drivers/infiniband/core/cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/idr.h>
  42 #include <linux/inetdevice.h>
  43 #include <linux/slab.h>
  44 #include <linux/module.h>
  45 #include <net/route.h>
  46
  47 #include <net/tcp.h>
  48 #include <net/ipv6.h>
  49
  50 #include <rdma/rdma_cm.h>
  51 #include <rdma/rdma_cm_ib.h>
  52 #include <rdma/rdma_netlink.h>
  53 #include <rdma/ib_cache.h>
  54 #include <rdma/ib_cm.h>
  55 #include <rdma/ib_sa.h>
  56 #include <rdma/iw_cm.h>
  57
  58 MODULE_AUTHOR("Sean Hefty");
  59 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  60 MODULE_LICENSE("Dual BSD/GPL");
  61
  62 #define CMA_CM_RESPONSE_TIMEOUT 20
  63 #define CMA_MAX_CM_RETRIES 15
  64 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  65 #define CMA_IBOE_PACKET_LIFETIME 18
  66
  67 static void cma_add_one(struct ib_device *device);
  68 static void cma_remove_one(struct ib_device *device);
  69
  70 static struct ib_client cma_client = {
  71         .name   = "cma",
  72         .add    = cma_add_one,
  73         .remove = cma_remove_one
  74 };
  75
  76 static struct ib_sa_client sa_client;
  77 static struct rdma_addr_client addr_client;
  78 static LIST_HEAD(dev_list);
  79 static LIST_HEAD(listen_any_list);
  80 static DEFINE_MUTEX(lock);
  81 static struct workqueue_struct *cma_wq;
  82 static DEFINE_IDR(sdp_ps);
  83 static DEFINE_IDR(tcp_ps);
  84 static DEFINE_IDR(udp_ps);
  85 static DEFINE_IDR(ipoib_ps);
  86 static DEFINE_IDR(ib_ps);
  87
  88 struct cma_device {
  89         struct list_head        list;
  90         struct ib_device        *device;
  91         struct completion       comp;
  92         atomic_t                refcount;
  93         struct list_head        id_list;
  94 };
  95
  96 struct rdma_bind_list {
  97         struct idr              *ps;
  98         struct hlist_head       owners;
  99         unsigned short          port;
 100 };
 101
 102 enum {
 103         CMA_OPTION_AFONLY,
 104 };
 105
 106 /*
 107  * Device removal can occur at anytime, so we need extra handling to
 108  * serialize notifying the user of device removal with other callbacks.
 109  * We do this by disabling removal notification while a callback is in process,
 110  * and reporting it after the callback completes.
 111  */
 112 struct rdma_id_private {
 113         struct rdma_cm_id       id;
 114
 115         struct rdma_bind_list   *bind_list;
 116         struct hlist_node       node;
 117         struct list_head        list; /* listen_any_list or cma_device.list */
 118         struct list_head        listen_list; /* per device listens */
 119         struct cma_device       *cma_dev;
 120         struct list_head        mc_list;
 121
 122         int                     internal_id;
 123         enum rdma_cm_state      state;
 124         spinlock_t              lock;
 125         struct mutex            qp_mutex;
 126
 127         struct completion       comp;
 128         atomic_t                refcount;
 129         struct mutex            handler_mutex;
 130
 131         int                     backlog;
 132         int                     timeout_ms;
 133         struct ib_sa_query      *query;
 134         int                     query_id;
 135         union {
 136                 struct ib_cm_id *ib;
 137                 struct iw_cm_id *iw;
 138         } cm_id;
 139
 140         u32                     seq_num;
 141         u32                     qkey;
 142         u32                     qp_num;
 143         pid_t                   owner;
 144         u32                     options;
 145         u8                      srq;
 146         u8                      tos;
 147         u8                      reuseaddr;
 148         u8                      afonly;
 149 };
 150
 151 struct cma_multicast {
 152         struct rdma_id_private *id_priv;
 153         union {
 154                 struct ib_sa_multicast *ib;
 155         } multicast;
 156         struct list_head        list;
 157         void                    *context;
 158         struct sockaddr_storage addr;
 159         struct kref             mcref;
 160 };
 161
 162 struct cma_work {
 163         struct work_struct      work;
 164         struct rdma_id_private  *id;
 165         enum rdma_cm_state      old_state;
 166         enum rdma_cm_state      new_state;
 167         struct rdma_cm_event    event;
 168 };
 169
 170 struct cma_ndev_work {
 171         struct work_struct      work;
 172         struct rdma_id_private  *id;
 173         struct rdma_cm_event    event;
 174 };
 175
 176 struct iboe_mcast_work {
 177         struct work_struct       work;
 178         struct rdma_id_private  *id;
 179         struct cma_multicast    *mc;
 180 };
 181
 182 union cma_ip_addr {
 183         struct in6_addr ip6;
 184         struct {
 185                 __be32 pad[3];
 186                 __be32 addr;
 187         } ip4;
 188 };
 189
 190 struct cma_hdr {
 191         u8 cma_version;
 192         u8 ip_version;  /* IP version: 7:4 */
 193         __be16 port;
 194         union cma_ip_addr src_addr;
 195         union cma_ip_addr dst_addr;
 196 };
 197
 198 struct sdp_hh {
 199         u8 bsdh[16];
 200         u8 sdp_version; /* Major version: 7:4 */
 201         u8 ip_version;  /* IP version: 7:4 */
 202         u8 sdp_specific1[10];
 203         __be16 port;
 204         __be16 sdp_specific2;
 205         union cma_ip_addr src_addr;
 206         union cma_ip_addr dst_addr;
 207 };
 208
 209 struct sdp_hah {
 210         u8 bsdh[16];
 211         u8 sdp_version;
 212 };
 213
 214 #define CMA_VERSION 0x00
 215 #define SDP_MAJ_VERSION 0x2
 216
 217 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 218 {
 219         unsigned long flags;
 220         int ret;
 221
 222         spin_lock_irqsave(&id_priv->lock, flags);
 223         ret = (id_priv->state == comp);
 224         spin_unlock_irqrestore(&id_priv->lock, flags);
 225         return ret;
 226 }
 227
 228 static int cma_comp_exch(struct rdma_id_private *id_priv,
 229                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 230 {
 231         unsigned long flags;
 232         int ret;
 233
 234         spin_lock_irqsave(&id_priv->lock, flags);
 235         if ((ret = (id_priv->state == comp)))
 236                 id_priv->state = exch;
 237         spin_unlock_irqrestore(&id_priv->lock, flags);
 238         return ret;
 239 }
 240
 241 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 242                                    enum rdma_cm_state exch)
 243 {
 244         unsigned long flags;
 245         enum rdma_cm_state old;
 246
 247         spin_lock_irqsave(&id_priv->lock, flags);
 248         old = id_priv->state;
 249         id_priv->state = exch;
 250         spin_unlock_irqrestore(&id_priv->lock, flags);
 251         return old;
 252 }
 253
 254 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 255 {
 256         return hdr->ip_version >> 4;
 257 }
 258
 259 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 260 {
 261         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 262 }
 263
 264 static inline u8 sdp_get_majv(u8 sdp_version)
 265 {
 266         return sdp_version >> 4;
 267 }
 268
 269 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 270 {
 271         return hh->ip_version >> 4;
 272 }
 273
 274 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 275 {
 276         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 277 }
 278
 279 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 280                               struct cma_device *cma_dev)
 281 {
 282         atomic_inc(&cma_dev->refcount);
 283         id_priv->cma_dev = cma_dev;
 284         id_priv->id.device = cma_dev->device;
 285         id_priv->id.route.addr.dev_addr.transport =
 286                 rdma_node_get_transport(cma_dev->device->node_type);
 287         list_add_tail(&id_priv->list, &cma_dev->id_list);
 288 }
 289
 290 static inline void cma_deref_dev(struct cma_device *cma_dev)
 291 {
 292         if (atomic_dec_and_test(&cma_dev->refcount))
 293                 complete(&cma_dev->comp);
 294 }
 295
 296 static inline void release_mc(struct kref *kref)
 297 {
 298         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 299
 300         kfree(mc->multicast.ib);
 301         kfree(mc);
 302 }
 303
 304 static void cma_release_dev(struct rdma_id_private *id_priv)
 305 {
 306         mutex_lock(&lock);
 307         list_del(&id_priv->list);
 308         cma_deref_dev(id_priv->cma_dev);
 309         id_priv->cma_dev = NULL;
 310         mutex_unlock(&lock);
 311 }
 312
 313 static int cma_set_qkey(struct rdma_id_private *id_priv)
 314 {
 315         struct ib_sa_mcmember_rec rec;
 316         int ret = 0;
 317
 318         if (id_priv->qkey)
 319                 return 0;
 320
 321         switch (id_priv->id.ps) {
 322         case RDMA_PS_UDP:
 323                 id_priv->qkey = RDMA_UDP_QKEY;
 324                 break;
 325         case RDMA_PS_IPOIB:
 326                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 327                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 328                                              id_priv->id.port_num, &rec.mgid,
 329                                              &rec);
 330                 if (!ret)
 331                         id_priv->qkey = be32_to_cpu(rec.qkey);
 332                 break;
 333         default:
 334                 break;
 335         }
 336         return ret;
 337 }
 338
 339 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 340 {
 341         int i;
 342         int err;
 343         struct ib_port_attr props;
 344         union ib_gid tmp;
 345
 346         err = ib_query_port(device, port_num, &props);
 347         if (err)
 348                 return err;
 349
 350         for (i = 0; i < props.gid_tbl_len; ++i) {
 351                 err = ib_query_gid(device, port_num, i, &tmp);
 352                 if (err)
 353                         return err;
 354                 if (!memcmp(&tmp, gid, sizeof tmp))
 355                         return 0;
 356         }
 357
 358         return -EADDRNOTAVAIL;
 359 }
 360
 361 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 362 {
 363         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 364         struct cma_device *cma_dev;
 365         union ib_gid gid, iboe_gid;
 366         int ret = -ENODEV;
 367         u8 port;
 368         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 369                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 370
 371         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 372             id_priv->id.ps == RDMA_PS_IPOIB)
 373                 return -EINVAL;
 374
 375         mutex_lock(&lock);
 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                                 }
 392                         }
 393                 }
 394         }
 395
 396 out:
 397         if (!ret)
 398                 cma_attach_to_dev(id_priv, cma_dev);
 399
 400         mutex_unlock(&lock);
 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 rdma_cm_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 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 422                                   void *context, enum rdma_port_space ps,
 423                                   enum ib_qp_type qp_type)
 424 {
 425         struct rdma_id_private *id_priv;
 426
 427         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 428         if (!id_priv)
 429                 return ERR_PTR(-ENOMEM);
 430
 431         id_priv->owner = task_pid_nr(current);
 432         id_priv->state = RDMA_CM_IDLE;
 433         id_priv->id.context = context;
 434         id_priv->id.event_handler = event_handler;
 435         id_priv->id.ps = ps;
 436         id_priv->id.qp_type = qp_type;
 437         spin_lock_init(&id_priv->lock);
 438         mutex_init(&id_priv->qp_mutex);
 439         init_completion(&id_priv->comp);
 440         atomic_set(&id_priv->refcount, 1);
 441         mutex_init(&id_priv->handler_mutex);
 442         INIT_LIST_HEAD(&id_priv->listen_list);
 443         INIT_LIST_HEAD(&id_priv->mc_list);
 444         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 445
 446         return &id_priv->id;
 447 }
 448 EXPORT_SYMBOL(rdma_create_id);
 449
 450 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 451 {
 452         struct ib_qp_attr qp_attr;
 453         int qp_attr_mask, ret;
 454
 455         qp_attr.qp_state = IB_QPS_INIT;
 456         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 457         if (ret)
 458                 return ret;
 459
 460         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 461         if (ret)
 462                 return ret;
 463
 464         qp_attr.qp_state = IB_QPS_RTR;
 465         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 466         if (ret)
 467                 return ret;
 468
 469         qp_attr.qp_state = IB_QPS_RTS;
 470         qp_attr.sq_psn = 0;
 471         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 472
 473         return ret;
 474 }
 475
 476 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 477 {
 478         struct ib_qp_attr qp_attr;
 479         int qp_attr_mask, ret;
 480
 481         qp_attr.qp_state = IB_QPS_INIT;
 482         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 483         if (ret)
 484                 return ret;
 485
 486         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 487 }
 488
 489 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 490                    struct ib_qp_init_attr *qp_init_attr)
 491 {
 492         struct rdma_id_private *id_priv;
 493         struct ib_qp *qp;
 494         int ret;
 495
 496         id_priv = container_of(id, struct rdma_id_private, id);
 497         if (id->device != pd->device)
 498                 return -EINVAL;
 499
 500         qp = ib_create_qp(pd, qp_init_attr);
 501         if (IS_ERR(qp))
 502                 return PTR_ERR(qp);
 503
 504         if (id->qp_type == IB_QPT_UD)
 505                 ret = cma_init_ud_qp(id_priv, qp);
 506         else
 507                 ret = cma_init_conn_qp(id_priv, qp);
 508         if (ret)
 509                 goto err;
 510
 511         id->qp = qp;
 512         id_priv->qp_num = qp->qp_num;
 513         id_priv->srq = (qp->srq != NULL);
 514         return 0;
 515 err:
 516         ib_destroy_qp(qp);
 517         return ret;
 518 }
 519 EXPORT_SYMBOL(rdma_create_qp);
 520
 521 void rdma_destroy_qp(struct rdma_cm_id *id)
 522 {
 523         struct rdma_id_private *id_priv;
 524
