drivers/infiniband/core/cm.c

   1 /*
   2  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. 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/dma-mapping.h>
  38 #include <linux/device.h>
  39 #include <linux/module.h>
  40 #include <linux/err.h>
  41 #include <linux/idr.h>
  42 #include <linux/interrupt.h>
  43 #include <linux/random.h>
  44 #include <linux/rbtree.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/slab.h>
  47 #include <linux/sysfs.h>
  48 #include <linux/workqueue.h>
  49 #include <linux/kdev_t.h>
  50 #include <linux/etherdevice.h>
  51
  52 #include <rdma/ib_cache.h>
  53 #include <rdma/ib_cm.h>
  54 #include "cm_msgs.h"
  55
  56 MODULE_AUTHOR("Sean Hefty");
  57 MODULE_DESCRIPTION("InfiniBand CM");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59
  60 static const char * const ibcm_rej_reason_strs[] = {
  61         [IB_CM_REJ_NO_QP]                       = "no QP",
  62         [IB_CM_REJ_NO_EEC]                      = "no EEC",
  63         [IB_CM_REJ_NO_RESOURCES]                = "no resources",
  64         [IB_CM_REJ_TIMEOUT]                     = "timeout",
  65         [IB_CM_REJ_UNSUPPORTED]                 = "unsupported",
  66         [IB_CM_REJ_INVALID_COMM_ID]             = "invalid comm ID",
  67         [IB_CM_REJ_INVALID_COMM_INSTANCE]       = "invalid comm instance",
  68         [IB_CM_REJ_INVALID_SERVICE_ID]          = "invalid service ID",
  69         [IB_CM_REJ_INVALID_TRANSPORT_TYPE]      = "invalid transport type",
  70         [IB_CM_REJ_STALE_CONN]                  = "stale conn",
  71         [IB_CM_REJ_RDC_NOT_EXIST]               = "RDC not exist",
  72         [IB_CM_REJ_INVALID_GID]                 = "invalid GID",
  73         [IB_CM_REJ_INVALID_LID]                 = "invalid LID",
  74         [IB_CM_REJ_INVALID_SL]                  = "invalid SL",
  75         [IB_CM_REJ_INVALID_TRAFFIC_CLASS]       = "invalid traffic class",
  76         [IB_CM_REJ_INVALID_HOP_LIMIT]           = "invalid hop limit",
  77         [IB_CM_REJ_INVALID_PACKET_RATE]         = "invalid packet rate",
  78         [IB_CM_REJ_INVALID_ALT_GID]             = "invalid alt GID",
  79         [IB_CM_REJ_INVALID_ALT_LID]             = "invalid alt LID",
  80         [IB_CM_REJ_INVALID_ALT_SL]              = "invalid alt SL",
  81         [IB_CM_REJ_INVALID_ALT_TRAFFIC_CLASS]   = "invalid alt traffic class",
  82         [IB_CM_REJ_INVALID_ALT_HOP_LIMIT]       = "invalid alt hop limit",
  83         [IB_CM_REJ_INVALID_ALT_PACKET_RATE]     = "invalid alt packet rate",
  84         [IB_CM_REJ_PORT_CM_REDIRECT]            = "port CM redirect",
  85         [IB_CM_REJ_PORT_REDIRECT]               = "port redirect",
  86         [IB_CM_REJ_INVALID_MTU]                 = "invalid MTU",
  87         [IB_CM_REJ_INSUFFICIENT_RESP_RESOURCES] = "insufficient resp resources",
  88         [IB_CM_REJ_CONSUMER_DEFINED]            = "consumer defined",
  89         [IB_CM_REJ_INVALID_RNR_RETRY]           = "invalid RNR retry",
  90         [IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID]     = "duplicate local comm ID",
  91         [IB_CM_REJ_INVALID_CLASS_VERSION]       = "invalid class version",
  92         [IB_CM_REJ_INVALID_FLOW_LABEL]          = "invalid flow label",
  93         [IB_CM_REJ_INVALID_ALT_FLOW_LABEL]      = "invalid alt flow label",
  94 };
  95
  96 const char *__attribute_const__ ibcm_reject_msg(int reason)
  97 {
  98         size_t index = reason;
  99
 100         if (index < ARRAY_SIZE(ibcm_rej_reason_strs) &&
 101             ibcm_rej_reason_strs[index])
 102                 return ibcm_rej_reason_strs[index];
 103         else
 104                 return "unrecognized reason";
 105 }
 106 EXPORT_SYMBOL(ibcm_reject_msg);
 107
 108 static void cm_add_one(struct ib_device *device);
 109 static void cm_remove_one(struct ib_device *device, void *client_data);
 110
 111 static struct ib_client cm_client = {
 112         .name   = "cm",
 113         .add    = cm_add_one,
 114         .remove = cm_remove_one
 115 };
 116
 117 static struct ib_cm {
 118         spinlock_t lock;
 119         struct list_head device_list;
 120         rwlock_t device_lock;
 121         struct rb_root listen_service_table;
 122         u64 listen_service_id;
 123         /* struct rb_root peer_service_table; todo: fix peer to peer */
 124         struct rb_root remote_qp_table;
 125         struct rb_root remote_id_table;
 126         struct rb_root remote_sidr_table;
 127         struct idr local_id_table;
 128         __be32 random_id_operand;
 129         struct list_head timewait_list;
 130         struct workqueue_struct *wq;
 131         /* Sync on cm change port state */
 132         spinlock_t state_lock;
 133 } cm;
 134
 135 /* Counter indexes ordered by attribute ID */
 136 enum {
 137         CM_REQ_COUNTER,
 138         CM_MRA_COUNTER,
 139         CM_REJ_COUNTER,
 140         CM_REP_COUNTER,
 141         CM_RTU_COUNTER,
 142         CM_DREQ_COUNTER,
 143         CM_DREP_COUNTER,
 144         CM_SIDR_REQ_COUNTER,
 145         CM_SIDR_REP_COUNTER,
 146         CM_LAP_COUNTER,
 147         CM_APR_COUNTER,
 148         CM_ATTR_COUNT,
 149         CM_ATTR_ID_OFFSET = 0x0010,
 150 };
 151
 152 enum {
 153         CM_XMIT,
 154         CM_XMIT_RETRIES,
 155         CM_RECV,
 156         CM_RECV_DUPLICATES,
 157         CM_COUNTER_GROUPS
 158 };
 159
 160 static char const counter_group_names[CM_COUNTER_GROUPS]
 161                                      [sizeof("cm_rx_duplicates")] = {
 162         "cm_tx_msgs", "cm_tx_retries",
 163         "cm_rx_msgs", "cm_rx_duplicates"
 164 };
 165
 166 struct cm_counter_group {
 167         struct kobject obj;
 168         atomic_long_t counter[CM_ATTR_COUNT];
 169 };
 170
 171 struct cm_counter_attribute {
 172         struct attribute attr;
 173         int index;
 174 };
 175
 176 #define CM_COUNTER_ATTR(_name, _index) \
 177 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 178         .attr = { .name = __stringify(_name), .mode = 0444 }, \
 179         .index = _index \
 180 }
 181
 182 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 183 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 184 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 185 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 186 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 187 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 188 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 189 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 190 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 191 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 192 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 193
 194 static struct attribute *cm_counter_default_attrs[] = {
 195         &cm_req_counter_attr.attr,
 196         &cm_mra_counter_attr.attr,
 197         &cm_rej_counter_attr.attr,
 198         &cm_rep_counter_attr.attr,
 199         &cm_rtu_counter_attr.attr,
 200         &cm_dreq_counter_attr.attr,
 201         &cm_drep_counter_attr.attr,
 202         &cm_sidr_req_counter_attr.attr,
 203         &cm_sidr_rep_counter_attr.attr,
 204         &cm_lap_counter_attr.attr,
 205         &cm_apr_counter_attr.attr,
 206         NULL
 207 };
 208
 209 struct cm_port {
 210         struct cm_device *cm_dev;
 211         struct ib_mad_agent *mad_agent;
 212         struct kobject port_obj;
 213         u8 port_num;
 214         struct list_head cm_priv_prim_list;
 215         struct list_head cm_priv_altr_list;
 216         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 217 };
 218
 219 struct cm_device {
 220         struct list_head list;
 221         struct ib_device *ib_device;
 222         struct device *device;
 223         u8 ack_delay;
 224         int going_down;
 225         struct cm_port *port[0];
 226 };
 227
 228 struct cm_av {
 229         struct cm_port *port;
 230         union ib_gid dgid;
 231         struct rdma_ah_attr ah_attr;
 232         u16 pkey_index;
 233         u8 timeout;
 234 };
 235
 236 struct cm_work {
 237         struct delayed_work work;
 238         struct list_head list;
 239         struct cm_port *port;
 240         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 241         __be32 local_id;                        /* Established / timewait */
 242         __be32 remote_id;
 243         struct ib_cm_event cm_event;
 244         struct ib_sa_path_rec path[0];
 245 };
 246
 247 struct cm_timewait_info {
 248         struct cm_work work;                    /* Must be first. */
 249         struct list_head list;
 250         struct rb_node remote_qp_node;
 251         struct rb_node remote_id_node;
 252         __be64 remote_ca_guid;
 253         __be32 remote_qpn;
 254         u8 inserted_remote_qp;
 255         u8 inserted_remote_id;
 256 };
 257
 258 struct cm_id_private {
 259         struct ib_cm_id id;
 260
 261         struct rb_node service_node;
 262         struct rb_node sidr_id_node;
 263         spinlock_t lock;        /* Do not acquire inside cm.lock */
 264         struct completion comp;
 265         atomic_t refcount;
 266         /* Number of clients sharing this ib_cm_id. Only valid for listeners.
