drivers/infiniband/core/cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/idr.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/slab.h>
  44 #include <linux/module.h>
  45 #include <net/route.h>
  46
  47 #include <net/tcp.h>
  48 #include <net/ipv6.h>
  49
  50 #include <rdma/rdma_cm.h>
  51 #include <rdma/rdma_cm_ib.h>
  52 #include <rdma/rdma_netlink.h>
  53 #include <rdma/ib_cache.h>
  54 #include <rdma/ib_cm.h>
  55 #include <rdma/ib_sa.h>
  56 #include <rdma/iw_cm.h>
  57
  58 MODULE_AUTHOR("Sean Hefty");
  59 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  60 MODULE_LICENSE("Dual BSD/GPL");
  61
  62 #define CMA_CM_RESPONSE_TIMEOUT 20
  63 #define CMA_MAX_CM_RETRIES 15
  64 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  65 #define CMA_IBOE_PACKET_LIFETIME 18
  66
  67 static void cma_add_one(struct ib_device *device);
  68 static void cma_remove_one(struct ib_device *device);
  69
  70 static struct ib_client cma_client = {
  71         .name   = "cma",
  72         .add    = cma_add_one,
  73         .remove = cma_remove_one
  74 };
  75
  76 static struct ib_sa_client sa_client;
  77 static struct rdma_addr_client addr_client;
  78 static LIST_HEAD(dev_list);
  79 static LIST_HEAD(listen_any_list);
  80 static DEFINE_MUTEX(lock);
  81 static struct workqueue_struct *cma_wq;
  82 static DEFINE_IDR(sdp_ps);
  83 static DEFINE_IDR(tcp_ps);
  84 static DEFINE_IDR(udp_ps);
  85 static DEFINE_IDR(ipoib_ps);
  86 static DEFINE_IDR(ib_ps);
  87
  88 struct cma_device {
  89         struct list_head        list;
  90         struct ib_device        *device;
  91         struct completion       comp;
  92         atomic_t                refcount;
  93         struct list_head        id_list;
  94 };
  95
  96 struct rdma_bind_list {
  97         struct idr              *ps;
  98         struct hlist_head       owners;
  99         unsigned short          port;
 100 };
 101
 102 /*
 103  * Device removal can occur at anytime, so we need extra handling to
 104  * serialize notifying the user of device removal with other callbacks.
 105  * We do this by disabling removal notification while a callback is in process,
 106  * and reporting it after the callback completes.
 107  */
 108 struct rdma_id_private {
 109         struct rdma_cm_id       id;
 110
 111         struct rdma_bind_list   *bind_list;
 112         struct hlist_node       node;
 113         struct list_head        list; /* listen_any_list or cma_device.list */
 114         struct list_head        listen_list; /* per device listens */
 115         struct cma_device       *cma_dev;
 116         struct list_head        mc_list;
 117
 118         int                     internal_id;
 119         enum rdma_cm_state      state;
 120         spinlock_t              lock;
 121         struct mutex            qp_mutex;
 122
 123         struct completion       comp;
 124         atomic_t                refcount;
 125         struct mutex            handler_mutex;
 126
 127         int                     backlog;
 128         int                     timeout_ms;
 129         struct ib_sa_query      *query;
 130         int                     query_id;
 131         union {
 132                 struct ib_cm_id *ib;
 133                 struct iw_cm_id *iw;
 134         } cm_id;
 135
 136         u32                     seq_num;
 137         u32                     qkey;
 138         u32                     qp_num;
 139         pid_t                   owner;
 140         u8                      srq;
 141         u8                      tos;
 142         u8                      reuseaddr;
 143 };
 144
 145 struct cma_multicast {
 146         struct rdma_id_private *id_priv;
 147         union {
 148                 struct ib_sa_multicast *ib;
 149         } multicast;
 150         struct list_head        list;
 151         void                    *context;
 152         struct sockaddr_storage addr;
 153         struct kref             mcref;
 154 };
 155
 156 struct cma_work {
 157         struct work_struct      work;
 158         struct rdma_id_private  *id;
 159         enum rdma_cm_state      old_state;
 160         enum rdma_cm_state      new_state;
 161         struct rdma_cm_event    event;
 162 };
 163
 164 struct cma_ndev_work {
 165         struct work_struct      work;
 166         struct rdma_id_private  *id;
 167         struct rdma_cm_event    event;
 168 };
 169
 170 struct iboe_mcast_work {
 171         struct work_struct       work;
 172         struct rdma_id_private  *id;
 173         struct cma_multicast    *mc;
 174 };
 175
 176 union cma_ip_addr {
 177         struct in6_addr ip6;
 178         struct {
 179                 __be32 pad[3];
 180                 __be32 addr;
 181         } ip4;
 182 };
 183
 184 struct cma_hdr {
 185         u8 cma_version;
 186         u8 ip_version;  /* IP version: 7:4 */
 187         __be16 port;
 188         union cma_ip_addr src_addr;
 189         union cma_ip_addr dst_addr;
 190 };
 191
 192 struct sdp_hh {
 193         u8 bsdh[16];
 194         u8 sdp_version; /* Major version: 7:4 */
 195         u8 ip_version;  /* IP version: 7:4 */
 196         u8 sdp_specific1[10];
 197         __be16 port;
 198         __be16 sdp_specific2;
 199         union cma_ip_addr src_addr;
 200         union cma_ip_addr dst_addr;
 201 };
 202
 203 struct sdp_hah {
 204         u8 bsdh[16];
 205         u8 sdp_version;
 206 };
 207
 208 #define CMA_VERSION 0x00
 209 #define SDP_MAJ_VERSION 0x2
 210
 211 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 212 {
 213         unsigned long flags;
 214         int ret;
 215
 216         spin_lock_irqsave(&id_priv->lock, flags);
 217         ret = (id_priv->state == comp);
 218         spin_unlock_irqrestore(&id_priv->lock, flags);
 219         return ret;
 220 }
 221
 222 static int cma_comp_exch(struct rdma_id_private *id_priv,
 223                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 224 {
 225         unsigned long flags;
 226         int ret;
 227
 228         spin_lock_irqsave(&id_priv->lock, flags);
 229         if ((ret = (id_priv->state == comp)))
 230                 id_priv->state = exch;
 231         spin_unlock_irqrestore(&id_priv->lock, flags);
 232         return ret;
 233 }
 234
 235 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 236                                    enum rdma_cm_state exch)
 237 {
 238         unsigned long flags;
 239         enum rdma_cm_state old;
 240
 241         spin_lock_irqsave(&id_priv->lock, flags);
 242         old = id_priv->state;
 243         id_priv->state = exch;
 244         spin_unlock_irqrestore(&id_priv->lock, flags);
 245         return old;
 246 }
 247
 248 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 249 {
 250         return hdr->ip_version >> 4;
 251 }
 252
 253 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 254 {
 255         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 256 }
 257
 258 static inline u8 sdp_get_majv(u8 sdp_version)
 259 {
 260         return sdp_version >> 4;
 261 }
 262
 263 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 264 {
 265         return hh->ip_version >> 4;
 266 }
 267
 268 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 269 {
 270         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 271 }
 272
 273 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 274                               struct cma_device *cma_dev)
 275 {
 276         atomic_inc(&cma_dev->refcount);
 277         id_priv->cma_dev = cma_dev;
 278         id_priv->id.device = cma_dev->device;
 279         id_priv->id.route.addr.dev_addr.transport =
 280                 rdma_node_get_transport(cma_dev->device->node_type);
 281         list_add_tail(&id_priv->list, &cma_dev->id_list);
 282 }
 283
 284 static inline void cma_deref_dev(struct cma_device *cma_dev)
 285 {
 286         if (atomic_dec_and_test(&cma_dev->refcount))
 287                 complete(&cma_dev->comp);
 288 }
 289
 290 static inline void release_mc(struct kref *kref)
 291 {
 292         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 293
 294         kfree(mc->multicast.ib);
 295         kfree(mc);
 296 }
 297
 298 static void cma_release_dev(struct rdma_id_private *id_priv)
 299 {
 300         mutex_lock(&lock);
 301         list_del(&id_priv->list);
 302         cma_deref_dev(id_priv->cma_dev);
 303         id_priv->cma_dev = NULL;
 304         mutex_unlock(&lock);
 305 }
 306
 307 static int cma_set_qkey(struct rdma_id_private *id_priv)
 308 {
 309         struct ib_sa_mcmember_rec rec;
 310         int ret = 0;
 311
 312         if (id_priv->qkey)
 313                 return 0;
 314
 315         switch (id_priv->id.ps) {
 316         case RDMA_PS_UDP:
 317                 id_priv->qkey = RDMA_UDP_QKEY;
 318                 break;
 319         case RDMA_PS_IPOIB:
 320                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 321                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 322                                              id_priv->id.port_num, &rec.mgid,
 323                                              &rec);
 324                 if (!ret)
 325                         id_priv->qkey = be32_to_cpu(rec.qkey);
 326                 break;
 327         default:
 328                 break;
 329         }
 330         return ret;
 331 }
 332
 333 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 334 {
 335         int i;
 336         int err;
 337         struct ib_port_attr props;
 338         union ib_gid tmp;
 339
 340         err = ib_query_port(device, port_num, &props);
 341         if (err)
 342                 return 1;
 343
 344         for (i = 0; i < props.gid_tbl_len; ++i) {
 345                 err = ib_query_gid(device, port_num, i, &tmp);
 346                 if (err)
 347                         return 1;
 348                 if (!memcmp(&tmp, gid, sizeof tmp))
 349                         return 0;
 350         }
 351
 352         return -EAGAIN;
 353 }
 354
 355 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 356 {
 357         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 358         struct cma_device *cma_dev;
 359         union ib_gid gid, iboe_gid;
 360         int ret = -ENODEV;
 361         u8 port;
 362         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 363                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 364
 365         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 366             id_priv->id.ps == RDMA_PS_IPOIB)
 367                 return -EINVAL;
 368
 369         mutex_lock(&lock);
 370         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 371         memcpy(&gid, dev_addr->src_dev_addr +
 372                rdma_addr_gid_offset(dev_addr), sizeof gid);
 373         list_for_each_entry(cma_dev, &dev_list, list) {
 374                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 375                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 376                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 377                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 378                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 379                                 else
 380                                         ret = find_gid_port(cma_dev->device, &gid, port);
 381
 382                                 if (!ret) {
 383                                         id_priv->id.port_num = port;
 384                                         goto out;
 385                                 } else if (ret == 1)
 386                                         break;
 387                         }
 388                 }
 389         }
 390
 391 out:
 392         if (!ret)
 393                 cma_attach_to_dev(id_priv, cma_dev);
 394
 395         mutex_unlock(&lock);
 396         return ret;
 397 }
 398
 399 static void cma_deref_id(struct rdma_id_private *id_priv)
 400 {
 401         if (atomic_dec_and_test(&id_priv->refcount))
 402                 complete(&id_priv->comp);
 403 }
 404
 405 static int cma_disable_callback(struct rdma_id_private *id_priv,
 406                                 enum rdma_cm_state state)
 407 {
 408         mutex_lock(&id_priv->handler_mutex);
 409         if (id_priv->state != state) {
 410                 mutex_unlock(&id_priv->handler_mutex);
 411                 return -EINVAL;
 412         }
 413         return 0;
 414 }
 415
 416 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 417                                   void *context, enum rdma_port_space ps,
 418                                   enum ib_qp_type qp_type)
 419 {
 420         struct rdma_id_private *id_priv;
 421
 422         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 423         if (!id_priv)
 424                 return ERR_PTR(-ENOMEM);
 425
 426         id_priv->owner = task_pid_nr(current);
 427         id_priv->state = RDMA_CM_IDLE;
 428         id_priv->id.context = context;
 429         id_priv->id.event_handler = event_handler;
 430         id_priv->id.ps = ps;
 431         id_priv->id.qp_type = qp_type;
 432         spin_lock_init(&id_priv->lock);
 433         mutex_init(&id_priv->qp_mutex);
 434         init_completion(&id_priv->comp);
 435         atomic_set(&id_priv->refcount, 1);
 436         mutex_init(&id_priv->handler_mutex);
 437         INIT_LIST_HEAD(&id_priv->listen_list);
 438         INIT_LIST_HEAD(&id_priv->mc_list);
 439         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 440
 441         return &id_priv->id;
 442 }
 443 EXPORT_SYMBOL(rdma_create_id);
 444
 445 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 446 {
 447         struct ib_qp_attr qp_attr;
 448         int qp_attr_mask, ret;
 449
 450         qp_attr.qp_state = IB_QPS_INIT;
 451         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 452         if (ret)
 453                 return ret;
 454
 455         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 456         if (ret)
 457                 return ret;
 458
 459         qp_attr.qp_state = IB_QPS_RTR;
 460         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 461         if (ret)
 462                 return ret;
 463
 464         qp_attr.qp_state = IB_QPS_RTS;
 465         qp_attr.sq_psn = 0;
 466         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 467
 468         return ret;
 469 }
 470
 471 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 472 {
 473         struct ib_qp_attr qp_attr;
 474         int qp_attr_mask, ret;
 475
 476         qp_attr.qp_state = IB_QPS_INIT;
 477         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 478         if (ret)
 479                 return ret;
 480
 481         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 482 }
 483
 484 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 485                    struct ib_qp_init_attr *qp_init_attr)
 486 {
 487         struct rdma_id_private *id_priv;
 488         struct ib_qp *qp;
 489         int ret;
 490
 491         id_priv = container_of(id, struct rdma_id_private, id);
 492         if (id->device != pd->device)
 493                 return -EINVAL;
 494
 495         qp = ib_create_qp(pd, qp_init_attr);
 496         if (IS_ERR(qp))
 497                 return PTR_ERR(qp);