 525         id_priv = container_of(id, struct rdma_id_private, id);
 526         mutex_lock(&id_priv->qp_mutex);
 527         ib_destroy_qp(id_priv->id.qp);
 528         id_priv->id.qp = NULL;
 529         mutex_unlock(&id_priv->qp_mutex);
 530 }
 531 EXPORT_SYMBOL(rdma_destroy_qp);
 532
 533 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 534                              struct rdma_conn_param *conn_param)
 535 {
 536         struct ib_qp_attr qp_attr;
 537         int qp_attr_mask, ret;
 538
 539         mutex_lock(&id_priv->qp_mutex);
 540         if (!id_priv->id.qp) {
 541                 ret = 0;
 542                 goto out;
 543         }
 544
 545         /* Need to update QP attributes from default values. */
 546         qp_attr.qp_state = IB_QPS_INIT;
 547         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 548         if (ret)
 549                 goto out;
 550
 551         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 552         if (ret)
 553                 goto out;
 554
 555         qp_attr.qp_state = IB_QPS_RTR;
 556         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 557         if (ret)
 558                 goto out;
 559
 560         if (conn_param)
 561                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 562         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 563 out:
 564         mutex_unlock(&id_priv->qp_mutex);
 565         return ret;
 566 }
 567
 568 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 569                              struct rdma_conn_param *conn_param)
 570 {
 571         struct ib_qp_attr qp_attr;
 572         int qp_attr_mask, ret;
 573
 574         mutex_lock(&id_priv->qp_mutex);
 575         if (!id_priv->id.qp) {
 576                 ret = 0;
 577                 goto out;
 578         }
 579
 580         qp_attr.qp_state = IB_QPS_RTS;
 581         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 582         if (ret)
 583                 goto out;
 584
 585         if (conn_param)
 586                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 587         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 588 out:
 589         mutex_unlock(&id_priv->qp_mutex);
 590         return ret;
 591 }
 592
 593 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 594 {
 595         struct ib_qp_attr qp_attr;
 596         int ret;
 597
 598         mutex_lock(&id_priv->qp_mutex);
 599         if (!id_priv->id.qp) {
 600                 ret = 0;
 601                 goto out;
 602         }
 603
 604         qp_attr.qp_state = IB_QPS_ERR;
 605         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 606 out:
 607         mutex_unlock(&id_priv->qp_mutex);
 608         return ret;
 609 }
 610
 611 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 612                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 613 {
 614         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 615         int ret;
 616         u16 pkey;
 617
 618         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 619             IB_LINK_LAYER_INFINIBAND)
 620                 pkey = ib_addr_get_pkey(dev_addr);
 621         else
 622                 pkey = 0xffff;
 623
 624         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 625                                   pkey, &qp_attr->pkey_index);
 626         if (ret)
 627                 return ret;
 628
 629         qp_attr->port_num = id_priv->id.port_num;
 630         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 631
 632         if (id_priv->id.qp_type == IB_QPT_UD) {
 633                 ret = cma_set_qkey(id_priv);
 634                 if (ret)
 635                         return ret;
 636
 637                 qp_attr->qkey = id_priv->qkey;
 638                 *qp_attr_mask |= IB_QP_QKEY;
 639         } else {
 640                 qp_attr->qp_access_flags = 0;
 641                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 642         }
 643         return 0;
 644 }
 645
 646 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 647                        int *qp_attr_mask)
 648 {
 649         struct rdma_id_private *id_priv;
 650         int ret = 0;
 651
 652         id_priv = container_of(id, struct rdma_id_private, id);
 653         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 654         case RDMA_TRANSPORT_IB:
 655                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 656                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 657                 else
 658                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 659                                                  qp_attr_mask);
 660                 if (qp_attr->qp_state == IB_QPS_RTR)
 661                         qp_attr->rq_psn = id_priv->seq_num;
 662                 break;
 663         case RDMA_TRANSPORT_IWARP:
 664                 if (!id_priv->cm_id.iw) {
 665                         qp_attr->qp_access_flags = 0;
 666                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 667                 } else
 668                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 669                                                  qp_attr_mask);
 670                 break;
 671         default:
 672                 ret = -ENOSYS;
 673                 break;
 674         }
 675
 676         return ret;
 677 }
 678 EXPORT_SYMBOL(rdma_init_qp_attr);
 679
 680 static inline int cma_zero_addr(struct sockaddr *addr)
 681 {
 682         struct in6_addr *ip6;
 683
 684         if (addr->sa_family == AF_INET)
 685                 return ipv4_is_zeronet(
 686                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
 687         else {
 688                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 689                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 690                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 691         }
 692 }
 693
 694 static inline int cma_loopback_addr(struct sockaddr *addr)
 695 {
 696         if (addr->sa_family == AF_INET)
 697                 return ipv4_is_loopback(
 698                         ((struct sockaddr_in *) addr)->sin_addr.s_addr);
 699         else
 700                 return ipv6_addr_loopback(
 701                         &((struct sockaddr_in6 *) addr)->sin6_addr);
 702 }
 703
 704 static inline int cma_any_addr(struct sockaddr *addr)
 705 {
 706         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 707 }
 708
 709 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 710 {
 711         if (src->sa_family != dst->sa_family)
 712                 return -1;
 713
 714         switch (src->sa_family) {
 715         case AF_INET:
 716                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 717                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 718         default:
 719                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 720                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 721         }
 722 }
 723
 724 static inline __be16 cma_port(struct sockaddr *addr)
 725 {
 726         if (addr->sa_family == AF_INET)
 727                 return ((struct sockaddr_in *) addr)->sin_port;
 728         else
 729                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 730 }
 731
 732 static inline int cma_any_port(struct sockaddr *addr)
 733 {
 734         return !cma_port(addr);
 735 }
 736
 737 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 738                             u8 *ip_ver, __be16 *port,
 739                             union cma_ip_addr **src, union cma_ip_addr **dst)
 740 {
 741         switch (ps) {
 742         case RDMA_PS_SDP:
 743                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 744                     SDP_MAJ_VERSION)
 745                         return -EINVAL;
 746
 747                 *ip_ver = sdp_get_ip_ver(hdr);
 748                 *port   = ((struct sdp_hh *) hdr)->port;
 749                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 750                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 751                 break;
 752         default:
 753                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 754                         return -EINVAL;
 755
 756                 *ip_ver = cma_get_ip_ver(hdr);
 757                 *port   = ((struct cma_hdr *) hdr)->port;
 758                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 759                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 760                 break;
 761         }
 762
 763         if (*ip_ver != 4 && *ip_ver != 6)
 764                 return -EINVAL;
 765         return 0;
 766 }
 767
 768 static void cma_save_net_info(struct rdma_addr *addr,
 769                               struct rdma_addr *listen_addr,
 770                               u8 ip_ver, __be16 port,
 771                               union cma_ip_addr *src, union cma_ip_addr *dst)
 772 {
 773         struct sockaddr_in *listen4, *ip4;
 774         struct sockaddr_in6 *listen6, *ip6;
 775
 776         switch (ip_ver) {
 777         case 4:
 778                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 779                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 780                 ip4->sin_family = listen4->sin_family;
 781                 ip4->sin_addr.s_addr = dst->ip4.addr;
 782                 ip4->sin_port = listen4->sin_port;
 783
 784                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 785                 ip4->sin_family = listen4->sin_family;
 786                 ip4->sin_addr.s_addr = src->ip4.addr;
 787                 ip4->sin_port = port;
 788                 break;
 789         case 6:
 790                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 791                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 792                 ip6->sin6_family = listen6->sin6_family;
 793                 ip6->sin6_addr = dst->ip6;
 794                 ip6->sin6_port = listen6->sin6_port;
 795
 796                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 797                 ip6->sin6_family = listen6->sin6_family;
 798                 ip6->sin6_addr = src->ip6;
 799                 ip6->sin6_port = port;
 800                 break;
 801         default:
 802                 break;
 803         }
 804 }
 805
 806 static inline int cma_user_data_offset(enum rdma_port_space ps)
 807 {
 808         switch (ps) {
 809         case RDMA_PS_SDP:
 810                 return 0;
 811         default:
 812                 return sizeof(struct cma_hdr);
 813         }
 814 }
 815
 816 static void cma_cancel_route(struct rdma_id_private *id_priv)
 817 {
 818         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 819         case IB_LINK_LAYER_INFINIBAND:
 820                 if (id_priv->query)
 821                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 822                 break;
 823         default:
 824                 break;
 825         }
 826 }
 827
 828 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 829 {
 830         struct rdma_id_private *dev_id_priv;
 831
 832         /*
 833          * Remove from listen_any_list to prevent added devices from spawning
 834          * additional listen requests.
 835          */
 836         mutex_lock(&lock);
 837         list_del(&id_priv->list);
 838
 839         while (!list_empty(&id_priv->listen_list)) {
 840                 dev_id_priv = list_entry(id_priv->listen_list.next,
 841                                          struct rdma_id_private, listen_list);
 842                 /* sync with device removal to avoid duplicate destruction */
 843                 list_del_init(&dev_id_priv->list);
 844                 list_del(&dev_id_priv->listen_list);
 845                 mutex_unlock(&lock);
 846
 847                 rdma_destroy_id(&dev_id_priv->id);
 848                 mutex_lock(&lock);
 849         }
 850         mutex_unlock(&lock);
 851 }
 852
 853 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 854                                  enum rdma_cm_state state)
 855 {
 856         switch (state) {
 857         case RDMA_CM_ADDR_QUERY:
 858                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 859                 break;
 860         case RDMA_CM_ROUTE_QUERY:
 861                 cma_cancel_route(id_priv);
 862                 break;
 863         case RDMA_CM_LISTEN:
 864                 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
 865                                 && !id_priv->cma_dev)
 866                         cma_cancel_listens(id_priv);
 867                 break;
 868         default:
 869                 break;
 870         }
 871 }
 872
 873 static void cma_release_port(struct rdma_id_private *id_priv)
 874 {
 875         struct rdma_bind_list *bind_list = id_priv->bind_list;
 876
 877         if (!bind_list)
 878                 return;
 879
 880         mutex_lock(&lock);
 881         hlist_del(&id_priv->node);
 882         if (hlist_empty(&bind_list->owners)) {
 883                 idr_remove(bind_list->ps, bind_list->port);
 884                 kfree(bind_list);
 885         }
 886         mutex_unlock(&lock);
 887 }
 888
 889 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 890 {
 891         struct cma_multicast *mc;
 892
 893         while (!list_empty(&id_priv->mc_list)) {
 894                 mc = container_of(id_priv->mc_list.next,
 895                                   struct cma_multicast, list);
 896                 list_del(&mc->list);
 897                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 898                 case IB_LINK_LAYER_INFINIBAND:
 899                         ib_sa_free_multicast(mc->multicast.ib);
 900                         kfree(mc);
 901                         break;
 902                 case IB_LINK_LAYER_ETHERNET:
 903                         kref_put(&mc->mcref, release_mc);
 904                         break;
 905                 default:
 906                         break;
 907                 }
 908         }
 909 }
 910
 911 void rdma_destroy_id(struct rdma_cm_id *id)
 912 {
 913         struct rdma_id_private *id_priv;
 914         enum rdma_cm_state state;
 915
 916         id_priv = container_of(id, struct rdma_id_private, id);
 917         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 918         cma_cancel_operation(id_priv, state);
 919
 920         /*
 921          * Wait for any active callback to finish.  New callbacks will find
 922          * the id_priv state set to destroying and abort.
