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