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