 923          */
 924         mutex_lock(&id_priv->handler_mutex);
 925         mutex_unlock(&id_priv->handler_mutex);
 926
 927         if (id_priv->cma_dev) {
 928                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 929                 case RDMA_TRANSPORT_IB:
 930                         if (id_priv->cm_id.ib)
 931                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 932                         break;
 933                 case RDMA_TRANSPORT_IWARP:
 934                         if (id_priv->cm_id.iw)
 935                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 936                         break;
 937                 default:
 938                         break;
 939                 }
 940                 cma_leave_mc_groups(id_priv);
 941                 cma_release_dev(id_priv);
 942         }
 943
 944         cma_release_port(id_priv);
 945         cma_deref_id(id_priv);
 946         wait_for_completion(&id_priv->comp);
 947
 948         if (id_priv->internal_id)
 949                 cma_deref_id(id_priv->id.context);
 950
 951         kfree(id_priv->id.route.path_rec);
 952         kfree(id_priv);
 953 }
 954 EXPORT_SYMBOL(rdma_destroy_id);
 955
 956 static int cma_rep_recv(struct rdma_id_private *id_priv)
 957 {
 958         int ret;
 959
 960         ret = cma_modify_qp_rtr(id_priv, NULL);
 961         if (ret)
 962                 goto reject;
 963
 964         ret = cma_modify_qp_rts(id_priv, NULL);
 965         if (ret)
 966                 goto reject;
 967
 968         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 969         if (ret)
 970                 goto reject;
 971
 972         return 0;
 973 reject:
 974         cma_modify_qp_err(id_priv);
 975         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 976                        NULL, 0, NULL, 0);
 977         return ret;
 978 }
 979
 980 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 981 {
 982         if (id_priv->id.ps == RDMA_PS_SDP &&
 983             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 984             SDP_MAJ_VERSION)
 985                 return -EINVAL;
 986
 987         return 0;
 988 }
 989
 990 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 991                                    struct ib_cm_rep_event_param *rep_data,
 992                                    void *private_data)
 993 {
 994         event->param.conn.private_data = private_data;
 995         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 996         event->param.conn.responder_resources = rep_data->responder_resources;
 997         event->param.conn.initiator_depth = rep_data->initiator_depth;
 998         event->param.conn.flow_control = rep_data->flow_control;
 999         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1000         event->param.conn.srq = rep_data->srq;
1001         event->param.conn.qp_num = rep_data->remote_qpn;
1002 }
1003
1004 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1005 {
1006         struct rdma_id_private *id_priv = cm_id->context;
1007         struct rdma_cm_event event;
1008         int ret = 0;
1009
1010         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1011                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1012             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1013                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1014                 return 0;
1015
1016         memset(&event, 0, sizeof event);
1017         switch (ib_event->event) {
1018         case IB_CM_REQ_ERROR:
1019         case IB_CM_REP_ERROR:
1020                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1021                 event.status = -ETIMEDOUT;
1022                 break;
1023         case IB_CM_REP_RECEIVED:
1024                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1025                 if (event.status)
1026                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1027                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1028                         event.status = cma_rep_recv(id_priv);
1029                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1030                                                      RDMA_CM_EVENT_ESTABLISHED;
1031                 } else
1032                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1033                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1034                                        ib_event->private_data);
1035                 break;
1036         case IB_CM_RTU_RECEIVED:
1037         case IB_CM_USER_ESTABLISHED:
1038                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1039                 break;
1040         case IB_CM_DREQ_ERROR:
1041                 event.status = -ETIMEDOUT; /* fall through */
1042         case IB_CM_DREQ_RECEIVED:
1043         case IB_CM_DREP_RECEIVED:
1044                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1045                                    RDMA_CM_DISCONNECT))
1046                         goto out;
1047                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1048                 break;
1049         case IB_CM_TIMEWAIT_EXIT:
1050                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1051                 break;
1052         case IB_CM_MRA_RECEIVED:
1053                 /* ignore event */
1054                 goto out;
1055         case IB_CM_REJ_RECEIVED:
1056                 cma_modify_qp_err(id_priv);
1057                 event.status = ib_event->param.rej_rcvd.reason;
1058                 event.event = RDMA_CM_EVENT_REJECTED;
1059                 event.param.conn.private_data = ib_event->private_data;
1060                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1061                 break;
1062         default:
1063                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1064                        ib_event->event);
1065                 goto out;
1066         }
1067
1068         ret = id_priv->id.event_handler(&id_priv->id, &event);
1069         if (ret) {
1070                 /* Destroy the CM ID by returning a non-zero value. */
1071                 id_priv->cm_id.ib = NULL;
1072                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1073                 mutex_unlock(&id_priv->handler_mutex);
1074                 rdma_destroy_id(&id_priv->id);
1075                 return ret;
1076         }
1077 out:
1078         mutex_unlock(&id_priv->handler_mutex);
1079         return ret;
1080 }
1081
1082 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1083                                                struct ib_cm_event *ib_event)
1084 {
1085         struct rdma_id_private *id_priv;
1086         struct rdma_cm_id *id;
1087         struct rdma_route *rt;
1088         union cma_ip_addr *src, *dst;
1089         __be16 port;
1090         u8 ip_ver;
1091         int ret;
1092
1093         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1094                              &ip_ver, &port, &src, &dst))
1095                 return NULL;
1096
1097         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1098                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1099         if (IS_ERR(id))
1100                 return NULL;
1101
1102         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1103                           ip_ver, port, src, dst);
1104
1105         rt = &id->route;
1106         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1107         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1108                                GFP_KERNEL);
1109         if (!rt->path_rec)
1110                 goto err;
1111
1112         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1113         if (rt->num_paths == 2)
1114                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1115
1116         if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1117                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1118                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1119                 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1120         } else {
1121                 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1122                                         &rt->addr.dev_addr);
1123                 if (ret)
1124                         goto err;
1125         }
1126         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1127
1128         id_priv = container_of(id, struct rdma_id_private, id);
1129         id_priv->state = RDMA_CM_CONNECT;
1130         return id_priv;
1131
1132 err:
1133         rdma_destroy_id(id);
1134         return NULL;
1135 }
1136
1137 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1138                                               struct ib_cm_event *ib_event)
1139 {
1140         struct rdma_id_private *id_priv;
1141         struct rdma_cm_id *id;
1142         union cma_ip_addr *src, *dst;
1143         __be16 port;
1144         u8 ip_ver;
1145         int ret;
1146
1147         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1148                             listen_id->ps, IB_QPT_UD);
1149         if (IS_ERR(id))
1150                 return NULL;
1151
1152
1153         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1154                              &ip_ver, &port, &src, &dst))
1155                 goto err;
1156
1157         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1158                           ip_ver, port, src, dst);
1159
1160         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1161                 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1162                                         &id->route.addr.dev_addr);
1163                 if (ret)
1164                         goto err;
1165         }
1166
1167         id_priv = container_of(id, struct rdma_id_private, id);
1168         id_priv->state = RDMA_CM_CONNECT;
1169         return id_priv;
1170 err:
1171         rdma_destroy_id(id);
1172         return NULL;
1173 }
1174
1175 static void cma_set_req_event_data(struct rdma_cm_event *event,
1176                                    struct ib_cm_req_event_param *req_data,
1177                                    void *private_data, int offset)
1178 {
1179         event->param.conn.private_data = private_data + offset;
1180         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1181         event->param.conn.responder_resources = req_data->responder_resources;
1182         event->param.conn.initiator_depth = req_data->initiator_depth;
1183         event->param.conn.flow_control = req_data->flow_control;
1184         event->param.conn.retry_count = req_data->retry_count;
1185         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1186         event->param.conn.srq = req_data->srq;
1187         event->param.conn.qp_num = req_data->remote_qpn;
1188 }
1189
1190 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1191 {
1192         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1193                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1194                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1195                  (id->qp_type == IB_QPT_UD)) ||
1196                 (!id->qp_type));
1197 }
1198
1199 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1200 {
1201         struct rdma_id_private *listen_id, *conn_id;
1202         struct rdma_cm_event event;
1203         int offset, ret;
1204
1205         listen_id = cm_id->context;
1206         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1207                 return -EINVAL;
1208
1209         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1210                 return -ECONNABORTED;
1211
1212         memset(&event, 0, sizeof event);
1213         offset = cma_user_data_offset(listen_id->id.ps);
1214         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1215         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1216                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1217                 event.param.ud.private_data = ib_event->private_data + offset;
1218                 event.param.ud.private_data_len =
1219                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1220         } else {
1221                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1222                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1223                                        ib_event->private_data, offset);
1224         }
1225         if (!conn_id) {
1226                 ret = -ENOMEM;
1227                 goto err1;
1228         }
1229
1230         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1231         ret = cma_acquire_dev(conn_id);
1232         if (ret)
1233                 goto err2;
1234
1235         conn_id->cm_id.ib = cm_id;
1236         cm_id->context = conn_id;
1237         cm_id->cm_handler = cma_ib_handler;
1238
1239         /*
1240          * Protect against the user destroying conn_id from another thread
1241          * until we're done accessing it.
1242          */
1243         atomic_inc(&conn_id->refcount);
1244         ret = conn_id->id.event_handler(&conn_id->id, &event);
1245         if (ret)
1246                 goto err3;
1247
1248         /*
1249          * Acquire mutex to prevent user executing rdma_destroy_id()
1250          * while we're accessing the cm_id.
1251          */
1252         mutex_lock(&lock);
1253         if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1254                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1255         mutex_unlock(&lock);
1256         mutex_unlock(&conn_id->handler_mutex);
1257         mutex_unlock(&listen_id->handler_mutex);
1258         cma_deref_id(conn_id);
1259         return 0;
1260
1261 err3:
1262         cma_deref_id(conn_id);
1263         /* Destroy the CM ID by returning a non-zero value. */
1264         conn_id->cm_id.ib = NULL;
1265 err2:
1266         cma_exch(conn_id, RDMA_CM_DESTROYING);
1267         mutex_unlock(&conn_id->handler_mutex);
1268 err1:
1269         mutex_unlock(&listen_id->handler_mutex);
1270         if (conn_id)
1271                 rdma_destroy_id(&conn_id->id);
1272         return ret;
1273 }
1274
1275 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1276 {
1277         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1278 }
1279
1280 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1281                                  struct ib_cm_compare_data *compare)
1282 {
1283         struct cma_hdr *cma_data, *cma_mask;
1284         struct sdp_hh *sdp_data, *sdp_mask;
1285         __be32 ip4_addr;
1286         struct in6_addr ip6_addr;
1287
1288         memset(compare, 0, sizeof *compare);
1289         cma_data = (void *) compare->data;
1290         cma_mask = (void *) compare->mask;
1291         sdp_data = (void *) compare->data;
1292         sdp_mask = (void *) compare->mask;
1293
1294         switch (addr->sa_family) {
1295         case AF_INET:
1296                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1297                 if (ps == RDMA_PS_SDP) {
1298                         sdp_set_ip_ver(sdp_data, 4);
1299                         sdp_set_ip_ver(sdp_mask, 0xF);
1300                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1301                         sdp_mask->dst_addr.ip4.addr = htonl(~0);
1302                 } else {
1303                         cma_set_ip_ver(cma_data, 4);
1304                         cma_set_ip_ver(cma_mask, 0xF);
1305                         if (!cma_any_addr(addr)) {
1306                                 cma_data->dst_addr.ip4.addr = ip4_addr;
1307                                 cma_mask->dst_addr.ip4.addr = htonl(~0);
1308                         }
1309                 }
1310                 break;
1311         case AF_INET6:
1312                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1313                 if (ps == RDMA_PS_SDP) {
1314                         sdp_set_ip_ver(sdp_data, 6);
1315                         sdp_set_ip_ver(sdp_mask, 0xF);
1316                         sdp_data->dst_addr.ip6 = ip6_addr;
1317                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1318                                sizeof sdp_mask->dst_addr.ip6);
1319                 } else {
1320                         cma_set_ip_ver(cma_data, 6);
1321                         cma_set_ip_ver(cma_mask, 0xF);
1322                         if (!cma_any_addr(addr)) {
1323                                 cma_data->dst_addr.ip6 = ip6_addr;
1324                                 memset(&cma_mask->dst_addr.ip6, 0xFF,
1325                                        sizeof cma_mask->dst_addr.ip6);
1326                         }
1327                 }
1328                 break;
1329         default:
1330                 break;
1331         }
1332 }
1333
1334 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1335 {
1336         struct rdma_id_private *id_priv = iw_id->context;
1337         struct rdma_cm_event event;
1338         struct sockaddr_in *sin;
1339         int ret = 0;
1340
1341         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1342                 return 0;
1343
1344         memset(&event, 0, sizeof event);
1345         switch (iw_event->event) {
1346         case IW_CM_EVENT_CLOSE:
1347                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1348                 break;
1349         case IW_CM_EVENT_CONNECT_REPLY:
1350                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1351                 *sin = iw_event->local_addr;
1352                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1353                 *sin = iw_event->remote_addr;
1354                 switch (iw_event->status) {
1355                 case 0:
1356                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1357                         event.param.conn.initiator_depth = iw_event->ird;
1358                         event.param.conn.responder_resources = iw_event->ord;
1359                         break;
1360                 case -ECONNRESET:
1361                 case -ECONNREFUSED:
1362                         event.event = RDMA_CM_EVENT_REJECTED;
1363                         break;
1364                 case -ETIMEDOUT:
1365                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1366                         break;
1367                 default:
1368                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1369                         break;
1370                 }
1371                 break;
1372         case IW_CM_EVENT_ESTABLISHED:
1373                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1374                 event.param.conn.initiator_depth = iw_event->ird;
1375                 event.param.conn.responder_resources = iw_event->ord;
1376                 break;
1377         default:
1378                 BUG_ON(1);
1379         }
1380
1381         event.status = iw_event->status;
1382         event.param.conn.private_data = iw_event->private_data;
1383         event.param.conn.private_data_len = iw_event->private_data_len;
1384         ret = id_priv->id.event_handler(&id_priv->id, &event);
1385         if (ret) {
1386                 /* Destroy the CM ID by returning a non-zero value. */
1387                 id_priv->cm_id.iw = NULL;
1388                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1389                 mutex_unlock(&id_priv->handler_mutex);
1390                 rdma_destroy_id(&id_priv->id);
1391                 return ret;
1392         }
1393
1394         mutex_unlock(&id_priv->handler_mutex);
1395         return ret;
1396 }
1397
1398 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1399                                struct iw_cm_event *iw_event)
1400 {
1401         struct rdma_cm_id *new_cm_id;
1402         struct rdma_id_private *listen_id, *conn_id;
1403         struct sockaddr_in *sin;
1404         struct net_device *dev = NULL;
1405         struct rdma_cm_event event;
1406         int ret;
1407         struct ib_device_attr attr;
1408
1409         listen_id = cm_id->context;
1410         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1411                 return -ECONNABORTED;
1412
1413         /* Create a new RDMA id for the new IW CM ID */
1414         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1415                                    listen_id->id.context,
1416                                    RDMA_PS_TCP, IB_QPT_RC);
1417         if (IS_ERR(new_cm_id)) {
1418                 ret = -ENOMEM;
1419                 goto out;
1420         }
1421         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1422         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1423         conn_id->state = RDMA_CM_CONNECT;
1424
1425         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1426         if (!dev) {
1427                 ret = -EADDRNOTAVAIL;
1428                 mutex_unlock(&conn_id->handler_mutex);
1429                 rdma_destroy_id(new_cm_id);
1430                 goto out;
1431         }
1432         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1433         if (ret) {
1434                 mutex_unlock(&conn_id->handler_mutex);
1435                 rdma_destroy_id(new_cm_id);
1436                 goto out;
1437         }
1438
1439         ret = cma_acquire_dev(conn_id);
1440         if (ret) {
1441                 mutex_unlock(&conn_id->handler_mutex);
1442                 rdma_destroy_id(new_cm_id);
1443                 goto out;
1444         }
1445
1446         conn_id->cm_id.iw = cm_id;
1447         cm_id->context = conn_id;
1448         cm_id->cm_handler = cma_iw_handler;
1449
1450         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1451         *sin = iw_event->local_addr;
1452         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1453         *sin = iw_event->remote_addr;
1454
1455         ret = ib_query_device(conn_id->id.device, &attr);
1456         if (ret) {
1457                 mutex_unlock(&conn_id->handler_mutex);
1458                 rdma_destroy_id(new_cm_id);
1459                 goto out;
1460         }
1461
1462         memset(&event, 0, sizeof event);
1463         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1464         event.param.conn.private_data = iw_event->private_data;
1465         event.param.conn.private_data_len = iw_event->private_data_len;
1466         event.param.conn.initiator_depth = iw_event->ird;
1467         event.param.conn.responder_resources = iw_event->ord;
1468
1469         /*
1470          * Protect against the user destroying conn_id from another thread
1471          * until we're done accessing it.
