git.oblomov.eu Git - linux-2.6/blob - 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 licensed under one of the following licenses:
   8  *
   9  * 1) under the terms of the "Common Public License 1.0" a copy of which is
  10  *    available from the Open Source Initiative, see
  11  *    http://www.opensource.org/licenses/cpl.php.
  12  *
  13  * 2) under the terms of the "The BSD License" a copy of which is
  14  *    available from the Open Source Initiative, see
  15  *    http://www.opensource.org/licenses/bsd-license.php.
  16  *
  17  * 3) under the terms of the "GNU General Public License (GPL) Version 2" a
  18  *    copy of which is available from the Open Source Initiative, see
  19  *    http://www.opensource.org/licenses/gpl-license.php.
  20  *
  21  * Licensee has the right to choose one of the above licenses.
  22  *
  23  * Redistributions of source code must retain the above copyright
  24  * notice and one of the license notices.
  25  *
  26  * Redistributions in binary form must reproduce both the above copyright
  27  * notice, one of the license notices in the documentation
  28  * and/or other materials provided with the distribution.
  29  *
  30  */
  31
  32 #include <linux/completion.h>
  33 #include <linux/in.h>
  34 #include <linux/in6.h>
  35 #include <linux/mutex.h>
  36 #include <linux/random.h>
  37 #include <linux/idr.h>
  38
  39 #include <net/tcp.h>
  40
  41 #include <rdma/rdma_cm.h>
  42 #include <rdma/rdma_cm_ib.h>
  43 #include <rdma/ib_cache.h>
  44 #include <rdma/ib_cm.h>
  45 #include <rdma/ib_sa.h>
  46
  47 MODULE_AUTHOR("Sean Hefty");
  48 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  49 MODULE_LICENSE("Dual BSD/GPL");
  50
  51 #define CMA_CM_RESPONSE_TIMEOUT 20
  52 #define CMA_MAX_CM_RETRIES 15
  53
  54 static void cma_add_one(struct ib_device *device);
  55 static void cma_remove_one(struct ib_device *device);
  56
  57 static struct ib_client cma_client = {
  58         .name   = "cma",
  59         .add    = cma_add_one,
  60         .remove = cma_remove_one
  61 };
  62
  63 static LIST_HEAD(dev_list);
  64 static LIST_HEAD(listen_any_list);
  65 static DEFINE_MUTEX(lock);
  66 static struct workqueue_struct *cma_wq;
  67 static DEFINE_IDR(sdp_ps);
  68 static DEFINE_IDR(tcp_ps);
  69
  70 struct cma_device {
  71         struct list_head        list;
  72         struct ib_device        *device;
  73         __be64                  node_guid;
  74         struct completion       comp;
  75         atomic_t                refcount;
  76         struct list_head        id_list;
  77 };
  78
  79 enum cma_state {
  80         CMA_IDLE,
  81         CMA_ADDR_QUERY,
  82         CMA_ADDR_RESOLVED,
  83         CMA_ROUTE_QUERY,
  84         CMA_ROUTE_RESOLVED,
  85         CMA_CONNECT,
  86         CMA_DISCONNECT,
  87         CMA_ADDR_BOUND,
  88         CMA_LISTEN,
  89         CMA_DEVICE_REMOVAL,
  90         CMA_DESTROYING
  91 };
  92
  93 struct rdma_bind_list {
  94         struct idr              *ps;
  95         struct hlist_head       owners;
  96         unsigned short          port;
  97 };
  98
  99 /*
 100  * Device removal can occur at anytime, so we need extra handling to
 101  * serialize notifying the user of device removal with other callbacks.
 102  * We do this by disabling removal notification while a callback is in process,
 103  * and reporting it after the callback completes.
 104  */
 105 struct rdma_id_private {
 106         struct rdma_cm_id       id;
 107
 108         struct rdma_bind_list   *bind_list;
 109         struct hlist_node       node;
 110         struct list_head        list;
 111         struct list_head        listen_list;
 112         struct cma_device       *cma_dev;
 113
 114         enum cma_state          state;
 115         spinlock_t              lock;
 116         struct completion       comp;
 117         atomic_t                refcount;
 118         wait_queue_head_t       wait_remove;
 119         atomic_t                dev_remove;
 120
 121         int                     backlog;
 122         int                     timeout_ms;
 123         struct ib_sa_query      *query;
 124         int                     query_id;
 125         union {
 126                 struct ib_cm_id *ib;
 127         } cm_id;
 128
 129         u32                     seq_num;
 130         u32                     qp_num;
 131         enum ib_qp_type         qp_type;
 132         u8                      srq;
 133 };
 134
 135 struct cma_work {
 136         struct work_struct      work;
 137         struct rdma_id_private  *id;
 138         enum cma_state          old_state;
 139         enum cma_state          new_state;
 140         struct rdma_cm_event    event;
 141 };
 142
 143 union cma_ip_addr {
 144         struct in6_addr ip6;
 145         struct {
 146                 __u32 pad[3];
 147                 __u32 addr;
 148         } ip4;
 149 };
 150
 151 struct cma_hdr {
 152         u8 cma_version;
 153         u8 ip_version;  /* IP version: 7:4 */
 154         __u16 port;
 155         union cma_ip_addr src_addr;
 156         union cma_ip_addr dst_addr;
 157 };
 158
 159 struct sdp_hh {
 160         u8 bsdh[16];
 161         u8 sdp_version; /* Major version: 7:4 */
 162         u8 ip_version;  /* IP version: 7:4 */
 163         u8 sdp_specific1[10];
 164         __u16 port;
 165         __u16 sdp_specific2;
 166         union cma_ip_addr src_addr;
 167         union cma_ip_addr dst_addr;
 168 };
 169
 170 struct sdp_hah {
 171         u8 bsdh[16];
 172         u8 sdp_version;
 173 };
 174
 175 #define CMA_VERSION 0x00
 176 #define SDP_MAJ_VERSION 0x2
 177
 178 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 179 {
 180         unsigned long flags;
 181         int ret;
 182
 183         spin_lock_irqsave(&id_priv->lock, flags);
 184         ret = (id_priv->state == comp);
 185         spin_unlock_irqrestore(&id_priv->lock, flags);
 186         return ret;
 187 }
 188
 189 static int cma_comp_exch(struct rdma_id_private *id_priv,
 190                          enum cma_state comp, enum cma_state exch)
 191 {
 192         unsigned long flags;
 193         int ret;
 194
 195         spin_lock_irqsave(&id_priv->lock, flags);
 196         if ((ret = (id_priv->state == comp)))
 197                 id_priv->state = exch;
 198         spin_unlock_irqrestore(&id_priv->lock, flags);
 199         return ret;
 200 }
 201
 202 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 203                                enum cma_state exch)
 204 {
 205         unsigned long flags;
 206         enum cma_state old;
 207
 208         spin_lock_irqsave(&id_priv->lock, flags);
 209         old = id_priv->state;
 210         id_priv->state = exch;
 211         spin_unlock_irqrestore(&id_priv->lock, flags);
 212         return old;
 213 }
 214
 215 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 216 {
 217         return hdr->ip_version >> 4;
 218 }
 219
 220 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 221 {
 222         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 223 }
 224
 225 static inline u8 sdp_get_majv(u8 sdp_version)
 226 {
 227         return sdp_version >> 4;
 228 }
 229
 230 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 231 {
 232         return hh->ip_version >> 4;
 233 }
 234
 235 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 236 {
 237         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 238 }
 239
 240 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 241                               struct cma_device *cma_dev)
 242 {
 243         atomic_inc(&cma_dev->refcount);
 244         id_priv->cma_dev = cma_dev;
 245         id_priv->id.device = cma_dev->device;
 246         list_add_tail(&id_priv->list, &cma_dev->id_list);
 247 }
 248
 249 static inline void cma_deref_dev(struct cma_device *cma_dev)
 250 {
 251         if (atomic_dec_and_test(&cma_dev->refcount))
 252                 complete(&cma_dev->comp);
 253 }
 254
 255 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 256 {
 257         list_del(&id_priv->list);
 258         cma_deref_dev(id_priv->cma_dev);
 259         id_priv->cma_dev = NULL;
 260 }
 261
 262 static int cma_acquire_ib_dev(struct rdma_id_private *id_priv)
 263 {
 264         struct cma_device *cma_dev;
 265         union ib_gid gid;
 266         int ret = -ENODEV;
 267
 268         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid),
 269
 270         mutex_lock(&lock);
 271         list_for_each_entry(cma_dev, &dev_list, list) {
 272                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 273                                          &id_priv->id.port_num, NULL);
