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
1795         if (src_addr->sin_family != dst_addr->sin_family)
1796                 return -EINVAL;
1797
1798         work = kzalloc(sizeof *work, GFP_KERNEL);
1799         if (!work)
1800                 return -ENOMEM;
1801
1802         work->id = id_priv;
1803         INIT_WORK(&work->work, cma_work_handler);
1804
1805         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1806         if (!route->path_rec) {
1807                 ret = -ENOMEM;
1808                 goto err1;
1809         }
1810
1811         route->num_paths = 1;
1812
1813         iboe_mac_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr);
1814         iboe_mac_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr);
1815
1816         route->path_rec->hop_limit = 1;
1817         route->path_rec->reversible = 1;
1818         route->path_rec->pkey = cpu_to_be16(0xffff);
1819         route->path_rec->mtu_selector = IB_SA_EQ;
1820
1821         if (addr->dev_addr.bound_dev_if)
1822                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1823         if (!ndev) {
1824                 ret = -ENODEV;
1825                 goto err2;
1826         }
1827
1828         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1829         route->path_rec->rate_selector = IB_SA_EQ;
1830         route->path_rec->rate = iboe_get_rate(ndev);
1831         dev_put(ndev);
1832         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1833         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1834         if (!route->path_rec->mtu) {
1835                 ret = -EINVAL;
1836                 goto err2;
1837         }
1838
1839         work->old_state = CMA_ROUTE_QUERY;
1840         work->new_state = CMA_ROUTE_RESOLVED;
1841         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1842         work->event.status = 0;
1843
1844         queue_work(cma_wq, &work->work);
1845
1846         return 0;
1847
1848 err2:
1849         kfree(route->path_rec);
1850         route->path_rec = NULL;
1851 err1:
1852         kfree(work);
1853         return ret;
1854 }
1855
1856 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1857 {
1858         struct rdma_id_private *id_priv;
1859         int ret;
1860
1861         id_priv = container_of(id, struct rdma_id_private, id);
1862         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1863                 return -EINVAL;
1864
1865         atomic_inc(&id_priv->refcount);
1866         switch (rdma_node_get_transport(id->device->node_type)) {
1867         case RDMA_TRANSPORT_IB:
1868                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1869                 case IB_LINK_LAYER_INFINIBAND:
1870                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1871                         break;
1872                 case IB_LINK_LAYER_ETHERNET:
1873                         ret = cma_resolve_iboe_route(id_priv);
1874                         break;
1875                 default:
1876                         ret = -ENOSYS;
1877                 }
1878                 break;
1879         case RDMA_TRANSPORT_IWARP:
1880                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1881                 break;
1882         default:
1883                 ret = -ENOSYS;
1884                 break;
1885         }
1886         if (ret)
1887                 goto err;
1888
1889         return 0;
1890 err:
1891         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1892         cma_deref_id(id_priv);
1893         return ret;
1894 }
1895 EXPORT_SYMBOL(rdma_resolve_route);
1896
1897 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1898 {
1899         struct cma_device *cma_dev;
1900         struct ib_port_attr port_attr;
1901         union ib_gid gid;
1902         u16 pkey;
1903         int ret;
1904         u8 p;
1905
1906         mutex_lock(&lock);
1907         if (list_empty(&dev_list)) {
1908                 ret = -ENODEV;
1909                 goto out;
1910         }
1911         list_for_each_entry(cma_dev, &dev_list, list)
1912                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1913                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1914                             port_attr.state == IB_PORT_ACTIVE)
1915                                 goto port_found;
1916
1917         p = 1;
1918         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1919
1920 port_found:
1921         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1922         if (ret)
1923                 goto out;
1924
1925         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1926         if (ret)
1927                 goto out;
1928
1929         id_priv->id.route.addr.dev_addr.dev_type =
1930                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1931                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1932
1933         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1934         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1935         id_priv->id.port_num = p;
1936         cma_attach_to_dev(id_priv, cma_dev);
1937 out:
1938         mutex_unlock(&lock);
1939         return ret;
1940 }
1941
1942 static void addr_handler(int status, struct sockaddr *src_addr,
1943                          struct rdma_dev_addr *dev_addr, void *context)
1944 {
1945         struct rdma_id_private *id_priv = context;
1946         struct rdma_cm_event event;
1947
1948         memset(&event, 0, sizeof event);
1949         mutex_lock(&id_priv->handler_mutex);
1950
1951         /*
1952          * Grab mutex to block rdma_destroy_id() from removing the device while
1953          * we're trying to acquire it.
