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.h>
  54 #include <rdma/ib_cache.h>
  55 #include <rdma/ib_cm.h>
  56 #include <rdma/ib_sa.h>
  57 #include <rdma/iw_cm.h>
  58
  59 MODULE_AUTHOR("Sean Hefty");
  60 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  61 MODULE_LICENSE("Dual BSD/GPL");
  62
  63 #define CMA_CM_RESPONSE_TIMEOUT 20
  64 #define CMA_MAX_CM_RETRIES 15
  65 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  66 #define CMA_IBOE_PACKET_LIFETIME 18
  67
  68 static void cma_add_one(struct ib_device *device);
  69 static void cma_remove_one(struct ib_device *device);
  70
  71 static struct ib_client cma_client = {
  72         .name   = "cma",
  73         .add    = cma_add_one,
  74         .remove = cma_remove_one
  75 };
  76
  77 static struct ib_sa_client sa_client;
  78 static struct rdma_addr_client addr_client;
  79 static LIST_HEAD(dev_list);
  80 static LIST_HEAD(listen_any_list);
  81 static DEFINE_MUTEX(lock);
  82 static struct workqueue_struct *cma_wq;
  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 #define CMA_VERSION 0x00
 199
 200 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 201 {
 202         unsigned long flags;
 203         int ret;
 204
 205         spin_lock_irqsave(&id_priv->lock, flags);
 206         ret = (id_priv->state == comp);
 207         spin_unlock_irqrestore(&id_priv->lock, flags);
 208         return ret;
 209 }
 210
 211 static int cma_comp_exch(struct rdma_id_private *id_priv,
 212                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 213 {
 214         unsigned long flags;
 215         int ret;
 216
 217         spin_lock_irqsave(&id_priv->lock, flags);
 218         if ((ret = (id_priv->state == comp)))
 219                 id_priv->state = exch;
 220         spin_unlock_irqrestore(&id_priv->lock, flags);
 221         return ret;
 222 }
 223
 224 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 225                                    enum rdma_cm_state exch)
 226 {
 227         unsigned long flags;
 228         enum rdma_cm_state old;
 229
 230         spin_lock_irqsave(&id_priv->lock, flags);
 231         old = id_priv->state;
 232         id_priv->state = exch;
 233         spin_unlock_irqrestore(&id_priv->lock, flags);
 234         return old;
 235 }
 236
 237 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 238 {
 239         return hdr->ip_version >> 4;
 240 }
 241
 242 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 243 {
 244         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 245 }
 246
 247 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 248                               struct cma_device *cma_dev)
 249 {
 250         atomic_inc(&cma_dev->refcount);
 251         id_priv->cma_dev = cma_dev;
 252         id_priv->id.device = cma_dev->device;
 253         id_priv->id.route.addr.dev_addr.transport =
 254                 rdma_node_get_transport(cma_dev->device->node_type);
 255         list_add_tail(&id_priv->list, &cma_dev->id_list);
 256 }
 257
 258 static inline void cma_deref_dev(struct cma_device *cma_dev)
 259 {
 260         if (atomic_dec_and_test(&cma_dev->refcount))
 261                 complete(&cma_dev->comp);
 262 }
 263
 264 static inline void release_mc(struct kref *kref)
 265 {
 266         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 267
 268         kfree(mc->multicast.ib);
 269         kfree(mc);
 270 }
 271
 272 static void cma_release_dev(struct rdma_id_private *id_priv)
 273 {
 274         mutex_lock(&lock);
 275         list_del(&id_priv->list);
 276         cma_deref_dev(id_priv->cma_dev);
 277         id_priv->cma_dev = NULL;
 278         mutex_unlock(&lock);
 279 }
 280
 281 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 282 {
 283         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 284 }
 285
 286 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 287 {
 288         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 289 }
 290
 291 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 292 {
 293         return id_priv->id.route.addr.src_addr.ss_family;
 294 }
 295
 296 static int cma_set_qkey(struct rdma_id_private *id_priv)
 297 {
 298         struct ib_sa_mcmember_rec rec;
 299         int ret = 0;
 300
 301         if (id_priv->qkey)
 302                 return 0;
 303
 304         switch (id_priv->id.ps) {
 305         case RDMA_PS_UDP:
 306                 id_priv->qkey = RDMA_UDP_QKEY;
 307                 break;
 308         case RDMA_PS_IPOIB:
 309                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 310                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 311                                              id_priv->id.port_num, &rec.mgid,
 312                                              &rec);
 313                 if (!ret)
 314                         id_priv->qkey = be32_to_cpu(rec.qkey);
 315                 break;
 316         default:
 317                 break;
 318         }
 319         return ret;
 320 }
 321
 322 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
 323 {
 324         int i;
 325         int err;
 326         struct ib_port_attr props;
 327         union ib_gid tmp;
 328
 329         err = ib_query_port(device, port_num, &props);
 330         if (err)
 331                 return err;
 332
 333         for (i = 0; i < props.gid_tbl_len; ++i) {
 334                 err = ib_query_gid(device, port_num, i, &tmp);
 335                 if (err)
 336                         return err;
 337                 if (!memcmp(&tmp, gid, sizeof tmp))
 338                         return 0;
 339         }
 340
 341         return -EADDRNOTAVAIL;
 342 }
 343
 344 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 345 {
 346         dev_addr->dev_type = ARPHRD_INFINIBAND;
 347         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 348         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 349 }
 350
 351 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 352 {
 353         int ret;
 354
 355         if (addr->sa_family != AF_IB) {
 356                 ret = rdma_translate_ip(addr, dev_addr);
 357         } else {
 358                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 359                 ret = 0;
 360         }
 361
 362         return ret;
 363 }
 364
 365 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 366 {
 367         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 368         struct cma_device *cma_dev;
 369         union ib_gid gid, iboe_gid;
 370         int ret = -ENODEV;
 371         u8 port;
 372         enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
 373                 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
 374
 375         if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
 376             id_priv->id.ps == RDMA_PS_IPOIB)
 377                 return -EINVAL;
 378
 379         mutex_lock(&lock);
 380         iboe_addr_get_sgid(dev_addr, &iboe_gid);
 381         memcpy(&gid, dev_addr->src_dev_addr +
 382                rdma_addr_gid_offset(dev_addr), sizeof gid);
 383         list_for_each_entry(cma_dev, &dev_list, list) {
 384                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 385                         if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
 386                                 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
 387                                     rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
 388                                         ret = find_gid_port(cma_dev->device, &iboe_gid, port);
 389                                 else
 390                                         ret = find_gid_port(cma_dev->device, &gid, port);
 391
 392                                 if (!ret) {
 393                                         id_priv->id.port_num = port;
 394                                         goto out;
 395                                 }
 396                         }
 397                 }
 398         }
 399
 400 out:
 401         if (!ret)
 402                 cma_attach_to_dev(id_priv, cma_dev);
 403
 404         mutex_unlock(&lock);
 405         return ret;
 406 }
 407
 408 /*
 409  * Select the source IB device and address to reach the destination IB address.
 410  */
 411 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 412 {
 413         struct cma_device *cma_dev, *cur_dev;
 414         struct sockaddr_ib *addr;
 415         union ib_gid gid, sgid, *dgid;
 416         u16 pkey, index;
 417         u8 port, p;
 418         int i;
 419
 420         cma_dev = NULL;
 421         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 422         dgid = (union ib_gid *) &addr->sib_addr;
 423         pkey = ntohs(addr->sib_pkey);
 424
 425         list_for_each_entry(cur_dev, &dev_list, list) {
 426                 if (rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
 427                         continue;
 428
 429                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
 430                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 431                                 continue;
 432
 433                         for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, &gid); i++) {
 434                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 435                                         cma_dev = cur_dev;
 436                                         sgid = gid;
 437                                         port = p;
 438                                         goto found;
 439                                 }
 440
 441                                 if (!cma_dev && (gid.global.subnet_prefix ==
 442                                                  dgid->global.subnet_prefix)) {
 443                                         cma_dev = cur_dev;
 444                                         sgid = gid;
 445                                         port = p;
 446                                 }
 447                         }
 448                 }
 449         }
 450
 451         if (!cma_dev)
 452                 return -ENODEV;
 453
 454 found:
 455         cma_attach_to_dev(id_priv, cma_dev);
 456         id_priv->id.port_num = port;
 457         addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
 458         memcpy(&addr->sib_addr, &sgid, sizeof sgid);
 459         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 460         return 0;
 461 }
 462
 463 static void cma_deref_id(struct rdma_id_private *id_priv)
 464 {
 465         if (atomic_dec_and_test(&id_priv->refcount))
 466                 complete(&id_priv->comp);
 467 }
 468
 469 static int cma_disable_callback(struct rdma_id_private *id_priv,
 470                                 enum rdma_cm_state state)
 471 {
 472         mutex_lock(&id_priv->handler_mutex);
 473         if (id_priv->state != state) {
 474                 mutex_unlock(&id_priv->handler_mutex);
 475                 return -EINVAL;
 476         }
 477         return 0;
 478 }
 479
 480 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 481                                   void *context, enum rdma_port_space ps,
 482                                   enum ib_qp_type qp_type)
 483 {
 484         struct rdma_id_private *id_priv;
 485
 486         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 487         if (!id_priv)
 488                 return ERR_PTR(-ENOMEM);
 489
 490         id_priv->owner = task_pid_nr(current);
 491         id_priv->state = RDMA_CM_IDLE;
 492         id_priv->id.context = context;
 493         id_priv->id.event_handler = event_handler;
 494         id_priv->id.ps = ps;
 495         id_priv->id.qp_type = qp_type;
 496         spin_lock_init(&id_priv->lock);
 497         mutex_init(&id_priv->qp_mutex);
 498         init_completion(&id_priv->comp);
 499         atomic_set(&id_priv->refcount, 1);
 500         mutex_init(&id_priv->handler_mutex);
 501         INIT_LIST_HEAD(&id_priv->listen_list);
 502         INIT_LIST_HEAD(&id_priv->mc_list);
 503         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 504
 505         return &id_priv->id;
 506 }
 507 EXPORT_SYMBOL(rdma_create_id);
 508
 509 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 510 {
 511         struct ib_qp_attr qp_attr;
 512         int qp_attr_mask, ret;
 513
 514         qp_attr.qp_state = IB_QPS_INIT;
 515         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 516         if (ret)
 517                 return ret;
 518
 519         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 520         if (ret)
 521                 return ret;
 522
 523         qp_attr.qp_state = IB_QPS_RTR;
 524         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 525         if (ret)
 526                 return ret;
 527
 528         qp_attr.qp_state = IB_QPS_RTS;
 529         qp_attr.sq_psn = 0;
 530         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 531
 532         return ret;
 533 }
 534
 535 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 536 {
 537         struct ib_qp_attr qp_attr;
 538         int qp_attr_mask, ret;
 539
 540         qp_attr.qp_state = IB_QPS_INIT;
 541         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 542         if (ret)
 543                 return ret;
 544
 545         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 546 }
 547
 548 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 549                    struct ib_qp_init_attr *qp_init_attr)
 550 {
 551         struct rdma_id_private *id_priv;
 552         struct ib_qp *qp;
 553         int ret;
 554
 555         id_priv = container_of(id, struct rdma_id_private, id);
 556         if (id->device != pd->device)
 557                 return -EINVAL;
 558
 559         qp = ib_create_qp(pd, qp_init_attr);
 560         if (IS_ERR(qp))
 561                 return PTR_ERR(qp);
 562
 563         if (id->qp_type == IB_QPT_UD)
 564                 ret = cma_init_ud_qp(id_priv, qp);
 565         else
 566                 ret = cma_init_conn_qp(id_priv, qp);
 567         if (ret)
 568                 goto err;
 569
 570         id->qp = qp;
 571         id_priv->qp_num = qp->qp_num;
 572         id_priv->srq = (qp->srq != NULL);
 573         return 0;
 574 err:
 575         ib_destroy_qp(qp);
 576         return ret;
 577 }
 578 EXPORT_SYMBOL(rdma_create_qp);
 579
 580 void rdma_destroy_qp(struct rdma_cm_id *id)
 581 {
 582         struct rdma_id_private *id_priv;
 583