 267          * Protected by the cm.lock spinlock. */
 268         int listen_sharecount;
 269
 270         struct ib_mad_send_buf *msg;
 271         struct cm_timewait_info *timewait_info;
 272         /* todo: use alternate port on send failure */
 273         struct cm_av av;
 274         struct cm_av alt_av;
 275
 276         void *private_data;
 277         __be64 tid;
 278         __be32 local_qpn;
 279         __be32 remote_qpn;
 280         enum ib_qp_type qp_type;
 281         __be32 sq_psn;
 282         __be32 rq_psn;
 283         int timeout_ms;
 284         enum ib_mtu path_mtu;
 285         __be16 pkey;
 286         u8 private_data_len;
 287         u8 max_cm_retries;
 288         u8 peer_to_peer;
 289         u8 responder_resources;
 290         u8 initiator_depth;
 291         u8 retry_count;
 292         u8 rnr_retry_count;
 293         u8 service_timeout;
 294         u8 target_ack_delay;
 295
 296         struct list_head prim_list;
 297         struct list_head altr_list;
 298         /* Indicates that the send port mad is registered and av is set */
 299         int prim_send_port_not_ready;
 300         int altr_send_port_not_ready;
 301
 302         struct list_head work_list;
 303         atomic_t work_count;
 304 };
 305
 306 static void cm_work_handler(struct work_struct *work);
 307
 308 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 309 {
 310         if (atomic_dec_and_test(&cm_id_priv->refcount))
 311                 complete(&cm_id_priv->comp);
 312 }
 313
 314 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 315                         struct ib_mad_send_buf **msg)
 316 {
 317         struct ib_mad_agent *mad_agent;
 318         struct ib_mad_send_buf *m;
 319         struct ib_ah *ah;
 320         struct cm_av *av;
 321         unsigned long flags, flags2;
 322         int ret = 0;
 323
 324         /* don't let the port to be released till the agent is down */
 325         spin_lock_irqsave(&cm.state_lock, flags2);
 326         spin_lock_irqsave(&cm.lock, flags);
 327         if (!cm_id_priv->prim_send_port_not_ready)
 328                 av = &cm_id_priv->av;
 329         else if (!cm_id_priv->altr_send_port_not_ready &&
 330                  (cm_id_priv->alt_av.port))
 331                 av = &cm_id_priv->alt_av;
 332         else {
 333                 pr_info("%s: not valid CM id\n", __func__);
 334                 ret = -ENODEV;
 335                 spin_unlock_irqrestore(&cm.lock, flags);
 336                 goto out;
 337         }
 338         spin_unlock_irqrestore(&cm.lock, flags);
 339         /* Make sure the port haven't released the mad yet */
 340         mad_agent = cm_id_priv->av.port->mad_agent;
 341         if (!mad_agent) {
 342                 pr_info("%s: not a valid MAD agent\n", __func__);
 343                 ret = -ENODEV;
 344                 goto out;
 345         }
 346         ah = rdma_create_ah(mad_agent->qp->pd, &av->ah_attr);
 347         if (IS_ERR(ah)) {
 348                 ret = PTR_ERR(ah);
 349                 goto out;
 350         }
 351
 352         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 353                                av->pkey_index,
 354                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 355                                GFP_ATOMIC,
 356                                IB_MGMT_BASE_VERSION);
 357         if (IS_ERR(m)) {
 358                 rdma_destroy_ah(ah);
 359                 ret = PTR_ERR(m);
 360                 goto out;
 361         }
 362
 363         /* Timeout set by caller if response is expected. */
 364         m->ah = ah;
 365         m->retries = cm_id_priv->max_cm_retries;
 366
 367         atomic_inc(&cm_id_priv->refcount);
 368         m->context[0] = cm_id_priv;
 369         *msg = m;
 370
 371 out:
 372         spin_unlock_irqrestore(&cm.state_lock, flags2);
 373         return ret;
 374 }
 375
 376 static int cm_alloc_response_msg(struct cm_port *port,
 377                                  struct ib_mad_recv_wc *mad_recv_wc,
 378                                  struct ib_mad_send_buf **msg)
 379 {
 380         struct ib_mad_send_buf *m;
 381         struct ib_ah *ah;
 382
 383         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 384                                   mad_recv_wc->recv_buf.grh, port->port_num);
 385         if (IS_ERR(ah))
 386                 return PTR_ERR(ah);
 387
 388         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 389                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 390                                GFP_ATOMIC,
 391                                IB_MGMT_BASE_VERSION);
 392         if (IS_ERR(m)) {
 393                 rdma_destroy_ah(ah);
 394                 return PTR_ERR(m);
 395         }
 396         m->ah = ah;
 397         *msg = m;
 398         return 0;
 399 }
 400
 401 static void cm_free_msg(struct ib_mad_send_buf *msg)
 402 {
 403         rdma_destroy_ah(msg->ah);
 404         if (msg->context[0])
 405                 cm_deref_id(msg->context[0]);
 406         ib_free_send_mad(msg);
 407 }
 408
 409 static void * cm_copy_private_data(const void *private_data,
 410                                    u8 private_data_len)
 411 {
 412         void *data;
 413
 414         if (!private_data || !private_data_len)
 415                 return NULL;
 416
 417         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 418         if (!data)
 419                 return ERR_PTR(-ENOMEM);
 420
 421         return data;
 422 }
 423
 424 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 425                                  void *private_data, u8 private_data_len)
 426 {
 427         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 428                 kfree(cm_id_priv->private_data);
 429
 430         cm_id_priv->private_data = private_data;
 431         cm_id_priv->private_data_len = private_data_len;
 432 }
 433
 434 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 435                                     struct ib_grh *grh, struct cm_av *av)
 436 {
 437         av->port = port;
 438         av->pkey_index = wc->pkey_index;
 439         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 440                            grh, &av->ah_attr);
 441 }
 442
 443 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av,
 444                               struct cm_id_private *cm_id_priv)
 445 {
 446         struct cm_device *cm_dev;
 447         struct cm_port *port = NULL;
 448         unsigned long flags;
 449         int ret;
 450         u8 p;
 451         struct net_device *ndev = ib_get_ndev_from_path(path);
 452
 453         read_lock_irqsave(&cm.device_lock, flags);
 454         list_for_each_entry(cm_dev, &cm.device_list, list) {
 455                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 456                                         path->gid_type, ndev, &p, NULL)) {
 457                         port = cm_dev->port[p-1];
 458                         break;
 459                 }
 460         }
 461         read_unlock_irqrestore(&cm.device_lock, flags);
 462
 463         if (ndev)
 464                 dev_put(ndev);
 465
 466         if (!port)
 467                 return -EINVAL;
 468
 469         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 470                                   be16_to_cpu(path->pkey), &av->pkey_index);
 471         if (ret)
 472                 return ret;
 473
 474         av->port = port;
 475         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 476                              &av->ah_attr);
 477         av->timeout = path->packet_life_time + 1;
 478
 479         spin_lock_irqsave(&cm.lock, flags);
 480         if (&cm_id_priv->av == av)
 481                 list_add_tail(&cm_id_priv->prim_list, &port->cm_priv_prim_list);
 482         else if (&cm_id_priv->alt_av == av)
 483                 list_add_tail(&cm_id_priv->altr_list, &port->cm_priv_altr_list);
 484         else
 485                 ret = -EINVAL;
 486
 487         spin_unlock_irqrestore(&cm.lock, flags);
 488
 489         return ret;
 490 }
 491
 492 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 493 {
 494         unsigned long flags;
 495         int id;
 496
 497         idr_preload(GFP_KERNEL);
 498         spin_lock_irqsave(&cm.lock, flags);
 499
 500         id = idr_alloc_cyclic(&cm.local_id_table, cm_id_priv, 0, 0, GFP_NOWAIT);
 501
 502         spin_unlock_irqrestore(&cm.lock, flags);
 503         idr_preload_end();
 504
 505         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 506         return id < 0 ? id : 0;
 507 }
 508
 509 static void cm_free_id(__be32 local_id)
 510 {
 511         spin_lock_irq(&cm.lock);
 512         idr_remove(&cm.local_id_table,
 513                    (__force int) (local_id ^ cm.random_id_operand));
 514         spin_unlock_irq(&cm.lock);
 515 }
 516
 517 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 518 {
 519         struct cm_id_private *cm_id_priv;
 520
 521         cm_id_priv = idr_find(&cm.local_id_table,
 522                               (__force int) (local_id ^ cm.random_id_operand));
 523         if (cm_id_priv) {
 524                 if (cm_id_priv->id.remote_id == remote_id)
 525                         atomic_inc(&cm_id_priv->refcount);
 526                 else
 527                         cm_id_priv = NULL;
 528         }
 529
 530         return cm_id_priv;
 531 }
 532
 533 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 534 {
 535         struct cm_id_private *cm_id_priv;
 536
 537         spin_lock_irq(&cm.lock);
 538         cm_id_priv = cm_get_id(local_id, remote_id);
 539         spin_unlock_irq(&cm.lock);
 540
 541         return cm_id_priv;
 542 }
 543
 544 /*
 545  * Trivial helpers to strip endian annotation and compare; the
 546  * endianness doesn't actually matter since we just need a stable
 547  * order for the RB tree.
 548  */
 549 static int be32_lt(__be32 a, __be32 b)
 550 {
 551         return (__force u32) a < (__force u32) b;
 552 }
 553
 554 static int be32_gt(__be32 a, __be32 b)
 555 {
 556         return (__force u32) a > (__force u32) b;
 557 }
 558
 559 static int be64_lt(__be64 a, __be64 b)
 560 {
 561         return (__force u64) a < (__force u64) b;
 562 }
 563
 564 static int be64_gt(__be64 a, __be64 b)
 565 {
 566         return (__force u64) a > (__force u64) b;
 567 }
 568
 569 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 570 {
 571         struct rb_node **link = &cm.listen_service_table.rb_node;
 572         struct rb_node *parent = NULL;
 573         struct cm_id_private *cur_cm_id_priv;
 574         __be64 service_id = cm_id_priv->id.service_id;
 575         __be64 service_mask = cm_id_priv->id.service_mask;
 576
 577         while (*link) {
 578                 parent = *link;
 579                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 580                                           service_node);
 581                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 582                     (service_mask & cur_cm_id_priv->id.service_id) &&
 583                     (cm_id_priv->id.device == cur_cm_id_priv->id.device))
 584                         return cur_cm_id_priv;
 585
 586                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 587                         link = &(*link)->rb_left;
 588                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 589                         link = &(*link)->rb_right;
 590                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 591                         link = &(*link)->rb_left;
 592                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 593                         link = &(*link)->rb_right;
 594                 else
 595                         link = &(*link)->rb_right;
 596         }
 597         rb_link_node(&cm_id_priv->service_node, parent, link);
 598         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 599         return NULL;
 600 }
 601
 602 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 603                                              __be64 service_id)
 604 {
 605         struct rb_node *node = cm.listen_service_table.rb_node;
 606         struct cm_id_private *cm_id_priv;
 607
 608         while (node) {
 609                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 610                 if ((cm_id_priv->id.service_mask & service_id) ==
 611                      cm_id_priv->id.service_id &&
 612                     (cm_id_priv->id.device == device))
 613                         return cm_id_priv;
 614
 615                 if (device < cm_id_priv->id.device)
 616                         node = node->rb_left;
 617                 else if (device > cm_id_priv->id.device)
 618                         node = node->rb_right;
 619                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 620                         node = node->rb_left;
 621                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 622                         node = node->rb_right;
 623                 else
 624                         node = node->rb_right;
 625         }
 626         return NULL;
 627 }
 628
 629 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 630                                                      *timewait_info)
 631 {
 632         struct rb_node **link = &cm.remote_id_table.rb_node;
 633         struct rb_node *parent = NULL;
 634         struct cm_timewait_info *cur_timewait_info;
 635         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 636         __be32 remote_id = timewait_info->work.remote_id;
 637
 638         while (*link) {
 639                 parent = *link;
 640                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 641                                              remote_id_node);
 642                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 643                         link = &(*link)->rb_left;
 644                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 645                         link = &(*link)->rb_right;
 646                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 647                         link = &(*link)->rb_left;
 648                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 649                         link = &(*link)->rb_right;
 650                 else
 651                         return cur_timewait_info;
 652         }
 653         timewait_info->inserted_remote_id = 1;
 654         rb_link_node(&timewait_info->remote_id_node, parent, link);
 655         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 656         return NULL;
 657 }
 658
 659 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 660                                                    __be32 remote_id)
 661 {
 662         struct rb_node *node = cm.remote_id_table.rb_node;
 663         struct cm_timewait_info *timewait_info;
 664
 665         while (node) {
 666                 timewait_info = rb_entry(node, struct cm_timewait_info,
 667                                          remote_id_node);
 668                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 669                         node = node->rb_left;
 670                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 671                         node = node->rb_right;
 672                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 673                         node = node->rb_left;
 674                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 675                         node = node->rb_right;
 676                 else
 677                         return timewait_info;
 678         }
 679         return NULL;
 680 }
 681
 682 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 683                                                       *timewait_info)
 684 {
 685         struct rb_node **link = &cm.remote_qp_table.rb_node;
 686         struct rb_node *parent = NULL;
 687         struct cm_timewait_info *cur_timewait_info;
 688         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 689         __be32 remote_qpn = timewait_info->remote_qpn;
 690
 691         while (*link) {
 692                 parent = *link;
 693                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 694                                              remote_qp_node);
 695                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 696                         link = &(*link)->rb_left;
 697                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 698                         link = &(*link)->rb_right;
 699                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 700                         link = &(*link)->rb_left;
 701                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 702                         link = &(*link)->rb_right;
 703                 else
 704                         return cur_timewait_info;
 705         }
 706         timewait_info->inserted_remote_qp = 1;
 707         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 708         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 709         return NULL;
 710 }
 711
 712 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 713                                                     *cm_id_priv)
 714 {
 715         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 716         struct rb_node *parent = NULL;
 717         struct cm_id_private *cur_cm_id_priv;
 718         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 719         __be32 remote_id = cm_id_priv->id.remote_id;
 720
 721         while (*link) {
 722                 parent = *link;
 723                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 724                                           sidr_id_node);
 725                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 726                         link = &(*link)->rb_left;
 727                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 728                         link = &(*link)->rb_right;
 729                 else {
 730                         int cmp;
 731                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 732                                      sizeof *port_gid);
 733                         if (cmp < 0)
 734                                 link = &(*link)->rb_left;
 735                         else if (cmp > 0)
 736                                 link = &(*link)->rb_right;
 737                         else
 738                                 return cur_cm_id_priv;
 739                 }
 740         }
 741         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 742         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 743         return NULL;
 744 }
 745
 746 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 747                                enum ib_cm_sidr_status status)
 748 {
 749         struct ib_cm_sidr_rep_param param;
 750
 751         memset(&param, 0, sizeof param);
 752         param.status = status;
 753         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 754 }
 755
 756 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 757                                  ib_cm_handler cm_handler,
 758                                  void *context)
 759 {
 760         struct cm_id_private *cm_id_priv;
 761         int ret;
 762
 763         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 764         if (!cm_id_priv)
 765                 return ERR_PTR(-ENOMEM);
 766
 767         cm_id_priv->id.state = IB_CM_IDLE;
 768         cm_id_priv->id.device = device;
 769         cm_id_priv->id.cm_handler = cm_handler;
 770         cm_id_priv->id.context = context;
 771         cm_id_priv->id.remote_cm_qpn = 1;
 772         ret = cm_alloc_id(cm_id_priv);
 773         if (ret)
 774                 goto error;
 775
 776         spin_lock_init(&cm_id_priv->lock);
 777         init_completion(&cm_id_priv->comp);
 778         INIT_LIST_HEAD(&cm_id_priv->work_list);
 779         INIT_LIST_HEAD(&cm_id_priv->prim_list);
 780         INIT_LIST_HEAD(&cm_id_priv->altr_list);
 781         atomic_set(&cm_id_priv->work_count, -1);
 782         atomic_set(&cm_id_priv->refcount, 1);
 783         return &cm_id_priv->id;
 784
 785 error:
 786         kfree(cm_id_priv);
 787         return ERR_PTR(-ENOMEM);
 788 }
 789 EXPORT_SYMBOL(ib_create_cm_id);
 790
 791 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 792 {
 793         struct cm_work *work;
 794
 795         if (list_empty(&cm_id_priv->work_list))
 796                 return NULL;
 797
 798         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 799         list_del(&work->list);
 800         return work;
 801 }
 802
 803 static void cm_free_work(struct cm_work *work)
 804 {
 805         if (work->mad_recv_wc)
 806                 ib_free_recv_mad(work->mad_recv_wc);
 807         kfree(work);
 808 }
 809
 810 static inline int cm_convert_to_ms(int iba_time)
 811 {
 812         /* approximate conversion to ms from 4.096us x 2^iba_time */
 813         return 1 << max(iba_time - 8, 0);
 814 }
 815
 816 /*
 817  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 818  * Because of how ack_timeout is stored, adding one doubles the timeout.