 274                 if (!ret) {
 275                         cma_attach_to_dev(id_priv, cma_dev);
 276                         break;
 277                 }
 278         }
 279         mutex_unlock(&lock);
 280         return ret;
 281 }
 282
 283 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 284 {
 285         switch (id_priv->id.route.addr.dev_addr.dev_type) {
 286         case IB_NODE_CA:
 287                 return cma_acquire_ib_dev(id_priv);
 288         default:
 289                 return -ENODEV;
 290         }
 291 }
 292
 293 static void cma_deref_id(struct rdma_id_private *id_priv)
 294 {
 295         if (atomic_dec_and_test(&id_priv->refcount))
 296                 complete(&id_priv->comp);
 297 }
 298
 299 static void cma_release_remove(struct rdma_id_private *id_priv)
 300 {
 301         if (atomic_dec_and_test(&id_priv->dev_remove))
 302                 wake_up(&id_priv->wait_remove);
 303 }
 304
 305 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 306                                   void *context, enum rdma_port_space ps)
 307 {
 308         struct rdma_id_private *id_priv;
 309
 310         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 311         if (!id_priv)
 312                 return ERR_PTR(-ENOMEM);
 313
 314         id_priv->state = CMA_IDLE;
 315         id_priv->id.context = context;
 316         id_priv->id.event_handler = event_handler;
 317         id_priv->id.ps = ps;
 318         spin_lock_init(&id_priv->lock);
 319         init_completion(&id_priv->comp);
 320         atomic_set(&id_priv->refcount, 1);
 321         init_waitqueue_head(&id_priv->wait_remove);
 322         atomic_set(&id_priv->dev_remove, 0);
 323         INIT_LIST_HEAD(&id_priv->listen_list);
 324         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 325
 326         return &id_priv->id;
 327 }
 328 EXPORT_SYMBOL(rdma_create_id);
 329
 330 static int cma_init_ib_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 331 {
 332         struct ib_qp_attr qp_attr;
 333         struct rdma_dev_addr *dev_addr;
 334         int ret;
 335
 336         dev_addr = &id_priv->id.route.addr.dev_addr;
 337         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 338                                   ib_addr_get_pkey(dev_addr),
 339                                   &qp_attr.pkey_index);
 340         if (ret)
 341                 return ret;
 342
 343         qp_attr.qp_state = IB_QPS_INIT;
 344         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 345         qp_attr.port_num = id_priv->id.port_num;
 346         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS |
 347                                           IB_QP_PKEY_INDEX | IB_QP_PORT);
 348 }
 349
 350 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 351                    struct ib_qp_init_attr *qp_init_attr)
 352 {
 353         struct rdma_id_private *id_priv;
 354         struct ib_qp *qp;
 355         int ret;
 356
 357         id_priv = container_of(id, struct rdma_id_private, id);
 358         if (id->device != pd->device)
 359                 return -EINVAL;
 360
 361         qp = ib_create_qp(pd, qp_init_attr);
 362         if (IS_ERR(qp))
 363                 return PTR_ERR(qp);
 364
 365         switch (id->device->node_type) {
 366         case IB_NODE_CA:
 367                 ret = cma_init_ib_qp(id_priv, qp);
 368                 break;
 369         default:
 370                 ret = -ENOSYS;
 371                 break;
 372         }
 373
 374         if (ret)
 375                 goto err;
 376
 377         id->qp = qp;
 378         id_priv->qp_num = qp->qp_num;
 379         id_priv->qp_type = qp->qp_type;
 380         id_priv->srq = (qp->srq != NULL);
 381         return 0;
 382 err:
 383         ib_destroy_qp(qp);
 384         return ret;
 385 }
 386 EXPORT_SYMBOL(rdma_create_qp);
 387
 388 void rdma_destroy_qp(struct rdma_cm_id *id)
 389 {
 390         ib_destroy_qp(id->qp);
 391 }
 392 EXPORT_SYMBOL(rdma_destroy_qp);
 393
 394 static int cma_modify_qp_rtr(struct rdma_cm_id *id)
 395 {
 396         struct ib_qp_attr qp_attr;
 397         int qp_attr_mask, ret;
 398
 399         if (!id->qp)
 400                 return 0;
 401
 402         /* Need to update QP attributes from default values. */
 403         qp_attr.qp_state = IB_QPS_INIT;
 404         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 405         if (ret)
 406                 return ret;
 407
 408         ret = ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 409         if (ret)
 410                 return ret;
 411
 412         qp_attr.qp_state = IB_QPS_RTR;
 413         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 414         if (ret)
 415                 return ret;
 416
 417         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 418 }
 419
 420 static int cma_modify_qp_rts(struct rdma_cm_id *id)
 421 {
 422         struct ib_qp_attr qp_attr;
 423         int qp_attr_mask, ret;
 424
 425         if (!id->qp)
 426                 return 0;
 427
 428         qp_attr.qp_state = IB_QPS_RTS;
 429         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 430         if (ret)
 431                 return ret;
 432
 433         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 434 }
 435
 436 static int cma_modify_qp_err(struct rdma_cm_id *id)
 437 {
 438         struct ib_qp_attr qp_attr;
 439
 440         if (!id->qp)
 441                 return 0;
 442
 443         qp_attr.qp_state = IB_QPS_ERR;
 444         return ib_modify_qp(id->qp, &qp_attr, IB_QP_STATE);
 445 }
 446
 447 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 448                        int *qp_attr_mask)
 449 {
 450         struct rdma_id_private *id_priv;
 451         int ret;
 452
 453         id_priv = container_of(id, struct rdma_id_private, id);
 454         switch (id_priv->id.device->node_type) {
 455         case IB_NODE_CA:
 456                 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 457                                          qp_attr_mask);
 458                 if (qp_attr->qp_state == IB_QPS_RTR)
 459                         qp_attr->rq_psn = id_priv->seq_num;
 460                 break;
 461         default:
 462                 ret = -ENOSYS;
 463                 break;
 464         }
 465
 466         return ret;
 467 }
 468 EXPORT_SYMBOL(rdma_init_qp_attr);
 469
 470 static inline int cma_zero_addr(struct sockaddr *addr)
 471 {
 472         struct in6_addr *ip6;
 473
 474         if (addr->sa_family == AF_INET)
 475                 return ZERONET(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 476         else {
 477                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 478                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 479                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 480         }
 481 }
 482
 483 static inline int cma_loopback_addr(struct sockaddr *addr)
 484 {
 485         return LOOPBACK(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 486 }
 487
 488 static inline int cma_any_addr(struct sockaddr *addr)
 489 {
 490         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 491 }
 492
 493 static inline int cma_any_port(struct sockaddr *addr)
 494 {
 495         return !((struct sockaddr_in *) addr)->sin_port;
 496 }
 497
 498 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 499                             u8 *ip_ver, __u16 *port,
 500                             union cma_ip_addr **src, union cma_ip_addr **dst)
 501 {
 502         switch (ps) {
 503         case RDMA_PS_SDP:
 504                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 505                     SDP_MAJ_VERSION)
 506                         return -EINVAL;
 507
 508                 *ip_ver = sdp_get_ip_ver(hdr);
 509                 *port   = ((struct sdp_hh *) hdr)->port;
 510                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 511                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 512                 break;
 513         default:
 514                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 515                         return -EINVAL;
 516
 517                 *ip_ver = cma_get_ip_ver(hdr);
 518                 *port   = ((struct cma_hdr *) hdr)->port;
 519                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 520                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 521                 break;
 522         }
 523
 524         if (*ip_ver != 4 && *ip_ver != 6)
 525                 return -EINVAL;
 526         return 0;
 527 }
 528
 529 static void cma_save_net_info(struct rdma_addr *addr,