 584         id_priv = container_of(id, struct rdma_id_private, id);
 585         mutex_lock(&id_priv->qp_mutex);
 586         ib_destroy_qp(id_priv->id.qp);
 587         id_priv->id.qp = NULL;
 588         mutex_unlock(&id_priv->qp_mutex);
 589 }
 590 EXPORT_SYMBOL(rdma_destroy_qp);
 591
 592 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 593                              struct rdma_conn_param *conn_param)
 594 {
 595         struct ib_qp_attr qp_attr;
 596         int qp_attr_mask, ret;
 597
 598         mutex_lock(&id_priv->qp_mutex);
 599         if (!id_priv->id.qp) {
 600                 ret = 0;
 601                 goto out;
 602         }
 603
 604         /* Need to update QP attributes from default values. */
 605         qp_attr.qp_state = IB_QPS_INIT;
 606         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 607         if (ret)
 608                 goto out;
 609
 610         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 611         if (ret)
 612                 goto out;
 613
 614         qp_attr.qp_state = IB_QPS_RTR;
 615         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 616         if (ret)
 617                 goto out;
 618
 619         if (conn_param)
 620                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 621         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 622 out:
 623         mutex_unlock(&id_priv->qp_mutex);
 624         return ret;
 625 }
 626
 627 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 628                              struct rdma_conn_param *conn_param)
 629 {
 630         struct ib_qp_attr qp_attr;
 631         int qp_attr_mask, ret;
 632
 633         mutex_lock(&id_priv->qp_mutex);
 634         if (!id_priv->id.qp) {
 635                 ret = 0;
 636                 goto out;
 637         }
 638
 639         qp_attr.qp_state = IB_QPS_RTS;
 640         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 641         if (ret)
 642                 goto out;
 643
 644         if (conn_param)
 645                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 646         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 647 out:
 648         mutex_unlock(&id_priv->qp_mutex);
 649         return ret;
 650 }
 651
 652 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 653 {
 654         struct ib_qp_attr qp_attr;
 655         int ret;
 656
 657         mutex_lock(&id_priv->qp_mutex);
 658         if (!id_priv->id.qp) {
 659                 ret = 0;
 660                 goto out;
 661         }
 662
 663         qp_attr.qp_state = IB_QPS_ERR;
 664         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 665 out:
 666         mutex_unlock(&id_priv->qp_mutex);
 667         return ret;
 668 }
 669
 670 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 671                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 672 {
 673         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 674         int ret;
 675         u16 pkey;
 676
 677         if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
 678             IB_LINK_LAYER_INFINIBAND)
 679                 pkey = ib_addr_get_pkey(dev_addr);
 680         else
 681                 pkey = 0xffff;
 682
 683         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 684                                   pkey, &qp_attr->pkey_index);
 685         if (ret)
 686                 return ret;
 687
 688         qp_attr->port_num = id_priv->id.port_num;
 689         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 690
 691         if (id_priv->id.qp_type == IB_QPT_UD) {
 692                 ret = cma_set_qkey(id_priv);
 693                 if (ret)
 694                         return ret;
 695
 696                 qp_attr->qkey = id_priv->qkey;
 697                 *qp_attr_mask |= IB_QP_QKEY;
 698         } else {
 699                 qp_attr->qp_access_flags = 0;
 700                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 701         }
 702         return 0;
 703 }
 704
 705 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 706                        int *qp_attr_mask)
 707 {
 708         struct rdma_id_private *id_priv;
 709         int ret = 0;
 710
 711         id_priv = container_of(id, struct rdma_id_private, id);
 712         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 713         case RDMA_TRANSPORT_IB:
 714                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 715                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 716                 else
 717                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 718                                                  qp_attr_mask);
 719                 if (qp_attr->qp_state == IB_QPS_RTR)
 720                         qp_attr->rq_psn = id_priv->seq_num;
 721                 break;
 722         case RDMA_TRANSPORT_IWARP:
 723                 if (!id_priv->cm_id.iw) {
 724                         qp_attr->qp_access_flags = 0;
 725                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 726                 } else
 727                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 728                                                  qp_attr_mask);
 729                 break;
 730         default:
 731                 ret = -ENOSYS;
 732                 break;
 733         }
 734
 735         return ret;
 736 }
 737 EXPORT_SYMBOL(rdma_init_qp_attr);
 738
 739 static inline int cma_zero_addr(struct sockaddr *addr)
 740 {
 741         switch (addr->sa_family) {
 742         case AF_INET:
 743                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
 744         case AF_INET6:
 745                 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
 746         case AF_IB:
 747                 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
 748         default:
 749                 return 0;
 750         }
 751 }
 752
 753 static inline int cma_loopback_addr(struct sockaddr *addr)
 754 {
 755         switch (addr->sa_family) {
 756         case AF_INET:
 757                 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 758         case AF_INET6:
 759                 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
 760         case AF_IB:
 761                 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
 762         default:
 763                 return 0;
 764         }
 765 }
 766
 767 static inline int cma_any_addr(struct sockaddr *addr)
 768 {
 769         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 770 }
 771
 772 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
 773 {
 774         if (src->sa_family != dst->sa_family)
 775                 return -1;
 776
 777         switch (src->sa_family) {
 778         case AF_INET:
 779                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
 780                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
 781         case AF_INET6:
 782                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
 783                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
 784         default:
 785                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
 786                                    &((struct sockaddr_ib *) dst)->sib_addr);
 787         }
 788 }
 789
 790 static __be16 cma_port(struct sockaddr *addr)
 791 {
 792         struct sockaddr_ib *sib;
 793
 794         switch (addr->sa_family) {
 795         case AF_INET:
 796                 return ((struct sockaddr_in *) addr)->sin_port;
 797         case AF_INET6:
 798                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 799         case AF_IB:
 800                 sib = (struct sockaddr_ib *) addr;
 801                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
 802                                     be64_to_cpu(sib->sib_sid_mask)));
 803         default:
 804                 return 0;
 805         }
 806 }
 807
 808 static inline int cma_any_port(struct sockaddr *addr)
 809 {
 810         return !cma_port(addr);
 811 }
 812
 813 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 814                             u8 *ip_ver, __be16 *port,
 815                             union cma_ip_addr **src, union cma_ip_addr **dst)
 816 {
 817         if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 818                 return -EINVAL;
 819
 820         *ip_ver = cma_get_ip_ver(hdr);
 821         *port   = ((struct cma_hdr *) hdr)->port;
 822         *src    = &((struct cma_hdr *) hdr)->src_addr;
 823         *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 824
 825         if (*ip_ver != 4 && *ip_ver != 6)
 826                 return -EINVAL;
 827         return 0;
 828 }
 829
 830 static void cma_save_net_info(struct rdma_addr *addr,
 831                               struct rdma_addr *listen_addr,
 832                               u8 ip_ver, __be16 port,
 833                               union cma_ip_addr *src, union cma_ip_addr *dst)
 834 {
 835         struct sockaddr_in *listen4, *ip4;
 836         struct sockaddr_in6 *listen6, *ip6;
 837
 838         switch (ip_ver) {
 839         case 4:
 840                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 841                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 842                 ip4->sin_family = listen4->sin_family;
 843                 ip4->sin_addr.s_addr = dst->ip4.addr;
 844                 ip4->sin_port = listen4->sin_port;
 845
 846                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 847                 ip4->sin_family = listen4->sin_family;
 848                 ip4->sin_addr.s_addr = src->ip4.addr;
 849                 ip4->sin_port = port;
 850                 break;
 851         case 6:
 852                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 853                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 854                 ip6->sin6_family = listen6->sin6_family;
 855                 ip6->sin6_addr = dst->ip6;
 856                 ip6->sin6_port = listen6->sin6_port;
 857
 858                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 859                 ip6->sin6_family = listen6->sin6_family;
 860                 ip6->sin6_addr = src->ip6;
 861                 ip6->sin6_port = port;
 862                 break;
 863         default:
 864                 break;
 865         }
 866 }
 867
 868 static inline int cma_user_data_offset(enum rdma_port_space ps)
 869 {
 870         return sizeof(struct cma_hdr);
 871 }
 872
 873 static void cma_cancel_route(struct rdma_id_private *id_priv)
 874 {
 875         switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
 876         case IB_LINK_LAYER_INFINIBAND:
 877                 if (id_priv->query)
 878                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 879                 break;
 880         default:
 881                 break;
 882         }
 883 }
 884
 885 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 886 {
 887         struct rdma_id_private *dev_id_priv;
 888
 889         /*
 890          * Remove from listen_any_list to prevent added devices from spawning
 891          * additional listen requests.
 892          */
 893         mutex_lock(&lock);
 894         list_del(&id_priv->list);
 895
 896         while (!list_empty(&id_priv->listen_list)) {
 897                 dev_id_priv = list_entry(id_priv->listen_list.next,
 898                                          struct rdma_id_private, listen_list);
 899                 /* sync with device removal to avoid duplicate destruction */
 900                 list_del_init(&dev_id_priv->list);
 901                 list_del(&dev_id_priv->listen_list);
 902                 mutex_unlock(&lock);
 903
 904                 rdma_destroy_id(&dev_id_priv->id);
 905                 mutex_lock(&lock);
 906         }
 907         mutex_unlock(&lock);
 908 }
 909
 910 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 911                                  enum rdma_cm_state state)
 912 {
 913         switch (state) {
 914         case RDMA_CM_ADDR_QUERY:
 915                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 916                 break;
 917         case RDMA_CM_ROUTE_QUERY:
 918                 cma_cancel_route(id_priv);
 919                 break;
 920         case RDMA_CM_LISTEN:
 921                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
 922                         cma_cancel_listens(id_priv);
 923                 break;
 924         default:
 925                 break;
 926         }
 927 }
 928
 929 static void cma_release_port(struct rdma_id_private *id_priv)
 930 {
 931         struct rdma_bind_list *bind_list = id_priv->bind_list;
 932
 933         if (!bind_list)
 934                 return;
 935
 936         mutex_lock(&lock);
 937         hlist_del(&id_priv->node);
 938         if (hlist_empty(&bind_list->owners)) {
 939                 idr_remove(bind_list->ps, bind_list->port);
 940                 kfree(bind_list);
 941         }
 942         mutex_unlock(&lock);
 943 }
 944
 945 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 946 {
 947         struct cma_multicast *mc;
 948
 949         while (!list_empty(&id_priv->mc_list)) {
 950                 mc = container_of(id_priv->mc_list.next,
 951                                   struct cma_multicast, list);
 952                 list_del(&mc->list);
 953                 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
 954                 case IB_LINK_LAYER_INFINIBAND:
 955                         ib_sa_free_multicast(mc->multicast.ib);
 956                         kfree(mc);
 957                         break;
 958                 case IB_LINK_LAYER_ETHERNET:
 959                         kref_put(&mc->mcref, release_mc);
 960                         break;
 961                 default:
 962                         break;
 963                 }
 964         }
 965 }
 966
 967 void rdma_destroy_id(struct rdma_cm_id *id)
 968 {
 969         struct rdma_id_private *id_priv;
 970         enum rdma_cm_state state;
 971
 972         id_priv = container_of(id, struct rdma_id_private, id);
 973         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
 974         cma_cancel_operation(id_priv, state);
 975
 976         /*
 977          * Wait for any active callback to finish.  New callbacks will find
 978          * the id_priv state set to destroying and abort.