 498
 499         if (id->qp_type == IB_QPT_UD)
 500                 ret = cma_init_ud_qp(id_priv, qp);
 501         else
 502                 ret = cma_init_conn_qp(id_priv, qp);
 503         if (ret)
 504                 goto err;
 505
 506         id->qp = qp;
 507         id_priv->qp_num = qp->qp_num;
 508         id_priv->srq = (qp->srq != NULL);
 509         return 0;
 510 err:
 511         ib_destroy_qp(qp);
 512         return ret;
 513 }
 514 EXPORT_SYMBOL(rdma_create_qp);
 515
 516 void rdma_destroy_qp(struct rdma_cm_id *id)
 517 {
 518         struct rdma_id_private *id_priv;
 519
 520         id_priv = container_of(id, struct rdma_id_private, id);
 521         mutex_lock(&id_priv->qp_mutex);
 522         ib_destroy_qp(id_priv->id.qp);
 523         id_priv->id.qp = NULL;
 524         mutex_unlock(&id_priv->qp_mutex);
 525 }
 526 EXPORT_SYMBOL(rdma_destroy_qp);
 527
 528 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 529                              struct rdma_conn_param *conn_param)
 530 {
 531         struct ib_qp_attr qp_attr;
 532         int qp_attr_mask, ret;
 533
 534         mutex_lock(&id_priv->qp_mutex);
 535         if (!id_priv->id.qp) {
 536                 ret = 0;
 537                 goto out;
 538         }
 539
 540         /* Need to update QP attributes from default values. */
 541         qp_attr.qp_state = IB_QPS_INIT;
 542         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 543         if (ret)
 544                 goto out;
 545
 546         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 547         if (ret)
 548                 goto out;
 549
 550         qp_attr.qp_state = IB_QPS_RTR;
 551         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 552         if (ret)
 553                 goto out;
 554
 555         if (conn_param)
 556                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 557         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 558 out:
 559         mutex_unlock(&id_priv->qp_mutex);
 560         return ret;
 561 }
 562
 563 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 564                              struct rdma_conn_param *conn_param)
 565 {
 566         struct ib_qp_attr qp_attr;
 567         int qp_attr_mask, ret;
 568
 569         mutex_lock(&id_priv->qp_mutex);
 570         if (!id_priv->id.qp) {
 571                 ret = 0;
 572                 goto out;
 573         }
 574
 575         qp_attr.qp_state = IB_QPS_RTS;
 576         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 577         if (ret)
 578                 goto out;
 579
 580         if (conn_param)
 581                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 582         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 583 out:
 584         mutex_unlock(&id_priv->qp_mutex);
 585         return ret;
 586 }
 587
 588 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 589 {
 590         struct ib_qp_attr qp_attr;
 591         int ret;
 592
 593         mutex_lock(&id_priv->qp_mutex);
 594         if (!id_priv->id.qp) {
 595                 ret = 0;
 596                 goto out;
 597         }
 598
 599         qp_attr.qp_state = IB_QPS_ERR;
 600         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 601 out:
 602         mutex_unlock(&id_priv->qp_mutex);
 603         return ret;
 604 }
 605
 606 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 607                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 608 {
 609         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 610         int ret;
 611         u16 pkey;
 612
 613         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 614             IB_LINK_LAYER_INFINIBAND)
 615                 pkey = ib_addr_get_pkey(dev_addr);
 616         else
 617                 pkey = 0xffff;
 618
 619         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 620                                   pkey, &qp_attr->pkey_index);
 621         if (ret)
 622                 return ret;
 623
 624         qp_attr->port_num = id_priv->id.port_num;
 625         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 626
 627         if (id_priv->id.qp_type == IB_QPT_UD) {
 628                 ret = cma_set_qkey(id_priv);
 629                 if (ret)
 630                         return ret;
 631
 632                 qp_attr->qkey = id_priv->qkey;
 633                 *qp_attr_mask |= IB_QP_QKEY;
 634         } else {
 635                 qp_attr->qp_access_flags = 0;
 636                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 637         }
 638         return 0;
 639 }
 640
 641 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 642                        int *qp_attr_mask)
 643 {
 644         struct rdma_id_private *id_priv;
 645         int ret = 0;
 646
 647         id_priv = container_of(id, struct rdma_id_private, id);
 648         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 649         case RDMA_TRANSPORT_IB:
 650                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 651                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 652                 else
 653                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 654                                                  qp_attr_mask);
 655                 if (qp_attr->qp_state == IB_QPS_RTR)
 656                         qp_attr->rq_psn = id_priv->seq_num;
 657                 break;
 658         case RDMA_TRANSPORT_IWARP:
 659                 if (!id_priv->cm_id.iw) {
 660                         qp_attr->qp_access_flags = 0;
 661                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 662                 } else
 663                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 664                                                  qp_attr_mask);
 665                 break;
 666         default:
 667                 ret = -ENOSYS;
 668                 break;
 669         }
 670
 671         return ret;
 672 }
 673 EXPORT_SYMBOL(rdma_init_qp_attr);
 674
 675 static inline int cma_zero_addr(struct sockaddr *addr)
 676 {
 677         struct in6_addr *ip6;
 678
 679         if (addr->sa_family == AF_INET)
 680                 return ipv4_is_zeronet(
 681                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 682         else {
 683                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 684                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 685                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 686         }
 687 }
 688
 689 static inline int cma_loopback_addr(struct sockaddr *addr)
 690 {
 691         if (addr->sa_family == AF_INET)
 692                 return ipv4_is_loopback(
 693                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 694         else
 695                 return ipv6_addr_loopback(
 696                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 697 }
 698
 699 static inline int cma_any_addr(struct sockaddr *addr)
 700 {
 701         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 702 }
 703
 704 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 705 {
 706         if (src->sa_family != dst->sa_family)
 707                 return -1;
 708
 709         switch (src->sa_family) {
 710         case AF_INET:
 711                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 712                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 713         default:
 714                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 715                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 716         }
 717 }
 718
 719 static inline __be16 cma_port(struct sockaddr *addr)
 720 {
 721         if (addr->sa_family == AF_INET)
 722                 return ((struct sockaddr_in *) addr)->sin_port;
 723         else
 724                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 725 }
 726
 727 static inline int cma_any_port(struct sockaddr *addr)
 728 {
 729         return !cma_port(addr);
 730 }
 731
 732 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 733                             u8 *ip_ver, __be16 *port,
 734                             union cma_ip_addr **src, union cma_ip_addr **dst)
 735 {
 736         switch (ps) {
 737         case RDMA_PS_SDP:
 738                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 739                     SDP_MAJ_VERSION)
 740                         return -EINVAL;
 741
 742                 *ip_ver = sdp_get_ip_ver(hdr);
 743                 *port   = ((struct sdp_hh *) hdr)->port;
 744                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 745                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 746                 break;
 747         default:
 748                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 749                         return -EINVAL;
 750
 751                 *ip_ver = cma_get_ip_ver(hdr);
 752                 *port   = ((struct cma_hdr *) hdr)->port;
 753                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 754                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 755                 break;
 756         }
 757
 758         if (*ip_ver != 4 && *ip_ver != 6)
 759                 return -EINVAL;
 760         return 0;
 761 }
 762
 763 static void cma_save_net_info(struct rdma_addr *addr,
 764                               struct rdma_addr *listen_addr,
 765                               u8 ip_ver, __be16 port,
 766                               union cma_ip_addr *src, union cma_ip_addr *dst)
 767 {
 768         struct sockaddr_in *listen4, *ip4;
 769         struct sockaddr_in6 *listen6, *ip6;
 770
 771         switch (ip_ver) {
 772         case 4:
 773                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 774                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 775                 ip4->sin_family = listen4->sin_family;
 776                 ip4->sin_addr.s_addr = dst->ip4.addr;
 777                 ip4->sin_port = listen4->sin_port;
 778
 779                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 780                 ip4->sin_family = listen4->sin_family;
 781                 ip4->sin_addr.s_addr = src->ip4.addr;
 782                 ip4->sin_port = port;
 783                 break;
 784         case 6:
 785                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 786                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 787                 ip6->sin6_family = listen6->sin6_family;
 788                 ip6->sin6_addr = dst->ip6;
 789                 ip6->sin6_port = listen6->sin6_port;
 790
 791                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 792                 ip6->sin6_family = listen6->sin6_family;
 793                 ip6->sin6_addr = src->ip6;
 794                 ip6->sin6_port = port;
 795                 break;
 796         default:
 797                 break;
 798         }
 799 }
 800
 801 static inline int cma_user_data_offset(enum rdma_port_space ps)
 802 {
 803         switch (ps) {
 804         case RDMA_PS_SDP:
 805                 return 0;
 806         default:
 807                 return sizeof(struct cma_hdr);
 808         }
 809 }
 810
 811 static void cma_cancel_route(struct rdma_id_private *id_priv)
 812 {
 813         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 814         case IB_LINK_LAYER_INFINIBAND:
 815                 if (id_priv->query)
 816                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 817                 break;
 818         default:
 819                 break;
 820         }
 821 }
 822
 823 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 824 {
 825         struct rdma_id_private *dev_id_priv;
 826
 827         /*
 828          * Remove from listen_any_list to prevent added devices from spawning
 829          * additional listen requests.
 830          */
 831         mutex_lock(&lock);
 832         list_del(&id_priv->list);
 833
 834         while (!list_empty(&id_priv->listen_list)) {
 835                 dev_id_priv = list_entry(id_priv->listen_list.next,
 836                                          struct rdma_id_private, listen_list);
 837                 /* sync with device removal to avoid duplicate destruction */
 838                 list_del_init(&dev_id_priv->list);
 839                 list_del(&dev_id_priv->listen_list);
 840                 mutex_unlock(&lock);
 841
 842                 rdma_destroy_id(&dev_id_priv->id);
 843                 mutex_lock(&lock);
 844         }
 845         mutex_unlock(&lock);
 846 }
 847
 848 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 849                                  enum rdma_cm_state state)
 850 {
 851         switch (state) {
 852         case RDMA_CM_ADDR_QUERY:
 853                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 854                 break;
 855         case RDMA_CM_ROUTE_QUERY:
 856                 cma_cancel_route(id_priv);
 857                 break;
 858         case RDMA_CM_LISTEN:
 859                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 860                                 && !id_priv->cma_dev)
 861                         cma_cancel_listens(id_priv);
 862                 break;
 863         default:
 864                 break;
 865         }
 866 }
 867
 868 static void cma_release_port(struct rdma_id_private *id_priv)
 869 {
 870         struct rdma_bind_list *bind_list = id_priv->bind_list;
 871
 872         if (!bind_list)
 873                 return;
 874
 875         mutex_lock(&lock);
 876         hlist_del(&id_priv->node);
 877         if (hlist_empty(&bind_list->owners)) {
 878                 idr_remove(bind_list->ps, bind_list->port);
 879                 kfree(bind_list);
 880         }
 881         mutex_unlock(&lock);
 882 }
 883
 884 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 885 {
 886         struct cma_multicast *mc;
 887
 888         while (!list_empty(&id_priv->mc_list)) {
 889                 mc = container_of(id_priv->mc_list.next,
 890                                   struct cma_multicast, list);
 891                 list_del(&mc->list);
 892                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 893                 case IB_LINK_LAYER_INFINIBAND:
 894                         ib_sa_free_multicast(mc->multicast.ib);
 895                         kfree(mc);
 896                         break;
 897                 case IB_LINK_LAYER_ETHERNET:
 898                         kref_put(&mc->mcref, release_mc);
 899                         break;
 900                 default:
 901                         break;
 902                 }
 903         }
 904 }
 905
 906 void rdma_destroy_id(struct rdma_cm_id *id)
 907 {
 908         struct rdma_id_private *id_priv;
 909         enum rdma_cm_state state;
 910
 911         id_priv = container_of(id, struct rdma_id_private, id);
 912         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 913         cma_cancel_operation(id_priv, state);
 914
 915         /*
 916          * Wait for any active callback to finish.  New callbacks will find
 917          * the id_priv state set to destroying and abort.
