drivers/infiniband/core/cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/idr.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/slab.h>
  44
  45 #include <net/tcp.h>
  46 #include <net/ipv6.h>
  47
  48 #include <rdma/rdma_cm.h>
  49 #include <rdma/rdma_cm_ib.h>
  50 #include <rdma/rdma_netlink.h>
  51 #include <rdma/ib_cache.h>
  52 #include <rdma/ib_cm.h>
  53 #include <rdma/ib_sa.h>
  54 #include <rdma/iw_cm.h>
  55
  56 MODULE_AUTHOR("Sean Hefty");
  57 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59
  60 #define CMA_CM_RESPONSE_TIMEOUT 20
  61 #define CMA_MAX_CM_RETRIES 15
  62 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  63 #define CMA_IBOE_PACKET_LIFETIME 18
  64
  65 static void cma_add_one(struct ib_device *device);
  66 static void cma_remove_one(struct ib_device *device);
  67
  68 static struct ib_client cma_client = {
  69         .name   = "cma",
  70         .add    = cma_add_one,
  71         .remove = cma_remove_one
  72 };
  73
  74 static struct ib_sa_client sa_client;
  75 static struct rdma_addr_client addr_client;
  76 static LIST_HEAD(dev_list);
  77 static LIST_HEAD(listen_any_list);
  78 static DEFINE_MUTEX(lock);
  79 static struct workqueue_struct *cma_wq;
  80 static DEFINE_IDR(sdp_ps);
  81 static DEFINE_IDR(tcp_ps);
  82 static DEFINE_IDR(udp_ps);
  83 static DEFINE_IDR(ipoib_ps);
  84
  85 struct cma_device {
  86         struct list_head        list;
  87         struct ib_device        *device;
  88         struct completion       comp;
  89         atomic_t                refcount;
  90         struct list_head        id_list;
  91 };
  92
  93 struct rdma_bind_list {
  94         struct idr              *ps;
  95         struct hlist_head       owners;
  96         unsigned short          port;
  97 };
  98
  99 /*
 100  * Device removal can occur at anytime, so we need extra handling to
 101  * serialize notifying the user of device removal with other callbacks.
 102  * We do this by disabling removal notification while a callback is in process,
 103  * and reporting it after the callback completes.
 104  */
 105 struct rdma_id_private {
 106         struct rdma_cm_id       id;
 107
 108         struct rdma_bind_list   *bind_list;
 109         struct hlist_node       node;
 110         struct list_head        list; /* listen_any_list or cma_device.list */
 111         struct list_head        listen_list; /* per device listens */
 112         struct cma_device       *cma_dev;
 113         struct list_head        mc_list;
 114
 115         int                     internal_id;
 116         enum rdma_cm_state      state;
 117         spinlock_t              lock;
 118         struct mutex            qp_mutex;
 119
 120         struct completion       comp;
 121         atomic_t                refcount;
 122         struct mutex            handler_mutex;
 123
 124         int                     backlog;
 125         int                     timeout_ms;
 126         struct ib_sa_query      *query;
 127         int                     query_id;
 128         union {
 129                 struct ib_cm_id *ib;
 130                 struct iw_cm_id *iw;
 131         } cm_id;
 132
 133         u32                     seq_num;
 134         u32                     qkey;
 135         u32                     qp_num;
 136         pid_t                   owner;
 137         u8                      srq;
 138         u8                      tos;
 139         u8                      reuseaddr;
 140 };
 141
 142 struct cma_multicast {
 143         struct rdma_id_private *id_priv;
 144         union {
 145                 struct ib_sa_multicast *ib;
 146         } multicast;
 147         struct list_head        list;
 148         void                    *context;
 149         struct sockaddr_storage addr;
 150         struct kref             mcref;
 151 };
 152
 153 struct cma_work {
 154         struct work_struct      work;
 155         struct rdma_id_private  *id;
 156         enum rdma_cm_state      old_state;
 157         enum rdma_cm_state      new_state;
 158         struct rdma_cm_event    event;
 159 };
 160
 161 struct cma_ndev_work {
 162         struct work_struct      work;
 163         struct rdma_id_private  *id;
 164         struct rdma_cm_event    event;
 165 };
 166
 167 struct iboe_mcast_work {
 168         struct work_struct       work;
 169         struct rdma_id_private  *id;
 170         struct cma_multicast    *mc;
 171 };
 172
 173 union cma_ip_addr {
 174         struct in6_addr ip6;
 175         struct {
 176                 __be32 pad[3];
 177                 __be32 addr;
 178         } ip4;
 179 };
 180
 181 struct cma_hdr {
 182         u8 cma_version;
 183         u8 ip_version;  /* IP version: 7:4 */
 184         __be16 port;
 185         union cma_ip_addr src_addr;
 186         union cma_ip_addr dst_addr;
 187 };
 188
 189 struct sdp_hh {
 190         u8 bsdh[16];
 191         u8 sdp_version; /* Major version: 7:4 */
 192         u8 ip_version;  /* IP version: 7:4 */
 193         u8 sdp_specific1[10];
 194         __be16 port;
 195         __be16 sdp_specific2;
 196         union cma_ip_addr src_addr;
 197         union cma_ip_addr dst_addr;
 198 };
 199
 200 struct sdp_hah {
 201         u8 bsdh[16];
 202         u8 sdp_version;
 203 };
 204
 205 #define CMA_VERSION 0x00
 206 #define SDP_MAJ_VERSION 0x2
 207
 208 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 209 {
 210         unsigned long flags;
 211         int ret;
 212
 213         spin_lock_irqsave(&id_priv->lock, flags);
 214         ret = (id_priv->state == comp);
 215         spin_unlock_irqrestore(&id_priv->lock, flags);
 216         return ret;
 217 }
 218
 219 static int cma_comp_exch(struct rdma_id_private *id_priv,
 220                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 221 {
 222         unsigned long flags;
 223         int ret;
 224
 225         spin_lock_irqsave(&id_priv->lock, flags);
 226         if ((ret = (id_priv->state == comp)))
 227                 id_priv->state = exch;
 228         spin_unlock_irqrestore(&id_priv->lock, flags);
 229         return ret;
 230 }
 231
 232 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 233                                    enum rdma_cm_state exch)
 234 {
 235         unsigned long flags;
 236         enum rdma_cm_state old;
 237
 238         spin_lock_irqsave(&id_priv->lock, flags);
 239         old = id_priv->state;
 240         id_priv->state = exch;
 241         spin_unlock_irqrestore(&id_priv->lock, flags);
 242         return old;
 243 }
 244
 245 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 246 {
 247         return hdr->ip_version >> 4;
 248 }
 249
 250 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 251 {
 252         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 253 }
 254
 255 static inline u8 sdp_get_majv(u8 sdp_version)
 256 {
 257         return sdp_version >> 4;
 258 }
 259
 260 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 261 {
 262         return hh->ip_version >> 4;
 263 }
 264
 265 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 266 {
 267         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 268 }
 269
 270 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 271                               struct cma_device *cma_dev)
 272 {
 273         atomic_inc(&cma_dev->refcount);
 274         id_priv->cma_dev = cma_dev;
 275         id_priv->id.device = cma_dev->device;
 276         id_priv->id.route.addr.dev_addr.transport =
 277                 rdma_node_get_transport(cma_dev->device->node_type);
 278         list_add_tail(&id_priv->list, &cma_dev->id_list);
 279 }
 280
 281 static inline void cma_deref_dev(struct cma_device *cma_dev)
 282 {
 283         if (atomic_dec_and_test(&cma_dev->refcount))
 284                 complete(&cma_dev->comp);
 285 }
 286
 287 static inline void release_mc(struct kref *kref)
 288 {
 289         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 290
 291         kfree(mc->multicast.ib);
 292         kfree(mc);
 293 }
 294
 295 static void cma_release_dev(struct rdma_id_private *id_priv)
 296 {
 297         mutex_lock(&lock);
 298         list_del(&id_priv->list);
 299         cma_deref_dev(id_priv->cma_dev);
 300         id_priv->cma_dev = NULL;
 301         mutex_unlock(&lock);
 302 }
 303
 304 static int cma_set_qkey(struct rdma_id_private *id_priv)
 305 {
 306         struct ib_sa_mcmember_rec rec;
 307         int ret = 0;
 308
 309         if (id_priv->qkey)
 310                 return 0;
 311
 312         switch (id_priv->id.ps) {
 313         case RDMA_PS_UDP:
 314                 id_priv->qkey = RDMA_UDP_QKEY;
 315                 break;
 316         case RDMA_PS_IPOIB:
 317                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 318                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 319                                              id_priv->id.port_num, &rec.mgid,
 320                                              &rec);
 321                 if (!ret)
 322                         id_priv->qkey = be32_to_cpu(rec.qkey);
 323                 break;
 324         default:
 325                 break;
 326         }
 327         return ret;
 328 }
 329
 330 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 331 {
 332         int i;
 333         int err;
 334         struct ib_port_attr props;
 335         union ib_gid tmp;
 336
 337         err = ib_query_port(device, port_num, &props);
 338         if (err)
 339                 return 1;
 340
 341         for (i = 0; i < props.gid_tbl_len; ++i) {
 342                 err = ib_query_gid(device, port_num, i, &tmp);
 343                 if (err)
 344                         return 1;
 345                 if (!memcmp(&tmp, gid, sizeof tmp))
 346                         return 0;
 347         }
 348
 349         return -EAGAIN;
 350 }
 351
 352 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 353 {
 354         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 355         struct cma_device *cma_dev;
 356         union ib_gid gid, iboe_gid;
 357         int ret = -ENODEV;
 358         u8 port;
 359         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 360                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 361
 362         mutex_lock(&lock);
 363         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 364         memcpy(&gid, dev_addr->src_dev_addr +
 365                rdma_addr_gid_offset(dev_addr), sizeof gid);
 366         list_for_each_entry(cma_dev, &dev_list, list) {
 367                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 368                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 369                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 370                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 371                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 372                                 else
 373                                         ret = find_gid_port(cma_dev->device, &gid, port);
 374
 375                                 if (!ret) {
 376                                         id_priv->id.port_num = port;
 377                                         goto out;
 378                                 } else if (ret == 1)
 379                                         break;
 380                         }
 381                 }
 382         }
 383
 384 out:
 385         if (!ret)
 386                 cma_attach_to_dev(id_priv, cma_dev);
 387
 388         mutex_unlock(&lock);
 389         return ret;
 390 }
 391
 392 static void cma_deref_id(struct rdma_id_private *id_priv)
 393 {
 394         if (atomic_dec_and_test(&id_priv->refcount))
 395                 complete(&id_priv->comp);
 396 }
 397
 398 static int cma_disable_callback(struct rdma_id_private *id_priv,
 399                                 enum rdma_cm_state state)
 400 {
 401         mutex_lock(&id_priv->handler_mutex);
 402         if (id_priv->state != state) {
 403                 mutex_unlock(&id_priv->handler_mutex);
 404                 return -EINVAL;
 405         }
 406         return 0;
 407 }
 408
 409 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 410                                   void *context, enum rdma_port_space ps,
 411                                   enum ib_qp_type qp_type)
 412 {
 413         struct rdma_id_private *id_priv;
 414
 415         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 416         if (!id_priv)
 417                 return ERR_PTR(-ENOMEM);
 418
 419         id_priv->owner = task_pid_nr(current);
 420         id_priv->state = RDMA_CM_IDLE;
 421         id_priv->id.context = context;
 422         id_priv->id.event_handler = event_handler;
 423         id_priv->id.ps = ps;
 424         id_priv->id.qp_type = qp_type;
 425         spin_lock_init(&id_priv->lock);
 426         mutex_init(&id_priv->qp_mutex);
 427         init_completion(&id_priv->comp);
 428         atomic_set(&id_priv->refcount, 1);
 429         mutex_init(&id_priv->handler_mutex);
 430         INIT_LIST_HEAD(&id_priv->listen_list);
 431         INIT_LIST_HEAD(&id_priv->mc_list);
 432         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 433
 434         return &id_priv->id;
 435 }
 436 EXPORT_SYMBOL(rdma_create_id);
 437
 438 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 439 {
 440         struct ib_qp_attr qp_attr;
 441         int qp_attr_mask, ret;
 442
 443         qp_attr.qp_state = IB_QPS_INIT;
 444         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 445         if (ret)
 446                 return ret;
 447
 448         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 449         if (ret)
 450                 return ret;
 451
 452         qp_attr.qp_state = IB_QPS_RTR;
 453         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 454         if (ret)
 455                 return ret;
 456
 457         qp_attr.qp_state = IB_QPS_RTS;
 458         qp_attr.sq_psn = 0;
 459         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 460
 461         return ret;
 462 }
 463
 464 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 465 {
 466         struct ib_qp_attr qp_attr;
 467         int qp_attr_mask, ret;
 468
 469         qp_attr.qp_state = IB_QPS_INIT;
 470         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 471         if (ret)
 472                 return ret;
 473
 474         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 475 }
 476
 477 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 478                    struct ib_qp_init_attr *qp_init_attr)
 479 {
 480         struct rdma_id_private *id_priv;
 481         struct ib_qp *qp;
 482         int ret;
 483
 484         id_priv = container_of(id, struct rdma_id_private, id);
 485         if (id->device != pd->device)
 486                 return -EINVAL;
 487
 488         qp = ib_create_qp(pd, qp_init_attr);
 489         if (IS_ERR(qp))
 490                 return PTR_ERR(qp);
 491
 492         if (id->qp_type == IB_QPT_UD)
 493                 ret = cma_init_ud_qp(id_priv, qp);
 494         else