 530                               struct rdma_addr *listen_addr,
 531                               u8 ip_ver, __u16 port,
 532                               union cma_ip_addr *src, union cma_ip_addr *dst)
 533 {
 534         struct sockaddr_in *listen4, *ip4;
 535         struct sockaddr_in6 *listen6, *ip6;
 536
 537         switch (ip_ver) {
 538         case 4:
 539                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 540                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 541                 ip4->sin_family = listen4->sin_family;
 542                 ip4->sin_addr.s_addr = dst->ip4.addr;
 543                 ip4->sin_port = listen4->sin_port;
 544
 545                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 546                 ip4->sin_family = listen4->sin_family;
 547                 ip4->sin_addr.s_addr = src->ip4.addr;
 548                 ip4->sin_port = port;
 549                 break;
 550         case 6:
 551                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 552                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 553                 ip6->sin6_family = listen6->sin6_family;
 554                 ip6->sin6_addr = dst->ip6;
 555                 ip6->sin6_port = listen6->sin6_port;
 556
 557                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 558                 ip6->sin6_family = listen6->sin6_family;
 559                 ip6->sin6_addr = src->ip6;
 560                 ip6->sin6_port = port;
 561                 break;
 562         default:
 563                 break;
 564         }
 565 }
 566
 567 static inline int cma_user_data_offset(enum rdma_port_space ps)
 568 {
 569         switch (ps) {
 570         case RDMA_PS_SDP:
 571                 return 0;
 572         default:
 573                 return sizeof(struct cma_hdr);
 574         }
 575 }
 576
 577 static int cma_notify_user(struct rdma_id_private *id_priv,
 578                            enum rdma_cm_event_type type, int status,
 579                            void *data, u8 data_len)
 580 {
 581         struct rdma_cm_event event;
 582
 583         event.event = type;
 584         event.status = status;
 585         event.private_data = data;
 586         event.private_data_len = data_len;
 587
 588         return id_priv->id.event_handler(&id_priv->id, &event);
 589 }
 590
 591 static void cma_cancel_route(struct rdma_id_private *id_priv)
 592 {
 593         switch (id_priv->id.device->node_type) {
 594         case IB_NODE_CA:
 595                 if (id_priv->query)
 596                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 597                 break;
 598         default:
 599                 break;
 600         }
 601 }
 602
 603 static inline int cma_internal_listen(struct rdma_id_private *id_priv)
 604 {
 605         return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
 606                cma_any_addr(&id_priv->id.route.addr.src_addr);
 607 }
 608
 609 static void cma_destroy_listen(struct rdma_id_private *id_priv)
 610 {
 611         cma_exch(id_priv, CMA_DESTROYING);
 612
 613         if (id_priv->cma_dev) {
 614                 switch (id_priv->id.device->node_type) {
 615                 case IB_NODE_CA:
 616                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 617                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 618                         break;
 619                 default:
 620                         break;
 621                 }
 622                 cma_detach_from_dev(id_priv);
 623         }
 624         list_del(&id_priv->listen_list);
 625
 626         cma_deref_id(id_priv);
 627         wait_for_completion(&id_priv->comp);
 628
 629         kfree(id_priv);
 630 }
 631
 632 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 633 {
 634         struct rdma_id_private *dev_id_priv;
 635
 636         mutex_lock(&lock);
 637         list_del(&id_priv->list);
 638
 639         while (!list_empty(&id_priv->listen_list)) {
 640                 dev_id_priv = list_entry(id_priv->listen_list.next,
 641                                          struct rdma_id_private, listen_list);
 642                 cma_destroy_listen(dev_id_priv);
 643         }
 644         mutex_unlock(&lock);
 645 }
 646
 647 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 648                                  enum cma_state state)
 649 {
 650         switch (state) {
 651         case CMA_ADDR_QUERY:
 652                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 653                 break;
 654         case CMA_ROUTE_QUERY:
 655                 cma_cancel_route(id_priv);
 656                 break;
 657         case CMA_LISTEN:
 658                 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
 659                     !id_priv->cma_dev)
 660                         cma_cancel_listens(id_priv);
 661                 break;
 662         default:
 663                 break;
 664         }
 665 }
 666
 667 static void cma_release_port(struct rdma_id_private *id_priv)
 668 {
 669         struct rdma_bind_list *bind_list = id_priv->bind_list;
 670
 671         if (!bind_list)
 672                 return;
 673
 674         mutex_lock(&lock);
 675         hlist_del(&id_priv->node);
 676         if (hlist_empty(&bind_list->owners)) {
 677                 idr_remove(bind_list->ps, bind_list->port);
 678                 kfree(bind_list);
 679         }
 680         mutex_unlock(&lock);
 681 }
 682
 683 void rdma_destroy_id(struct rdma_cm_id *id)
 684 {
 685         struct rdma_id_private *id_priv;
 686         enum cma_state state;
 687
 688         id_priv = container_of(id, struct rdma_id_private, id);
 689         state = cma_exch(id_priv, CMA_DESTROYING);
 690         cma_cancel_operation(id_priv, state);
 691
 692         if (id_priv->cma_dev) {
 693                 switch (id->device->node_type) {
 694                 case IB_NODE_CA:
 695                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 696                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 697                         break;
 698                 default:
 699                         break;
 700                 }
 701                 mutex_lock(&lock);
 702                 cma_detach_from_dev(id_priv);
 703                 mutex_unlock(&lock);
 704         }
 705
 706         cma_release_port(id_priv);
 707         cma_deref_id(id_priv);
 708         wait_for_completion(&id_priv->comp);
 709
 710         kfree(id_priv->id.route.path_rec);
 711         kfree(id_priv);
 712 }
 713 EXPORT_SYMBOL(rdma_destroy_id);
 714
 715 static int cma_rep_recv(struct rdma_id_private *id_priv)
 716 {
 717         int ret;
 718
 719         ret = cma_modify_qp_rtr(&id_priv->id);
 720         if (ret)
 721                 goto reject;
 722
 723         ret = cma_modify_qp_rts(&id_priv->id);
 724         if (ret)
 725                 goto reject;
 726
 727         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 728         if (ret)
 729                 goto reject;
 730
 731         return 0;
 732 reject:
 733         cma_modify_qp_err(&id_priv->id);
 734         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 735                        NULL, 0, NULL, 0);
 736         return ret;
 737 }
 738
 739 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 740 {
 741         if (id_priv->id.ps == RDMA_PS_SDP &&
 742             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 743             SDP_MAJ_VERSION)
 744                 return -EINVAL;
 745
 746         return 0;
 747 }
 748
 749 static int cma_rtu_recv(struct rdma_id_private *id_priv)
 750 {
 751         int ret;
 752
 753         ret = cma_modify_qp_rts(&id_priv->id);
 754         if (ret)
 755                 goto reject;
 756
 757         return 0;
 758 reject:
 759         cma_modify_qp_err(&id_priv->id);
 760         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 761                        NULL, 0, NULL, 0);
 762         return ret;
 763 }
 764
 765 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 766 {
 767         struct rdma_id_private *id_priv = cm_id->context;
 768         enum rdma_cm_event_type event;
 769         u8 private_data_len = 0;
 770         int ret = 0, status = 0;
 771
 772         atomic_inc(&id_priv->dev_remove);
 773         if (!cma_comp(id_priv, CMA_CONNECT))
 774                 goto out;
 775
 776         switch (ib_event->event) {
 777         case IB_CM_REQ_ERROR:
 778         case IB_CM_REP_ERROR:
 779                 event = RDMA_CM_EVENT_UNREACHABLE;
 780                 status = -ETIMEDOUT;
 781                 break;
 782         case IB_CM_REP_RECEIVED:
 783                 status = cma_verify_rep(id_priv, ib_event->private_data);
 784                 if (status)
 785                         event = RDMA_CM_EVENT_CONNECT_ERROR;
 786                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 787                         status = cma_rep_recv(id_priv);
 788                         event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
 789                                          RDMA_CM_EVENT_ESTABLISHED;
 790                 } else
 791                         event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 792                 private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 793                 break;
 794         case IB_CM_RTU_RECEIVED:
 795                 status = cma_rtu_recv(id_priv);
 796                 event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
 797                                  RDMA_CM_EVENT_ESTABLISHED;
 798                 break;
 799         case IB_CM_DREQ_ERROR:
 800                 status = -ETIMEDOUT; /* fall through */
 801         case IB_CM_DREQ_RECEIVED:
 802         case IB_CM_DREP_RECEIVED:
 803                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 804                         goto out;
 805                 event = RDMA_CM_EVENT_DISCONNECTED;
 806                 break;
 807         case IB_CM_TIMEWAIT_EXIT:
 808         case IB_CM_MRA_RECEIVED:
 809                 /* ignore event */
 810                 goto out;
 811         case IB_CM_REJ_RECEIVED:
 812                 cma_modify_qp_err(&id_priv->id);
 813                 status = ib_event->param.rej_rcvd.reason;
 814                 event = RDMA_CM_EVENT_REJECTED;
 815                 private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 816                 break;
 817         default:
 818                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
 819                        ib_event->event);
 820                 goto out;
 821         }
 822
 823         ret = cma_notify_user(id_priv, event, status, ib_event->private_data,
 824                               private_data_len);
 825         if (ret) {
 826                 /* Destroy the CM ID by returning a non-zero value. */
 827                 id_priv->cm_id.ib = NULL;
 828                 cma_exch(id_priv, CMA_DESTROYING);
 829                 cma_release_remove(id_priv);
 830                 rdma_destroy_id(&id_priv->id);
 831                 return ret;
 832         }
 833 out:
 834         cma_release_remove(id_priv);
 835         return ret;
 836 }
 837
 838 static struct rdma_id_private *cma_new_id(struct rdma_cm_id *listen_id,
 839                                           struct ib_cm_event *ib_event)
 840 {
 841         struct rdma_id_private *id_priv;
 842         struct rdma_cm_id *id;
 843         struct rdma_route *rt;
 844         union cma_ip_addr *src, *dst;
 845         __u16 port;
 846         u8 ip_ver;
 847
 848         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 849                             listen_id->ps);
 850         if (IS_ERR(id))
 851                 return NULL;
 852
 853         rt = &id->route;
 854         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
 855         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths, GFP_KERNEL);
 856         if (!rt->path_rec)
 857                 goto err;
 858
 859         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 860                              &ip_ver, &port, &src, &dst))
 861                 goto err;
 862
 863         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 864                           ip_ver, port, src, dst);
 865         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
 866         if (rt->num_paths == 2)
 867                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
 868
 869         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
 870         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
 871         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
 872         rt->addr.dev_addr.dev_type = IB_NODE_CA;
 873
 874         id_priv = container_of(id, struct rdma_id_private, id);
 875         id_priv->state = CMA_CONNECT;
 876         return id_priv;
 877 err:
 878         rdma_destroy_id(id);
 879         return NULL;
 880 }
 881
 882 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 883 {
 884         struct rdma_id_private *listen_id, *conn_id;
 885         int offset, ret;
 886
 887         listen_id = cm_id->context;
 888         atomic_inc(&listen_id->dev_remove);
 889         if (!cma_comp(listen_id, CMA_LISTEN)) {
 890                 ret = -ECONNABORTED;
 891                 goto out;
 892         }
 893
 894         conn_id = cma_new_id(&listen_id->id, ib_event);
 895         if (!conn_id) {
 896                 ret = -ENOMEM;
 897                 goto out;
 898         }
 899
 900         atomic_inc(&conn_id->dev_remove);
 901         ret = cma_acquire_ib_dev(conn_id);
 902         if (ret) {
 903                 ret = -ENODEV;
 904                 cma_release_remove(conn_id);
 905                 rdma_destroy_id(&conn_id->id);
 906                 goto out;
 907         }
 908
 909         conn_id->cm_id.ib = cm_id;
 910         cm_id->context = conn_id;
 911         cm_id->cm_handler = cma_ib_handler;
 912
 913         offset = cma_user_data_offset(listen_id->id.ps);
 914         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
 915                               ib_event->private_data + offset,
 916                               IB_CM_REQ_PRIVATE_DATA_SIZE - offset);
 917         if (ret) {
 918                 /* Destroy the CM ID by returning a non-zero value. */
 919                 conn_id->cm_id.ib = NULL;
 920                 cma_exch(conn_id, CMA_DESTROYING);
 921                 cma_release_remove(conn_id);
 922                 rdma_destroy_id(&conn_id->id);
 923         }
 924 out:
 925         cma_release_remove(listen_id);
 926         return ret;
 927 }
 928
 929 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
 930 {
 931         return cpu_to_be64(((u64)ps << 16) +
 932                be16_to_cpu(((struct sockaddr_in *) addr)->sin_port));
 933 }
 934
 935 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
 936                                  struct ib_cm_compare_data *compare)
 937 {
 938         struct cma_hdr *cma_data, *cma_mask;
 939         struct sdp_hh *sdp_data, *sdp_mask;
 940         __u32 ip4_addr;
 941         struct in6_addr ip6_addr;
 942
 943         memset(compare, 0, sizeof *compare);
 944         cma_data = (void *) compare->data;
 945         cma_mask = (void *) compare->mask;
 946         sdp_data = (void *) compare->data;
 947         sdp_mask = (void *) compare->mask;
 948
 949         switch (addr->sa_family) {
 950         case AF_INET:
 951                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
 952                 if (ps == RDMA_PS_SDP) {
 953                         sdp_set_ip_ver(sdp_data, 4);
 954                         sdp_set_ip_ver(sdp_mask, 0xF);
 955                         sdp_data->dst_addr.ip4.addr = ip4_addr;
 956                         sdp_mask->dst_addr.ip4.addr = ~0;
 957                 } else {
 958                         cma_set_ip_ver(cma_data, 4);
 959                         cma_set_ip_ver(cma_mask, 0xF);
 960                         cma_data->dst_addr.ip4.addr = ip4_addr;
 961                         cma_mask->dst_addr.ip4.addr = ~0;
 962                 }
 963                 break;
 964         case AF_INET6:
 965                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
 966                 if (ps == RDMA_PS_SDP) {
 967                         sdp_set_ip_ver(sdp_data, 6);
 968                         sdp_set_ip_ver(sdp_mask, 0xF);
 969                         sdp_data->dst_addr.ip6 = ip6_addr;
 970                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
 971                                sizeof sdp_mask->dst_addr.ip6);
 972                 } else {
 973                         cma_set_ip_ver(cma_data, 6);
 974                         cma_set_ip_ver(cma_mask, 0xF);
 975                         cma_data->dst_addr.ip6 = ip6_addr;
 976                         memset(&cma_mask->dst_addr.ip6, 0xFF,
 977                                sizeof cma_mask->dst_addr.ip6);
 978                 }
 979                 break;
 980         default:
 981                 break;
 982         }
 983 }
 984
 985 static int cma_ib_listen(struct rdma_id_private *id_priv)
 986 {
 987         struct ib_cm_compare_data compare_data;
 988         struct sockaddr *addr;
 989         __be64 svc_id;
 990         int ret;
 991