1954          */
1955         mutex_lock(&lock);
1956         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1957                 mutex_unlock(&lock);
1958                 goto out;
1959         }
1960
1961         if (!status && !id_priv->cma_dev)
1962                 status = cma_acquire_dev(id_priv);
1963         mutex_unlock(&lock);
1964
1965         if (status) {
1966                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1967                         goto out;
1968                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1969                 event.status = status;
1970         } else {
1971                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1972                        ip_addr_size(src_addr));
1973                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1974         }
1975
1976         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1977                 cma_exch(id_priv, CMA_DESTROYING);
1978                 mutex_unlock(&id_priv->handler_mutex);
1979                 cma_deref_id(id_priv);
1980                 rdma_destroy_id(&id_priv->id);
1981                 return;
1982         }
1983 out:
1984         mutex_unlock(&id_priv->handler_mutex);
1985         cma_deref_id(id_priv);
1986 }
1987
1988 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1989 {
1990         struct cma_work *work;
1991         struct sockaddr *src, *dst;
1992         union ib_gid gid;
1993         int ret;
1994
1995         work = kzalloc(sizeof *work, GFP_KERNEL);
1996         if (!work)
1997                 return -ENOMEM;
1998
1999         if (!id_priv->cma_dev) {
2000                 ret = cma_bind_loopback(id_priv);
2001                 if (ret)
2002                         goto err;
2003         }
2004
2005         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2006         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2007
2008         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2009         if (cma_zero_addr(src)) {
2010                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2011                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2012                         ((struct sockaddr_in *) src)->sin_addr.s_addr =
2013                                 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
2014                 } else {
2015                         ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
2016                                        &((struct sockaddr_in6 *) dst)->sin6_addr);
2017                 }
2018         }
2019
2020         work->id = id_priv;
2021         INIT_WORK(&work->work, cma_work_handler);
2022         work->old_state = CMA_ADDR_QUERY;
2023         work->new_state = CMA_ADDR_RESOLVED;
2024         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2025         queue_work(cma_wq, &work->work);
2026         return 0;
2027 err:
2028         kfree(work);
2029         return ret;
2030 }
2031
2032 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2033                          struct sockaddr *dst_addr)
2034 {
2035         if (!src_addr || !src_addr->sa_family) {
2036                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2037                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2038                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2039                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2040                 }
2041         }
2042         return rdma_bind_addr(id, src_addr);
2043 }
2044
2045 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2046                       struct sockaddr *dst_addr, int timeout_ms)
2047 {
2048         struct rdma_id_private *id_priv;
2049         int ret;
2050
2051         id_priv = container_of(id, struct rdma_id_private, id);
2052         if (id_priv->state == CMA_IDLE) {
2053                 ret = cma_bind_addr(id, src_addr, dst_addr);
2054                 if (ret)
2055                         return ret;
2056         }
2057
2058         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
2059                 return -EINVAL;
2060
2061         atomic_inc(&id_priv->refcount);
2062         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2063         if (cma_any_addr(dst_addr))
2064                 ret = cma_resolve_loopback(id_priv);
2065         else
2066                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2067                                       dst_addr, &id->route.addr.dev_addr,
2068                                       timeout_ms, addr_handler, id_priv);
2069         if (ret)
2070                 goto err;
2071
2072         return 0;
2073 err:
2074         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
2075         cma_deref_id(id_priv);
2076         return ret;
2077 }
2078 EXPORT_SYMBOL(rdma_resolve_addr);
2079
2080 static void cma_bind_port(struct rdma_bind_list *bind_list,
2081                           struct rdma_id_private *id_priv)
2082 {
2083         struct sockaddr_in *sin;
2084
2085         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2086         sin->sin_port = htons(bind_list->port);
2087         id_priv->bind_list = bind_list;
2088         hlist_add_head(&id_priv->node, &bind_list->owners);
2089 }
2090
2091 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2092                           unsigned short snum)
2093 {
2094         struct rdma_bind_list *bind_list;
2095         int port, ret;
2096
2097         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2098         if (!bind_list)
2099                 return -ENOMEM;
2100
2101         do {
2102                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2103         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2104
2105         if (ret)
2106                 goto err1;
2107
2108         if (port != snum) {
2109                 ret = -EADDRNOTAVAIL;
2110                 goto err2;
2111         }
2112
2113         bind_list->ps = ps;
2114         bind_list->port = (unsigned short) port;
2115         cma_bind_port(bind_list, id_priv);
2116         return 0;
2117 err2:
2118         idr_remove(ps, port);
2119 err1:
2120         kfree(bind_list);
2121         return ret;
2122 }
2123
2124 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2125 {
2126         static unsigned int last_used_port;
2127         int low, high, remaining;
2128         unsigned int rover;
2129
2130         inet_get_local_port_range(&low, &high);
2131         remaining = (high - low) + 1;
2132         rover = net_random() % remaining + low;
2133 retry:
2134         if (last_used_port != rover &&
2135             !idr_find(ps, (unsigned short) rover)) {
2136                 int ret = cma_alloc_port(ps, id_priv, rover);
2137                 /*
2138                  * Remember previously used port number in order to avoid
2139                  * re-using same port immediately after it is closed.
2140                  */
2141                 if (!ret)
2142                         last_used_port = rover;
2143                 if (ret != -EADDRNOTAVAIL)
2144                         return ret;
2145         }
2146         if (--remaining) {
2147                 rover++;
2148                 if ((rover < low) || (rover > high))
2149                         rover = low;
2150                 goto retry;
2151         }
2152         return -EADDRNOTAVAIL;
2153 }
2154
2155 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2156 {
2157         struct rdma_id_private *cur_id;
2158         struct sockaddr_in *sin, *cur_sin;
2159         struct rdma_bind_list *bind_list;
2160         struct hlist_node *node;
2161         unsigned short snum;
2162
2163         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2164         snum = ntohs(sin->sin_port);
2165         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2166                 return -EACCES;
2167
2168         bind_list = idr_find(ps, snum);
2169         if (!bind_list)
2170                 return cma_alloc_port(ps, id_priv, snum);
2171
2172         /*
2173          * We don't support binding to any address if anyone is bound to
2174          * a specific address on the same port.