 819  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 820  * increment it (round up) only if the other is within 50%.
 821  */
 822 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 823 {
 824         int ack_timeout = packet_life_time + 1;
 825
 826         if (ack_timeout >= ca_ack_delay)
 827                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 828         else
 829                 ack_timeout = ca_ack_delay +
 830                               (ack_timeout >= (ca_ack_delay - 1));
 831
 832         return min(31, ack_timeout);
 833 }
 834
 835 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 836 {
 837         if (timewait_info->inserted_remote_id) {
 838                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 839                 timewait_info->inserted_remote_id = 0;
 840         }
 841
 842         if (timewait_info->inserted_remote_qp) {
 843                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 844                 timewait_info->inserted_remote_qp = 0;
 845         }
 846 }
 847
 848 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 849 {
 850         struct cm_timewait_info *timewait_info;
 851
 852         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 853         if (!timewait_info)
 854                 return ERR_PTR(-ENOMEM);
 855
 856         timewait_info->work.local_id = local_id;
 857         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 858         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 859         return timewait_info;
 860 }
 861
 862 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 863 {
 864         int wait_time;
 865         unsigned long flags;
 866         struct cm_device *cm_dev;
 867
 868         cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
 869         if (!cm_dev)
 870                 return;
 871
 872         spin_lock_irqsave(&cm.lock, flags);
 873         cm_cleanup_timewait(cm_id_priv->timewait_info);
 874         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 875         spin_unlock_irqrestore(&cm.lock, flags);
 876
 877         /*
 878          * The cm_id could be destroyed by the user before we exit timewait.
 879          * To protect against this, we search for the cm_id after exiting
 880          * timewait before notifying the user that we've exited timewait.
 881          */
 882         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 883         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 884
 885         /* Check if the device started its remove_one */
 886         spin_lock_irqsave(&cm.lock, flags);
 887         if (!cm_dev->going_down)
 888                 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 889                                    msecs_to_jiffies(wait_time));
 890         spin_unlock_irqrestore(&cm.lock, flags);
 891
 892         cm_id_priv->timewait_info = NULL;
 893 }
 894
 895 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 896 {
 897         unsigned long flags;
 898
 899         cm_id_priv->id.state = IB_CM_IDLE;
 900         if (cm_id_priv->timewait_info) {
 901                 spin_lock_irqsave(&cm.lock, flags);
 902                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 903                 spin_unlock_irqrestore(&cm.lock, flags);
 904                 kfree(cm_id_priv->timewait_info);
 905                 cm_id_priv->timewait_info = NULL;
 906         }
 907 }
 908
 909 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 910 {
 911         struct cm_id_private *cm_id_priv;
 912         struct cm_work *work;
 913
 914         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 915 retest:
 916         spin_lock_irq(&cm_id_priv->lock);
 917         switch (cm_id->state) {
 918         case IB_CM_LISTEN:
 919                 spin_unlock_irq(&cm_id_priv->lock);
 920
 921                 spin_lock_irq(&cm.lock);
 922                 if (--cm_id_priv->listen_sharecount > 0) {
 923                         /* The id is still shared. */
 924                         cm_deref_id(cm_id_priv);
 925                         spin_unlock_irq(&cm.lock);
 926                         return;
 927                 }
 928                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 929                 spin_unlock_irq(&cm.lock);
 930                 break;
 931         case IB_CM_SIDR_REQ_SENT:
 932                 cm_id->state = IB_CM_IDLE;
 933                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 934                 spin_unlock_irq(&cm_id_priv->lock);
 935                 break;
 936         case IB_CM_SIDR_REQ_RCVD:
 937                 spin_unlock_irq(&cm_id_priv->lock);
 938                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 939                 spin_lock_irq(&cm.lock);
 940                 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node))
 941                         rb_erase(&cm_id_priv->sidr_id_node,
 942                                  &cm.remote_sidr_table);
 943                 spin_unlock_irq(&cm.lock);
 944                 break;
 945         case IB_CM_REQ_SENT:
 946         case IB_CM_MRA_REQ_RCVD:
 947                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 948                 spin_unlock_irq(&cm_id_priv->lock);
 949                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 950                                &cm_id_priv->id.device->node_guid,
 951                                sizeof cm_id_priv->id.device->node_guid,
 952                                NULL, 0);
 953                 break;
 954         case IB_CM_REQ_RCVD:
 955                 if (err == -ENOMEM) {
 956                         /* Do not reject to allow future retries. */
 957                         cm_reset_to_idle(cm_id_priv);
 958                         spin_unlock_irq(&cm_id_priv->lock);
 959                 } else {
 960                         spin_unlock_irq(&cm_id_priv->lock);
 961                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 962                                        NULL, 0, NULL, 0);
 963                 }
 964                 break;
 965         case IB_CM_REP_SENT:
 966         case IB_CM_MRA_REP_RCVD:
 967                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 968                 /* Fall through */
 969         case IB_CM_MRA_REQ_SENT:
 970         case IB_CM_REP_RCVD:
 971         case IB_CM_MRA_REP_SENT:
 972                 spin_unlock_irq(&cm_id_priv->lock);
 973                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 974                                NULL, 0, NULL, 0);
 975                 break;
 976         case IB_CM_ESTABLISHED:
 977                 spin_unlock_irq(&cm_id_priv->lock);
 978                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT)
 979                         break;
 980                 ib_send_cm_dreq(cm_id, NULL, 0);
 981                 goto retest;
 982         case IB_CM_DREQ_SENT:
 983                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 984                 cm_enter_timewait(cm_id_priv);
 985                 spin_unlock_irq(&cm_id_priv->lock);
 986                 break;
 987         case IB_CM_DREQ_RCVD:
 988                 spin_unlock_irq(&cm_id_priv->lock);
 989                 ib_send_cm_drep(cm_id, NULL, 0);
 990                 break;
 991         default:
 992                 spin_unlock_irq(&cm_id_priv->lock);
 993                 break;
 994         }
 995
 996         spin_lock_irq(&cm.lock);
 997         if (!list_empty(&cm_id_priv->altr_list) &&
 998             (!cm_id_priv->altr_send_port_not_ready))
 999                 list_del(&cm_id_priv->altr_list);
1000         if (!list_empty(&cm_id_priv->prim_list) &&
1001             (!cm_id_priv->prim_send_port_not_ready))
1002                 list_del(&cm_id_priv->prim_list);
1003         spin_unlock_irq(&cm.lock);
1004
1005         cm_free_id(cm_id->local_id);
1006         cm_deref_id(cm_id_priv);
1007         wait_for_completion(&cm_id_priv->comp);
1008         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
1009                 cm_free_work(work);
1010         kfree(cm_id_priv->private_data);
1011         kfree(cm_id_priv);
1012 }
1013
1014 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
1015 {
1016         cm_destroy_id(cm_id, 0);
1017 }
1018 EXPORT_SYMBOL(ib_destroy_cm_id);
1019
1020 /**
1021  * __ib_cm_listen - Initiates listening on the specified service ID for
1022  *   connection and service ID resolution requests.
1023  * @cm_id: Connection identifier associated with the listen request.
1024  * @service_id: Service identifier matched against incoming connection
1025  *   and service ID resolution requests.  The service ID should be specified
1026  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1027  *   assign a service ID to the caller.
1028  * @service_mask: Mask applied to service ID used to listen across a
1029  *   range of service IDs.  If set to 0, the service ID is matched
1030  *   exactly.  This parameter is ignored if %service_id is set to
1031  *   IB_CM_ASSIGN_SERVICE_ID.
