drivers/infiniband/core/cma.c

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