2175          */
2176         if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2177                 return -EADDRNOTAVAIL;
2178
2179         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2180                 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2181                         return -EADDRNOTAVAIL;
2182
2183                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2184                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2185                         return -EADDRINUSE;
2186         }
2187
2188         cma_bind_port(bind_list, id_priv);
2189         return 0;
2190 }
2191
2192 static int cma_get_port(struct rdma_id_private *id_priv)
2193 {
2194         struct idr *ps;
2195         int ret;
2196
2197         switch (id_priv->id.ps) {
2198         case RDMA_PS_SDP:
2199                 ps = &sdp_ps;
2200                 break;
2201         case RDMA_PS_TCP:
2202                 ps = &tcp_ps;
2203                 break;
2204         case RDMA_PS_UDP:
2205                 ps = &udp_ps;
2206                 break;
2207         case RDMA_PS_IPOIB:
2208                 ps = &ipoib_ps;
2209                 break;
2210         default:
2211                 return -EPROTONOSUPPORT;
2212         }
2213
2214         mutex_lock(&lock);
2215         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2216                 ret = cma_alloc_any_port(ps, id_priv);
2217         else
2218                 ret = cma_use_port(ps, id_priv);
2219         mutex_unlock(&lock);
2220
2221         return ret;
2222 }
2223
2224 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2225                                struct sockaddr *addr)
2226 {
2227 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2228         struct sockaddr_in6 *sin6;
2229
2230         if (addr->sa_family != AF_INET6)
2231                 return 0;
2232
2233         sin6 = (struct sockaddr_in6 *) addr;
2234         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2235             !sin6->sin6_scope_id)
2236                         return -EINVAL;
2237
2238         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2239 #endif
2240         return 0;
2241 }
2242
2243 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2244 {
2245         struct rdma_id_private *id_priv;
2246         int ret;
2247
2248         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2249                 return -EAFNOSUPPORT;
2250
2251         id_priv = container_of(id, struct rdma_id_private, id);
2252         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2253                 return -EINVAL;
2254
2255         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2256         if (ret)
2257                 goto err1;
2258
2259         if (!cma_any_addr(addr)) {
2260                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2261                 if (ret)
2262                         goto err1;
2263
2264                 mutex_lock(&lock);
2265                 ret = cma_acquire_dev(id_priv);
2266                 mutex_unlock(&lock);
2267                 if (ret)
2268                         goto err1;
2269         }
2270
2271         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2272         ret = cma_get_port(id_priv);
2273         if (ret)
2274                 goto err2;
2275
2276         return 0;
2277 err2:
2278         if (id_priv->cma_dev) {
2279                 mutex_lock(&lock);
2280                 cma_detach_from_dev(id_priv);
2281                 mutex_unlock(&lock);
2282         }
2283 err1:
2284         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2285         return ret;
2286 }
2287 EXPORT_SYMBOL(rdma_bind_addr);
2288
2289 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2290                           struct rdma_route *route)
2291 {
2292         struct cma_hdr *cma_hdr;
2293         struct sdp_hh *sdp_hdr;
2294
2295         if (route->addr.src_addr.ss_family == AF_INET) {
2296                 struct sockaddr_in *src4, *dst4;
2297
2298                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2299                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2300
2301                 switch (ps) {
2302                 case RDMA_PS_SDP:
2303                         sdp_hdr = hdr;
2304                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2305                                 return -EINVAL;
2306                         sdp_set_ip_ver(sdp_hdr, 4);
2307                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2308                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2309                         sdp_hdr->port = src4->sin_port;
2310                         break;
2311                 default:
2312                         cma_hdr = hdr;
2313                         cma_hdr->cma_version = CMA_VERSION;
2314                         cma_set_ip_ver(cma_hdr, 4);
2315                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2316                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2317                         cma_hdr->port = src4->sin_port;
2318                         break;
2319                 }
2320         } else {
2321                 struct sockaddr_in6 *src6, *dst6;
2322
2323                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2324                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2325
2326                 switch (ps) {
2327                 case RDMA_PS_SDP:
2328                         sdp_hdr = hdr;
2329                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2330                                 return -EINVAL;
2331                         sdp_set_ip_ver(sdp_hdr, 6);
2332                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2333                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2334                         sdp_hdr->port = src6->sin6_port;
2335                         break;
2336                 default:
2337                         cma_hdr = hdr;
2338                         cma_hdr->cma_version = CMA_VERSION;
2339                         cma_set_ip_ver(cma_hdr, 6);
2340                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2341                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2342                         cma_hdr->port = src6->sin6_port;
2343                         break;
2344                 }
2345         }
2346         return 0;
2347 }
2348
2349 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2350                                 struct ib_cm_event *ib_event)
2351 {
2352         struct rdma_id_private *id_priv = cm_id->context;
2353         struct rdma_cm_event event;
2354         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2355         int ret = 0;
2356
2357         if (cma_disable_callback(id_priv, CMA_CONNECT))
2358                 return 0;
2359
2360         memset(&event, 0, sizeof event);
2361         switch (ib_event->event) {
2362         case IB_CM_SIDR_REQ_ERROR:
2363                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2364                 event.status = -ETIMEDOUT;
