drivers/infiniband/core/cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/idr.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/slab.h>
  44 #include <linux/module.h>
  45
  46 #include <net/tcp.h>
  47 #include <net/ipv6.h>
  48
  49 #include <rdma/rdma_cm.h>
  50 #include <rdma/rdma_cm_ib.h>
  51 #include <rdma/rdma_netlink.h>
  52 #include <rdma/ib_cache.h>
  53 #include <rdma/ib_cm.h>
  54 #include <rdma/ib_sa.h>
  55 #include <rdma/iw_cm.h>
  56
  57 MODULE_AUTHOR("Sean Hefty");
  58 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  59 MODULE_LICENSE("Dual BSD/GPL");
  60
  61 #define CMA_CM_RESPONSE_TIMEOUT 20
  62 #define CMA_MAX_CM_RETRIES 15
  63 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  64 #define CMA_IBOE_PACKET_LIFETIME 18
  65
  66 static void cma_add_one(struct ib_device *device);
  67 static void cma_remove_one(struct ib_device *device);
  68
  69 static struct ib_client cma_client = {
  70         .name   = "cma",
  71         .add    = cma_add_one,
  72         .remove = cma_remove_one
  73 };
  74
  75 static struct ib_sa_client sa_client;
  76 static struct rdma_addr_client addr_client;
  77 static LIST_HEAD(dev_list);
  78 static LIST_HEAD(listen_any_list);
  79 static DEFINE_MUTEX(lock);
  80 static struct workqueue_struct *cma_wq;
  81 static DEFINE_IDR(sdp_ps);
  82 static DEFINE_IDR(tcp_ps);
  83 static DEFINE_IDR(udp_ps);
  84 static DEFINE_IDR(ipoib_ps);
  85 static DEFINE_IDR(ib_ps);
  86
  87 struct cma_device {
  88         struct list_head        list;
  89         struct ib_device        *device;
  90         struct completion       comp;
  91         atomic_t                refcount;
  92         struct list_head        id_list;
  93 };
  94
  95 struct rdma_bind_list {
  96         struct idr              *ps;
  97         struct hlist_head       owners;
  98         unsigned short          port;
  99 };
 100
 101 /*
 102  * Device removal can occur at anytime, so we need extra handling to
 103  * serialize notifying the user of device removal with other callbacks.
 104  * We do this by disabling removal notification while a callback is in process,
 105  * and reporting it after the callback completes.
 106  */
 107 struct rdma_id_private {
 108         struct rdma_cm_id       id;
 109
 110         struct rdma_bind_list   *bind_list;
 111         struct hlist_node       node;
 112         struct list_head        list; /* listen_any_list or cma_device.list */
 113         struct list_head        listen_list; /* per device listens */
 114         struct cma_device       *cma_dev;
 115         struct list_head        mc_list;
 116
 117         int                     internal_id;
 118         enum rdma_cm_state      state;
 119         spinlock_t              lock;
 120         struct mutex            qp_mutex;
 121
 122         struct completion       comp;
 123         atomic_t                refcount;
 124         struct mutex            handler_mutex;
 125
 126         int                     backlog;
 127         int                     timeout_ms;
 128         struct ib_sa_query      *query;
 129         int                     query_id;
 130         union {
 131                 struct ib_cm_id *ib;
 132                 struct iw_cm_id *iw;
 133         } cm_id;
 134
 135         u32                     seq_num;
 136         u32                     qkey;
 137         u32                     qp_num;
 138         pid_t                   owner;
 139         u8                      srq;
 140         u8                      tos;
 141         u8                      reuseaddr;
 142 };
 143
 144 struct cma_multicast {
 145         struct rdma_id_private *id_priv;
 146         union {
 147                 struct ib_sa_multicast *ib;
 148         } multicast;
 149         struct list_head        list;
 150         void                    *context;
 151         struct sockaddr_storage addr;
 152         struct kref             mcref;
 153 };
 154
 155 struct cma_work {
 156         struct work_struct      work;
 157         struct rdma_id_private  *id;
 158         enum rdma_cm_state      old_state;
 159         enum rdma_cm_state      new_state;
 160         struct rdma_cm_event    event;
 161 };
 162
 163 struct cma_ndev_work {
 164         struct work_struct      work;
 165         struct rdma_id_private  *id;
 166         struct rdma_cm_event    event;
 167 };
 168
 169 struct iboe_mcast_work {
 170         struct work_struct       work;
 171         struct rdma_id_private  *id;
 172         struct cma_multicast    *mc;
 173 };
 174
 175 union cma_ip_addr {
 176         struct in6_addr ip6;
 177         struct {
 178                 __be32 pad[3];
 179                 __be32 addr;
 180         } ip4;
 181 };
 182
 183 struct cma_hdr {
 184         u8 cma_version;
 185         u8 ip_version;  /* IP version: 7:4 */
 186         __be16 port;
 187         union cma_ip_addr src_addr;
 188         union cma_ip_addr dst_addr;
 189 };
 190
 191 struct sdp_hh {
 192         u8 bsdh[16];
 193         u8 sdp_version; /* Major version: 7:4 */
 194         u8 ip_version;  /* IP version: 7:4 */
 195         u8 sdp_specific1[10];
 196         __be16 port;
 197         __be16 sdp_specific2;
 198         union cma_ip_addr src_addr;
 199         union cma_ip_addr dst_addr;
 200 };
 201
 202 struct sdp_hah {
 203         u8 bsdh[16];
 204         u8 sdp_version;
 205 };
 206
 207 #define CMA_VERSION 0x00
 208 #define SDP_MAJ_VERSION 0x2
 209
 210 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 211 {
 212         unsigned long flags;
 213         int ret;
 214
 215         spin_lock_irqsave(&id_priv->lock, flags);
 216         ret = (id_priv->state == comp);
 217         spin_unlock_irqrestore(&id_priv->lock, flags);
 218         return ret;
 219 }
 220
 221 static int cma_comp_exch(struct rdma_id_private *id_priv,
 222                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 223 {
 224         unsigned long flags;
 225         int ret;
 226
 227         spin_lock_irqsave(&id_priv->lock, flags);
 228         if ((ret = (id_priv->state == comp)))
 229                 id_priv->state = exch;
 230         spin_unlock_irqrestore(&id_priv->lock, flags);
 231         return ret;
 232 }
 233
 234 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 235                                    enum rdma_cm_state exch)
 236 {
 237         unsigned long flags;
 238         enum rdma_cm_state old;
 239
 240         spin_lock_irqsave(&id_priv->lock, flags);
 241         old = id_priv->state;
 242         id_priv->state = exch;
 243         spin_unlock_irqrestore(&id_priv->lock, flags);
 244         return old;
 245 }
 246
 247 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 248 {
 249         return hdr->ip_version >> 4;
 250 }
 251
 252 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 253 {
 254         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 255 }
 256
 257 static inline u8 sdp_get_majv(u8 sdp_version)
 258 {
 259         return sdp_version >> 4;
 260 }
 261
 262 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 263 {
 264         return hh->ip_version >> 4;
 265 }
 266
 267 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 268 {
 269         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 270 }
 271
 272 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 273                               struct cma_device *cma_dev)
 274 {
 275         atomic_inc(&cma_dev->refcount);
 276         id_priv->cma_dev = cma_dev;
 277         id_priv->id.device = cma_dev->device;
 278         id_priv->id.route.addr.dev_addr.transport =
 279                 rdma_node_get_transport(cma_dev->device->node_type);
 280         list_add_tail(&id_priv->list, &cma_dev->id_list);
 281 }
 282
 283 static inline void cma_deref_dev(struct cma_device *cma_dev)
 284 {
 285         if (atomic_dec_and_test(&cma_dev->refcount))
 286                 complete(&cma_dev->comp);
 287 }
 288
 289 static inline void release_mc(struct kref *kref)
 290 {
 291         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 292
 293         kfree(mc->multicast.ib);
 294         kfree(mc);
 295 }
 296
 297 static void cma_release_dev(struct rdma_id_private *id_priv)
 298 {
 299         mutex_lock(&lock);
 300         list_del(&id_priv->list);
 301         cma_deref_dev(id_priv->cma_dev);
 302         id_priv->cma_dev = NULL;
 303         mutex_unlock(&lock);
 304 }
 305
 306 static int cma_set_qkey(struct rdma_id_private *id_priv)
 307 {
 308         struct ib_sa_mcmember_rec rec;
 309         int ret = 0;
 310
 311         if (id_priv->qkey)
 312                 return 0;
 313
 314         switch (id_priv->id.ps) {
 315         case RDMA_PS_UDP:
 316                 id_priv->qkey = RDMA_UDP_QKEY;
 317                 break;
 318         case RDMA_PS_IPOIB:
 319                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 320                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 321                                              id_priv->id.port_num, &rec.mgid,
 322                                              &rec);
 323                 if (!ret)
 324                         id_priv->qkey = be32_to_cpu(rec.qkey);
 325                 break;
 326         default:
 327                 break;
 328         }
 329         return ret;
 330 }
 331
 332 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 333 {
 334         int i;
 335         int err;
 336         struct ib_port_attr props;
 337         union ib_gid tmp;
 338
 339         err = ib_query_port(device, port_num, &props);
 340         if (err)
 341                 return 1;
 342
 343         for (i = 0; i < props.gid_tbl_len; ++i) {
 344                 err = ib_query_gid(device, port_num, i, &tmp);
 345                 if (err)
 346                         return 1;
 347                 if (!memcmp(&tmp, gid, sizeof tmp))
 348                         return 0;
 349         }
 350
 351         return -EAGAIN;
 352 }
 353
 354 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 355 {
 356         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 357         struct cma_device *cma_dev;
 358         union ib_gid gid, iboe_gid;
 359         int ret = -ENODEV;
 360         u8 port;
 361         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 362                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 363
 364         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 365             id_priv->id.ps == RDMA_PS_IPOIB)
 366                 return -EINVAL;
 367
 368         mutex_lock(&lock);
 369         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 370         memcpy(&gid, dev_addr->src_dev_addr +
 371                rdma_addr_gid_offset(dev_addr), sizeof gid);
 372         list_for_each_entry(cma_dev, &dev_list, list) {
 373                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 374                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 375                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 376                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 377                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 378                                 else
 379                                         ret = find_gid_port(cma_dev->device, &gid, port);
 380
 381                                 if (!ret) {
 382                                         id_priv->id.port_num = port;
 383                                         goto out;
 384                                 } else if (ret == 1)
 385                                         break;
 386                         }
 387                 }
 388         }
 389
 390 out:
 391         if (!ret)
 392                 cma_attach_to_dev(id_priv, cma_dev);
 393
 394         mutex_unlock(&lock);
 395         return ret;
 396 }
 397
 398 static void cma_deref_id(struct rdma_id_private *id_priv)
 399 {
 400         if (atomic_dec_and_test(&id_priv->refcount))
 401                 complete(&id_priv->comp);
 402 }
 403
 404 static int cma_disable_callback(struct rdma_id_private *id_priv,
 405                                 enum rdma_cm_state state)
 406 {
 407         mutex_lock(&id_priv->handler_mutex);
 408         if (id_priv->state != state) {
 409                 mutex_unlock(&id_priv->handler_mutex);
 410                 return -EINVAL;
 411         }
 412         return 0;
 413 }
 414
 415 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 416                                   void *context, enum rdma_port_space ps,
 417                                   enum ib_qp_type qp_type)
 418 {
 419         struct rdma_id_private *id_priv;
 420
 421         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 422         if (!id_priv)
 423                 return ERR_PTR(-ENOMEM);
 424
 425         id_priv->owner = task_pid_nr(current);
 426         id_priv->state = RDMA_CM_IDLE;
 427         id_priv->id.context = context;
 428         id_priv->id.event_handler = event_handler;
 429         id_priv->id.ps = ps;
 430         id_priv->id.qp_type = qp_type;
 431         spin_lock_init(&id_priv->lock);
 432         mutex_init(&id_priv->qp_mutex);
 433         init_completion(&id_priv->comp);
 434         atomic_set(&id_priv->refcount, 1);
 435         mutex_init(&id_priv->handler_mutex);
 436         INIT_LIST_HEAD(&id_priv->listen_list);
 437         INIT_LIST_HEAD(&id_priv->mc_list);
 438         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 439
 440         return &id_priv->id;
 441 }
 442 EXPORT_SYMBOL(rdma_create_id);
 443
 444 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 445 {
 446         struct ib_qp_attr qp_attr;
 447         int qp_attr_mask, ret;
 448
 449         qp_attr.qp_state = IB_QPS_INIT;
 450         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 451         if (ret)
 452                 return ret;
 453
 454         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 455         if (ret)
 456                 return ret;
 457
 458         qp_attr.qp_state = IB_QPS_RTR;
 459         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 460         if (ret)
 461                 return ret;
 462
 463         qp_attr.qp_state = IB_QPS_RTS;
 464         qp_attr.sq_psn = 0;
 465         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 466
 467         return ret;
 468 }
 469
 470 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 471 {
 472         struct ib_qp_attr qp_attr;
 473         int qp_attr_mask, ret;
 474
 475         qp_attr.qp_state = IB_QPS_INIT;
 476         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 477         if (ret)
 478                 return ret;
 479
 480         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 481 }
 482
 483 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 484                    struct ib_qp_init_attr *qp_init_attr)
 485 {
 486         struct rdma_id_private *id_priv;
 487         struct ib_qp *qp;
 488         int ret;
 489
 490         id_priv = container_of(id, struct rdma_id_private, id);
 491         if (id->device != pd->device)
 492                 return -EINVAL;
 493
 494         qp = ib_create_qp(pd, qp_init_attr);
 495         if (IS_ERR(qp))
 496                 return PTR_ERR(qp);
 497
 498         if (id->qp_type == IB_QPT_UD)