 495                 ret = cma_init_conn_qp(id_priv, qp);
 496         if (ret)
 497                 goto err;
 498
 499         id->qp = qp;
 500         id_priv->qp_num = qp->qp_num;
 501         id_priv->srq = (qp->srq != NULL);
 502         return 0;
 503 err:
 504         ib_destroy_qp(qp);
 505         return ret;
 506 }
 507 EXPORT_SYMBOL(rdma_create_qp);
 508
 509 void rdma_destroy_qp(struct rdma_cm_id *id)
 510 {
 511         struct rdma_id_private *id_priv;
 512
 513         id_priv = container_of(id, struct rdma_id_private, id);
 514         mutex_lock(&id_priv->qp_mutex);
 515         ib_destroy_qp(id_priv->id.qp);
 516         id_priv->id.qp = NULL;
 517         mutex_unlock(&id_priv->qp_mutex);
 518 }
 519 EXPORT_SYMBOL(rdma_destroy_qp);
 520
 521 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 522                              struct rdma_conn_param *conn_param)
 523 {
 524         struct ib_qp_attr qp_attr;
 525         int qp_attr_mask, ret;
 526
 527         mutex_lock(&id_priv->qp_mutex);
 528         if (!id_priv->id.qp) {
 529                 ret = 0;
 530                 goto out;
 531         }
 532
 533         /* Need to update QP attributes from default values. */
 534         qp_attr.qp_state = IB_QPS_INIT;
 535         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 536         if (ret)
 537                 goto out;
 538
 539         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 540         if (ret)
 541                 goto out;
 542
 543         qp_attr.qp_state = IB_QPS_RTR;
 544         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 545         if (ret)
 546                 goto out;
 547
 548         if (conn_param)
 549                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 550         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 551 out:
 552         mutex_unlock(&id_priv->qp_mutex);
 553         return ret;
 554 }
 555
 556 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 557                              struct rdma_conn_param *conn_param)
 558 {
 559         struct ib_qp_attr qp_attr;
 560         int qp_attr_mask, ret;
 561
 562         mutex_lock(&id_priv->qp_mutex);
 563         if (!id_priv->id.qp) {
 564                 ret = 0;
 565                 goto out;
 566         }
 567
 568         qp_attr.qp_state = IB_QPS_RTS;
 569         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 570         if (ret)
 571                 goto out;
 572
 573         if (conn_param)
 574                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 575         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 576 out:
 577         mutex_unlock(&id_priv->qp_mutex);
 578         return ret;
 579 }
 580
 581 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 582 {
 583         struct ib_qp_attr qp_attr;
 584         int ret;
 585
 586         mutex_lock(&id_priv->qp_mutex);
 587         if (!id_priv->id.qp) {
 588                 ret = 0;
 589                 goto out;
 590         }
 591
 592         qp_attr.qp_state = IB_QPS_ERR;
 593         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 594 out:
 595         mutex_unlock(&id_priv->qp_mutex);
 596         return ret;
 597 }
 598
 599 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 600                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 601 {
 602         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 603         int ret;
 604         u16 pkey;
 605
 606         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 607             IB_LINK_LAYER_INFINIBAND)
 608                 pkey = ib_addr_get_pkey(dev_addr);
 609         else
 610                 pkey = 0xffff;
 611
 612         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 613                                   pkey, &qp_attr->pkey_index);
 614         if (ret)
 615                 return ret;
 616
 617         qp_attr->port_num = id_priv->id.port_num;
 618         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 619
 620         if (id_priv->id.qp_type == IB_QPT_UD) {
 621                 ret = cma_set_qkey(id_priv);
 622                 if (ret)
 623                         return ret;
 624
 625                 qp_attr->qkey = id_priv->qkey;
 626                 *qp_attr_mask |= IB_QP_QKEY;
 627         } else {
 628                 qp_attr->qp_access_flags = 0;
 629                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 630         }
 631         return 0;
 632 }
 633
 634 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 635                        int *qp_attr_mask)
 636 {
 637         struct rdma_id_private *id_priv;
 638         int ret = 0;
 639
 640         id_priv = container_of(id, struct rdma_id_private, id);
 641         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 642         case RDMA_TRANSPORT_IB:
 643                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 644                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 645                 else
 646                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 647                                                  qp_attr_mask);
 648                 if (qp_attr->qp_state == IB_QPS_RTR)
 649                         qp_attr->rq_psn = id_priv->seq_num;
 650                 break;
 651         case RDMA_TRANSPORT_IWARP:
 652                 if (!id_priv->cm_id.iw) {
 653                         qp_attr->qp_access_flags = 0;
 654                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 655                 } else
 656                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 657                                                  qp_attr_mask);
 658                 break;
 659         default:
 660                 ret = -ENOSYS;
 661                 break;
 662         }
 663
 664         return ret;
 665 }
 666 EXPORT_SYMBOL(rdma_init_qp_attr);
 667
 668 static inline int cma_zero_addr(struct sockaddr *addr)
 669 {
 670         struct in6_addr *ip6;
 671
 672         if (addr->sa_family == AF_INET)
 673                 return ipv4_is_zeronet(
 674                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 675         else {
 676                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 677                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 678                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 679         }
 680 }
 681
 682 static inline int cma_loopback_addr(struct sockaddr *addr)
 683 {
 684         if (addr->sa_family == AF_INET)
 685                 return ipv4_is_loopback(
 686                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 687         else
 688                 return ipv6_addr_loopback(
 689                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 690 }
 691
 692 static inline int cma_any_addr(struct sockaddr *addr)
 693 {
 694         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 695 }
 696
 697 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 698 {
 699         if (src->sa_family != dst->sa_family)
 700                 return -1;
 701
 702         switch (src->sa_family) {
 703         case AF_INET:
 704                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 705                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 706         default:
 707                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 708                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 709         }
 710 }
 711
 712 static inline __be16 cma_port(struct sockaddr *addr)
 713 {
 714         if (addr->sa_family == AF_INET)
 715                 return ((struct sockaddr_in *) addr)->sin_port;
 716         else
 717                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 718 }
 719
 720 static inline int cma_any_port(struct sockaddr *addr)
 721 {
 722         return !cma_port(addr);
 723 }
 724
 725 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 726                             u8 *ip_ver, __be16 *port,
 727                             union cma_ip_addr **src, union cma_ip_addr **dst)
 728 {
 729         switch (ps) {
 730         case RDMA_PS_SDP:
 731                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 732                     SDP_MAJ_VERSION)
 733                         return -EINVAL;
 734
 735                 *ip_ver = sdp_get_ip_ver(hdr);
 736                 *port   = ((struct sdp_hh *) hdr)->port;
 737                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 738                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 739                 break;
 740         default:
 741                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 742                         return -EINVAL;
 743
 744                 *ip_ver = cma_get_ip_ver(hdr);
 745                 *port   = ((struct cma_hdr *) hdr)->port;
 746                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 747                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 748                 break;
 749         }
 750
 751         if (*ip_ver != 4 && *ip_ver != 6)
 752                 return -EINVAL;
 753         return 0;
 754 }
 755
 756 static void cma_save_net_info(struct rdma_addr *addr,
 757                               struct rdma_addr *listen_addr,
 758                               u8 ip_ver, __be16 port,
 759                               union cma_ip_addr *src, union cma_ip_addr *dst)
 760 {
 761         struct sockaddr_in *listen4, *ip4;
 762         struct sockaddr_in6 *listen6, *ip6;
 763
 764         switch (ip_ver) {
 765         case 4:
 766                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 767                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 768                 ip4->sin_family = listen4->sin_family;
 769                 ip4->sin_addr.s_addr = dst->ip4.addr;
 770                 ip4->sin_port = listen4->sin_port;
 771
 772                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 773                 ip4->sin_family = listen4->sin_family;
 774                 ip4->sin_addr.s_addr = src->ip4.addr;
 775                 ip4->sin_port = port;
 776                 break;
 777         case 6:
 778                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 779                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 780                 ip6->sin6_family = listen6->sin6_family;
 781                 ip6->sin6_addr = dst->ip6;
 782                 ip6->sin6_port = listen6->sin6_port;
 783
 784                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 785                 ip6->sin6_family = listen6->sin6_family;
 786                 ip6->sin6_addr = src->ip6;
 787                 ip6->sin6_port = port;
 788                 break;
 789         default:
 790                 break;
 791         }
 792 }
 793
 794 static inline int cma_user_data_offset(enum rdma_port_space ps)
 795 {
 796         switch (ps) {
 797         case RDMA_PS_SDP:
 798                 return 0;
 799         default:
 800                 return sizeof(struct cma_hdr);
 801         }
 802 }
 803
 804 static void cma_cancel_route(struct rdma_id_private *id_priv)
 805 {
 806         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 807         case IB_LINK_LAYER_INFINIBAND:
 808                 if (id_priv->query)
 809                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 810                 break;
 811         default:
 812                 break;
 813         }
 814 }
 815
 816 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 817 {
 818         struct rdma_id_private *dev_id_priv;
 819
 820         /*
 821          * Remove from listen_any_list to prevent added devices from spawning
 822          * additional listen requests.
 823          */
 824         mutex_lock(&lock);
 825         list_del(&id_priv->list);
 826
 827         while (!list_empty(&id_priv->listen_list)) {
 828                 dev_id_priv = list_entry(id_priv->listen_list.next,
 829                                          struct rdma_id_private, listen_list);
 830                 /* sync with device removal to avoid duplicate destruction */
 831                 list_del_init(&dev_id_priv->list);
 832                 list_del(&dev_id_priv->listen_list);
 833                 mutex_unlock(&lock);
 834
 835                 rdma_destroy_id(&dev_id_priv->id);
 836                 mutex_lock(&lock);
 837         }
 838         mutex_unlock(&lock);
 839 }
 840
 841 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 842                                  enum rdma_cm_state state)
 843 {
 844         switch (state) {
 845         case RDMA_CM_ADDR_QUERY:
 846                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 847                 break;
 848         case RDMA_CM_ROUTE_QUERY:
 849                 cma_cancel_route(id_priv);
 850                 break;
 851         case RDMA_CM_LISTEN:
 852                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 853                                 && !id_priv->cma_dev)
 854                         cma_cancel_listens(id_priv);
 855                 break;
 856         default:
 857                 break;
 858         }
 859 }
 860
 861 static void cma_release_port(struct rdma_id_private *id_priv)
 862 {
 863         struct rdma_bind_list *bind_list = id_priv->bind_list;
 864
 865         if (!bind_list)
 866                 return;
 867
 868         mutex_lock(&lock);
 869         hlist_del(&id_priv->node);
 870         if (hlist_empty(&bind_list->owners)) {
 871                 idr_remove(bind_list->ps, bind_list->port);
 872                 kfree(bind_list);
 873         }
 874         mutex_unlock(&lock);
 875 }
 876
 877 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 878 {
 879         struct cma_multicast *mc;
 880
 881         while (!list_empty(&id_priv->mc_list)) {
 882                 mc = container_of(id_priv->mc_list.next,
 883                                   struct cma_multicast, list);
 884                 list_del(&mc->list);
 885                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 886                 case IB_LINK_LAYER_INFINIBAND:
 887                         ib_sa_free_multicast(mc->multicast.ib);
 888                         kfree(mc);
 889                         break;
 890                 case IB_LINK_LAYER_ETHERNET:
 891                         kref_put(&mc->mcref, release_mc);
 892                         break;
 893                 default:
 894                         break;
 895                 }
 896         }
 897 }
 898
 899 void rdma_destroy_id(struct rdma_cm_id *id)
 900 {
 901         struct rdma_id_private *id_priv;
 902         enum rdma_cm_state state;
 903
 904         id_priv = container_of(id, struct rdma_id_private, id);
 905         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 906         cma_cancel_operation(id_priv, state);
 907
 908         /*
 909          * Wait for any active callback to finish.  New callbacks will find
 910          * the id_priv state set to destroying and abort.