1472          */
1473         atomic_inc(&conn_id->refcount);
1474         ret = conn_id->id.event_handler(&conn_id->id, &event);
1475         if (ret) {
1476                 /* User wants to destroy the CM ID */
1477                 conn_id->cm_id.iw = NULL;
1478                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1479                 mutex_unlock(&conn_id->handler_mutex);
1480                 cma_deref_id(conn_id);
1481                 rdma_destroy_id(&conn_id->id);
1482                 goto out;
1483         }
1484
1485         mutex_unlock(&conn_id->handler_mutex);
1486         cma_deref_id(conn_id);
1487
1488 out:
1489         if (dev)
1490                 dev_put(dev);
1491         mutex_unlock(&listen_id->handler_mutex);
1492         return ret;
1493 }
1494
1495 static int cma_ib_listen(struct rdma_id_private *id_priv)
1496 {
1497         struct ib_cm_compare_data compare_data;
1498         struct sockaddr *addr;
1499         struct ib_cm_id *id;
1500         __be64 svc_id;
1501         int ret;
1502
1503         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1504         if (IS_ERR(id))
1505                 return PTR_ERR(id);
1506
1507         id_priv->cm_id.ib = id;
1508
1509         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1510         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1511         if (cma_any_addr(addr) && !id_priv->afonly)
1512                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1513         else {
1514                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1515                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1516         }
1517
1518         if (ret) {
1519                 ib_destroy_cm_id(id_priv->cm_id.ib);
1520                 id_priv->cm_id.ib = NULL;
1521         }
1522
1523         return ret;
1524 }
1525
1526 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1527 {
1528         int ret;
1529         struct sockaddr_in *sin;
1530         struct iw_cm_id *id;
1531
1532         id = iw_create_cm_id(id_priv->id.device,
1533                              iw_conn_req_handler,
1534                              id_priv);
1535         if (IS_ERR(id))
1536                 return PTR_ERR(id);
1537
1538         id_priv->cm_id.iw = id;
1539
1540         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1541         id_priv->cm_id.iw->local_addr = *sin;
1542
1543         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1544
1545         if (ret) {
1546                 iw_destroy_cm_id(id_priv->cm_id.iw);
1547                 id_priv->cm_id.iw = NULL;
1548         }
1549
1550         return ret;
1551 }
1552
1553 static int cma_listen_handler(struct rdma_cm_id *id,
1554                               struct rdma_cm_event *event)
1555 {
1556         struct rdma_id_private *id_priv = id->context;
1557
1558         id->context = id_priv->id.context;
1559         id->event_handler = id_priv->id.event_handler;
1560         return id_priv->id.event_handler(id, event);
1561 }
1562
1563 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1564                               struct cma_device *cma_dev)
1565 {
1566         struct rdma_id_private *dev_id_priv;
1567         struct rdma_cm_id *id;
1568         int ret;
1569
1570         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1571                             id_priv->id.qp_type);
1572         if (IS_ERR(id))
1573                 return;
1574
1575         dev_id_priv = container_of(id, struct rdma_id_private, id);
1576
1577         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1578         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1579                ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1580
1581         cma_attach_to_dev(dev_id_priv, cma_dev);
1582         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1583         atomic_inc(&id_priv->refcount);
1584         dev_id_priv->internal_id = 1;
1585         dev_id_priv->afonly = id_priv->afonly;
1586
1587         ret = rdma_listen(id, id_priv->backlog);
1588         if (ret)
1589                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1590                        "listening on device %s\n", ret, cma_dev->device->name);
1591 }
1592
1593 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1594 {
1595         struct cma_device *cma_dev;
1596
1597         mutex_lock(&lock);
1598         list_add_tail(&id_priv->list, &listen_any_list);
1599         list_for_each_entry(cma_dev, &dev_list, list)
1600                 cma_listen_on_dev(id_priv, cma_dev);
1601         mutex_unlock(&lock);
1602 }
1603
1604 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1605 {
1606         struct rdma_id_private *id_priv;
1607
1608         id_priv = container_of(id, struct rdma_id_private, id);
1609         id_priv->tos = (u8) tos;
1610 }
1611 EXPORT_SYMBOL(rdma_set_service_type);
1612
1613 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1614                               void *context)
1615 {
1616         struct cma_work *work = context;
1617         struct rdma_route *route;
1618
1619         route = &work->id->id.route;
1620
1621         if (!status) {
1622                 route->num_paths = 1;
1623                 *route->path_rec = *path_rec;
1624         } else {
1625                 work->old_state = RDMA_CM_ROUTE_QUERY;
1626                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1627                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1628                 work->event.status = status;
1629         }
1630
1631         queue_work(cma_wq, &work->work);
1632 }
1633
1634 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1635                               struct cma_work *work)
1636 {
1637         struct rdma_addr *addr = &id_priv->id.route.addr;
1638         struct ib_sa_path_rec path_rec;
1639         ib_sa_comp_mask comp_mask;
1640         struct sockaddr_in6 *sin6;
1641
1642         memset(&path_rec, 0, sizeof path_rec);
1643         rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1644         rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1645         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1646         path_rec.numb_path = 1;
1647         path_rec.reversible = 1;
1648         path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1649                                                         (struct sockaddr *) &addr->dst_addr);
1650
1651         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1652                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1653                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1654
1655         if (addr->src_addr.ss_family == AF_INET) {
1656                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1657                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1658         } else {
1659                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1660                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1661                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1662         }
1663
1664         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1665                                                id_priv->id.port_num, &path_rec,
1666                                                comp_mask, timeout_ms,
1667                                                GFP_KERNEL, cma_query_handler,
1668                                                work, &id_priv->query);
1669
1670         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1671 }
1672
1673 static void cma_work_handler(struct work_struct *_work)
1674 {
1675         struct cma_work *work = container_of(_work, struct cma_work, work);
1676         struct rdma_id_private *id_priv = work->id;
1677         int destroy = 0;
1678
1679         mutex_lock(&id_priv->handler_mutex);
1680         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1681                 goto out;
1682
1683         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1684                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1685                 destroy = 1;
1686         }
1687 out:
1688         mutex_unlock(&id_priv->handler_mutex);
1689         cma_deref_id(id_priv);
1690         if (destroy)
1691                 rdma_destroy_id(&id_priv->id);
1692         kfree(work);
1693 }
1694
1695 static void cma_ndev_work_handler(struct work_struct *_work)
1696 {
1697         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1698         struct rdma_id_private *id_priv = work->id;
1699         int destroy = 0;
1700
1701         mutex_lock(&id_priv->handler_mutex);
1702         if (id_priv->state == RDMA_CM_DESTROYING ||
1703             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1704                 goto out;
1705
1706         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1707                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1708                 destroy = 1;
1709         }
1710
1711 out:
1712         mutex_unlock(&id_priv->handler_mutex);
1713         cma_deref_id(id_priv);
1714         if (destroy)
1715                 rdma_destroy_id(&id_priv->id);
1716         kfree(work);
1717 }
1718
1719 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1720 {
1721         struct rdma_route *route = &id_priv->id.route;
1722         struct cma_work *work;
1723         int ret;
1724
1725         work = kzalloc(sizeof *work, GFP_KERNEL);
1726         if (!work)
1727                 return -ENOMEM;
1728
1729         work->id = id_priv;
1730         INIT_WORK(&work->work, cma_work_handler);
1731         work->old_state = RDMA_CM_ROUTE_QUERY;
1732         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1733         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1734
1735         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1736         if (!route->path_rec) {
1737                 ret = -ENOMEM;
1738                 goto err1;
1739         }
1740
1741         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1742         if (ret)
1743                 goto err2;
1744
1745         return 0;
1746 err2:
1747         kfree(route->path_rec);
1748         route->path_rec = NULL;
1749 err1:
1750         kfree(work);
1751         return ret;
1752 }
1753
1754 int rdma_set_ib_paths(struct rdma_cm_id *id,
1755                       struct ib_sa_path_rec *path_rec, int num_paths)
1756 {
1757         struct rdma_id_private *id_priv;
1758         int ret;
1759
1760         id_priv = container_of(id, struct rdma_id_private, id);
1761         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1762                            RDMA_CM_ROUTE_RESOLVED))
1763                 return -EINVAL;
1764
1765         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1766                                      GFP_KERNEL);
1767         if (!id->route.path_rec) {
1768                 ret = -ENOMEM;
1769                 goto err;
1770         }
1771
1772         id->route.num_paths = num_paths;
1773         return 0;
1774 err:
1775         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1776         return ret;
1777 }
1778 EXPORT_SYMBOL(rdma_set_ib_paths);
1779
1780 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1781 {
1782         struct cma_work *work;
1783
1784         work = kzalloc(sizeof *work, GFP_KERNEL);
1785         if (!work)
1786                 return -ENOMEM;
1787
1788         work->id = id_priv;
1789         INIT_WORK(&work->work, cma_work_handler);
1790         work->old_state = RDMA_CM_ROUTE_QUERY;
1791         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1792         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1793         queue_work(cma_wq, &work->work);
1794         return 0;
1795 }
1796
1797 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1798 {
1799         struct rdma_route *route = &id_priv->id.route;
1800         struct rdma_addr *addr = &route->addr;
1801         struct cma_work *work;
1802         int ret;
1803         struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1804         struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1805         struct net_device *ndev = NULL;
1806         u16 vid;
1807
1808         if (src_addr->sin_family != dst_addr->sin_family)
1809                 return -EINVAL;
1810
1811         work = kzalloc(sizeof *work, GFP_KERNEL);
1812         if (!work)
1813                 return -ENOMEM;
1814
1815         work->id = id_priv;
1816         INIT_WORK(&work->work, cma_work_handler);
1817
1818         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1819         if (!route->path_rec) {
1820                 ret = -ENOMEM;
1821                 goto err1;
1822         }
1823
1824         route->num_paths = 1;
1825
1826         if (addr->dev_addr.bound_dev_if)
1827                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1828         if (!ndev) {
1829                 ret = -ENODEV;
1830                 goto err2;
1831         }
1832
1833         vid = rdma_vlan_dev_vlan_id(ndev);
1834
1835         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1836         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1837
1838         route->path_rec->hop_limit = 1;
1839         route->path_rec->reversible = 1;
1840         route->path_rec->pkey = cpu_to_be16(0xffff);
1841         route->path_rec->mtu_selector = IB_SA_EQ;
1842         route->path_rec->sl = netdev_get_prio_tc_map(
1843                         ndev->priv_flags & IFF_802_1Q_VLAN ?