 499                 ret = cma_init_ud_qp(id_priv, qp);
 500         else
 501                 ret = cma_init_conn_qp(id_priv, qp);
 502         if (ret)
 503                 goto err;
 504
 505         id->qp = qp;
 506         id_priv->qp_num = qp->qp_num;
 507         id_priv->srq = (qp->srq != NULL);
 508         return 0;
 509 err:
 510         ib_destroy_qp(qp);
 511         return ret;
 512 }
 513 EXPORT_SYMBOL(rdma_create_qp);
 514
 515 void rdma_destroy_qp(struct rdma_cm_id *id)
 516 {
 517         struct rdma_id_private *id_priv;
 518
 519         id_priv = container_of(id, struct rdma_id_private, id);
 520         mutex_lock(&id_priv->qp_mutex);
 521         ib_destroy_qp(id_priv->id.qp);
 522         id_priv->id.qp = NULL;
 523         mutex_unlock(&id_priv->qp_mutex);
 524 }
 525 EXPORT_SYMBOL(rdma_destroy_qp);
 526
 527 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 528                              struct rdma_conn_param *conn_param)
 529 {
 530         struct ib_qp_attr qp_attr;
 531         int qp_attr_mask, ret;
 532
 533         mutex_lock(&id_priv->qp_mutex);
 534         if (!id_priv->id.qp) {
 535                 ret = 0;
 536                 goto out;
 537         }
 538
 539         /* Need to update QP attributes from default values. */
 540         qp_attr.qp_state = IB_QPS_INIT;
 541         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 542         if (ret)
 543                 goto out;
 544
 545         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 546         if (ret)
 547                 goto out;
 548
 549         qp_attr.qp_state = IB_QPS_RTR;
 550         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 551         if (ret)
 552                 goto out;
 553
 554         if (conn_param)
 555                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 556         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 557 out:
 558         mutex_unlock(&id_priv->qp_mutex);
 559         return ret;
 560 }
 561
 562 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 563                              struct rdma_conn_param *conn_param)
 564 {
 565         struct ib_qp_attr qp_attr;
 566         int qp_attr_mask, ret;
 567
 568         mutex_lock(&id_priv->qp_mutex);
 569         if (!id_priv->id.qp) {
 570                 ret = 0;
 571                 goto out;
 572         }
 573
 574         qp_attr.qp_state = IB_QPS_RTS;
 575         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 576         if (ret)
 577                 goto out;
 578
 579         if (conn_param)
 580                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 581         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 582 out:
 583         mutex_unlock(&id_priv->qp_mutex);
 584         return ret;
 585 }
 586
 587 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 588 {
 589         struct ib_qp_attr qp_attr;
 590         int ret;
 591
 592         mutex_lock(&id_priv->qp_mutex);
 593         if (!id_priv->id.qp) {
 594                 ret = 0;
 595                 goto out;
 596         }
 597
 598         qp_attr.qp_state = IB_QPS_ERR;
 599         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 600 out:
 601         mutex_unlock(&id_priv->qp_mutex);
 602         return ret;
 603 }
 604
 605 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 606                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 607 {
 608         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 609         int ret;
 610         u16 pkey;
 611
 612         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 613             IB_LINK_LAYER_INFINIBAND)
 614                 pkey = ib_addr_get_pkey(dev_addr);
 615         else
 616                 pkey = 0xffff;
 617
 618         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 619                                   pkey, &qp_attr->pkey_index);
 620         if (ret)
 621                 return ret;
 622
 623         qp_attr->port_num = id_priv->id.port_num;
 624         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 625
 626         if (id_priv->id.qp_type == IB_QPT_UD) {
 627                 ret = cma_set_qkey(id_priv);
 628                 if (ret)
 629                         return ret;
 630
 631                 qp_attr->qkey = id_priv->qkey;
 632                 *qp_attr_mask |= IB_QP_QKEY;
 633         } else {
 634                 qp_attr->qp_access_flags = 0;
 635                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 636         }
 637         return 0;
 638 }
 639
 640 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 641                        int *qp_attr_mask)
 642 {
 643         struct rdma_id_private *id_priv;
 644         int ret = 0;
 645
 646         id_priv = container_of(id, struct rdma_id_private, id);
 647         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 648         case RDMA_TRANSPORT_IB:
 649                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 650                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 651                 else
 652                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 653                                                  qp_attr_mask);
 654                 if (qp_attr->qp_state == IB_QPS_RTR)
 655                         qp_attr->rq_psn = id_priv->seq_num;
 656                 break;
 657         case RDMA_TRANSPORT_IWARP:
 658                 if (!id_priv->cm_id.iw) {
 659                         qp_attr->qp_access_flags = 0;
 660                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 661                 } else
 662                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 663                                                  qp_attr_mask);
 664                 qp_attr->port_num = id_priv->id.port_num;
 665                 *qp_attr_mask |= IB_QP_PORT;
 666                 break;
 667         default:
 668                 ret = -ENOSYS;
 669                 break;
 670         }
 671
 672         return ret;
 673 }
 674 EXPORT_SYMBOL(rdma_init_qp_attr);
 675
 676 static inline int cma_zero_addr(struct sockaddr *addr)
 677 {
 678         struct in6_addr *ip6;
 679
 680         if (addr->sa_family == AF_INET)
 681                 return ipv4_is_zeronet(
 682                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 683         else {
 684                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 685                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 686                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 687         }
 688 }
 689
 690 static inline int cma_loopback_addr(struct sockaddr *addr)
 691 {
 692         if (addr->sa_family == AF_INET)
 693                 return ipv4_is_loopback(
 694                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 695         else
 696                 return ipv6_addr_loopback(
 697                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 698 }
 699
 700 static inline int cma_any_addr(struct sockaddr *addr)
 701 {
 702         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 703 }
 704
 705 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 706 {
 707         if (src->sa_family != dst->sa_family)
 708                 return -1;
 709
 710         switch (src->sa_family) {
 711         case AF_INET:
 712                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 713                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 714         default:
 715                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 716                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 717         }
 718 }
 719
 720 static inline __be16 cma_port(struct sockaddr *addr)
 721 {
 722         if (addr->sa_family == AF_INET)
 723                 return ((struct sockaddr_in *) addr)->sin_port;
 724         else
 725                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 726 }
 727
 728 static inline int cma_any_port(struct sockaddr *addr)
 729 {
 730         return !cma_port(addr);
 731 }
 732
 733 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 734                             u8 *ip_ver, __be16 *port,
 735                             union cma_ip_addr **src, union cma_ip_addr **dst)
 736 {
 737         switch (ps) {
 738         case RDMA_PS_SDP:
 739                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 740                     SDP_MAJ_VERSION)
 741                         return -EINVAL;
 742
 743                 *ip_ver = sdp_get_ip_ver(hdr);
 744                 *port   = ((struct sdp_hh *) hdr)->port;
 745                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 746                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 747                 break;
 748         default:
 749                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 750                         return -EINVAL;
 751
 752                 *ip_ver = cma_get_ip_ver(hdr);
 753                 *port   = ((struct cma_hdr *) hdr)->port;
 754                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 755                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 756                 break;
 757         }
 758
 759         if (*ip_ver != 4 && *ip_ver != 6)
 760                 return -EINVAL;
 761         return 0;
 762 }
 763
 764 static void cma_save_net_info(struct rdma_addr *addr,
 765                               struct rdma_addr *listen_addr,
 766                               u8 ip_ver, __be16 port,
 767                               union cma_ip_addr *src, union cma_ip_addr *dst)
 768 {
 769         struct sockaddr_in *listen4, *ip4;
 770         struct sockaddr_in6 *listen6, *ip6;
 771
 772         switch (ip_ver) {
 773         case 4:
 774                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 775                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 776                 ip4->sin_family = listen4->sin_family;
 777                 ip4->sin_addr.s_addr = dst->ip4.addr;
 778                 ip4->sin_port = listen4->sin_port;
 779
 780                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 781                 ip4->sin_family = listen4->sin_family;
 782                 ip4->sin_addr.s_addr = src->ip4.addr;
 783                 ip4->sin_port = port;
 784                 break;
 785         case 6:
 786                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 787                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 788                 ip6->sin6_family = listen6->sin6_family;
 789                 ip6->sin6_addr = dst->ip6;
 790                 ip6->sin6_port = listen6->sin6_port;
 791
 792                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 793                 ip6->sin6_family = listen6->sin6_family;
 794                 ip6->sin6_addr = src->ip6;
 795                 ip6->sin6_port = port;
 796                 break;
 797         default:
 798                 break;
 799         }
 800 }
 801
 802 static inline int cma_user_data_offset(enum rdma_port_space ps)
 803 {
 804         switch (ps) {
 805         case RDMA_PS_SDP:
 806                 return 0;
 807         default:
 808                 return sizeof(struct cma_hdr);
 809         }
 810 }
 811
 812 static void cma_cancel_route(struct rdma_id_private *id_priv)
 813 {
 814         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 815         case IB_LINK_LAYER_INFINIBAND:
 816                 if (id_priv->query)
 817                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 818                 break;
 819         default:
 820                 break;
 821         }
 822 }
 823
 824 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 825 {
 826         struct rdma_id_private *dev_id_priv;
 827
 828         /*
 829          * Remove from listen_any_list to prevent added devices from spawning
 830          * additional listen requests.
 831          */
 832         mutex_lock(&lock);
 833         list_del(&id_priv->list);
 834
 835         while (!list_empty(&id_priv->listen_list)) {
 836                 dev_id_priv = list_entry(id_priv->listen_list.next,
 837                                          struct rdma_id_private, listen_list);
 838                 /* sync with device removal to avoid duplicate destruction */
 839                 list_del_init(&dev_id_priv->list);
 840                 list_del(&dev_id_priv->listen_list);
 841                 mutex_unlock(&lock);
 842
 843                 rdma_destroy_id(&dev_id_priv->id);
 844                 mutex_lock(&lock);
 845         }
 846         mutex_unlock(&lock);
 847 }
 848
 849 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 850                                  enum rdma_cm_state state)
 851 {
 852         switch (state) {
 853         case RDMA_CM_ADDR_QUERY:
 854                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 855                 break;
 856         case RDMA_CM_ROUTE_QUERY:
 857                 cma_cancel_route(id_priv);
 858                 break;
 859         case RDMA_CM_LISTEN:
 860                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 861                                 && !id_priv->cma_dev)
 862                         cma_cancel_listens(id_priv);
 863                 break;
 864         default:
 865                 break;
 866         }
 867 }
 868
 869 static void cma_release_port(struct rdma_id_private *id_priv)
 870 {
 871         struct rdma_bind_list *bind_list = id_priv->bind_list;
 872
 873         if (!bind_list)
 874                 return;
 875
 876         mutex_lock(&lock);
 877         hlist_del(&id_priv->node);
 878         if (hlist_empty(&bind_list->owners)) {
 879                 idr_remove(bind_list->ps, bind_list->port);
 880                 kfree(bind_list);
 881         }
 882         mutex_unlock(&lock);
 883 }
 884
 885 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 886 {
 887         struct cma_multicast *mc;
 888
 889         while (!list_empty(&id_priv->mc_list)) {
 890                 mc = container_of(id_priv->mc_list.next,
 891                                   struct cma_multicast, list);
 892                 list_del(&mc->list);
 893                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 894                 case IB_LINK_LAYER_INFINIBAND:
 895                         ib_sa_free_multicast(mc->multicast.ib);
 896                         kfree(mc);
 897                         break;
 898                 case IB_LINK_LAYER_ETHERNET:
 899                         kref_put(&mc->mcref, release_mc);
 900                         break;
 901                 default:
 902                         break;
 903                 }
 904         }
 905 }
 906
 907 void rdma_destroy_id(struct rdma_cm_id *id)
 908 {
 909         struct rdma_id_private *id_priv;
 910         enum rdma_cm_state state;
 911
 912         id_priv = container_of(id, struct rdma_id_private, id);
 913         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 914         cma_cancel_operation(id_priv, state);
 915
 916         /*
 917          * Wait for any active callback to finish.  New callbacks will find
 918          * the id_priv state set to destroying and abort.