1032  */
1033 static int __ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id,
1034                           __be64 service_mask)
1035 {
1036         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
1037         int ret = 0;
1038
1039         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
1040         service_id &= service_mask;
1041         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
1042             (service_id != IB_CM_ASSIGN_SERVICE_ID))
1043                 return -EINVAL;
1044
1045         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1046         if (cm_id->state != IB_CM_IDLE)
1047                 return -EINVAL;
1048
1049         cm_id->state = IB_CM_LISTEN;
1050         ++cm_id_priv->listen_sharecount;
1051
1052         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
1053                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
1054                 cm_id->service_mask = ~cpu_to_be64(0);
1055         } else {
1056                 cm_id->service_id = service_id;
1057                 cm_id->service_mask = service_mask;
1058         }
1059         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
1060
1061         if (cur_cm_id_priv) {
1062                 cm_id->state = IB_CM_IDLE;
1063                 --cm_id_priv->listen_sharecount;
1064                 ret = -EBUSY;
1065         }
1066         return ret;
1067 }
1068
1069 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask)
1070 {
1071         unsigned long flags;
1072         int ret;
1073
1074         spin_lock_irqsave(&cm.lock, flags);
1075         ret = __ib_cm_listen(cm_id, service_id, service_mask);
1076         spin_unlock_irqrestore(&cm.lock, flags);
1077
1078         return ret;
1079 }
1080 EXPORT_SYMBOL(ib_cm_listen);
1081
1082 /**
1083  * Create a new listening ib_cm_id and listen on the given service ID.
1084  *
1085  * If there's an existing ID listening on that same device and service ID,
1086  * return it.
1087  *
1088  * @device: Device associated with the cm_id.  All related communication will
1089  * be associated with the specified device.
1090  * @cm_handler: Callback invoked to notify the user of CM events.
1091  * @service_id: Service identifier matched against incoming connection
1092  *   and service ID resolution requests.  The service ID should be specified
1093  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1094  *   assign a service ID to the caller.
1095  *
1096  * Callers should call ib_destroy_cm_id when done with the listener ID.
1097  */
1098 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1099                                      ib_cm_handler cm_handler,
1100                                      __be64 service_id)
1101 {
1102         struct cm_id_private *cm_id_priv;
1103         struct ib_cm_id *cm_id;
1104         unsigned long flags;
1105         int err = 0;
1106
1107         /* Create an ID in advance, since the creation may sleep */
1108         cm_id = ib_create_cm_id(device, cm_handler, NULL);
1109         if (IS_ERR(cm_id))
1110                 return cm_id;
1111
1112         spin_lock_irqsave(&cm.lock, flags);
1113
1114         if (service_id == IB_CM_ASSIGN_SERVICE_ID)
1115                 goto new_id;
1116
1117         /* Find an existing ID */
1118         cm_id_priv = cm_find_listen(device, service_id);
1119         if (cm_id_priv) {
1120                 if (cm_id->cm_handler != cm_handler || cm_id->context) {
1121                         /* Sharing an ib_cm_id with different handlers is not
1122                          * supported */
1123                         spin_unlock_irqrestore(&cm.lock, flags);
1124                         return ERR_PTR(-EINVAL);
1125                 }
1126                 atomic_inc(&cm_id_priv->refcount);
1127                 ++cm_id_priv->listen_sharecount;
1128                 spin_unlock_irqrestore(&cm.lock, flags);
1129
1130                 ib_destroy_cm_id(cm_id);
1131                 cm_id = &cm_id_priv->id;
1132                 return cm_id;
1133         }
1134
1135 new_id:
1136         /* Use newly created ID */
1137         err = __ib_cm_listen(cm_id, service_id, 0);
1138
1139         spin_unlock_irqrestore(&cm.lock, flags);
1140
1141         if (err) {
1142                 ib_destroy_cm_id(cm_id);
1143                 return ERR_PTR(err);
1144         }
1145         return cm_id;
1146 }
1147 EXPORT_SYMBOL(ib_cm_insert_listen);
1148
1149 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
1150                           enum cm_msg_sequence msg_seq)
1151 {
1152         u64 hi_tid, low_tid;
1153
1154         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1155         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
1156                           (msg_seq << 30));
1157         return cpu_to_be64(hi_tid | low_tid);
1158 }
1159
1160 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1161                               __be16 attr_id, __be64 tid)
1162 {
1163         hdr->base_version  = IB_MGMT_BASE_VERSION;
1164         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1165         hdr->class_version = IB_CM_CLASS_VERSION;
1166         hdr->method        = IB_MGMT_METHOD_SEND;
1167         hdr->attr_id       = attr_id;
1168         hdr->tid           = tid;
1169 }
1170
1171 static void cm_format_req(struct cm_req_msg *req_msg,
1172                           struct cm_id_private *cm_id_priv,
1173                           struct ib_cm_req_param *param)
1174 {
1175         struct ib_sa_path_rec *pri_path = param->primary_path;
1176         struct ib_sa_path_rec *alt_path = param->alternate_path;
1177
1178         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1179                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1180
1181         req_msg->local_comm_id = cm_id_priv->id.local_id;
1182         req_msg->service_id = param->service_id;
1183         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1184         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1185         cm_req_set_init_depth(req_msg, param->initiator_depth);
1186         cm_req_set_remote_resp_timeout(req_msg,
1187                                        param->remote_cm_response_timeout);
1188         cm_req_set_qp_type(req_msg, param->qp_type);
1189         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1190         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1191         cm_req_set_local_resp_timeout(req_msg,
1192                                       param->local_cm_response_timeout);
1193         req_msg->pkey = param->primary_path->pkey;
1194         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1195         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1196
1197         if (param->qp_type != IB_QPT_XRC_INI) {
1198                 cm_req_set_resp_res(req_msg, param->responder_resources);
1199                 cm_req_set_retry_count(req_msg, param->retry_count);
1200                 cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1201                 cm_req_set_srq(req_msg, param->srq);
1202         }
1203
1204         if (pri_path->hop_limit <= 1) {
1205                 req_msg->primary_local_lid = pri_path->slid;
1206                 req_msg->primary_remote_lid = pri_path->dlid;
1207         } else {
1208                 /* Work-around until there's a way to obtain remote LID info */
1209                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1210                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1211         }
1212         req_msg->primary_local_gid = pri_path->sgid;
1213         req_msg->primary_remote_gid = pri_path->dgid;
1214         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1215         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1216         req_msg->primary_traffic_class = pri_path->traffic_class;
1217         req_msg->primary_hop_limit = pri_path->hop_limit;
1218         cm_req_set_primary_sl(req_msg, pri_path->sl);
1219         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1220         cm_req_set_primary_local_ack_timeout(req_msg,
1221                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1222                                pri_path->packet_life_time));
1223
1224         if (alt_path) {
1225                 if (alt_path->hop_limit <= 1) {
1226                         req_msg->alt_local_lid = alt_path->slid;
1227                         req_msg->alt_remote_lid = alt_path->dlid;
1228                 } else {
1229                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1230                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1231                 }
1232                 req_msg->alt_local_gid = alt_path->sgid;
1233                 req_msg->alt_remote_gid = alt_path->dgid;
1234                 cm_req_set_alt_flow_label(req_msg,
1235                                           alt_path->flow_label);
1236                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1237                 req_msg->alt_traffic_class = alt_path->traffic_class;
1238                 req_msg->alt_hop_limit = alt_path->hop_limit;
1239                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1240                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1241                 cm_req_set_alt_local_ack_timeout(req_msg,
1242                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1243                                        alt_path->packet_life_time));
1244         }
1245
1246         if (param->private_data && param->private_data_len)
1247                 memcpy(req_msg->private_data, param->private_data,
1248                        param->private_data_len);
1249 }
1250
1251 static int cm_validate_req_param(struct ib_cm_req_param *param)
1252 {
1253         /* peer-to-peer not supported */
1254         if (param->peer_to_peer)
1255                 return -EINVAL;
1256
1257         if (!param->primary_path)
1258                 return -EINVAL;
1259
1260         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1261             param->qp_type != IB_QPT_XRC_INI)
1262                 return -EINVAL;
1263
1264         if (param->private_data &&
1265             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1266                 return -EINVAL;
1267
1268         if (param->alternate_path &&
1269             (param->alternate_path->pkey != param->primary_path->pkey ||
1270              param->alternate_path->mtu != param->primary_path->mtu))
1271                 return -EINVAL;
1272
1273         return 0;
1274 }
1275
1276 int ib_send_cm_req(struct ib_cm_id *cm_id,
1277                    struct ib_cm_req_param *param)
1278 {
1279         struct cm_id_private *cm_id_priv;
1280         struct cm_req_msg *req_msg;
1281         unsigned long flags;
1282         int ret;
1283
1284         ret = cm_validate_req_param(param);
1285         if (ret)
1286                 return ret;
1287
1288         /* Verify that we're not in timewait. */
1289         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1290         spin_lock_irqsave(&cm_id_priv->lock, flags);
1291         if (cm_id->state != IB_CM_IDLE) {
1292                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1293                 ret = -EINVAL;
1294                 goto out;
1295         }
1296         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1297
1298         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1299                                                             id.local_id);
1300         if (IS_ERR(cm_id_priv->timewait_info)) {
1301                 ret = PTR_ERR(cm_id_priv->timewait_info);
1302                 goto out;
1303         }
1304
1305         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av,
1306                                  cm_id_priv);
1307         if (ret)
1308                 goto error1;
1309         if (param->alternate_path) {
1310                 ret = cm_init_av_by_path(param->alternate_path,
1311                                          &cm_id_priv->alt_av, cm_id_priv);
1312                 if (ret)
1313                         goto error1;
1314         }
1315         cm_id->service_id = param->service_id;
1316         cm_id->service_mask = ~cpu_to_be64(0);
1317         cm_id_priv->timeout_ms = cm_convert_to_ms(
1318                                     param->primary_path->packet_life_time) * 2 +
1319                                  cm_convert_to_ms(
1320                                     param->remote_cm_response_timeout);
1321         cm_id_priv->max_cm_retries = param->max_cm_retries;
1322         cm_id_priv->initiator_depth = param->initiator_depth;
1323         cm_id_priv->responder_resources = param->responder_resources;
1324         cm_id_priv->retry_count = param->retry_count;
1325         cm_id_priv->path_mtu = param->primary_path->mtu;
1326         cm_id_priv->pkey = param->primary_path->pkey;
1327         cm_id_priv->qp_type = param->qp_type;
1328
1329         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1330         if (ret)
1331                 goto error1;
1332
1333         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1334         cm_format_req(req_msg, cm_id_priv, param);
1335         cm_id_priv->tid = req_msg->hdr.tid;
1336         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1337         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1338
1339         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1340         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1341
1342         spin_lock_irqsave(&cm_id_priv->lock, flags);
1343         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1344         if (ret) {
1345                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1346                 goto error2;
1347         }
1348         BUG_ON(cm_id->state != IB_CM_IDLE);
1349         cm_id->state = IB_CM_REQ_SENT;
1350         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1351         return 0;
1352
1353 error2: cm_free_msg(cm_id_priv->msg);
1354 error1: kfree(cm_id_priv->timewait_info);
1355 out:    return ret;
1356 }
1357 EXPORT_SYMBOL(ib_send_cm_req);
1358
1359 static int cm_issue_rej(struct cm_port *port,
1360                         struct ib_mad_recv_wc *mad_recv_wc,
1361                         enum ib_cm_rej_reason reason,
1362                         enum cm_msg_response msg_rejected,
1363                         void *ari, u8 ari_length)
1364 {
1365         struct ib_mad_send_buf *msg = NULL;
1366         struct cm_rej_msg *rej_msg, *rcv_msg;
1367         int ret;
1368
1369         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1370         if (ret)
