {
3088                 err = -ENODEV;
3089                 goto out2;
3090         }
3091         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3092         mc->multicast.ib->rec.hop_limit = 1;
3093         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3094         dev_put(ndev);
3095         if (!mc->multicast.ib->rec.mtu) {
3096                 err = -EINVAL;
3097                 goto out2;
3098         }
3099         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3100         work->id = id_priv;
3101         work->mc = mc;
3102         INIT_WORK(&work->work, iboe_mcast_work_handler);
3103         kref_get(&mc->mcref);
3104         queue_work(cma_wq, &work->work);
3105
3106         return 0;
3107
3108 out2:
3109         kfree(mc->multicast.ib);
3110 out1:
3111         kfree(work);
3112         return err;
3113 }
3114
3115 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3116                         void *context)
3117 {
3118         struct rdma_id_private *id_priv;
3119         struct cma_multicast *mc;
3120         int ret;
3121
3122         id_priv = container_of(id, struct rdma_id_private, id);
3123         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3124             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3125                 return -EINVAL;
3126
3127         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3128         if (!mc)
3129                 return -ENOMEM;
3130
3131         memcpy(&mc->addr, addr, ip_addr_size(addr));
3132         mc->context = context;
3133         mc->id_priv = id_priv;
3134
3135         spin_lock(&id_priv->lock);
3136         list_add(&mc->list, &id_priv->mc_list);
3137         spin_unlock(&id_priv->lock);
3138
3139         switch (rdma_node_get_transport(id->device->node_type)) {
3140         case RDMA_TRANSPORT_IB:
3141                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3142                 case IB_LINK_LAYER_INFINIBAND:
3143                         ret = cma_join_ib_multicast(id_priv, mc);
3144                         break;
3145                 case IB_LINK_LAYER_ETHERNET:
3146                         kref_init(&mc->mcref);
3147                         ret = cma_iboe_join_multicast(id_priv, mc);
3148                         break;
3149                 default:
3150                         ret = -EINVAL;
3151                 }
3152                 break;
3153         default:
3154                 ret = -ENOSYS;
3155                 break;
3156         }
3157
3158         if (ret) {
3159                 spin_lock_irq(&id_priv->lock);
3160                 list_del(&mc->list);
3161                 spin_unlock_irq(&id_priv->lock);
3162                 kfree(mc);
3163         }
3164         return ret;
3165 }
3166 EXPORT_SYMBOL(rdma_join_multicast);
3167
3168 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3169 {
3170         struct rdma_id_private *id_priv;
3171         struct cma_multicast *mc;
3172
3173         id_priv = container_of(id, struct rdma_id_private, id);
3174         spin_lock_irq(&id_priv->lock);
3175         list_for_each_entry(mc, &id_priv->mc_list, list) {
3176                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3177                         list_del(&mc->list);
3178                         spin_unlock_irq(&id_priv->lock);
3179
3180                         if (id->qp)
3181                                 ib_detach_mcast(id->qp,
3182                                                 &mc->multicast.ib->rec.mgid,
3183                                                 mc->multicast.ib->rec.mlid);
3184                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3185                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3186                                 case IB_LINK_LAYER_INFINIBAND:
3187                                         ib_sa_free_multicast(mc->multicast.ib);
3188                                         kfree(mc);
3189                                         break;
3190                                 case IB_LINK_LAYER_ETHERNET:
3191                                         kref_put(&mc->mcref, release_mc);
3192                                         break;
3193                                 default:
3194                                         break;
3195                                 }
3196                         }
3197                         return;
3198                 }
3199         }
3200         spin_unlock_irq(&id_priv->lock);
3201 }
3202 EXPORT_SYMBOL(rdma_leave_multicast);
3203
3204 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3205 {
3206         struct rdma_dev_addr *dev_addr;
3207         struct cma_ndev_work *work;
3208
3209         dev_addr = &id_priv->id.route.addr.dev_addr;
3210
3211         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3212             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3213                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3214                        ndev->name, &id_priv->id);
3215                 work = kzalloc(sizeof *work, GFP_KERNEL);
3216                 if (!work)