 919          */
 920         mutex_lock(&id_priv->handler_mutex);
 921         mutex_unlock(&id_priv->handler_mutex);
 922
 923         if (id_priv->cma_dev) {
 924                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 925                 case RDMA_TRANSPORT_IB:
 926                         if (id_priv->cm_id.ib)
 927                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 928                         break;
 929                 case RDMA_TRANSPORT_IWARP:
 930                         if (id_priv->cm_id.iw)
 931                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 932                         break;
 933                 default:
 934                         break;
 935                 }
 936                 cma_leave_mc_groups(id_priv);
 937                 cma_release_dev(id_priv);
 938         }
 939
 940         cma_release_port(id_priv);
 941         cma_deref_id(id_priv);
 942         wait_for_completion(&id_priv->comp);
 943
 944         if (id_priv->internal_id)
 945                 cma_deref_id(id_priv->id.context);
 946
 947         kfree(id_priv->id.route.path_rec);
 948         kfree(id_priv);
 949 }
 950 EXPORT_SYMBOL(rdma_destroy_id);
 951
 952 static int cma_rep_recv(struct rdma_id_private *id_priv)
 953 {
 954         int ret;
 955
 956         ret = cma_modify_qp_rtr(id_priv, NULL);
 957         if (ret)
 958                 goto reject;
 959
 960         ret = cma_modify_qp_rts(id_priv, NULL);
 961         if (ret)
 962                 goto reject;
 963
 964         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 965         if (ret)
 966                 goto reject;
 967
 968         return 0;
 969 reject:
 970         cma_modify_qp_err(id_priv);
 971         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 972                        NULL, 0, NULL, 0);
 973         return ret;
 974 }
 975
 976 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 977 {
 978         if (id_priv->id.ps == RDMA_PS_SDP &&
 979             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 980             SDP_MAJ_VERSION)
 981                 return -EINVAL;
 982
 983         return 0;
 984 }
 985
 986 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 987                                    struct ib_cm_rep_event_param *rep_data,
 988                                    void *private_data)
 989 {
 990         event->param.conn.private_data = private_data;
 991         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 992         event->param.conn.responder_resources = rep_data->responder_resources;
 993         event->param.conn.initiator_depth = rep_data->initiator_depth;
 994         event->param.conn.flow_control = rep_data->flow_control;
 995         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 996         event->param.conn.srq = rep_data->srq;
 997         event->param.conn.qp_num = rep_data->remote_qpn;
 998 }
 999
1000 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1001 {
1002         struct rdma_id_private *id_priv = cm_id->context;
1003         struct rdma_cm_event event;
1004         int ret = 0;
1005
1006         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1007                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1008             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1009                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1010                 return 0;
1011
1012         memset(&event, 0, sizeof event);
1013         switch (ib_event->event) {
1014         case IB_CM_REQ_ERROR:
1015         case IB_CM_REP_ERROR:
1016                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1017                 event.status = -ETIMEDOUT;
1018                 break;
1019         case IB_CM_REP_RECEIVED:
1020                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1021                 if (event.status)
1022                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1023                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1024                         event.status = cma_rep_recv(id_priv);
1025                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1026                                                      RDMA_CM_EVENT_ESTABLISHED;
1027                 } else
1028                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1029                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1030                                        ib_event->private_data);
1031                 break;
1032         case IB_CM_RTU_RECEIVED:
1033         case IB_CM_USER_ESTABLISHED:
1034                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1035                 break;
1036         case IB_CM_DREQ_ERROR:
1037                 event.status = -ETIMEDOUT; /* fall through */
1038         case IB_CM_DREQ_RECEIVED:
1039         case IB_CM_DREP_RECEIVED:
1040                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1041                                    RDMA_CM_DISCONNECT))
1042                         goto out;
1043                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1044                 break;
1045         case IB_CM_TIMEWAIT_EXIT:
1046                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1047                 break;
1048         case IB_CM_MRA_RECEIVED:
1049                 /* ignore event */
1050                 goto out;
1051         case IB_CM_REJ_RECEIVED:
1052                 cma_modify_qp_err(id_priv);
1053                 event.status = ib_event->param.rej_rcvd.reason;
1054                 event.event = RDMA_CM_EVENT_REJECTED;
1055                 event.param.conn.private_data = ib_event->private_data;
1056                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1057                 break;
1058         default:
1059                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1060                        ib_event->event);
1061                 goto out;
1062         }
1063
1064         ret = id_priv->id.event_handler(&id_priv->id, &event);
1065         if (ret) {
1066                 /* Destroy the CM ID by returning a non-zero value. */
1067                 id_priv->cm_id.ib = NULL;
1068                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1069                 mutex_unlock(&id_priv->handler_mutex);
1070                 rdma_destroy_id(&id_priv->id);
1071                 return ret;
1072         }
1073 out:
1074         mutex_unlock(&id_priv->handler_mutex);
1075         return ret;
1076 }
1077
1078 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1079                                                struct ib_cm_event *ib_event)
1080 {
1081         struct rdma_id_private *id_priv;
1082         struct rdma_cm_id *id;
1083         struct rdma_route *rt;
1084         union cma_ip_addr *src, *dst;
1085         __be16 port;
1086         u8 ip_ver;
1087         int ret;
1088
1089         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1090                              &ip_ver, &port, &src, &dst))
1091                 return NULL;
1092
1093         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1094                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1095         if (IS_ERR(id))
1096                 return NULL;
1097
1098         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1099                           ip_ver, port, src, dst);
1100
1101         rt = &id->route;
1102         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1103         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1104                                GFP_KERNEL);
1105         if (!rt->path_rec)
1106                 goto err;
1107
1108         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1109         if (rt->num_paths == 2)
1110                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1111
1112         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1113                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1114                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1115                 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1116         } else {
1117                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1118                                         &rt->addr.dev_addr);
1119                 if (ret)
1120                         goto err;
1121         }
1122         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1123
1124         id_priv = container_of(id, struct rdma_id_private, id);
1125         id_priv->state = RDMA_CM_CONNECT;
1126         return id_priv;
1127
1128 err:
1129         rdma_destroy_id(id);
1130         return NULL;
1131 }
1132
1133 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1134                                               struct ib_cm_event *ib_event)
1135 {
1136         struct rdma_id_private *id_priv;
1137         struct rdma_cm_id *id;
1138         union cma_ip_addr *src, *dst;
1139         __be16 port;
1140         u8 ip_ver;
1141         int ret;
1142
1143         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1144                             listen_id->ps, IB_QPT_UD);
1145         if (IS_ERR(id))
1146                 return NULL;
1147
1148
1149         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1150                              &ip_ver, &port, &src, &dst))
1151                 goto err;
1152
1153         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1154                           ip_ver, port, src, dst);
1155
1156         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1157                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1158                                         &id->route.addr.dev_addr);
1159                 if (ret)
1160                         goto err;
1161         }
1162
1163         id_priv = container_of(id, struct rdma_id_private, id);
1164         id_priv->state = RDMA_CM_CONNECT;
1165         return id_priv;
1166 err:
1167         rdma_destroy_id(id);
1168         return NULL;
1169 }
1170
1171 static void cma_set_req_event_data(struct rdma_cm_event *event,
1172                                    struct ib_cm_req_event_param *req_data,
1173                                    void *private_data, int offset)
1174 {
1175         event->param.conn.private_data = private_data + offset;
1176         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1177         event->param.conn.responder_resources = req_data->responder_resources;
1178         event->param.conn.initiator_depth = req_data->initiator_depth;
1179         event->param.conn.flow_control = req_data->flow_control;
1180         event->param.conn.retry_count = req_data->retry_count;
1181         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1182         event->param.conn.srq = req_data->srq;
1183         event->param.conn.qp_num = req_data->remote_qpn;
1184 }
1185
1186 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1187 {
1188         return (((ib_event->event == IB_CM_REQ_RECEIVED) ||
1189                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1190                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1191                  (id->qp_type == IB_QPT_UD)) ||
1192                 (!id->qp_type));
1193 }
1194
1195 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1196 {
1197         struct rdma_id_private *listen_id, *conn_id;
1198         struct rdma_cm_event event;
1199         int offset, ret;
1200
1201         listen_id = cm_id->context;
1202         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1203                 return -EINVAL;
1204
1205         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1206                 return -ECONNABORTED;
1207
1208         memset(&event, 0, sizeof event);
1209         offset = cma_user_data_offset(listen_id->id.ps);
1210         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1211         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1212                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1213                 event.param.ud.private_data = ib_event->private_data + offset;
1214                 event.param.ud.private_data_len =
1215                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1216         } else {
1217                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1218                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1219                                        ib_event->private_data, offset);
1220         }
1221         if (!conn_id) {
1222                 ret = -ENOMEM;
1223                 goto out;
1224         }
1225
1226         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1227         ret = cma_acquire_dev(conn_id);
1228         if (ret)
1229                 goto release_conn_id;
1230
1231         conn_id->cm_id.ib = cm_id;
1232         cm_id->context = conn_id;
1233         cm_id->cm_handler = cma_ib_handler;
1234
1235         /*
1236          * Protect against the user destroying conn_id from another thread
1237          * until we're done accessing it.
1238          */
1239         atomic_inc(&conn_id->refcount);
1240         ret = conn_id->id.event_handler(&conn_id->id, &event);
1241         if (!ret) {
1242                 /*
1243                  * Acquire mutex to prevent user executing rdma_destroy_id()
1244                  * while we're accessing the cm_id.