 911          */
 912         mutex_lock(&id_priv->handler_mutex);
 913         mutex_unlock(&id_priv->handler_mutex);
 914
 915         if (id_priv->cma_dev) {
 916                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 917                 case RDMA_TRANSPORT_IB:
 918                         if (id_priv->cm_id.ib)
 919                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 920                         break;
 921                 case RDMA_TRANSPORT_IWARP:
 922                         if (id_priv->cm_id.iw)
 923                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 924                         break;
 925                 default:
 926                         break;
 927                 }
 928                 cma_leave_mc_groups(id_priv);
 929                 cma_release_dev(id_priv);
 930         }
 931
 932         cma_release_port(id_priv);
 933         cma_deref_id(id_priv);
 934         wait_for_completion(&id_priv->comp);
 935
 936         if (id_priv->internal_id)
 937                 cma_deref_id(id_priv->id.context);
 938
 939         kfree(id_priv->id.route.path_rec);
 940         kfree(id_priv);
 941 }
 942 EXPORT_SYMBOL(rdma_destroy_id);
 943
 944 static int cma_rep_recv(struct rdma_id_private *id_priv)
 945 {
 946         int ret;
 947
 948         ret = cma_modify_qp_rtr(id_priv, NULL);
 949         if (ret)
 950                 goto reject;
 951
 952         ret = cma_modify_qp_rts(id_priv, NULL);
 953         if (ret)
 954                 goto reject;
 955
 956         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 957         if (ret)
 958                 goto reject;
 959
 960         return 0;
 961 reject:
 962         cma_modify_qp_err(id_priv);
 963         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 964                        NULL, 0, NULL, 0);
 965         return ret;
 966 }
 967
 968 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 969 {
 970         if (id_priv->id.ps == RDMA_PS_SDP &&
 971             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 972             SDP_MAJ_VERSION)
 973                 return -EINVAL;
 974
 975         return 0;
 976 }
 977
 978 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 979                                    struct ib_cm_rep_event_param *rep_data,
 980                                    void *private_data)
 981 {
 982         event->param.conn.private_data = private_data;
 983         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 984         event->param.conn.responder_resources = rep_data->responder_resources;
 985         event->param.conn.initiator_depth = rep_data->initiator_depth;
 986         event->param.conn.flow_control = rep_data->flow_control;
 987         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 988         event->param.conn.srq = rep_data->srq;
 989         event->param.conn.qp_num = rep_data->remote_qpn;
 990 }
 991
 992 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 993 {
 994         struct rdma_id_private *id_priv = cm_id->context;
 995         struct rdma_cm_event event;
 996         int ret = 0;
 997
 998         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
 999                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1000             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1001                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1002                 return 0;
1003
1004         memset(&event, 0, sizeof event);
1005         switch (ib_event->event) {
1006         case IB_CM_REQ_ERROR:
1007         case IB_CM_REP_ERROR:
1008                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1009                 event.status = -ETIMEDOUT;
1010                 break;
1011         case IB_CM_REP_RECEIVED:
1012                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1013                 if (event.status)
1014                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1015                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1016                         event.status = cma_rep_recv(id_priv);
1017                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1018                                                      RDMA_CM_EVENT_ESTABLISHED;
1019                 } else
1020                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1021                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1022                                        ib_event->private_data);
1023                 break;
1024         case IB_CM_RTU_RECEIVED:
1025         case IB_CM_USER_ESTABLISHED:
1026                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1027                 break;
1028         case IB_CM_DREQ_ERROR:
1029                 event.status = -ETIMEDOUT; /* fall through */
1030         case IB_CM_DREQ_RECEIVED:
1031         case IB_CM_DREP_RECEIVED:
1032                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1033                                    RDMA_CM_DISCONNECT))
1034                         goto out;
1035                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1036                 break;
1037         case IB_CM_TIMEWAIT_EXIT:
1038                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1039                 break;
1040         case IB_CM_MRA_RECEIVED:
1041                 /* ignore event */
1042                 goto out;
1043         case IB_CM_REJ_RECEIVED:
1044                 cma_modify_qp_err(id_priv);
1045                 event.status = ib_event->param.rej_rcvd.reason;
1046                 event.event = RDMA_CM_EVENT_REJECTED;
1047                 event.param.conn.private_data = ib_event->private_data;
1048                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1049                 break;
1050         default:
1051                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1052                        ib_event->event);
1053                 goto out;
1054         }
1055
1056         ret = id_priv->id.event_handler(&id_priv->id, &event);
1057         if (ret) {
1058                 /* Destroy the CM ID by returning a non-zero value. */
1059                 id_priv->cm_id.ib = NULL;
1060                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1061                 mutex_unlock(&id_priv->handler_mutex);
1062                 rdma_destroy_id(&id_priv->id);
1063                 return ret;
1064         }
1065 out:
1066         mutex_unlock(&id_priv->handler_mutex);
1067         return ret;
1068 }
1069
1070 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1071                                                struct ib_cm_event *ib_event)
1072 {
1073         struct rdma_id_private *id_priv;
1074         struct rdma_cm_id *id;
1075         struct rdma_route *rt;
1076         union cma_ip_addr *src, *dst;
1077         __be16 port;
1078         u8 ip_ver;
1079         int ret;
1080
1081         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1082                              &ip_ver, &port, &src, &dst))
1083                 return NULL;
1084
1085         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1086                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1087         if (IS_ERR(id))
1088                 return NULL;
1089
1090         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1091                           ip_ver, port, src, dst);
1092
1093         rt = &id->route;
1094         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1095         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1096                                GFP_KERNEL);
1097         if (!rt->path_rec)
1098                 goto err;
1099
1100         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1101         if (rt->num_paths == 2)
1102                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1103
1104         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1105                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1106                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1107                 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1108         } else {
1109                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1110                                         &rt->addr.dev_addr);
1111                 if (ret)
1112                         goto err;
1113         }
1114         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1115
1116         id_priv = container_of(id, struct rdma_id_private, id);
1117         id_priv->state = RDMA_CM_CONNECT;
1118         return id_priv;
1119
1120 err:
1121         rdma_destroy_id(id);
1122         return NULL;
1123 }
1124
1125 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1126                                               struct ib_cm_event *ib_event)
1127 {
1128         struct rdma_id_private *id_priv;
1129         struct rdma_cm_id *id;
1130         union cma_ip_addr *src, *dst;
1131         __be16 port;
1132         u8 ip_ver;
1133         int ret;
1134
1135         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1136                             listen_id->ps, IB_QPT_UD);
1137         if (IS_ERR(id))
1138                 return NULL;
1139
1140
1141         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1142                              &ip_ver, &port, &src, &dst))
1143                 goto err;
1144
1145         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1146                           ip_ver, port, src, dst);
1147
1148         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1149                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1150                                         &id->route.addr.dev_addr);
1151                 if (ret)
1152                         goto err;
1153         }
1154
1155         id_priv = container_of(id, struct rdma_id_private, id);
1156         id_priv->state = RDMA_CM_CONNECT;
1157         return id_priv;
1158 err:
1159         rdma_destroy_id(id);
1160         return NULL;
1161 }
1162
1163 static void cma_set_req_event_data(struct rdma_cm_event *event,
1164                                    struct ib_cm_req_event_param *req_data,
1165                                    void *private_data, int offset)
1166 {
1167         event->param.conn.private_data = private_data + offset;
1168         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1169         event->param.conn.responder_resources = req_data->responder_resources;
1170         event->param.conn.initiator_depth = req_data->initiator_depth;
1171         event->param.conn.flow_control = req_data->flow_control;
1172         event->param.conn.retry_count = req_data->retry_count;
1173         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1174         event->param.conn.srq = req_data->srq;
1175         event->param.conn.qp_num = req_data->remote_qpn;
1176 }
1177
1178 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1179 {
1180         struct rdma_id_private *listen_id, *conn_id;
1181         struct rdma_cm_event event;
1182         int offset, ret;
1183
1184         listen_id = cm_id->context;
1185         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1186                 return -ECONNABORTED;
1187
1188         memset(&event, 0, sizeof event);
1189         offset = cma_user_data_offset(listen_id->id.ps);
1190         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1191         if (listen_id->id.qp_type == IB_QPT_UD) {
1192                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1193                 event.param.ud.private_data = ib_event->private_data + offset;
1194                 event.param.ud.private_data_len =
1195                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1196         } else {
1197                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1198                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1199                                        ib_event->private_data, offset);
1200         }
1201         if (!conn_id) {
1202                 ret = -ENOMEM;
1203                 goto out;
1204         }
1205
1206         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1207         ret = cma_acquire_dev(conn_id);
1208         if (ret)
1209                 goto release_conn_id;
1210
1211         conn_id->cm_id.ib = cm_id;
1212         cm_id->context = conn_id;
1213         cm_id->cm_handler = cma_ib_handler;
1214
1215         /*
1216          * Protect against the user destroying conn_id from another thread
1217          * until we're done accessing it.
1218          */
1219         atomic_inc(&conn_id->refcount);
1220         ret = conn_id->id.event_handler(&conn_id->id, &event);
1221         if (!ret) {
1222                 /*
1223                  * Acquire mutex to prevent user executing rdma_destroy_id()
1224                  * while we're accessing the cm_id.