1371                 return ret;
1372
1373         /* We just need common CM header information.  Cast to any message. */
1374         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1375         rej_msg = (struct cm_rej_msg *) msg->mad;
1376
1377         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1378         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1379         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1380         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1381         rej_msg->reason = cpu_to_be16(reason);
1382
1383         if (ari && ari_length) {
1384                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1385                 memcpy(rej_msg->ari, ari, ari_length);
1386         }
1387
1388         ret = ib_post_send_mad(msg, NULL);
1389         if (ret)
1390                 cm_free_msg(msg);
1391
1392         return ret;
1393 }
1394
1395 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1396                                     __be32 local_qpn, __be32 remote_qpn)
1397 {
1398         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1399                 ((local_ca_guid == remote_ca_guid) &&
1400                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1401 }
1402
1403 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1404                                             struct ib_sa_path_rec *primary_path,
1405                                             struct ib_sa_path_rec *alt_path)
1406 {
1407         memset(primary_path, 0, sizeof *primary_path);
1408         primary_path->dgid = req_msg->primary_local_gid;
1409         primary_path->sgid = req_msg->primary_remote_gid;
1410         primary_path->dlid = req_msg->primary_local_lid;
1411         primary_path->slid = req_msg->primary_remote_lid;
1412         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1413         primary_path->hop_limit = req_msg->primary_hop_limit;
1414         primary_path->traffic_class = req_msg->primary_traffic_class;
1415         primary_path->reversible = 1;
1416         primary_path->pkey = req_msg->pkey;
1417         primary_path->sl = cm_req_get_primary_sl(req_msg);
1418         primary_path->mtu_selector = IB_SA_EQ;
1419         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1420         primary_path->rate_selector = IB_SA_EQ;
1421         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1422         primary_path->packet_life_time_selector = IB_SA_EQ;
1423         primary_path->packet_life_time =
1424                 cm_req_get_primary_local_ack_timeout(req_msg);
1425         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1426         primary_path->service_id = req_msg->service_id;
1427
1428         if (req_msg->alt_local_lid) {
1429                 memset(alt_path, 0, sizeof *alt_path);
1430                 alt_path->dgid = req_msg->alt_local_gid;
1431                 alt_path->sgid = req_msg->alt_remote_gid;
1432                 alt_path->dlid = req_msg->alt_local_lid;
1433                 alt_path->slid = req_msg->alt_remote_lid;
1434                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1435                 alt_path->hop_limit = req_msg->alt_hop_limit;
1436                 alt_path->traffic_class = req_msg->alt_traffic_class;
1437                 alt_path->reversible = 1;
1438                 alt_path->pkey = req_msg->pkey;
1439                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1440                 alt_path->mtu_selector = IB_SA_EQ;
1441                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1442                 alt_path->rate_selector = IB_SA_EQ;
1443                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1444                 alt_path->packet_life_time_selector = IB_SA_EQ;
1445                 alt_path->packet_life_time =
1446                         cm_req_get_alt_local_ack_timeout(req_msg);
1447                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1448                 alt_path->service_id = req_msg->service_id;
1449         }
1450 }
1451
1452 static u16 cm_get_bth_pkey(struct cm_work *work)
1453 {
1454         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1455         u8 port_num = work->port->port_num;
1456         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1457         u16 pkey;
1458         int ret;
1459
1460         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1461         if (ret) {
1462                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1463                                      port_num, pkey_index, ret);
1464                 return 0;
1465         }
1466
1467         return pkey;
1468 }
1469
1470 static void cm_format_req_event(struct cm_work *work,
1471                                 struct cm_id_private *cm_id_priv,
1472                                 struct ib_cm_id *listen_id)
1473 {
1474         struct cm_req_msg *req_msg;
1475         struct ib_cm_req_event_param *param;
1476
1477         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1478         param = &work->cm_event.param.req_rcvd;
1479         param->listen_id = listen_id;
1480         param->bth_pkey = cm_get_bth_pkey(work);
1481         param->port = cm_id_priv->av.port->port_num;
1482         param->primary_path = &work->path[0];
1483         if (req_msg->alt_local_lid)
1484                 param->alternate_path = &work->path[1];
1485         else
1486                 param->alternate_path = NULL;
1487         param->remote_ca_guid = req_msg->local_ca_guid;
1488         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1489         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1490         param->qp_type = cm_req_get_qp_type(req_msg);
1491         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1492         param->responder_resources = cm_req_get_init_depth(req_msg);
1493         param->initiator_depth = cm_req_get_resp_res(req_msg);
1494         param->local_cm_response_timeout =
1495                                         cm_req_get_remote_resp_timeout(req_msg);
1496         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1497         param->remote_cm_response_timeout =
1498                                         cm_req_get_local_resp_timeout(req_msg);
1499         param->retry_count = cm_req_get_retry_count(req_msg);
1500         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1501         param->srq = cm_req_get_srq(req_msg);
1502         work->cm_event.private_data = &req_msg->private_data;
1503 }
1504
1505 static void cm_process_work(struct cm_id_private *cm_id_priv,
1506                             struct cm_work *work)
1507 {
1508         int ret;
1509
1510         /* We will typically only have the current event to report. */
1511         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1512         cm_free_work(work);
1513
1514         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1515                 spin_lock_irq(&cm_id_priv->lock);
1516                 work = cm_dequeue_work(cm_id_priv);
1517                 spin_unlock_irq(&cm_id_priv->lock);
1518                 BUG_ON(!work);
1519                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1520                                                 &work->cm_event);
1521                 cm_free_work(work);
1522         }
1523         cm_deref_id(cm_id_priv);
1524         if (ret)
1525                 cm_destroy_id(&cm_id_priv->id, ret);
1526 }
1527
1528 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1529                           struct cm_id_private *cm_id_priv,
1530                           enum cm_msg_response msg_mraed, u8 service_timeout,
1531                           const void *private_data, u8 private_data_len)
1532 {
1533         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1534         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1535         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1536         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1537         cm_mra_set_service_timeout(mra_msg, service_timeout);
1538
1539         if (private_data && private_data_len)
1540                 memcpy(mra_msg->private_data, private_data, private_data_len);
1541 }
1542
1543 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1544                           struct cm_id_private *cm_id_priv,
1545                           enum ib_cm_rej_reason reason,
1546                           void *ari,
1547                           u8 ari_length,
1548                           const void *private_data,
1549                           u8 private_data_len)
1550 {
1551         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1552         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1553
1554         switch(cm_id_priv->id.state) {
1555         case IB_CM_REQ_RCVD:
1556                 rej_msg->local_comm_id = 0;
1557                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1558                 break;
1559         case IB_CM_MRA_REQ_SENT:
1560                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1561                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1562                 break;
1563         case IB_CM_REP_RCVD:
1564         case IB_CM_MRA_REP_SENT:
1565                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1566                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1567                 break;
1568         default:
1569                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1570                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1571                 break;
1572         }
1573
1574         rej_msg->reason = cpu_to_be16(reason);
1575         if (ari && ari_length) {
1576                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1577                 memcpy(rej_msg->ari, ari, ari_length);
1578         }
1579
1580         if (private_data && private_data_len)
1581                 memcpy(rej_msg->private_data, private_data, private_data_len);
1582 }
1583
1584 static void cm_dup_req_handler(struct cm_work *work,
1585                                struct cm_id_private *cm_id_priv)
1586 {
1587         struct ib_mad_send_buf *msg = NULL;
1588         int ret;
1589
1590         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1591                         counter[CM_REQ_COUNTER]);
1592
1593         /* Quick state check to discard duplicate REQs. */
1594         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1595                 return;
1596
1597         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1598         if (ret)
1599                 return;
1600
1601         spin_lock_irq(&cm_id_priv->lock);
1602         switch (cm_id_priv->id.state) {
1603         case IB_CM_MRA_REQ_SENT:
1604                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1605                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1606                               cm_id_priv->private_data,
1607                               cm_id_priv->private_data_len);
1608                 break;
1609         case IB_CM_TIMEWAIT:
1610                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1611                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1612                 break;
1613         default:
1614                 goto unlock;
1615         }
1616         spin_unlock_irq(&cm_id_priv->lock);
1617
1618         ret = ib_post_send_mad(msg, NULL);
1619         if (ret)
1620                 goto free;
1621         return;
1622
1623 unlock: spin_unlock_irq(&cm_id_priv->lock);
1624 free:   cm_free_msg(msg);
1625 }
1626
1627 static struct cm_id_private * cm_match_req(struct cm_work *work,
1628                                            struct cm_id_private *cm_id_priv)
1629 {
1630         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1631         struct cm_timewait_info *timewait_info;
1632         struct cm_req_msg *req_msg;
1633         struct ib_cm_id *cm_id;
1634
1635         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1636
1637         /* Check for possible duplicate REQ. */
1638         spin_lock_irq(&cm.lock);
1639         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1640         if (timewait_info) {
1641                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1642                                            timewait_info->work.remote_id);
1643                 spin_unlock_irq(&cm.lock);
1644                 if (cur_cm_id_priv) {
1645                         cm_dup_req_handler(work, cur_cm_id_priv);
1646                         cm_deref_id(cur_cm_id_priv);
1647                 }
1648                 return NULL;
1649         }
1650
1651         /* Check for stale connections. */
1652         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1653         if (timewait_info) {
1654                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1655                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1656                                            timewait_info->work.remote_id);
1657
1658                 spin_unlock_irq(&cm.lock);
1659                 cm_issue_rej(work->port, work->mad_recv_wc,
1660                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1661                              NULL, 0);
1662                 if (cur_cm_id_priv) {
1663                         cm_id = &cur_cm_id_priv->id;
1664                         ib_send_cm_dreq(cm_id, NULL, 0);
1665                         cm_deref_id(cur_cm_id_priv);
1666                 }
1667                 return NULL;
1668         }
1669
1670         /* Find matching listen request. */
1671         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1672                                            req_msg->service_id);
1673         if (!listen_cm_id_priv) {
1674                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1675                 spin_unlock_irq(&cm.lock);
1676                 cm_issue_rej(work->port, work->mad_recv_wc,
1677                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1678                              NULL, 0);
1679                 goto out;
1680         }
1681         atomic_inc(&listen_cm_id_priv->refcount);
1682         atomic_inc(&cm_id_priv->refcount);
1683         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1684         atomic_inc(&cm_id_priv->work_count);
1685         spin_unlock_irq(&cm.lock);
1686 out:
1687         return listen_cm_id_priv;
1688 }
1689
1690 /*
1691  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1692  * we need to override the LID/SL data in the REQ with the LID information
1693  * in the work completion.