1245                  */
1246                 mutex_lock(&lock);
1247                 if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1248                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1249                 mutex_unlock(&lock);
1250                 mutex_unlock(&conn_id->handler_mutex);
1251                 cma_deref_id(conn_id);
1252                 goto out;
1253         }
1254         cma_deref_id(conn_id);
1255
1256         /* Destroy the CM ID by returning a non-zero value. */
1257         conn_id->cm_id.ib = NULL;
1258
1259 release_conn_id:
1260         cma_exch(conn_id, RDMA_CM_DESTROYING);
1261         mutex_unlock(&conn_id->handler_mutex);
1262         rdma_destroy_id(&conn_id->id);
1263
1264 out:
1265         mutex_unlock(&listen_id->handler_mutex);
1266         return ret;
1267 }
1268
1269 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1270 {
1271         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1272 }
1273
1274 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1275                                  struct ib_cm_compare_data *compare)
1276 {
1277         struct cma_hdr *cma_data, *cma_mask;
1278         struct sdp_hh *sdp_data, *sdp_mask;
1279         __be32 ip4_addr;
1280         struct in6_addr ip6_addr;
1281
1282         memset(compare, 0, sizeof *compare);
1283         cma_data = (void *) compare->data;
1284         cma_mask = (void *) compare->mask;
1285         sdp_data = (void *) compare->data;
1286         sdp_mask = (void *) compare->mask;
1287
1288         switch (addr->sa_family) {
1289         case AF_INET:
1290                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1291                 if (ps == RDMA_PS_SDP) {
1292                         sdp_set_ip_ver(sdp_data, 4);
1293                         sdp_set_ip_ver(sdp_mask, 0xF);
1294                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1295                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1296                 } else {
1297                         cma_set_ip_ver(cma_data, 4);
1298                         cma_set_ip_ver(cma_mask, 0xF);
1299                         cma_data->dst_addr.ip4.addr = ip4_addr;
1300                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1301                 }
1302                 break;
1303         case AF_INET6:
1304                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1305                 if (ps == RDMA_PS_SDP) {
1306                         sdp_set_ip_ver(sdp_data, 6);
1307                         sdp_set_ip_ver(sdp_mask, 0xF);
1308                         sdp_data->dst_addr.ip6 = ip6_addr;
1309                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1310                                sizeof sdp_mask->dst_addr.ip6);
1311                 } else {
1312                         cma_set_ip_ver(cma_data, 6);
1313                         cma_set_ip_ver(cma_mask, 0xF);
1314                         cma_data->dst_addr.ip6 = ip6_addr;
1315                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1316                                sizeof cma_mask->dst_addr.ip6);
1317                 }
1318                 break;
1319         default:
1320                 break;
1321         }
1322 }
1323
1324 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1325 {
1326         struct rdma_id_private *id_priv = iw_id->context;
1327         struct rdma_cm_event event;
1328         struct sockaddr_in *sin;
1329         int ret = 0;
1330
1331         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1332                 return 0;
1333
1334         memset(&event, 0, sizeof event);
1335         switch (iw_event->event) {
1336         case IW_CM_EVENT_CLOSE:
1337                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1338                 break;
1339         case IW_CM_EVENT_CONNECT_REPLY:
1340                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1341                 *sin = iw_event->local_addr;
1342                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1343                 *sin = iw_event->remote_addr;
1344                 switch (iw_event->status) {
1345                 case 0:
1346                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1347                         event.param.conn.initiator_depth = iw_event->ird;
1348                         event.param.conn.responder_resources = iw_event->ord;
1349                         break;
1350                 case -ECONNRESET:
1351                 case -ECONNREFUSED:
1352                         event.event = RDMA_CM_EVENT_REJECTED;
1353                         break;
1354                 case -ETIMEDOUT:
1355                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1356                         break;
1357                 default:
1358                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1359                         break;
1360                 }
1361                 break;
1362         case IW_CM_EVENT_ESTABLISHED:
1363                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1364                 event.param.conn.initiator_depth = iw_event->ird;
1365                 event.param.conn.responder_resources = iw_event->ord;
1366                 break;
1367         default:
1368                 BUG_ON(1);
1369         }
1370
1371         event.status = iw_event->status;
1372         event.param.conn.private_data = iw_event->private_data;
1373         event.param.conn.private_data_len = iw_event->private_data_len;
1374         ret = id_priv->id.event_handler(&id_priv->id, &event);
1375         if (ret) {
1376                 /* Destroy the CM ID by returning a non-zero value. */
1377                 id_priv->cm_id.iw = NULL;
1378                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1379                 mutex_unlock(&id_priv->handler_mutex);
1380                 rdma_destroy_id(&id_priv->id);
1381                 return ret;
1382         }
1383
1384         mutex_unlock(&id_priv->handler_mutex);
1385         return ret;
1386 }
1387
1388 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1389                                struct iw_cm_event *iw_event)
1390 {
1391         struct rdma_cm_id *new_cm_id;
1392         struct rdma_id_private *listen_id, *conn_id;
1393         struct sockaddr_in *sin;
1394         struct net_device *dev = NULL;
1395         struct rdma_cm_event event;
1396         int ret;
1397         struct ib_device_attr attr;
1398
1399         listen_id = cm_id->context;
1400         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1401                 return -ECONNABORTED;
1402
1403         /* Create a new RDMA id for the new IW CM ID */
1404         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1405                                    listen_id->id.context,
1406                                    RDMA_PS_TCP, IB_QPT_RC);
1407         if (IS_ERR(new_cm_id)) {
1408                 ret = -ENOMEM;
1409                 goto out;
1410         }
1411         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1412         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1413         conn_id->state = RDMA_CM_CONNECT;
1414
1415         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1416         if (!dev) {
1417                 ret = -EADDRNOTAVAIL;
1418                 mutex_unlock(&conn_id->handler_mutex);
1419                 rdma_destroy_id(new_cm_id);
1420                 goto out;
1421         }
1422         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1423         if (ret) {
1424                 mutex_unlock(&conn_id->handler_mutex);
1425                 rdma_destroy_id(new_cm_id);
1426                 goto out;
1427         }
1428
1429         ret = cma_acquire_dev(conn_id);
1430         if (ret) {
1431                 mutex_unlock(&conn_id->handler_mutex);
1432                 rdma_destroy_id(new_cm_id);
1433                 goto out;
1434         }
1435
1436         conn_id->cm_id.iw = cm_id;
1437         cm_id->context = conn_id;
1438         cm_id->cm_handler = cma_iw_handler;
1439
1440         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1441         *sin = iw_event->local_addr;
1442         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1443         *sin = iw_event->remote_addr;
1444
1445         ret = ib_query_device(conn_id->id.device, &attr);
1446         if (ret) {
1447                 mutex_unlock(&conn_id->handler_mutex);
1448                 rdma_destroy_id(new_cm_id);
1449                 goto out;
1450         }
1451
1452         memset(&event, 0, sizeof event);
1453         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1454         event.param.conn.private_data = iw_event->private_data;
1455         event.param.conn.private_data_len = iw_event->private_data_len;
1456         event.param.conn.initiator_depth = iw_event->ird;
1457         event.param.conn.responder_resources = iw_event->ord;
1458
1459         /*
1460          * Protect against the user destroying conn_id from another thread
1461          * until we're done accessing it.
1462          */
1463         atomic_inc(&conn_id->refcount);
1464         ret = conn_id->id.event_handler(&conn_id->id, &event);
1465         if (ret) {
1466                 /* User wants to destroy the CM ID */
1467                 conn_id->cm_id.iw = NULL;
1468                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1469                 mutex_unlock(&conn_id->handler_mutex);
1470                 cma_deref_id(conn_id);
1471                 rdma_destroy_id(&conn_id->id);
1472                 goto out;
1473         }
1474
1475         mutex_unlock(&conn_id->handler_mutex);
1476         cma_deref_id(conn_id);
1477
1478 out:
1479         if (dev)
1480                 dev_put(dev);
1481         mutex_unlock(&listen_id->handler_mutex);
1482         return ret;
1483 }
1484
1485 static int cma_ib_listen(struct rdma_id_private *id_priv)
1486 {
1487         struct ib_cm_compare_data compare_data;
1488         struct sockaddr *addr;
1489         struct ib_cm_id *id;
1490         __be64 svc_id;
1491         int ret;
1492
1493         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1494         if (IS_ERR(id))
1495                 return PTR_ERR(id);
1496
1497         id_priv->cm_id.ib = id;
1498
1499         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1500         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1501         if (cma_any_addr(addr))
1502                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1503         else {
1504                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1505                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1506         }
1507
1508         if (ret) {
1509                 ib_destroy_cm_id(id_priv->cm_id.ib);
1510                 id_priv->cm_id.ib = NULL;
1511         }
1512
1513         return ret;
1514 }
1515
1516 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1517 {
1518         int ret;
1519         struct sockaddr_in *sin;
1520         struct iw_cm_id *id;
1521
1522         id = iw_create_cm_id(id_priv->id.device,
1523                              iw_conn_req_handler,
1524                              id_priv);
1525         if (IS_ERR(id))
1526                 return PTR_ERR(id);
1527
1528         id_priv->cm_id.iw = id;
1529
1530         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1531         id_priv->cm_id.iw->local_addr = *sin;
1532
1533         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1534
1535         if (ret) {
1536                 iw_destroy_cm_id(id_priv->cm_id.iw);
1537                 id_priv->cm_id.iw = NULL;
1538         }
1539
1540         return ret;
1541 }
1542
1543 static int cma_listen_handler(struct rdma_cm_id *id,
1544                               struct rdma_cm_event *event)
1545 {
1546         struct rdma_id_private *id_priv = id->context;
1547
1548         id->context = id_priv->id.context;
1549         id->event_handler = id_priv->id.event_handler;
1550         return id_priv->id.event_handler(id, event);
1551 }
1552
1553 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1554                               struct cma_device *cma_dev)
1555 {
1556         struct rdma_id_private *dev_id_priv;
1557         struct rdma_cm_id *id;
1558         int ret;
1559
1560         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1561                             id_priv->id.qp_type);
1562         if (IS_ERR(id))
1563                 return;
1564
1565         dev_id_priv = container_of(id, struct rdma_id_private, id);
1566
1567         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1568         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1569                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1570
1571         cma_attach_to_dev(dev_id_priv, cma_dev);
1572         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1573         atomic_inc(&id_priv->refcount);
1574         dev_id_priv->internal_id = 1;
1575
1576         ret = rdma_listen(id, id_priv->backlog);
1577         if (ret)
1578                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1579                        "listening on device %s\n", ret, cma_dev->device->name);
1580 }
1581
1582 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1583 {
1584         struct cma_device *cma_dev;
1585
1586         mutex_lock(&lock);
1587         list_add_tail(&id_priv->list, &listen_any_list);
1588         list_for_each_entry(cma_dev, &dev_list, list)
1589                 cma_listen_on_dev(id_priv, cma_dev);
1590         mutex_unlock(&lock);
1591 }
1592
1593 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1594 {
1595         struct rdma_id_private *id_priv;
1596
1597         id_priv = container_of(id, struct rdma_id_private, id);
1598         id_priv->tos = (u8) tos;
1599 }
1600 EXPORT_SYMBOL(rdma_set_service_type);
1601
1602 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1603                               void *context)
1604 {
1605         struct cma_work *work = context;
1606         struct rdma_route *route;
1607
1608         route = &work->id->id.route;
1609
1610         if (!status) {
1611                 route->num_paths = 1;
1612                 *route->path_rec = *path_rec;
1613         } else {
1614                 work->old_state = RDMA_CM_ROUTE_QUERY;
1615                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1616                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1617                 work->event.status = status;
1618         }
1619
1620         queue_work(cma_wq, &work->work);
1621 }
1622
1623 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1624                               struct cma_work *work)
1625 {
1626         struct rdma_addr *addr = &id_priv->id.route.addr;
1627         struct ib_sa_path_rec path_rec;
1628         ib_sa_comp_mask comp_mask;
1629         struct sockaddr_in6 *sin6;
1630
1631         memset(&path_rec, 0, sizeof path_rec);
1632         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1633         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1634         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1635         path_rec.numb_path = 1;
1636         path_rec.reversible = 1;
1637         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1638                                                         (struct sockaddr *) &addr->dst_addr);
1639
1640         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1641                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1642                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1643
1644         if (addr->src_addr.ss_family == AF_INET) {
1645                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1646                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1647         } else {
1648                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1649                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1650                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1651         }
1652
1653         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1654                                                id_priv->id.port_num, &path_rec,
1655                                                comp_mask, timeout_ms,
1656                                                GFP_KERNEL, cma_query_handler,
1657                                                work, &id_priv->query);
1658
1659         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1660 }
1661
1662 static void cma_work_handler(struct work_struct *_work)
1663 {
1664         struct cma_work *work = container_of(_work, struct cma_work, work);
1665         struct rdma_id_private *id_priv = work->id;
1666         int destroy = 0;
1667
1668         mutex_lock(&id_priv->handler_mutex);
1669         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1670                 goto out;
1671
1672         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1673                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1674                 destroy = 1;
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 void cma_ndev_work_handler(struct work_struct *_work)
1685 {
1686         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1687         struct rdma_id_private *id_priv = work->id;
1688         int destroy = 0;
1689
1690         mutex_lock(&id_priv->handler_mutex);