 918          */
 919         mutex_lock(&id_priv->handler_mutex);
 920         mutex_unlock(&id_priv->handler_mutex);
 921
 922         if (id_priv->cma_dev) {
 923                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 924                 case RDMA_TRANSPORT_IB:
 925                         if (id_priv->cm_id.ib)
 926                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 927                         break;
 928                 case RDMA_TRANSPORT_IWARP:
 929                         if (id_priv->cm_id.iw)
 930                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 931                         break;
 932                 default:
 933                         break;
 934                 }
 935                 cma_leave_mc_groups(id_priv);
 936                 cma_release_dev(id_priv);
 937         }
 938
 939         cma_release_port(id_priv);
 940         cma_deref_id(id_priv);
 941         wait_for_completion(&id_priv->comp);
 942
 943         if (id_priv->internal_id)
 944                 cma_deref_id(id_priv->id.context);
 945
 946         kfree(id_priv->id.route.path_rec);
 947         kfree(id_priv);
 948 }
 949 EXPORT_SYMBOL(rdma_destroy_id);
 950
 951 static int cma_rep_recv(struct rdma_id_private *id_priv)
 952 {
 953         int ret;
 954
 955         ret = cma_modify_qp_rtr(id_priv, NULL);
 956         if (ret)
 957                 goto reject;
 958
 959         ret = cma_modify_qp_rts(id_priv, NULL);
 960         if (ret)
 961                 goto reject;
 962
 963         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 964         if (ret)
 965                 goto reject;
 966
 967         return 0;
 968 reject:
 969         cma_modify_qp_err(id_priv);
 970         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 971                        NULL, 0, NULL, 0);
 972         return ret;
 973 }
 974
 975 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 976 {
 977         if (id_priv->id.ps == RDMA_PS_SDP &&
 978             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 979             SDP_MAJ_VERSION)
 980                 return -EINVAL;
 981
 982         return 0;
 983 }
 984
 985 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 986                                    struct ib_cm_rep_event_param *rep_data,
 987                                    void *private_data)
 988 {
 989         event->param.conn.private_data = private_data;
 990         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 991         event->param.conn.responder_resources = rep_data->responder_resources;
 992         event->param.conn.initiator_depth = rep_data->initiator_depth;
 993         event->param.conn.flow_control = rep_data->flow_control;
 994         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 995         event->param.conn.srq = rep_data->srq;
 996         event->param.conn.qp_num = rep_data->remote_qpn;
 997 }
 998
 999 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1000 {
1001         struct rdma_id_private *id_priv = cm_id->context;
1002         struct rdma_cm_event event;
1003         int ret = 0;
1004
1005         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1006                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1007             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1008                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1009                 return 0;
1010
1011         memset(&event, 0, sizeof event);
1012         switch (ib_event->event) {
1013         case IB_CM_REQ_ERROR:
1014         case IB_CM_REP_ERROR:
1015                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1016                 event.status = -ETIMEDOUT;
1017                 break;
1018         case IB_CM_REP_RECEIVED:
1019                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1020                 if (event.status)
1021                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1022                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1023                         event.status = cma_rep_recv(id_priv);
1024                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1025                                                      RDMA_CM_EVENT_ESTABLISHED;
1026                 } else
1027                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1028                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1029                                        ib_event->private_data);
1030                 break;
1031         case IB_CM_RTU_RECEIVED:
1032         case IB_CM_USER_ESTABLISHED:
1033                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1034                 break;
1035         case IB_CM_DREQ_ERROR:
1036                 event.status = -ETIMEDOUT; /* fall through */
1037         case IB_CM_DREQ_RECEIVED:
1038         case IB_CM_DREP_RECEIVED:
1039                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1040                                    RDMA_CM_DISCONNECT))
1041                         goto out;
1042                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1043                 break;
1044         case IB_CM_TIMEWAIT_EXIT:
1045                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1046                 break;
1047         case IB_CM_MRA_RECEIVED:
1048                 /* ignore event */
1049                 goto out;
1050         case IB_CM_REJ_RECEIVED:
1051                 cma_modify_qp_err(id_priv);
1052                 event.status = ib_event->param.rej_rcvd.reason;
1053                 event.event = RDMA_CM_EVENT_REJECTED;
1054                 event.param.conn.private_data = ib_event->private_data;
1055                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1056                 break;
1057         default:
1058                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1059                        ib_event->event);
1060                 goto out;
1061         }
1062
1063         ret = id_priv->id.event_handler(&id_priv->id, &event);
1064         if (ret) {
1065                 /* Destroy the CM ID by returning a non-zero value. */
1066                 id_priv->cm_id.ib = NULL;
1067                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1068                 mutex_unlock(&id_priv->handler_mutex);
1069                 rdma_destroy_id(&id_priv->id);
1070                 return ret;
1071         }
1072 out:
1073         mutex_unlock(&id_priv->handler_mutex);
1074         return ret;
1075 }
1076
1077 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1078                                                struct ib_cm_event *ib_event)
1079 {
1080         struct rdma_id_private *id_priv;
1081         struct rdma_cm_id *id;
1082         struct rdma_route *rt;
1083         union cma_ip_addr *src, *dst;
1084         __be16 port;
1085         u8 ip_ver;
1086         int ret;
1087
1088         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1089                              &ip_ver, &port, &src, &dst))
1090                 return NULL;
1091
1092         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1093                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1094         if (IS_ERR(id))
1095                 return NULL;
1096
1097         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1098                           ip_ver, port, src, dst);
1099
1100         rt = &id->route;
1101         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1102         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1103                                GFP_KERNEL);
1104         if (!rt->path_rec)
1105                 goto err;
1106
1107         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1108         if (rt->num_paths == 2)
1109                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1110
1111         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1112                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1113                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1114                 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1115         } else {
1116                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1117                                         &rt->addr.dev_addr);
1118                 if (ret)
1119                         goto err;
1120         }
1121         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1122
1123         id_priv = container_of(id, struct rdma_id_private, id);
1124         id_priv->state = RDMA_CM_CONNECT;
1125         return id_priv;
1126
1127 err:
1128         rdma_destroy_id(id);
1129         return NULL;
1130 }
1131
1132 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1133                                               struct ib_cm_event *ib_event)
1134 {
1135         struct rdma_id_private *id_priv;
1136         struct rdma_cm_id *id;
1137         union cma_ip_addr *src, *dst;
1138         __be16 port;
1139         u8 ip_ver;
1140         int ret;
1141
1142         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1143                             listen_id->ps, IB_QPT_UD);
1144         if (IS_ERR(id))
1145                 return NULL;
1146
1147
1148         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1149                              &ip_ver, &port, &src, &dst))
1150                 goto err;
1151
1152         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1153                           ip_ver, port, src, dst);
1154
1155         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1156                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1157                                         &id->route.addr.dev_addr);
1158                 if (ret)
1159                         goto err;
1160         }
1161
1162         id_priv = container_of(id, struct rdma_id_private, id);
1163         id_priv->state = RDMA_CM_CONNECT;
1164         return id_priv;
1165 err:
1166         rdma_destroy_id(id);
1167         return NULL;
1168 }
1169
1170 static void cma_set_req_event_data(struct rdma_cm_event *event,
1171                                    struct ib_cm_req_event_param *req_data,
1172                                    void *private_data, int offset)
1173 {
1174         event->param.conn.private_data = private_data + offset;
1175         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1176         event->param.conn.responder_resources = req_data->responder_resources;
1177         event->param.conn.initiator_depth = req_data->initiator_depth;
1178         event->param.conn.flow_control = req_data->flow_control;
1179         event->param.conn.retry_count = req_data->retry_count;
1180         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1181         event->param.conn.srq = req_data->srq;
1182         event->param.conn.qp_num = req_data->remote_qpn;
1183 }
1184
1185 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1186 {
1187         return (((ib_event->event == IB_CM_REQ_RECEIVED) ||
1188                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1189                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1190                  (id->qp_type == IB_QPT_UD)) ||
1191                 (!id->qp_type));
1192 }
1193
1194 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1195 {
1196         struct rdma_id_private *listen_id, *conn_id;
1197         struct rdma_cm_event event;
1198         int offset, ret;
1199
1200         listen_id = cm_id->context;
1201         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1202                 return -EINVAL;
1203
1204         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1205                 return -ECONNABORTED;
1206
1207         memset(&event, 0, sizeof event);
1208         offset = cma_user_data_offset(listen_id->id.ps);
1209         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1210         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1211                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1212                 event.param.ud.private_data = ib_event->private_data + offset;
1213                 event.param.ud.private_data_len =
1214                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1215         } else {
1216                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1217                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1218                                        ib_event->private_data, offset);
1219         }
1220         if (!conn_id) {
1221                 ret = -ENOMEM;
1222                 goto out;
1223         }
1224
1225         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1226         ret = cma_acquire_dev(conn_id);
1227         if (ret)
1228                 goto release_conn_id;
1229
1230         conn_id->cm_id.ib = cm_id;
1231         cm_id->context = conn_id;
1232         cm_id->cm_handler = cma_ib_handler;
1233
1234         /*
1235          * Protect against the user destroying conn_id from another thread
1236          * until we're done accessing it.
1237          */
1238         atomic_inc(&conn_id->refcount);
1239         ret = conn_id->id.event_handler(&conn_id->id, &event);
1240         if (!ret) {
1241                 /*
1242                  * Acquire mutex to prevent user executing rdma_destroy_id()
1243                  * while we're accessing the cm_id.