1694  */
1695 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1696 {
1697         if (!cm_req_get_primary_subnet_local(req_msg)) {
1698                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1699                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1700                         cm_req_set_primary_sl(req_msg, wc->sl);
1701                 }
1702
1703                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1704                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1705         }
1706
1707         if (!cm_req_get_alt_subnet_local(req_msg)) {
1708                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1709                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1710                         cm_req_set_alt_sl(req_msg, wc->sl);
1711                 }
1712
1713                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1714                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1715         }
1716 }
1717
1718 static int cm_req_handler(struct cm_work *work)
1719 {
1720         struct ib_cm_id *cm_id;
1721         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1722         struct cm_req_msg *req_msg;
1723         union ib_gid gid;
1724         struct ib_gid_attr gid_attr;
1725         const struct ib_global_route *grh;
1726         int ret;
1727
1728         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1729
1730         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1731         if (IS_ERR(cm_id))
1732                 return PTR_ERR(cm_id);
1733
1734         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1735         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1736         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1737                                 work->mad_recv_wc->recv_buf.grh,
1738                                 &cm_id_priv->av);
1739         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1740                                                             id.local_id);
1741         if (IS_ERR(cm_id_priv->timewait_info)) {
1742                 ret = PTR_ERR(cm_id_priv->timewait_info);
1743                 goto destroy;
1744         }
1745         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1746         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1747         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1748
1749         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1750         if (!listen_cm_id_priv) {
1751                 ret = -EINVAL;
1752                 kfree(cm_id_priv->timewait_info);
1753                 goto destroy;
1754         }
1755
1756         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1757         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1758         cm_id_priv->id.service_id = req_msg->service_id;
1759         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1760
1761         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1762         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1763
1764         memcpy(work->path[0].dmac, cm_id_priv->av.ah_attr.dmac, ETH_ALEN);
1765         grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
1766         work->path[0].hop_limit = grh->hop_limit;
1767         ret = ib_get_cached_gid(work->port->cm_dev->ib_device,
1768                                 work->port->port_num,
1769                                 grh->sgid_index,
1770                                 &gid, &gid_attr);
1771         if (!ret) {
1772                 if (gid_attr.ndev) {
1773                         work->path[0].ifindex = gid_attr.ndev->ifindex;
1774                         work->path[0].net = dev_net(gid_attr.ndev);
1775                         dev_put(gid_attr.ndev);
1776                 }
1777                 work->path[0].gid_type = gid_attr.gid_type;
1778                 ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av,
1779                                          cm_id_priv);
1780         }
1781         if (ret) {
1782                 int err = ib_get_cached_gid(work->port->cm_dev->ib_device,
1783                                             work->port->port_num, 0,
1784                                             &work->path[0].sgid,
1785                                             &gid_attr);
1786                 if (!err && gid_attr.ndev) {
1787                         work->path[0].ifindex = gid_attr.ndev->ifindex;
1788                         work->path[0].net = dev_net(gid_attr.ndev);
1789                         dev_put(gid_attr.ndev);
1790                 }
1791                 work->path[0].gid_type = gid_attr.gid_type;
1792                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1793                                &work->path[0].sgid, sizeof work->path[0].sgid,
1794                                NULL, 0);
1795                 goto rejected;
1796         }
1797         if (req_msg->alt_local_lid) {
1798                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av,
1799                                          cm_id_priv);
1800                 if (ret) {
1801                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1802                                        &work->path[0].sgid,
1803                                        sizeof work->path[0].sgid, NULL, 0);
1804                         goto rejected;
1805                 }
1806         }
1807         cm_id_priv->tid = req_msg->hdr.tid;
1808         cm_id_priv->timeout_ms = cm_convert_to_ms(
1809                                         cm_req_get_local_resp_timeout(req_msg));
1810         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1811         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1812         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1813         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1814         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1815         cm_id_priv->pkey = req_msg->pkey;
1816         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1817         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1818         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1819         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1820
1821         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1822         cm_process_work(cm_id_priv, work);
1823         cm_deref_id(listen_cm_id_priv);
1824         return 0;
1825
1826 rejected:
1827         atomic_dec(&cm_id_priv->refcount);
1828         cm_deref_id(listen_cm_id_priv);
1829 destroy:
1830         ib_destroy_cm_id(cm_id);
1831         return ret;
1832 }
1833
1834 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1835                           struct cm_id_private *cm_id_priv,
1836                           struct ib_cm_rep_param *param)
1837 {
1838         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1839         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1840         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1841         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1842         rep_msg->resp_resources = param->responder_resources;
1843         cm_rep_set_target_ack_delay(rep_msg,
1844                                     cm_id_priv->av.port->cm_dev->ack_delay);
1845         cm_rep_set_failover(rep_msg, param->failover_accepted);
1846         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1847         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1848
1849         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1850                 rep_msg->initiator_depth = param->initiator_depth;
1851                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1852                 cm_rep_set_srq(rep_msg, param->srq);
1853                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1854         } else {
1855                 cm_rep_set_srq(rep_msg, 1);
1856                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1857         }
1858
1859         if (param->private_data && param->private_data_len)
1860                 memcpy(rep_msg->private_data, param->private_data,
1861                        param->private_data_len);
1862 }
1863
1864 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1865                    struct ib_cm_rep_param *param)
1866 {
1867         struct cm_id_private *cm_id_priv;
1868         struct ib_mad_send_buf *msg;
1869         struct cm_rep_msg *rep_msg;
1870         unsigned long flags;
1871         int ret;
1872
1873         if (param->private_data &&
1874             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1875                 return -EINVAL;
1876
1877         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1878         spin_lock_irqsave(&cm_id_priv->lock, flags);
1879         if (cm_id->state != IB_CM_REQ_RCVD &&
1880             cm_id->state != IB_CM_MRA_REQ_SENT) {
1881                 ret = -EINVAL;
1882                 goto out;
1883         }
1884
1885         ret = cm_alloc_msg(cm_id_priv, &msg);
1886         if (ret)
1887                 goto out;
1888
1889         rep_msg = (struct cm_rep_msg *) msg->mad;
1890         cm_format_rep(rep_msg, cm_id_priv, param);
1891         msg->timeout_ms = cm_id_priv->timeout_ms;
1892         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1893
1894         ret = ib_post_send_mad(msg, NULL);
1895         if (ret) {
1896                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1897                 cm_free_msg(msg);
1898                 return ret;
1899         }
1900
1901         cm_id->state = IB_CM_REP_SENT;
1902         cm_id_priv->msg = msg;
1903         cm_id_priv->initiator_depth = param->initiator_depth;
1904         cm_id_priv->responder_resources = param->responder_resources;
1905         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1906         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
1907
1908 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1909         return ret;
1910 }
1911 EXPORT_SYMBOL(ib_send_cm_rep);
1912
1913 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1914                           struct cm_id_private *cm_id_priv,
1915                           const void *private_data,
1916                           u8 private_data_len)
1917 {
1918         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1919         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1920         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1921
1922         if (private_data && private_data_len)
1923                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1924 }
1925
1926 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1927                    const void *private_data,
1928                    u8 private_data_len)
1929 {
1930         struct cm_id_private *cm_id_priv;
1931         struct ib_mad_send_buf *msg;
1932         unsigned long flags;
1933         void *data;
1934         int ret;
1935
1936         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1937                 return -EINVAL;
1938
1939         data = cm_copy_private_data(private_data, private_data_len);
1940         if (IS_ERR(data))
1941                 return PTR_ERR(data);
1942
1943         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1944         spin_lock_irqsave(&cm_id_priv->lock, flags);
1945         if (cm_id->state != IB_CM_REP_RCVD &&
1946             cm_id->state != IB_CM_MRA_REP_SENT) {
1947                 ret = -EINVAL;
1948                 goto error;
1949         }
1950
1951         ret = cm_alloc_msg(cm_id_priv, &msg);
1952         if (ret)
1953                 goto error;
1954
1955         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1956                       private_data, private_data_len);
1957
1958         ret = ib_post_send_mad(msg, NULL);
1959         if (ret) {
1960                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1961                 cm_free_msg(msg);
1962                 kfree(data);
1963                 return ret;
1964         }
1965
1966         cm_id->state = IB_CM_ESTABLISHED;
1967         cm_set_private_data(cm_id_priv, data, private_data_len);
1968         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1969         return 0;
1970
1971 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1972         kfree(data);
1973         return ret;
1974 }
1975 EXPORT_SYMBOL(ib_send_cm_rtu);
1976
1977 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
1978 {
1979         struct cm_rep_msg *rep_msg;
1980         struct ib_cm_rep_event_param *param;
1981
1982         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1983         param = &work->cm_event.param.rep_rcvd;
1984         param->remote_ca_guid = rep_msg->local_ca_guid;
1985         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1986         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
1987         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1988         param->responder_resources = rep_msg->initiator_depth;
1989         param->initiator_depth = rep_msg->resp_resources;
1990         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1991         param->failover_accepted = cm_rep_get_failover(rep_msg);
1992         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1993         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1994         param->srq = cm_rep_get_srq(rep_msg);
1995         work->cm_event.private_data = &rep_msg->private_data;
1996 }
1997
1998 static void cm_dup_rep_handler(struct cm_work *work)
1999 {
2000         struct cm_id_private *cm_id_priv;
2001         struct cm_rep_msg *rep_msg;
2002         struct ib_mad_send_buf *msg = NULL;
2003         int ret;
2004
2005         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2006         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
2007                                    rep_msg->local_comm_id);
2008         if (!cm_id_priv)
2009                 return;
2010
2011         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2012                         counter[CM_REP_COUNTER]);
2013         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2014         if (ret)
2015                 goto deref;
2016
2017         spin_lock_irq(&cm_id_priv->lock);
2018         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2019                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2020                               cm_id_priv->private_data,
2021                               cm_id_priv->private_data_len);
2022         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2023                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2024                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2025                               cm_id_priv->private_data,
2026                               cm_id_priv->private_data_len);
2027         else
2028                 goto unlock;
2029         spin_unlock_irq(&cm_id_priv->lock);
2030
2031         ret = ib_post_send_mad(msg, NULL);
2032         if (ret)
2033                 goto free;
2034         goto deref;
2035
2036 unlock: spin_unlock_irq(&cm_id_priv->lock);
2037 free:   cm_free_msg(msg);
2038 deref:  cm_deref_id(cm_id_priv);
2039 }
2040
2041 static int cm_rep_handler(struct cm_work *work)
2042 {
2043         struct cm_id_private *cm_id_priv;
2044         struct cm_rep_msg *rep_msg;
2045         int ret;
2046         struct cm_id_private *cur_cm_id_priv;
2047         struct ib_cm_id *cm_id;
2048         struct cm_timewait_info *timewait_info;
2049
2050         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2051         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
2052         if (!cm_id_priv) {
2053                 cm_dup_rep_handler(work);
2054                 return -EINVAL;
2055         }
2056
2057         cm_format_rep_event(work, cm_id_priv->qp_type);
2058
2059         spin_lock_irq(&cm_id_priv->lock);
2060         switch (cm_id_priv->id.state) {
2061         case IB_CM_REQ_SENT:
2062         case IB_CM_MRA_REQ_RCVD:
2063                 break;
2064         default:
2065                 spin_unlock_irq(&cm_id_priv->lock);
2066                 ret = -EINVAL;
2067                 goto error;
2068         }
2069
2070         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
2071         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
2072         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2073
2074         spin_lock(&cm.lock);
2075         /* Check for duplicate REP. */
2076         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2077                 spin_unlock(&cm.lock);
2078                 spin_unlock_irq(&cm_id_priv->lock);
2079                 ret = -EINVAL;
2080                 goto error;
2081         }