1844                                 vlan_dev_real_dev(ndev) : ndev,
1845                         rt_tos2priority(id_priv->tos));
1846
1847         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1848         route->path_rec->rate_selector = IB_SA_EQ;
1849         route->path_rec->rate = iboe_get_rate(ndev);
1850         dev_put(ndev);
1851         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1852         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1853         if (!route->path_rec->mtu) {
1854                 ret = -EINVAL;
1855                 goto err2;
1856         }
1857
1858         work->old_state = RDMA_CM_ROUTE_QUERY;
1859         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1860         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1861         work->event.status = 0;
1862
1863         queue_work(cma_wq, &work->work);
1864
1865         return 0;
1866
1867 err2:
1868         kfree(route->path_rec);
1869         route->path_rec = NULL;
1870 err1:
1871         kfree(work);
1872         return ret;
1873 }
1874
1875 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1876 {
1877         struct rdma_id_private *id_priv;
1878         int ret;
1879
1880         id_priv = container_of(id, struct rdma_id_private, id);
1881         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1882                 return -EINVAL;
1883
1884         atomic_inc(&id_priv->refcount);
1885         switch (rdma_node_get_transport(id->device->node_type)) {
1886         case RDMA_TRANSPORT_IB:
1887                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1888                 case IB_LINK_LAYER_INFINIBAND:
1889                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1890                         break;
1891                 case IB_LINK_LAYER_ETHERNET:
1892                         ret = cma_resolve_iboe_route(id_priv);
1893                         break;
1894                 default:
1895                         ret = -ENOSYS;
1896                 }
1897                 break;
1898         case RDMA_TRANSPORT_IWARP:
1899                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1900                 break;
1901         default:
1902                 ret = -ENOSYS;
1903                 break;
1904         }
1905         if (ret)
1906                 goto err;
1907
1908         return 0;
1909 err:
1910         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1911         cma_deref_id(id_priv);
1912         return ret;
1913 }
1914 EXPORT_SYMBOL(rdma_resolve_route);
1915
1916 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1917 {
1918         struct cma_device *cma_dev;
1919         struct ib_port_attr port_attr;
1920         union ib_gid gid;
1921         u16 pkey;
1922         int ret;
1923         u8 p;
1924
1925         mutex_lock(&lock);
1926         if (list_empty(&dev_list)) {
1927                 ret = -ENODEV;
1928                 goto out;
1929         }
1930         list_for_each_entry(cma_dev, &dev_list, list)
1931                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1932                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1933                             port_attr.state == IB_PORT_ACTIVE)
1934                                 goto port_found;
1935
1936         p = 1;
1937         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1938
1939 port_found:
1940         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1941         if (ret)
1942                 goto out;
1943
1944         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1945         if (ret)
1946                 goto out;
1947
1948         id_priv->id.route.addr.dev_addr.dev_type =
1949                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1950                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1951
1952         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1953         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1954         id_priv->id.port_num = p;
1955         cma_attach_to_dev(id_priv, cma_dev);
1956 out:
1957         mutex_unlock(&lock);
1958         return ret;
1959 }
1960
1961 static void addr_handler(int status, struct sockaddr *src_addr,
1962                          struct rdma_dev_addr *dev_addr, void *context)
1963 {
1964         struct rdma_id_private *id_priv = context;
1965         struct rdma_cm_event event;
1966
1967         memset(&event, 0, sizeof event);
1968         mutex_lock(&id_priv->handler_mutex);
1969         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1970                            RDMA_CM_ADDR_RESOLVED))
1971                 goto out;
1972
1973         if (!status && !id_priv->cma_dev)
1974                 status = cma_acquire_dev(id_priv);
1975
1976         if (status) {
1977                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1978                                    RDMA_CM_ADDR_BOUND))
1979                         goto out;
1980                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1981                 event.status = status;
1982         } else {
1983                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1984                        ip_addr_size(src_addr));
1985                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1986         }
1987
1988         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1989                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1990                 mutex_unlock(&id_priv->handler_mutex);
1991                 cma_deref_id(id_priv);
1992                 rdma_destroy_id(&id_priv->id);
1993                 return;
1994         }
1995 out:
1996         mutex_unlock(&id_priv->handler_mutex);
1997         cma_deref_id(id_priv);
1998 }
1999
2000 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2001 {
2002         struct cma_work *work;
2003         struct sockaddr *src, *dst;
2004         union ib_gid gid;
2005         int ret;
2006
2007         work = kzalloc(sizeof *work, GFP_KERNEL);
2008         if (!work)
2009                 return -ENOMEM;
2010
2011         if (!id_priv->cma_dev) {
2012                 ret = cma_bind_loopback(id_priv);
2013                 if (ret)
2014                         goto err;
2015         }
2016
2017         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2018         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2019
2020         src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2021         if (cma_zero_addr(src)) {
2022                 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2023                 if ((src->sa_family = dst->sa_family) == AF_INET) {
2024                         ((struct sockaddr_in *)src)->sin_addr =
2025                                 ((struct sockaddr_in *)dst)->sin_addr;
2026                 } else {
2027                         ((struct sockaddr_in6 *)src)->sin6_addr =
2028                                 ((struct sockaddr_in6 *)dst)->sin6_addr;
2029                 }
2030         }
2031
2032         work->id = id_priv;
2033         INIT_WORK(&work->work, cma_work_handler);
2034         work->old_state = RDMA_CM_ADDR_QUERY;
2035         work->new_state = RDMA_CM_ADDR_RESOLVED;
2036         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2037         queue_work(cma_wq, &work->work);
2038         return 0;
2039 err:
2040         kfree(work);
2041         return ret;
2042 }
2043
2044 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2045                          struct sockaddr *dst_addr)
2046 {
2047         if (!src_addr || !src_addr->sa_family) {
2048                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2049                 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2050                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2051                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2052                 }
2053         }
2054         return rdma_bind_addr(id, src_addr);
2055 }
2056
2057 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2058                       struct sockaddr *dst_addr, int timeout_ms)
2059 {
2060         struct rdma_id_private *id_priv;
2061         int ret;
2062
2063         id_priv = container_of(id, struct rdma_id_private, id);
2064         if (id_priv->state == RDMA_CM_IDLE) {
2065                 ret = cma_bind_addr(id, src_addr, dst_addr);
2066                 if (ret)
2067                         return ret;
2068         }
2069
2070         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2071                 return -EINVAL;
2072
2073         atomic_inc(&id_priv->refcount);
2074         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2075         if (cma_any_addr(dst_addr))
2076                 ret = cma_resolve_loopback(id_priv);
2077         else
2078                 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2079                                       dst_addr, &id->route.addr.dev_addr,
2080                                       timeout_ms, addr_handler, id_priv);
2081         if (ret)
2082                 goto err;
2083
2084         return 0;
2085 err:
2086         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2087         cma_deref_id(id_priv);
2088         return ret;
2089 }
2090 EXPORT_SYMBOL(rdma_resolve_addr);
2091
2092 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2093 {
2094         struct rdma_id_private *id_priv;
2095         unsigned long flags;
2096         int ret;
2097
2098         id_priv = container_of(id, struct rdma_id_private, id);
2099         spin_lock_irqsave(&id_priv->lock, flags);
2100         if (id_priv->state == RDMA_CM_IDLE) {
2101                 id_priv->reuseaddr = reuse;
2102                 ret = 0;
2103         } else {
2104                 ret = -EINVAL;
2105         }
2106         spin_unlock_irqrestore(&id_priv->lock, flags);
2107         return ret;
2108 }
2109 EXPORT_SYMBOL(rdma_set_reuseaddr);
2110
2111 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2112 {
2113         struct rdma_id_private *id_priv;
2114         unsigned long flags;
2115         int ret;
2116
2117         id_priv = container_of(id, struct rdma_id_private, id);
2118         spin_lock_irqsave(&id_priv->lock, flags);
2119         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2120                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2121                 id_priv->afonly = afonly;
2122                 ret = 0;
2123         } else {
2124                 ret = -EINVAL;
2125         }
2126         spin_unlock_irqrestore(&id_priv->lock, flags);
2127         return ret;
2128 }
2129 EXPORT_SYMBOL(rdma_set_afonly);
2130
2131 static void cma_bind_port(struct rdma_bind_list *bind_list,
2132                           struct rdma_id_private *id_priv)
2133 {
2134         struct sockaddr_in *sin;
2135
2136         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2137         sin->sin_port = htons(bind_list->port);
2138         id_priv->bind_list = bind_list;
2139         hlist_add_head(&id_priv->node, &bind_list->owners);
2140 }
2141
2142 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2143                           unsigned short snum)
2144 {
2145         struct rdma_bind_list *bind_list;
2146         int port, ret;
2147
2148         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2149         if (!bind_list)
2150                 return -ENOMEM;
2151
2152         do {
2153                 ret = idr_get_new_above(ps, bind_list, snum, &port);
2154         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2155
2156         if (ret)
2157                 goto err1;
2158
2159         if (port != snum) {
2160                 ret = -EADDRNOTAVAIL;
2161                 goto err2;
2162         }
2163
2164         bind_list->ps = ps;
2165         bind_list->port = (unsigned short) port;
2166         cma_bind_port(bind_list, id_priv);
2167         return 0;
2168 err2:
2169         idr_remove(ps, port);
2170 err1:
2171         kfree(bind_list);
2172         return ret;
2173 }
2174
2175 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2176 {
2177         static unsigned int last_used_port;
2178         int low, high, remaining;
2179         unsigned int rover;
2180
2181         inet_get_local_port_range(&low, &high);
2182         remaining = (high - low) + 1;
2183         rover = net_random() % remaining + low;
2184 retry:
2185         if (last_used_port != rover &&
2186             !idr_find(ps, (unsigned short) rover)) {
2187                 int ret = cma_alloc_port(ps, id_priv, rover);
2188                 /*
2189                  * Remember previously used port number in order to avoid
2190                  * re-using same port immediately after it is closed.
2191                  */
2192                 if (!ret)
2193                         last_used_port = rover;
2194                 if (ret != -EADDRNOTAVAIL)
2195                         return ret;
2196         }
2197         if (--remaining) {
2198                 rover++;
2199                 if ((rover < low) || (rover > high))
2200                         rover = low;
2201                 goto retry;
2202         }
2203         return -EADDRNOTAVAIL;
2204 }
2205
2206 /*
2207  * Check that the requested port is available.  This is called when trying to
2208  * bind to a specific port, or when trying to listen on a bound port.  In
2209  * the latter case, the provided id_priv may already be on the bind_list, but
2210  * we still need to check that it's okay to start listening.