2365                 break;
2366         case IB_CM_SIDR_REP_RECEIVED:
2367                 event.param.ud.private_data = ib_event->private_data;
2368                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2369                 if (rep->status != IB_SIDR_SUCCESS) {
2370                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2371                         event.status = ib_event->param.sidr_rep_rcvd.status;
2372                         break;
2373                 }
2374                 ret = cma_set_qkey(id_priv);
2375                 if (ret) {
2376                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2377                         event.status = -EINVAL;
2378                         break;
2379                 }
2380                 if (id_priv->qkey != rep->qkey) {
2381                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2382                         event.status = -EINVAL;
2383                         break;
2384                 }
2385                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2386                                      id_priv->id.route.path_rec,
2387                                      &event.param.ud.ah_attr);
2388                 event.param.ud.qp_num = rep->qpn;
2389                 event.param.ud.qkey = rep->qkey;
2390                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2391                 event.status = 0;
2392                 break;
2393         default:
2394                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2395                        ib_event->event);
2396                 goto out;
2397         }
2398
2399         ret = id_priv->id.event_handler(&id_priv->id, &event);
2400         if (ret) {
2401                 /* Destroy the CM ID by returning a non-zero value. */
2402                 id_priv->cm_id.ib = NULL;
2403                 cma_exch(id_priv, CMA_DESTROYING);
2404                 mutex_unlock(&id_priv->handler_mutex);
2405                 rdma_destroy_id(&id_priv->id);
2406                 return ret;
2407         }
2408 out:
2409         mutex_unlock(&id_priv->handler_mutex);
2410         return ret;
2411 }
2412
2413 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2414                               struct rdma_conn_param *conn_param)
2415 {
2416         struct ib_cm_sidr_req_param req;
2417         struct rdma_route *route;
2418         int ret;
2419
2420         req.private_data_len = sizeof(struct cma_hdr) +
2421                                conn_param->private_data_len;
2422         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2423         if (!req.private_data)
2424                 return -ENOMEM;
2425
2426         if (conn_param->private_data && conn_param->private_data_len)
2427                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2428                        conn_param->private_data, conn_param->private_data_len);
2429
2430         route = &id_priv->id.route;
2431         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2432         if (ret)
2433                 goto out;
2434
2435         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2436                                             cma_sidr_rep_handler, id_priv);
2437         if (IS_ERR(id_priv->cm_id.ib)) {
2438                 ret = PTR_ERR(id_priv->cm_id.ib);
2439                 goto out;
2440         }
2441
2442         req.path = route->path_rec;
2443         req.service_id = cma_get_service_id(id_priv->id.ps,
2444                                             (struct sockaddr *) &route->addr.dst_addr);
2445         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2446         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2447
2448         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2449         if (ret) {
2450                 ib_destroy_cm_id(id_priv->cm_id.ib);
2451                 id_priv->cm_id.ib = NULL;
2452         }
2453 out:
2454         kfree(req.private_data);
2455         return ret;
2456 }
2457
2458 static int cma_connect_ib(struct rdma_id_private *id_priv,
2459                           struct rdma_conn_param *conn_param)
2460 {
2461         struct ib_cm_req_param req;
2462         struct rdma_route *route;
2463         void *private_data;
2464         int offset, ret;
2465
2466         memset(&req, 0, sizeof req);
2467         offset = cma_user_data_offset(id_priv->id.ps);
2468         req.private_data_len = offset + conn_param->private_data_len;
2469         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2470         if (!private_data)
2471                 return -ENOMEM;
2472
2473         if (conn_param->private_data && conn_param->private_data_len)
2474                 memcpy(private_data + offset, conn_param->private_data,
2475                        conn_param->private_data_len);
2476
2477         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2478                                             id_priv);
2479         if (IS_ERR(id_priv->cm_id.ib)) {
2480                 ret = PTR_ERR(id_priv->cm_id.ib);
2481                 goto out;
2482         }
2483
2484         route = &id_priv->id.route;
2485         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2486         if (ret)
2487                 goto out;
2488         req.private_data = private_data;
2489
2490         req.primary_path = &route->path_rec[0];
2491         if (route->num_paths == 2)
2492                 req.alternate_path = &route->path_rec[1];
2493
2494         req.service_id = cma_get_service_id(id_priv->id.ps,
2495                                             (struct sockaddr *) &route->addr.dst_addr);
2496         req.qp_num = id_priv->qp_num;
2497         req.qp_type = IB_QPT_RC;
2498         req.starting_psn = id_priv->seq_num;
2499         req.responder_resources = conn_param->responder_resources;
2500         req.initiator_depth = conn_param->initiator_depth;
2501         req.flow_control = conn_param->flow_control;
2502         req.retry_count = conn_param->retry_count;
2503         req.rnr_retry_count = conn_param->rnr_retry_count;
2504         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2505         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2506         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2507         req.srq = id_priv->srq ? 1 : 0;
2508
2509         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2510 out:
2511         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2512                 ib_destroy_cm_id(id_priv->cm_id.ib);
2513                 id_priv->cm_id.ib = NULL;
2514         }
2515
2516         kfree(private_data);
2517         return ret;
2518 }
2519
2520 static int cma_connect_iw(struct rdma_id_private *id_priv,
2521                           struct rdma_conn_param *conn_param)
2522 {
2523         struct iw_cm_id *cm_id;
2524         struct sockaddr_in* sin;
2525         int ret;
2526         struct iw_cm_conn_param iw_param;
2527
2528         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2529         if (IS_ERR(cm_id)) {
2530                 ret = PTR_ERR(cm_id);
2531                 goto out;
2532         }
2533