1691         if (id_priv->state == RDMA_CM_DESTROYING ||
1692             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1693                 goto out;
1694
1695         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1696                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1697                 destroy = 1;
1698         }
1699
1700 out:
1701         mutex_unlock(&id_priv->handler_mutex);
1702         cma_deref_id(id_priv);
1703         if (destroy)
1704                 rdma_destroy_id(&id_priv->id);
1705         kfree(work);
1706 }
1707
1708 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1709 {
1710         struct rdma_route *route = &id_priv->id.route;
1711         struct cma_work *work;
1712         int ret;
1713
1714         work = kzalloc(sizeof *work, GFP_KERNEL);
1715         if (!work)
1716                 return -ENOMEM;
1717
1718         work->id = id_priv;
1719         INIT_WORK(&work->work, cma_work_handler);
1720         work->old_state = RDMA_CM_ROUTE_QUERY;
1721         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1722         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1723
1724         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1725         if (!route->path_rec) {
1726                 ret = -ENOMEM;
1727                 goto err1;
1728         }
1729
1730         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1731         if (ret)
1732                 goto err2;
1733
1734         return 0;
1735 err2:
1736         kfree(route->path_rec);
1737         route->path_rec = NULL;
1738 err1:
1739         kfree(work);
1740         return ret;
1741 }
1742
1743 int rdma_set_ib_paths(struct rdma_cm_id *id,
1744                       struct ib_sa_path_rec *path_rec, int num_paths)
1745 {
1746         struct rdma_id_private *id_priv;
1747         int ret;
1748
1749         id_priv = container_of(id, struct rdma_id_private, id);
1750         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1751                            RDMA_CM_ROUTE_RESOLVED))
1752                 return -EINVAL;
1753
1754         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1755                                      GFP_KERNEL);
1756         if (!id->route.path_rec) {
1757                 ret = -ENOMEM;
1758                 goto err;
1759         }
1760
1761         id->route.num_paths = num_paths;
1762         return 0;
1763 err:
1764         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1765         return ret;
1766 }
1767 EXPORT_SYMBOL(rdma_set_ib_paths);
1768
1769 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1770 {
1771         struct cma_work *work;
1772
1773         work = kzalloc(sizeof *work, GFP_KERNEL);
1774         if (!work)
1775                 return -ENOMEM;
1776
1777         work->id = id_priv;
1778         INIT_WORK(&work->work, cma_work_handler);
1779         work->old_state = RDMA_CM_ROUTE_QUERY;
1780         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1781         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1782         queue_work(cma_wq, &work->work);
1783         return 0;
1784 }
1785
1786 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1787 {
1788         struct rdma_route *route = &id_priv->id.route;
1789         struct rdma_addr *addr = &route->addr;
1790         struct cma_work *work;
1791         int ret;
1792         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1793         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1794         struct net_device *ndev = NULL;
1795         u16 vid;
1796
1797         if (src_addr->sin_family != dst_addr->sin_family)
1798                 return -EINVAL;
1799
1800         work = kzalloc(sizeof *work, GFP_KERNEL);
1801         if (!work)
1802                 return -ENOMEM;
1803
1804         work->id = id_priv;
1805         INIT_WORK(&work->work, cma_work_handler);
1806
1807         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1808         if (!route->path_rec) {
1809                 ret = -ENOMEM;
1810                 goto err1;
1811         }
1812
1813         route->num_paths = 1;
1814
1815         if (addr->dev_addr.bound_dev_if)
1816                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1817         if (!ndev) {
1818                 ret = -ENODEV;
1819                 goto err2;
1820         }
1821
1822         vid = rdma_vlan_dev_vlan_id(ndev);
1823
1824         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1825         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1826
1827         route->path_rec->hop_limit = 1;
1828         route->path_rec->reversible = 1;
1829         route->path_rec->pkey = cpu_to_be16(0xffff);
1830         route->path_rec->mtu_selector = IB_SA_EQ;
1831         route->path_rec->sl = id_priv->tos >> 5;
1832
1833         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1834         route->path_rec->rate_selector = IB_SA_EQ;
1835         route->path_rec->rate = iboe_get_rate(ndev);
1836         dev_put(ndev);
1837         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1838         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1839         if (!route->path_rec->mtu) {
1840                 ret = -EINVAL;
1841                 goto err2;
1842         }
1843
1844         work->old_state = RDMA_CM_ROUTE_QUERY;
1845         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1846         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1847         work->event.status = 0;
1848
1849         queue_work(cma_wq, &work->work);
1850
1851         return 0;
1852
1853 err2:
1854         kfree(route->path_rec);
1855         route->path_rec = NULL;
1856 err1:
1857         kfree(work);
1858         return ret;
1859 }
1860
1861 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1862 {
1863         struct rdma_id_private *id_priv;
1864         int ret;
1865
1866         id_priv = container_of(id, struct rdma_id_private, id);
1867         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1868                 return -EINVAL;
1869
1870         atomic_inc(&id_priv->refcount);
1871         switch (rdma_node_get_transport(id->device->node_type)) {
1872         case RDMA_TRANSPORT_IB:
1873                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1874                 case IB_LINK_LAYER_INFINIBAND:
1875                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1876                         break;
1877                 case IB_LINK_LAYER_ETHERNET:
1878                         ret = cma_resolve_iboe_route(id_priv);
1879                         break;
1880                 default:
1881                         ret = -ENOSYS;
1882                 }
1883                 break;
1884         case RDMA_TRANSPORT_IWARP:
1885                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1886                 break;
1887         default:
1888                 ret = -ENOSYS;
1889                 break;
1890         }
1891         if (ret)
1892                 goto err;
1893
1894         return 0;
1895 err:
1896         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1897         cma_deref_id(id_priv);
1898         return ret;
1899 }
1900 EXPORT_SYMBOL(rdma_resolve_route);
1901
1902 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1903 {
1904         struct cma_device *cma_dev;
1905         struct ib_port_attr port_attr;
1906         union ib_gid gid;
1907         u16 pkey;
1908         int ret;
1909         u8 p;
1910
1911         mutex_lock(&lock);
1912         if (list_empty(&dev_list)) {
1913                 ret = -ENODEV;
1914                 goto out;
1915         }
1916         list_for_each_entry(cma_dev, &dev_list, list)
1917                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1918                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1919                             port_attr.state == IB_PORT_ACTIVE)
1920                                 goto port_found;
1921
1922         p = 1;
1923         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1924
1925 port_found:
1926         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1927         if (ret)
1928                 goto out;
1929
1930         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1931         if (ret)
1932                 goto out;
1933
1934         id_priv->id.route.addr.dev_addr.dev_type =
1935                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1936                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1937
1938         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1939         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1940         id_priv->id.port_num = p;
1941         cma_attach_to_dev(id_priv, cma_dev);
1942 out:
1943         mutex_unlock(&lock);
1944         return ret;
1945 }
1946
1947 static void addr_handler(int status, struct sockaddr *src_addr,
1948                          struct rdma_dev_addr *dev_addr, void *context)
1949 {
1950         struct rdma_id_private *id_priv = context;
1951         struct rdma_cm_event event;
1952
1953         memset(&event, 0, sizeof event);
1954         mutex_lock(&id_priv->handler_mutex);
1955         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1956                            RDMA_CM_ADDR_RESOLVED))
1957                 goto out;
1958
1959         if (!status && !id_priv->cma_dev)
1960                 status = cma_acquire_dev(id_priv);
1961
1962         if (status) {
1963                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1964                                    RDMA_CM_ADDR_BOUND))
1965                         goto out;
1966                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1967                 event.status = status;
1968         } else {
1969                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1970                        ip_addr_size(src_addr));
1971                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1972         }
1973
1974         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1975                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1976                 mutex_unlock(&id_priv->handler_mutex);
1977                 cma_deref_id(id_priv);
1978                 rdma_destroy_id(&id_priv->id);
1979                 return;
1980         }
1981 out:
1982         mutex_unlock(&id_priv->handler_mutex);
1983         cma_deref_id(id_priv);
1984 }
1985
1986 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1987 {
1988         struct cma_work *work;
1989         struct sockaddr *src, *dst;
1990         union ib_gid gid;
1991         int ret;
1992
1993         work = kzalloc(sizeof *work, GFP_KERNEL);
1994         if (!work)
1995                 return -ENOMEM;
1996
1997         if (!id_priv->cma_dev) {
1998                 ret = cma_bind_loopback(id_priv);
1999                 if (ret)
2000                         goto err;
2001         }
2002
2003         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2004         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2005
2006         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2007         if (cma_zero_addr(src)) {
2008                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2009                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2010                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
2011                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
2012                 } else {
2013                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
2014                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
2015                 }
2016         }
2017
2018         work->id = id_priv;
2019         INIT_WORK(&work->work, cma_work_handler);
2020         work->old_state = RDMA_CM_ADDR_QUERY;
2021         work->new_state = RDMA_CM_ADDR_RESOLVED;
2022         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2023         queue_work(cma_wq, &work->work);
2024         return 0;
2025 err:
2026         kfree(work);
2027         return ret;
2028 }
2029
2030 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2031                          struct sockaddr *dst_addr)
2032 {
2033         if (!src_addr || !src_addr->sa_family) {
2034                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2035                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2036                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2037                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2038                 }
2039         }
2040         return rdma_bind_addr(id, src_addr);
2041 }
2042
2043 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2044                       struct sockaddr *dst_addr, int timeout_ms)
2045 {
2046         struct rdma_id_private *id_priv;
2047         int ret;
2048
2049         id_priv = container_of(id, struct rdma_id_private, id);
2050         if (id_priv->state == RDMA_CM_IDLE) {
2051                 ret = cma_bind_addr(id, src_addr, dst_addr);
2052                 if (ret)
2053                         return ret;
2054         }
2055
2056         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2057                 return -EINVAL;
2058
2059         atomic_inc(&id_priv->refcount);
2060         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2061         if (cma_any_addr(dst_addr))
2062                 ret = cma_resolve_loopback(id_priv);
2063         else
2064                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2065                                       dst_addr, &id->route.addr.dev_addr,
2066                                       timeout_ms, addr_handler, id_priv);
2067         if (ret)
2068                 goto err;
2069
2070         return 0;
2071 err:
2072         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2073         cma_deref_id(id_priv);
2074         return ret;
2075 }
2076 EXPORT_SYMBOL(rdma_resolve_addr);
2077
2078 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2079 {
2080         struct rdma_id_private *id_priv;
2081         unsigned long flags;
2082         int ret;
2083
2084         id_priv = container_of(id, struct rdma_id_private, id);
2085         spin_lock_irqsave(&id_priv->lock, flags);
2086         if (id_priv->state == RDMA_CM_IDLE) {
2087                 id_priv->reuseaddr = reuse;
2088                 ret = 0;
2089         } else {
2090                 ret = -EINVAL;
2091         }
2092         spin_unlock_irqrestore(&id_priv->lock, flags);
2093         return ret;
2094 }
2095 EXPORT_SYMBOL(rdma_set_reuseaddr);
2096
2097 static void cma_bind_port(struct rdma_bind_list *bind_list,
2098                           struct rdma_id_private *id_priv)
2099 {
2100         struct sockaddr_in *sin;
2101
2102         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2103         sin->sin_port = htons(bind_list->port);
2104         id_priv->bind_list = bind_list;
2105         hlist_add_head(&id_priv->node, &bind_list->owners);
2106 }
2107
2108 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2109                           unsigned short snum)
2110 {
2111         struct rdma_bind_list *bind_list;
2112         int port, ret;
2113
2114         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2115         if (!bind_list)
2116                 return -ENOMEM;
2117
2118         do {
2119                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2120         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2121
2122         if (ret)
2123                 goto err1;
2124
2125         if (port != snum) {
2126                 ret = -EADDRNOTAVAIL;
2127                 goto err2;
2128         }
2129
2130         bind_list->ps = ps;
2131         bind_list->port = (unsigned short) port;
2132         cma_bind_port(bind_list, id_priv);
2133         return 0;
2134 err2:
2135         idr_remove(ps, port);
2136 err1:
2137         kfree(bind_list);
2138         return ret;
2139 }
2140
2141 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2142 {
2143         static unsigned int last_used_port;
2144         int low, high, remaining;
2145         unsigned int rover;
2146
2147         inet_get_local_port_range(&low, &high);
2148         remaining = (high - low) + 1;
2149         rover = net_random() % remaining + low;
2150 retry:
2151         if (last_used_port != rover &&
2152             !idr_find(ps, (unsigned short) rover)) {
2153                 int ret = cma_alloc_port(ps, id_priv, rover);
2154                 /*
2155                  * Remember previously used port number in order to avoid
2156                  * re-using same port immediately after it is closed.
2157                  */
2158                 if (!ret)
2159                         last_used_port = rover;
2160                 if (ret != -EADDRNOTAVAIL)
2161                         return ret;
2162         }
2163         if (--remaining) {
2164                 rover++;
2165                 if ((rover < low) || (rover > high))
2166                         rover = low;
2167                 goto retry;
2168         }
2169         return -EADDRNOTAVAIL;
2170 }
2171
2172 /*
2173  * Check that the requested port is available.  This is called when trying to
2174  * bind to a specific port, or when trying to listen on a bound port.  In
2175  * the latter case, the provided id_priv may already be on the bind_list, but
2176  * we still need to check that it's okay to start listening.