2211  */
2212 static int cma_check_port(struct rdma_bind_list *bind_list,
2213                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2214 {
2215         struct rdma_id_private *cur_id;
2216         struct sockaddr *addr, *cur_addr;
2217         struct hlist_node *node;
2218
2219         addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2220         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2221                 if (id_priv == cur_id)
2222                         continue;
2223
2224                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2225                     cur_id->reuseaddr)
2226                         continue;
2227
2228                 cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2229                 if (id_priv->afonly && cur_id->afonly &&
2230                     (addr->sa_family != cur_addr->sa_family))
2231                         continue;
2232
2233                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2234                         return -EADDRNOTAVAIL;
2235
2236                 if (!cma_addr_cmp(addr, cur_addr))
2237                         return -EADDRINUSE;
2238         }
2239         return 0;
2240 }
2241
2242 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2243 {
2244         struct rdma_bind_list *bind_list;
2245         unsigned short snum;
2246         int ret;
2247
2248         snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2249         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2250                 return -EACCES;
2251
2252         bind_list = idr_find(ps, snum);
2253         if (!bind_list) {
2254                 ret = cma_alloc_port(ps, id_priv, snum);
2255         } else {
2256                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2257                 if (!ret)
2258                         cma_bind_port(bind_list, id_priv);
2259         }
2260         return ret;
2261 }
2262
2263 static int cma_bind_listen(struct rdma_id_private *id_priv)
2264 {
2265         struct rdma_bind_list *bind_list = id_priv->bind_list;
2266         int ret = 0;
2267
2268         mutex_lock(&lock);
2269         if (bind_list->owners.first->next)
2270                 ret = cma_check_port(bind_list, id_priv, 0);
2271         mutex_unlock(&lock);
2272         return ret;
2273 }
2274
2275 static int cma_get_port(struct rdma_id_private *id_priv)
2276 {
2277         struct idr *ps;
2278         int ret;
2279
2280         switch (id_priv->id.ps) {
2281         case RDMA_PS_SDP:
2282                 ps = &sdp_ps;
2283                 break;
2284         case RDMA_PS_TCP:
2285                 ps = &tcp_ps;
2286                 break;
2287         case RDMA_PS_UDP:
2288                 ps = &udp_ps;
2289                 break;
2290         case RDMA_PS_IPOIB:
2291                 ps = &ipoib_ps;
2292                 break;
2293         case RDMA_PS_IB:
2294                 ps = &ib_ps;
2295                 break;
2296         default:
2297                 return -EPROTONOSUPPORT;
2298         }
2299
2300         mutex_lock(&lock);
2301         if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2302                 ret = cma_alloc_any_port(ps, id_priv);
2303         else
2304                 ret = cma_use_port(ps, id_priv);
2305         mutex_unlock(&lock);
2306
2307         return ret;
2308 }
2309
2310 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2311                                struct sockaddr *addr)
2312 {
2313 #if IS_ENABLED(CONFIG_IPV6)
2314         struct sockaddr_in6 *sin6;
2315
2316         if (addr->sa_family != AF_INET6)
2317                 return 0;
2318
2319         sin6 = (struct sockaddr_in6 *) addr;
2320         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2321             !sin6->sin6_scope_id)
2322                         return -EINVAL;
2323
2324         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2325 #endif
2326         return 0;
2327 }
2328
2329 int rdma_listen(struct rdma_cm_id *id, int backlog)
2330 {
2331         struct rdma_id_private *id_priv;
2332         int ret;
2333
2334         id_priv = container_of(id, struct rdma_id_private, id);
2335         if (id_priv->state == RDMA_CM_IDLE) {
2336                 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2337                 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2338                 if (ret)
2339                         return ret;
2340         }
2341
2342         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2343                 return -EINVAL;
2344
2345         if (id_priv->reuseaddr) {
2346                 ret = cma_bind_listen(id_priv);
2347                 if (ret)
2348                         goto err;
2349         }
2350
2351         id_priv->backlog = backlog;
2352         if (id->device) {
2353                 switch (rdma_node_get_transport(id->device->node_type)) {
2354                 case RDMA_TRANSPORT_IB:
2355                         ret = cma_ib_listen(id_priv);
2356                         if (ret)
2357                                 goto err;
2358                         break;
2359                 case RDMA_TRANSPORT_IWARP:
2360                         ret = cma_iw_listen(id_priv, backlog);
2361                         if (ret)
2362                                 goto err;
2363                         break;
2364                 default:
2365                         ret = -ENOSYS;
2366                         goto err;
2367                 }
2368         } else
2369                 cma_listen_on_all(id_priv);
2370
2371         return 0;
2372 err:
2373         id_priv->backlog = 0;
2374         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2375         return ret;
2376 }
2377 EXPORT_SYMBOL(rdma_listen);
2378
2379 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2380 {
2381         struct rdma_id_private *id_priv;
2382         int ret;
2383
2384         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2385                 return -EAFNOSUPPORT;
2386
2387         id_priv = container_of(id, struct rdma_id_private, id);
2388         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2389                 return -EINVAL;
2390
2391         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2392         if (ret)
2393                 goto err1;
2394
2395         if (!cma_any_addr(addr)) {
2396                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2397                 if (ret)
2398                         goto err1;
2399
2400                 ret = cma_acquire_dev(id_priv);
2401                 if (ret)
2402                         goto err1;
2403         }
2404
2405         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2406         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2407                 if (addr->sa_family == AF_INET)
2408                         id_priv->afonly = 1;
2409 #if IS_ENABLED(CONFIG_IPV6)
2410                 else if (addr->sa_family == AF_INET6)
2411                         id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2412 #endif
2413         }
2414         ret = cma_get_port(id_priv);
2415         if (ret)
2416                 goto err2;
2417
2418         return 0;
2419 err2:
2420         if (id_priv->cma_dev)
2421                 cma_release_dev(id_priv);
2422 err1:
2423         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2424         return ret;
2425 }
2426 EXPORT_SYMBOL(rdma_bind_addr);
2427
2428 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2429                           struct rdma_route *route)
2430 {
2431         struct cma_hdr *cma_hdr;
2432         struct sdp_hh *sdp_hdr;
2433
2434         if (route->addr.src_addr.ss_family == AF_INET) {
2435                 struct sockaddr_in *src4, *dst4;
2436
2437                 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2438                 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2439
2440                 switch (ps) {
2441                 case RDMA_PS_SDP:
2442                         sdp_hdr = hdr;
2443                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2444                                 return -EINVAL;
2445                         sdp_set_ip_ver(sdp_hdr, 4);
2446                         sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2447                         sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2448                         sdp_hdr->port = src4->sin_port;
2449                         break;
2450                 default:
2451                         cma_hdr = hdr;
2452                         cma_hdr->cma_version = CMA_VERSION;
2453                         cma_set_ip_ver(cma_hdr, 4);
2454                         cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2455                         cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2456                         cma_hdr->port = src4->sin_port;
2457                         break;
2458                 }
2459         } else {
2460                 struct sockaddr_in6 *src6, *dst6;
2461
2462                 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2463                 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2464
2465                 switch (ps) {
2466                 case RDMA_PS_SDP:
2467                         sdp_hdr = hdr;
2468                         if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2469                                 return -EINVAL;
2470                         sdp_set_ip_ver(sdp_hdr, 6);
2471                         sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2472                         sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2473                         sdp_hdr->port = src6->sin6_port;
2474                         break;
2475                 default:
2476                         cma_hdr = hdr;
2477                         cma_hdr->cma_version = CMA_VERSION;
2478                         cma_set_ip_ver(cma_hdr, 6);
2479                         cma_hdr->src_addr.ip6 = src6->sin6_addr;
2480                         cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2481                         cma_hdr->port = src6->sin6_port;
2482                         break;
2483                 }
2484         }
2485         return 0;
2486 }
2487
2488 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2489                                 struct ib_cm_event *ib_event)
2490 {
2491         struct rdma_id_private *id_priv = cm_id->context;
2492         struct rdma_cm_event event;
2493         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2494         int ret = 0;
2495
2496         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2497                 return 0;
2498
2499         memset(&event, 0, sizeof event);
2500         switch (ib_event->event) {
2501         case IB_CM_SIDR_REQ_ERROR:
2502                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2503                 event.status = -ETIMEDOUT;
2504                 break;
2505         case IB_CM_SIDR_REP_RECEIVED:
2506                 event.param.ud.private_data = ib_event->private_data;
2507                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2508                 if (rep->status != IB_SIDR_SUCCESS) {
2509                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2510                         event.status = ib_event->param.sidr_rep_rcvd.status;
2511                         break;
2512                 }
2513                 ret = cma_set_qkey(id_priv);
2514                 if (ret) {
2515                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2516                         event.status = -EINVAL;
2517                         break;
2518                 }
2519                 if (id_priv->qkey != rep->qkey) {
2520                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2521                         event.status = -EINVAL;
2522                         break;
2523                 }
2524                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2525                                      id_priv->id.route.path_rec,
2526                                      &event.param.ud.ah_attr);
2527                 event.param.ud.qp_num = rep->qpn;
2528                 event.param.ud.qkey = rep->qkey;
2529                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2530                 event.status = 0;
2531                 break;
2532         default:
2533                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2534                        ib_event->event);
2535                 goto out;
2536         }
2537
2538         ret = id_priv->id.event_handler(&id_priv->id, &event);
2539         if (ret) {
2540                 /* Destroy the CM ID by returning a non-zero value. */
2541                 id_priv->cm_id.ib = NULL;
2542                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2543                 mutex_unlock(&id_priv->handler_mutex);
2544                 rdma_destroy_id(&id_priv->id);
2545                 return ret;
2546         }
2547 out:
2548         mutex_unlock(&id_priv->handler_mutex);
2549         return ret;
2550 }
2551
2552 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2553                               struct rdma_conn_param *conn_param)
2554 {
2555         struct ib_cm_sidr_req_param req;
2556         struct rdma_route *route;
2557         struct ib_cm_id *id;
2558         int ret;
2559
2560         req.private_data_len = sizeof(struct cma_hdr) +
2561                                conn_param->private_data_len;
2562         if (req.private_data_len < conn_param->private_data_len)
2563                 return -EINVAL;
2564
2565         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2566         if (!req.private_data)
2567                 return -ENOMEM;
2568
2569         if (conn_param->private_data && conn_param->private_data_len)
2570                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2571                        conn_param->private_data, conn_param->private_data_len);
2572
2573         route = &id_priv->id.route;
2574         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2575         if (ret)
2576                 goto out;
2577
2578         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2579                              id_priv);
2580         if (IS_ERR(id)) {
2581                 ret = PTR_ERR(id);
2582                 goto out;
2583         }
2584         id_priv->cm_id.ib = id;
2585
2586         req.path = route->path_rec;
2587         req.service_id = cma_get_service_id(id_priv->id.ps,
2588                                             (struct sockaddr *) &route->addr.dst_addr);
2589         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2590         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2591
2592         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2593         if (ret) {
2594                 ib_destroy_cm_id(id_priv->cm_id.ib);
2595                 id_priv->cm_id.ib = NULL;
2596         }
2597 out:
2598         kfree(req.private_data);
2599         return ret;
2600 }
2601
2602 static int cma_connect_ib(struct rdma_id_private *id_priv,
2603                           struct rdma_conn_param *conn_param)
2604 {
2605         struct ib_cm_req_param req;
2606         struct rdma_route *route;
2607         void *private_data;
2608         struct ib_cm_id *id;
2609         int offset, ret;
2610
2611         memset(&req, 0, sizeof req);
2612         offset = cma_user_data_offset(id_priv->id.ps);
2613         req.private_data_len = offset + conn_param->private_data_len;
2614         if (req.private_data_len < conn_param->private_data_len)
2615                 return -EINVAL;
2616
2617         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2618         if (!private_data)
2619                 return -ENOMEM;
2620
2621         if (conn_param->private_data && conn_param->private_data_len)
2622                 memcpy(private_data + offset, conn_param->private_data,
2623                        conn_param->private_data_len);
2624
2625         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2626         if (IS_ERR(id)) {
2627                 ret = PTR_ERR(id);
2628                 goto out;
2629         }
2630         id_priv->cm_id.ib = id;
2631
2632         route = &id_priv->id.route;
2633         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2634         if (ret)
2635                 goto out;
2636         req.private_data = private_data;
2637
2638         req.primary_path = &route->path_rec[0];
2639         if (route->num_paths == 2)
2640                 req.alternate_path = &route->path_rec[1];
2641
2642         req.service_id = cma_get_service_id(id_priv->id.ps,
2643                                             (struct sockaddr *) &route->addr.dst_addr);
2644         req.qp_num = id_priv->qp_num;
2645         req.qp_type = id_priv->id.qp_type;
2646         req.starting_psn = id_priv->seq_num;
2647         req.responder_resources = conn_param->responder_resources;
2648         req.initiator_depth = conn_param->initiator_depth;
2649         req.flow_control = conn_param->flow_control;
2650         req.retry_count = min_t(u8, 7, conn_param->retry_count);