2534         id_priv->cm_id.iw = cm_id;
2535
2536         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2537         cm_id->local_addr = *sin;
2538
2539         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2540         cm_id->remote_addr = *sin;
2541
2542         ret = cma_modify_qp_rtr(id_priv, conn_param);
2543         if (ret)
2544                 goto out;
2545
2546         iw_param.ord = conn_param->initiator_depth;
2547         iw_param.ird = conn_param->responder_resources;
2548         iw_param.private_data = conn_param->private_data;
2549         iw_param.private_data_len = conn_param->private_data_len;
2550         if (id_priv->id.qp)
2551                 iw_param.qpn = id_priv->qp_num;
2552         else
2553                 iw_param.qpn = conn_param->qp_num;
2554         ret = iw_cm_connect(cm_id, &iw_param);
2555 out:
2556         if (ret && !IS_ERR(cm_id)) {
2557                 iw_destroy_cm_id(cm_id);
2558                 id_priv->cm_id.iw = NULL;
2559         }
2560         return ret;
2561 }
2562
2563 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2564 {
2565         struct rdma_id_private *id_priv;
2566         int ret;
2567
2568         id_priv = container_of(id, struct rdma_id_private, id);
2569         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2570                 return -EINVAL;
2571
2572         if (!id->qp) {
2573                 id_priv->qp_num = conn_param->qp_num;
2574                 id_priv->srq = conn_param->srq;
2575         }
2576
2577         switch (rdma_node_get_transport(id->device->node_type)) {
2578         case RDMA_TRANSPORT_IB:
2579                 if (cma_is_ud_ps(id->ps))
2580                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2581                 else
2582                         ret = cma_connect_ib(id_priv, conn_param);
2583                 break;
2584         case RDMA_TRANSPORT_IWARP:
2585                 ret = cma_connect_iw(id_priv, conn_param);
2586                 break;
2587         default:
2588                 ret = -ENOSYS;
2589                 break;
2590         }
2591         if (ret)
2592                 goto err;
2593
2594         return 0;
2595 err:
2596         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2597         return ret;
2598 }
2599 EXPORT_SYMBOL(rdma_connect);
2600
2601 static int cma_accept_ib(struct rdma_id_private *id_priv,
2602                          struct rdma_conn_param *conn_param)
2603 {
2604         struct ib_cm_rep_param rep;
2605         int ret;
2606
2607         ret = cma_modify_qp_rtr(id_priv, conn_param);
2608         if (ret)
2609                 goto out;
2610
2611         ret = cma_modify_qp_rts(id_priv, conn_param);
2612         if (ret)
2613                 goto out;
2614
2615         memset(&rep, 0, sizeof rep);
2616         rep.qp_num = id_priv->qp_num;
2617         rep.starting_psn = id_priv->seq_num;
2618         rep.private_data = conn_param->private_data;
2619         rep.private_data_len = conn_param->private_data_len;
2620         rep.responder_resources = conn_param->responder_resources;
2621         rep.initiator_depth = conn_param->initiator_depth;
2622         rep.failover_accepted = 0;
2623         rep.flow_control = conn_param->flow_control;
2624         rep.rnr_retry_count = conn_param->rnr_retry_count;
2625         rep.srq = id_priv->srq ? 1 : 0;
2626
2627         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2628 out:
2629         return ret;
2630 }
2631
2632 static int cma_accept_iw(struct rdma_id_private *id_priv,
2633                   struct rdma_conn_param *conn_param)
2634 {
2635         struct iw_cm_conn_param iw_param;
2636         int ret;
2637
2638         ret = cma_modify_qp_rtr(id_priv, conn_param);
2639         if (ret)
2640                 return ret;
2641
2642         iw_param.ord = conn_param->initiator_depth;
2643         iw_param.ird = conn_param->responder_resources;
2644         iw_param.private_data = conn_param->private_data;
2645         iw_param.private_data_len = conn_param->private_data_len;
2646         if (id_priv->id.qp) {
2647                 iw_param.qpn = id_priv->qp_num;
2648         } else
2649                 iw_param.qpn = conn_param->qp_num;
2650
2651         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2652 }
2653
2654 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2655                              enum ib_cm_sidr_status status,
2656                              const void *private_data, int private_data_len)
2657 {
2658         struct ib_cm_sidr_rep_param rep;
2659         int ret;
2660
2661         memset(&rep, 0, sizeof rep);
2662         rep.status = status;
2663         if (status == IB_SIDR_SUCCESS) {
2664                 ret = cma_set_qkey(id_priv);
2665                 if (ret)
2666                         return ret;
2667                 rep.qp_num = id_priv->qp_num;
2668                 rep.qkey = id_priv->qkey;
2669         }
2670         rep.private_data = private_data;
2671         rep.private_data_len = private_data_len;
2672
2673         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2674 }
2675
2676 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2677 {
2678         struct rdma_id_private *id_priv;
2679         int ret;
2680
2681         id_priv = container_of(id, struct rdma_id_private, id);
2682         if (!cma_comp(id_priv, CMA_CONNECT))
2683                 return -EINVAL;
2684
2685         if (!id->qp && conn_param) {
2686                 id_priv->qp_num = conn_param->qp_num;
2687                 id_priv->srq = conn_param->srq;
2688         }
2689
2690         switch (rdma_node_get_transport(id->device->node_type)) {
2691         case RDMA_TRANSPORT_IB:
2692                 if (cma_is_ud_ps(id->ps))
2693                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2694                                                 conn_param->private_data,
2695                                                 conn_param->private_data_len);
2696                 else if (conn_param)
2697                         ret = cma_accept_ib(id_priv, conn_param);
2698                 else
2699                         ret = cma_rep_recv(id_priv);
2700                 break;
2701         case RDMA_TRANSPORT_IWARP:
2702                 ret = cma_accept_iw(id_priv, conn_param);
2703                 break;
2704         default:
2705                 ret = -ENOSYS;
2706                 break;
2707         }
2708
2709         if (ret)
2710                 goto reject;
2711
2712         return 0;
2713 reject:
2714         cma_modify_qp_err(id_priv);
2715         rdma_reject(id, NULL, 0);
2716         return ret;
2717 }
2718 EXPORT_SYMBOL(rdma_accept);
2719
2720 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2721 {
2722         struct rdma_id_private *id_priv;
2723         int ret;
2724
2725         id_priv = container_of(id, struct rdma_id_private, id);
2726         if (!cma_has_cm_dev(id_priv))
2727                 return -EINVAL;
2728
2729         switch (id->device->node_type) {
2730         case RDMA_NODE_IB_CA:
2731                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2732                 break;