2082         /* Check for a stale connection. */
2083         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2084         if (timewait_info) {
2085                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
2086                          &cm.remote_id_table);
2087                 cm_id_priv->timewait_info->inserted_remote_id = 0;
2088                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
2089                                            timewait_info->work.remote_id);
2090
2091                 spin_unlock(&cm.lock);
2092                 spin_unlock_irq(&cm_id_priv->lock);
2093                 cm_issue_rej(work->port, work->mad_recv_wc,
2094                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2095                              NULL, 0);
2096                 ret = -EINVAL;
2097                 if (cur_cm_id_priv) {
2098                         cm_id = &cur_cm_id_priv->id;
2099                         ib_send_cm_dreq(cm_id, NULL, 0);
2100                         cm_deref_id(cur_cm_id_priv);
2101                 }
2102
2103                 goto error;
2104         }
2105         spin_unlock(&cm.lock);
2106
2107         cm_id_priv->id.state = IB_CM_REP_RCVD;
2108         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
2109         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2110         cm_id_priv->initiator_depth = rep_msg->resp_resources;
2111         cm_id_priv->responder_resources = rep_msg->initiator_depth;
2112         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
2113         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2114         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2115         cm_id_priv->av.timeout =
2116                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2117                                        cm_id_priv->av.timeout - 1);
2118         cm_id_priv->alt_av.timeout =
2119                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2120                                        cm_id_priv->alt_av.timeout - 1);
2121
2122         /* todo: handle peer_to_peer */
2123
2124         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2125         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2126         if (!ret)
2127                 list_add_tail(&work->list, &cm_id_priv->work_list);
2128         spin_unlock_irq(&cm_id_priv->lock);
2129
2130         if (ret)
2131                 cm_process_work(cm_id_priv, work);
2132         else
2133                 cm_deref_id(cm_id_priv);
2134         return 0;
2135
2136 error:
2137         cm_deref_id(cm_id_priv);
2138         return ret;
2139 }
2140
2141 static int cm_establish_handler(struct cm_work *work)
2142 {
2143         struct cm_id_private *cm_id_priv;
2144         int ret;
2145
2146         /* See comment in cm_establish about lookup. */
2147         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2148         if (!cm_id_priv)
2149                 return -EINVAL;
2150
2151         spin_lock_irq(&cm_id_priv->lock);
2152         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2153                 spin_unlock_irq(&cm_id_priv->lock);
2154                 goto out;
2155         }
2156
2157         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2158         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2159         if (!ret)
2160                 list_add_tail(&work->list, &cm_id_priv->work_list);
2161         spin_unlock_irq(&cm_id_priv->lock);
2162
2163         if (ret)
2164                 cm_process_work(cm_id_priv, work);
2165         else
2166                 cm_deref_id(cm_id_priv);
2167         return 0;
2168 out:
2169         cm_deref_id(cm_id_priv);
2170         return -EINVAL;
2171 }
2172
2173 static int cm_rtu_handler(struct cm_work *work)
2174 {
2175         struct cm_id_private *cm_id_priv;
2176         struct cm_rtu_msg *rtu_msg;
2177         int ret;
2178
2179         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2180         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
2181                                    rtu_msg->local_comm_id);
2182         if (!cm_id_priv)
2183                 return -EINVAL;
2184
2185         work->cm_event.private_data = &rtu_msg->private_data;
2186
2187         spin_lock_irq(&cm_id_priv->lock);
2188         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2189             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2190                 spin_unlock_irq(&cm_id_priv->lock);
2191                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2192                                 counter[CM_RTU_COUNTER]);
2193                 goto out;
2194         }
2195         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2196
2197         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2198         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2199         if (!ret)
2200                 list_add_tail(&work->list, &cm_id_priv->work_list);
2201         spin_unlock_irq(&cm_id_priv->lock);
2202
2203         if (ret)
2204                 cm_process_work(cm_id_priv, work);
2205         else
2206                 cm_deref_id(cm_id_priv);
2207         return 0;
2208 out:
2209         cm_deref_id(cm_id_priv);
2210         return -EINVAL;
2211 }
2212
2213 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2214                           struct cm_id_private *cm_id_priv,
2215                           const void *private_data,
2216                           u8 private_data_len)
2217 {
2218         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2219                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
2220         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
2221         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
2222         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2223
2224         if (private_data && private_data_len)
2225                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2226 }
2227
2228 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2229                     const void *private_data,
2230                     u8 private_data_len)
2231 {
2232         struct cm_id_private *cm_id_priv;
2233         struct ib_mad_send_buf *msg;
2234         unsigned long flags;
2235         int ret;
2236
2237         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2238                 return -EINVAL;
2239
2240         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2241         spin_lock_irqsave(&cm_id_priv->lock, flags);
2242         if (cm_id->state != IB_CM_ESTABLISHED) {
2243                 ret = -EINVAL;
2244                 goto out;
2245         }
2246
2247         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2248             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2249                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2250
2251         ret = cm_alloc_msg(cm_id_priv, &msg);
2252         if (ret) {
2253                 cm_enter_timewait(cm_id_priv);
2254                 goto out;
2255         }
2256
2257         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2258                        private_data, private_data_len);
2259         msg->timeout_ms = cm_id_priv->timeout_ms;
2260         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2261
2262         ret = ib_post_send_mad(msg, NULL);
2263         if (ret) {
2264                 cm_enter_timewait(cm_id_priv);
2265                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2266                 cm_free_msg(msg);
2267                 return ret;
2268         }
2269
2270         cm_id->state = IB_CM_DREQ_SENT;
2271         cm_id_priv->msg = msg;
2272 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2273         return ret;
2274 }
2275 EXPORT_SYMBOL(ib_send_cm_dreq);
2276
2277 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2278                           struct cm_id_private *cm_id_priv,
2279                           const void *private_data,
2280                           u8 private_data_len)
2281 {
2282         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2283         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2284         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2285
2286         if (private_data && private_data_len)
2287                 memcpy(drep_msg->private_data, private_data, private_data_len);
2288 }
2289
2290 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2291                     const void *private_data,
2292                     u8 private_data_len)
2293 {
2294         struct cm_id_private *cm_id_priv;
2295         struct ib_mad_send_buf *msg;
2296         unsigned long flags;
2297         void *data;
2298         int ret;
2299
2300         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2301                 return -EINVAL;
2302
2303         data = cm_copy_private_data(private_data, private_data_len);
2304         if (IS_ERR(data))
2305                 return PTR_ERR(data);
2306
2307         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2308         spin_lock_irqsave(&cm_id_priv->lock, flags);
2309         if (cm_id->state != IB_CM_DREQ_RCVD) {
2310                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2311                 kfree(data);
2312                 return -EINVAL;
2313         }
2314
2315         cm_set_private_data(cm_id_priv, data, private_data_len);
2316         cm_enter_timewait(cm_id_priv);
2317
2318         ret = cm_alloc_msg(cm_id_priv, &msg);
2319         if (ret)
2320                 goto out;
2321
2322         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2323                        private_data, private_data_len);
2324
2325         ret = ib_post_send_mad(msg, NULL);
2326         if (ret) {
2327                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2328                 cm_free_msg(msg);
2329                 return ret;
2330         }
2331
2332 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2333         return ret;
2334 }
2335 EXPORT_SYMBOL(ib_send_cm_drep);
2336
2337 static int cm_issue_drep(struct cm_port *port,
2338                          struct ib_mad_recv_wc *mad_recv_wc)
2339 {
2340         struct ib_mad_send_buf *msg = NULL;
2341         struct cm_dreq_msg *dreq_msg;
2342         struct cm_drep_msg *drep_msg;
2343         int ret;
2344
2345         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2346         if (ret)
2347                 return ret;
2348
2349         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2350         drep_msg = (struct cm_drep_msg *) msg->mad;
2351
2352         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2353         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2354         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2355
2356         ret = ib_post_send_mad(msg, NULL);
2357         if (ret)
2358                 cm_free_msg(msg);
2359
2360         return ret;
2361 }
2362
2363 static int cm_dreq_handler(struct cm_work *work)
2364 {
2365         struct cm_id_private *cm_id_priv;
2366         struct cm_dreq_msg *dreq_msg;
2367         struct ib_mad_send_buf *msg = NULL;
2368         int ret;
2369
2370         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2371         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2372                                    dreq_msg->local_comm_id);
2373         if (!cm_id_priv) {
2374                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2375                                 counter[CM_DREQ_COUNTER]);
2376                 cm_issue_drep(work->port, work->mad_recv_wc);
2377                 return -EINVAL;
2378         }
2379
2380         work->cm_event.private_data = &dreq_msg->private_data;
2381
2382         spin_lock_irq(&cm_id_priv->lock);
2383         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2384                 goto unlock;
2385
2386         switch (cm_id_priv->id.state) {
2387         case IB_CM_REP_SENT:
2388         case IB_CM_DREQ_SENT:
2389                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2390                 break;
2391         case IB_CM_ESTABLISHED:
2392                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2393                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2394                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2395                 break;
2396         case IB_CM_MRA_REP_RCVD:
2397                 break;
2398         case IB_CM_TIMEWAIT:
2399                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2400                                 counter[CM_DREQ_COUNTER]);
2401                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2402                         goto unlock;
2403
2404                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2405                                cm_id_priv->private_data,
2406                                cm_id_priv->private_data_len);
2407                 spin_unlock_irq(&cm_id_priv->lock);
2408
2409                 if (ib_post_send_mad(msg, NULL))
2410                         cm_free_msg(msg);
2411                 goto deref;
2412         case IB_CM_DREQ_RCVD:
2413                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2414                                 counter[CM_DREQ_COUNTER]);
2415                 goto unlock;
2416         default:
2417                 goto unlock;
2418         }
2419         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2420         cm_id_priv->tid = dreq_msg->hdr.tid;
2421         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2422         if (!ret)
2423                 list_add_tail(&work->list, &cm_id_priv->work_list);
2424         spin_unlock_irq(&cm_id_priv->lock);
2425
2426         if (ret)
2427                 cm_process_work(cm_id_priv, work);
2428         else
2429                 cm_deref_id(cm_id_priv);
2430         return 0;
2431
2432 unlock: spin_unlock_irq(&cm_id_priv->lock);
2433 deref:  cm_deref_id(cm_id_priv);
2434         return -EINVAL;
2435 }
2436
2437 static int cm_drep_handler(struct cm_work *work)
2438 {
2439         struct cm_id_private *cm_id_priv;
2440         struct cm_drep_msg *drep_msg;
2441         int ret;
2442
2443         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2444         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2445                                    drep_msg->local_comm_id);
2446         if (!cm_id_priv)
2447                 return -EINVAL;
2448
2449         work->cm_event.private_data = &drep_msg->private_data;
2450
2451         spin_lock_irq(&cm_id_priv->lock);
2452         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2453             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2454                 spin_unlock_irq(&cm_id_priv->lock);
2455                 goto out;
2456         }
2457         cm_enter_timewait(cm_id_priv);
2458
2459         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2460         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2461         if (!ret)
2462                 list_add_tail(&work->list, &cm_id_priv->work_list);
2463         spin_unlock_irq(&cm_id_priv->lock);
2464
2465         if (ret)
2466                 cm_process_work(cm_id_priv, work);
2467         else
2468                 cm_deref_id(cm_id_priv);
2469         return 0;
2470 out:
2471         cm_deref_id(cm_id_priv);
2472         return -EINVAL;
2473 }
2474
2475 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2476                    enum ib_cm_rej_reason reason,
2477                    void *ari,
2478                    u8 ari_length,
2479                    const void *private_data,
2480                    u8 private_data_len)
2481 {
2482         struct cm_id_private *cm_id_priv;
2483         struct ib_mad_send_buf *msg;
2484         unsigned long flags;
2485         int ret;
2486
2487         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2488             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2489                 return -EINVAL;
2490
2491         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2492
2493         spin_lock_irqsave(&cm_id_priv->lock, flags);
2494         switch (cm_id->state) {
2495         case IB_CM_REQ_SENT:
2496         case IB_CM_MRA_REQ_RCVD:
2497         case IB_CM_REQ_RCVD:
2498         case IB_CM_MRA_REQ_SENT:
2499         case IB_CM_REP_RCVD:
2500         case IB_CM_MRA_REP_SENT:
2501                 ret = cm_alloc_msg(cm_id_priv, &msg);
2502                 if (!ret)
2503                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2504                                       cm_id_priv, reason, ari, ari_length,
2505                                       private_data, private_data_len);
2506
2507                 cm_reset_to_idle(cm_id_priv);