2651         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2652         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2653         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2654         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2655         req.srq = id_priv->srq ? 1 : 0;
2656
2657         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2658 out:
2659         if (ret && !IS_ERR(id)) {
2660                 ib_destroy_cm_id(id);
2661                 id_priv->cm_id.ib = NULL;
2662         }
2663
2664         kfree(private_data);
2665         return ret;
2666 }
2667
2668 static int cma_connect_iw(struct rdma_id_private *id_priv,
2669                           struct rdma_conn_param *conn_param)
2670 {
2671         struct iw_cm_id *cm_id;
2672         struct sockaddr_in* sin;
2673         int ret;
2674         struct iw_cm_conn_param iw_param;
2675
2676         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2677         if (IS_ERR(cm_id))
2678                 return PTR_ERR(cm_id);
2679
2680         id_priv->cm_id.iw = cm_id;
2681
2682         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2683         cm_id->local_addr = *sin;
2684
2685         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2686         cm_id->remote_addr = *sin;
2687
2688         ret = cma_modify_qp_rtr(id_priv, conn_param);
2689         if (ret)
2690                 goto out;
2691
2692         if (conn_param) {
2693                 iw_param.ord = conn_param->initiator_depth;
2694                 iw_param.ird = conn_param->responder_resources;
2695                 iw_param.private_data = conn_param->private_data;
2696                 iw_param.private_data_len = conn_param->private_data_len;
2697                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2698         } else {
2699                 memset(&iw_param, 0, sizeof iw_param);
2700                 iw_param.qpn = id_priv->qp_num;
2701         }
2702         ret = iw_cm_connect(cm_id, &iw_param);
2703 out:
2704         if (ret) {
2705                 iw_destroy_cm_id(cm_id);
2706                 id_priv->cm_id.iw = NULL;
2707         }
2708         return ret;
2709 }
2710
2711 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2712 {
2713         struct rdma_id_private *id_priv;
2714         int ret;
2715
2716         id_priv = container_of(id, struct rdma_id_private, id);
2717         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2718                 return -EINVAL;
2719
2720         if (!id->qp) {
2721                 id_priv->qp_num = conn_param->qp_num;
2722                 id_priv->srq = conn_param->srq;
2723         }
2724
2725         switch (rdma_node_get_transport(id->device->node_type)) {
2726         case RDMA_TRANSPORT_IB:
2727                 if (id->qp_type == IB_QPT_UD)
2728                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2729                 else
2730                         ret = cma_connect_ib(id_priv, conn_param);
2731                 break;
2732         case RDMA_TRANSPORT_IWARP:
2733                 ret = cma_connect_iw(id_priv, conn_param);
2734                 break;
2735         default:
2736                 ret = -ENOSYS;
2737                 break;
2738         }
2739         if (ret)
2740                 goto err;
2741
2742         return 0;
2743 err:
2744         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2745         return ret;
2746 }
2747 EXPORT_SYMBOL(rdma_connect);
2748
2749 static int cma_accept_ib(struct rdma_id_private *id_priv,
2750                          struct rdma_conn_param *conn_param)
2751 {
2752         struct ib_cm_rep_param rep;
2753         int ret;
2754
2755         ret = cma_modify_qp_rtr(id_priv, conn_param);
2756         if (ret)
2757                 goto out;
2758
2759         ret = cma_modify_qp_rts(id_priv, conn_param);
2760         if (ret)
2761                 goto out;
2762
2763         memset(&rep, 0, sizeof rep);
2764         rep.qp_num = id_priv->qp_num;
2765         rep.starting_psn = id_priv->seq_num;
2766         rep.private_data = conn_param->private_data;
2767         rep.private_data_len = conn_param->private_data_len;
2768         rep.responder_resources = conn_param->responder_resources;
2769         rep.initiator_depth = conn_param->initiator_depth;
2770         rep.failover_accepted = 0;
2771         rep.flow_control = conn_param->flow_control;
2772         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2773         rep.srq = id_priv->srq ? 1 : 0;
2774
2775         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2776 out:
2777         return ret;
2778 }
2779
2780 static int cma_accept_iw(struct rdma_id_private *id_priv,
2781                   struct rdma_conn_param *conn_param)
2782 {
2783         struct iw_cm_conn_param iw_param;
2784         int ret;
2785
2786         ret = cma_modify_qp_rtr(id_priv, conn_param);
2787         if (ret)
2788                 return ret;
2789
2790         iw_param.ord = conn_param->initiator_depth;
2791         iw_param.ird = conn_param->responder_resources;
2792         iw_param.private_data = conn_param->private_data;
2793         iw_param.private_data_len = conn_param->private_data_len;
2794         if (id_priv->id.qp) {
2795                 iw_param.qpn = id_priv->qp_num;
2796         } else
2797                 iw_param.qpn = conn_param->qp_num;
2798
2799         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2800 }
2801
2802 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2803                              enum ib_cm_sidr_status status,
2804                              const void *private_data, int private_data_len)
2805 {
2806         struct ib_cm_sidr_rep_param rep;
2807         int ret;
2808
2809         memset(&rep, 0, sizeof rep);
2810         rep.status = status;
2811         if (status == IB_SIDR_SUCCESS) {
2812                 ret = cma_set_qkey(id_priv);
2813                 if (ret)
2814                         return ret;
2815                 rep.qp_num = id_priv->qp_num;
2816                 rep.qkey = id_priv->qkey;
2817         }
2818         rep.private_data = private_data;
2819         rep.private_data_len = private_data_len;
2820
2821         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2822 }
2823
2824 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2825 {
2826         struct rdma_id_private *id_priv;
2827         int ret;
2828
2829         id_priv = container_of(id, struct rdma_id_private, id);
2830
2831         id_priv->owner = task_pid_nr(current);
2832
2833         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2834                 return -EINVAL;
2835
2836         if (!id->qp && conn_param) {
2837                 id_priv->qp_num = conn_param->qp_num;
2838                 id_priv->srq = conn_param->srq;
2839         }
2840
2841         switch (rdma_node_get_transport(id->device->node_type)) {
2842         case RDMA_TRANSPORT_IB:
2843                 if (id->qp_type == IB_QPT_UD) {
2844                         if (conn_param)
2845                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2846                                                         conn_param->private_data,
2847                                                         conn_param->private_data_len);
2848                         else
2849                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2850                                                         NULL, 0);
2851                 } else {
2852                         if (conn_param)
2853                                 ret = cma_accept_ib(id_priv, conn_param);
2854                         else
2855                                 ret = cma_rep_recv(id_priv);
2856                 }
2857                 break;
2858         case RDMA_TRANSPORT_IWARP:
2859                 ret = cma_accept_iw(id_priv, conn_param);
2860                 break;
2861         default:
2862                 ret = -ENOSYS;
2863                 break;
2864         }
2865
2866         if (ret)
2867                 goto reject;
2868
2869         return 0;
2870 reject:
2871         cma_modify_qp_err(id_priv);
2872         rdma_reject(id, NULL, 0);
2873         return ret;
2874 }
2875 EXPORT_SYMBOL(rdma_accept);
2876
2877 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2878 {
2879         struct rdma_id_private *id_priv;
2880         int ret;
2881
2882         id_priv = container_of(id, struct rdma_id_private, id);
2883         if (!id_priv->cm_id.ib)
2884                 return -EINVAL;
2885
2886         switch (id->device->node_type) {
2887         case RDMA_NODE_IB_CA:
2888                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2889                 break;
2890         default:
2891                 ret = 0;
2892                 break;
2893         }
2894         return ret;
2895 }
2896 EXPORT_SYMBOL(rdma_notify);
2897
2898 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2899                 u8 private_data_len)
2900 {
2901         struct rdma_id_private *id_priv;
2902         int ret;
2903
2904         id_priv = container_of(id, struct rdma_id_private, id);
2905         if (!id_priv->cm_id.ib)
2906                 return -EINVAL;
2907
2908         switch (rdma_node_get_transport(id->device->node_type)) {
2909         case RDMA_TRANSPORT_IB:
2910                 if (id->qp_type == IB_QPT_UD)
2911                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2912                                                 private_data, private_data_len);
2913                 else
2914                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2915                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2916                                              0, private_data, private_data_len);
2917                 break;
2918         case RDMA_TRANSPORT_IWARP:
2919                 ret = iw_cm_reject(id_priv->cm_id.iw,
2920                                    private_data, private_data_len);
2921                 break;
2922         default:
2923                 ret = -ENOSYS;
2924                 break;
2925         }
2926         return ret;
2927 }
2928 EXPORT_SYMBOL(rdma_reject);
2929
2930 int rdma_disconnect(struct rdma_cm_id *id)
2931 {
2932         struct rdma_id_private *id_priv;
2933         int ret;
2934
2935         id_priv = container_of(id, struct rdma_id_private, id);
2936         if (!id_priv->cm_id.ib)
2937                 return -EINVAL;
2938
2939         switch (rdma_node_get_transport(id->device->node_type)) {
2940         case RDMA_TRANSPORT_IB:
2941                 ret = cma_modify_qp_err(id_priv);
2942                 if (ret)
2943                         goto out;
2944                 /* Initiate or respond to a disconnect. */
2945                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2946                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2947                 break;
2948         case RDMA_TRANSPORT_IWARP:
2949                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2950                 break;
2951         default:
2952                 ret = -EINVAL;
2953                 break;
2954         }
2955 out:
2956         return ret;
2957 }
2958 EXPORT_SYMBOL(rdma_disconnect);
2959
2960 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2961 {
2962         struct rdma_id_private *id_priv;
2963         struct cma_multicast *mc = multicast->context;
2964         struct rdma_cm_event event;
2965         int ret;
2966
2967         id_priv = mc->id_priv;
2968         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2969             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2970                 return 0;
2971
2972         mutex_lock(&id_priv->qp_mutex);
2973         if (!status && id_priv->id.qp)
2974                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2975                                          be16_to_cpu(multicast->rec.mlid));
2976         mutex_unlock(&id_priv->qp_mutex);
2977
2978         memset(&event, 0, sizeof event);
2979         event.status = status;
2980         event.param.ud.private_data = mc->context;
2981         if (!status) {
2982                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2983                 ib_init_ah_from_mcmember(id_priv->id.device,
2984                                          id_priv->id.port_num, &multicast->rec,
2985                                          &event.param.ud.ah_attr);
2986                 event.param.ud.qp_num = 0xFFFFFF;
2987                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2988         } else
2989                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2990
2991         ret = id_priv->id.event_handler(&id_priv->id, &event);
2992         if (ret) {
2993                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2994                 mutex_unlock(&id_priv->handler_mutex);
2995                 rdma_destroy_id(&id_priv->id);
2996                 return 0;
2997         }
2998
2999         mutex_unlock(&id_priv->handler_mutex);
3000         return 0;
3001 }
3002
3003 static void cma_set_mgid(struct rdma_id_private *id_priv,
3004                          struct sockaddr *addr, union ib_gid *mgid)
3005 {
3006         unsigned char mc_map[MAX_ADDR_LEN];
3007         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3008         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3009         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3010
3011         if (cma_any_addr(addr)) {
3012                 memset(mgid, 0, sizeof *mgid);
3013         } else if ((addr->sa_family == AF_INET6) &&
3014                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3015                                                                  0xFF10A01B)) {
3016                 /* IPv6 address is an SA assigned MGID. */
3017                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3018         } else if ((addr->sa_family == AF_INET6)) {
3019                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3020                 if (id_priv->id.ps == RDMA_PS_UDP)
3021                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3022                 *mgid = *(union ib_gid *) (mc_map + 4);
3023         } else {
3024                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3025                 if (id_priv->id.ps == RDMA_PS_UDP)
3026                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3027                 *mgid = *(union ib_gid *) (mc_map + 4);
3028         }
3029 }
3030
3031 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3032                                  struct cma_multicast *mc)
3033 {
3034         struct ib_sa_mcmember_rec rec;
3035         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3036         ib_sa_comp_mask comp_mask;
3037         int ret;
3038
3039         ib_addr_get_mgid(dev_addr, &rec.mgid);
3040         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3041                                      &rec.mgid, &rec);
3042         if (ret)
3043                 return ret;
3044
3045         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3046         if (id_priv->id.ps == RDMA_PS_UDP)
3047                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3048         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3049         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3050         rec.join_state = 1;
3051
3052         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3053                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3054                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3055                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3056                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3057
3058         if (id_priv->id.ps == RDMA_PS_IPOIB)
3059                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3060                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3061                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3062                              IB_SA_MCMEMBER_REC_MTU |
3063                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
3064
3065         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3066                                                 id_priv->id.port_num, &rec,
3067                                                 comp_mask, GFP_KERNEL,
3068                                                 cma_ib_mc_handler, mc);
3069         return PTR_RET(mc->multicast.ib);
3070 }
3071
3072 static void iboe_mcast_work_handler(struct work_struct *work)
3073 {
3074         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3075         struct cma_multicast *mc = mw->mc;
3076         struct ib_sa_multicast *m = mc->multicast.ib;
3077
3078         mc->multicast.ib->context = mc;
3079         cma_ib_mc_handler(0, m);
3080         kref_put(&mc->mcref, release_mc);
3081         kfree(mw);
3082 }
3083
3084 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3085 {
3086         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3087         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3088
3089         if (cma_any_addr(addr)) {
3090                 memset(mgid, 0, sizeof *mgid);
3091         } else if (addr->sa_family == AF_INET6) {
3092                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3093         } else {
3094                 mgid->raw[0] = 0xff;