3217                         return -ENOMEM;
3218
3219                 INIT_WORK(&work->work, cma_ndev_work_handler);
3220                 work->id = id_priv;
3221                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3222                 atomic_inc(&id_priv->refcount);
3223                 queue_work(cma_wq, &work->work);
3224         }
3225
3226         return 0;
3227 }
3228
3229 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3230                                void *ctx)
3231 {
3232         struct net_device *ndev = (struct net_device *)ctx;
3233         struct cma_device *cma_dev;
3234         struct rdma_id_private *id_priv;
3235         int ret = NOTIFY_DONE;
3236
3237         if (dev_net(ndev) != &init_net)
3238                 return NOTIFY_DONE;
3239
3240         if (event != NETDEV_BONDING_FAILOVER)
3241                 return NOTIFY_DONE;
3242
3243         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3244                 return NOTIFY_DONE;
3245
3246         mutex_lock(&lock);
3247         list_for_each_entry(cma_dev, &dev_list, list)
3248                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3249                         ret = cma_netdev_change(ndev, id_priv);
3250                         if (ret)
3251                                 goto out;
3252                 }
3253
3254 out:
3255         mutex_unlock(&lock);
3256         return ret;
3257 }
3258
3259 static struct notifier_block cma_nb = {
3260         .notifier_call = cma_netdev_callback
3261 };
3262
3263 static void cma_add_one(struct ib_device *device)
3264 {
3265         struct cma_device *cma_dev;
3266         struct rdma_id_private *id_priv;
3267
3268         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3269         if (!cma_dev)
3270                 return;
3271
3272         cma_dev->device = device;
3273
3274         init_completion(&cma_dev->comp);
3275         atomic_set(&cma_dev->refcount, 1);
3276         INIT_LIST_HEAD(&cma_dev->id_list);
3277         ib_set_client_data(device, &cma_client, cma_dev);
3278
3279         mutex_lock(&lock);
3280         list_add_tail(&cma_dev->list, &dev_list);
3281         list_for_each_entry(id_priv, &listen_any_list, list)
3282                 cma_listen_on_dev(id_priv, cma_dev);
3283         mutex_unlock(&lock);
3284 }
3285
3286 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3287 {
3288         struct rdma_cm_event event;
3289         enum rdma_cm_state state;
3290         int ret = 0;
3291
3292         /* Record that we want to remove the device */
3293         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3294         if (state == RDMA_CM_DESTROYING)
3295                 return 0;
3296
3297         cma_cancel_operation(id_priv, state);
3298         mutex_lock(&id_priv->handler_mutex);
3299
3300         /* Check for destruction from another callback. */
3301         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3302                 goto out;
3303
3304         memset(&event, 0, sizeof event);
3305         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3306         ret = id_priv->id.event_handler(&id_priv->id, &event);
3307 out:
3308         mutex_unlock(&id_priv->handler_mutex);
3309         return ret;
3310 }
3311
3312 static void cma_process_remove(struct cma_device *cma_dev)
3313 {
3314         struct rdma_id_private *id_priv;
3315         int ret;
3316
3317         mutex_lock(&lock);
3318         while (!list_empty(&cma_dev->id_list)) {
3319                 id_priv = list_entry(cma_dev->id_list.next,
3320                                      struct rdma_id_private, list);
3321
3322                 list_del(&id_priv->listen_list);
3323                 list_del_init(&id_priv->list);
3324                 atomic_inc(&id_priv->refcount);
3325                 mutex_unlock(&lock);
3326
3327                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3328                 cma_deref_id(id_priv);
3329                 if (ret)
3330                         rdma_destroy_id(&id_priv->id);
3331
3332                 mutex_lock(&lock);
3333         }
3334         mutex_unlock(&lock);
3335
3336         cma_deref_dev(cma_dev);
3337         wait_for_completion(&cma_dev->comp);
3338 }
3339
3340 static void cma_remove_one(struct ib_device *device)
3341 {
3342         struct cma_device *cma_dev;
3343
3344         cma_dev = ib_get_client_data(device, &cma_client);
3345         if (!cma_dev)
3346                 return;
3347
3348         mutex_lock(&lock);
3349         list_del(&cma_dev->list);
3350         mutex_unlock(&lock);
3351
3352         cma_process_remove(cma_dev);
3353         kfree(cma_dev);
3354 }
3355
3356 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3357 {
3358         struct nlmsghdr *nlh;
3359         struct rdma_cm_id_stats *id_stats;
3360         struct rdma_id_private *id_priv;
3361         struct rdma_cm_id *id = NULL;
3362         struct cma_device *cma_dev;
3363         int i_dev = 0, i_id = 0;
3364
3365         /*
3366          * We export all of the IDs as a sequence of messages.  Each
3367          * ID gets its own netlink message.