1225                  */
1226                 mutex_lock(&lock);
1227                 if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1228                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1229                 mutex_unlock(&lock);
1230                 mutex_unlock(&conn_id->handler_mutex);
1231                 cma_deref_id(conn_id);
1232                 goto out;
1233         }
1234         cma_deref_id(conn_id);
1235
1236         /* Destroy the CM ID by returning a non-zero value. */
1237         conn_id->cm_id.ib = NULL;
1238
1239 release_conn_id:
1240         cma_exch(conn_id, RDMA_CM_DESTROYING);
1241         mutex_unlock(&conn_id->handler_mutex);
1242         rdma_destroy_id(&conn_id->id);
1243
1244 out:
1245         mutex_unlock(&listen_id->handler_mutex);
1246         return ret;
1247 }
1248
1249 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1250 {
1251         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1252 }
1253
1254 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1255                                  struct ib_cm_compare_data *compare)
1256 {
1257         struct cma_hdr *cma_data, *cma_mask;
1258         struct sdp_hh *sdp_data, *sdp_mask;
1259         __be32 ip4_addr;
1260         struct in6_addr ip6_addr;
1261
1262         memset(compare, 0, sizeof *compare);
1263         cma_data = (void *) compare->data;
1264         cma_mask = (void *) compare->mask;
1265         sdp_data = (void *) compare->data;
1266         sdp_mask = (void *) compare->mask;
1267
1268         switch (addr->sa_family) {
1269         case AF_INET:
1270                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1271                 if (ps == RDMA_PS_SDP) {
1272                         sdp_set_ip_ver(sdp_data, 4);
1273                         sdp_set_ip_ver(sdp_mask, 0xF);
1274                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1275                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1276                 } else {
1277                         cma_set_ip_ver(cma_data, 4);
1278                         cma_set_ip_ver(cma_mask, 0xF);
1279                         cma_data->dst_addr.ip4.addr = ip4_addr;
1280                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1281                 }
1282                 break;
1283         case AF_INET6:
1284                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1285                 if (ps == RDMA_PS_SDP) {
1286                         sdp_set_ip_ver(sdp_data, 6);
1287                         sdp_set_ip_ver(sdp_mask, 0xF);
1288                         sdp_data->dst_addr.ip6 = ip6_addr;
1289                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1290                                sizeof sdp_mask->dst_addr.ip6);
1291                 } else {
1292                         cma_set_ip_ver(cma_data, 6);
1293                         cma_set_ip_ver(cma_mask, 0xF);
1294                         cma_data->dst_addr.ip6 = ip6_addr;
1295                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1296                                sizeof cma_mask->dst_addr.ip6);
1297                 }
1298                 break;
1299         default:
1300                 break;
1301         }
1302 }
1303
1304 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1305 {
1306         struct rdma_id_private *id_priv = iw_id->context;
1307         struct rdma_cm_event event;
1308         struct sockaddr_in *sin;
1309         int ret = 0;
1310
1311         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1312                 return 0;
1313
1314         memset(&event, 0, sizeof event);
1315         switch (iw_event->event) {
1316         case IW_CM_EVENT_CLOSE:
1317                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1318                 break;
1319         case IW_CM_EVENT_CONNECT_REPLY:
1320                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1321                 *sin = iw_event->local_addr;
1322                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1323                 *sin = iw_event->remote_addr;
1324                 switch (iw_event->status) {
1325                 case 0:
1326                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1327                         break;
1328                 case -ECONNRESET:
1329                 case -ECONNREFUSED:
1330                         event.event = RDMA_CM_EVENT_REJECTED;
1331                         break;
1332                 case -ETIMEDOUT:
1333                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1334                         break;
1335                 default:
1336                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1337                         break;
1338                 }
1339                 break;
1340         case IW_CM_EVENT_ESTABLISHED:
1341                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1342                 break;
1343         default:
1344                 BUG_ON(1);
1345         }
1346
1347         event.status = iw_event->status;
1348         event.param.conn.private_data = iw_event->private_data;
1349         event.param.conn.private_data_len = iw_event->private_data_len;
1350         ret = id_priv->id.event_handler(&id_priv->id, &event);
1351         if (ret) {
1352                 /* Destroy the CM ID by returning a non-zero value. */
1353                 id_priv->cm_id.iw = NULL;
1354                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1355                 mutex_unlock(&id_priv->handler_mutex);
1356                 rdma_destroy_id(&id_priv->id);
1357                 return ret;
1358         }
1359
1360         mutex_unlock(&id_priv->handler_mutex);
1361         return ret;
1362 }
1363
1364 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1365                                struct iw_cm_event *iw_event)
1366 {
1367         struct rdma_cm_id *new_cm_id;
1368         struct rdma_id_private *listen_id, *conn_id;
1369         struct sockaddr_in *sin;
1370         struct net_device *dev = NULL;
1371         struct rdma_cm_event event;
1372         int ret;
1373         struct ib_device_attr attr;
1374
1375         listen_id = cm_id->context;
1376         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1377                 return -ECONNABORTED;
1378
1379         /* Create a new RDMA id for the new IW CM ID */
1380         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1381                                    listen_id->id.context,
1382                                    RDMA_PS_TCP, IB_QPT_RC);
1383         if (IS_ERR(new_cm_id)) {
1384                 ret = -ENOMEM;
1385                 goto out;
1386         }
1387         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1388         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1389         conn_id->state = RDMA_CM_CONNECT;
1390
1391         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1392         if (!dev) {
1393                 ret = -EADDRNOTAVAIL;
1394                 mutex_unlock(&conn_id->handler_mutex);
1395                 rdma_destroy_id(new_cm_id);
1396                 goto out;
1397         }
1398         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1399         if (ret) {
1400                 mutex_unlock(&conn_id->handler_mutex);
1401                 rdma_destroy_id(new_cm_id);
1402                 goto out;
1403         }
1404
1405         ret = cma_acquire_dev(conn_id);
1406         if (ret) {
1407                 mutex_unlock(&conn_id->handler_mutex);
1408                 rdma_destroy_id(new_cm_id);
1409                 goto out;
1410         }
1411
1412         conn_id->cm_id.iw = cm_id;
1413         cm_id->context = conn_id;
1414         cm_id->cm_handler = cma_iw_handler;
1415
1416         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1417         *sin = iw_event->local_addr;
1418         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1419         *sin = iw_event->remote_addr;
1420
1421         ret = ib_query_device(conn_id->id.device, &attr);
1422         if (ret) {
1423                 mutex_unlock(&conn_id->handler_mutex);
1424                 rdma_destroy_id(new_cm_id);
1425                 goto out;
1426         }
1427
1428         memset(&event, 0, sizeof event);
1429         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1430         event.param.conn.private_data = iw_event->private_data;
1431         event.param.conn.private_data_len = iw_event->private_data_len;
1432         event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1433         event.param.conn.responder_resources = attr.max_qp_rd_atom;
1434
1435         /*
1436          * Protect against the user destroying conn_id from another thread
1437          * until we're done accessing it.
1438          */
1439         atomic_inc(&conn_id->refcount);
1440         ret = conn_id->id.event_handler(&conn_id->id, &event);
1441         if (ret) {
1442                 /* User wants to destroy the CM ID */
1443                 conn_id->cm_id.iw = NULL;
1444                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1445                 mutex_unlock(&conn_id->handler_mutex);
1446                 cma_deref_id(conn_id);
1447                 rdma_destroy_id(&conn_id->id);
1448                 goto out;
1449         }
1450
1451         mutex_unlock(&conn_id->handler_mutex);
1452         cma_deref_id(conn_id);
1453
1454 out:
1455         if (dev)
1456                 dev_put(dev);
1457         mutex_unlock(&listen_id->handler_mutex);
1458         return ret;
1459 }
1460
1461 static int cma_ib_listen(struct rdma_id_private *id_priv)
1462 {
1463         struct ib_cm_compare_data compare_data;
1464         struct sockaddr *addr;
1465         struct ib_cm_id *id;
1466         __be64 svc_id;
1467         int ret;
1468
1469         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1470         if (IS_ERR(id))
1471                 return PTR_ERR(id);
1472
1473         id_priv->cm_id.ib = id;
1474
1475         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1476         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1477         if (cma_any_addr(addr))
1478                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1479         else {
1480                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1481                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1482         }
1483
1484         if (ret) {
1485                 ib_destroy_cm_id(id_priv->cm_id.ib);
1486                 id_priv->cm_id.ib = NULL;
1487         }
1488
1489         return ret;
1490 }
1491
1492 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1493 {
1494         int ret;
1495         struct sockaddr_in *sin;
1496         struct iw_cm_id *id;
1497
1498         id = iw_create_cm_id(id_priv->id.device,
1499                              iw_conn_req_handler,
1500                              id_priv);
1501         if (IS_ERR(id))
1502                 return PTR_ERR(id);
1503
1504         id_priv->cm_id.iw = id;
1505
1506         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1507         id_priv->cm_id.iw->local_addr = *sin;
1508
1509         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1510
1511         if (ret) {
1512                 iw_destroy_cm_id(id_priv->cm_id.iw);
1513                 id_priv->cm_id.iw = NULL;
1514         }
1515
1516         return ret;
1517 }
1518
1519 static int cma_listen_handler(struct rdma_cm_id *id,
1520                               struct rdma_cm_event *event)
1521 {
1522         struct rdma_id_private *id_priv = id->context;
1523
1524         id->context = id_priv->id.context;
1525         id->event_handler = id_priv->id.event_handler;
1526         return id_priv->id.event_handler(id, event);
1527 }
1528
1529 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1530                               struct cma_device *cma_dev)
1531 {
1532         struct rdma_id_private *dev_id_priv;
1533         struct rdma_cm_id *id;
1534         int ret;
1535
1536         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1537                             id_priv->id.qp_type);
1538         if (IS_ERR(id))
1539                 return;
1540
1541         dev_id_priv = container_of(id, struct rdma_id_private, id);
1542
1543         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1544         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1545                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1546
1547         cma_attach_to_dev(dev_id_priv, cma_dev);
1548         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1549         atomic_inc(&id_priv->refcount);
1550         dev_id_priv->internal_id = 1;
1551
1552         ret = rdma_listen(id, id_priv->backlog);
1553         if (ret)
1554                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1555                        "listening on device %s\n", ret, cma_dev->device->name);
1556 }
1557
1558 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1559 {
1560         struct cma_device *cma_dev;
1561
1562         mutex_lock(&lock);
1563         list_add_tail(&id_priv->list, &listen_any_list);
1564         list_for_each_entry(cma_dev, &dev_list, list)
1565                 cma_listen_on_dev(id_priv, cma_dev);
1566         mutex_unlock(&lock);
1567 }
1568
1569 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1570 {
1571         struct rdma_id_private *id_priv;
1572
1573         id_priv = container_of(id, struct rdma_id_private, id);
1574         id_priv->tos = (u8) tos;
1575 }
1576 EXPORT_SYMBOL(rdma_set_service_type);
1577
1578 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1579                               void *context)
1580 {
1581         struct cma_work *work = context;
1582         struct rdma_route *route;
1583
1584         route = &work->id->id.route;
1585
1586         if (!status) {
1587                 route->num_paths = 1;
1588                 *route->path_rec = *path_rec;
1589         } else {
1590                 work->old_state = RDMA_CM_ROUTE_QUERY;
1591                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1592                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1593                 work->event.status = status;
1594         }
1595
1596         queue_work(cma_wq, &work->work);
1597 }
1598
1599 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1600                               struct cma_work *work)
1601 {
1602         struct rdma_addr *addr = &id_priv->id.route.addr;
1603         struct ib_sa_path_rec path_rec;
1604         ib_sa_comp_mask comp_mask;
1605         struct sockaddr_in6 *sin6;
1606
1607         memset(&path_rec, 0, sizeof path_rec);
1608         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1609         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1610         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1611         path_rec.numb_path = 1;
1612         path_rec.reversible = 1;
1613         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1614                                                         (struct sockaddr *) &addr->dst_addr);
1615
1616         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1617                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1618                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1619
1620         if (addr->src_addr.ss_family == AF_INET) {
1621                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1622                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1623         } else {
1624                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1625                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1626                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1627         }
1628
1629         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1630                                                id_priv->id.port_num, &path_rec,
1631                                                comp_mask, timeout_ms,
1632                                                GFP_KERNEL, cma_query_handler,
1633                                                work, &id_priv->query);
1634
1635         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1636 }
1637
1638 static void cma_work_handler(struct work_struct *_work)
1639 {
1640         struct cma_work *work = container_of(_work, struct cma_work, work);
1641         struct rdma_id_private *id_priv = work->id;
1642         int destroy = 0;
1643
1644         mutex_lock(&id_priv->handler_mutex);
1645         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1646                 goto out;
1647
1648         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1649                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1650                 destroy = 1;
1651         }
1652 out:
1653         mutex_unlock(&id_priv->handler_mutex);
1654         cma_deref_id(id_priv);
1655         if (destroy)
1656                 rdma_destroy_id(&id_priv->id);
1657         kfree(work);
1658 }
1659
1660 static void cma_ndev_work_handler(struct work_struct *_work)
1661 {
1662         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1663         struct rdma_id_private *id_priv = work->id;
1664         int destroy = 0;
1665