 979          */
 980         mutex_lock(&id_priv->handler_mutex);
 981         mutex_unlock(&id_priv->handler_mutex);
 982
 983         if (id_priv->cma_dev) {
 984                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 985                 case RDMA_TRANSPORT_IB:
 986                         if (id_priv->cm_id.ib)
 987                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 988                         break;
 989                 case RDMA_TRANSPORT_IWARP:
 990                         if (id_priv->cm_id.iw)
 991                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 992                         break;
 993                 default:
 994                         break;
 995                 }
 996                 cma_leave_mc_groups(id_priv);
 997                 cma_release_dev(id_priv);
 998         }
 999
1000         cma_release_port(id_priv);
1001         cma_deref_id(id_priv);
1002         wait_for_completion(&id_priv->comp);
1003
1004         if (id_priv->internal_id)
1005                 cma_deref_id(id_priv->id.context);
1006
1007         kfree(id_priv->id.route.path_rec);
1008         kfree(id_priv);
1009 }
1010 EXPORT_SYMBOL(rdma_destroy_id);
1011
1012 static int cma_rep_recv(struct rdma_id_private *id_priv)
1013 {
1014         int ret;
1015
1016         ret = cma_modify_qp_rtr(id_priv, NULL);
1017         if (ret)
1018                 goto reject;
1019
1020         ret = cma_modify_qp_rts(id_priv, NULL);
1021         if (ret)
1022                 goto reject;
1023
1024         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1025         if (ret)
1026                 goto reject;
1027
1028         return 0;
1029 reject:
1030         cma_modify_qp_err(id_priv);
1031         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1032                        NULL, 0, NULL, 0);
1033         return ret;
1034 }
1035
1036 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1037                                    struct ib_cm_rep_event_param *rep_data,
1038                                    void *private_data)
1039 {
1040         event->param.conn.private_data = private_data;
1041         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1042         event->param.conn.responder_resources = rep_data->responder_resources;
1043         event->param.conn.initiator_depth = rep_data->initiator_depth;
1044         event->param.conn.flow_control = rep_data->flow_control;
1045         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1046         event->param.conn.srq = rep_data->srq;
1047         event->param.conn.qp_num = rep_data->remote_qpn;
1048 }
1049
1050 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1051 {
1052         struct rdma_id_private *id_priv = cm_id->context;
1053         struct rdma_cm_event event;
1054         int ret = 0;
1055
1056         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1057                 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1058             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1059                 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1060                 return 0;
1061
1062         memset(&event, 0, sizeof event);
1063         switch (ib_event->event) {
1064         case IB_CM_REQ_ERROR:
1065         case IB_CM_REP_ERROR:
1066                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1067                 event.status = -ETIMEDOUT;
1068                 break;
1069         case IB_CM_REP_RECEIVED:
1070                 if (id_priv->id.qp) {
1071                         event.status = cma_rep_recv(id_priv);
1072                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1073                                                      RDMA_CM_EVENT_ESTABLISHED;
1074                 } else {
1075                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1076                 }
1077                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1078                                        ib_event->private_data);
1079                 break;
1080         case IB_CM_RTU_RECEIVED:
1081         case IB_CM_USER_ESTABLISHED:
1082                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1083                 break;
1084         case IB_CM_DREQ_ERROR:
1085                 event.status = -ETIMEDOUT; /* fall through */
1086         case IB_CM_DREQ_RECEIVED:
1087         case IB_CM_DREP_RECEIVED:
1088                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1089                                    RDMA_CM_DISCONNECT))
1090                         goto out;
1091                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1092                 break;
1093         case IB_CM_TIMEWAIT_EXIT:
1094                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1095                 break;
1096         case IB_CM_MRA_RECEIVED:
1097                 /* ignore event */
1098                 goto out;
1099         case IB_CM_REJ_RECEIVED:
1100                 cma_modify_qp_err(id_priv);
1101                 event.status = ib_event->param.rej_rcvd.reason;
1102                 event.event = RDMA_CM_EVENT_REJECTED;
1103                 event.param.conn.private_data = ib_event->private_data;
1104                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1105                 break;
1106         default:
1107                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1108                        ib_event->event);
1109                 goto out;
1110         }
1111
1112         ret = id_priv->id.event_handler(&id_priv->id, &event);
1113         if (ret) {
1114                 /* Destroy the CM ID by returning a non-zero value. */
1115                 id_priv->cm_id.ib = NULL;
1116                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1117                 mutex_unlock(&id_priv->handler_mutex);
1118                 rdma_destroy_id(&id_priv->id);
1119                 return ret;
1120         }
1121 out:
1122         mutex_unlock(&id_priv->handler_mutex);
1123         return ret;
1124 }
1125
1126 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1127                                                struct ib_cm_event *ib_event)
1128 {
1129         struct rdma_id_private *id_priv;
1130         struct rdma_cm_id *id;
1131         struct rdma_route *rt;
1132         union cma_ip_addr *src, *dst;
1133         __be16 port;
1134         u8 ip_ver;
1135         int ret;
1136
1137         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1138                              &ip_ver, &port, &src, &dst))
1139                 return NULL;
1140
1141         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1142                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1143         if (IS_ERR(id))
1144                 return NULL;
1145
1146         id_priv = container_of(id, struct rdma_id_private, id);
1147         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1148                           ip_ver, port, src, dst);
1149
1150         rt = &id->route;
1151         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1152         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1153                                GFP_KERNEL);
1154         if (!rt->path_rec)
1155                 goto err;
1156
1157         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1158         if (rt->num_paths == 2)
1159                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1160
1161         if (cma_any_addr(cma_src_addr(id_priv))) {
1162                 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1163                 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1164                 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1165         } else {
1166                 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1167                 if (ret)
1168                         goto err;
1169         }
1170         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1171
1172         id_priv->state = RDMA_CM_CONNECT;
1173         return id_priv;
1174
1175 err:
1176         rdma_destroy_id(id);
1177         return NULL;
1178 }
1179
1180 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1181                                               struct ib_cm_event *ib_event)
1182 {
1183         struct rdma_id_private *id_priv;
1184         struct rdma_cm_id *id;
1185         union cma_ip_addr *src, *dst;
1186         __be16 port;
1187         u8 ip_ver;
1188         int ret;
1189
1190         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1191                             listen_id->ps, IB_QPT_UD);
1192         if (IS_ERR(id))
1193                 return NULL;
1194
1195         id_priv = container_of(id, struct rdma_id_private, id);
1196         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1197                              &ip_ver, &port, &src, &dst))
1198                 goto err;
1199
1200         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1201                           ip_ver, port, src, dst);
1202
1203         if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1204                 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1205                 if (ret)
1206                         goto err;
1207         }
1208
1209         id_priv->state = RDMA_CM_CONNECT;
1210         return id_priv;
1211 err:
1212         rdma_destroy_id(id);
1213         return NULL;
1214 }
1215
1216 static void cma_set_req_event_data(struct rdma_cm_event *event,
1217                                    struct ib_cm_req_event_param *req_data,
1218                                    void *private_data, int offset)
1219 {
1220         event->param.conn.private_data = private_data + offset;
1221         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1222         event->param.conn.responder_resources = req_data->responder_resources;
1223         event->param.conn.initiator_depth = req_data->initiator_depth;
1224         event->param.conn.flow_control = req_data->flow_control;
1225         event->param.conn.retry_count = req_data->retry_count;
1226         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1227         event->param.conn.srq = req_data->srq;
1228         event->param.conn.qp_num = req_data->remote_qpn;
1229 }
1230
1231 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1232 {
1233         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1234                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1235                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1236                  (id->qp_type == IB_QPT_UD)) ||
1237                 (!id->qp_type));
1238 }
1239
1240 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1241 {
1242         struct rdma_id_private *listen_id, *conn_id;
1243         struct rdma_cm_event event;
1244         int offset, ret;
1245
1246         listen_id = cm_id->context;
1247         if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1248                 return -EINVAL;
1249
1250         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1251                 return -ECONNABORTED;
1252
1253         memset(&event, 0, sizeof event);
1254         offset = cma_user_data_offset(listen_id->id.ps);
1255         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1256         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1257                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1258                 event.param.ud.private_data = ib_event->private_data + offset;
1259                 event.param.ud.private_data_len =
1260                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1261         } else {
1262                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1263                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1264                                        ib_event->private_data, offset);
1265         }
1266         if (!conn_id) {
1267                 ret = -ENOMEM;
1268                 goto err1;
1269         }
1270
1271         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1272         ret = cma_acquire_dev(conn_id);
1273         if (ret)
1274                 goto err2;
1275
1276         conn_id->cm_id.ib = cm_id;
1277         cm_id->context = conn_id;
1278         cm_id->cm_handler = cma_ib_handler;
1279
1280         /*
1281          * Protect against the user destroying conn_id from another thread
1282          * until we're done accessing it.
1283          */
1284         atomic_inc(&conn_id->refcount);
1285         ret = conn_id->id.event_handler(&conn_id->id, &event);
1286         if (ret)
1287                 goto err3;
1288
1289         /*
1290          * Acquire mutex to prevent user executing rdma_destroy_id()
1291          * while we're accessing the cm_id.
1292          */
1293         mutex_lock(&lock);
1294         if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1295                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1296         mutex_unlock(&lock);
1297         mutex_unlock(&conn_id->handler_mutex);
1298         mutex_unlock(&listen_id->handler_mutex);
1299         cma_deref_id(conn_id);
1300         return 0;
1301
1302 err3:
1303         cma_deref_id(conn_id);
1304         /* Destroy the CM ID by returning a non-zero value. */
1305         conn_id->cm_id.ib = NULL;
1306 err2:
1307         cma_exch(conn_id, RDMA_CM_DESTROYING);
1308         mutex_unlock(&conn_id->handler_mutex);
1309 err1:
1310         mutex_unlock(&listen_id->handler_mutex);
1311         if (conn_id)
1312                 rdma_destroy_id(&conn_id->id);
1313         return ret;
1314 }
1315
1316 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1317 {
1318         if (addr->sa_family == AF_IB)
1319                 return ((struct sockaddr_ib *) addr)->sib_sid;
1320
1321         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1322 }
1323
1324 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1325                                  struct ib_cm_compare_data *compare)
1326 {
1327         struct cma_hdr *cma_data, *cma_mask;
1328         __be32 ip4_addr;
1329         struct in6_addr ip6_addr;
1330
1331         memset(compare, 0, sizeof *compare);
1332         cma_data = (void *) compare->data;
1333         cma_mask = (void *) compare->mask;
1334
1335         switch (addr->sa_family) {
1336         case AF_INET:
1337                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1338                 cma_set_ip_ver(cma_data, 4);
1339                 cma_set_ip_ver(cma_mask, 0xF);
1340                 if (!cma_any_addr(addr)) {
1341                         cma_data->dst_addr.ip4.addr = ip4_addr;
1342                         cma_mask->dst_addr.ip4.addr = htonl(~0);
1343                 }
1344                 break;
1345         case AF_INET6:
1346                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1347                 cma_set_ip_ver(cma_data, 6);
1348                 cma_set_ip_ver(cma_mask, 0xF);
1349                 if (!cma_any_addr(addr)) {
1350                         cma_data->dst_addr.ip6 = ip6_addr;
1351                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1352                                sizeof cma_mask->dst_addr.ip6);
1353                 }
1354                 break;
1355         default:
1356                 break;
1357         }
1358 }
1359
1360 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1361 {
1362         struct rdma_id_private *id_priv = iw_id->context;
1363         struct rdma_cm_event event;
1364         struct sockaddr_in *sin;
1365         int ret = 0;
1366
1367         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1368                 return 0;
1369
1370         memset(&event, 0, sizeof event);
1371         switch (iw_event->event) {
1372         case IW_CM_EVENT_CLOSE:
1373                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1374                 break;
1375         case IW_CM_EVENT_CONNECT_REPLY:
1376                 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1377                 *sin = iw_event->local_addr;
1378                 sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
1379                 *sin = iw_event->remote_addr;
1380                 switch (iw_event->status) {
1381                 case 0:
1382                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1383                         event.param.conn.initiator_depth = iw_event->ird;
1384                         event.param.conn.responder_resources = iw_event->ord;
1385                         break;
1386                 case -ECONNRESET:
1387                 case -ECONNREFUSED:
1388                         event.event = RDMA_CM_EVENT_REJECTED;
1389                         break;
1390                 case -ETIMEDOUT:
1391                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1392                         break;
1393                 default:
1394                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1395                         break;
1396                 }
1397                 break;
1398         case IW_CM_EVENT_ESTABLISHED:
1399                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1400                 event.param.conn.initiator_depth = iw_event->ird;
1401                 event.param.conn.responder_resources = iw_event->ord;
1402                 break;
1403         default:
1404                 BUG_ON(1);
1405         }
1406
1407         event.status = iw_event->status;
1408         event.param.conn.private_data = iw_event->private_data;
1409         event.param.conn.private_data_len = iw_event->private_data_len;
1410         ret = id_priv->id.event_handler(&id_priv->id, &event);
1411         if (ret) {
1412                 /* Destroy the CM ID by returning a non-zero value. */
1413                 id_priv->cm_id.iw = NULL;
1414                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1415                 mutex_unlock(&id_priv->handler_mutex);
1416                 rdma_destroy_id(&id_priv->id);
1417                 return ret;
1418         }
1419
1420         mutex_unlock(&id_priv->handler_mutex);
1421         return ret;
1422 }
1423
1424 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1425                                struct iw_cm_event *iw_event)
1426 {
1427         struct rdma_cm_id *new_cm_id;
1428         struct rdma_id_private *listen_id, *conn_id;
1429         struct sockaddr_in *sin;
1430         struct net_device *dev = NULL;
1431         struct rdma_cm_event event;
1432         int ret;
1433         struct ib_device_attr attr;
1434
1435         listen_id = cm_id->context;
1436         if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1437                 return -ECONNABORTED;
1438
1439         /* Create a new RDMA id for the new IW CM ID */
1440         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1441                                    listen_id->id.context,
1442                                    RDMA_PS_TCP, IB_QPT_RC);
1443         if (IS_ERR(new_cm_id)) {
1444                 ret = -ENOMEM;
1445                 goto out;
1446         }
1447         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1448         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1449         conn_id->state = RDMA_CM_CONNECT;
1450
1451         dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1452         if (!dev) {
1453                 ret = -EADDRNOTAVAIL;
1454                 mutex_unlock(&conn_id->handler_mutex);
1455                 rdma_destroy_id(new_cm_id);
1456                 goto out;
1457         }
1458         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1459         if (ret) {
1460                 mutex_unlock(&conn_id->handler_mutex);
1461                 rdma_destroy_id(new_cm_id);
1462                 goto out;
1463         }
1464
1465         ret = cma_acquire_dev(conn_id);
1466         if (ret) {
1467                 mutex_unlock(&conn_id->handler_mutex);
1468                 rdma_destroy_id(new_cm_id);
1469                 goto out;
1470         }
1471
1472         conn_id->cm_id.iw = cm_id;
1473         cm_id->context = conn_id;
1474         cm_id->cm_handler = cma_iw_handler;
1475
1476         sin = (struct sockaddr_in *) cma_src_addr(conn_id);
1477         *sin = iw_event->local_addr;
1478         sin = (struct sockaddr_in *) cma_dst_addr(conn_id);
1479         *sin = iw_event->remote_addr;
1480
1481         ret = ib_query_device(conn_id->id.device, &attr);
1482         if (ret) {
1483                 mutex_unlock(&conn_id->handler_mutex);
1484                 rdma_destroy_id(new_cm_id);
1485                 goto out;
1486         }
1487
1488         memset(&event, 0, sizeof event);
1489         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1490         event.param.conn.private_data = iw_event->private_data;
1491         event.param.conn.private_data_len = iw_event->private_data_len;
1492         event.param.conn.initiator_depth = iw_event->ird;
1493         event.param.conn.responder_resources = iw_event->ord;
1494
1495         /*
1496          * Protect against the user destroying conn_id from another thread
1497          * until we're done accessing it.