 992         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
 993                                             id_priv);
 994         if (IS_ERR(id_priv->cm_id.ib))
 995                 return PTR_ERR(id_priv->cm_id.ib);
 996
 997         addr = &id_priv->id.route.addr.src_addr;
 998         svc_id = cma_get_service_id(id_priv->id.ps, addr);
 999         if (cma_any_addr(addr))
1000                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1001         else {
1002                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1003                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1004         }
1005
1006         if (ret) {
1007                 ib_destroy_cm_id(id_priv->cm_id.ib);
1008                 id_priv->cm_id.ib = NULL;
1009         }
1010
1011         return ret;
1012 }
1013
1014 static int cma_listen_handler(struct rdma_cm_id *id,
1015                               struct rdma_cm_event *event)
1016 {
1017         struct rdma_id_private *id_priv = id->context;
1018
1019         id->context = id_priv->id.context;
1020         id->event_handler = id_priv->id.event_handler;
1021         return id_priv->id.event_handler(id, event);
1022 }
1023
1024 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1025                               struct cma_device *cma_dev)
1026 {
1027         struct rdma_id_private *dev_id_priv;
1028         struct rdma_cm_id *id;
1029         int ret;
1030
1031         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1032         if (IS_ERR(id))
1033                 return;
1034
1035         dev_id_priv = container_of(id, struct rdma_id_private, id);
1036
1037         dev_id_priv->state = CMA_ADDR_BOUND;
1038         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1039                ip_addr_size(&id_priv->id.route.addr.src_addr));
1040
1041         cma_attach_to_dev(dev_id_priv, cma_dev);
1042         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1043
1044         ret = rdma_listen(id, id_priv->backlog);
1045         if (ret)
1046                 goto err;
1047
1048         return;
1049 err:
1050         cma_destroy_listen(dev_id_priv);
1051 }
1052
1053 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1054 {
1055         struct cma_device *cma_dev;
1056
1057         mutex_lock(&lock);
1058         list_add_tail(&id_priv->list, &listen_any_list);
1059         list_for_each_entry(cma_dev, &dev_list, list)
1060                 cma_listen_on_dev(id_priv, cma_dev);
1061         mutex_unlock(&lock);
1062 }
1063
1064 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1065 {
1066         struct sockaddr_in addr_in;
1067
1068         memset(&addr_in, 0, sizeof addr_in);
1069         addr_in.sin_family = af;
1070         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1071 }
1072
1073 int rdma_listen(struct rdma_cm_id *id, int backlog)
1074 {
1075         struct rdma_id_private *id_priv;
1076         int ret;
1077
1078         id_priv = container_of(id, struct rdma_id_private, id);
1079         if (id_priv->state == CMA_IDLE) {
1080                 ret = cma_bind_any(id, AF_INET);
1081                 if (ret)
1082                         return ret;
1083         }
1084
1085         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1086                 return -EINVAL;
1087
1088         id_priv->backlog = backlog;
1089         if (id->device) {
1090                 switch (id->device->node_type) {
1091                 case IB_NODE_CA:
1092                         ret = cma_ib_listen(id_priv);
1093                         if (ret)
1094                                 goto err;
1095                         break;
1096                 default:
1097                         ret = -ENOSYS;
1098                         goto err;
1099                 }
1100         } else
1101                 cma_listen_on_all(id_priv);
1102
1103         return 0;
1104 err:
1105         id_priv->backlog = 0;
1106         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1107         return ret;
1108 }
1109 EXPORT_SYMBOL(rdma_listen);
1110
1111 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1112                               void *context)
1113 {
1114         struct cma_work *work = context;
1115         struct rdma_route *route;
1116
1117         route = &work->id->id.route;
1118
1119         if (!status) {
1120                 route->num_paths = 1;
1121                 *route->path_rec = *path_rec;
1122         } else {
1123                 work->old_state = CMA_ROUTE_QUERY;
1124                 work->new_state = CMA_ADDR_RESOLVED;
1125                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1126         }
1127
1128         queue_work(cma_wq, &work->work);
1129 }
1130
1131 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1132                               struct cma_work *work)
1133 {
1134         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1135         struct ib_sa_path_rec path_rec;
1136
1137         memset(&path_rec, 0, sizeof path_rec);
1138         ib_addr_get_sgid(addr, &path_rec.sgid);
1139         ib_addr_get_dgid(addr, &path_rec.dgid);
1140         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1141         path_rec.numb_path = 1;
1142
1143         id_priv->query_id = ib_sa_path_rec_get(id_priv->id.device,
1144                                 id_priv->id.port_num, &path_rec,
1145                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1146                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
1147                                 timeout_ms, GFP_KERNEL,
1148                                 cma_query_handler, work, &id_priv->query);
1149
1150         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1151 }
1152
1153 static void cma_work_handler(void *data)
1154 {
1155         struct cma_work *work = data;
1156         struct rdma_id_private *id_priv = work->id;
1157         int destroy = 0;
1158
1159         atomic_inc(&id_priv->dev_remove);
1160         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1161                 goto out;
1162
1163         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1164                 cma_exch(id_priv, CMA_DESTROYING);
1165                 destroy = 1;
1166         }
1167 out:
1168         cma_release_remove(id_priv);
1169         cma_deref_id(id_priv);
1170         if (destroy)
1171                 rdma_destroy_id(&id_priv->id);
1172         kfree(work);
1173 }
1174
1175 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1176 {
1177         struct rdma_route *route = &id_priv->id.route;
1178         struct cma_work *work;
1179         int ret;
1180
1181         work = kzalloc(sizeof *work, GFP_KERNEL);
1182         if (!work)
1183                 return -ENOMEM;
1184
1185         work->id = id_priv;
1186         INIT_WORK(&work->work, cma_work_handler, work);
1187         work->old_state = CMA_ROUTE_QUERY;
1188         work->new_state = CMA_ROUTE_RESOLVED;
1189         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1190
1191         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1192         if (!route->path_rec) {
1193                 ret = -ENOMEM;
1194                 goto err1;
1195         }
1196
1197         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1198         if (ret)
1199                 goto err2;
1200
1201         return 0;
1202 err2:
1203         kfree(route->path_rec);
1204         route->path_rec = NULL;
1205 err1:
1206         kfree(work);
1207         return ret;
1208 }
1209
1210 int rdma_set_ib_paths(struct rdma_cm_id *id,
1211                       struct ib_sa_path_rec *path_rec, int num_paths)
1212 {
1213         struct rdma_id_private *id_priv;
1214         int ret;
1215
1216         id_priv = container_of(id, struct rdma_id_private, id);
1217         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1218                 return -EINVAL;
1219
1220         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1221         if (!id->route.path_rec) {
1222                 ret = -ENOMEM;
1223                 goto err;
1224         }
1225
1226         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1227         return 0;
1228 err:
1229         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1230         return ret;
1231 }
1232 EXPORT_SYMBOL(rdma_set_ib_paths);
1233
1234 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1235 {
1236         struct rdma_id_private *id_priv;
1237         int ret;
1238
1239         id_priv = container_of(id, struct rdma_id_private, id);
1240         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1241                 return -EINVAL;
1242
1243         atomic_inc(&id_priv->refcount);