2177  */
2178 static int cma_check_port(struct rdma_bind_list *bind_list,
2179                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2180 {
2181         struct rdma_id_private *cur_id;
2182         struct sockaddr *addr, *cur_addr;
2183         struct hlist_node *node;
2184
2185         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2186         if (cma_any_addr(addr) && !reuseaddr)
2187                 return -EADDRNOTAVAIL;
2188
2189         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2190                 if (id_priv == cur_id)
2191                         continue;
2192
2193                 if ((cur_id->state == RDMA_CM_LISTEN) ||
2194                     !reuseaddr || !cur_id->reuseaddr) {
2195                         cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2196                         if (cma_any_addr(cur_addr))
2197                                 return -EADDRNOTAVAIL;
2198
2199                         if (!cma_addr_cmp(addr, cur_addr))
2200                                 return -EADDRINUSE;
2201                 }
2202         }
2203         return 0;
2204 }
2205
2206 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2207 {
2208         struct rdma_bind_list *bind_list;
2209         unsigned short snum;
2210         int ret;
2211
2212         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2213         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2214                 return -EACCES;
2215
2216         bind_list = idr_find(ps, snum);
2217         if (!bind_list) {
2218                 ret = cma_alloc_port(ps, id_priv, snum);
2219         } else {
2220                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2221                 if (!ret)
2222                         cma_bind_port(bind_list, id_priv);
2223         }
2224         return ret;
2225 }
2226
2227 static int cma_bind_listen(struct rdma_id_private *id_priv)
2228 {
2229         struct rdma_bind_list *bind_list = id_priv->bind_list;
2230         int ret = 0;
2231
2232         mutex_lock(&lock);
2233         if (bind_list->owners.first->next)
2234                 ret = cma_check_port(bind_list, id_priv, 0);
2235         mutex_unlock(&lock);
2236         return ret;
2237 }
2238
2239 static int cma_get_port(struct rdma_id_private *id_priv)
2240 {
2241         struct idr *ps;
2242         int ret;
2243
2244         switch (id_priv->id.ps) {
2245         case RDMA_PS_SDP:
2246                 ps = &sdp_ps;
2247                 break;
2248         case RDMA_PS_TCP:
2249                 ps = &tcp_ps;
2250                 break;
2251         case RDMA_PS_UDP:
2252                 ps = &udp_ps;
2253                 break;
2254         case RDMA_PS_IPOIB:
2255                 ps = &ipoib_ps;
2256                 break;
2257         case RDMA_PS_IB:
2258                 ps = &ib_ps;
2259                 break;
2260         default:
2261                 return -EPROTONOSUPPORT;
2262         }
2263
2264         mutex_lock(&lock);
2265         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2266                 ret = cma_alloc_any_port(ps, id_priv);
2267         else
2268                 ret = cma_use_port(ps, id_priv);
2269         mutex_unlock(&lock);
2270
2271         return ret;
2272 }
2273
2274 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2275                                struct sockaddr *addr)
2276 {
2277 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2278         struct sockaddr_in6 *sin6;
2279
2280         if (addr->sa_family != AF_INET6)
2281                 return 0;
2282
2283         sin6 = (struct sockaddr_in6 *) addr;
2284         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2285             !sin6->sin6_scope_id)
2286                         return -EINVAL;
2287
2288         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2289 #endif
2290         return 0;
2291 }
2292
2293 int rdma_listen(struct rdma_cm_id *id, int backlog)
2294 {
2295         struct rdma_id_private *id_priv;
2296         int ret;
2297
2298         id_priv = container_of(id, struct rdma_id_private, id);
2299         if (id_priv->state == RDMA_CM_IDLE) {
2300                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2301                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2302                 if (ret)
2303                         return ret;
2304         }
2305
2306         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2307                 return -EINVAL;
2308
2309         if (id_priv->reuseaddr) {
2310                 ret = cma_bind_listen(id_priv);
2311                 if (ret)
2312                         goto err;
2313         }
2314
2315         id_priv->backlog = backlog;
2316         if (id->device) {
2317                 switch (rdma_node_get_transport(id->device->node_type)) {
2318                 case RDMA_TRANSPORT_IB:
2319                         ret = cma_ib_listen(id_priv);
2320                         if (ret)
2321                                 goto err;
2322                         break;
2323                 case RDMA_TRANSPORT_IWARP:
2324                         ret = cma_iw_listen(id_priv, backlog);
2325                         if (ret)
2326                                 goto err;
2327                         break;
2328                 default:
2329                         ret = -ENOSYS;
2330                         goto err;
2331                 }
2332         } else
2333                 cma_listen_on_all(id_priv);
2334
2335         return 0;
2336 err:
2337         id_priv->backlog = 0;
2338         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2339         return ret;
2340 }
2341 EXPORT_SYMBOL(rdma_listen);
2342
2343 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2344 {
2345         struct rdma_id_private *id_priv;
2346         int ret;
2347
2348         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2349                 return -EAFNOSUPPORT;
2350
2351         id_priv = container_of(id, struct rdma_id_private, id);
2352         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2353                 return -EINVAL;
2354
2355         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2356         if (ret)
2357                 goto err1;
2358
2359         if (!cma_any_addr(addr)) {
2360                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2361                 if (ret)
2362                         goto err1;
2363
2364                 ret = cma_acquire_dev(id_priv);
2365                 if (ret)
2366                         goto err1;
2367         }
2368
2369         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2370         ret = cma_get_port(id_priv);
2371         if (ret)
2372                 goto err2;
2373
2374         return 0;
2375 err2:
2376         if (id_priv->cma_dev)
2377                 cma_release_dev(id_priv);
2378 err1:
2379         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2380         return ret;
2381 }
2382 EXPORT_SYMBOL(rdma_bind_addr);
2383
2384 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2385                           struct rdma_route *route)
2386 {
2387         struct cma_hdr *cma_hdr;
2388         struct sdp_hh *sdp_hdr;
2389
2390         if (route->addr.src_addr.ss_family == AF_INET) {
2391                 struct sockaddr_in *src4, *dst4;
2392
2393                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2394                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2395
2396                 switch (ps) {
2397                 case RDMA_PS_SDP:
2398                         sdp_hdr = hdr;
2399                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2400                                 return -EINVAL;
2401                         sdp_set_ip_ver(sdp_hdr, 4);
2402                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2403                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2404                         sdp_hdr->port = src4->sin_port;
2405                         break;
2406                 default:
2407                         cma_hdr = hdr;
2408                         cma_hdr->cma_version = CMA_VERSION;
2409                         cma_set_ip_ver(cma_hdr, 4);
2410                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2411                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2412                         cma_hdr->port = src4->sin_port;
2413                         break;
2414                 }
2415         } else {
2416                 struct sockaddr_in6 *src6, *dst6;
2417
2418                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2419                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2420
2421                 switch (ps) {
2422                 case RDMA_PS_SDP:
2423                         sdp_hdr = hdr;
2424                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2425                                 return -EINVAL;
2426                         sdp_set_ip_ver(sdp_hdr, 6);
2427                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2428                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2429                         sdp_hdr->port = src6->sin6_port;
2430                         break;
2431                 default:
2432                         cma_hdr = hdr;
2433                         cma_hdr->cma_version = CMA_VERSION;
2434                         cma_set_ip_ver(cma_hdr, 6);
2435                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2436                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2437                         cma_hdr->port = src6->sin6_port;
2438                         break;
2439                 }
2440         }
2441         return 0;
2442 }
2443
2444 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2445                                 struct ib_cm_event *ib_event)
2446 {
2447         struct rdma_id_private *id_priv = cm_id->context;
2448         struct rdma_cm_event event;
2449         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2450         int ret = 0;
2451
2452         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2453                 return 0;
2454
2455         memset(&event, 0, sizeof event);
2456         switch (ib_event->event) {
2457         case IB_CM_SIDR_REQ_ERROR:
2458                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2459                 event.status = -ETIMEDOUT;
2460                 break;
2461         case IB_CM_SIDR_REP_RECEIVED:
2462                 event.param.ud.private_data = ib_event->private_data;
2463                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2464                 if (rep->status != IB_SIDR_SUCCESS) {
2465                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2466                         event.status = ib_event->param.sidr_rep_rcvd.status;
2467                         break;
2468                 }
2469                 ret = cma_set_qkey(id_priv);
2470                 if (ret) {
2471                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2472                         event.status = -EINVAL;
2473                         break;
2474                 }
2475                 if (id_priv->qkey != rep->qkey) {
2476                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2477                         event.status = -EINVAL;
2478                         break;
2479                 }
2480                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2481                                      id_priv->id.route.path_rec,
2482                                      &event.param.ud.ah_attr);
2483                 event.param.ud.qp_num = rep->qpn;
2484                 event.param.ud.qkey = rep->qkey;
2485                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2486                 event.status = 0;
2487                 break;
2488         default:
2489                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2490                        ib_event->event);
2491                 goto out;
2492         }
2493
2494         ret = id_priv->id.event_handler(&id_priv->id, &event);
2495         if (ret) {
2496                 /* Destroy the CM ID by returning a non-zero value. */
2497                 id_priv->cm_id.ib = NULL;
2498                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2499                 mutex_unlock(&id_priv->handler_mutex);
2500                 rdma_destroy_id(&id_priv->id);
2501                 return ret;
2502         }
2503 out:
2504         mutex_unlock(&id_priv->handler_mutex);
2505         return ret;
2506 }
2507
2508 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2509                               struct rdma_conn_param *conn_param)
2510 {
2511         struct ib_cm_sidr_req_param req;
2512         struct rdma_route *route;
2513         struct ib_cm_id *id;
2514         int ret;
2515
2516         req.private_data_len = sizeof(struct cma_hdr) +
2517                                conn_param->private_data_len;
2518         if (req.private_data_len < conn_param->private_data_len)
2519                 return -EINVAL;
2520
2521         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2522         if (!req.private_data)
2523                 return -ENOMEM;
2524
2525         if (conn_param->private_data && conn_param->private_data_len)
2526                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2527                        conn_param->private_data, conn_param->private_data_len);
2528
2529         route = &id_priv->id.route;
2530         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2531         if (ret)
2532                 goto out;
2533
2534         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2535                              id_priv);
2536         if (IS_ERR(id)) {
2537                 ret = PTR_ERR(id);
2538                 goto out;
2539         }
2540         id_priv->cm_id.ib = id;
2541
2542         req.path = route->path_rec;
2543         req.service_id = cma_get_service_id(id_priv->id.ps,
2544                                             (struct sockaddr *) &route->addr.dst_addr);
2545         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2546         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2547
2548         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2549         if (ret) {
2550                 ib_destroy_cm_id(id_priv->cm_id.ib);
2551                 id_priv->cm_id.ib = NULL;
2552         }
2553 out:
2554         kfree(req.private_data);
2555         return ret;
2556 }
2557
2558 static int cma_connect_ib(struct rdma_id_private *id_priv,
2559                           struct rdma_conn_param *conn_param)
2560 {
2561         struct ib_cm_req_param req;
2562         struct rdma_route *route;
2563         void *private_data;
2564         struct ib_cm_id *id;
2565         int offset, ret;
2566
2567         memset(&req, 0, sizeof req);
2568         offset = cma_user_data_offset(id_priv->id.ps);
2569         req.private_data_len = offset + conn_param->private_data_len;
2570         if (req.private_data_len < conn_param->private_data_len)
2571                 return -EINVAL;
2572
2573         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2574         if (!private_data)
2575                 return -ENOMEM;
2576
2577         if (conn_param->private_data && conn_param->private_data_len)
2578                 memcpy(private_data + offset, conn_param->private_data,
2579                        conn_param->private_data_len);
2580
2581         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2582         if (IS_ERR(id)) {
2583                 ret = PTR_ERR(id);
2584                 goto out;
2585         }
2586         id_priv->cm_id.ib = id;
2587
2588         route = &id_priv->id.route;
2589         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2590         if (ret)
2591                 goto out;
2592         req.private_data = private_data;
2593
2594         req.primary_path = &route->path_rec[0];
2595         if (route->num_paths == 2)
2596                 req.alternate_path = &route->path_rec[1];
2597
2598         req.service_id = cma_get_service_id(id_priv->id.ps,
2599                                             (struct sockaddr *) &route->addr.dst_addr);
2600         req.qp_num = id_priv->qp_num;
2601         req.qp_type = id_priv->id.qp_type;
2602         req.starting_psn = id_priv->seq_num;
2603         req.responder_resources = conn_param->responder_resources;
2604         req.initiator_depth = conn_param->initiator_depth;
2605         req.flow_control = conn_param->flow_control;
2606         req.retry_count = conn_param->retry_count;
2607         req.rnr_retry_count = conn_param->rnr_retry_count;
2608         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2609         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2610         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2611         req.srq = id_priv->srq ? 1 : 0;
2612
2613         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2614 out:
2615         if (ret && !IS_ERR(id)) {