2733         default:
2734                 ret = 0;
2735                 break;
2736         }
2737         return ret;
2738 }
2739 EXPORT_SYMBOL(rdma_notify);
2740
2741 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2742                 u8 private_data_len)
2743 {
2744         struct rdma_id_private *id_priv;
2745         int ret;
2746
2747         id_priv = container_of(id, struct rdma_id_private, id);
2748         if (!cma_has_cm_dev(id_priv))
2749                 return -EINVAL;
2750
2751         switch (rdma_node_get_transport(id->device->node_type)) {
2752         case RDMA_TRANSPORT_IB:
2753                 if (cma_is_ud_ps(id->ps))
2754                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2755                                                 private_data, private_data_len);
2756                 else
2757                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2758                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2759                                              0, private_data, private_data_len);
2760                 break;
2761         case RDMA_TRANSPORT_IWARP:
2762                 ret = iw_cm_reject(id_priv->cm_id.iw,
2763                                    private_data, private_data_len);
2764                 break;
2765         default:
2766                 ret = -ENOSYS;
2767                 break;
2768         }
2769         return ret;
2770 }
2771 EXPORT_SYMBOL(rdma_reject);
2772
2773 int rdma_disconnect(struct rdma_cm_id *id)
2774 {
2775         struct rdma_id_private *id_priv;
2776         int ret;
2777
2778         id_priv = container_of(id, struct rdma_id_private, id);
2779         if (!cma_has_cm_dev(id_priv))
2780                 return -EINVAL;
2781
2782         switch (rdma_node_get_transport(id->device->node_type)) {
2783         case RDMA_TRANSPORT_IB:
2784                 ret = cma_modify_qp_err(id_priv);
2785                 if (ret)
2786                         goto out;
2787                 /* Initiate or respond to a disconnect. */
2788                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2789                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2790                 break;
2791         case RDMA_TRANSPORT_IWARP:
2792                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2793                 break;
2794         default:
2795                 ret = -EINVAL;
2796                 break;
2797         }
2798 out:
2799         return ret;
2800 }
2801 EXPORT_SYMBOL(rdma_disconnect);
2802
2803 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2804 {
2805         struct rdma_id_private *id_priv;
2806         struct cma_multicast *mc = multicast->context;
2807         struct rdma_cm_event event;
2808         int ret;
2809
2810         id_priv = mc->id_priv;
2811         if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2812             cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2813                 return 0;
2814
2815         mutex_lock(&id_priv->qp_mutex);
2816         if (!status && id_priv->id.qp)
2817                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2818                                          multicast->rec.mlid);
2819         mutex_unlock(&id_priv->qp_mutex);
2820
2821         memset(&event, 0, sizeof event);
2822         event.status = status;
2823         event.param.ud.private_data = mc->context;
2824         if (!status) {
2825                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2826                 ib_init_ah_from_mcmember(id_priv->id.device,
2827                                          id_priv->id.port_num, &multicast->rec,
2828                                          &event.param.ud.ah_attr);
2829                 event.param.ud.qp_num = 0xFFFFFF;
2830                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2831         } else
2832                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2833
2834         ret = id_priv->id.event_handler(&id_priv->id, &event);
2835         if (ret) {
2836                 cma_exch(id_priv, CMA_DESTROYING);
2837                 mutex_unlock(&id_priv->handler_mutex);
2838                 rdma_destroy_id(&id_priv->id);
2839                 return 0;
2840         }
2841
2842         mutex_unlock(&id_priv->handler_mutex);
2843         return 0;
2844 }
2845
2846 static void cma_set_mgid(struct rdma_id_private *id_priv,
2847                          struct sockaddr *addr, union ib_gid *mgid)
2848 {
2849         unsigned char mc_map[MAX_ADDR_LEN];
2850         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2851         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2852         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2853
2854         if (cma_any_addr(addr)) {
2855                 memset(mgid, 0, sizeof *mgid);
2856         } else if ((addr->sa_family == AF_INET6) &&
2857                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2858                                                                  0xFF10A01B)) {
2859                 /* IPv6 address is an SA assigned MGID. */
2860                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2861         } else if ((addr->sa_family == AF_INET6)) {
2862                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2863                 if (id_priv->id.ps == RDMA_PS_UDP)
2864                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2865                 *mgid = *(union ib_gid *) (mc_map + 4);
2866         } else {
2867                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2868                 if (id_priv->id.ps == RDMA_PS_UDP)
2869                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2870                 *mgid = *(union ib_gid *) (mc_map + 4);
2871         }
2872 }
2873
2874 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2875                                  struct cma_multicast *mc)
2876 {
2877         struct ib_sa_mcmember_rec rec;
2878         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2879         ib_sa_comp_mask comp_mask;
2880         int ret;
2881
2882         ib_addr_get_mgid(dev_addr, &rec.mgid);
2883         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2884                                      &rec.mgid, &rec);
2885         if (ret)
2886                 return ret;
2887
2888         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2889         if (id_priv->id.ps == RDMA_PS_UDP)
2890                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2891         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2892         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2893         rec.join_state = 1;
2894
2895         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2896                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2897                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2898                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2899                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2900
2901         if (id_priv->id.ps == RDMA_PS_IPOIB)