1244         switch (id->device->node_type) {
1245         case IB_NODE_CA:
1246                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1247                 break;
1248         default:
1249                 ret = -ENOSYS;
1250                 break;
1251         }
1252         if (ret)
1253                 goto err;
1254
1255         return 0;
1256 err:
1257         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1258         cma_deref_id(id_priv);
1259         return ret;
1260 }
1261 EXPORT_SYMBOL(rdma_resolve_route);
1262
1263 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1264 {
1265         struct cma_device *cma_dev;
1266         struct ib_port_attr port_attr;
1267         union ib_gid gid;
1268         u16 pkey;
1269         int ret;
1270         u8 p;
1271
1272         mutex_lock(&lock);
1273         list_for_each_entry(cma_dev, &dev_list, list)
1274                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1275                         if (!ib_query_port (cma_dev->device, p, &port_attr) &&
1276                             port_attr.state == IB_PORT_ACTIVE)
1277                                 goto port_found;
1278
1279         if (!list_empty(&dev_list)) {
1280                 p = 1;
1281                 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1282         } else {
1283                 ret = -ENODEV;
1284                 goto out;
1285         }
1286
1287 port_found:
1288         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1289         if (ret)
1290                 goto out;
1291
1292         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1293         if (ret)
1294                 goto out;
1295
1296         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1297         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1298         id_priv->id.port_num = p;
1299         cma_attach_to_dev(id_priv, cma_dev);
1300 out:
1301         mutex_unlock(&lock);
1302         return ret;
1303 }
1304
1305 static void addr_handler(int status, struct sockaddr *src_addr,
1306                          struct rdma_dev_addr *dev_addr, void *context)
1307 {
1308         struct rdma_id_private *id_priv = context;
1309         enum rdma_cm_event_type event;
1310
1311         atomic_inc(&id_priv->dev_remove);
1312         if (!id_priv->cma_dev && !status)
1313                 status = cma_acquire_dev(id_priv);
1314
1315         if (status) {
1316                 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND))
1317                         goto out;
1318                 event = RDMA_CM_EVENT_ADDR_ERROR;
1319         } else {
1320                 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED))
1321                         goto out;
1322                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1323                        ip_addr_size(src_addr));
1324                 event = RDMA_CM_EVENT_ADDR_RESOLVED;
1325         }
1326
1327         if (cma_notify_user(id_priv, event, status, NULL, 0)) {
1328                 cma_exch(id_priv, CMA_DESTROYING);
1329                 cma_release_remove(id_priv);
1330                 cma_deref_id(id_priv);
1331                 rdma_destroy_id(&id_priv->id);
1332                 return;
1333         }
1334 out:
1335         cma_release_remove(id_priv);
1336         cma_deref_id(id_priv);
1337 }
1338
1339 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1340 {
1341         struct cma_work *work;
1342         struct sockaddr_in *src_in, *dst_in;
1343         union ib_gid gid;
1344         int ret;
1345
1346         work = kzalloc(sizeof *work, GFP_KERNEL);
1347         if (!work)
1348                 return -ENOMEM;
1349
1350         if (!id_priv->cma_dev) {
1351                 ret = cma_bind_loopback(id_priv);
1352                 if (ret)
1353                         goto err;
1354         }
1355
1356         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1357         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1358
1359         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1360                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1361                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1362                 src_in->sin_family = dst_in->sin_family;
1363                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1364         }
1365
1366         work->id = id_priv;
1367         INIT_WORK(&work->work, cma_work_handler, work);
1368         work->old_state = CMA_ADDR_QUERY;
1369         work->new_state = CMA_ADDR_RESOLVED;
1370         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1371         queue_work(cma_wq, &work->work);
1372         return 0;
1373 err:
1374         kfree(work);
1375         return ret;
1376 }
1377
1378 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1379                          struct sockaddr *dst_addr)
1380 {
1381         if (src_addr && src_addr->sa_family)
1382                 return rdma_bind_addr(id, src_addr);
1383         else
1384                 return cma_bind_any(id, dst_addr->sa_family);
1385 }
1386
1387 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1388                       struct sockaddr *dst_addr, int timeout_ms)
1389 {
1390         struct rdma_id_private *id_priv;
1391         int ret;
1392
1393         id_priv = container_of(id, struct rdma_id_private, id);
1394         if (id_priv->state == CMA_IDLE) {
1395                 ret = cma_bind_addr(id, src_addr, dst_addr);
1396                 if (ret)
1397                         return ret;
1398         }
1399
1400         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1401                 return -EINVAL;
1402
1403         atomic_inc(&id_priv->refcount);
1404         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1405         if (cma_any_addr(dst_addr))
1406                 ret = cma_resolve_loopback(id_priv);
1407         else
1408                 ret = rdma_resolve_ip(&id->route.addr.src_addr, dst_addr,
1409                                       &id->route.addr.dev_addr,
1410                                       timeout_ms, addr_handler, id_priv);
1411         if (ret)
1412                 goto err;
1413
1414         return 0;
1415 err:
1416         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1417         cma_deref_id(id_priv);
1418         return ret;
1419 }
1420 EXPORT_SYMBOL(rdma_resolve_addr);
1421
1422 static void cma_bind_port(struct rdma_bind_list *bind_list,
1423                           struct rdma_id_private *id_priv)
1424 {
1425         struct sockaddr_in *sin;
1426
1427         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1428         sin->sin_port = htons(bind_list->port);
1429         id_priv->bind_list = bind_list;
1430         hlist_add_head(&id_priv->node, &bind_list->owners);
1431 }
1432
1433 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1434                           unsigned short snum)
1435 {
1436         struct rdma_bind_list *bind_list;
1437         int port, start, ret;
1438
1439         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1440         if (!bind_list)
1441                 return -ENOMEM;
1442
1443         start = snum ? snum : sysctl_local_port_range[0];
1444
1445         do {
1446                 ret = idr_get_new_above(ps, bind_list, start, &port);
1447         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1448
1449         if (ret)
1450                 goto err;
1451
1452         if ((snum && port != snum) ||
1453             (!snum && port > sysctl_local_port_range[1])) {
1454                 idr_remove(ps, port);
1455                 ret = -EADDRNOTAVAIL;
1456                 goto err;
1457         }
1458
1459         bind_list->ps = ps;
1460         bind_list->port = (unsigned short) port;
1461         cma_bind_port(bind_list, id_priv);
1462         return 0;
1463 err:
1464         kfree(bind_list);
1465         return ret;
1466 }
1467
1468 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1469 {
1470         struct rdma_id_private *cur_id;
1471         struct sockaddr_in *sin, *cur_sin;
1472         struct rdma_bind_list *bind_list;
1473         struct hlist_node *node;
1474         unsigned short snum;
1475
1476         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1477         snum = ntohs(sin->sin_port);
1478         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1479                 return -EACCES;
1480
1481         bind_list = idr_find(ps, snum);
1482         if (!bind_list)
1483                 return cma_alloc_port(ps, id_priv, snum);
1484
1485         /*
1486          * We don't support binding to any address if anyone is bound to
1487          * a specific address on the same port.