1498          */
1499         atomic_inc(&conn_id->refcount);
1500         ret = conn_id->id.event_handler(&conn_id->id, &event);
1501         if (ret) {
1502                 /* User wants to destroy the CM ID */
1503                 conn_id->cm_id.iw = NULL;
1504                 cma_exch(conn_id, RDMA_CM_DESTROYING);
1505                 mutex_unlock(&conn_id->handler_mutex);
1506                 cma_deref_id(conn_id);
1507                 rdma_destroy_id(&conn_id->id);
1508                 goto out;
1509         }
1510
1511         mutex_unlock(&conn_id->handler_mutex);
1512         cma_deref_id(conn_id);
1513
1514 out:
1515         if (dev)
1516                 dev_put(dev);
1517         mutex_unlock(&listen_id->handler_mutex);
1518         return ret;
1519 }
1520
1521 static int cma_ib_listen(struct rdma_id_private *id_priv)
1522 {
1523         struct ib_cm_compare_data compare_data;
1524         struct sockaddr *addr;
1525         struct ib_cm_id *id;
1526         __be64 svc_id;
1527         int ret;
1528
1529         id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1530         if (IS_ERR(id))
1531                 return PTR_ERR(id);
1532
1533         id_priv->cm_id.ib = id;
1534
1535         addr = cma_src_addr(id_priv);
1536         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1537         if (cma_any_addr(addr) && !id_priv->afonly)
1538                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1539         else {
1540                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1541                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1542         }
1543
1544         if (ret) {
1545                 ib_destroy_cm_id(id_priv->cm_id.ib);
1546                 id_priv->cm_id.ib = NULL;
1547         }
1548
1549         return ret;
1550 }
1551
1552 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1553 {
1554         int ret;
1555         struct sockaddr_in *sin;
1556         struct iw_cm_id *id;
1557
1558         id = iw_create_cm_id(id_priv->id.device,
1559                              iw_conn_req_handler,
1560                              id_priv);
1561         if (IS_ERR(id))
1562                 return PTR_ERR(id);
1563
1564         id_priv->cm_id.iw = id;
1565
1566         sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1567         id_priv->cm_id.iw->local_addr = *sin;
1568
1569         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1570
1571         if (ret) {
1572                 iw_destroy_cm_id(id_priv->cm_id.iw);
1573                 id_priv->cm_id.iw = NULL;
1574         }
1575
1576         return ret;
1577 }
1578
1579 static int cma_listen_handler(struct rdma_cm_id *id,
1580                               struct rdma_cm_event *event)
1581 {
1582         struct rdma_id_private *id_priv = id->context;
1583
1584         id->context = id_priv->id.context;
1585         id->event_handler = id_priv->id.event_handler;
1586         return id_priv->id.event_handler(id, event);
1587 }
1588
1589 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1590                               struct cma_device *cma_dev)
1591 {
1592         struct rdma_id_private *dev_id_priv;
1593         struct rdma_cm_id *id;
1594         int ret;
1595
1596         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1597                             id_priv->id.qp_type);
1598         if (IS_ERR(id))
1599                 return;
1600
1601         dev_id_priv = container_of(id, struct rdma_id_private, id);
1602
1603         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1604         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1605                rdma_addr_size(cma_src_addr(id_priv)));
1606
1607         cma_attach_to_dev(dev_id_priv, cma_dev);
1608         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1609         atomic_inc(&id_priv->refcount);
1610         dev_id_priv->internal_id = 1;
1611         dev_id_priv->afonly = id_priv->afonly;
1612
1613         ret = rdma_listen(id, id_priv->backlog);
1614         if (ret)
1615                 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1616                        "listening on device %s\n", ret, cma_dev->device->name);
1617 }
1618
1619 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1620 {
1621         struct cma_device *cma_dev;
1622
1623         mutex_lock(&lock);
1624         list_add_tail(&id_priv->list, &listen_any_list);
1625         list_for_each_entry(cma_dev, &dev_list, list)
1626                 cma_listen_on_dev(id_priv, cma_dev);
1627         mutex_unlock(&lock);
1628 }
1629
1630 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1631 {
1632         struct rdma_id_private *id_priv;
1633
1634         id_priv = container_of(id, struct rdma_id_private, id);
1635         id_priv->tos = (u8) tos;
1636 }
1637 EXPORT_SYMBOL(rdma_set_service_type);
1638
1639 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1640                               void *context)
1641 {
1642         struct cma_work *work = context;
1643         struct rdma_route *route;
1644
1645         route = &work->id->id.route;
1646
1647         if (!status) {
1648                 route->num_paths = 1;
1649                 *route->path_rec = *path_rec;
1650         } else {
1651                 work->old_state = RDMA_CM_ROUTE_QUERY;
1652                 work->new_state = RDMA_CM_ADDR_RESOLVED;
1653                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1654                 work->event.status = status;
1655         }
1656
1657         queue_work(cma_wq, &work->work);
1658 }
1659
1660 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1661                               struct cma_work *work)
1662 {
1663         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1664         struct ib_sa_path_rec path_rec;
1665         ib_sa_comp_mask comp_mask;
1666         struct sockaddr_in6 *sin6;
1667         struct sockaddr_ib *sib;
1668
1669         memset(&path_rec, 0, sizeof path_rec);
1670         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1671         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1672         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1673         path_rec.numb_path = 1;
1674         path_rec.reversible = 1;
1675         path_rec.service_id = cma_get_service_id(id_priv->id.ps, cma_dst_addr(id_priv));
1676
1677         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1678                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1679                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1680
1681         switch (cma_family(id_priv)) {
1682         case AF_INET:
1683                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1684                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1685                 break;
1686         case AF_INET6:
1687                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1688                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1689                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1690                 break;
1691         case AF_IB:
1692                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1693                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1694                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1695                 break;
1696         }
1697
1698         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1699                                                id_priv->id.port_num, &path_rec,
1700                                                comp_mask, timeout_ms,
1701                                                GFP_KERNEL, cma_query_handler,
1702                                                work, &id_priv->query);
1703
1704         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1705 }
1706
1707 static void cma_work_handler(struct work_struct *_work)
1708 {
1709         struct cma_work *work = container_of(_work, struct cma_work, work);
1710         struct rdma_id_private *id_priv = work->id;
1711         int destroy = 0;
1712
1713         mutex_lock(&id_priv->handler_mutex);
1714         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1715                 goto out;
1716
1717         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1718                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1719                 destroy = 1;
1720         }
1721 out:
1722         mutex_unlock(&id_priv->handler_mutex);
1723         cma_deref_id(id_priv);
1724         if (destroy)
1725                 rdma_destroy_id(&id_priv->id);
1726         kfree(work);
1727 }
1728
1729 static void cma_ndev_work_handler(struct work_struct *_work)
1730 {
1731         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1732         struct rdma_id_private *id_priv = work->id;
1733         int destroy = 0;
1734
1735         mutex_lock(&id_priv->handler_mutex);
1736         if (id_priv->state == RDMA_CM_DESTROYING ||
1737             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1738                 goto out;
1739
1740         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1741                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1742                 destroy = 1;
1743         }
1744
1745 out:
1746         mutex_unlock(&id_priv->handler_mutex);
1747         cma_deref_id(id_priv);
1748         if (destroy)
1749                 rdma_destroy_id(&id_priv->id);
1750         kfree(work);
1751 }
1752
1753 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1754 {
1755         struct rdma_route *route = &id_priv->id.route;
1756         struct cma_work *work;
1757         int ret;
1758
1759         work = kzalloc(sizeof *work, GFP_KERNEL);
1760         if (!work)
1761                 return -ENOMEM;
1762
1763         work->id = id_priv;
1764         INIT_WORK(&work->work, cma_work_handler);
1765         work->old_state = RDMA_CM_ROUTE_QUERY;
1766         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1767         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1768
1769         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1770         if (!route->path_rec) {
1771                 ret = -ENOMEM;
1772                 goto err1;
1773         }
1774
1775         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1776         if (ret)
1777                 goto err2;
1778
1779         return 0;
1780 err2:
1781         kfree(route->path_rec);
1782         route->path_rec = NULL;
1783 err1:
1784         kfree(work);
1785         return ret;
1786 }
1787
1788 int rdma_set_ib_paths(struct rdma_cm_id *id,
1789                       struct ib_sa_path_rec *path_rec, int num_paths)
1790 {
1791         struct rdma_id_private *id_priv;
1792         int ret;
1793
1794         id_priv = container_of(id, struct rdma_id_private, id);
1795         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1796                            RDMA_CM_ROUTE_RESOLVED))
1797                 return -EINVAL;
1798
1799         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1800                                      GFP_KERNEL);
1801         if (!id->route.path_rec) {
1802                 ret = -ENOMEM;
1803                 goto err;
1804         }
1805
1806         id->route.num_paths = num_paths;
1807         return 0;
1808 err:
1809         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1810         return ret;
1811 }
1812 EXPORT_SYMBOL(rdma_set_ib_paths);
1813
1814 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1815 {
1816         struct cma_work *work;
1817
1818         work = kzalloc(sizeof *work, GFP_KERNEL);
1819         if (!work)
1820                 return -ENOMEM;
1821
1822         work->id = id_priv;
1823         INIT_WORK(&work->work, cma_work_handler);
1824         work->old_state = RDMA_CM_ROUTE_QUERY;
1825         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1826         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1827         queue_work(cma_wq, &work->work);
1828         return 0;
1829 }
1830
1831 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1832 {
1833         struct rdma_route *route = &id_priv->id.route;
1834         struct rdma_addr *addr = &route->addr;
1835         struct cma_work *work;
1836         int ret;
1837         struct net_device *ndev = NULL;
1838         u16 vid;
1839
1840         work = kzalloc(sizeof *work, GFP_KERNEL);
1841         if (!work)
1842                 return -ENOMEM;
1843
1844         work->id = id_priv;
1845         INIT_WORK(&work->work, cma_work_handler);
1846
1847         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1848         if (!route->path_rec) {
1849                 ret = -ENOMEM;
1850                 goto err1;
1851         }
1852
1853         route->num_paths = 1;
1854
1855         if (addr->dev_addr.bound_dev_if)
1856                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1857         if (!ndev) {
1858                 ret = -ENODEV;
1859                 goto err2;
1860         }
1861
1862         vid = rdma_vlan_dev_vlan_id(ndev);
1863
1864         iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1865         iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1866
1867         route->path_rec->hop_limit = 1;
1868         route->path_rec->reversible = 1;
1869         route->path_rec->pkey = cpu_to_be16(0xffff);
1870         route->path_rec->mtu_selector = IB_SA_EQ;
1871         route->path_rec->sl = netdev_get_prio_tc_map(
1872                         ndev->priv_flags & IFF_802_1Q_VLAN ?