3368          */
3369         mutex_lock(&lock);
3370
3371         list_for_each_entry(cma_dev, &dev_list, list) {
3372                 if (i_dev < cb->args[0]) {
3373                         i_dev++;
3374                         continue;
3375                 }
3376
3377                 i_id = 0;
3378                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3379                         if (i_id < cb->args[1]) {
3380                                 i_id++;
3381                                 continue;
3382                         }
3383
3384                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3385                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3386                                                 RDMA_NL_RDMA_CM_ID_STATS);
3387                         if (!id_stats)
3388                                 goto out;
3389
3390                         memset(id_stats, 0, sizeof *id_stats);
3391                         id = &id_priv->id;
3392                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3393                         id_stats->port_num = id->port_num;
3394                         id_stats->bound_dev_if =
3395                                 id->route.addr.dev_addr.bound_dev_if;
3396
3397                         if (id->route.addr.src_addr.ss_family == AF_INET) {
3398                                 if (ibnl_put_attr(skb, nlh,
3399                                                   sizeof(struct sockaddr_in),
3400                                                   &id->route.addr.src_addr,
3401                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3402                                         goto out;
3403                                 }
3404                                 if (ibnl_put_attr(skb, nlh,
3405                                                   sizeof(struct sockaddr_in),
3406                                                   &id->route.addr.dst_addr,
3407                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3408                                         goto out;
3409                                 }
3410                         } else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3411                                 if (ibnl_put_attr(skb, nlh,
3412                                                   sizeof(struct sockaddr_in6),
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_in6),
3419                                                   &id->route.addr.dst_addr,
3420                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3421                                         goto out;
3422                                 }
3423                         }
3424
3425                         id_stats->pid           = id_priv->owner;
3426                         id_stats->port_space    = id->ps;
3427                         id_stats->cm_state      = id_priv->state;
3428                         id_stats->qp_num        = id_priv->qp_num;
3429                         id_stats->qp_type       = id->qp_type;
3430
3431                         i_id++;
3432                 }
3433
3434                 cb->args[1] = 0;
3435                 i_dev++;
3436         }
3437
3438 out:
3439         mutex_unlock(&lock);
3440         cb->args[0] = i_dev;
3441         cb->args[1] = i_id;
3442
3443         return skb->len;
3444 }
3445
3446 static const struct ibnl_client_cbs cma_cb_table[] = {
3447         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats },
3448 };
3449
3450 static int __init cma_init(void)
3451 {
3452         int ret;
3453
3454         cma_wq = create_singlethread_workqueue("rdma_cm");
3455         if (!cma_wq)
3456                 return -ENOMEM;
3457
3458         ib_sa_register_client(&sa_client);
3459         rdma_addr_register_client(&addr_client);
3460         register_netdevice_notifier(&cma_nb);
3461
3462         ret = ib_register_client(&cma_client);
3463         if (ret)
3464                 goto err;
3465
3466         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3467                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3468
3469         return 0;
3470
3471 err:
3472         unregister_netdevice_notifier(&cma_nb);
3473         rdma_addr_unregister_client(&addr_client);
3474         ib_sa_unregister_client(&sa_client);
3475         destroy_workqueue(cma_wq);
3476         return ret;
3477 }
3478
3479 static void __exit cma_cleanup(void)
3480 {
3481         ibnl_remove_client(RDMA_NL_RDMA_CM);
3482         ib_unregister_client(&cma_client);
3483         unregister_netdevice_notifier(&cma_nb);
3484         rdma_addr_unregister_client(&addr_client);
3485         ib_sa_unregister_client(&sa_client);
3486         destroy_workqueue(cma_wq);
3487         idr_destroy(&sdp_ps);
3488         idr_destroy(&tcp_ps);
3489         idr_destroy(&udp_ps);
3490         idr_destroy(&ipoib_ps);
3491         idr_destroy(&ib_ps);
3492 }
3493
3494 module_init(cma_init);
3495 module_exit(cma_cleanup);