1244                  */
1245                 mutex_lock(&lock);
1246                 if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1247                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1248                 mutex_unlock(&lock);
1249                 mutex_unlock(&conn_id->handler_mutex);
1250                 cma_deref_id(conn_id);
1251                 goto out;
1252         }
1253         cma_deref_id(conn_id);
1254
1255         /* Destroy the CM ID by returning a non-zero value. */
1256         conn_id->cm_id.ib = NULL;
1257
1258 release_conn_id:
1259         cma_exch(conn_id, RDMA_CM_DESTROYING);
1260         mutex_unlock(&conn_id->handler_mutex);
1261         rdma_destroy_id(&conn_id->id);
1262
1263 out:
1264         mutex_unlock(&listen_id->handler_mutex);
1265         return ret;
1266 }
1267
1268 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1269 {
1270         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1271 }
1272
1273 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1274                                  struct ib_cm_compare_data *compare)
1275 {
1276         struct cma_hdr *cma_data, *cma_mask;
1277         struct sdp_hh *sdp_data, *sdp_mask;
1278         __be32 ip4_addr;
1279         struct in6_addr ip6_addr;
1280
1281         memset(compare, 0, sizeof *compare);
1282         cma_data = (void *) compare->data;
1283         cma_mask = (void *) compare->mask;
1284         sdp_data = (void *) compare->data;
1285         sdp_mask = (void *) compare->mask;
1286
1287         switch (addr->sa_family) {
1288         case AF_INET:
1289                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1290                 if (ps == RDMA_PS_SDP) {
1291                         sdp_set_ip_ver(sdp_data, 4);
1292                         sdp_set_ip_ver(sdp_mask, 0xF);
1293                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1294                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1295                 } else {
1296                         cma_set_ip_ver(cma_data, 4);
1297                         cma_set_ip_ver(cma_mask, 0xF);
1298                         cma_data->dst_addr.ip4.addr = ip4_addr;
1299                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1300                 }
1301                 break;
1302         case AF_INET6:
1303                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1304                 if (ps == RDMA_PS_SDP) {
1305                         sdp_set_ip_ver(sdp_data, 6);
1306                         sdp_set_ip_ver(sdp_mask, 0xF);
1307                         sdp_data->dst_addr.ip6 = ip6_addr;
1308                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1309                                sizeof sdp_mask->dst_addr.ip6);
1310                 } else {
1311                         cma_set_ip_ver(cma_data, 6);
1312                         cma_set_ip_ver(cma_mask, 0xF);
1313                         cma_data->dst_addr.ip6 = ip6_addr;
1314                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1315                                sizeof cma_mask->dst_addr.ip6);
1316                 }
1317                 break;
1318         default:
1319                 break;
1320         }
1321 }
1322
1323 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1324 {
1325         struct rdma_id_private *id_priv = iw_id->context;
1326         struct rdma_cm_event event;
1327         struct sockaddr_in *sin;
1328         int ret = 0;
1329
1330         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1331                 return 0;
1332
1333         memset(&event, 0, sizeof event);
1334         switch (iw_event->event) {
1335         case IW_CM_EVENT_CLOSE:
1336                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1337                 break;
1338         case IW_CM_EVENT_CONNECT_REPLY:
1339                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1340                 *sin = iw_event->local_addr;
1341                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1342                 *sin = iw_event->remote_addr;
1343                 switch (iw_event->status) {
1344                 case 0:
1345                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1346                         event.param.conn.initiator_depth = iw_event->ird;
1347                         event.param.conn.responder_resources = iw_event->ord;
1348                         break;
1349                 case -ECONNRESET:
1350                 case -ECONNREFUSED:
1351                         event.event = RDMA_CM_EVENT_REJECTED;
1352                         break;
1353                 case -ETIMEDOUT:
1354                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1355                         break;
1356                 default:
1357                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1358                         break;
1359                 }
1360                 break;
1361         case IW_CM_EVENT_ESTABLISHED:
1362                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1363                 event.param.conn.initiator_depth = iw_event->ird;
1364                 event.param.conn.responder_resources = iw_event->ord;
1365                 break;
1366         default:
1367                 BUG_ON(1);
1368         }
1369
1370         event.status = iw_event->status;
1371         event.param.conn.private_data = iw_event->private_data;
1372         event.param.conn.private_data_len = iw_event->private_data_len;
1373         ret = id_priv->id.event_handler(&id_priv->id, &event);
1374         if (ret) {
1375                 /* Destroy the CM ID by returning a non-zero value. */
1376                 id_priv->cm_id.iw = NULL;
1377                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1378                 mutex_unlock(&id_priv->handler_mutex);
1379                 rdma_destroy_id(&id_priv->id);
1380                 return ret;
1381         }
1382
1383         mutex_unlock(&id_priv->handler_mutex);
1384         return ret;
1385 }
1386
1387 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1388                                struct iw_cm_event *iw_event)
1389 {
1390         struct rdma_cm_id *new_cm_id;
1391         struct rdma_id_private *listen_id, *conn_id;
1392         struct sockaddr_in *sin;
1393         struct net_device *dev = NULL;
1394         struct rdma_cm_event event;
1395         int ret;
1396         struct ib_device_attr attr;
1397
1398         listen_id = cm_id->context;
1399         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1400                 return -ECONNABORTED;
1401
1402         /* Create a new RDMA id for the new IW CM ID */
1403         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1404                                    listen_id->id.context,
1405                                    RDMA_PS_TCP, IB_QPT_RC);
1406         if (IS_ERR(new_cm_id)) {
1407                 ret = -ENOMEM;
1408                 goto out;
1409         }
1410         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1411         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1412         conn_id->state = RDMA_CM_CONNECT;
1413
1414         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1415         if (!dev) {
1416                 ret = -EADDRNOTAVAIL;
1417                 mutex_unlock(&conn_id->handler_mutex);
1418                 rdma_destroy_id(new_cm_id);
1419                 goto out;
1420         }
1421         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1422         if (ret) {
1423                 mutex_unlock(&conn_id->handler_mutex);
1424                 rdma_destroy_id(new_cm_id);
1425                 goto out;
1426         }
1427
1428         ret = cma_acquire_dev(conn_id);
1429         if (ret) {
1430                 mutex_unlock(&conn_id->handler_mutex);
1431                 rdma_destroy_id(new_cm_id);
1432                 goto out;
1433         }
1434
1435         conn_id->cm_id.iw = cm_id;
1436         cm_id->context = conn_id;
1437         cm_id->cm_handler = cma_iw_handler;
1438
1439         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1440         *sin = iw_event->local_addr;
1441         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1442         *sin = iw_event->remote_addr;
1443
1444         ret = ib_query_device(conn_id->id.device, &attr);
1445         if (ret) {
1446                 mutex_unlock(&conn_id->handler_mutex);
1447                 rdma_destroy_id(new_cm_id);
1448                 goto out;
1449         }
1450
1451         memset(&event, 0, sizeof event);
1452         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1453         event.param.conn.private_data = iw_event->private_data;
1454         event.param.conn.private_data_len = iw_event->private_data_len;
1455         event.param.conn.initiator_depth = iw_event->ird;
1456         event.param.conn.responder_resources = iw_event->ord;
1457
1458         /*
1459          * Protect against the user destroying conn_id from another thread
1460          * until we're done accessing it.
1461          */
1462         atomic_inc(&conn_id->refcount);
1463         ret = conn_id->id.event_handler(&conn_id->id, &event);
1464         if (ret) {
1465                 /* User wants to destroy the CM ID */
1466                 conn_id->cm_id.iw = NULL;
1467                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1468                 mutex_unlock(&conn_id->handler_mutex);
1469                 cma_deref_id(conn_id);
1470                 rdma_destroy_id(&conn_id->id);
1471                 goto out;
1472         }
1473
1474         mutex_unlock(&conn_id->handler_mutex);
1475         cma_deref_id(conn_id);
1476
1477 out:
1478         if (dev)
1479                 dev_put(dev);
1480         mutex_unlock(&listen_id->handler_mutex);
1481         return ret;
1482 }
1483
1484 static int cma_ib_listen(struct rdma_id_private *id_priv)
1485 {
1486         struct ib_cm_compare_data compare_data;
1487         struct sockaddr *addr;
1488         struct ib_cm_id *id;
1489         __be64 svc_id;
1490         int ret;
1491
1492         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1493         if (IS_ERR(id))
1494                 return PTR_ERR(id);
1495
1496         id_priv->cm_id.ib = id;
1497
1498         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1499         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1500         if (cma_any_addr(addr))
1501                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1502         else {
1503                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1504                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1505         }
1506
1507         if (ret) {
1508                 ib_destroy_cm_id(id_priv->cm_id.ib);
1509                 id_priv->cm_id.ib = NULL;
1510         }
1511
1512         return ret;
1513 }
1514
1515 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1516 {
1517         int ret;
1518         struct sockaddr_in *sin;
1519         struct iw_cm_id *id;
1520
1521         id = iw_create_cm_id(id_priv->id.device,
1522                              iw_conn_req_handler,
1523                              id_priv);
1524         if (IS_ERR(id))
1525                 return PTR_ERR(id);
1526
1527         id_priv->cm_id.iw = id;
1528
1529         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1530         id_priv->cm_id.iw->local_addr = *sin;
1531
1532         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1533
1534         if (ret) {
1535                 iw_destroy_cm_id(id_priv->cm_id.iw);
1536                 id_priv->cm_id.iw = NULL;
1537         }
1538
1539         return ret;
1540 }
1541
1542 static int cma_listen_handler(struct rdma_cm_id *id,
1543                               struct rdma_cm_event *event)
1544 {
1545         struct rdma_id_private *id_priv = id->context;
1546
1547         id->context = id_priv->id.context;
1548         id->event_handler = id_priv->id.event_handler;
1549         return id_priv->id.event_handler(id, event);
1550 }
1551
1552 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1553                               struct cma_device *cma_dev)
1554 {
1555         struct rdma_id_private *dev_id_priv;
1556         struct rdma_cm_id *id;
1557         int ret;
1558
1559         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1560                             id_priv->id.qp_type);
1561         if (IS_ERR(id))
1562                 return;
1563
1564         dev_id_priv = container_of(id, struct rdma_id_private, id);
1565
1566         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1567         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1568                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1569
1570         cma_attach_to_dev(dev_id_priv, cma_dev);
1571         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1572         atomic_inc(&id_priv->refcount);
1573         dev_id_priv->internal_id = 1;
1574
1575         ret = rdma_listen(id, id_priv->backlog);
1576         if (ret)
1577                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1578                        "listening on device %s\n", ret, cma_dev->device->name);
1579 }
1580
1581 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1582 {
1583         struct cma_device *cma_dev;
1584
1585         mutex_lock(&lock);
1586         list_add_tail(&id_priv->list, &listen_any_list);
1587         list_for_each_entry(cma_dev, &dev_list, list)
1588                 cma_listen_on_dev(id_priv, cma_dev);
1589         mutex_unlock(&lock);
1590 }
1591
1592 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1593 {
1594         struct rdma_id_private *id_priv;
1595
1596         id_priv = container_of(id, struct rdma_id_private, id);
1597         id_priv->tos = (u8) tos;
1598 }
1599 EXPORT_SYMBOL(rdma_set_service_type);
1600
1601 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1602                               void *context)
1603 {
1604         struct cma_work *work = context;
1605         struct rdma_route *route;
1606
1607         route = &work->id->id.route;
1608
1609         if (!status) {
1610                 route->num_paths = 1;
1611                 *route->path_rec = *path_rec;
1612         } else {
1613                 work->old_state = RDMA_CM_ROUTE_QUERY;
1614                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1615                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1616                 work->event.status = status;
1617         }
1618
1619         queue_work(cma_wq, &work->work);
1620 }
1621
1622 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1623                               struct cma_work *work)
1624 {
1625         struct rdma_addr *addr = &id_priv->id.route.addr;
1626         struct ib_sa_path_rec path_rec;
1627         ib_sa_comp_mask comp_mask;
1628         struct sockaddr_in6 *sin6;
1629
1630         memset(&path_rec, 0, sizeof path_rec);
1631         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1632         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1633         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1634         path_rec.numb_path = 1;
1635         path_rec.reversible = 1;
1636         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1637                                                         (struct sockaddr *) &addr->dst_addr);
1638
1639         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1640                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1641                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1642
1643         if (addr->src_addr.ss_family == AF_INET) {
1644                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1645                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1646         } else {
1647                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1648                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1649                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1650         }
1651
1652         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1653                                                id_priv->id.port_num, &path_rec,
1654                                                comp_mask, timeout_ms,
1655                                                GFP_KERNEL, cma_query_handler,
1656                                                work, &id_priv->query);
1657
1658         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1659 }
1660
1661 static void cma_work_handler(struct work_struct *_work)
1662 {
1663         struct cma_work *work = container_of(_work, struct cma_work, work);
1664         struct rdma_id_private *id_priv = work->id;
1665         int destroy = 0;
1666
1667         mutex_lock(&id_priv->handler_mutex);
1668         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1669                 goto out;
1670
1671         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1672                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1673                 destroy = 1;
1674         }
1675 out:
1676         mutex_unlock(&id_priv->handler_mutex);
1677         cma_deref_id(id_priv);
1678         if (destroy)
1679                 rdma_destroy_id(&id_priv->id);
1680         kfree(work);
1681 }
1682
1683 static void cma_ndev_work_handler(struct work_struct *_work)
1684 {
1685         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1686         struct rdma_id_private *id_priv = work->id;
1687         int destroy = 0;
1688
1689         mutex_lock(&id_priv->handler_mutex);
1690         if (id_priv->state == RDMA_CM_DESTROYING ||
1691             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1692                 goto out;
1693
1694         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1695                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1696                 destroy = 1;
1697         }
1698
1699 out:
1700         mutex_unlock(&id_priv->handler_mutex);
1701         cma_deref_id(id_priv);
1702         if (destroy)
1703                 rdma_destroy_id(&id_priv->id);
1704         kfree(work);
1705 }
1706
1707 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1708 {
1709         struct rdma_route *route = &id_priv->id.route;
1710         struct cma_work *work;
1711         int ret;
1712
1713         work = kzalloc(sizeof *work, GFP_KERNEL);
1714         if (!work)
1715                 return -ENOMEM;
1716
1717         work->id = id_priv;
1718         INIT_WORK(&work->work, cma_work_handler);
1719         work->old_state = RDMA_CM_ROUTE_QUERY;
1720         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1721         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1722
1723         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1724         if (!route->path_rec) {
1725                 ret = -ENOMEM;
1726                 goto err1;
1727         }
1728
1729         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1730         if (ret)
1731                 goto err2;
1732
1733         return 0;
1734 err2:
1735         kfree(route->path_rec);
1736         route->path_rec = NULL;
1737 err1:
1738         kfree(work);
1739         return ret;
1740 }
1741
1742 int rdma_set_ib_paths(struct rdma_cm_id *id,
1743                       struct ib_sa_path_rec *path_rec, int num_paths)
1744 {
1745         struct rdma_id_private *id_priv;
1746         int ret;
1747
1748         id_priv = container_of(id, struct rdma_id_private, id);
1749         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1750                            RDMA_CM_ROUTE_RESOLVED))
1751                 return -EINVAL;
1752
1753         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1754                                      GFP_KERNEL);
1755         if (!id->route.path_rec) {
1756                 ret = -ENOMEM;
1757                 goto err;
1758         }
1759
1760         id->route.num_paths = num_paths;
1761         return 0;
1762 err:
1763         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1764         return ret;
1765 }
1766 EXPORT_SYMBOL(rdma_set_ib_paths);
1767
1768 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1769 {
1770         struct cma_work *work;
1771
1772         work = kzalloc(sizeof *work, GFP_KERNEL);
1773         if (!work)
1774                 return -ENOMEM;
1775
1776         work->id = id_priv;
1777         INIT_WORK(&work->work, cma_work_handler);
1778         work->old_state = RDMA_CM_ROUTE_QUERY;
1779         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1780         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1781         queue_work(cma_wq, &work->work);
1782         return 0;
1783 }
1784
1785 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1786 {
1787         struct rdma_route *route = &id_priv->id.route;
1788         struct rdma_addr *addr = &route->addr;
1789         struct cma_work *work;
1790         int ret;
1791         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1792         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1793         struct net_device *ndev = NULL;
1794         u16 vid;
1795
1796         if (src_addr->sin_family != dst_addr->sin_family)
1797                 return -EINVAL;
1798
1799         work = kzalloc(sizeof *work, GFP_KERNEL);
1800         if (!work)
1801                 return -ENOMEM;
1802
1803         work->id = id_priv;
1804         INIT_WORK(&work->work, cma_work_handler);
1805
1806         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1807         if (!route->path_rec) {
1808                 ret = -ENOMEM;
1809                 goto err1;
1810         }
1811
1812         route->num_paths = 1;
1813
1814         if (addr->dev_addr.bound_dev_if)
1815                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1816         if (!ndev) {
1817                 ret = -ENODEV;
1818                 goto err2;
1819         }
1820
1821         vid = rdma_vlan_dev_vlan_id(ndev);
1822
1823         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1824         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1825
1826         route->path_rec->hop_limit = 1;
1827         route->path_rec->reversible = 1;
1828         route->path_rec->pkey = cpu_to_be16(0xffff);
1829         route->path_rec->mtu_selector = IB_SA_EQ;
1830         route->path_rec->sl = netdev_get_prio_tc_map(
1831                         ndev->priv_flags & IFF_802_1Q_VLAN ?