1873                                 vlan_dev_real_dev(ndev) : ndev,
1874                         rt_tos2priority(id_priv->tos));
1875
1876         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1877         route->path_rec->rate_selector = IB_SA_EQ;
1878         route->path_rec->rate = iboe_get_rate(ndev);
1879         dev_put(ndev);
1880         route->path_rec->packet_life_time_selector = IB_SA_EQ;
1881         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1882         if (!route->path_rec->mtu) {
1883                 ret = -EINVAL;
1884                 goto err2;
1885         }
1886
1887         work->old_state = RDMA_CM_ROUTE_QUERY;
1888         work->new_state = RDMA_CM_ROUTE_RESOLVED;
1889         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1890         work->event.status = 0;
1891
1892         queue_work(cma_wq, &work->work);
1893
1894         return 0;
1895
1896 err2:
1897         kfree(route->path_rec);
1898         route->path_rec = NULL;
1899 err1:
1900         kfree(work);
1901         return ret;
1902 }
1903
1904 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1905 {
1906         struct rdma_id_private *id_priv;
1907         int ret;
1908
1909         id_priv = container_of(id, struct rdma_id_private, id);
1910         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1911                 return -EINVAL;
1912
1913         atomic_inc(&id_priv->refcount);
1914         switch (rdma_node_get_transport(id->device->node_type)) {
1915         case RDMA_TRANSPORT_IB:
1916                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1917                 case IB_LINK_LAYER_INFINIBAND:
1918                         ret = cma_resolve_ib_route(id_priv, timeout_ms);
1919                         break;
1920                 case IB_LINK_LAYER_ETHERNET:
1921                         ret = cma_resolve_iboe_route(id_priv);
1922                         break;
1923                 default:
1924                         ret = -ENOSYS;
1925                 }
1926                 break;
1927         case RDMA_TRANSPORT_IWARP:
1928                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1929                 break;
1930         default:
1931                 ret = -ENOSYS;
1932                 break;
1933         }
1934         if (ret)
1935                 goto err;
1936
1937         return 0;
1938 err:
1939         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1940         cma_deref_id(id_priv);
1941         return ret;
1942 }
1943 EXPORT_SYMBOL(rdma_resolve_route);
1944
1945 static void cma_set_loopback(struct sockaddr *addr)
1946 {
1947         switch (addr->sa_family) {
1948         case AF_INET:
1949                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1950                 break;
1951         case AF_INET6:
1952                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
1953                               0, 0, 0, htonl(1));
1954                 break;
1955         default:
1956                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
1957                             0, 0, 0, htonl(1));
1958                 break;
1959         }
1960 }
1961
1962 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1963 {
1964         struct cma_device *cma_dev, *cur_dev;
1965         struct ib_port_attr port_attr;
1966         union ib_gid gid;
1967         u16 pkey;
1968         int ret;
1969         u8 p;
1970
1971         cma_dev = NULL;
1972         mutex_lock(&lock);
1973         list_for_each_entry(cur_dev, &dev_list, list) {
1974                 if (cma_family(id_priv) == AF_IB &&
1975                     rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
1976                         continue;
1977
1978                 if (!cma_dev)
1979                         cma_dev = cur_dev;
1980
1981                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
1982                         if (!ib_query_port(cur_dev->device, p, &port_attr) &&
1983                             port_attr.state == IB_PORT_ACTIVE) {
1984                                 cma_dev = cur_dev;
1985                                 goto port_found;
1986                         }
1987                 }
1988         }
1989
1990         if (!cma_dev) {
1991                 ret = -ENODEV;
1992                 goto out;
1993         }
1994
1995         p = 1;
1996
1997 port_found:
1998         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1999         if (ret)
2000                 goto out;
2001
2002         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2003         if (ret)
2004                 goto out;
2005
2006         id_priv->id.route.addr.dev_addr.dev_type =
2007                 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2008                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2009
2010         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2011         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2012         id_priv->id.port_num = p;
2013         cma_attach_to_dev(id_priv, cma_dev);
2014         cma_set_loopback(cma_src_addr(id_priv));
2015 out:
2016         mutex_unlock(&lock);
2017         return ret;
2018 }
2019
2020 static void addr_handler(int status, struct sockaddr *src_addr,
2021                          struct rdma_dev_addr *dev_addr, void *context)
2022 {
2023         struct rdma_id_private *id_priv = context;
2024         struct rdma_cm_event event;
2025
2026         memset(&event, 0, sizeof event);
2027         mutex_lock(&id_priv->handler_mutex);
2028         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2029                            RDMA_CM_ADDR_RESOLVED))
2030                 goto out;
2031
2032         if (!status && !id_priv->cma_dev)
2033                 status = cma_acquire_dev(id_priv);
2034
2035         if (status) {
2036                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2037                                    RDMA_CM_ADDR_BOUND))
2038                         goto out;
2039                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2040                 event.status = status;
2041         } else {
2042                 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2043                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2044         }
2045
2046         if (id_priv->id.event_handler(&id_priv->id, &event)) {
2047                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2048                 mutex_unlock(&id_priv->handler_mutex);
2049                 cma_deref_id(id_priv);
2050                 rdma_destroy_id(&id_priv->id);
2051                 return;
2052         }
2053 out:
2054         mutex_unlock(&id_priv->handler_mutex);
2055         cma_deref_id(id_priv);
2056 }
2057
2058 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2059 {
2060         struct cma_work *work;
2061         union ib_gid gid;
2062         int ret;
2063
2064         work = kzalloc(sizeof *work, GFP_KERNEL);
2065         if (!work)
2066                 return -ENOMEM;
2067
2068         if (!id_priv->cma_dev) {
2069                 ret = cma_bind_loopback(id_priv);
2070                 if (ret)
2071                         goto err;
2072         }
2073
2074         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2075         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2076
2077         work->id = id_priv;
2078         INIT_WORK(&work->work, cma_work_handler);
2079         work->old_state = RDMA_CM_ADDR_QUERY;
2080         work->new_state = RDMA_CM_ADDR_RESOLVED;
2081         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2082         queue_work(cma_wq, &work->work);
2083         return 0;
2084 err:
2085         kfree(work);
2086         return ret;
2087 }
2088
2089 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2090 {
2091         struct cma_work *work;
2092         int ret;
2093
2094         work = kzalloc(sizeof *work, GFP_KERNEL);
2095         if (!work)
2096                 return -ENOMEM;
2097
2098         if (!id_priv->cma_dev) {
2099                 ret = cma_resolve_ib_dev(id_priv);
2100                 if (ret)
2101                         goto err;
2102         }
2103
2104         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2105                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2106
2107         work->id = id_priv;
2108         INIT_WORK(&work->work, cma_work_handler);
2109         work->old_state = RDMA_CM_ADDR_QUERY;
2110         work->new_state = RDMA_CM_ADDR_RESOLVED;
2111         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2112         queue_work(cma_wq, &work->work);
2113         return 0;
2114 err:
2115         kfree(work);
2116         return ret;
2117 }
2118
2119 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2120                          struct sockaddr *dst_addr)
2121 {
2122         if (!src_addr || !src_addr->sa_family) {
2123                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2124                 src_addr->sa_family = dst_addr->sa_family;
2125                 if (dst_addr->sa_family == AF_INET6) {
2126                         ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2127                                 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2128                 } else if (dst_addr->sa_family == AF_IB) {
2129                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
2130                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2131                 }
2132         }
2133         return rdma_bind_addr(id, src_addr);
2134 }
2135
2136 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2137                       struct sockaddr *dst_addr, int timeout_ms)
2138 {
2139         struct rdma_id_private *id_priv;
2140         int ret;
2141
2142         id_priv = container_of(id, struct rdma_id_private, id);
2143         if (id_priv->state == RDMA_CM_IDLE) {
2144                 ret = cma_bind_addr(id, src_addr, dst_addr);
2145                 if (ret)
2146                         return ret;
2147         }
2148
2149         if (cma_family(id_priv) != dst_addr->sa_family)
2150                 return -EINVAL;
2151
2152         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2153                 return -EINVAL;
2154
2155         atomic_inc(&id_priv->refcount);
2156         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2157         if (cma_any_addr(dst_addr)) {
2158                 ret = cma_resolve_loopback(id_priv);
2159         } else {
2160                 if (dst_addr->sa_family == AF_IB) {
2161                         ret = cma_resolve_ib_addr(id_priv);
2162                 } else {
2163                         ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2164                                               dst_addr, &id->route.addr.dev_addr,
2165                                               timeout_ms, addr_handler, id_priv);
2166                 }
2167         }
2168         if (ret)
2169                 goto err;
2170
2171         return 0;
2172 err:
2173         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2174         cma_deref_id(id_priv);
2175         return ret;
2176 }
2177 EXPORT_SYMBOL(rdma_resolve_addr);
2178
2179 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2180 {
2181         struct rdma_id_private *id_priv;
2182         unsigned long flags;
2183         int ret;
2184
2185         id_priv = container_of(id, struct rdma_id_private, id);
2186         spin_lock_irqsave(&id_priv->lock, flags);
2187         if (reuse || id_priv->state == RDMA_CM_IDLE) {
2188                 id_priv->reuseaddr = reuse;
2189                 ret = 0;
2190         } else {
2191                 ret = -EINVAL;
2192         }
2193         spin_unlock_irqrestore(&id_priv->lock, flags);
2194         return ret;
2195 }
2196 EXPORT_SYMBOL(rdma_set_reuseaddr);
2197
2198 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2199 {
2200         struct rdma_id_private *id_priv;
2201         unsigned long flags;
2202         int ret;
2203
2204         id_priv = container_of(id, struct rdma_id_private, id);
2205         spin_lock_irqsave(&id_priv->lock, flags);
2206         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2207                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2208                 id_priv->afonly = afonly;
2209                 ret = 0;
2210         } else {
2211                 ret = -EINVAL;
2212         }
2213         spin_unlock_irqrestore(&id_priv->lock, flags);
2214         return ret;
2215 }
2216 EXPORT_SYMBOL(rdma_set_afonly);
2217
2218 static void cma_bind_port(struct rdma_bind_list *bind_list,
2219                           struct rdma_id_private *id_priv)
2220 {
2221         struct sockaddr *addr;
2222         struct sockaddr_ib *sib;
2223         u64 sid, mask;
2224         __be16 port;
2225
2226         addr = cma_src_addr(id_priv);
2227         port = htons(bind_list->port);
2228
2229         switch (addr->sa_family) {
2230         case AF_INET:
2231                 ((struct sockaddr_in *) addr)->sin_port = port;
2232                 break;
2233         case AF_INET6:
2234                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2235                 break;
2236         case AF_IB:
2237                 sib = (struct sockaddr_ib *) addr;
2238                 sid = be64_to_cpu(sib->sib_sid);
2239                 mask = be64_to_cpu(sib->sib_sid_mask);
2240                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2241                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2242                 break;
2243         }
2244         id_priv->bind_list = bind_list;
2245         hlist_add_head(&id_priv->node, &bind_list->owners);
2246 }
2247
2248 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2249                           unsigned short snum)
2250 {
2251         struct rdma_bind_list *bind_list;
2252         int ret;
2253
2254         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2255         if (!bind_list)
2256                 return -ENOMEM;
2257
2258         ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2259         if (ret < 0)
2260                 goto err;
2261
2262         bind_list->ps = ps;
2263         bind_list->port = (unsigned short)ret;
2264         cma_bind_port(bind_list, id_priv);
2265         return 0;
2266 err:
2267         kfree(bind_list);
2268         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2269 }
2270
2271 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2272 {
2273         static unsigned int last_used_port;
2274         int low, high, remaining;
2275         unsigned int rover;
2276
2277         inet_get_local_port_range(&low, &high);
2278         remaining = (high - low) + 1;
2279         rover = net_random() % remaining + low;
2280 retry:
2281         if (last_used_port != rover &&
2282             !idr_find(ps, (unsigned short) rover)) {
2283                 int ret = cma_alloc_port(ps, id_priv, rover);
2284                 /*
2285                  * Remember previously used port number in order to avoid
2286                  * re-using same port immediately after it is closed.
2287                  */
2288                 if (!ret)
2289                         last_used_port = rover;
2290                 if (ret != -EADDRNOTAVAIL)
2291                         return ret;
2292         }
2293         if (--remaining) {
2294                 rover++;
2295                 if ((rover < low) || (rover > high))
2296                         rover = low;
2297                 goto retry;
2298         }
2299         return -EADDRNOTAVAIL;
2300 }
2301
2302 /*
2303  * Check that the requested port is available.  This is called when trying to
2304  * bind to a specific port, or when trying to listen on a bound port.  In
2305  * the latter case, the provided id_priv may already be on the bind_list, but
2306  * we still need to check that it's okay to start listening.