1666         mutex_lock(&id_priv->handler_mutex);
1667         if (id_priv->state == RDMA_CM_DESTROYING ||
1668             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
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
1676 out:
1677         mutex_unlock(&id_priv->handler_mutex);
1678         cma_deref_id(id_priv);
1679         if (destroy)
1680                 rdma_destroy_id(&id_priv->id);
1681         kfree(work);
1682 }
1683
1684 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1685 {
1686         struct rdma_route *route = &id_priv->id.route;
1687         struct cma_work *work;
1688         int ret;
1689
1690         work = kzalloc(sizeof *work, GFP_KERNEL);
1691         if (!work)
1692                 return -ENOMEM;
1693
1694         work->id = id_priv;
1695         INIT_WORK(&work->work, cma_work_handler);
1696         work->old_state = RDMA_CM_ROUTE_QUERY;
1697         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1698         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1699
1700         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1701         if (!route->path_rec) {
1702                 ret = -ENOMEM;
1703                 goto err1;
1704         }
1705
1706         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1707         if (ret)
1708                 goto err2;
1709
1710         return 0;
1711 err2:
1712         kfree(route->path_rec);
1713         route->path_rec = NULL;
1714 err1:
1715         kfree(work);
1716         return ret;
1717 }
1718
1719 int rdma_set_ib_paths(struct rdma_cm_id *id,
1720                       struct ib_sa_path_rec *path_rec, int num_paths)
1721 {
1722         struct rdma_id_private *id_priv;
1723         int ret;
1724
1725         id_priv = container_of(id, struct rdma_id_private, id);
1726         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1727                            RDMA_CM_ROUTE_RESOLVED))
1728                 return -EINVAL;
1729
1730         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1731                                      GFP_KERNEL);
1732         if (!id->route.path_rec) {
1733                 ret = -ENOMEM;
1734                 goto err;
1735         }
1736
1737         id->route.num_paths = num_paths;
1738         return 0;
1739 err:
1740         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1741         return ret;
1742 }
1743 EXPORT_SYMBOL(rdma_set_ib_paths);
1744
1745 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1746 {
1747         struct cma_work *work;
1748
1749         work = kzalloc(sizeof *work, GFP_KERNEL);
1750         if (!work)
1751                 return -ENOMEM;
1752
1753         work->id = id_priv;
1754         INIT_WORK(&work->work, cma_work_handler);
1755         work->old_state = RDMA_CM_ROUTE_QUERY;
1756         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1757         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1758         queue_work(cma_wq, &work->work);
1759         return 0;
1760 }
1761
1762 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1763 {
1764         struct rdma_route *route = &id_priv->id.route;
1765         struct rdma_addr *addr = &route->addr;
1766         struct cma_work *work;
1767         int ret;
1768         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1769         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1770         struct net_device *ndev = NULL;
1771         u16 vid;
1772
1773         if (src_addr->sin_family != dst_addr->sin_family)
1774                 return -EINVAL;
1775
1776         work = kzalloc(sizeof *work, GFP_KERNEL);
1777         if (!work)
1778                 return -ENOMEM;
1779
1780         work->id = id_priv;
1781         INIT_WORK(&work->work, cma_work_handler);
1782
1783         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1784         if (!route->path_rec) {
1785                 ret = -ENOMEM;
1786                 goto err1;
1787         }
1788
1789         route->num_paths = 1;
1790
1791         if (addr->dev_addr.bound_dev_if)
1792                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1793         if (!ndev) {
1794                 ret = -ENODEV;
1795                 goto err2;
1796         }
1797
1798         vid = rdma_vlan_dev_vlan_id(ndev);
1799
1800         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1801         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1802
1803         route->path_rec->hop_limit = 1;
1804         route->path_rec->reversible = 1;
1805         route->path_rec->pkey = cpu_to_be16(0xffff);
1806         route->path_rec->mtu_selector = IB_SA_EQ;
1807         route->path_rec->sl = id_priv->tos >> 5;
1808
1809         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1810         route->path_rec->rate_selector = IB_SA_EQ;
1811         route->path_rec->rate = iboe_get_rate(ndev);
1812         dev_put(ndev);
1813         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1814         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1815         if (!route->path_rec->mtu) {
1816                 ret = -EINVAL;
1817                 goto err2;
1818         }
1819
1820         work->old_state = RDMA_CM_ROUTE_QUERY;
1821         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1822         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1823         work->event.status = 0;
1824
1825         queue_work(cma_wq, &work->work);
1826
1827         return 0;
1828
1829 err2:
1830         kfree(route->path_rec);
1831         route->path_rec = NULL;
1832 err1:
1833         kfree(work);
1834         return ret;
1835 }
1836
1837 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1838 {
1839         struct rdma_id_private *id_priv;
1840         int ret;
1841
1842         id_priv = container_of(id, struct rdma_id_private, id);
1843         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1844                 return -EINVAL;
1845
1846         atomic_inc(&id_priv->refcount);
1847         switch (rdma_node_get_transport(id->device->node_type)) {
1848         case RDMA_TRANSPORT_IB:
1849                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1850                 case IB_LINK_LAYER_INFINIBAND:
1851                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1852                         break;
1853                 case IB_LINK_LAYER_ETHERNET:
1854                         ret = cma_resolve_iboe_route(id_priv);
1855                         break;
1856                 default:
1857                         ret = -ENOSYS;
1858                 }
1859                 break;
1860         case RDMA_TRANSPORT_IWARP:
1861                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1862                 break;
1863         default:
1864                 ret = -ENOSYS;
1865                 break;
1866         }
1867         if (ret)
1868                 goto err;
1869
1870         return 0;
1871 err:
1872         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1873         cma_deref_id(id_priv);
1874         return ret;
1875 }
1876 EXPORT_SYMBOL(rdma_resolve_route);
1877
1878 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1879 {
1880         struct cma_device *cma_dev;
1881         struct ib_port_attr port_attr;
1882         union ib_gid gid;
1883         u16 pkey;
1884         int ret;
1885         u8 p;
1886
1887         mutex_lock(&lock);
1888         if (list_empty(&dev_list)) {
1889                 ret = -ENODEV;
1890                 goto out;
1891         }
1892         list_for_each_entry(cma_dev, &dev_list, list)
1893                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1894                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1895                             port_attr.state == IB_PORT_ACTIVE)
1896                                 goto port_found;
1897
1898         p = 1;
1899         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1900
1901 port_found:
1902         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1903         if (ret)
1904                 goto out;
1905
1906         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1907         if (ret)
1908                 goto out;
1909
1910         id_priv->id.route.addr.dev_addr.dev_type =
1911                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1912                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1913
1914         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1915         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1916         id_priv->id.port_num = p;
1917         cma_attach_to_dev(id_priv, cma_dev);
1918 out:
1919         mutex_unlock(&lock);
1920         return ret;
1921 }
1922
1923 static void addr_handler(int status, struct sockaddr *src_addr,
1924                          struct rdma_dev_addr *dev_addr, void *context)
1925 {
1926         struct rdma_id_private *id_priv = context;
1927         struct rdma_cm_event event;
1928
1929         memset(&event, 0, sizeof event);
1930         mutex_lock(&id_priv->handler_mutex);
1931         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1932                            RDMA_CM_ADDR_RESOLVED))
1933                 goto out;
1934
1935         if (!status && !id_priv->cma_dev)
1936                 status = cma_acquire_dev(id_priv);
1937
1938         if (status) {
1939                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1940                                    RDMA_CM_ADDR_BOUND))
1941                         goto out;
1942                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1943                 event.status = status;
1944         } else {
1945                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1946                        ip_addr_size(src_addr));
1947                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1948         }
1949
1950         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1951                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1952                 mutex_unlock(&id_priv->handler_mutex);
1953                 cma_deref_id(id_priv);
1954                 rdma_destroy_id(&id_priv->id);
1955                 return;
1956         }
1957 out:
1958         mutex_unlock(&id_priv->handler_mutex);
1959         cma_deref_id(id_priv);
1960 }
1961
1962 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1963 {
1964         struct cma_work *work;
1965         struct sockaddr *src, *dst;
1966         union ib_gid gid;
1967         int ret;
1968
1969         work = kzalloc(sizeof *work, GFP_KERNEL);
1970         if (!work)
1971                 return -ENOMEM;
1972
1973         if (!id_priv->cma_dev) {
1974                 ret = cma_bind_loopback(id_priv);
1975                 if (ret)
1976                         goto err;
1977         }
1978
1979         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1980         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1981
1982         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1983         if (cma_zero_addr(src)) {
1984                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
1985                 if ((src->sa_family = dst->sa_family) == AF_INET) {
1986                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
1987                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1988                 } else {
1989                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
1990                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
1991                 }
1992         }
1993
1994         work->id = id_priv;
1995         INIT_WORK(&work->work, cma_work_handler);
1996         work->old_state = RDMA_CM_ADDR_QUERY;
1997         work->new_state = RDMA_CM_ADDR_RESOLVED;
1998         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1999         queue_work(cma_wq, &work->work);
2000         return 0;
2001 err:
2002         kfree(work);
2003         return ret;
2004 }
2005
2006 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2007                          struct sockaddr *dst_addr)
2008 {
2009         if (!src_addr || !src_addr->sa_family) {
2010                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2011                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2012                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2013                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2014                 }
2015         }
2016         return rdma_bind_addr(id, src_addr);
2017 }
2018
2019 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2020                       struct sockaddr *dst_addr, int timeout_ms)
2021 {
2022         struct rdma_id_private *id_priv;
2023         int ret;
2024
2025         id_priv = container_of(id, struct rdma_id_private, id);
2026         if (id_priv->state == RDMA_CM_IDLE) {
2027                 ret = cma_bind_addr(id, src_addr, dst_addr);
2028                 if (ret)
2029                         return ret;
2030         }
2031
2032         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2033                 return -EINVAL;
2034
2035         atomic_inc(&id_priv->refcount);
2036         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2037         if (cma_any_addr(dst_addr))
2038                 ret = cma_resolve_loopback(id_priv);
2039         else
2040                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2041                                       dst_addr, &id->route.addr.dev_addr,
2042                                       timeout_ms, addr_handler, id_priv);
2043         if (ret)
2044                 goto err;
2045
2046         return 0;
2047 err:
2048         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2049         cma_deref_id(id_priv);
2050         return ret;
2051 }
2052 EXPORT_SYMBOL(rdma_resolve_addr);
2053
2054 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2055 {
2056         struct rdma_id_private *id_priv;
2057         unsigned long flags;
2058         int ret;
2059
2060         id_priv = container_of(id, struct rdma_id_private, id);
2061         spin_lock_irqsave(&id_priv->lock, flags);
2062         if (id_priv->state == RDMA_CM_IDLE) {
2063                 id_priv->reuseaddr = reuse;
2064                 ret = 0;
2065         } else {
2066                 ret = -EINVAL;
2067         }
2068         spin_unlock_irqrestore(&id_priv->lock, flags);
2069         return ret;
2070 }
2071 EXPORT_SYMBOL(rdma_set_reuseaddr);
2072
2073 static void cma_bind_port(struct rdma_bind_list *bind_list,
2074                           struct rdma_id_private *id_priv)
2075 {
2076         struct sockaddr_in *sin;
2077
2078         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2079         sin->sin_port = htons(bind_list->port);
2080         id_priv->bind_list = bind_list;
2081         hlist_add_head(&id_priv->node, &bind_list->owners);
2082 }
2083
2084 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2085                           unsigned short snum)
2086 {
2087         struct rdma_bind_list *bind_list;
2088         int port, ret;
2089
2090         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2091         if (!bind_list)
2092                 return -ENOMEM;
2093
2094         do {
2095                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2096         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2097
2098         if (ret)
2099                 goto err1;
2100
2101         if (port != snum) {
2102                 ret = -EADDRNOTAVAIL;
2103                 goto err2;
2104         }
2105
2106         bind_list->ps = ps;
2107         bind_list->port = (unsigned short) port;
2108         cma_bind_port(bind_list, id_priv);
2109         return 0;
2110 err2:
2111         idr_remove(ps, port);
2112 err1:
2113         kfree(bind_list);
2114         return ret;
2115 }
2116
2117 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2118 {
2119         static unsigned int last_used_port;
2120         int low, high, remaining;
2121         unsigned int rover;
2122
2123         inet_get_local_port_range(&low, &high);
2124         remaining = (high - low) + 1;
2125         rover = net_random() % remaining + low;
2126 retry:
2127         if (last_used_port != rover &&
2128             !idr_find(ps, (unsigned short) rover)) {
2129                 int ret = cma_alloc_port(ps, id_priv, rover);
2130                 /*
2131                  * Remember previously used port number in order to avoid
2132                  * re-using same port immediately after it is closed.
2133                  */
2134                 if (!ret)
2135                         last_used_port = rover;
2136                 if (ret != -EADDRNOTAVAIL)
2137                         return ret;
2138         }
2139         if (--remaining) {
2140                 rover++;
2141                 if ((rover < low) || (rover > high))
2142                         rover = low;
2143                 goto retry;
2144         }
2145         return -EADDRNOTAVAIL;
2146 }
2147
2148 /*
2149  * Check that the requested port is available.  This is called when trying to
2150  * bind to a specific port, or when trying to listen on a bound port.  In
2151  * the latter case, the provided id_priv may already be on the bind_list, but
2152  * we still need to check that it's okay to start listening.