3095                 mgid->raw[1] = 0x0e;
3096                 mgid->raw[2] = 0;
3097                 mgid->raw[3] = 0;
3098                 mgid->raw[4] = 0;
3099                 mgid->raw[5] = 0;
3100                 mgid->raw[6] = 0;
3101                 mgid->raw[7] = 0;
3102                 mgid->raw[8] = 0;
3103                 mgid->raw[9] = 0;
3104                 mgid->raw[10] = 0xff;
3105                 mgid->raw[11] = 0xff;
3106                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3107         }
3108 }
3109
3110 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3111                                    struct cma_multicast *mc)
3112 {
3113         struct iboe_mcast_work *work;
3114         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3115         int err;
3116         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3117         struct net_device *ndev = NULL;
3118
3119         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3120                 return -EINVAL;
3121
3122         work = kzalloc(sizeof *work, GFP_KERNEL);
3123         if (!work)
3124                 return -ENOMEM;
3125
3126         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3127         if (!mc->multicast.ib) {
3128                 err = -ENOMEM;
3129                 goto out1;
3130         }
3131
3132         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3133
3134         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3135         if (id_priv->id.ps == RDMA_PS_UDP)
3136                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3137
3138         if (dev_addr->bound_dev_if)
3139                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3140         if (!ndev) {
3141                 err = -ENODEV;
3142                 goto out2;
3143         }
3144         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3145         mc->multicast.ib->rec.hop_limit = 1;
3146         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3147         dev_put(ndev);
3148         if (!mc->multicast.ib->rec.mtu) {
3149                 err = -EINVAL;
3150                 goto out2;
3151         }
3152         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3153         work->id = id_priv;
3154         work->mc = mc;
3155         INIT_WORK(&work->work, iboe_mcast_work_handler);
3156         kref_get(&mc->mcref);
3157         queue_work(cma_wq, &work->work);
3158
3159         return 0;
3160
3161 out2:
3162         kfree(mc->multicast.ib);
3163 out1:
3164         kfree(work);
3165         return err;
3166 }
3167
3168 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3169                         void *context)
3170 {
3171         struct rdma_id_private *id_priv;
3172         struct cma_multicast *mc;
3173         int ret;
3174
3175         id_priv = container_of(id, struct rdma_id_private, id);
3176         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3177             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3178                 return -EINVAL;
3179
3180         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3181         if (!mc)
3182                 return -ENOMEM;
3183
3184         memcpy(&mc->addr, addr, ip_addr_size(addr));
3185         mc->context = context;
3186         mc->id_priv = id_priv;
3187
3188         spin_lock(&id_priv->lock);
3189         list_add(&mc->list, &id_priv->mc_list);
3190         spin_unlock(&id_priv->lock);
3191
3192         switch (rdma_node_get_transport(id->device->node_type)) {
3193         case RDMA_TRANSPORT_IB:
3194                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3195                 case IB_LINK_LAYER_INFINIBAND:
3196                         ret = cma_join_ib_multicast(id_priv, mc);
3197                         break;
3198                 case IB_LINK_LAYER_ETHERNET:
3199                         kref_init(&mc->mcref);
3200                         ret = cma_iboe_join_multicast(id_priv, mc);
3201                         break;
3202                 default:
3203                         ret = -EINVAL;
3204                 }
3205                 break;
3206         default:
3207                 ret = -ENOSYS;
3208                 break;
3209         }
3210
3211         if (ret) {
3212                 spin_lock_irq(&id_priv->lock);
3213                 list_del(&mc->list);
3214                 spin_unlock_irq(&id_priv->lock);
3215                 kfree(mc);
3216         }
3217         return ret;
3218 }
3219 EXPORT_SYMBOL(rdma_join_multicast);
3220
3221 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3222 {
3223         struct rdma_id_private *id_priv;
3224         struct cma_multicast *mc;
3225
3226         id_priv = container_of(id, struct rdma_id_private, id);
3227         spin_lock_irq(&id_priv->lock);
3228         list_for_each_entry(mc, &id_priv->mc_list, list) {
3229                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3230                         list_del(&mc->list);
3231                         spin_unlock_irq(&id_priv->lock);
3232
3233                         if (id->qp)
3234                                 ib_detach_mcast(id->qp,
3235                                                 &mc->multicast.ib->rec.mgid,
3236                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3237                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3238                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3239                                 case IB_LINK_LAYER_INFINIBAND:
3240                                         ib_sa_free_multicast(mc->multicast.ib);
3241                                         kfree(mc);
3242                                         break;
3243                                 case IB_LINK_LAYER_ETHERNET:
3244                                         kref_put(&mc->mcref, release_mc);
3245                                         break;
3246                                 default:
3247                                         break;
3248                                 }
3249                         }
3250                         return;
3251                 }
3252         }
3253         spin_unlock_irq(&id_priv->lock);
3254 }
3255 EXPORT_SYMBOL(rdma_leave_multicast);
3256
3257 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3258 {
3259         struct rdma_dev_addr *dev_addr;
3260         struct cma_ndev_work *work;
3261
3262         dev_addr = &id_priv->id.route.addr.dev_addr;
3263
3264         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3265             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3266                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3267                        ndev->name, &id_priv->id);
3268                 work = kzalloc(sizeof *work, GFP_KERNEL);
3269                 if (!work)
3270                         return -ENOMEM;
3271
3272                 INIT_WORK(&work->work, cma_ndev_work_handler);
3273                 work->id = id_priv;
3274                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3275                 atomic_inc(&id_priv->refcount);
3276                 queue_work(cma_wq, &work->work);
3277         }
3278
3279         return 0;
3280 }
3281
3282 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3283                                void *ctx)
3284 {
3285         struct net_device *ndev = (struct net_device *)ctx;
3286         struct cma_device *cma_dev;
3287         struct rdma_id_private *id_priv;
3288         int ret = NOTIFY_DONE;
3289
3290         if (dev_net(ndev) != &init_net)
3291                 return NOTIFY_DONE;
3292
3293         if (event != NETDEV_BONDING_FAILOVER)
3294                 return NOTIFY_DONE;
3295
3296         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3297                 return NOTIFY_DONE;
3298
3299         mutex_lock(&lock);
3300         list_for_each_entry(cma_dev, &dev_list, list)
3301                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3302                         ret = cma_netdev_change(ndev, id_priv);
3303                         if (ret)
3304                                 goto out;
3305                 }
3306
3307 out:
3308         mutex_unlock(&lock);
3309         return ret;
3310 }
3311
3312 static struct notifier_block cma_nb = {
3313         .notifier_call = cma_netdev_callback
3314 };
3315
3316 static void cma_add_one(struct ib_device *device)
3317 {
3318         struct cma_device *cma_dev;
3319         struct rdma_id_private *id_priv;
3320
3321         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3322         if (!cma_dev)
3323                 return;
3324
3325         cma_dev->device = device;
3326
3327         init_completion(&cma_dev->comp);
3328         atomic_set(&cma_dev->refcount, 1);
3329         INIT_LIST_HEAD(&cma_dev->id_list);
3330         ib_set_client_data(device, &cma_client, cma_dev);
3331
3332         mutex_lock(&lock);
3333         list_add_tail(&cma_dev->list, &dev_list);
3334         list_for_each_entry(id_priv, &listen_any_list, list)
3335                 cma_listen_on_dev(id_priv, cma_dev);
3336         mutex_unlock(&lock);
3337 }
3338
3339 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3340 {
3341         struct rdma_cm_event event;
3342         enum rdma_cm_state state;
3343         int ret = 0;
3344
3345         /* Record that we want to remove the device */
3346         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3347         if (state == RDMA_CM_DESTROYING)
3348                 return 0;
3349
3350         cma_cancel_operation(id_priv, state);
3351         mutex_lock(&id_priv->handler_mutex);
3352
3353         /* Check for destruction from another callback. */
3354         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3355                 goto out;
3356
3357         memset(&event, 0, sizeof event);
3358         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3359         ret = id_priv->id.event_handler(&id_priv->id, &event);
3360 out:
3361         mutex_unlock(&id_priv->handler_mutex);
3362         return ret;
3363 }
3364
3365 static void cma_process_remove(struct cma_device *cma_dev)
3366 {
3367         struct rdma_id_private *id_priv;
3368         int ret;
3369
3370         mutex_lock(&lock);
3371         while (!list_empty(&cma_dev->id_list)) {
3372                 id_priv = list_entry(cma_dev->id_list.next,
3373                                      struct rdma_id_private, list);
3374
3375                 list_del(&id_priv->listen_list);
3376                 list_del_init(&id_priv->list);
3377                 atomic_inc(&id_priv->refcount);
3378                 mutex_unlock(&lock);
3379
3380                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3381                 cma_deref_id(id_priv);
3382                 if (ret)
3383                         rdma_destroy_id(&id_priv->id);
3384
3385                 mutex_lock(&lock);
3386         }
3387         mutex_unlock(&lock);
3388
3389         cma_deref_dev(cma_dev);
3390         wait_for_completion(&cma_dev->comp);
3391 }
3392
3393 static void cma_remove_one(struct ib_device *device)
3394 {
3395         struct cma_device *cma_dev;
3396
3397         cma_dev = ib_get_client_data(device, &cma_client);
3398         if (!cma_dev)
3399                 return;
3400
3401         mutex_lock(&lock);
3402         list_del(&cma_dev->list);
3403         mutex_unlock(&lock);
3404
3405         cma_process_remove(cma_dev);
3406         kfree(cma_dev);
3407 }
3408
3409 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3410 {
3411         struct nlmsghdr *nlh;
3412         struct rdma_cm_id_stats *id_stats;
3413         struct rdma_id_private *id_priv;
3414         struct rdma_cm_id *id = NULL;
3415         struct cma_device *cma_dev;
3416         int i_dev = 0, i_id = 0;
3417
3418         /*
3419          * We export all of the IDs as a sequence of messages.  Each
3420          * ID gets its own netlink message.
3421          */
3422         mutex_lock(&lock);
3423
3424         list_for_each_entry(cma_dev, &dev_list, list) {
3425                 if (i_dev < cb->args[0]) {
3426                         i_dev++;
3427                         continue;
3428                 }
3429
3430                 i_id = 0;
3431                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3432                         if (i_id < cb->args[1]) {
3433                                 i_id++;
3434                                 continue;
3435                         }
3436
3437                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3438                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3439                                                 RDMA_NL_RDMA_CM_ID_STATS);
3440                         if (!id_stats)
3441                                 goto out;
3442
3443                         memset(id_stats, 0, sizeof *id_stats);
3444                         id = &id_priv->id;
3445                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3446                         id_stats->port_num = id->port_num;
3447                         id_stats->bound_dev_if =
3448                                 id->route.addr.dev_addr.bound_dev_if;
3449
3450                         if (id->route.addr.src_addr.ss_family == AF_INET) {
3451                                 if (ibnl_put_attr(skb, nlh,
3452                                                   sizeof(struct sockaddr_in),
3453                                                   &id->route.addr.src_addr,
3454                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3455                                         goto out;
3456                                 }
3457                                 if (ibnl_put_attr(skb, nlh,
3458                                                   sizeof(struct sockaddr_in),
3459                                                   &id->route.addr.dst_addr,
3460                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3461                                         goto out;
3462                                 }
3463                         } else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3464                                 if (ibnl_put_attr(skb, nlh,
3465                                                   sizeof(struct sockaddr_in6),
3466                                                   &id->route.addr.src_addr,
3467                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3468                                         goto out;
3469                                 }
3470                                 if (ibnl_put_attr(skb, nlh,
3471                                                   sizeof(struct sockaddr_in6),
3472                                                   &id->route.addr.dst_addr,
3473                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3474                                         goto out;
3475                                 }
3476                         }
3477
3478                         id_stats->pid           = id_priv->owner;
3479                         id_stats->port_space    = id->ps;
3480                         id_stats->cm_state      = id_priv->state;
3481                         id_stats->qp_num        = id_priv->qp_num;
3482                         id_stats->qp_type       = id->qp_type;
3483
3484                         i_id++;
3485                 }
3486
3487                 cb->args[1] = 0;
3488                 i_dev++;
3489         }
3490
3491 out:
3492         mutex_unlock(&lock);
3493         cb->args[0] = i_dev;
3494         cb->args[1] = i_id;
3495
3496         return skb->len;
3497 }
3498
3499 static const struct ibnl_client_cbs cma_cb_table[] = {
3500         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3501                                        .module = THIS_MODULE },
3502 };
3503
3504 static int __init cma_init(void)
3505 {
3506         int ret;
3507
3508         cma_wq = create_singlethread_workqueue("rdma_cm");
3509         if (!cma_wq)
3510                 return -ENOMEM;
3511
3512         ib_sa_register_client(&sa_client);
3513         rdma_addr_register_client(&addr_client);
3514         register_netdevice_notifier(&cma_nb);
3515
3516         ret = ib_register_client(&cma_client);
3517         if (ret)
3518                 goto err;
3519
3520         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3521                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3522
3523         return 0;
3524
3525 err:
3526         unregister_netdevice_notifier(&cma_nb);
3527         rdma_addr_unregister_client(&addr_client);
3528         ib_sa_unregister_client(&sa_client);
3529         destroy_workqueue(cma_wq);
3530         return ret;
3531 }
3532
3533 static void __exit cma_cleanup(void)
3534 {
3535         ibnl_remove_client(RDMA_NL_RDMA_CM);
3536         ib_unregister_client(&cma_client);
3537         unregister_netdevice_notifier(&cma_nb);
3538         rdma_addr_unregister_client(&addr_client);
3539         ib_sa_unregister_client(&sa_client);
3540         destroy_workqueue(cma_wq);
3541         idr_destroy(&sdp_ps);
3542         idr_destroy(&tcp_ps);
3543         idr_destroy(&udp_ps);
3544         idr_destroy(&ipoib_ps);
3545         idr_destroy(&ib_ps);
3546 }
3547
3548 module_init(cma_init);
3549 module_exit(cma_cleanup);