2307  */
2308 static int cma_check_port(struct rdma_bind_list *bind_list,
2309                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2310 {
2311         struct rdma_id_private *cur_id;
2312         struct sockaddr *addr, *cur_addr;
2313
2314         addr = cma_src_addr(id_priv);
2315         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2316                 if (id_priv == cur_id)
2317                         continue;
2318
2319                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2320                     cur_id->reuseaddr)
2321                         continue;
2322
2323                 cur_addr = cma_src_addr(cur_id);
2324                 if (id_priv->afonly && cur_id->afonly &&
2325                     (addr->sa_family != cur_addr->sa_family))
2326                         continue;
2327
2328                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2329                         return -EADDRNOTAVAIL;
2330
2331                 if (!cma_addr_cmp(addr, cur_addr))
2332                         return -EADDRINUSE;
2333         }
2334         return 0;
2335 }
2336
2337 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2338 {
2339         struct rdma_bind_list *bind_list;
2340         unsigned short snum;
2341         int ret;
2342
2343         snum = ntohs(cma_port(cma_src_addr(id_priv)));
2344         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2345                 return -EACCES;
2346
2347         bind_list = idr_find(ps, snum);
2348         if (!bind_list) {
2349                 ret = cma_alloc_port(ps, id_priv, snum);
2350         } else {
2351                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2352                 if (!ret)
2353                         cma_bind_port(bind_list, id_priv);
2354         }
2355         return ret;
2356 }
2357
2358 static int cma_bind_listen(struct rdma_id_private *id_priv)
2359 {
2360         struct rdma_bind_list *bind_list = id_priv->bind_list;
2361         int ret = 0;
2362
2363         mutex_lock(&lock);
2364         if (bind_list->owners.first->next)
2365                 ret = cma_check_port(bind_list, id_priv, 0);
2366         mutex_unlock(&lock);
2367         return ret;
2368 }
2369
2370 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2371 {
2372         switch (id_priv->id.ps) {
2373         case RDMA_PS_TCP:
2374                 return &tcp_ps;
2375         case RDMA_PS_UDP:
2376                 return &udp_ps;
2377         case RDMA_PS_IPOIB:
2378                 return &ipoib_ps;
2379         case RDMA_PS_IB:
2380                 return &ib_ps;
2381         default:
2382                 return NULL;
2383         }
2384 }
2385
2386 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2387 {
2388         struct idr *ps = NULL;
2389         struct sockaddr_ib *sib;
2390         u64 sid_ps, mask, sid;
2391
2392         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2393         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2394         sid = be64_to_cpu(sib->sib_sid) & mask;
2395
2396         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2397                 sid_ps = RDMA_IB_IP_PS_IB;
2398                 ps = &ib_ps;
2399         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2400                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2401                 sid_ps = RDMA_IB_IP_PS_TCP;
2402                 ps = &tcp_ps;
2403         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2404                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2405                 sid_ps = RDMA_IB_IP_PS_UDP;
2406                 ps = &udp_ps;
2407         }
2408
2409         if (ps) {
2410                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2411                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2412                                                 be64_to_cpu(sib->sib_sid_mask));
2413         }
2414         return ps;
2415 }
2416
2417 static int cma_get_port(struct rdma_id_private *id_priv)
2418 {
2419         struct idr *ps;
2420         int ret;
2421
2422         if (cma_family(id_priv) != AF_IB)
2423                 ps = cma_select_inet_ps(id_priv);
2424         else
2425                 ps = cma_select_ib_ps(id_priv);
2426         if (!ps)
2427                 return -EPROTONOSUPPORT;
2428
2429         mutex_lock(&lock);
2430         if (cma_any_port(cma_src_addr(id_priv)))
2431                 ret = cma_alloc_any_port(ps, id_priv);
2432         else
2433                 ret = cma_use_port(ps, id_priv);
2434         mutex_unlock(&lock);
2435
2436         return ret;
2437 }
2438
2439 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2440                                struct sockaddr *addr)
2441 {
2442 #if IS_ENABLED(CONFIG_IPV6)
2443         struct sockaddr_in6 *sin6;
2444
2445         if (addr->sa_family != AF_INET6)
2446                 return 0;
2447
2448         sin6 = (struct sockaddr_in6 *) addr;
2449         if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2450             !sin6->sin6_scope_id)
2451                         return -EINVAL;
2452
2453         dev_addr->bound_dev_if = sin6->sin6_scope_id;
2454 #endif
2455         return 0;
2456 }
2457
2458 int rdma_listen(struct rdma_cm_id *id, int backlog)
2459 {
2460         struct rdma_id_private *id_priv;
2461         int ret;
2462
2463         id_priv = container_of(id, struct rdma_id_private, id);
2464         if (id_priv->state == RDMA_CM_IDLE) {
2465                 id->route.addr.src_addr.ss_family = AF_INET;
2466                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2467                 if (ret)
2468                         return ret;
2469         }
2470
2471         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2472                 return -EINVAL;
2473
2474         if (id_priv->reuseaddr) {
2475                 ret = cma_bind_listen(id_priv);
2476                 if (ret)
2477                         goto err;
2478         }
2479
2480         id_priv->backlog = backlog;
2481         if (id->device) {
2482                 switch (rdma_node_get_transport(id->device->node_type)) {
2483                 case RDMA_TRANSPORT_IB:
2484                         ret = cma_ib_listen(id_priv);
2485                         if (ret)
2486                                 goto err;
2487                         break;
2488                 case RDMA_TRANSPORT_IWARP:
2489                         ret = cma_iw_listen(id_priv, backlog);
2490                         if (ret)
2491                                 goto err;
2492                         break;
2493                 default:
2494                         ret = -ENOSYS;
2495                         goto err;
2496                 }
2497         } else
2498                 cma_listen_on_all(id_priv);
2499
2500         return 0;
2501 err:
2502         id_priv->backlog = 0;
2503         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2504         return ret;
2505 }
2506 EXPORT_SYMBOL(rdma_listen);
2507
2508 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2509 {
2510         struct rdma_id_private *id_priv;
2511         int ret;
2512
2513         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2514             addr->sa_family != AF_IB)
2515                 return -EAFNOSUPPORT;
2516
2517         id_priv = container_of(id, struct rdma_id_private, id);
2518         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2519                 return -EINVAL;
2520
2521         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2522         if (ret)
2523                 goto err1;
2524
2525         if (!cma_any_addr(addr)) {
2526                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2527                 if (ret)
2528                         goto err1;
2529
2530                 ret = cma_acquire_dev(id_priv);
2531                 if (ret)
2532                         goto err1;
2533         }
2534
2535         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2536         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2537                 if (addr->sa_family == AF_INET)
2538                         id_priv->afonly = 1;
2539 #if IS_ENABLED(CONFIG_IPV6)
2540                 else if (addr->sa_family == AF_INET6)
2541                         id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2542 #endif
2543         }
2544         ret = cma_get_port(id_priv);
2545         if (ret)
2546                 goto err2;
2547
2548         return 0;
2549 err2:
2550         if (id_priv->cma_dev)
2551                 cma_release_dev(id_priv);
2552 err1:
2553         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2554         return ret;
2555 }
2556 EXPORT_SYMBOL(rdma_bind_addr);
2557
2558 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2559 {
2560         struct cma_hdr *cma_hdr;
2561
2562         cma_hdr = hdr;
2563         cma_hdr->cma_version = CMA_VERSION;
2564         if (cma_family(id_priv) == AF_INET) {
2565                 struct sockaddr_in *src4, *dst4;
2566
2567                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2568                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2569
2570                 cma_set_ip_ver(cma_hdr, 4);
2571                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2572                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2573                 cma_hdr->port = src4->sin_port;
2574         } else {
2575                 struct sockaddr_in6 *src6, *dst6;
2576
2577                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2578                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2579
2580                 cma_set_ip_ver(cma_hdr, 6);
2581                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2582                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2583                 cma_hdr->port = src6->sin6_port;
2584         }
2585         return 0;
2586 }
2587
2588 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2589                                 struct ib_cm_event *ib_event)
2590 {
2591         struct rdma_id_private *id_priv = cm_id->context;
2592         struct rdma_cm_event event;
2593         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2594         int ret = 0;
2595
2596         if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2597                 return 0;
2598
2599         memset(&event, 0, sizeof event);
2600         switch (ib_event->event) {
2601         case IB_CM_SIDR_REQ_ERROR:
2602                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2603                 event.status = -ETIMEDOUT;
2604                 break;
2605         case IB_CM_SIDR_REP_RECEIVED:
2606                 event.param.ud.private_data = ib_event->private_data;
2607                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2608                 if (rep->status != IB_SIDR_SUCCESS) {
2609                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2610                         event.status = ib_event->param.sidr_rep_rcvd.status;
2611                         break;
2612                 }
2613                 ret = cma_set_qkey(id_priv);
2614                 if (ret) {
2615                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
2616                         event.status = -EINVAL;
2617                         break;
2618                 }
2619                 if (id_priv->qkey != rep->qkey) {
2620                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2621                         event.status = -EINVAL;
2622                         break;
2623                 }
2624                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2625                                      id_priv->id.route.path_rec,
2626                                      &event.param.ud.ah_attr);
2627                 event.param.ud.qp_num = rep->qpn;
2628                 event.param.ud.qkey = rep->qkey;
2629                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2630                 event.status = 0;
2631                 break;
2632         default:
2633                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2634                        ib_event->event);
2635                 goto out;
2636         }
2637
2638         ret = id_priv->id.event_handler(&id_priv->id, &event);
2639         if (ret) {
2640                 /* Destroy the CM ID by returning a non-zero value. */
2641                 id_priv->cm_id.ib = NULL;
2642                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2643                 mutex_unlock(&id_priv->handler_mutex);
2644                 rdma_destroy_id(&id_priv->id);
2645                 return ret;
2646         }
2647 out:
2648         mutex_unlock(&id_priv->handler_mutex);
2649         return ret;
2650 }
2651
2652 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2653                               struct rdma_conn_param *conn_param)
2654 {
2655         struct ib_cm_sidr_req_param req;
2656         struct ib_cm_id *id;
2657         int ret;
2658
2659         req.private_data_len = sizeof(struct cma_hdr) +
2660                                conn_param->private_data_len;
2661         if (req.private_data_len < conn_param->private_data_len)
2662                 return -EINVAL;
2663
2664         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2665         if (!req.private_data)
2666                 return -ENOMEM;
2667
2668         if (conn_param->private_data && conn_param->private_data_len)
2669                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2670                        conn_param->private_data, conn_param->private_data_len);
2671
2672         ret = cma_format_hdr((void *) req.private_data, id_priv);
2673         if (ret)
2674                 goto out;
2675
2676         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2677                              id_priv);
2678         if (IS_ERR(id)) {
2679                 ret = PTR_ERR(id);
2680                 goto out;
2681         }
2682         id_priv->cm_id.ib = id;
2683
2684         req.path = id_priv->id.route.path_rec;
2685         req.service_id = cma_get_service_id(id_priv->id.ps, cma_dst_addr(id_priv));
2686         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2687         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2688
2689         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2690         if (ret) {
2691                 ib_destroy_cm_id(id_priv->cm_id.ib);
2692                 id_priv->cm_id.ib = NULL;
2693         }
2694 out:
2695         kfree(req.private_data);
2696         return ret;
2697 }
2698
2699 static int cma_connect_ib(struct rdma_id_private *id_priv,
2700                           struct rdma_conn_param *conn_param)
2701 {
2702         struct ib_cm_req_param req;
2703         struct rdma_route *route;
2704         void *private_data;
2705         struct ib_cm_id *id;
2706         int offset, ret;
2707
2708         memset(&req, 0, sizeof req);
2709         offset = cma_user_data_offset(id_priv->id.ps);
2710         req.private_data_len = offset + conn_param->private_data_len;
2711         if (req.private_data_len < conn_param->private_data_len)
2712                 return -EINVAL;
2713
2714         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2715         if (!private_data)
2716                 return -ENOMEM;
2717
2718         if (conn_param->private_data && conn_param->private_data_len)
2719                 memcpy(private_data + offset, conn_param->private_data,
2720                        conn_param->private_data_len);
2721
2722         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2723         if (IS_ERR(id)) {
2724                 ret = PTR_ERR(id);
2725                 goto out;
2726         }
2727         id_priv->cm_id.ib = id;
2728
2729         route = &id_priv->id.route;
2730         ret = cma_format_hdr(private_data, id_priv);
2731         if (ret)
2732                 goto out;
2733         req.private_data = private_data;
2734
2735         req.primary_path = &route->path_rec[0];
2736         if (route->num_paths == 2)
2737                 req.alternate_path = &route->path_rec[1];
2738
2739         req.service_id = cma_get_service_id(id_priv->id.ps, cma_dst_addr(id_priv));
2740         req.qp_num = id_priv->qp_num;
2741         req.qp_type = id_priv->id.qp_type;
2742         req.starting_psn = id_priv->seq_num;
2743         req.responder_resources = conn_param->responder_resources;
2744         req.initiator_depth = conn_param->initiator_depth;
2745         req.flow_control = conn_param->flow_control;
2746         req.retry_count = min_t(u8, 7, conn_param->retry_count);
2747         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2748         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2749         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2750         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2751         req.srq = id_priv->srq ? 1 : 0;
2752
2753         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2754 out:
2755         if (ret && !IS_ERR(id)) {
2756                 ib_destroy_cm_id(id);