1832                                 vlan_dev_real_dev(ndev) : ndev,
1833                         rt_tos2priority(id_priv->tos));
1834
1835         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1836         route->path_rec->rate_selector = IB_SA_EQ;
1837         route->path_rec->rate = iboe_get_rate(ndev);
1838         dev_put(ndev);
1839         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1840         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1841         if (!route->path_rec->mtu) {
1842                 ret = -EINVAL;
1843                 goto err2;
1844         }
1845
1846         work->old_state = RDMA_CM_ROUTE_QUERY;
1847         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1848         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1849         work->event.status = 0;
1850
1851         queue_work(cma_wq, &work->work);
1852
1853         return 0;
1854
1855 err2:
1856         kfree(route->path_rec);
1857         route->path_rec = NULL;
1858 err1:
1859         kfree(work);
1860         return ret;
1861 }
1862
1863 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1864 {
1865         struct rdma_id_private *id_priv;
1866         int ret;
1867
1868         id_priv = container_of(id, struct rdma_id_private, id);
1869         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1870                 return -EINVAL;
1871
1872         atomic_inc(&id_priv->refcount);
1873         switch (rdma_node_get_transport(id->device->node_type)) {
1874         case RDMA_TRANSPORT_IB:
1875                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1876                 case IB_LINK_LAYER_INFINIBAND:
1877                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1878                         break;
1879                 case IB_LINK_LAYER_ETHERNET:
1880                         ret = cma_resolve_iboe_route(id_priv);
1881                         break;
1882                 default:
1883                         ret = -ENOSYS;
1884                 }
1885                 break;
1886         case RDMA_TRANSPORT_IWARP:
1887                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1888                 break;
1889         default:
1890                 ret = -ENOSYS;
1891                 break;
1892         }
1893         if (ret)
1894                 goto err;
1895
1896         return 0;
1897 err:
1898         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1899         cma_deref_id(id_priv);
1900         return ret;
1901 }
1902 EXPORT_SYMBOL(rdma_resolve_route);
1903
1904 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1905 {
1906         struct cma_device *cma_dev;
1907         struct ib_port_attr port_attr;
1908         union ib_gid gid;
1909         u16 pkey;
1910         int ret;
1911         u8 p;
1912
1913         mutex_lock(&lock);
1914         if (list_empty(&dev_list)) {
1915                 ret = -ENODEV;
1916                 goto out;
1917         }
1918         list_for_each_entry(cma_dev, &dev_list, list)
1919                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1920                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1921                             port_attr.state == IB_PORT_ACTIVE)
1922                                 goto port_found;
1923
1924         p = 1;
1925         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1926
1927 port_found:
1928         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1929         if (ret)
1930                 goto out;
1931
1932         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1933         if (ret)
1934                 goto out;
1935
1936         id_priv->id.route.addr.dev_addr.dev_type =
1937                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1938                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1939
1940         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1941         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1942         id_priv->id.port_num = p;
1943         cma_attach_to_dev(id_priv, cma_dev);
1944 out:
1945         mutex_unlock(&lock);
1946         return ret;
1947 }
1948
1949 static void addr_handler(int status, struct sockaddr *src_addr,
1950                          struct rdma_dev_addr *dev_addr, void *context)
1951 {
1952         struct rdma_id_private *id_priv = context;
1953         struct rdma_cm_event event;
1954
1955         memset(&event, 0, sizeof event);
1956         mutex_lock(&id_priv->handler_mutex);
1957         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1958                            RDMA_CM_ADDR_RESOLVED))
1959                 goto out;
1960
1961         if (!status && !id_priv->cma_dev)
1962                 status = cma_acquire_dev(id_priv);
1963
1964         if (status) {
1965                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1966                                    RDMA_CM_ADDR_BOUND))
1967                         goto out;
1968                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1969                 event.status = status;
1970         } else {
1971                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1972                        ip_addr_size(src_addr));
1973                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1974         }
1975
1976         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1977                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1978                 mutex_unlock(&id_priv->handler_mutex);
1979                 cma_deref_id(id_priv);
1980                 rdma_destroy_id(&id_priv->id);
1981                 return;
1982         }
1983 out:
1984         mutex_unlock(&id_priv->handler_mutex);
1985         cma_deref_id(id_priv);
1986 }
1987
1988 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1989 {
1990         struct cma_work *work;
1991         struct sockaddr *src, *dst;
1992         union ib_gid gid;
1993         int ret;
1994
1995         work = kzalloc(sizeof *work, GFP_KERNEL);
1996         if (!work)
1997                 return -ENOMEM;
1998
1999         if (!id_priv->cma_dev) {
2000                 ret = cma_bind_loopback(id_priv);
2001                 if (ret)
2002                         goto err;
2003         }
2004
2005         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2006         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2007
2008         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2009         if (cma_zero_addr(src)) {
2010                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2011                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2012                         ((struct sockaddr_in *)src)->sin_addr =
2013                                 ((struct sockaddr_in *)dst)->sin_addr;
2014                 } else {
2015                         ((struct sockaddr_in6 *)src)->sin6_addr =
2016                                 ((struct sockaddr_in6 *)dst)->sin6_addr;
2017                 }
2018         }
2019
2020         work->id = id_priv;
2021         INIT_WORK(&work->work, cma_work_handler);
2022         work->old_state = RDMA_CM_ADDR_QUERY;
2023         work->new_state = RDMA_CM_ADDR_RESOLVED;
2024         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2025         queue_work(cma_wq, &work->work);
2026         return 0;
2027 err:
2028         kfree(work);
2029         return ret;
2030 }
2031
2032 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2033                          struct sockaddr *dst_addr)
2034 {
2035         if (!src_addr || !src_addr->sa_family) {
2036                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2037                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2038                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2039                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2040                 }
2041         }
2042         return rdma_bind_addr(id, src_addr);
2043 }
2044
2045 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2046                       struct sockaddr *dst_addr, int timeout_ms)
2047 {
2048         struct rdma_id_private *id_priv;
2049         int ret;
2050
2051         id_priv = container_of(id, struct rdma_id_private, id);
2052         if (id_priv->state == RDMA_CM_IDLE) {
2053                 ret = cma_bind_addr(id, src_addr, dst_addr);
2054                 if (ret)
2055                         return ret;
2056         }
2057
2058         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2059                 return -EINVAL;
2060
2061         atomic_inc(&id_priv->refcount);
2062         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2063         if (cma_any_addr(dst_addr))
2064                 ret = cma_resolve_loopback(id_priv);
2065         else
2066                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2067                                       dst_addr, &id->route.addr.dev_addr,
2068                                       timeout_ms, addr_handler, id_priv);
2069         if (ret)
2070                 goto err;
2071
2072         return 0;
2073 err:
2074         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2075         cma_deref_id(id_priv);
2076         return ret;
2077 }
2078 EXPORT_SYMBOL(rdma_resolve_addr);
2079
2080 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2081 {
2082         struct rdma_id_private *id_priv;
2083         unsigned long flags;
2084         int ret;
2085
2086         id_priv = container_of(id, struct rdma_id_private, id);
2087         spin_lock_irqsave(&id_priv->lock, flags);
2088         if (id_priv->state == RDMA_CM_IDLE) {
2089                 id_priv->reuseaddr = reuse;
2090                 ret = 0;
2091         } else {
2092                 ret = -EINVAL;
2093         }
2094         spin_unlock_irqrestore(&id_priv->lock, flags);
2095         return ret;
2096 }
2097 EXPORT_SYMBOL(rdma_set_reuseaddr);
2098
2099 static void cma_bind_port(struct rdma_bind_list *bind_list,
2100                           struct rdma_id_private *id_priv)
2101 {
2102         struct sockaddr_in *sin;
2103
2104         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2105         sin->sin_port = htons(bind_list->port);
2106         id_priv->bind_list = bind_list;
2107         hlist_add_head(&id_priv->node, &bind_list->owners);
2108 }
2109
2110 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2111                           unsigned short snum)
2112 {
2113         struct rdma_bind_list *bind_list;
2114         int port, ret;
2115
2116         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2117         if (!bind_list)
2118                 return -ENOMEM;
2119
2120         do {
2121                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2122         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2123
2124         if (ret)
2125                 goto err1;
2126
2127         if (port != snum) {
2128                 ret = -EADDRNOTAVAIL;
2129                 goto err2;
2130         }
2131
2132         bind_list->ps = ps;
2133         bind_list->port = (unsigned short) port;
2134         cma_bind_port(bind_list, id_priv);
2135         return 0;
2136 err2:
2137         idr_remove(ps, port);
2138 err1:
2139         kfree(bind_list);
2140         return ret;
2141 }
2142
2143 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2144 {
2145         static unsigned int last_used_port;
2146         int low, high, remaining;
2147         unsigned int rover;
2148
2149         inet_get_local_port_range(&low, &high);
2150         remaining = (high - low) + 1;
2151         rover = net_random() % remaining + low;
2152 retry:
2153         if (last_used_port != rover &&
2154             !idr_find(ps, (unsigned short) rover)) {
2155                 int ret = cma_alloc_port(ps, id_priv, rover);
2156                 /*
2157                  * Remember previously used port number in order to avoid
2158                  * re-using same port immediately after it is closed.
2159                  */
2160                 if (!ret)
2161                         last_used_port = rover;
2162                 if (ret != -EADDRNOTAVAIL)
2163                         return ret;
2164         }
2165         if (--remaining) {
2166                 rover++;
2167                 if ((rover < low) || (rover > high))
2168                         rover = low;
2169                 goto retry;
2170         }
2171         return -EADDRNOTAVAIL;
2172 }
2173
2174 /*
2175  * Check that the requested port is available.  This is called when trying to
2176  * bind to a specific port, or when trying to listen on a bound port.  In
2177  * the latter case, the provided id_priv may already be on the bind_list, but
2178  * we still need to check that it's okay to start listening.