2616                 ib_destroy_cm_id(id);
2617                 id_priv->cm_id.ib = NULL;
2618         }
2619
2620         kfree(private_data);
2621         return ret;
2622 }
2623
2624 static int cma_connect_iw(struct rdma_id_private *id_priv,
2625                           struct rdma_conn_param *conn_param)
2626 {
2627         struct iw_cm_id *cm_id;
2628         struct sockaddr_in* sin;
2629         int ret;
2630         struct iw_cm_conn_param iw_param;
2631
2632         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2633         if (IS_ERR(cm_id))
2634                 return PTR_ERR(cm_id);
2635
2636         id_priv->cm_id.iw = cm_id;
2637
2638         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2639         cm_id->local_addr = *sin;
2640
2641         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2642         cm_id->remote_addr = *sin;
2643
2644         ret = cma_modify_qp_rtr(id_priv, conn_param);
2645         if (ret)
2646                 goto out;
2647
2648         if (conn_param) {
2649                 iw_param.ord = conn_param->initiator_depth;
2650                 iw_param.ird = conn_param->responder_resources;
2651                 iw_param.private_data = conn_param->private_data;
2652                 iw_param.private_data_len = conn_param->private_data_len;
2653                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2654         } else {
2655                 memset(&iw_param, 0, sizeof iw_param);
2656                 iw_param.qpn = id_priv->qp_num;
2657         }
2658         ret = iw_cm_connect(cm_id, &iw_param);
2659 out:
2660         if (ret) {
2661                 iw_destroy_cm_id(cm_id);
2662                 id_priv->cm_id.iw = NULL;
2663         }
2664         return ret;
2665 }
2666
2667 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2668 {
2669         struct rdma_id_private *id_priv;
2670         int ret;
2671
2672         id_priv = container_of(id, struct rdma_id_private, id);
2673         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2674                 return -EINVAL;
2675
2676         if (!id->qp) {
2677                 id_priv->qp_num = conn_param->qp_num;
2678                 id_priv->srq = conn_param->srq;
2679         }
2680
2681         switch (rdma_node_get_transport(id->device->node_type)) {
2682         case RDMA_TRANSPORT_IB:
2683                 if (id->qp_type == IB_QPT_UD)
2684                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2685                 else
2686                         ret = cma_connect_ib(id_priv, conn_param);
2687                 break;
2688         case RDMA_TRANSPORT_IWARP:
2689                 ret = cma_connect_iw(id_priv, conn_param);
2690                 break;
2691         default:
2692                 ret = -ENOSYS;
2693                 break;
2694         }
2695         if (ret)
2696                 goto err;
2697
2698         return 0;
2699 err:
2700         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2701         return ret;
2702 }
2703 EXPORT_SYMBOL(rdma_connect);
2704
2705 static int cma_accept_ib(struct rdma_id_private *id_priv,
2706                          struct rdma_conn_param *conn_param)
2707 {
2708         struct ib_cm_rep_param rep;
2709         int ret;
2710
2711         ret = cma_modify_qp_rtr(id_priv, conn_param);
2712         if (ret)
2713                 goto out;
2714
2715         ret = cma_modify_qp_rts(id_priv, conn_param);
2716         if (ret)
2717                 goto out;
2718
2719         memset(&rep, 0, sizeof rep);
2720         rep.qp_num = id_priv->qp_num;
2721         rep.starting_psn = id_priv->seq_num;
2722         rep.private_data = conn_param->private_data;
2723         rep.private_data_len = conn_param->private_data_len;
2724         rep.responder_resources = conn_param->responder_resources;
2725         rep.initiator_depth = conn_param->initiator_depth;
2726         rep.failover_accepted = 0;
2727         rep.flow_control = conn_param->flow_control;
2728         rep.rnr_retry_count = conn_param->rnr_retry_count;
2729         rep.srq = id_priv->srq ? 1 : 0;
2730
2731         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2732 out:
2733         return ret;
2734 }
2735
2736 static int cma_accept_iw(struct rdma_id_private *id_priv,
2737                   struct rdma_conn_param *conn_param)
2738 {
2739         struct iw_cm_conn_param iw_param;
2740         int ret;
2741
2742         if (!conn_param)
2743                 return -EINVAL;
2744
2745         ret = cma_modify_qp_rtr(id_priv, conn_param);
2746         if (ret)
2747                 return ret;
2748
2749         iw_param.ord = conn_param->initiator_depth;
2750         iw_param.ird = conn_param->responder_resources;
2751         iw_param.private_data = conn_param->private_data;
2752         iw_param.private_data_len = conn_param->private_data_len;
2753         if (id_priv->id.qp) {
2754                 iw_param.qpn = id_priv->qp_num;
2755         } else
2756                 iw_param.qpn = conn_param->qp_num;
2757
2758         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2759 }
2760
2761 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2762                              enum ib_cm_sidr_status status,
2763                              const void *private_data, int private_data_len)
2764 {
2765         struct ib_cm_sidr_rep_param rep;
2766         int ret;
2767
2768         memset(&rep, 0, sizeof rep);
2769         rep.status = status;
2770         if (status == IB_SIDR_SUCCESS) {
2771                 ret = cma_set_qkey(id_priv);
2772                 if (ret)
2773                         return ret;
2774                 rep.qp_num = id_priv->qp_num;
2775                 rep.qkey = id_priv->qkey;
2776         }
2777         rep.private_data = private_data;
2778         rep.private_data_len = private_data_len;
2779
2780         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2781 }
2782
2783 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2784 {
2785         struct rdma_id_private *id_priv;
2786         int ret;
2787
2788         id_priv = container_of(id, struct rdma_id_private, id);
2789
2790         id_priv->owner = task_pid_nr(current);
2791
2792         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2793                 return -EINVAL;
2794
2795         if (!id->qp && conn_param) {
2796                 id_priv->qp_num = conn_param->qp_num;
2797                 id_priv->srq = conn_param->srq;
2798         }
2799
2800         switch (rdma_node_get_transport(id->device->node_type)) {
2801         case RDMA_TRANSPORT_IB:
2802                 if (id->qp_type == IB_QPT_UD) {
2803                         if (conn_param)
2804                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2805                                                         conn_param->private_data,
2806                                                         conn_param->private_data_len);
2807                         else
2808                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2809                                                         NULL, 0);
2810                 } else {
2811                         if (conn_param)
2812                                 ret = cma_accept_ib(id_priv, conn_param);
2813                         else
2814                                 ret = cma_rep_recv(id_priv);
2815                 }
2816                 break;
2817         case RDMA_TRANSPORT_IWARP:
2818                 ret = cma_accept_iw(id_priv, conn_param);
2819                 break;
2820         default:
2821                 ret = -ENOSYS;
2822                 break;
2823         }
2824
2825         if (ret)
2826                 goto reject;
2827
2828         return 0;
2829 reject:
2830         cma_modify_qp_err(id_priv);
2831         rdma_reject(id, NULL, 0);
2832         return ret;
2833 }
2834 EXPORT_SYMBOL(rdma_accept);
2835
2836 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2837 {
2838         struct rdma_id_private *id_priv;
2839         int ret;
2840
2841         id_priv = container_of(id, struct rdma_id_private, id);
2842         if (!id_priv->cm_id.ib)
2843                 return -EINVAL;
2844
2845         switch (id->device->node_type) {
2846         case RDMA_NODE_IB_CA:
2847                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2848                 break;
2849         default:
2850                 ret = 0;
2851                 break;
2852         }
2853         return ret;
2854 }
2855 EXPORT_SYMBOL(rdma_notify);
2856
2857 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2858                 u8 private_data_len)
2859 {
2860         struct rdma_id_private *id_priv;
2861         int ret;
2862
2863         id_priv = container_of(id, struct rdma_id_private, id);
2864         if (!id_priv->cm_id.ib)
2865                 return -EINVAL;
2866
2867         switch (rdma_node_get_transport(id->device->node_type)) {
2868         case RDMA_TRANSPORT_IB:
2869                 if (id->qp_type == IB_QPT_UD)
2870                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2871                                                 private_data, private_data_len);
2872                 else
2873                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2874                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2875                                              0, private_data, private_data_len);
2876                 break;
2877         case RDMA_TRANSPORT_IWARP:
2878                 ret = iw_cm_reject(id_priv->cm_id.iw,
2879                                    private_data, private_data_len);
2880                 break;
2881         default:
2882                 ret = -ENOSYS;
2883                 break;
2884         }
2885         return ret;
2886 }
2887 EXPORT_SYMBOL(rdma_reject);
2888
2889 int rdma_disconnect(struct rdma_cm_id *id)
2890 {
2891         struct rdma_id_private *id_priv;
2892         int ret;
2893
2894         id_priv = container_of(id, struct rdma_id_private, id);
2895         if (!id_priv->cm_id.ib)
2896                 return -EINVAL;
2897
2898         switch (rdma_node_get_transport(id->device->node_type)) {
2899         case RDMA_TRANSPORT_IB:
2900                 ret = cma_modify_qp_err(id_priv);
2901                 if (ret)
2902                         goto out;
2903                 /* Initiate or respond to a disconnect. */
2904                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2905                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2906                 break;
2907         case RDMA_TRANSPORT_IWARP:
2908                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2909                 break;
2910         default:
2911                 ret = -EINVAL;
2912                 break;
2913         }
2914 out:
2915         return ret;
2916 }
2917 EXPORT_SYMBOL(rdma_disconnect);
2918
2919 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2920 {
2921         struct rdma_id_private *id_priv;
2922         struct cma_multicast *mc = multicast->context;
2923         struct rdma_cm_event event;
2924         int ret;
2925
2926         id_priv = mc->id_priv;
2927         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2928             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2929                 return 0;
2930
2931         mutex_lock(&id_priv->qp_mutex);
2932         if (!status && id_priv->id.qp)
2933                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2934                                          multicast->rec.mlid);
2935         mutex_unlock(&id_priv->qp_mutex);
2936
2937         memset(&event, 0, sizeof event);
2938         event.status = status;
2939         event.param.ud.private_data = mc->context;
2940         if (!status) {
2941                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2942                 ib_init_ah_from_mcmember(id_priv->id.device,
2943                                          id_priv->id.port_num, &multicast->rec,
2944                                          &event.param.ud.ah_attr);
2945                 event.param.ud.qp_num = 0xFFFFFF;
2946                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2947         } else
2948                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2949
2950         ret = id_priv->id.event_handler(&id_priv->id, &event);
2951         if (ret) {
2952                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2953                 mutex_unlock(&id_priv->handler_mutex);
2954                 rdma_destroy_id(&id_priv->id);
2955                 return 0;
2956         }
2957
2958         mutex_unlock(&id_priv->handler_mutex);
2959         return 0;
2960 }
2961
2962 static void cma_set_mgid(struct rdma_id_private *id_priv,
2963                          struct sockaddr *addr, union ib_gid *mgid)
2964 {
2965         unsigned char mc_map[MAX_ADDR_LEN];
2966         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2967         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2968         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2969
2970         if (cma_any_addr(addr)) {
2971                 memset(mgid, 0, sizeof *mgid);
2972         } else if ((addr->sa_family == AF_INET6) &&
2973                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2974                                                                  0xFF10A01B)) {
2975                 /* IPv6 address is an SA assigned MGID. */
2976                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2977         } else if ((addr->sa_family == AF_INET6)) {
2978                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2979                 if (id_priv->id.ps == RDMA_PS_UDP)
2980                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2981                 *mgid = *(union ib_gid *) (mc_map + 4);
2982         } else {
2983                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2984                 if (id_priv->id.ps == RDMA_PS_UDP)
2985                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2986                 *mgid = *(union ib_gid *) (mc_map + 4);
2987         }
2988 }
2989
2990 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2991                                  struct cma_multicast *mc)
2992 {
2993         struct ib_sa_mcmember_rec rec;
2994         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2995         ib_sa_comp_mask comp_mask;
2996         int ret;
2997
2998         ib_addr_get_mgid(dev_addr, &rec.mgid);
2999         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3000                                      &rec.mgid, &rec);
3001         if (ret)
3002                 return ret;
3003
3004         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3005         if (id_priv->id.ps == RDMA_PS_UDP)
3006                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3007         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3008         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3009         rec.join_state = 1;
3010
3011         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3012                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3013                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3014                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3015                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3016
3017         if (id_priv->id.ps == RDMA_PS_IPOIB)
3018                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3019                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
3020
3021         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3022                                                 id_priv->id.port_num, &rec,
3023                                                 comp_mask, GFP_KERNEL,
3024                                                 cma_ib_mc_handler, mc);
3025         if (IS_ERR(mc->multicast.ib))
3026                 return PTR_ERR(mc->multicast.ib);
3027
3028         return 0;
3029 }
3030
3031 static void iboe_mcast_work_handler(struct work_struct *work)
3032 {
3033         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3034         struct cma_multicast *mc = mw->mc;
3035         struct ib_sa_multicast *m = mc->multicast.ib;
3036
3037         mc->multicast.ib->context = mc;
3038         cma_ib_mc_handler(0, m);
3039         kref_put(&mc->mcref, release_mc);
3040         kfree(mw);
3041 }
3042
3043 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3044 {
3045         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3046         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3047
3048         if (cma_any_addr(addr)) {
3049                 memset(mgid, 0, sizeof *mgid);
3050         } else if (addr->sa_family == AF_INET6) {
3051                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3052         } else {
3053                 mgid->raw[0] = 0xff;
3054                 mgid->raw[1] = 0x0e;
3055                 mgid->raw[2] = 0;
3056                 mgid->raw[3] = 0;
3057                 mgid->raw[4] = 0;