2757                 id_priv->cm_id.ib = NULL;
2758         }
2759
2760         kfree(private_data);
2761         return ret;
2762 }
2763
2764 static int cma_connect_iw(struct rdma_id_private *id_priv,
2765                           struct rdma_conn_param *conn_param)
2766 {
2767         struct iw_cm_id *cm_id;
2768         struct sockaddr_in* sin;
2769         int ret;
2770         struct iw_cm_conn_param iw_param;
2771
2772         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2773         if (IS_ERR(cm_id))
2774                 return PTR_ERR(cm_id);
2775
2776         id_priv->cm_id.iw = cm_id;
2777
2778         sin = (struct sockaddr_in *) cma_src_addr(id_priv);
2779         cm_id->local_addr = *sin;
2780
2781         sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
2782         cm_id->remote_addr = *sin;
2783
2784         ret = cma_modify_qp_rtr(id_priv, conn_param);
2785         if (ret)
2786                 goto out;
2787
2788         if (conn_param) {
2789                 iw_param.ord = conn_param->initiator_depth;
2790                 iw_param.ird = conn_param->responder_resources;
2791                 iw_param.private_data = conn_param->private_data;
2792                 iw_param.private_data_len = conn_param->private_data_len;
2793                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2794         } else {
2795                 memset(&iw_param, 0, sizeof iw_param);
2796                 iw_param.qpn = id_priv->qp_num;
2797         }
2798         ret = iw_cm_connect(cm_id, &iw_param);
2799 out:
2800         if (ret) {
2801                 iw_destroy_cm_id(cm_id);
2802                 id_priv->cm_id.iw = NULL;
2803         }
2804         return ret;
2805 }
2806
2807 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2808 {
2809         struct rdma_id_private *id_priv;
2810         int ret;
2811
2812         id_priv = container_of(id, struct rdma_id_private, id);
2813         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2814                 return -EINVAL;
2815
2816         if (!id->qp) {
2817                 id_priv->qp_num = conn_param->qp_num;
2818                 id_priv->srq = conn_param->srq;
2819         }
2820
2821         switch (rdma_node_get_transport(id->device->node_type)) {
2822         case RDMA_TRANSPORT_IB:
2823                 if (id->qp_type == IB_QPT_UD)
2824                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2825                 else
2826                         ret = cma_connect_ib(id_priv, conn_param);
2827                 break;
2828         case RDMA_TRANSPORT_IWARP:
2829                 ret = cma_connect_iw(id_priv, conn_param);
2830                 break;
2831         default:
2832                 ret = -ENOSYS;
2833                 break;
2834         }
2835         if (ret)
2836                 goto err;
2837
2838         return 0;
2839 err:
2840         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2841         return ret;
2842 }
2843 EXPORT_SYMBOL(rdma_connect);
2844
2845 static int cma_accept_ib(struct rdma_id_private *id_priv,
2846                          struct rdma_conn_param *conn_param)
2847 {
2848         struct ib_cm_rep_param rep;
2849         int ret;
2850
2851         ret = cma_modify_qp_rtr(id_priv, conn_param);
2852         if (ret)
2853                 goto out;
2854
2855         ret = cma_modify_qp_rts(id_priv, conn_param);
2856         if (ret)
2857                 goto out;
2858
2859         memset(&rep, 0, sizeof rep);
2860         rep.qp_num = id_priv->qp_num;
2861         rep.starting_psn = id_priv->seq_num;
2862         rep.private_data = conn_param->private_data;
2863         rep.private_data_len = conn_param->private_data_len;
2864         rep.responder_resources = conn_param->responder_resources;
2865         rep.initiator_depth = conn_param->initiator_depth;
2866         rep.failover_accepted = 0;
2867         rep.flow_control = conn_param->flow_control;
2868         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2869         rep.srq = id_priv->srq ? 1 : 0;
2870
2871         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2872 out:
2873         return ret;
2874 }
2875
2876 static int cma_accept_iw(struct rdma_id_private *id_priv,
2877                   struct rdma_conn_param *conn_param)
2878 {
2879         struct iw_cm_conn_param iw_param;
2880         int ret;
2881
2882         ret = cma_modify_qp_rtr(id_priv, conn_param);
2883         if (ret)
2884                 return ret;
2885
2886         iw_param.ord = conn_param->initiator_depth;
2887         iw_param.ird = conn_param->responder_resources;
2888         iw_param.private_data = conn_param->private_data;
2889         iw_param.private_data_len = conn_param->private_data_len;
2890         if (id_priv->id.qp) {
2891                 iw_param.qpn = id_priv->qp_num;
2892         } else
2893                 iw_param.qpn = conn_param->qp_num;
2894
2895         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2896 }
2897
2898 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2899                              enum ib_cm_sidr_status status,
2900                              const void *private_data, int private_data_len)
2901 {
2902         struct ib_cm_sidr_rep_param rep;
2903         int ret;
2904
2905         memset(&rep, 0, sizeof rep);
2906         rep.status = status;
2907         if (status == IB_SIDR_SUCCESS) {
2908                 ret = cma_set_qkey(id_priv);
2909                 if (ret)
2910                         return ret;
2911                 rep.qp_num = id_priv->qp_num;
2912                 rep.qkey = id_priv->qkey;
2913         }
2914         rep.private_data = private_data;
2915         rep.private_data_len = private_data_len;
2916
2917         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2918 }
2919
2920 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2921 {
2922         struct rdma_id_private *id_priv;
2923         int ret;
2924
2925         id_priv = container_of(id, struct rdma_id_private, id);
2926
2927         id_priv->owner = task_pid_nr(current);
2928
2929         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2930                 return -EINVAL;
2931
2932         if (!id->qp && conn_param) {
2933                 id_priv->qp_num = conn_param->qp_num;
2934                 id_priv->srq = conn_param->srq;
2935         }
2936
2937         switch (rdma_node_get_transport(id->device->node_type)) {
2938         case RDMA_TRANSPORT_IB:
2939                 if (id->qp_type == IB_QPT_UD) {
2940                         if (conn_param)
2941                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2942                                                         conn_param->private_data,
2943                                                         conn_param->private_data_len);
2944                         else
2945                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2946                                                         NULL, 0);
2947                 } else {
2948                         if (conn_param)
2949                                 ret = cma_accept_ib(id_priv, conn_param);
2950                         else
2951                                 ret = cma_rep_recv(id_priv);
2952                 }
2953                 break;
2954         case RDMA_TRANSPORT_IWARP:
2955                 ret = cma_accept_iw(id_priv, conn_param);
2956                 break;
2957         default:
2958                 ret = -ENOSYS;
2959                 break;
2960         }
2961
2962         if (ret)
2963                 goto reject;
2964
2965         return 0;
2966 reject:
2967         cma_modify_qp_err(id_priv);
2968         rdma_reject(id, NULL, 0);
2969         return ret;
2970 }
2971 EXPORT_SYMBOL(rdma_accept);
2972
2973 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2974 {
2975         struct rdma_id_private *id_priv;
2976         int ret;
2977
2978         id_priv = container_of(id, struct rdma_id_private, id);
2979         if (!id_priv->cm_id.ib)
2980                 return -EINVAL;
2981
2982         switch (id->device->node_type) {
2983         case RDMA_NODE_IB_CA:
2984                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2985                 break;
2986         default:
2987                 ret = 0;
2988                 break;
2989         }
2990         return ret;
2991 }
2992 EXPORT_SYMBOL(rdma_notify);
2993
2994 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2995                 u8 private_data_len)
2996 {
2997         struct rdma_id_private *id_priv;
2998         int ret;
2999
3000         id_priv = container_of(id, struct rdma_id_private, id);
3001         if (!id_priv->cm_id.ib)
3002                 return -EINVAL;
3003
3004         switch (rdma_node_get_transport(id->device->node_type)) {
3005         case RDMA_TRANSPORT_IB:
3006                 if (id->qp_type == IB_QPT_UD)
3007                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
3008                                                 private_data, private_data_len);
3009                 else
3010                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
3011                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
3012                                              0, private_data, private_data_len);
3013                 break;
3014         case RDMA_TRANSPORT_IWARP:
3015                 ret = iw_cm_reject(id_priv->cm_id.iw,
3016                                    private_data, private_data_len);
3017                 break;
3018         default:
3019                 ret = -ENOSYS;
3020                 break;
3021         }
3022         return ret;
3023 }
3024 EXPORT_SYMBOL(rdma_reject);
3025
3026 int rdma_disconnect(struct rdma_cm_id *id)
3027 {
3028         struct rdma_id_private *id_priv;
3029         int ret;
3030
3031         id_priv = container_of(id, struct rdma_id_private, id);
3032         if (!id_priv->cm_id.ib)
3033                 return -EINVAL;
3034
3035         switch (rdma_node_get_transport(id->device->node_type)) {
3036         case RDMA_TRANSPORT_IB:
3037                 ret = cma_modify_qp_err(id_priv);
3038                 if (ret)
3039                         goto out;
3040                 /* Initiate or respond to a disconnect. */
3041                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3042                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3043                 break;
3044         case RDMA_TRANSPORT_IWARP:
3045                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3046                 break;
3047         default:
3048                 ret = -EINVAL;
3049                 break;
3050         }
3051 out:
3052         return ret;
3053 }
3054 EXPORT_SYMBOL(rdma_disconnect);
3055
3056 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3057 {
3058         struct rdma_id_private *id_priv;
3059         struct cma_multicast *mc = multicast->context;
3060         struct rdma_cm_event event;
3061         int ret;
3062
3063         id_priv = mc->id_priv;
3064         if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3065             cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3066                 return 0;
3067
3068         mutex_lock(&id_priv->qp_mutex);
3069         if (!status && id_priv->id.qp)
3070                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3071                                          be16_to_cpu(multicast->rec.mlid));
3072         mutex_unlock(&id_priv->qp_mutex);
3073
3074         memset(&event, 0, sizeof event);
3075         event.status = status;
3076         event.param.ud.private_data = mc->context;
3077         if (!status) {
3078                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3079                 ib_init_ah_from_mcmember(id_priv->id.device,
3080                                          id_priv->id.port_num, &multicast->rec,
3081                                          &event.param.ud.ah_attr);
3082                 event.param.ud.qp_num = 0xFFFFFF;
3083                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3084         } else
3085                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3086
3087         ret = id_priv->id.event_handler(&id_priv->id, &event);
3088         if (ret) {
3089                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3090                 mutex_unlock(&id_priv->handler_mutex);
3091                 rdma_destroy_id(&id_priv->id);
3092                 return 0;
3093         }
3094
3095         mutex_unlock(&id_priv->handler_mutex);
3096         return 0;
3097 }
3098
3099 static void cma_set_mgid(struct rdma_id_private *id_priv,
3100                          struct sockaddr *addr, union ib_gid *mgid)
3101 {
3102         unsigned char mc_map[MAX_ADDR_LEN];
3103         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3104         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3105         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3106
3107         if (cma_any_addr(addr)) {
3108                 memset(mgid, 0, sizeof *mgid);
3109         } else if ((addr->sa_family == AF_INET6) &&
3110                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3111                                                                  0xFF10A01B)) {
3112                 /* IPv6 address is an SA assigned MGID. */
3113                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3114         } else if ((addr->sa_family == AF_INET6)) {
3115                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3116                 if (id_priv->id.ps == RDMA_PS_UDP)
3117                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3118                 *mgid = *(union ib_gid *) (mc_map + 4);
3119         } else {
3120                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3121                 if (id_priv->id.ps == RDMA_PS_UDP)
3122                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3123                 *mgid = *(union ib_gid *) (mc_map + 4);
3124         }
3125 }
3126
3127 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3128                                  struct cma_multicast *mc)
3129 {
3130         struct ib_sa_mcmember_rec rec;
3131         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3132         ib_sa_comp_mask comp_mask;
3133         int ret;
3134
3135         ib_addr_get_mgid(dev_addr, &rec.mgid);
3136         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3137                                      &rec.mgid, &rec);
3138         if (ret)
3139                 return ret;
3140
3141         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3142         if (id_priv->id.ps == RDMA_PS_UDP)
3143                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3144         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3145         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3146         rec.join_state = 1;
3147
3148         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3149                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3150                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3151                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3152                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3153
3154         if (id_priv->id.ps == RDMA_PS_IPOIB)
3155                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3156                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3157                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3158                              IB_SA_MCMEMBER_REC_MTU |
3159                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
3160
3161         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3162                                                 id_priv->id.port_num, &rec,
3163                                                 comp_mask, GFP_KERNEL,
3164                                                 cma_ib_mc_handler, mc);
3165         return PTR_RET(mc->multicast.ib);
3166 }
3167
3168 static void iboe_mcast_work_handler(struct work_struct *work)
3169 {
3170         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3171         struct cma_multicast *mc = mw->mc;
3172         struct ib_sa_multicast *m = mc->multicast.ib;
3173
3174         mc->multicast.ib->context = mc;
3175         cma_ib_mc_handler(0, m);
3176         kref_put(&mc->mcref, release_mc);
3177         kfree(mw);
3178 }
3179
3180 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3181 {
3182         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3183         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3184
3185         if (cma_any_addr(addr)) {
3186                 memset(mgid, 0, sizeof *mgid);
3187         } else if (addr->sa_family == AF_INET6) {
3188                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3189         } else {
3190                 mgid->raw[0] = 0xff;
3191                 mgid->raw[1] = 0x0e;
3192                 mgid->raw[2] = 0;
3193                 mgid->raw[3] = 0;
3194                 mgid->raw[4] = 0;