2179  */
2180 static int cma_check_port(struct rdma_bind_list *bind_list,
2181                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2182 {
2183         struct rdma_id_private *cur_id;
2184         struct sockaddr *addr, *cur_addr;
2185         struct hlist_node *node;
2186
2187         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2188         if (cma_any_addr(addr) && !reuseaddr)
2189                 return -EADDRNOTAVAIL;
2190
2191         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2192                 if (id_priv == cur_id)
2193                         continue;
2194
2195                 if ((cur_id->state == RDMA_CM_LISTEN) ||
2196                     !reuseaddr || !cur_id->reuseaddr) {
2197                         cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2198                         if (cma_any_addr(cur_addr))
2199                                 return -EADDRNOTAVAIL;
2200
2201                         if (!cma_addr_cmp(addr, cur_addr))
2202                                 return -EADDRINUSE;
2203                 }
2204         }
2205         return 0;
2206 }
2207
2208 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2209 {
2210         struct rdma_bind_list *bind_list;
2211         unsigned short snum;
2212         int ret;
2213
2214         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2215         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2216                 return -EACCES;
2217
2218         bind_list = idr_find(ps, snum);
2219         if (!bind_list) {
2220                 ret = cma_alloc_port(ps, id_priv, snum);
2221         } else {
2222                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2223                 if (!ret)
2224                         cma_bind_port(bind_list, id_priv);
2225         }
2226         return ret;
2227 }
2228
2229 static int cma_bind_listen(struct rdma_id_private *id_priv)
2230 {
2231         struct rdma_bind_list *bind_list = id_priv->bind_list;
2232         int ret = 0;
2233
2234         mutex_lock(&lock);
2235         if (bind_list->owners.first->next)
2236                 ret = cma_check_port(bind_list, id_priv, 0);
2237         mutex_unlock(&lock);
2238         return ret;
2239 }
2240
2241 static int cma_get_port(struct rdma_id_private *id_priv)
2242 {
2243         struct idr *ps;
2244         int ret;
2245
2246         switch (id_priv->id.ps) {
2247         case RDMA_PS_SDP:
2248                 ps = &sdp_ps;
2249                 break;
2250         case RDMA_PS_TCP:
2251                 ps = &tcp_ps;
2252                 break;
2253         case RDMA_PS_UDP:
2254                 ps = &udp_ps;
2255                 break;
2256         case RDMA_PS_IPOIB:
2257                 ps = &ipoib_ps;
2258                 break;
2259         case RDMA_PS_IB:
2260                 ps = &ib_ps;
2261                 break;
2262         default:
2263                 return -EPROTONOSUPPORT;
2264         }
2265
2266         mutex_lock(&lock);
2267         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2268                 ret = cma_alloc_any_port(ps, id_priv);
2269         else
2270                 ret = cma_use_port(ps, id_priv);
2271         mutex_unlock(&lock);
2272
2273         return ret;
2274 }
2275
2276 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2277                                struct sockaddr *addr)
2278 {
2279 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2280         struct sockaddr_in6 *sin6;
2281
2282         if (addr->sa_family != AF_INET6)
2283                 return 0;
2284
2285         sin6 = (struct sockaddr_in6 *) addr;
2286         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2287             !sin6->sin6_scope_id)
2288                         return -EINVAL;
2289
2290         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2291 #endif
2292         return 0;
2293 }
2294
2295 int rdma_listen(struct rdma_cm_id *id, int backlog)
2296 {
2297         struct rdma_id_private *id_priv;
2298         int ret;
2299
2300         id_priv = container_of(id, struct rdma_id_private, id);
2301         if (id_priv->state == RDMA_CM_IDLE) {
2302                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2303                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2304                 if (ret)
2305                         return ret;
2306         }
2307
2308         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2309                 return -EINVAL;
2310
2311         if (id_priv->reuseaddr) {
2312                 ret = cma_bind_listen(id_priv);
2313                 if (ret)
2314                         goto err;
2315         }
2316
2317         id_priv->backlog = backlog;
2318         if (id->device) {
2319                 switch (rdma_node_get_transport(id->device->node_type)) {
2320                 case RDMA_TRANSPORT_IB:
2321                         ret = cma_ib_listen(id_priv);
2322                         if (ret)
2323                                 goto err;
2324                         break;
2325                 case RDMA_TRANSPORT_IWARP:
2326                         ret = cma_iw_listen(id_priv, backlog);
2327                         if (ret)
2328                                 goto err;
2329                         break;
2330                 default:
2331                         ret = -ENOSYS;
2332                         goto err;
2333                 }
2334         } else
2335                 cma_listen_on_all(id_priv);
2336
2337         return 0;
2338 err:
2339         id_priv->backlog = 0;
2340         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2341         return ret;
2342 }
2343 EXPORT_SYMBOL(rdma_listen);
2344
2345 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2346 {
2347         struct rdma_id_private *id_priv;
2348         int ret;
2349
2350         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2351                 return -EAFNOSUPPORT;
2352
2353         id_priv = container_of(id, struct rdma_id_private, id);
2354         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2355                 return -EINVAL;
2356
2357         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2358         if (ret)
2359                 goto err1;
2360
2361         if (!cma_any_addr(addr)) {
2362                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2363                 if (ret)
2364                         goto err1;
2365
2366                 ret = cma_acquire_dev(id_priv);
2367                 if (ret)
2368                         goto err1;
2369         }
2370
2371         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2372         ret = cma_get_port(id_priv);
2373         if (ret)
2374                 goto err2;
2375
2376         return 0;
2377 err2:
2378         if (id_priv->cma_dev)
2379                 cma_release_dev(id_priv);
2380 err1:
2381         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2382         return ret;
2383 }
2384 EXPORT_SYMBOL(rdma_bind_addr);
2385
2386 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2387                           struct rdma_route *route)
2388 {
2389         struct cma_hdr *cma_hdr;
2390         struct sdp_hh *sdp_hdr;
2391
2392         if (route->addr.src_addr.ss_family == AF_INET) {
2393                 struct sockaddr_in *src4, *dst4;
2394
2395                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2396                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2397
2398                 switch (ps) {
2399                 case RDMA_PS_SDP:
2400                         sdp_hdr = hdr;
2401                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2402                                 return -EINVAL;
2403                         sdp_set_ip_ver(sdp_hdr, 4);
2404                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2405                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2406                         sdp_hdr->port = src4->sin_port;
2407                         break;
2408                 default:
2409                         cma_hdr = hdr;
2410                         cma_hdr->cma_version = CMA_VERSION;
2411                         cma_set_ip_ver(cma_hdr, 4);
2412                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2413                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2414                         cma_hdr->port = src4->sin_port;
2415                         break;
2416                 }
2417         } else {
2418                 struct sockaddr_in6 *src6, *dst6;
2419
2420                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2421                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2422
2423                 switch (ps) {
2424                 case RDMA_PS_SDP:
2425                         sdp_hdr = hdr;
2426                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2427                                 return -EINVAL;
2428                         sdp_set_ip_ver(sdp_hdr, 6);
2429                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2430                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2431                         sdp_hdr->port = src6->sin6_port;
2432                         break;
2433                 default:
2434                         cma_hdr = hdr;
2435                         cma_hdr->cma_version = CMA_VERSION;
2436                         cma_set_ip_ver(cma_hdr, 6);
2437                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2438                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2439                         cma_hdr->port = src6->sin6_port;
2440                         break;
2441                 }
2442         }
2443         return 0;
2444 }
2445
2446 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2447                                 struct ib_cm_event *ib_event)
2448 {
2449         struct rdma_id_private *id_priv = cm_id->context;
2450         struct rdma_cm_event event;
2451         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2452         int ret = 0;
2453
2454         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2455                 return 0;
2456
2457         memset(&event, 0, sizeof event);
2458         switch (ib_event->event) {
2459         case IB_CM_SIDR_REQ_ERROR:
2460                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2461                 event.status = -ETIMEDOUT;
2462                 break;
2463         case IB_CM_SIDR_REP_RECEIVED:
2464                 event.param.ud.private_data = ib_event->private_data;
2465                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2466                 if (rep->status != IB_SIDR_SUCCESS) {
2467                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2468                         event.status = ib_event->param.sidr_rep_rcvd.status;
2469                         break;
2470                 }
2471                 ret = cma_set_qkey(id_priv);
2472                 if (ret) {
2473                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2474                         event.status = -EINVAL;
2475                         break;
2476                 }
2477                 if (id_priv->qkey != rep->qkey) {
2478                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2479                         event.status = -EINVAL;
2480                         break;
2481                 }
2482                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2483                                      id_priv->id.route.path_rec,
2484                                      &event.param.ud.ah_attr);
2485                 event.param.ud.qp_num = rep->qpn;
2486                 event.param.ud.qkey = rep->qkey;
2487                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2488                 event.status = 0;
2489                 break;
2490         default:
2491                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2492                        ib_event->event);
2493                 goto out;
2494         }
2495
2496         ret = id_priv->id.event_handler(&id_priv->id, &event);
2497         if (ret) {
2498                 /* Destroy the CM ID by returning a non-zero value. */
2499                 id_priv->cm_id.ib = NULL;
2500                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2501                 mutex_unlock(&id_priv->handler_mutex);
2502                 rdma_destroy_id(&id_priv->id);
2503                 return ret;
2504         }
2505 out:
2506         mutex_unlock(&id_priv->handler_mutex);
2507         return ret;
2508 }
2509
2510 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2511                               struct rdma_conn_param *conn_param)
2512 {
2513         struct ib_cm_sidr_req_param req;
2514         struct rdma_route *route;
2515         struct ib_cm_id *id;
2516         int ret;
2517
2518         req.private_data_len = sizeof(struct cma_hdr) +
2519                                conn_param->private_data_len;
2520         if (req.private_data_len < conn_param->private_data_len)
2521                 return -EINVAL;
2522
2523         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2524         if (!req.private_data)
2525                 return -ENOMEM;
2526
2527         if (conn_param->private_data && conn_param->private_data_len)
2528                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2529                        conn_param->private_data, conn_param->private_data_len);
2530
2531         route = &id_priv->id.route;
2532         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2533         if (ret)
2534                 goto out;
2535
2536         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2537                              id_priv);
2538         if (IS_ERR(id)) {
2539                 ret = PTR_ERR(id);
2540                 goto out;
2541         }
2542         id_priv->cm_id.ib = id;
2543
2544         req.path = route->path_rec;
2545         req.service_id = cma_get_service_id(id_priv->id.ps,
2546                                             (struct sockaddr *) &route->addr.dst_addr);
2547         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2548         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2549
2550         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2551         if (ret) {
2552                 ib_destroy_cm_id(id_priv->cm_id.ib);
2553                 id_priv->cm_id.ib = NULL;
2554         }
2555 out:
2556         kfree(req.private_data);
2557         return ret;
2558 }
2559
2560 static int cma_connect_ib(struct rdma_id_private *id_priv,
2561                           struct rdma_conn_param *conn_param)
2562 {
2563         struct ib_cm_req_param req;
2564         struct rdma_route *route;
2565         void *private_data;
2566         struct ib_cm_id *id;
2567         int offset, ret;
2568
2569         memset(&req, 0, sizeof req);
2570         offset = cma_user_data_offset(id_priv->id.ps);
2571         req.private_data_len = offset + conn_param->private_data_len;
2572         if (req.private_data_len < conn_param->private_data_len)
2573                 return -EINVAL;
2574
2575         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2576         if (!private_data)
2577                 return -ENOMEM;
2578
2579         if (conn_param->private_data && conn_param->private_data_len)
2580                 memcpy(private_data + offset, conn_param->private_data,
2581                        conn_param->private_data_len);
2582
2583         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2584         if (IS_ERR(id)) {
2585                 ret = PTR_ERR(id);
2586                 goto out;
2587         }
2588         id_priv->cm_id.ib = id;
2589
2590         route = &id_priv->id.route;
2591         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2592         if (ret)
2593                 goto out;
2594         req.private_data = private_data;
2595
2596         req.primary_path = &route->path_rec[0];
2597         if (route->num_paths == 2)
2598                 req.alternate_path = &route->path_rec[1];
2599
2600         req.service_id = cma_get_service_id(id_priv->id.ps,
2601                                             (struct sockaddr *) &route->addr.dst_addr);
2602         req.qp_num = id_priv->qp_num;
2603         req.qp_type = id_priv->id.qp_type;
2604         req.starting_psn = id_priv->seq_num;
2605         req.responder_resources = conn_param->responder_resources;
2606         req.initiator_depth = conn_param->initiator_depth;
2607         req.flow_control = conn_param->flow_control;
2608         req.retry_count = conn_param->retry_count;
2609         req.rnr_retry_count = conn_param->rnr_retry_count;
2610         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2611         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2612         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2613         req.srq = id_priv->srq ? 1 : 0;
2614