3058                 mgid->raw[5] = 0;
3059                 mgid->raw[6] = 0;
3060                 mgid->raw[7] = 0;
3061                 mgid->raw[8] = 0;
3062                 mgid->raw[9] = 0;
3063                 mgid->raw[10] = 0xff;
3064                 mgid->raw[11] = 0xff;
3065                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3066         }
3067 }
3068
3069 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3070                                    struct cma_multicast *mc)
3071 {
3072         struct iboe_mcast_work *work;
3073         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3074         int err;
3075         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3076         struct net_device *ndev = NULL;
3077
3078         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3079                 return -EINVAL;
3080
3081         work = kzalloc(sizeof *work, GFP_KERNEL);
3082         if (!work)
3083                 return -ENOMEM;
3084
3085         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3086         if (!mc->multicast.ib) {
3087                 err = -ENOMEM;
3088                 goto out1;
3089         }
3090
3091         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3092
3093         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3094         if (id_priv->id.ps == RDMA_PS_UDP)
3095                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3096
3097         if (dev_addr->bound_dev_if)
3098                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3099         if (!ndev) {
3100                 err = -ENODEV;
3101                 goto out2;
3102         }
3103         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3104         mc->multicast.ib->rec.hop_limit = 1;
3105         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3106         dev_put(ndev);
3107         if (!mc->multicast.ib->rec.mtu) {
3108                 err = -EINVAL;
3109                 goto out2;
3110         }
3111         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3112         work->id = id_priv;
3113         work->mc = mc;
3114         INIT_WORK(&work->work, iboe_mcast_work_handler);
3115         kref_get(&mc->mcref);
3116         queue_work(cma_wq, &work->work);
3117
3118         return 0;
3119
3120 out2:
3121         kfree(mc->multicast.ib);
3122 out1:
3123         kfree(work);
3124         return err;
3125 }
3126
3127 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3128                         void *context)
3129 {
3130         struct rdma_id_private *id_priv;
3131         struct cma_multicast *mc;
3132         int ret;
3133
3134         id_priv = container_of(id, struct rdma_id_private, id);
3135         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3136             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3137                 return -EINVAL;
3138
3139         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3140         if (!mc)
3141                 return -ENOMEM;
3142
3143         memcpy(&mc->addr, addr, ip_addr_size(addr));
3144         mc->context = context;
3145         mc->id_priv = id_priv;
3146
3147         spin_lock(&id_priv->lock);
3148         list_add(&mc->list, &id_priv->mc_list);
3149         spin_unlock(&id_priv->lock);
3150
3151         switch (rdma_node_get_transport(id->device->node_type)) {
3152         case RDMA_TRANSPORT_IB:
3153                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3154                 case IB_LINK_LAYER_INFINIBAND:
3155                         ret = cma_join_ib_multicast(id_priv, mc);
3156                         break;
3157                 case IB_LINK_LAYER_ETHERNET:
3158                         kref_init(&mc->mcref);
3159                         ret = cma_iboe_join_multicast(id_priv, mc);
3160                         break;
3161                 default:
3162                         ret = -EINVAL;
3163                 }
3164                 break;
3165         default:
3166                 ret = -ENOSYS;
3167                 break;
3168         }
3169
3170         if (ret) {
3171                 spin_lock_irq(&id_priv->lock);
3172                 list_del(&mc->list);
3173                 spin_unlock_irq(&id_priv->lock);
3174                 kfree(mc);
3175         }
3176         return ret;
3177 }
3178 EXPORT_SYMBOL(rdma_join_multicast);
3179
3180 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3181 {
3182         struct rdma_id_private *id_priv;
3183         struct cma_multicast *mc;
3184
3185         id_priv = container_of(id, struct rdma_id_private, id);
3186         spin_lock_irq(&id_priv->lock);
3187         list_for_each_entry(mc, &id_priv->mc_list, list) {
3188                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3189                         list_del(&mc->list);
3190                         spin_unlock_irq(&id_priv->lock);
3191
3192                         if (id->qp)
3193                                 ib_detach_mcast(id->qp,
3194                                                 &mc->multicast.ib->rec.mgid,
3195                                                 mc->multicast.ib->rec.mlid);
3196                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3197                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3198                                 case IB_LINK_LAYER_INFINIBAND:
3199                                         ib_sa_free_multicast(mc->multicast.ib);
3200                                         kfree(mc);
3201                                         break;
3202                                 case IB_LINK_LAYER_ETHERNET:
3203                                         kref_put(&mc->mcref, release_mc);
3204                                         break;
3205                                 default:
3206                                         break;
3207                                 }
3208                         }
3209                         return;
3210                 }
3211         }
3212         spin_unlock_irq(&id_priv->lock);
3213 }
3214 EXPORT_SYMBOL(rdma_leave_multicast);
3215
3216 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3217 {
3218         struct rdma_dev_addr *dev_addr;
3219         struct cma_ndev_work *work;
3220
3221         dev_addr = &id_priv->id.route.addr.dev_addr;
3222
3223         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3224             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3225                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3226                        ndev->name, &id_priv->id);
3227                 work = kzalloc(sizeof *work, GFP_KERNEL);
3228                 if (!work)
3229                         return -ENOMEM;
3230
3231                 INIT_WORK(&work->work, cma_ndev_work_handler);
3232                 work->id = id_priv;
3233                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3234                 atomic_inc(&id_priv->refcount);
3235                 queue_work(cma_wq, &work->work);
3236         }
3237
3238         return 0;
3239 }
3240
3241 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3242                                void *ctx)
3243 {
3244         struct net_device *ndev = (struct net_device *)ctx;
3245         struct cma_device *cma_dev;
3246         struct rdma_id_private *id_priv;
3247         int ret = NOTIFY_DONE;
3248
3249         if (dev_net(ndev) != &init_net)
3250                 return NOTIFY_DONE;
3251
3252         if (event != NETDEV_BONDING_FAILOVER)
3253                 return NOTIFY_DONE;
3254
3255         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3256                 return NOTIFY_DONE;
3257
3258         mutex_lock(&lock);
3259         list_for_each_entry(cma_dev, &dev_list, list)
3260                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3261                         ret = cma_netdev_change(ndev, id_priv);
3262                         if (ret)
3263                                 goto out;
3264                 }
3265
3266 out:
3267         mutex_unlock(&lock);
3268         return ret;
3269 }
3270
3271 static struct notifier_block cma_nb = {
3272         .notifier_call = cma_netdev_callback
3273 };
3274
3275 static void cma_add_one(struct ib_device *device)
3276 {
3277         struct cma_device *cma_dev;
3278         struct rdma_id_private *id_priv;
3279
3280         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3281         if (!cma_dev)
3282                 return;
3283
3284         cma_dev->device = device;
3285
3286         init_completion(&cma_dev->comp);
3287         atomic_set(&cma_dev->refcount, 1);
3288         INIT_LIST_HEAD(&cma_dev->id_list);
3289         ib_set_client_data(device, &cma_client, cma_dev);
3290
3291         mutex_lock(&lock);
3292         list_add_tail(&cma_dev->list, &dev_list);
3293         list_for_each_entry(id_priv, &listen_any_list, list)
3294                 cma_listen_on_dev(id_priv, cma_dev);
3295         mutex_unlock(&lock);
3296 }
3297
3298 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3299 {
3300         struct rdma_cm_event event;
3301         enum rdma_cm_state state;
3302         int ret = 0;
3303
3304         /* Record that we want to remove the device */
3305         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3306         if (state == RDMA_CM_DESTROYING)
3307                 return 0;
3308
3309         cma_cancel_operation(id_priv, state);
3310         mutex_lock(&id_priv->handler_mutex);
3311
3312         /* Check for destruction from another callback. */
3313         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3314                 goto out;
3315
3316         memset(&event, 0, sizeof event);
3317         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3318         ret = id_priv->id.event_handler(&id_priv->id, &event);
3319 out:
3320         mutex_unlock(&id_priv->handler_mutex);
3321         return ret;
3322 }
3323
3324 static void cma_process_remove(struct cma_device *cma_dev)
3325 {
3326         struct rdma_id_private *id_priv;
3327         int ret;
3328
3329         mutex_lock(&lock);
3330         while (!list_empty(&cma_dev->id_list)) {
3331                 id_priv = list_entry(cma_dev->id_list.next,
3332                                      struct rdma_id_private, list);
3333
3334                 list_del(&id_priv->listen_list);
3335                 list_del_init(&id_priv->list);
3336                 atomic_inc(&id_priv->refcount);
3337                 mutex_unlock(&lock);
3338
3339                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3340                 cma_deref_id(id_priv);
3341                 if (ret)
3342                         rdma_destroy_id(&id_priv->id);
3343
3344                 mutex_lock(&lock);
3345         }
3346         mutex_unlock(&lock);
3347
3348         cma_deref_dev(cma_dev);
3349         wait_for_completion(&cma_dev->comp);
3350 }
3351
3352 static void cma_remove_one(struct ib_device *device)
3353 {
3354         struct cma_device *cma_dev;
3355
3356         cma_dev = ib_get_client_data(device, &cma_client);
3357         if (!cma_dev)
3358                 return;
3359
3360         mutex_lock(&lock);
3361         list_del(&cma_dev->list);
3362         mutex_unlock(&lock);
3363
3364         cma_process_remove(cma_dev);
3365         kfree(cma_dev);
3366 }
3367
3368 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3369 {
3370         struct nlmsghdr *nlh;
3371         struct rdma_cm_id_stats *id_stats;
3372         struct rdma_id_private *id_priv;
3373         struct rdma_cm_id *id = NULL;
3374         struct cma_device *cma_dev;
3375         int i_dev = 0, i_id = 0;
3376
3377         /*
3378          * We export all of the IDs as a sequence of messages.  Each
3379          * ID gets its own netlink message.
3380          */
3381         mutex_lock(&lock);
3382
3383         list_for_each_entry(cma_dev, &dev_list, list) {
3384                 if (i_dev < cb->args[0]) {
3385                         i_dev++;
3386                         continue;
3387                 }
3388
3389                 i_id = 0;
3390                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3391                         if (i_id < cb->args[1]) {
3392                                 i_id++;
3393                                 continue;
3394                         }
3395
3396                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3397                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3398                                                 RDMA_NL_RDMA_CM_ID_STATS);
3399                         if (!id_stats)
3400                                 goto out;
3401
3402                         memset(id_stats, 0, sizeof *id_stats);
3403                         id = &id_priv->id;
3404                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3405                         id_stats->port_num = id->port_num;
3406                         id_stats->bound_dev_if =
3407                                 id->route.addr.dev_addr.bound_dev_if;
3408
3409                         if (id->route.addr.src_addr.ss_family == AF_INET) {
3410                                 if (ibnl_put_attr(skb, nlh,
3411                                                   sizeof(struct sockaddr_in),
3412                                                   &id->route.addr.src_addr,
3413                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3414                                         goto out;
3415                                 }
3416                                 if (ibnl_put_attr(skb, nlh,
3417                                                   sizeof(struct sockaddr_in),
3418                                                   &id->route.addr.dst_addr,
3419                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3420                                         goto out;
3421                                 }
3422                         } else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3423                                 if (ibnl_put_attr(skb, nlh,
3424                                                   sizeof(struct sockaddr_in6),
3425                                                   &id->route.addr.src_addr,
3426                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3427                                         goto out;
3428                                 }
3429                                 if (ibnl_put_attr(skb, nlh,
3430                                                   sizeof(struct sockaddr_in6),
3431                                                   &id->route.addr.dst_addr,
3432                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3433                                         goto out;
3434                                 }
3435                         }
3436
3437                         id_stats->pid           = id_priv->owner;
3438                         id_stats->port_space    = id->ps;
3439                         id_stats->cm_state      = id_priv->state;
3440                         id_stats->qp_num        = id_priv->qp_num;
3441                         id_stats->qp_type       = id->qp_type;
3442
3443                         i_id++;
3444                 }
3445
3446                 cb->args[1] = 0;
3447                 i_dev++;
3448         }
3449
3450 out:
3451         mutex_unlock(&lock);
3452         cb->args[0] = i_dev;
3453         cb->args[1] = i_id;
3454
3455         return skb->len;
3456 }
3457
3458 static const struct ibnl_client_cbs cma_cb_table[] = {
3459         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats },
3460 };
3461
3462 static int __init cma_init(void)
3463 {
3464         int ret;
3465
3466         cma_wq = create_singlethread_workqueue("rdma_cm");
3467         if (!cma_wq)
3468                 return -ENOMEM;
3469
3470         ib_sa_register_client(&sa_client);
3471         rdma_addr_register_client(&addr_client);
3472         register_netdevice_notifier(&cma_nb);
3473
3474         ret = ib_register_client(&cma_client);
3475         if (ret)
3476                 goto err;
3477
3478         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3479                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3480
3481         return 0;
3482
3483 err:
3484         unregister_netdevice_notifier(&cma_nb);
3485         rdma_addr_unregister_client(&addr_client);
3486         ib_sa_unregister_client(&sa_client);
3487         destroy_workqueue(cma_wq);
3488         return ret;
3489 }
3490
3491 static void __exit cma_cleanup(void)
3492 {
3493         ibnl_remove_client(RDMA_NL_RDMA_CM);
3494         ib_unregister_client(&cma_client);
3495         unregister_netdevice_notifier(&cma_nb);
3496         rdma_addr_unregister_client(&addr_client);
3497         ib_sa_unregister_client(&sa_client);
3498         destroy_workqueue(cma_wq);
3499         idr_destroy(&sdp_ps);
3500         idr_destroy(&tcp_ps);
3501         idr_destroy(&udp_ps);
3502         idr_destroy(&ipoib_ps);
3503         idr_destroy(&ib_ps);
3504 }
3505
3506 module_init(cma_init);
3507 module_exit(cma_cleanup);