1488          */
1489         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1490                 return -EADDRNOTAVAIL;
1491
1492         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1493                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1494                         return -EADDRNOTAVAIL;
1495
1496                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1497                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1498                         return -EADDRINUSE;
1499         }
1500
1501         cma_bind_port(bind_list, id_priv);
1502         return 0;
1503 }
1504
1505 static int cma_get_port(struct rdma_id_private *id_priv)
1506 {
1507         struct idr *ps;
1508         int ret;
1509
1510         switch (id_priv->id.ps) {
1511         case RDMA_PS_SDP:
1512                 ps = &sdp_ps;
1513                 break;
1514         case RDMA_PS_TCP:
1515                 ps = &tcp_ps;
1516                 break;
1517         default:
1518                 return -EPROTONOSUPPORT;
1519         }
1520
1521         mutex_lock(&lock);
1522         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1523                 ret = cma_alloc_port(ps, id_priv, 0);
1524         else
1525                 ret = cma_use_port(ps, id_priv);
1526         mutex_unlock(&lock);
1527
1528         return ret;
1529 }
1530
1531 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1532 {
1533         struct rdma_id_private *id_priv;
1534         int ret;
1535
1536         if (addr->sa_family != AF_INET)
1537                 return -EAFNOSUPPORT;
1538
1539         id_priv = container_of(id, struct rdma_id_private, id);
1540         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1541                 return -EINVAL;
1542
1543         if (!cma_any_addr(addr)) {
1544                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1545                 if (!ret)
1546                         ret = cma_acquire_dev(id_priv);
1547                 if (ret)
1548                         goto err;
1549         }
1550
1551         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
1552         ret = cma_get_port(id_priv);
1553         if (ret)
1554                 goto err;
1555
1556         return 0;
1557 err:
1558         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
1559         return ret;
1560 }
1561 EXPORT_SYMBOL(rdma_bind_addr);
1562
1563 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
1564                           struct rdma_route *route)
1565 {
1566         struct sockaddr_in *src4, *dst4;
1567         struct cma_hdr *cma_hdr;
1568         struct sdp_hh *sdp_hdr;
1569
1570         src4 = (struct sockaddr_in *) &route->addr.src_addr;
1571         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
1572
1573         switch (ps) {
1574         case RDMA_PS_SDP:
1575                 sdp_hdr = hdr;
1576                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
1577                         return -EINVAL;
1578                 sdp_set_ip_ver(sdp_hdr, 4);
1579                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1580                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1581                 sdp_hdr->port = src4->sin_port;
1582                 break;
1583         default:
1584                 cma_hdr = hdr;
1585                 cma_hdr->cma_version = CMA_VERSION;
1586                 cma_set_ip_ver(cma_hdr, 4);
1587                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1588                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1589                 cma_hdr->port = src4->sin_port;
1590                 break;
1591         }
1592         return 0;
1593 }
1594
1595 static int cma_connect_ib(struct rdma_id_private *id_priv,
1596                           struct rdma_conn_param *conn_param)
1597 {
1598         struct ib_cm_req_param req;
1599         struct rdma_route *route;
1600         void *private_data;
1601         int offset, ret;
1602
1603         memset(&req, 0, sizeof req);
1604         offset = cma_user_data_offset(id_priv->id.ps);
1605         req.private_data_len = offset + conn_param->private_data_len;
1606         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
1607         if (!private_data)
1608                 return -ENOMEM;
1609
1610         if (conn_param->private_data && conn_param->private_data_len)
1611                 memcpy(private_data + offset, conn_param->private_data,
1612                        conn_param->private_data_len);
1613
1614         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
1615                                             id_priv);
1616         if (IS_ERR(id_priv->cm_id.ib)) {
1617                 ret = PTR_ERR(id_priv->cm_id.ib);
1618                 goto out;
1619         }
1620
1621         route = &id_priv->id.route;
1622         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
1623         if (ret)
1624                 goto out;
1625         req.private_data = private_data;
1626
1627         req.primary_path = &route->path_rec[0];
1628         if (route->num_paths == 2)
1629                 req.alternate_path = &route->path_rec[1];
1630
1631         req.service_id = cma_get_service_id(id_priv->id.ps,
1632                                             &route->addr.dst_addr);
1633         req.qp_num = id_priv->qp_num;
1634         req.qp_type = id_priv->qp_type;
1635         req.starting_psn = id_priv->seq_num;
1636         req.responder_resources = conn_param->responder_resources;
1637         req.initiator_depth = conn_param->initiator_depth;
1638         req.flow_control = conn_param->flow_control;
1639         req.retry_count = conn_param->retry_count;
1640         req.rnr_retry_count = conn_param->rnr_retry_count;
1641         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1642         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1643         req.max_cm_retries = CMA_MAX_CM_RETRIES;
1644         req.srq = id_priv->srq ? 1 : 0;
1645
1646         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
1647 out:
1648         kfree(private_data);
1649         return ret;
1650 }
1651
1652 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1653 {
1654         struct rdma_id_private *id_priv;
1655         int ret;
1656
1657         id_priv = container_of(id, struct rdma_id_private, id);
1658         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
1659                 return -EINVAL;
1660
1661         if (!id->qp) {
1662                 id_priv->qp_num = conn_param->qp_num;
1663                 id_priv->qp_type = conn_param->qp_type;
1664                 id_priv->srq = conn_param->srq;
1665         }
1666
1667         switch (id->device->node_type) {
1668         case IB_NODE_CA:
1669                 ret = cma_connect_ib(id_priv, conn_param);
1670                 break;
1671         default:
1672                 ret = -ENOSYS;
1673                 break;
1674         }
1675         if (ret)
1676                 goto err;
1677
1678         return 0;
1679 err:
1680         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
1681         return ret;
1682 }
1683 EXPORT_SYMBOL(rdma_connect);
1684
1685 static int cma_accept_ib(struct rdma_id_private *id_priv,
1686                          struct rdma_conn_param *conn_param)
1687 {
1688         struct ib_cm_rep_param rep;
1689         int ret;
1690
1691         ret = cma_modify_qp_rtr(&id_priv->id);
1692         if (ret)
1693                 return ret;
1694
1695         memset(&rep, 0, sizeof rep);
1696         rep.qp_num = id_priv->qp_num;
1697         rep.starting_psn = id_priv->seq_num;