2902                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2903                              IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2904
2905         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2906                                                 id_priv->id.port_num, &rec,
2907                                                 comp_mask, GFP_KERNEL,
2908                                                 cma_ib_mc_handler, mc);
2909         if (IS_ERR(mc->multicast.ib))
2910                 return PTR_ERR(mc->multicast.ib);
2911
2912         return 0;
2913 }
2914
2915 static void iboe_mcast_work_handler(struct work_struct *work)
2916 {
2917         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
2918         struct cma_multicast *mc = mw->mc;
2919         struct ib_sa_multicast *m = mc->multicast.ib;
2920
2921         mc->multicast.ib->context = mc;
2922         cma_ib_mc_handler(0, m);
2923         kref_put(&mc->mcref, release_mc);
2924         kfree(mw);
2925 }
2926
2927 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
2928 {
2929         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
2930         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
2931
2932         if (cma_any_addr(addr)) {
2933                 memset(mgid, 0, sizeof *mgid);
2934         } else if (addr->sa_family == AF_INET6) {
2935                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2936         } else {
2937                 mgid->raw[0] = 0xff;
2938                 mgid->raw[1] = 0x0e;
2939                 mgid->raw[2] = 0;
2940                 mgid->raw[3] = 0;
2941                 mgid->raw[4] = 0;
2942                 mgid->raw[5] = 0;
2943                 mgid->raw[6] = 0;
2944                 mgid->raw[7] = 0;
2945                 mgid->raw[8] = 0;
2946                 mgid->raw[9] = 0;
2947                 mgid->raw[10] = 0xff;
2948                 mgid->raw[11] = 0xff;
2949                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
2950         }
2951 }
2952
2953 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
2954                                    struct cma_multicast *mc)
2955 {
2956         struct iboe_mcast_work *work;
2957         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2958         int err;
2959         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
2960         struct net_device *ndev = NULL;
2961
2962         if (cma_zero_addr((struct sockaddr *)&mc->addr))
2963                 return -EINVAL;
2964
2965         work = kzalloc(sizeof *work, GFP_KERNEL);
2966         if (!work)
2967                 return -ENOMEM;
2968
2969         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
2970         if (!mc->multicast.ib) {
2971                 err = -ENOMEM;
2972                 goto out1;
2973         }
2974
2975         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
2976
2977         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
2978         if (id_priv->id.ps == RDMA_PS_UDP)
2979                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2980
2981         if (dev_addr->bound_dev_if)
2982                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
2983         if (!ndev) {
2984                 err = -ENODEV;
2985                 goto out2;
2986         }
2987         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
2988         mc->multicast.ib->rec.hop_limit = 1;
2989         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
2990         dev_put(ndev);
2991         if (!mc->multicast.ib->rec.mtu) {
2992                 err = -EINVAL;
2993                 goto out2;
2994         }
2995         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
2996         work->id = id_priv;
2997         work->mc = mc;
2998         INIT_WORK(&work->work, iboe_mcast_work_handler);
2999         kref_get(&mc->mcref);
3000         queue_work(cma_wq, &work->work);
3001
3002         return 0;
3003
3004 out2:
3005         kfree(mc->multicast.ib);
3006 out1:
3007         kfree(work);
3008         return err;
3009 }
3010
3011 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3012                         void *context)
3013 {
3014         struct rdma_id_private *id_priv;
3015         struct cma_multicast *mc;
3016         int ret;
3017
3018         id_priv = container_of(id, struct rdma_id_private, id);
3019         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
3020             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
3021                 return -EINVAL;
3022
3023         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3024         if (!mc)
3025                 return -ENOMEM;
3026
3027         memcpy(&mc->addr, addr, ip_addr_size(addr));
3028         mc->context = context;
3029         mc->id_priv = id_priv;
3030
3031         spin_lock(&id_priv->lock);
3032         list_add(&mc->list, &id_priv->mc_list);
3033         spin_unlock(&id_priv->lock);
3034
3035         switch (rdma_node_get_transport(id->device->node_type)) {
3036         case RDMA_TRANSPORT_IB:
3037                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3038                 case IB_LINK_LAYER_INFINIBAND:
3039                         ret = cma_join_ib_multicast(id_priv, mc);
3040                         break;
3041                 case IB_LINK_LAYER_ETHERNET:
3042                         kref_init(&mc->mcref);
3043                         ret = cma_iboe_join_multicast(id_priv, mc);
3044                         break;
3045                 default:
3046                         ret = -EINVAL;
3047                 }
3048                 break;
3049         default:
3050                 ret = -ENOSYS;
3051                 break;
3052         }
3053
3054         if (ret) {
3055                 spin_lock_irq(&id_priv->lock);
3056                 list_del(&mc->list);
3057                 spin_unlock_irq(&id_priv->lock);
3058                 kfree(mc);
3059         }
3060         return ret;
3061 }
3062 EXPORT_SYMBOL(rdma_join_multicast);
3063
3064 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3065 {
3066         struct rdma_id_private *id_priv;
3067         struct cma_multicast *mc;
3068
3069         id_priv = container_of(id, struct rdma_id_private, id);
3070         spin_lock_irq(&id_priv->lock);
3071         list_for_each_entry(mc, &id_priv->mc_list, list) {
3072                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3073                         list_del(&mc->list);
3074                         spin_unlock_irq(&id_priv->lock);
3075
3076                         if (id->qp)
3077                                 ib_detach_mcast(id->qp,
3078                                                 &mc->multicast.ib->rec.mgid,
3079                                                 mc->multicast.ib->rec.mlid);
3080                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3081                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3082                                 case IB_LINK_LAYER_INFINIBAND:
3083                                         ib_sa_free_multicast(mc->multicast.ib);