2153  */
2154 static int cma_check_port(struct rdma_bind_list *bind_list,
2155                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2156 {
2157         struct rdma_id_private *cur_id;
2158         struct sockaddr *addr, *cur_addr;
2159         struct hlist_node *node;
2160
2161         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2162         if (cma_any_addr(addr) && !reuseaddr)
2163                 return -EADDRNOTAVAIL;
2164
2165         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2166                 if (id_priv == cur_id)
2167                         continue;
2168
2169                 if ((cur_id->state == RDMA_CM_LISTEN) ||
2170                     !reuseaddr || !cur_id->reuseaddr) {
2171                         cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2172                         if (cma_any_addr(cur_addr))
2173                                 return -EADDRNOTAVAIL;
2174
2175                         if (!cma_addr_cmp(addr, cur_addr))
2176                                 return -EADDRINUSE;
2177                 }
2178         }
2179         return 0;
2180 }
2181
2182 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2183 {
2184         struct rdma_bind_list *bind_list;
2185         unsigned short snum;
2186         int ret;
2187
2188         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2189         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2190                 return -EACCES;
2191
2192         bind_list = idr_find(ps, snum);
2193         if (!bind_list) {
2194                 ret = cma_alloc_port(ps, id_priv, snum);
2195         } else {
2196                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2197                 if (!ret)
2198                         cma_bind_port(bind_list, id_priv);
2199         }
2200         return ret;
2201 }
2202
2203 static int cma_bind_listen(struct rdma_id_private *id_priv)
2204 {
2205         struct rdma_bind_list *bind_list = id_priv->bind_list;
2206         int ret = 0;
2207
2208         mutex_lock(&lock);
2209         if (bind_list->owners.first->next)
2210                 ret = cma_check_port(bind_list, id_priv, 0);
2211         mutex_unlock(&lock);
2212         return ret;
2213 }
2214
2215 static int cma_get_port(struct rdma_id_private *id_priv)
2216 {
2217         struct idr *ps;
2218         int ret;
2219
2220         switch (id_priv->id.ps) {
2221         case RDMA_PS_SDP:
2222                 ps = &sdp_ps;
2223                 break;
2224         case RDMA_PS_TCP:
2225                 ps = &tcp_ps;
2226                 break;
2227         case RDMA_PS_UDP:
2228                 ps = &udp_ps;
2229                 break;
2230         case RDMA_PS_IPOIB:
2231                 ps = &ipoib_ps;
2232                 break;
2233         default:
2234                 return -EPROTONOSUPPORT;
2235         }
2236
2237         mutex_lock(&lock);
2238         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2239                 ret = cma_alloc_any_port(ps, id_priv);
2240         else
2241                 ret = cma_use_port(ps, id_priv);
2242         mutex_unlock(&lock);
2243
2244         return ret;
2245 }
2246
2247 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2248                                struct sockaddr *addr)
2249 {
2250 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2251         struct sockaddr_in6 *sin6;
2252
2253         if (addr->sa_family != AF_INET6)
2254                 return 0;
2255
2256         sin6 = (struct sockaddr_in6 *) addr;
2257         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2258             !sin6->sin6_scope_id)
2259                         return -EINVAL;
2260
2261         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2262 #endif
2263         return 0;
2264 }
2265
2266 int rdma_listen(struct rdma_cm_id *id, int backlog)
2267 {
2268         struct rdma_id_private *id_priv;
2269         int ret;
2270
2271         id_priv = container_of(id, struct rdma_id_private, id);
2272         if (id_priv->state == RDMA_CM_IDLE) {
2273                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2274                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2275                 if (ret)
2276                         return ret;
2277         }
2278
2279         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2280                 return -EINVAL;
2281
2282         if (id_priv->reuseaddr) {
2283                 ret = cma_bind_listen(id_priv);
2284                 if (ret)
2285                         goto err;
2286         }
2287
2288         id_priv->backlog = backlog;
2289         if (id->device) {
2290                 switch (rdma_node_get_transport(id->device->node_type)) {
2291                 case RDMA_TRANSPORT_IB:
2292                         ret = cma_ib_listen(id_priv);
2293                         if (ret)
2294                                 goto err;
2295                         break;
2296                 case RDMA_TRANSPORT_IWARP:
2297                         ret = cma_iw_listen(id_priv, backlog);
2298                         if (ret)
2299                                 goto err;
2300                         break;
2301                 default:
2302                         ret = -ENOSYS;
2303                         goto err;
2304                 }
2305         } else
2306                 cma_listen_on_all(id_priv);
2307
2308         return 0;
2309 err:
2310         id_priv->backlog = 0;
2311         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2312         return ret;
2313 }
2314 EXPORT_SYMBOL(rdma_listen);
2315
2316 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2317 {
2318         struct rdma_id_private *id_priv;
2319         int ret;
2320
2321         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2322                 return -EAFNOSUPPORT;
2323
2324         id_priv = container_of(id, struct rdma_id_private, id);
2325         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2326                 return -EINVAL;
2327
2328         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2329         if (ret)
2330                 goto err1;
2331
2332         if (!cma_any_addr(addr)) {
2333                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2334                 if (ret)
2335                         goto err1;
2336
2337                 ret = cma_acquire_dev(id_priv);
2338                 if (ret)
2339                         goto err1;
2340         }
2341
2342         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2343         ret = cma_get_port(id_priv);
2344         if (ret)
2345                 goto err2;
2346
2347         return 0;
2348 err2:
2349         if (id_priv->cma_dev)
2350                 cma_release_dev(id_priv);
2351 err1:
2352         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2353         return ret;
2354 }
2355 EXPORT_SYMBOL(rdma_bind_addr);
2356
2357 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2358                           struct rdma_route *route)
2359 {
2360         struct cma_hdr *cma_hdr;
2361         struct sdp_hh *sdp_hdr;
2362
2363         if (route->addr.src_addr.ss_family == AF_INET) {
2364                 struct sockaddr_in *src4, *dst4;
2365
2366                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2367                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2368
2369                 switch (ps) {
2370                 case RDMA_PS_SDP:
2371                         sdp_hdr = hdr;
2372                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2373                                 return -EINVAL;
2374                         sdp_set_ip_ver(sdp_hdr, 4);
2375                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2376                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2377                         sdp_hdr->port = src4->sin_port;
2378                         break;
2379                 default:
2380                         cma_hdr = hdr;
2381                         cma_hdr->cma_version = CMA_VERSION;
2382                         cma_set_ip_ver(cma_hdr, 4);
2383                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2384                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2385                         cma_hdr->port = src4->sin_port;
2386                         break;
2387                 }
2388         } else {
2389                 struct sockaddr_in6 *src6, *dst6;
2390
2391                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2392                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2393
2394                 switch (ps) {
2395                 case RDMA_PS_SDP:
2396                         sdp_hdr = hdr;
2397                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2398                                 return -EINVAL;
2399                         sdp_set_ip_ver(sdp_hdr, 6);
2400                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2401                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2402                         sdp_hdr->port = src6->sin6_port;
2403                         break;
2404                 default:
2405                         cma_hdr = hdr;
2406                         cma_hdr->cma_version = CMA_VERSION;
2407                         cma_set_ip_ver(cma_hdr, 6);
2408                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2409                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2410                         cma_hdr->port = src6->sin6_port;
2411                         break;
2412                 }
2413         }
2414         return 0;
2415 }
2416
2417 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2418                                 struct ib_cm_event *ib_event)
2419 {
2420         struct rdma_id_private *id_priv = cm_id->context;
2421         struct rdma_cm_event event;
2422         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2423         int ret = 0;
2424
2425         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2426                 return 0;
2427
2428         memset(&event, 0, sizeof event);
2429         switch (ib_event->event) {
2430         case IB_CM_SIDR_REQ_ERROR:
2431                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2432                 event.status = -ETIMEDOUT;
2433                 break;
2434         case IB_CM_SIDR_REP_RECEIVED:
2435                 event.param.ud.private_data = ib_event->private_data;
2436                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2437                 if (rep->status != IB_SIDR_SUCCESS) {
2438                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2439                         event.status = ib_event->param.sidr_rep_rcvd.status;
2440                         break;
2441                 }
2442                 ret = cma_set_qkey(id_priv);
2443                 if (ret) {
2444                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2445                         event.status = -EINVAL;
2446                         break;
2447                 }
2448                 if (id_priv->qkey != rep->qkey) {
2449                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2450                         event.status = -EINVAL;
2451                         break;
2452                 }
2453                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2454                                      id_priv->id.route.path_rec,
2455                                      &event.param.ud.ah_attr);
2456                 event.param.ud.qp_num = rep->qpn;
2457                 event.param.ud.qkey = rep->qkey;
2458                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2459                 event.status = 0;
2460                 break;
2461         default:
2462                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2463                        ib_event->event);
2464                 goto out;
2465         }
2466
2467         ret = id_priv->id.event_handler(&id_priv->id, &event);
2468         if (ret) {
2469                 /* Destroy the CM ID by returning a non-zero value. */
2470                 id_priv->cm_id.ib = NULL;
2471                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2472                 mutex_unlock(&id_priv->handler_mutex);
2473                 rdma_destroy_id(&id_priv->id);
2474                 return ret;
2475         }
2476 out:
2477         mutex_unlock(&id_priv->handler_mutex);
2478         return ret;
2479 }
2480
2481 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2482                               struct rdma_conn_param *conn_param)
2483 {
2484         struct ib_cm_sidr_req_param req;
2485         struct rdma_route *route;
2486         struct ib_cm_id *id;
2487         int ret;
2488
2489         req.private_data_len = sizeof(struct cma_hdr) +
2490                                conn_param->private_data_len;
2491         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2492         if (!req.private_data)
2493                 return -ENOMEM;
2494
2495         if (conn_param->private_data && conn_param->private_data_len)
2496                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2497                        conn_param->private_data, conn_param->private_data_len);
2498
2499         route = &id_priv->id.route;
2500         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2501         if (ret)
2502                 goto out;
2503
2504         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2505                              id_priv);
2506         if (IS_ERR(id)) {
2507                 ret = PTR_ERR(id);
2508                 goto out;
2509         }
2510         id_priv->cm_id.ib = id;
2511
2512         req.path = route->path_rec;
2513         req.service_id = cma_get_service_id(id_priv->id.ps,
2514                                             (struct sockaddr *) &route->addr.dst_addr);
2515         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2516         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2517
2518         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2519         if (ret) {
2520                 ib_destroy_cm_id(id_priv->cm_id.ib);
2521                 id_priv->cm_id.ib = NULL;
2522         }
2523 out:
2524         kfree(req.private_data);
2525         return ret;
2526 }
2527
2528 static int cma_connect_ib(struct rdma_id_private *id_priv,
2529                           struct rdma_conn_param *conn_param)
2530 {
2531         struct ib_cm_req_param req;
2532         struct rdma_route *route;
2533         void *private_data;
2534         struct ib_cm_id *id;
2535         int offset, ret;
2536
2537         memset(&req, 0, sizeof req);
2538         offset = cma_user_data_offset(id_priv->id.ps);
2539         req.private_data_len = offset + conn_param->private_data_len;
2540         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2541         if (!private_data)
2542                 return -ENOMEM;
2543
2544         if (conn_param->private_data && conn_param->private_data_len)
2545                 memcpy(private_data + offset, conn_param->private_data,
2546                        conn_param->private_data_len);
2547
2548         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2549         if (IS_ERR(id)) {
2550                 ret = PTR_ERR(id);
2551                 goto out;
2552         }
2553         id_priv->cm_id.ib = id;
2554
2555         route = &id_priv->id.route;
2556         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2557         if (ret)
2558                 goto out;
2559         req.private_data = private_data;
2560
2561         req.primary_path = &route->path_rec[0];
2562         if (route->num_paths == 2)
2563                 req.alternate_path = &route->path_rec[1];
2564
2565         req.service_id = cma_get_service_id(id_priv->id.ps,
2566                                             (struct sockaddr *) &route->addr.dst_addr);
2567         req.qp_num = id_priv->qp_num;
2568         req.qp_type = IB_QPT_RC;
2569         req.starting_psn = id_priv->seq_num;
2570         req.responder_resources = conn_param->responder_resources;
2571         req.initiator_depth = conn_param->initiator_depth;
2572         req.flow_control = conn_param->flow_control;
2573         req.retry_count = conn_param->retry_count;
2574         req.rnr_retry_count = conn_param->rnr_retry_count;
2575         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2576         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2577         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2578         req.srq = id_priv->srq ? 1 : 0;
2579
2580         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2581 out:
2582         if (ret && !IS_ERR(id)) {
2583                 ib_destroy_cm_id(id);