1698         rep.private_data = conn_param->private_data;
1699         rep.private_data_len = conn_param->private_data_len;
1700         rep.responder_resources = conn_param->responder_resources;
1701         rep.initiator_depth = conn_param->initiator_depth;
1702         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
1703         rep.failover_accepted = 0;
1704         rep.flow_control = conn_param->flow_control;
1705         rep.rnr_retry_count = conn_param->rnr_retry_count;
1706         rep.srq = id_priv->srq ? 1 : 0;
1707
1708         return ib_send_cm_rep(id_priv->cm_id.ib, &rep);
1709 }
1710
1711 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1712 {
1713         struct rdma_id_private *id_priv;
1714         int ret;
1715
1716         id_priv = container_of(id, struct rdma_id_private, id);
1717         if (!cma_comp(id_priv, CMA_CONNECT))
1718                 return -EINVAL;
1719
1720         if (!id->qp && conn_param) {
1721                 id_priv->qp_num = conn_param->qp_num;
1722                 id_priv->qp_type = conn_param->qp_type;
1723                 id_priv->srq = conn_param->srq;
1724         }
1725
1726         switch (id->device->node_type) {
1727         case IB_NODE_CA:
1728                 if (conn_param)
1729                         ret = cma_accept_ib(id_priv, conn_param);
1730                 else
1731                         ret = cma_rep_recv(id_priv);
1732                 break;
1733         default:
1734                 ret = -ENOSYS;
1735                 break;
1736         }
1737
1738         if (ret)
1739                 goto reject;
1740
1741         return 0;
1742 reject:
1743         cma_modify_qp_err(id);
1744         rdma_reject(id, NULL, 0);
1745         return ret;
1746 }
1747 EXPORT_SYMBOL(rdma_accept);
1748
1749 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
1750                 u8 private_data_len)
1751 {
1752         struct rdma_id_private *id_priv;
1753         int ret;
1754
1755         id_priv = container_of(id, struct rdma_id_private, id);
1756         if (!cma_comp(id_priv, CMA_CONNECT))
1757                 return -EINVAL;
1758
1759         switch (id->device->node_type) {
1760         case IB_NODE_CA:
1761                 ret = ib_send_cm_rej(id_priv->cm_id.ib,
1762                                      IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
1763                                      private_data, private_data_len);
1764                 break;
1765         default:
1766                 ret = -ENOSYS;
1767                 break;
1768         }
1769         return ret;
1770 }
1771 EXPORT_SYMBOL(rdma_reject);
1772
1773 int rdma_disconnect(struct rdma_cm_id *id)
1774 {
1775         struct rdma_id_private *id_priv;
1776         int ret;
1777
1778         id_priv = container_of(id, struct rdma_id_private, id);
1779         if (!cma_comp(id_priv, CMA_CONNECT) &&
1780             !cma_comp(id_priv, CMA_DISCONNECT))
1781                 return -EINVAL;
1782
1783         ret = cma_modify_qp_err(id);
1784         if (ret)
1785                 goto out;
1786
1787         switch (id->device->node_type) {
1788         case IB_NODE_CA:
1789                 /* Initiate or respond to a disconnect. */
1790                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
1791                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
1792                 break;
1793         default:
1794                 break;
1795         }
1796 out:
1797         return ret;
1798 }
1799 EXPORT_SYMBOL(rdma_disconnect);
1800
1801 static void cma_add_one(struct ib_device *device)
1802 {
1803         struct cma_device *cma_dev;
1804         struct rdma_id_private *id_priv;
1805
1806         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
1807         if (!cma_dev)
1808                 return;
1809
1810         cma_dev->device = device;
1811         cma_dev->node_guid = device->node_guid;
1812         if (!cma_dev->node_guid)
1813                 goto err;
1814
1815         init_completion(&cma_dev->comp);
1816         atomic_set(&cma_dev->refcount, 1);
1817         INIT_LIST_HEAD(&cma_dev->id_list);
1818         ib_set_client_data(device, &cma_client, cma_dev);
1819
1820         mutex_lock(&lock);
1821         list_add_tail(&cma_dev->list, &dev_list);
1822         list_for_each_entry(id_priv, &listen_any_list, list)
1823                 cma_listen_on_dev(id_priv, cma_dev);
1824         mutex_unlock(&lock);
1825         return;
1826 err:
1827         kfree(cma_dev);
1828 }
1829
1830 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
1831 {
1832         enum cma_state state;
1833
1834         /* Record that we want to remove the device */
1835         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
1836         if (state == CMA_DESTROYING)
1837                 return 0;
1838
1839         cma_cancel_operation(id_priv, state);
1840         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
1841
1842         /* Check for destruction from another callback. */
1843         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
1844                 return 0;
1845
1846         return cma_notify_user(id_priv, RDMA_CM_EVENT_DEVICE_REMOVAL,
1847                                0, NULL, 0);
1848 }
1849
1850 static void cma_process_remove(struct cma_device *cma_dev)
1851 {
1852         struct list_head remove_list;
1853         struct rdma_id_private *id_priv;
1854         int ret;
1855
1856         INIT_LIST_HEAD(&remove_list);
1857
1858         mutex_lock(&lock);
1859         while (!list_empty(&cma_dev->id_list)) {
1860                 id_priv = list_entry(cma_dev->id_list.next,
1861                                      struct rdma_id_private, list);
1862
1863                 if (cma_internal_listen(id_priv)) {
1864                         cma_destroy_listen(id_priv);
1865                         continue;
1866                 }
1867
1868                 list_del(&id_priv->list);
1869                 list_add_tail(&id_priv->list, &remove_list);
1870                 atomic_inc(&id_priv->refcount);
1871                 mutex_unlock(&lock);
1872
1873                 ret = cma_remove_id_dev(id_priv);
1874                 cma_deref_id(id_priv);
1875                 if (ret)
1876                         rdma_destroy_id(&id_priv->id);
1877
1878                 mutex_lock(&lock);
1879         }
1880         mutex_unlock(&lock);
1881
1882         cma_deref_dev(cma_dev);
1883         wait_for_completion(&cma_dev->comp);
1884 }
1885
1886 static void cma_remove_one(struct ib_device *device)
1887 {
1888         struct cma_device *cma_dev;
1889
1890         cma_dev = ib_get_client_data(device, &cma_client);
1891         if (!cma_dev)
1892                 return;
1893
1894         mutex_lock(&lock);
1895         list_del(&cma_dev->list);
1896         mutex_unlock(&lock);
1897
1898         cma_process_remove(cma_dev);
1899         kfree(cma_dev);
1900 }
1901
1902 static int cma_init(void)
1903 {
1904         int ret;
1905
1906         cma_wq = create_singlethread_workqueue("rdma_cm_wq");
1907         if (!cma_wq)
1908                 return -ENOMEM;
1909
1910         ret = ib_register_client(&cma_client);
1911         if (ret)
1912                 goto err;
1913         return 0;
1914
1915 err:
1916         destroy_workqueue(cma_wq);
1917         return ret;
1918 }
1919
1920 static void cma_cleanup(void)
1921 {
1922         ib_unregister_client(&cma_client);
1923         destroy_workqueue(cma_wq);
1924         idr_destroy(&sdp_ps);
1925         idr_destroy(&tcp_ps);
1926 }
1927
1928 module_init(cma_init);
1929 module_exit(cma_cleanup);