2615         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2616 out:
2617         if (ret && !IS_ERR(id)) {
2618                 ib_destroy_cm_id(id);
2619                 id_priv->cm_id.ib = NULL;
2620         }
2621
2622         kfree(private_data);
2623         return ret;
2624 }
2625
2626 static int cma_connect_iw(struct rdma_id_private *id_priv,
2627                           struct rdma_conn_param *conn_param)
2628 {
2629         struct iw_cm_id *cm_id;
2630         struct sockaddr_in* sin;
2631         int ret;
2632         struct iw_cm_conn_param iw_param;
2633
2634         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2635         if (IS_ERR(cm_id))
2636                 return PTR_ERR(cm_id);
2637
2638         id_priv->cm_id.iw = cm_id;
2639
2640         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2641         cm_id->local_addr = *sin;
2642
2643         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2644         cm_id->remote_addr = *sin;
2645
2646         ret = cma_modify_qp_rtr(id_priv, conn_param);
2647         if (ret)
2648                 goto out;
2649
2650         if (conn_param) {
2651                 iw_param.ord = conn_param->initiator_depth;
2652                 iw_param.ird = conn_param->responder_resources;
2653                 iw_param.private_data = conn_param->private_data;
2654                 iw_param.private_data_len = conn_param->private_data_len;
2655                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2656         } else {
2657                 memset(&iw_param, 0, sizeof iw_param);
2658                 iw_param.qpn = id_priv->qp_num;
2659         }
2660         ret = iw_cm_connect(cm_id, &iw_param);
2661 out:
2662         if (ret) {
2663                 iw_destroy_cm_id(cm_id);
2664                 id_priv->cm_id.iw = NULL;
2665         }
2666         return ret;
2667 }
2668
2669 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2670 {
2671         struct rdma_id_private *id_priv;
2672         int ret;
2673
2674         id_priv = container_of(id, struct rdma_id_private, id);
2675         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2676                 return -EINVAL;
2677
2678         if (!id->qp) {
2679                 id_priv->qp_num = conn_param->qp_num;
2680                 id_priv->srq = conn_param->srq;
2681         }
2682
2683         switch (rdma_node_get_transport(id->device->node_type)) {
2684         case RDMA_TRANSPORT_IB:
2685                 if (id->qp_type == IB_QPT_UD)
2686                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2687                 else
2688                         ret = cma_connect_ib(id_priv, conn_param);
2689                 break;
2690         case RDMA_TRANSPORT_IWARP:
2691                 ret = cma_connect_iw(id_priv, conn_param);
2692                 break;
2693         default:
2694                 ret = -ENOSYS;
2695                 break;
2696         }
2697         if (ret)
2698                 goto err;
2699
2700         return 0;
2701 err:
2702         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2703         return ret;
2704 }
2705 EXPORT_SYMBOL(rdma_connect);
2706
2707 static int cma_accept_ib(struct rdma_id_private *id_priv,
2708                          struct rdma_conn_param *conn_param)
2709 {
2710         struct ib_cm_rep_param rep;
2711         int ret;
2712
2713         ret = cma_modify_qp_rtr(id_priv, conn_param);
2714         if (ret)
2715                 goto out;
2716
2717         ret = cma_modify_qp_rts(id_priv, conn_param);
2718         if (ret)
2719                 goto out;
2720
2721         memset(&rep, 0, sizeof rep);
2722         rep.qp_num = id_priv->qp_num;
2723         rep.starting_psn = id_priv->seq_num;
2724         rep.private_data = conn_param->private_data;
2725         rep.private_data_len = conn_param->private_data_len;
2726         rep.responder_resources = conn_param->responder_resources;
2727         rep.initiator_depth = conn_param->initiator_depth;
2728         rep.failover_accepted = 0;
2729         rep.flow_control = conn_param->flow_control;
2730         rep.rnr_retry_count = conn_param->rnr_retry_count;
2731         rep.srq = id_priv->srq ? 1 : 0;
2732
2733         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2734 out:
2735         return ret;
2736 }
2737
2738 static int cma_accept_iw(struct rdma_id_private *id_priv,
2739                   struct rdma_conn_param *conn_param)
2740 {
2741         struct iw_cm_conn_param iw_param;
2742         int ret;
2743
2744         ret = cma_modify_qp_rtr(id_priv, conn_param);
2745         if (ret)
2746                 return ret;
2747
2748         iw_param.ord = conn_param->initiator_depth;
2749         iw_param.ird = conn_param->responder_resources;
2750         iw_param.private_data = conn_param->private_data;
2751         iw_param.private_data_len = conn_param->private_data_len;
2752         if (id_priv->id.qp) {
2753                 iw_param.qpn = id_priv->qp_num;
2754         } else
2755                 iw_param.qpn = conn_param->qp_num;
2756
2757         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2758 }
2759
2760 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2761                              enum ib_cm_sidr_status status,
2762                              const void *private_data, int private_data_len)
2763 {
2764         struct ib_cm_sidr_rep_param rep;
2765         int ret;
2766
2767         memset(&rep, 0, sizeof rep);
2768         rep.status = status;
2769         if (status == IB_SIDR_SUCCESS) {
2770                 ret = cma_set_qkey(id_priv);
2771                 if (ret)
2772                         return ret;
2773                 rep.qp_num = id_priv->qp_num;
2774                 rep.qkey = id_priv->qkey;
2775         }
2776         rep.private_data = private_data;
2777         rep.private_data_len = private_data_len;
2778
2779         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2780 }
2781
2782 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2783 {
2784         struct rdma_id_private *id_priv;
2785         int ret;
2786
2787         id_priv = container_of(id, struct rdma_id_private, id);
2788
2789         id_priv->owner = task_pid_nr(current);
2790
2791         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2792                 return -EINVAL;
2793
2794         if (!id->qp && conn_param) {
2795                 id_priv->qp_num = conn_param->qp_num;
2796                 id_priv->srq = conn_param->srq;
2797         }
2798
2799         switch (rdma_node_get_transport(id->device->node_type)) {
2800         case RDMA_TRANSPORT_IB:
2801                 if (id->qp_type == IB_QPT_UD) {
2802                         if (conn_param)
2803                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2804                                                         conn_param->private_data,
2805                                                         conn_param->private_data_len);
2806                         else
2807                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2808                                                         NULL, 0);
2809                 } else {
2810                         if (conn_param)
2811                                 ret = cma_accept_ib(id_priv, conn_param);
2812                         else
2813                                 ret = cma_rep_recv(id_priv);
2814                 }
2815                 break;
2816         case RDMA_TRANSPORT_IWARP:
2817                 ret = cma_accept_iw(id_priv, conn_param);
2818                 break;
2819         default:
2820                 ret = -ENOSYS;
2821                 break;
2822         }
2823
2824         if (ret)
2825                 goto reject;
2826
2827         return 0;
2828 reject:
2829         cma_modify_qp_err(id_priv);
2830         rdma_reject(id, NULL, 0);
2831         return ret;
2832 }
2833 EXPORT_SYMBOL(rdma_accept);
2834
2835 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2836 {
2837         struct rdma_id_private *id_priv;
2838         int ret;
2839
2840         id_priv = container_of(id, struct rdma_id_private, id);
2841         if (!id_priv->cm_id.ib)
2842                 return -EINVAL;
2843
2844         switch (id->device->node_type) {
2845         case RDMA_NODE_IB_CA:
2846                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2847                 break;
2848         default:
2849                 ret = 0;
2850                 break;
2851         }
2852         return ret;
2853 }
2854 EXPORT_SYMBOL(rdma_notify);
2855
2856 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2857                 u8 private_data_len)
2858 {
2859         struct rdma_id_private *id_priv;
2860         int ret;
2861
2862         id_priv = container_of(id, struct rdma_id_private, id);
2863         if (!id_priv->cm_id.ib)
2864                 return -EINVAL;
2865
2866         switch (rdma_node_get_transport(id->device->node_type)) {
2867         case RDMA_TRANSPORT_IB:
2868                 if (id->qp_type == IB_QPT_UD)
2869                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2870                                                 private_data, private_data_len);
2871                 else
2872                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2873                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2874                                              0, private_data, private_data_len);
2875                 break;
2876         case RDMA_TRANSPORT_IWARP:
2877                 ret = iw_cm_reject(id_priv->cm_id.iw,
2878                                    private_data, private_data_len);
2879                 break;
2880         default:
2881                 ret = -ENOSYS;
2882                 break;
2883         }
2884         return ret;
2885 }
2886 EXPORT_SYMBOL(rdma_reject);
2887
2888 int rdma_disconnect(struct rdma_cm_id *id)
2889 {
2890         struct rdma_id_private *id_priv;
2891         int ret;
2892
2893         id_priv = container_of(id, struct rdma_id_private, id);
2894         if (!id_priv->cm_id.ib)
2895                 return -EINVAL;
2896
2897         switch (rdma_node_get_transport(id->device->node_type)) {
2898         case RDMA_TRANSPORT_IB:
2899                 ret = cma_modify_qp_err(id_priv);
2900                 if (ret)
2901                         goto out;
2902                 /* Initiate or respond to a disconnect. */
2903                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2904                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2905                 break;
2906         case RDMA_TRANSPORT_IWARP:
2907                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2908                 break;
2909         default:
2910                 ret = -EINVAL;
2911                 break;
2912         }
2913 out:
2914         return ret;
2915 }
2916 EXPORT_SYMBOL(rdma_disconnect);
2917
2918 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2919 {
2920         struct rdma_id_private *id_priv;
2921         struct cma_multicast *mc = multicast->context;
2922         struct rdma_cm_event event;
2923         int ret;
2924
2925         id_priv = mc->id_priv;
2926         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2927             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2928                 return 0;
2929
2930         mutex_lock(&id_priv->qp_mutex);
2931         if (!status && id_priv->id.qp)
2932                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2933                                          be16_to_cpu(multicast->rec.mlid));
2934         mutex_unlock(&id_priv->qp_mutex);
2935
2936         memset(&event, 0, sizeof event);
2937         event.status = status;
2938         event.param.ud.private_data = mc->context;
2939         if (!status) {
2940                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2941                 ib_init_ah_from_mcmember(id_priv->id.device,
2942                                          id_priv->id.port_num, &multicast->rec,
2943                                          &event.param.ud.ah_attr);
2944                 event.param.ud.qp_num = 0xFFFFFF;
2945                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2946         } else
2947                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2948
2949         ret = id_priv->id.event_handler(&id_priv->id, &event);
2950         if (ret) {
2951                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2952                 mutex_unlock(&id_priv->handler_mutex);
2953                 rdma_destroy_id(&id_priv->id);
2954                 return 0;
2955         }
2956
2957         mutex_unlock(&id_priv->handler_mutex);
2958         return 0;
2959 }
2960
2961 static void cma_set_mgid(struct rdma_id_private *id_priv,
2962                          struct sockaddr *addr, union ib_gid *mgid)
2963 {
2964         unsigned char mc_map[MAX_ADDR_LEN];
2965         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2966         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2967         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2968
2969         if (cma_any_addr(addr)) {
2970                 memset(mgid, 0, sizeof *mgid);
2971         } else if ((addr->sa_family == AF_INET6) &&
2972                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2973                                                                  0xFF10A01B)) {
2974                 /* IPv6 address is an SA assigned MGID. */
2975                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2976         } else if ((addr->sa_family == AF_INET6)) {
2977                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2978                 if (id_priv->id.ps == RDMA_PS_UDP)
2979                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2980                 *mgid = *(union ib_gid *) (mc_map + 4);
2981         } else {
2982                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2983                 if (id_priv->id.ps == RDMA_PS_UDP)
2984                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2985                 *mgid = *(union ib_gid *) (mc_map + 4);
2986         }
2987 }
2988
2989 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2990                                  struct cma_multicast *mc)
2991 {
2992         struct ib_sa_mcmember_rec rec;
2993         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2994         ib_sa_comp_mask comp_mask;
2995         int ret;
2996
2997         ib_addr_get_mgid(dev_addr, &rec.mgid);
2998         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2999                                      &rec.mgid, &rec);
3000         if (ret)
3001                 return ret;
3002
3003         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3004         if (id_priv->id.ps == RDMA_PS_UDP)
3005                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3006         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3007         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3008         rec.join_state = 1;
3009
3010         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3011                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3012                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3013                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3014                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3015
3016         if (id_priv->id.ps == RDMA_PS_IPOIB)
3017                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3018                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
3019
3020         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3021                                                 id_priv->id.port_num, &rec,
3022                                                 comp_mask, GFP_KERNEL,
3023                                                 cma_ib_mc_handler, mc);
3024         if (IS_ERR(mc->multicast.ib))
3025                 return PTR_ERR(mc->multicast.ib);
3026
3027         return 0;
3028 }
3029
3030 static void iboe_mcast_work_handler(struct work_struct *work)
3031 {
3032         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3033         struct cma_multicast *mc = mw->mc;
3034         struct ib_sa_multicast *m = mc->multicast.ib;
3035
3036         mc->multicast.ib->context = mc;
3037         cma_ib_mc_handler(0, m);
3038         kref_put(&mc->mcref, release_mc);
3039         kfree(mw);
3040 }
3041
3042 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3043 {
3044         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3045         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3046
3047         if (cma_any_addr(addr)) {
3048                 memset(mgid, 0, sizeof *mgid);
3049         } else if (addr->sa_family == AF_INET6) {
3050                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3051         } else {
3052                 mgid->raw[0] = 0xff;
3053                 mgid->raw[1] = 0x0e;
3054                 mgid->raw[2] = 0;