2584                 id_priv->cm_id.ib = NULL;
2585         }
2586
2587         kfree(private_data);
2588         return ret;
2589 }
2590
2591 static int cma_connect_iw(struct rdma_id_private *id_priv,
2592                           struct rdma_conn_param *conn_param)
2593 {
2594         struct iw_cm_id *cm_id;
2595         struct sockaddr_in* sin;
2596         int ret;
2597         struct iw_cm_conn_param iw_param;
2598
2599         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2600         if (IS_ERR(cm_id))
2601                 return PTR_ERR(cm_id);
2602
2603         id_priv->cm_id.iw = cm_id;
2604
2605         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2606         cm_id->local_addr = *sin;
2607
2608         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2609         cm_id->remote_addr = *sin;
2610
2611         ret = cma_modify_qp_rtr(id_priv, conn_param);
2612         if (ret)
2613                 goto out;
2614
2615         iw_param.ord = conn_param->initiator_depth;
2616         iw_param.ird = conn_param->responder_resources;
2617         iw_param.private_data = conn_param->private_data;
2618         iw_param.private_data_len = conn_param->private_data_len;
2619         if (id_priv->id.qp)
2620                 iw_param.qpn = id_priv->qp_num;
2621         else
2622                 iw_param.qpn = conn_param->qp_num;
2623         ret = iw_cm_connect(cm_id, &iw_param);
2624 out:
2625         if (ret) {
2626                 iw_destroy_cm_id(cm_id);
2627                 id_priv->cm_id.iw = NULL;
2628         }
2629         return ret;
2630 }
2631
2632 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2633 {
2634         struct rdma_id_private *id_priv;
2635         int ret;
2636
2637         id_priv = container_of(id, struct rdma_id_private, id);
2638         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2639                 return -EINVAL;
2640
2641         if (!id->qp) {
2642                 id_priv->qp_num = conn_param->qp_num;
2643                 id_priv->srq = conn_param->srq;
2644         }
2645
2646         switch (rdma_node_get_transport(id->device->node_type)) {
2647         case RDMA_TRANSPORT_IB:
2648                 if (id->qp_type == IB_QPT_UD)
2649                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2650                 else
2651                         ret = cma_connect_ib(id_priv, conn_param);
2652                 break;
2653         case RDMA_TRANSPORT_IWARP:
2654                 ret = cma_connect_iw(id_priv, conn_param);
2655                 break;
2656         default:
2657                 ret = -ENOSYS;
2658                 break;
2659         }
2660         if (ret)
2661                 goto err;
2662
2663         return 0;
2664 err:
2665         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2666         return ret;
2667 }
2668 EXPORT_SYMBOL(rdma_connect);
2669
2670 static int cma_accept_ib(struct rdma_id_private *id_priv,
2671                          struct rdma_conn_param *conn_param)
2672 {
2673         struct ib_cm_rep_param rep;
2674         int ret;
2675
2676         ret = cma_modify_qp_rtr(id_priv, conn_param);
2677         if (ret)
2678                 goto out;
2679
2680         ret = cma_modify_qp_rts(id_priv, conn_param);
2681         if (ret)
2682                 goto out;
2683
2684         memset(&rep, 0, sizeof rep);
2685         rep.qp_num = id_priv->qp_num;
2686         rep.starting_psn = id_priv->seq_num;
2687         rep.private_data = conn_param->private_data;
2688         rep.private_data_len = conn_param->private_data_len;
2689         rep.responder_resources = conn_param->responder_resources;
2690         rep.initiator_depth = conn_param->initiator_depth;
2691         rep.failover_accepted = 0;
2692         rep.flow_control = conn_param->flow_control;
2693         rep.rnr_retry_count = conn_param->rnr_retry_count;
2694         rep.srq = id_priv->srq ? 1 : 0;
2695
2696         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2697 out:
2698         return ret;
2699 }
2700
2701 static int cma_accept_iw(struct rdma_id_private *id_priv,
2702                   struct rdma_conn_param *conn_param)
2703 {
2704         struct iw_cm_conn_param iw_param;
2705         int ret;
2706
2707         ret = cma_modify_qp_rtr(id_priv, conn_param);
2708         if (ret)
2709                 return ret;
2710
2711         iw_param.ord = conn_param->initiator_depth;
2712         iw_param.ird = conn_param->responder_resources;
2713         iw_param.private_data = conn_param->private_data;
2714         iw_param.private_data_len = conn_param->private_data_len;
2715         if (id_priv->id.qp) {
2716                 iw_param.qpn = id_priv->qp_num;
2717         } else
2718                 iw_param.qpn = conn_param->qp_num;
2719
2720         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2721 }
2722
2723 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2724                              enum ib_cm_sidr_status status,
2725                              const void *private_data, int private_data_len)
2726 {
2727         struct ib_cm_sidr_rep_param rep;
2728         int ret;
2729
2730         memset(&rep, 0, sizeof rep);
2731         rep.status = status;
2732         if (status == IB_SIDR_SUCCESS) {
2733                 ret = cma_set_qkey(id_priv);
2734                 if (ret)
2735                         return ret;
2736                 rep.qp_num = id_priv->qp_num;
2737                 rep.qkey = id_priv->qkey;
2738         }
2739         rep.private_data = private_data;
2740         rep.private_data_len = private_data_len;
2741
2742         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2743 }
2744
2745 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2746 {
2747         struct rdma_id_private *id_priv;
2748         int ret;
2749
2750         id_priv = container_of(id, struct rdma_id_private, id);
2751
2752         id_priv->owner = task_pid_nr(current);
2753
2754         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2755                 return -EINVAL;
2756
2757         if (!id->qp && conn_param) {
2758                 id_priv->qp_num = conn_param->qp_num;
2759                 id_priv->srq = conn_param->srq;
2760         }
2761
2762         switch (rdma_node_get_transport(id->device->node_type)) {
2763         case RDMA_TRANSPORT_IB:
2764                 if (id->qp_type == IB_QPT_UD)
2765                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2766                                                 conn_param->private_data,
2767                                                 conn_param->private_data_len);
2768                 else if (conn_param)
2769                         ret = cma_accept_ib(id_priv, conn_param);
2770                 else
2771                         ret = cma_rep_recv(id_priv);
2772                 break;
2773         case RDMA_TRANSPORT_IWARP:
2774                 ret = cma_accept_iw(id_priv, conn_param);
2775                 break;
2776         default:
2777                 ret = -ENOSYS;
2778                 break;
2779         }
2780
2781         if (ret)
2782                 goto reject;
2783
2784         return 0;
2785 reject:
2786         cma_modify_qp_err(id_priv);
2787         rdma_reject(id, NULL, 0);
2788         return ret;
2789 }
2790 EXPORT_SYMBOL(rdma_accept);
2791
2792 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2793 {
2794         struct rdma_id_private *id_priv;
2795         int ret;
2796
2797         id_priv = container_of(id, struct rdma_id_private, id);
2798         if (!id_priv->cm_id.ib)
2799                 return -EINVAL;
2800
2801         switch (id->device->node_type) {
2802         case RDMA_NODE_IB_CA:
2803                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2804                 break;
2805         default:
2806                 ret = 0;
2807                 break;
2808         }
2809         return ret;
2810 }
2811 EXPORT_SYMBOL(rdma_notify);
2812
2813 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2814                 u8 private_data_len)
2815 {
2816         struct rdma_id_private *id_priv;
2817         int ret;
2818
2819         id_priv = container_of(id, struct rdma_id_private, id);
2820         if (!id_priv->cm_id.ib)
2821                 return -EINVAL;
2822
2823         switch (rdma_node_get_transport(id->device->node_type)) {
2824         case RDMA_TRANSPORT_IB:
2825                 if (id->qp_type == IB_QPT_UD)
2826                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2827                                                 private_data, private_data_len);
2828                 else
2829                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2830                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2831                                              0, private_data, private_data_len);
2832                 break;
2833         case RDMA_TRANSPORT_IWARP:
2834                 ret = iw_cm_reject(id_priv->cm_id.iw,
2835                                    private_data, private_data_len);
2836                 break;
2837         default:
2838                 ret = -ENOSYS;
2839                 break;
2840         }
2841         return ret;
2842 }
2843 EXPORT_SYMBOL(rdma_reject);
2844
2845 int rdma_disconnect(struct rdma_cm_id *id)
2846 {
2847         struct rdma_id_private *id_priv;
2848         int ret;
2849
2850         id_priv = container_of(id, struct rdma_id_private, id);
2851         if (!id_priv->cm_id.ib)
2852                 return -EINVAL;
2853
2854         switch (rdma_node_get_transport(id->device->node_type)) {
2855         case RDMA_TRANSPORT_IB:
2856                 ret = cma_modify_qp_err(id_priv);
2857                 if (ret)
2858                         goto out;
2859                 /* Initiate or respond to a disconnect. */
2860                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2861                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2862                 break;
2863         case RDMA_TRANSPORT_IWARP:
2864                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2865                 break;
2866         default:
2867                 ret = -EINVAL;
2868                 break;
2869         }
2870 out:
2871         return ret;
2872 }
2873 EXPORT_SYMBOL(rdma_disconnect);
2874
2875 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2876 {
2877         struct rdma_id_private *id_priv;
2878         struct cma_multicast *mc = multicast->context;
2879         struct rdma_cm_event event;
2880         int ret;
2881
2882         id_priv = mc->id_priv;
2883         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2884             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2885                 return 0;
2886
2887         mutex_lock(&id_priv->qp_mutex);
2888         if (!status && id_priv->id.qp)
2889                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2890                                          multicast->rec.mlid);
2891         mutex_unlock(&id_priv->qp_mutex);
2892
2893         memset(&event, 0, sizeof event);
2894         event.status = status;
2895         event.param.ud.private_data = mc->context;
2896         if (!status) {
2897                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2898                 ib_init_ah_from_mcmember(id_priv->id.device,
2899                                          id_priv->id.port_num, &multicast->rec,
2900                                          &event.param.ud.ah_attr);
2901                 event.param.ud.qp_num = 0xFFFFFF;
2902                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2903         } else
2904                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2905
2906         ret = id_priv->id.event_handler(&id_priv->id, &event);
2907         if (ret) {
2908                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2909                 mutex_unlock(&id_priv->handler_mutex);
2910                 rdma_destroy_id(&id_priv->id);
2911                 return 0;
2912         }
2913
2914         mutex_unlock(&id_priv->handler_mutex);
2915         return 0;
2916 }
2917
2918 static void cma_set_mgid(struct rdma_id_private *id_priv,
2919                          struct sockaddr *addr, union ib_gid *mgid)
2920 {
2921         unsigned char mc_map[MAX_ADDR_LEN];
2922         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2923         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2924         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2925
2926         if (cma_any_addr(addr)) {
2927                 memset(mgid, 0, sizeof *mgid);
2928         } else if ((addr->sa_family == AF_INET6) &&
2929                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2930                                                                  0xFF10A01B)) {
2931                 /* IPv6 address is an SA assigned MGID. */
2932                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2933         } else if ((addr->sa_family == AF_INET6)) {
2934                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2935                 if (id_priv->id.ps == RDMA_PS_UDP)
2936                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2937                 *mgid = *(union ib_gid *) (mc_map + 4);
2938         } else {
2939                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2940                 if (id_priv->id.ps == RDMA_PS_UDP)
2941                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2942                 *mgid = *(union ib_gid *) (mc_map + 4);
2943         }
2944 }
2945
2946 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2947                                  struct cma_multicast *mc)
2948 {
2949         struct ib_sa_mcmember_rec rec;
2950         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2951         ib_sa_comp_mask comp_mask;
2952         int ret;
2953
2954         ib_addr_get_mgid(dev_addr, &rec.mgid);
2955         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2956                                      &rec.mgid, &rec);
2957         if (ret)
2958                 return ret;
2959
2960         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2961         if (id_priv->id.ps == RDMA_PS_UDP)
2962                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2963         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2964         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2965         rec.join_state = 1;
2966
2967         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2968                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2969                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2970                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2971                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2972
2973         if (id_priv->id.ps == RDMA_PS_IPOIB)
2974                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2975                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2976
2977         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2978                                                 id_priv->id.port_num, &rec,
2979                                                 comp_mask, GFP_KERNEL,
2980                                                 cma_ib_mc_handler, mc);
2981         if (IS_ERR(mc->multicast.ib))
2982                 return PTR_ERR(mc->multicast.ib);
2983
2984         return 0;
2985 }
2986
2987 static void iboe_mcast_work_handler(struct work_struct *work)
2988 {
2989         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
2990         struct cma_multicast *mc = mw->mc;
2991         struct ib_sa_multicast *m = mc->multicast.ib;
2992
2993         mc->multicast.ib->context = mc;
2994         cma_ib_mc_handler(0, m);
2995         kref_put(&mc->mcref, release_mc);
2996         kfree(mw);
2997 }
2998
2999 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3000 {
3001         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3002         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3003
3004         if (cma_any_addr(addr)) {
3005                 memset(mgid, 0, sizeof *mgid);
3006         } else if (addr->sa_family == AF_INET6) {
3007                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3008         } else {
3009                 mgid->raw[0] = 0xff;
3010                 mgid->raw[1] = 0x0e;
3011                 mgid->raw[2] = 0;
3012                 mgid->raw[3] = 0;
3013                 mgid->raw[4] = 0;
3014                 mgid->raw[5] = 0;
3015                 mgid->raw[6] = 0;
3016                 mgid->raw[7] = 0;
3017                 mgid->raw[8] = 0;
3018                 mgid->raw[9] = 0;
3019                 mgid->raw[10] = 0xff;