3084                                         kfree(mc);
3085                                         break;
3086                                 case IB_LINK_LAYER_ETHERNET:
3087                                         kref_put(&mc->mcref, release_mc);
3088                                         break;
3089                                 default:
3090                                         break;
3091                                 }
3092                         }
3093                         return;
3094                 }
3095         }
3096         spin_unlock_irq(&id_priv->lock);
3097 }
3098 EXPORT_SYMBOL(rdma_leave_multicast);
3099
3100 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3101 {
3102         struct rdma_dev_addr *dev_addr;
3103         struct cma_ndev_work *work;
3104
3105         dev_addr = &id_priv->id.route.addr.dev_addr;
3106
3107         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3108             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3109                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3110                        ndev->name, &id_priv->id);
3111                 work = kzalloc(sizeof *work, GFP_KERNEL);
3112                 if (!work)
3113                         return -ENOMEM;
3114
3115                 INIT_WORK(&work->work, cma_ndev_work_handler);
3116                 work->id = id_priv;
3117                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3118                 atomic_inc(&id_priv->refcount);
3119                 queue_work(cma_wq, &work->work);
3120         }
3121
3122         return 0;
3123 }
3124
3125 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3126                                void *ctx)
3127 {
3128         struct net_device *ndev = (struct net_device *)ctx;
3129         struct cma_device *cma_dev;
3130         struct rdma_id_private *id_priv;
3131         int ret = NOTIFY_DONE;
3132
3133         if (dev_net(ndev) != &init_net)
3134                 return NOTIFY_DONE;
3135
3136         if (event != NETDEV_BONDING_FAILOVER)
3137                 return NOTIFY_DONE;
3138
3139         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3140                 return NOTIFY_DONE;
3141
3142         mutex_lock(&lock);
3143         list_for_each_entry(cma_dev, &dev_list, list)
3144                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3145                         ret = cma_netdev_change(ndev, id_priv);
3146                         if (ret)
3147                                 goto out;
3148                 }
3149
3150 out:
3151         mutex_unlock(&lock);
3152         return ret;
3153 }
3154
3155 static struct notifier_block cma_nb = {
3156         .notifier_call = cma_netdev_callback
3157 };
3158
3159 static void cma_add_one(struct ib_device *device)
3160 {
3161         struct cma_device *cma_dev;
3162         struct rdma_id_private *id_priv;
3163
3164         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3165         if (!cma_dev)
3166                 return;
3167
3168         cma_dev->device = device;
3169
3170         init_completion(&cma_dev->comp);
3171         atomic_set(&cma_dev->refcount, 1);
3172         INIT_LIST_HEAD(&cma_dev->id_list);
3173         ib_set_client_data(device, &cma_client, cma_dev);
3174
3175         mutex_lock(&lock);
3176         list_add_tail(&cma_dev->list, &dev_list);
3177         list_for_each_entry(id_priv, &listen_any_list, list)
3178                 cma_listen_on_dev(id_priv, cma_dev);
3179         mutex_unlock(&lock);
3180 }
3181
3182 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3183 {
3184         struct rdma_cm_event event;
3185         enum cma_state state;
3186         int ret = 0;
3187
3188         /* Record that we want to remove the device */
3189         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
3190         if (state == CMA_DESTROYING)
3191                 return 0;
3192
3193         cma_cancel_operation(id_priv, state);
3194         mutex_lock(&id_priv->handler_mutex);
3195
3196         /* Check for destruction from another callback. */
3197         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
3198                 goto out;
3199
3200         memset(&event, 0, sizeof event);
3201         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3202         ret = id_priv->id.event_handler(&id_priv->id, &event);
3203 out:
3204         mutex_unlock(&id_priv->handler_mutex);
3205         return ret;
3206 }
3207
3208 static void cma_process_remove(struct cma_device *cma_dev)
3209 {
3210         struct rdma_id_private *id_priv;
3211         int ret;
3212
3213         mutex_lock(&lock);
3214         while (!list_empty(&cma_dev->id_list)) {
3215                 id_priv = list_entry(cma_dev->id_list.next,
3216                                      struct rdma_id_private, list);
3217
3218                 list_del(&id_priv->listen_list);
3219                 list_del_init(&id_priv->list);
3220                 atomic_inc(&id_priv->refcount);
3221                 mutex_unlock(&lock);
3222
3223                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3224                 cma_deref_id(id_priv);
3225                 if (ret)
3226                         rdma_destroy_id(&id_priv->id);
3227
3228                 mutex_lock(&lock);
3229         }
3230         mutex_unlock(&lock);
3231
3232         cma_deref_dev(cma_dev);
3233         wait_for_completion(&cma_dev->comp);
3234 }
3235
3236 static void cma_remove_one(struct ib_device *device)
3237 {
3238         struct cma_device *cma_dev;
3239
3240         cma_dev = ib_get_client_data(device, &cma_client);
3241         if (!cma_dev)
3242                 return;
3243
3244         mutex_lock(&lock);
3245         list_del(&cma_dev->list);
3246         mutex_unlock(&lock);
3247
3248         cma_process_remove(cma_dev);
3249         kfree(cma_dev);
3250 }
3251
3252 static int __init cma_init(void)
3253 {
3254         int ret;
3255
3256         cma_wq = create_singlethread_workqueue("rdma_cm");
3257         if (!cma_wq)
3258                 return -ENOMEM;
3259
3260         ib_sa_register_client(&sa_client);
3261         rdma_addr_register_client(&addr_client);
3262         register_netdevice_notifier(&cma_nb);
3263
3264         ret = ib_register_client(&cma_client);
3265         if (ret)
3266                 goto err;
3267         return 0;
3268
3269 err:
3270         unregister_netdevice_notifier(&cma_nb);
3271         rdma_addr_unregister_client(&addr_client);
3272         ib_sa_unregister_client(&sa_client);
3273         destroy_workqueue(cma_wq);
3274         return ret;
3275 }
3276
3277 static void __exit cma_cleanup(void)
3278 {
3279         ib_unregister_client(&cma_client);
3280         unregister_netdevice_notifier(&cma_nb);
3281         rdma_addr_unregister_client(&addr_client);
3282         ib_sa_unregister_client(&sa_client);
3283         destroy_workqueue(cma_wq);
3284         idr_destroy(&sdp_ps);
3285         idr_destroy(&tcp_ps);
3286         idr_destroy(&udp_ps);
3287         idr_destroy(&ipoib_ps);
3288 }
3289
3290 module_init(cma_init);
3291 module_exit(cma_cleanup);