3055                 mgid->raw[3] = 0;
3056                 mgid->raw[4] = 0;
3057                 mgid->raw[5] = 0;
3058                 mgid->raw[6] = 0;
3059                 mgid->raw[7] = 0;
3060                 mgid->raw[8] = 0;
3061                 mgid->raw[9] = 0;
3062                 mgid->raw[10] = 0xff;
3063                 mgid->raw[11] = 0xff;
3064                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3065         }
3066 }
3067
3068 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3069                                    struct cma_multicast *mc)
3070 {
3071         struct iboe_mcast_work *work;
3072         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3073         int err;
3074         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3075         struct net_device *ndev = NULL;
3076
3077         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3078                 return -EINVAL;
3079
3080         work = kzalloc(sizeof *work, GFP_KERNEL);
3081         if (!work)
3082                 return -ENOMEM;
3083
3084         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3085         if (!mc->multicast.ib) {
3086                 err = -ENOMEM;
3087                 goto out1;
3088         }
3089
3090         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3091
3092         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3093         if (id_priv->id.ps == RDMA_PS_UDP)
3094                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3095
3096         if (dev_addr->bound_dev_if)
3097                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3098         if (!ndev) {
3099                 err = -ENODEV;
3100                 goto out2;
3101         }
3102         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3103         mc->multicast.ib->rec.hop_limit = 1;
3104         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3105         dev_put(ndev);
3106         if (!mc->multicast.ib->rec.mtu) {
3107                 err = -EINVAL;
3108                 goto out2;
3109         }
3110         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3111         work->id = id_priv;
3112         work->mc = mc;
3113         INIT_WORK(&work->work, iboe_mcast_work_handler);
3114         kref_get(&mc->mcref);
3115         queue_work(cma_wq, &work->work);
3116
3117         return 0;
3118
3119 out2:
3120         kfree(mc->multicast.ib);
3121 out1:
3122         kfree(work);
3123         return err;
3124 }
3125
3126 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3127                         void *context)
3128 {
3129         struct rdma_id_private *id_priv;
3130         struct cma_multicast *mc;
3131         int ret;
3132
3133         id_priv = container_of(id, struct rdma_id_private, id);
3134         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3135             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3136                 return -EINVAL;
3137
3138         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3139         if (!mc)
3140                 return -ENOMEM;
3141
3142         memcpy(&mc->addr, addr, ip_addr_size(addr));
3143         mc->context = context;
3144         mc->id_priv = id_priv;
3145
3146         spin_lock(&id_priv->lock);
3147         list_add(&mc->list, &id_priv->mc_list);
3148         spin_unlock(&id_priv->lock);
3149
3150         switch (rdma_node_get_transport(id->device->node_type)) {
3151         case RDMA_TRANSPORT_IB:
3152                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3153                 case IB_LINK_LAYER_INFINIBAND:
3154                         ret = cma_join_ib_multicast(id_priv, mc);
3155                         break;
3156                 case IB_LINK_LAYER_ETHERNET:
3157                         kref_init(&mc->mcref);
3158                         ret = cma_iboe_join_multicast(id_priv, mc);
3159                         break;
3160                 default:
3161                         ret = -EINVAL;
3162                 }
3163                 break;
3164         default:
3165                 ret = -ENOSYS;
3166                 break;
3167         }
3168
3169         if (ret) {
3170                 spin_lock_irq(&id_priv->lock);
3171                 list_del(&mc->list);
3172                 spin_unlock_irq(&id_priv->lock);
3173                 kfree(mc);
3174         }
3175         return ret;
3176 }
3177 EXPORT_SYMBOL(rdma_join_multicast);
3178
3179 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3180 {
3181         struct rdma_id_private *id_priv;
3182         struct cma_multicast *mc;
3183
3184         id_priv = container_of(id, struct rdma_id_private, id);
3185         spin_lock_irq(&id_priv->lock);
3186         list_for_each_entry(mc, &id_priv->mc_list, list) {
3187                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3188                         list_del(&mc->list);
3189                         spin_unlock_irq(&id_priv->lock);
3190
3191                         if (id->qp)
3192                                 ib_detach_mcast(id->qp,
3193                                                 &mc->multicast.ib->rec.mgid,
3194                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3195                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3196                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3197                                 case IB_LINK_LAYER_INFINIBAND:
3198                                         ib_sa_free_multicast(mc->multicast.ib);
3199                                         kfree(mc);
3200                                         break;
3201                                 case IB_LINK_LAYER_ETHERNET:
3202                                         kref_put(&mc->mcref, release_mc);
3203                                         break;
3204                                 default:
3205                                         break;
3206                                 }
3207                         }
3208                         return;
3209                 }
3210         }
3211         spin_unlock_irq(&id_priv->lock);
3212 }
3213 EXPORT_SYMBOL(rdma_leave_multicast);
3214
3215 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3216 {
3217         struct rdma_dev_addr *dev_addr;
3218         struct cma_ndev_work *work;
3219
3220         dev_addr = &id_priv->id.route.addr.dev_addr;
3221
3222         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3223             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3224                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3225                        ndev->name, &id_priv->id);
3226                 work = kzalloc(sizeof *work, GFP_KERNEL);
3227                 if (!work)
3228                         return -ENOMEM;
3229
3230                 INIT_WORK(&work->work, cma_ndev_work_handler);
3231                 work->id = id_priv;
3232                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3233                 atomic_inc(&id_priv->refcount);
3234                 queue_work(cma_wq, &work->work);
3235         }
3236
3237         return 0;
3238 }
3239
3240 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3241                                void *ctx)
3242 {
3243         struct net_device *ndev = (struct net_device *)ctx;
3244         struct cma_device *cma_dev;
3245         struct rdma_id_private *id_priv;
3246         int ret = NOTIFY_DONE;
3247
3248         if (dev_net(ndev) != &init_net)
3249                 return NOTIFY_DONE;
3250
3251         if (event != NETDEV_BONDING_FAILOVER)
3252                 return NOTIFY_DONE;
3253
3254         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3255                 return NOTIFY_DONE;
3256
3257         mutex_lock(&lock);
3258         list_for_each_entry(cma_dev, &dev_list, list)
3259                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3260                         ret = cma_netdev_change(ndev, id_priv);
3261                         if (ret)
3262                                 goto out;
3263                 }
3264
3265 out:
3266         mutex_unlock(&lock);
3267         return ret;
3268 }
3269
3270 static struct notifier_block cma_nb = {
3271         .notifier_call = cma_netdev_callback
3272 };
3273
3274 static void cma_add_one(struct ib_device *device)
3275 {
3276         struct cma_device *cma_dev;
3277         struct rdma_id_private *id_priv;
3278
3279         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3280         if (!cma_dev)
3281                 return;
3282
3283         cma_dev->device = device;
3284
3285         init_completion(&cma_dev->comp);
3286         atomic_set(&cma_dev->refcount, 1);
3287         INIT_LIST_HEAD(&cma_dev->id_list);
3288         ib_set_client_data(device, &cma_client, cma_dev);
3289
3290         mutex_lock(&lock);
3291         list_add_tail(&cma_dev->list, &dev_list);
3292         list_for_each_entry(id_priv, &listen_any_list, list)
3293                 cma_listen_on_dev(id_priv, cma_dev);
3294         mutex_unlock(&lock);
3295 }
3296
3297 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3298 {
3299         struct rdma_cm_event event;
3300         enum rdma_cm_state state;
3301         int ret = 0;
3302
3303         /* Record that we want to remove the device */
3304         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3305         if (state == RDMA_CM_DESTROYING)
3306                 return 0;
3307
3308         cma_cancel_operation(id_priv, state);
3309         mutex_lock(&id_priv->handler_mutex);
3310
3311         /* Check for destruction from another callback. */
3312         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3313                 goto out;
3314
3315         memset(&event, 0, sizeof event);
3316         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3317         ret = id_priv->id.event_handler(&id_priv->id, &event);
3318 out:
3319         mutex_unlock(&id_priv->handler_mutex);
3320         return ret;
3321 }
3322
3323 static void cma_process_remove(struct cma_device *cma_dev)
3324 {
3325         struct rdma_id_private *id_priv;
3326         int ret;
3327
3328         mutex_lock(&lock);
3329         while (!list_empty(&cma_dev->id_list)) {
3330                 id_priv = list_entry(cma_dev->id_list.next,
3331                                      struct rdma_id_private, list);
3332
3333                 list_del(&id_priv->listen_list);
3334                 list_del_init(&id_priv->list);
3335                 atomic_inc(&id_priv->refcount);
3336                 mutex_unlock(&lock);
3337
3338                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3339                 cma_deref_id(id_priv);
3340                 if (ret)
3341                         rdma_destroy_id(&id_priv->id);
3342
3343                 mutex_lock(&lock);
3344         }
3345         mutex_unlock(&lock);
3346
3347         cma_deref_dev(cma_dev);
3348         wait_for_completion(&cma_dev->comp);
3349 }
3350
3351 static void cma_remove_one(struct ib_device *device)
3352 {
3353         struct cma_device *cma_dev;
3354
3355         cma_dev = ib_get_client_data(device, &cma_client);
3356         if (!cma_dev)
3357                 return;
3358
3359         mutex_lock(&lock);
3360         list_del(&cma_dev->list);
3361         mutex_unlock(&lock);
3362
3363         cma_process_remove(cma_dev);
3364         kfree(cma_dev);
3365 }
3366
3367 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3368 {
3369         struct nlmsghdr *nlh;
3370         struct rdma_cm_id_stats *id_stats;
3371         struct rdma_id_private *id_priv;
3372         struct rdma_cm_id *id = NULL;
3373         struct cma_device *cma_dev;
3374         int i_dev = 0, i_id = 0;
3375
3376         /*
3377          * We export all of the IDs as a sequence of messages.  Each
3378          * ID gets its own netlink message.
3379          */
3380         mutex_lock(&lock);
3381
3382         list_for_each_entry(cma_dev, &dev_list, list) {
3383                 if (i_dev < cb->args[0]) {
3384                         i_dev++;
3385                         continue;
3386                 }
3387
3388                 i_id = 0;
3389                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3390                         if (i_id < cb->args[1]) {
3391                                 i_id++;
3392                                 continue;
3393                         }
3394
3395                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3396                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3397                                                 RDMA_NL_RDMA_CM_ID_STATS);
3398                         if (!id_stats)
3399                                 goto out;
3400
3401                         memset(id_stats, 0, sizeof *id_stats);
3402                         id = &id_priv->id;
3403                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3404                         id_stats->port_num = id->port_num;
3405                         id_stats->bound_dev_if =
3406                                 id->route.addr.dev_addr.bound_dev_if;
3407
3408                         if (id->route.addr.src_addr.ss_family == AF_INET) {
3409                                 if (ibnl_put_attr(skb, nlh,
3410                                                   sizeof(struct sockaddr_in),
3411                                                   &id->route.addr.src_addr,
3412                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3413                                         goto out;
3414                                 }
3415                                 if (ibnl_put_attr(skb, nlh,
3416                                                   sizeof(struct sockaddr_in),
3417                                                   &id->route.addr.dst_addr,
3418                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3419                                         goto out;
3420                                 }
3421                         } else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3422                                 if (ibnl_put_attr(skb, nlh,
3423                                                   sizeof(struct sockaddr_in6),
3424                                                   &id->route.addr.src_addr,
3425                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3426                                         goto out;
3427                                 }
3428                                 if (ibnl_put_attr(skb, nlh,
3429                                                   sizeof(struct sockaddr_in6),
3430                                                   &id->route.addr.dst_addr,
3431                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3432                                         goto out;
3433                                 }
3434                         }
3435
3436                         id_stats->pid           = id_priv->owner;
3437                         id_stats->port_space    = id->ps;
3438                         id_stats->cm_state      = id_priv->state;
3439                         id_stats->qp_num        = id_priv->qp_num;
3440                         id_stats->qp_type       = id->qp_type;
3441
3442                         i_id++;
3443                 }
3444
3445                 cb->args[1] = 0;
3446                 i_dev++;
3447         }
3448
3449 out:
3450         mutex_unlock(&lock);
3451         cb->args[0] = i_dev;
3452         cb->args[1] = i_id;
3453
3454         return skb->len;
3455 }
3456
3457 static const struct ibnl_client_cbs cma_cb_table[] = {
3458         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats },
3459 };
3460
3461 static int __init cma_init(void)
3462 {
3463         int ret;
3464
3465         cma_wq = create_singlethread_workqueue("rdma_cm");
3466         if (!cma_wq)
3467                 return -ENOMEM;
3468
3469         ib_sa_register_client(&sa_client);
3470         rdma_addr_register_client(&addr_client);
3471         register_netdevice_notifier(&cma_nb);
3472
3473         ret = ib_register_client(&cma_client);
3474         if (ret)
3475                 goto err;
3476
3477         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3478                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3479
3480         return 0;
3481
3482 err:
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         return ret;
3488 }
3489
3490 static void __exit cma_cleanup(void)
3491 {
3492         ibnl_remove_client(RDMA_NL_RDMA_CM);
3493         ib_unregister_client(&cma_client);
3494         unregister_netdevice_notifier(&cma_nb);
3495         rdma_addr_unregister_client(&addr_client);
3496         ib_sa_unregister_client(&sa_client);
3497         destroy_workqueue(cma_wq);
3498         idr_destroy(&sdp_ps);
3499         idr_destroy(&tcp_ps);
3500         idr_destroy(&udp_ps);
3501         idr_destroy(&ipoib_ps);
3502         idr_destroy(&ib_ps);
3503 }
3504
3505 module_init(cma_init);
3506 module_exit(cma_cleanup);