3020                 mgid->raw[11] = 0xff;
3021                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3022         }
3023 }
3024
3025 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3026                                    struct cma_multicast *mc)
3027 {
3028         struct iboe_mcast_work *work;
3029         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3030         int err;
3031         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3032         struct net_device *ndev = NULL;
3033
3034         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3035                 return -EINVAL;
3036
3037         work = kzalloc(sizeof *work, GFP_KERNEL);
3038         if (!work)
3039                 return -ENOMEM;
3040
3041         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3042         if (!mc->multicast.ib) {
3043                 err = -ENOMEM;
3044                 goto out1;
3045         }
3046
3047         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3048
3049         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3050         if (id_priv->id.ps == RDMA_PS_UDP)
3051                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3052
3053         if (dev_addr->bound_dev_if)
3054                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3055         if (!ndev) {
3056                 err = -ENODEV;
3057                 goto out2;
3058         }
3059         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3060         mc->multicast.ib->rec.hop_limit = 1;
3061         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3062         dev_put(ndev);
3063         if (!mc->multicast.ib->rec.mtu) {
3064                 err = -EINVAL;
3065                 goto out2;
3066         }
3067         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3068         work->id = id_priv;
3069         work->mc = mc;
3070         INIT_WORK(&work->work, iboe_mcast_work_handler);
3071         kref_get(&mc->mcref);
3072         queue_work(cma_wq, &work->work);
3073
3074         return 0;
3075
3076 out2:
3077         kfree(mc->multicast.ib);
3078 out1:
3079         kfree(work);
3080         return err;
3081 }
3082
3083 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3084                         void *context)
3085 {
3086         struct rdma_id_private *id_priv;
3087         struct cma_multicast *mc;
3088         int ret;
3089
3090         id_priv = container_of(id, struct rdma_id_private, id);
3091         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3092             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3093                 return -EINVAL;
3094
3095         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3096         if (!mc)
3097                 return -ENOMEM;
3098
3099         memcpy(&mc->addr, addr, ip_addr_size(addr));
3100         mc->context = context;
3101         mc->id_priv = id_priv;
3102
3103         spin_lock(&id_priv->lock);
3104         list_add(&mc->list, &id_priv->mc_list);
3105         spin_unlock(&id_priv->lock);
3106
3107         switch (rdma_node_get_transport(id->device->node_type)) {
3108         case RDMA_TRANSPORT_IB:
3109                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3110                 case IB_LINK_LAYER_INFINIBAND:
3111                         ret = cma_join_ib_multicast(id_priv, mc);
3112                         break;
3113                 case IB_LINK_LAYER_ETHERNET:
3114                         kref_init(&mc->mcref);
3115                         ret = cma_iboe_join_multicast(id_priv, mc);
3116                         break;
3117                 default:
3118                         ret = -EINVAL;
3119                 }
3120                 break;
3121         default:
3122                 ret = -ENOSYS;
3123                 break;
3124         }
3125
3126         if (ret) {
3127                 spin_lock_irq(&id_priv->lock);
3128                 list_del(&mc->list);
3129                 spin_unlock_irq(&id_priv->lock);
3130                 kfree(mc);
3131         }
3132         return ret;
3133 }
3134 EXPORT_SYMBOL(rdma_join_multicast);
3135
3136 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3137 {
3138         struct rdma_id_private *id_priv;
3139         struct cma_multicast *mc;
3140
3141         id_priv = container_of(id, struct rdma_id_private, id);
3142         spin_lock_irq(&id_priv->lock);
3143         list_for_each_entry(mc, &id_priv->mc_list, list) {
3144                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3145                         list_del(&mc->list);
3146                         spin_unlock_irq(&id_priv->lock);
3147
3148                         if (id->qp)
3149                                 ib_detach_mcast(id->qp,
3150                                                 &mc->multicast.ib->rec.mgid,
3151                                                 mc->multicast.ib->rec.mlid);
3152                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3153                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3154                                 case IB_LINK_LAYER_INFINIBAND:
3155                                         ib_sa_free_multicast(mc->multicast.ib);
3156                                         kfree(mc);
3157                                         break;
3158                                 case IB_LINK_LAYER_ETHERNET:
3159                                         kref_put(&mc->mcref, release_mc);
3160                                         break;
3161                                 default:
3162                                         break;
3163                                 }
3164                         }
3165                         return;
3166                 }
3167         }
3168         spin_unlock_irq(&id_priv->lock);
3169 }
3170 EXPORT_SYMBOL(rdma_leave_multicast);
3171
3172 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3173 {
3174         struct rdma_dev_addr *dev_addr;
3175         struct cma_ndev_work *work;
3176
3177         dev_addr = &id_priv->id.route.addr.dev_addr;
3178
3179         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3180             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3181                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3182                        ndev->name, &id_priv->id);
3183                 work = kzalloc(sizeof *work, GFP_KERNEL);
3184                 if (!work)
3185                         return -ENOMEM;
3186
3187                 INIT_WORK(&work->work, cma_ndev_work_handler);
3188                 work->id = id_priv;
3189                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3190                 atomic_inc(&id_priv->refcount);
3191                 queue_work(cma_wq, &work->work);
3192         }
3193
3194         return 0;
3195 }
3196
3197 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3198                                void *ctx)
3199 {
3200         struct net_device *ndev = (struct net_device *)ctx;
3201         struct cma_device *cma_dev;
3202         struct rdma_id_private *id_priv;
3203         int ret = NOTIFY_DONE;
3204
3205         if (dev_net(ndev) != &init_net)
3206                 return NOTIFY_DONE;
3207
3208         if (event != NETDEV_BONDING_FAILOVER)
3209                 return NOTIFY_DONE;
3210
3211         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3212                 return NOTIFY_DONE;
3213
3214         mutex_lock(&lock);
3215         list_for_each_entry(cma_dev, &dev_list, list)
3216                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3217                         ret = cma_netdev_change(ndev, id_priv);
3218                         if (ret)
3219                                 goto out;
3220                 }
3221
3222 out:
3223         mutex_unlock(&lock);
3224         return ret;
3225 }
3226
3227 static struct notifier_block cma_nb = {
3228         .notifier_call = cma_netdev_callback
3229 };
3230
3231 static void cma_add_one(struct ib_device *device)
3232 {
3233         struct cma_device *cma_dev;
3234         struct rdma_id_private *id_priv;
3235
3236         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3237         if (!cma_dev)
3238                 return;
3239
3240         cma_dev->device = device;
3241
3242         init_completion(&cma_dev->comp);
3243         atomic_set(&cma_dev->refcount, 1);
3244         INIT_LIST_HEAD(&cma_dev->id_list);
3245         ib_set_client_data(device, &cma_client, cma_dev);
3246
3247         mutex_lock(&lock);
3248         list_add_tail(&cma_dev->list, &dev_list);
3249         list_for_each_entry(id_priv, &listen_any_list, list)
3250                 cma_listen_on_dev(id_priv, cma_dev);
3251         mutex_unlock(&lock);
3252 }
3253
3254 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3255 {
3256         struct rdma_cm_event event;
3257         enum rdma_cm_state state;
3258         int ret = 0;
3259
3260         /* Record that we want to remove the device */
3261         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3262         if (state == RDMA_CM_DESTROYING)
3263                 return 0;
3264
3265         cma_cancel_operation(id_priv, state);
3266         mutex_lock(&id_priv->handler_mutex);
3267
3268         /* Check for destruction from another callback. */
3269         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3270                 goto out;
3271
3272         memset(&event, 0, sizeof event);
3273         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3274         ret = id_priv->id.event_handler(&id_priv->id, &event);
3275 out:
3276         mutex_unlock(&id_priv->handler_mutex);
3277         return ret;
3278 }
3279
3280 static void cma_process_remove(struct cma_device *cma_dev)
3281 {
3282         struct rdma_id_private *id_priv;
3283         int ret;
3284
3285         mutex_lock(&lock);
3286         while (!list_empty(&cma_dev->id_list)) {
3287                 id_priv = list_entry(cma_dev->id_list.next,
3288                                      struct rdma_id_private, list);
3289
3290                 list_del(&id_priv->listen_list);
3291                 list_del_init(&id_priv->list);
3292                 atomic_inc(&id_priv->refcount);
3293                 mutex_unlock(&lock);
3294
3295                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3296                 cma_deref_id(id_priv);
3297                 if (ret)
3298                         rdma_destroy_id(&id_priv->id);
3299
3300                 mutex_lock(&lock);
3301         }
3302         mutex_unlock(&lock);
3303
3304         cma_deref_dev(cma_dev);
3305         wait_for_completion(&cma_dev->comp);
3306 }
3307
3308 static void cma_remove_one(struct ib_device *device)
3309 {
3310         struct cma_device *cma_dev;
3311
3312         cma_dev = ib_get_client_data(device, &cma_client);
3313         if (!cma_dev)
3314                 return;
3315
3316         mutex_lock(&lock);
3317         list_del(&cma_dev->list);
3318         mutex_unlock(&lock);
3319
3320         cma_process_remove(cma_dev);
3321         kfree(cma_dev);
3322 }
3323
3324 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3325 {
3326         struct nlmsghdr *nlh;
3327         struct rdma_cm_id_stats *id_stats;
3328         struct rdma_id_private *id_priv;
3329         struct rdma_cm_id *id = NULL;
3330         struct cma_device *cma_dev;
3331         int i_dev = 0, i_id = 0;
3332
3333         /*
3334          * We export all of the IDs as a sequence of messages.  Each
3335          * ID gets its own netlink message.
3336          */
3337         mutex_lock(&lock);
3338
3339         list_for_each_entry(cma_dev, &dev_list, list) {
3340                 if (i_dev < cb->args[0]) {
3341                         i_dev++;
3342                         continue;
3343                 }
3344
3345                 i_id = 0;
3346                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3347                         if (i_id < cb->args[1]) {
3348                                 i_id++;
3349                                 continue;
3350                         }
3351
3352                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3353                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3354                                                 RDMA_NL_RDMA_CM_ID_STATS);
3355                         if (!id_stats)
3356                                 goto out;
3357
3358                         memset(id_stats, 0, sizeof *id_stats);
3359                         id = &id_priv->id;
3360                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3361                         id_stats->port_num = id->port_num;
3362                         id_stats->bound_dev_if =
3363                                 id->route.addr.dev_addr.bound_dev_if;
3364
3365                         if (id->route.addr.src_addr.ss_family == AF_INET) {
3366                                 if (ibnl_put_attr(skb, nlh,
3367                                                   sizeof(struct sockaddr_in),
3368                                                   &id->route.addr.src_addr,
3369                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3370                                         goto out;
3371                                 }
3372                                 if (ibnl_put_attr(skb, nlh,
3373                                                   sizeof(struct sockaddr_in),
3374                                                   &id->route.addr.dst_addr,
3375                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3376                                         goto out;
3377                                 }
3378                         } else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3379                                 if (ibnl_put_attr(skb, nlh,
3380                                                   sizeof(struct sockaddr_in6),
3381                                                   &id->route.addr.src_addr,
3382                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3383                                         goto out;
3384                                 }
3385                                 if (ibnl_put_attr(skb, nlh,
3386                                                   sizeof(struct sockaddr_in6),
3387                                                   &id->route.addr.dst_addr,
3388                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3389                                         goto out;
3390                                 }
3391                         }
3392
3393                         id_stats->pid           = id_priv->owner;
3394                         id_stats->port_space    = id->ps;
3395                         id_stats->cm_state      = id_priv->state;
3396                         id_stats->qp_num        = id_priv->qp_num;
3397                         id_stats->qp_type       = id->qp_type;
3398
3399                         i_id++;
3400                 }
3401
3402                 cb->args[1] = 0;
3403                 i_dev++;
3404         }
3405
3406 out:
3407         mutex_unlock(&lock);
3408         cb->args[0] = i_dev;
3409         cb->args[1] = i_id;
3410
3411         return skb->len;
3412 }
3413
3414 static const struct ibnl_client_cbs cma_cb_table[] = {
3415         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats },
3416 };
3417
3418 static int __init cma_init(void)
3419 {
3420         int ret;
3421
3422         cma_wq = create_singlethread_workqueue("rdma_cm");
3423         if (!cma_wq)
3424                 return -ENOMEM;
3425
3426         ib_sa_register_client(&sa_client);
3427         rdma_addr_register_client(&addr_client);
3428         register_netdevice_notifier(&cma_nb);
3429
3430         ret = ib_register_client(&cma_client);
3431         if (ret)
3432                 goto err;
3433
3434         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3435                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3436
3437         return 0;
3438
3439 err:
3440         unregister_netdevice_notifier(&cma_nb);
3441         rdma_addr_unregister_client(&addr_client);
3442         ib_sa_unregister_client(&sa_client);
3443         destroy_workqueue(cma_wq);
3444         return ret;
3445 }
3446
3447 static void __exit cma_cleanup(void)
3448 {
3449         ibnl_remove_client(RDMA_NL_RDMA_CM);
3450         ib_unregister_client(&cma_client);
3451         unregister_netdevice_notifier(&cma_nb);
3452         rdma_addr_unregister_client(&addr_client);
3453         ib_sa_unregister_client(&sa_client);
3454         destroy_workqueue(cma_wq);
3455         idr_destroy(&sdp_ps);
3456         idr_destroy(&tcp_ps);
3457         idr_destroy(&udp_ps);
3458         idr_destroy(&ipoib_ps);
3459 }
3460
3461 module_init(cma_init);
3462 module_exit(cma_cleanup);