3195                 mgid->raw[5] = 0;
3196                 mgid->raw[6] = 0;
3197                 mgid->raw[7] = 0;
3198                 mgid->raw[8] = 0;
3199                 mgid->raw[9] = 0;
3200                 mgid->raw[10] = 0xff;
3201                 mgid->raw[11] = 0xff;
3202                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3203         }
3204 }
3205
3206 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3207                                    struct cma_multicast *mc)
3208 {
3209         struct iboe_mcast_work *work;
3210         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3211         int err;
3212         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3213         struct net_device *ndev = NULL;
3214
3215         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3216                 return -EINVAL;
3217
3218         work = kzalloc(sizeof *work, GFP_KERNEL);
3219         if (!work)
3220                 return -ENOMEM;
3221
3222         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3223         if (!mc->multicast.ib) {
3224                 err = -ENOMEM;
3225                 goto out1;
3226         }
3227
3228         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3229
3230         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3231         if (id_priv->id.ps == RDMA_PS_UDP)
3232                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3233
3234         if (dev_addr->bound_dev_if)
3235                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3236         if (!ndev) {
3237                 err = -ENODEV;
3238                 goto out2;
3239         }
3240         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3241         mc->multicast.ib->rec.hop_limit = 1;
3242         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3243         dev_put(ndev);
3244         if (!mc->multicast.ib->rec.mtu) {
3245                 err = -EINVAL;
3246                 goto out2;
3247         }
3248         iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3249         work->id = id_priv;
3250         work->mc = mc;
3251         INIT_WORK(&work->work, iboe_mcast_work_handler);
3252         kref_get(&mc->mcref);
3253         queue_work(cma_wq, &work->work);
3254
3255         return 0;
3256
3257 out2:
3258         kfree(mc->multicast.ib);
3259 out1:
3260         kfree(work);
3261         return err;
3262 }
3263
3264 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3265                         void *context)
3266 {
3267         struct rdma_id_private *id_priv;
3268         struct cma_multicast *mc;
3269         int ret;
3270
3271         id_priv = container_of(id, struct rdma_id_private, id);
3272         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3273             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3274                 return -EINVAL;
3275
3276         mc = kmalloc(sizeof *mc, GFP_KERNEL);
3277         if (!mc)
3278                 return -ENOMEM;
3279
3280         memcpy(&mc->addr, addr, rdma_addr_size(addr));
3281         mc->context = context;
3282         mc->id_priv = id_priv;
3283
3284         spin_lock(&id_priv->lock);
3285         list_add(&mc->list, &id_priv->mc_list);
3286         spin_unlock(&id_priv->lock);
3287
3288         switch (rdma_node_get_transport(id->device->node_type)) {
3289         case RDMA_TRANSPORT_IB:
3290                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3291                 case IB_LINK_LAYER_INFINIBAND:
3292                         ret = cma_join_ib_multicast(id_priv, mc);
3293                         break;
3294                 case IB_LINK_LAYER_ETHERNET:
3295                         kref_init(&mc->mcref);
3296                         ret = cma_iboe_join_multicast(id_priv, mc);
3297                         break;
3298                 default:
3299                         ret = -EINVAL;
3300                 }
3301                 break;
3302         default:
3303                 ret = -ENOSYS;
3304                 break;
3305         }
3306
3307         if (ret) {
3308                 spin_lock_irq(&id_priv->lock);
3309                 list_del(&mc->list);
3310                 spin_unlock_irq(&id_priv->lock);
3311                 kfree(mc);
3312         }
3313         return ret;
3314 }
3315 EXPORT_SYMBOL(rdma_join_multicast);
3316
3317 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3318 {
3319         struct rdma_id_private *id_priv;
3320         struct cma_multicast *mc;
3321
3322         id_priv = container_of(id, struct rdma_id_private, id);
3323         spin_lock_irq(&id_priv->lock);
3324         list_for_each_entry(mc, &id_priv->mc_list, list) {
3325                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3326                         list_del(&mc->list);
3327                         spin_unlock_irq(&id_priv->lock);
3328
3329                         if (id->qp)
3330                                 ib_detach_mcast(id->qp,
3331                                                 &mc->multicast.ib->rec.mgid,
3332                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
3333                         if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3334                                 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3335                                 case IB_LINK_LAYER_INFINIBAND:
3336                                         ib_sa_free_multicast(mc->multicast.ib);
3337                                         kfree(mc);
3338                                         break;
3339                                 case IB_LINK_LAYER_ETHERNET:
3340                                         kref_put(&mc->mcref, release_mc);
3341                                         break;
3342                                 default:
3343                                         break;
3344                                 }
3345                         }
3346                         return;
3347                 }
3348         }
3349         spin_unlock_irq(&id_priv->lock);
3350 }
3351 EXPORT_SYMBOL(rdma_leave_multicast);
3352
3353 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3354 {
3355         struct rdma_dev_addr *dev_addr;
3356         struct cma_ndev_work *work;
3357
3358         dev_addr = &id_priv->id.route.addr.dev_addr;
3359
3360         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3361             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3362                 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3363                        ndev->name, &id_priv->id);
3364                 work = kzalloc(sizeof *work, GFP_KERNEL);
3365                 if (!work)
3366                         return -ENOMEM;
3367
3368                 INIT_WORK(&work->work, cma_ndev_work_handler);
3369                 work->id = id_priv;
3370                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3371                 atomic_inc(&id_priv->refcount);
3372                 queue_work(cma_wq, &work->work);
3373         }
3374
3375         return 0;
3376 }
3377
3378 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3379                                void *ptr)
3380 {
3381         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3382         struct cma_device *cma_dev;
3383         struct rdma_id_private *id_priv;
3384         int ret = NOTIFY_DONE;
3385
3386         if (dev_net(ndev) != &init_net)
3387                 return NOTIFY_DONE;
3388
3389         if (event != NETDEV_BONDING_FAILOVER)
3390                 return NOTIFY_DONE;
3391
3392         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3393                 return NOTIFY_DONE;
3394
3395         mutex_lock(&lock);
3396         list_for_each_entry(cma_dev, &dev_list, list)
3397                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3398                         ret = cma_netdev_change(ndev, id_priv);
3399                         if (ret)
3400                                 goto out;
3401                 }
3402
3403 out:
3404         mutex_unlock(&lock);
3405         return ret;
3406 }
3407
3408 static struct notifier_block cma_nb = {
3409         .notifier_call = cma_netdev_callback
3410 };
3411
3412 static void cma_add_one(struct ib_device *device)
3413 {
3414         struct cma_device *cma_dev;
3415         struct rdma_id_private *id_priv;
3416
3417         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3418         if (!cma_dev)
3419                 return;
3420
3421         cma_dev->device = device;
3422
3423         init_completion(&cma_dev->comp);
3424         atomic_set(&cma_dev->refcount, 1);
3425         INIT_LIST_HEAD(&cma_dev->id_list);
3426         ib_set_client_data(device, &cma_client, cma_dev);
3427
3428         mutex_lock(&lock);
3429         list_add_tail(&cma_dev->list, &dev_list);
3430         list_for_each_entry(id_priv, &listen_any_list, list)
3431                 cma_listen_on_dev(id_priv, cma_dev);
3432         mutex_unlock(&lock);
3433 }
3434
3435 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3436 {
3437         struct rdma_cm_event event;
3438         enum rdma_cm_state state;
3439         int ret = 0;
3440
3441         /* Record that we want to remove the device */
3442         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3443         if (state == RDMA_CM_DESTROYING)
3444                 return 0;
3445
3446         cma_cancel_operation(id_priv, state);
3447         mutex_lock(&id_priv->handler_mutex);
3448
3449         /* Check for destruction from another callback. */
3450         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3451                 goto out;
3452
3453         memset(&event, 0, sizeof event);
3454         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3455         ret = id_priv->id.event_handler(&id_priv->id, &event);
3456 out:
3457         mutex_unlock(&id_priv->handler_mutex);
3458         return ret;
3459 }
3460
3461 static void cma_process_remove(struct cma_device *cma_dev)
3462 {
3463         struct rdma_id_private *id_priv;
3464         int ret;
3465
3466         mutex_lock(&lock);
3467         while (!list_empty(&cma_dev->id_list)) {
3468                 id_priv = list_entry(cma_dev->id_list.next,
3469                                      struct rdma_id_private, list);
3470
3471                 list_del(&id_priv->listen_list);
3472                 list_del_init(&id_priv->list);
3473                 atomic_inc(&id_priv->refcount);
3474                 mutex_unlock(&lock);
3475
3476                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3477                 cma_deref_id(id_priv);
3478                 if (ret)
3479                         rdma_destroy_id(&id_priv->id);
3480
3481                 mutex_lock(&lock);
3482         }
3483         mutex_unlock(&lock);
3484
3485         cma_deref_dev(cma_dev);
3486         wait_for_completion(&cma_dev->comp);
3487 }
3488
3489 static void cma_remove_one(struct ib_device *device)
3490 {
3491         struct cma_device *cma_dev;
3492
3493         cma_dev = ib_get_client_data(device, &cma_client);
3494         if (!cma_dev)
3495                 return;
3496
3497         mutex_lock(&lock);
3498         list_del(&cma_dev->list);
3499         mutex_unlock(&lock);
3500
3501         cma_process_remove(cma_dev);
3502         kfree(cma_dev);
3503 }
3504
3505 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3506 {
3507         struct nlmsghdr *nlh;
3508         struct rdma_cm_id_stats *id_stats;
3509         struct rdma_id_private *id_priv;
3510         struct rdma_cm_id *id = NULL;
3511         struct cma_device *cma_dev;
3512         int i_dev = 0, i_id = 0;
3513
3514         /*
3515          * We export all of the IDs as a sequence of messages.  Each
3516          * ID gets its own netlink message.
3517          */
3518         mutex_lock(&lock);
3519
3520         list_for_each_entry(cma_dev, &dev_list, list) {
3521                 if (i_dev < cb->args[0]) {
3522                         i_dev++;
3523                         continue;
3524                 }
3525
3526                 i_id = 0;
3527                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3528                         if (i_id < cb->args[1]) {
3529                                 i_id++;
3530                                 continue;
3531                         }
3532
3533                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3534                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
3535                                                 RDMA_NL_RDMA_CM_ID_STATS);
3536                         if (!id_stats)
3537                                 goto out;
3538
3539                         memset(id_stats, 0, sizeof *id_stats);
3540                         id = &id_priv->id;
3541                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
3542                         id_stats->port_num = id->port_num;
3543                         id_stats->bound_dev_if =
3544                                 id->route.addr.dev_addr.bound_dev_if;
3545
3546                         if (cma_family(id_priv) == AF_INET) {
3547                                 if (ibnl_put_attr(skb, nlh,
3548                                                   sizeof(struct sockaddr_in),
3549                                                   cma_src_addr(id_priv),
3550                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3551                                         goto out;
3552                                 }
3553                                 if (ibnl_put_attr(skb, nlh,
3554                                                   sizeof(struct sockaddr_in),
3555                                                   cma_dst_addr(id_priv),
3556                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3557                                         goto out;
3558                                 }
3559                         } else if (cma_family(id_priv) == AF_INET6) {
3560                                 if (ibnl_put_attr(skb, nlh,
3561                                                   sizeof(struct sockaddr_in6),
3562                                                   cma_src_addr(id_priv),
3563                                                   RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3564                                         goto out;
3565                                 }
3566                                 if (ibnl_put_attr(skb, nlh,
3567                                                   sizeof(struct sockaddr_in6),
3568                                                   cma_dst_addr(id_priv),
3569                                                   RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3570                                         goto out;
3571                                 }
3572                         }
3573
3574                         id_stats->pid           = id_priv->owner;
3575                         id_stats->port_space    = id->ps;
3576                         id_stats->cm_state      = id_priv->state;
3577                         id_stats->qp_num        = id_priv->qp_num;
3578                         id_stats->qp_type       = id->qp_type;
3579
3580                         i_id++;
3581                 }
3582
3583                 cb->args[1] = 0;
3584                 i_dev++;
3585         }
3586
3587 out:
3588         mutex_unlock(&lock);
3589         cb->args[0] = i_dev;
3590         cb->args[1] = i_id;
3591
3592         return skb->len;
3593 }
3594
3595 static const struct ibnl_client_cbs cma_cb_table[] = {
3596         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3597                                        .module = THIS_MODULE },
3598 };
3599
3600 static int __init cma_init(void)
3601 {
3602         int ret;
3603
3604         cma_wq = create_singlethread_workqueue("rdma_cm");
3605         if (!cma_wq)
3606                 return -ENOMEM;
3607
3608         ib_sa_register_client(&sa_client);
3609         rdma_addr_register_client(&addr_client);
3610         register_netdevice_notifier(&cma_nb);
3611
3612         ret = ib_register_client(&cma_client);
3613         if (ret)
3614                 goto err;
3615
3616         if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3617                 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3618
3619         return 0;
3620
3621 err:
3622         unregister_netdevice_notifier(&cma_nb);
3623         rdma_addr_unregister_client(&addr_client);
3624         ib_sa_unregister_client(&sa_client);
3625         destroy_workqueue(cma_wq);
3626         return ret;
3627 }
3628
3629 static void __exit cma_cleanup(void)
3630 {
3631         ibnl_remove_client(RDMA_NL_RDMA_CM);
3632         ib_unregister_client(&cma_client);
3633         unregister_netdevice_notifier(&cma_nb);
3634         rdma_addr_unregister_client(&addr_client);
3635         ib_sa_unregister_client(&sa_client);
3636         destroy_workqueue(cma_wq);
3637         idr_destroy(&tcp_ps);
3638         idr_destroy(&udp_ps);
3639         idr_destroy(&ipoib_ps);
3640         idr_destroy(&ib_ps);
3641 }
3642
3643 module_init(cma_init);
3644 module_exit(cma_cleanup);