2508                 break;
2509         case IB_CM_REP_SENT:
2510         case IB_CM_MRA_REP_RCVD:
2511                 ret = cm_alloc_msg(cm_id_priv, &msg);
2512                 if (!ret)
2513                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2514                                       cm_id_priv, reason, ari, ari_length,
2515                                       private_data, private_data_len);
2516
2517                 cm_enter_timewait(cm_id_priv);
2518                 break;
2519         default:
2520                 ret = -EINVAL;
2521                 goto out;
2522         }
2523
2524         if (ret)
2525                 goto out;
2526
2527         ret = ib_post_send_mad(msg, NULL);
2528         if (ret)
2529                 cm_free_msg(msg);
2530
2531 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2532         return ret;
2533 }
2534 EXPORT_SYMBOL(ib_send_cm_rej);
2535
2536 static void cm_format_rej_event(struct cm_work *work)
2537 {
2538         struct cm_rej_msg *rej_msg;
2539         struct ib_cm_rej_event_param *param;
2540
2541         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2542         param = &work->cm_event.param.rej_rcvd;
2543         param->ari = rej_msg->ari;
2544         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2545         param->reason = __be16_to_cpu(rej_msg->reason);
2546         work->cm_event.private_data = &rej_msg->private_data;
2547 }
2548
2549 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2550 {
2551         struct cm_timewait_info *timewait_info;
2552         struct cm_id_private *cm_id_priv;
2553         __be32 remote_id;
2554
2555         remote_id = rej_msg->local_comm_id;
2556
2557         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2558                 spin_lock_irq(&cm.lock);
2559                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2560                                                   remote_id);
2561                 if (!timewait_info) {
2562                         spin_unlock_irq(&cm.lock);
2563                         return NULL;
2564                 }
2565                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2566                                       (timewait_info->work.local_id ^
2567                                        cm.random_id_operand));
2568                 if (cm_id_priv) {
2569                         if (cm_id_priv->id.remote_id == remote_id)
2570                                 atomic_inc(&cm_id_priv->refcount);
2571                         else
2572                                 cm_id_priv = NULL;
2573                 }
2574                 spin_unlock_irq(&cm.lock);
2575         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2576                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2577         else
2578                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2579
2580         return cm_id_priv;
2581 }
2582
2583 static int cm_rej_handler(struct cm_work *work)
2584 {
2585         struct cm_id_private *cm_id_priv;
2586         struct cm_rej_msg *rej_msg;
2587         int ret;
2588
2589         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2590         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2591         if (!cm_id_priv)
2592                 return -EINVAL;
2593
2594         cm_format_rej_event(work);
2595
2596         spin_lock_irq(&cm_id_priv->lock);
2597         switch (cm_id_priv->id.state) {
2598         case IB_CM_REQ_SENT:
2599         case IB_CM_MRA_REQ_RCVD:
2600         case IB_CM_REP_SENT:
2601         case IB_CM_MRA_REP_RCVD:
2602                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2603                 /* fall through */
2604         case IB_CM_REQ_RCVD:
2605         case IB_CM_MRA_REQ_SENT:
2606                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2607                         cm_enter_timewait(cm_id_priv);
2608                 else
2609                         cm_reset_to_idle(cm_id_priv);
2610                 break;
2611         case IB_CM_DREQ_SENT:
2612                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2613                 /* fall through */
2614         case IB_CM_REP_RCVD:
2615         case IB_CM_MRA_REP_SENT:
2616                 cm_enter_timewait(cm_id_priv);
2617                 break;
2618         case IB_CM_ESTABLISHED:
2619                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2620                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2621                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2622                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2623                                               cm_id_priv->msg);
2624                         cm_enter_timewait(cm_id_priv);
2625                         break;
2626                 }
2627                 /* fall through */
2628         default:
2629                 spin_unlock_irq(&cm_id_priv->lock);
2630                 ret = -EINVAL;
2631                 goto out;
2632         }
2633
2634         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2635         if (!ret)
2636                 list_add_tail(&work->list, &cm_id_priv->work_list);
2637         spin_unlock_irq(&cm_id_priv->lock);
2638
2639         if (ret)
2640                 cm_process_work(cm_id_priv, work);
2641         else
2642                 cm_deref_id(cm_id_priv);
2643         return 0;
2644 out:
2645         cm_deref_id(cm_id_priv);
2646         return -EINVAL;
2647 }
2648
2649 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2650                    u8 service_timeout,
2651                    const void *private_data,
2652                    u8 private_data_len)
2653 {
2654         struct cm_id_private *cm_id_priv;
2655         struct ib_mad_send_buf *msg;
2656         enum ib_cm_state cm_state;
2657         enum ib_cm_lap_state lap_state;
2658         enum cm_msg_response msg_response;
2659         void *data;
2660         unsigned long flags;
2661         int ret;
2662
2663         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2664                 return -EINVAL;
2665
2666         data = cm_copy_private_data(private_data, private_data_len);
2667         if (IS_ERR(data))
2668                 return PTR_ERR(data);
2669
2670         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2671
2672         spin_lock_irqsave(&cm_id_priv->lock, flags);
2673         switch(cm_id_priv->id.state) {
2674         case IB_CM_REQ_RCVD:
2675                 cm_state = IB_CM_MRA_REQ_SENT;
2676                 lap_state = cm_id->lap_state;
2677                 msg_response = CM_MSG_RESPONSE_REQ;
2678                 break;
2679         case IB_CM_REP_RCVD:
2680                 cm_state = IB_CM_MRA_REP_SENT;
2681                 lap_state = cm_id->lap_state;
2682                 msg_response = CM_MSG_RESPONSE_REP;
2683                 break;
2684         case IB_CM_ESTABLISHED:
2685                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2686                         cm_state = cm_id->state;
2687                         lap_state = IB_CM_MRA_LAP_SENT;
2688                         msg_response = CM_MSG_RESPONSE_OTHER;
2689                         break;
2690                 }
2691         default:
2692                 ret = -EINVAL;
2693                 goto error1;
2694         }
2695
2696         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2697                 ret = cm_alloc_msg(cm_id_priv, &msg);
2698                 if (ret)
2699                         goto error1;
2700
2701                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2702                               msg_response, service_timeout,
2703                               private_data, private_data_len);
2704                 ret = ib_post_send_mad(msg, NULL);
2705                 if (ret)
2706                         goto error2;
2707         }
2708
2709         cm_id->state = cm_state;
2710         cm_id->lap_state = lap_state;
2711         cm_id_priv->service_timeout = service_timeout;
2712         cm_set_private_data(cm_id_priv, data, private_data_len);
2713         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2714         return 0;
2715
2716 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2717         kfree(data);
2718         return ret;
2719
2720 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2721         kfree(data);
2722         cm_free_msg(msg);
2723         return ret;
2724 }
2725 EXPORT_SYMBOL(ib_send_cm_mra);
2726
2727 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2728 {
2729         switch (cm_mra_get_msg_mraed(mra_msg)) {
2730         case CM_MSG_RESPONSE_REQ:
2731                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2732         case CM_MSG_RESPONSE_REP:
2733         case CM_MSG_RESPONSE_OTHER:
2734                 return cm_acquire_id(mra_msg->remote_comm_id,
2735                                      mra_msg->local_comm_id);
2736         default:
2737                 return NULL;
2738         }
2739 }
2740
2741 static int cm_mra_handler(struct cm_work *work)
2742 {
2743         struct cm_id_private *cm_id_priv;
2744         struct cm_mra_msg *mra_msg;
2745         int timeout, ret;
2746
2747         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2748         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2749         if (!cm_id_priv)
2750                 return -EINVAL;
2751
2752         work->cm_event.private_data = &mra_msg->private_data;
2753         work->cm_event.param.mra_rcvd.service_timeout =
2754                                         cm_mra_get_service_timeout(mra_msg);
2755         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2756                   cm_convert_to_ms(cm_id_priv->av.timeout);
2757
2758         spin_lock_irq(&cm_id_priv->lock);
2759         switch (cm_id_priv->id.state) {
2760         case IB_CM_REQ_SENT:
2761                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2762                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2763                                   cm_id_priv->msg, timeout))
2764                         goto out;
2765                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2766                 break;
2767         case IB_CM_REP_SENT:
2768                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2769                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2770                                   cm_id_priv->msg, timeout))
2771                         goto out;
2772                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2773                 break;
2774         case IB_CM_ESTABLISHED:
2775                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2776                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2777                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2778                                   cm_id_priv->msg, timeout)) {
2779                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2780                                 atomic_long_inc(&work->port->
2781                                                 counter_group[CM_RECV_DUPLICATES].
2782                                                 counter[CM_MRA_COUNTER]);
2783                         goto out;
2784                 }
2785                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2786                 break;
2787         case IB_CM_MRA_REQ_RCVD:
2788         case IB_CM_MRA_REP_RCVD:
2789                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2790                                 counter[CM_MRA_COUNTER]);
2791                 /* fall through */
2792         default:
2793                 goto out;
2794         }
2795
2796         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2797                                       cm_id_priv->id.state;
2798         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2799         if (!ret)
2800                 list_add_tail(&work->list, &cm_id_priv->work_list);
2801         spin_unlock_irq(&cm_id_priv->lock);
2802
2803         if (ret)
2804                 cm_process_work(cm_id_priv, work);
2805         else
2806                 cm_deref_id(cm_id_priv);
2807         return 0;
2808 out:
2809         spin_unlock_irq(&cm_id_priv->lock);
2810         cm_deref_id(cm_id_priv);
2811         return -EINVAL;
2812 }
2813
2814 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2815                           struct cm_id_private *cm_id_priv,
2816                           struct ib_sa_path_rec *alternate_path,
2817                           const void *private_data,
2818                           u8 private_data_len)
2819 {
2820         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2821                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2822         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2823         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2824         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2825         /* todo: need remote CM response timeout */
2826         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2827         lap_msg->alt_local_lid = alternate_path->slid;
2828         lap_msg->alt_remote_lid = alternate_path->dlid;
2829         lap_msg->alt_local_gid = alternate_path->sgid;
2830         lap_msg->alt_remote_gid = alternate_path->dgid;
2831         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2832         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2833         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2834         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2835         cm_lap_set_sl(lap_msg, alternate_path->sl);
2836         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2837         cm_lap_set_local_ack_timeout(lap_msg,
2838                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2839                                alternate_path->packet_life_time));
2840
2841         if (private_data && private_data_len)
2842                 memcpy(lap_msg->private_data, private_data, private_data_len);
2843 }
2844
2845 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2846                    struct ib_sa_path_rec *alternate_path,
2847                    const void *private_data,
2848                    u8 private_data_len)
2849 {
2850         struct cm_id_private *cm_id_priv;
2851         struct ib_mad_send_buf *msg;
2852         unsigned long flags;
2853         int ret;
2854
2855         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2856                 return -EINVAL;
2857
2858         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2859         spin_lock_irqsave(&cm_id_priv->lock, flags);
2860         if (cm_id->state != IB_CM_ESTABLISHED ||
2861             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2862              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2863                 ret = -EINVAL;
2864                 goto out;
2865         }
2866
2867         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av,
2868                                  cm_id_priv);
2869         if (ret)
2870                 goto out;
2871         cm_id_priv->alt_av.timeout =
2872                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2873                                        cm_id_priv->alt_av.timeout - 1);
2874
2875         ret = cm_alloc_msg(cm_id_priv, &msg);
2876         if (ret)
2877                 goto out;
2878
2879         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2880                       alternate_path, private_data, private_data_len);
2881         msg->timeout_ms = cm_id_priv->timeout_ms;
2882         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2883
2884         ret = ib_post_send_mad(msg, NULL);
2885         if (ret) {
2886                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2887                 cm_free_msg(msg);
2888                 return ret;
2889         }
2890
2891         cm_id->lap_state = IB_CM_LAP_SENT;
2892         cm_id_priv->msg = msg;
2893
2894 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2895         return ret;
2896 }
2897 EXPORT_SYMBOL(ib_send_cm_lap);
2898
2899 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2900                                     struct ib_sa_path_rec *path,
2901                                     struct cm_lap_msg *lap_msg)
2902 {
2903         memset(path, 0, sizeof *path);
2904         path->dgid = lap_msg->alt_local_gid;
2905         path->sgid = lap_msg->alt_remote_gid;
2906         path->dlid = lap_msg->alt_local_lid;
2907         path->slid = lap_msg->alt_remote_lid;
2908         path->flow_label = cm_lap_get_flow_label(lap_msg);
2909         path->hop_limit = lap_msg->alt_hop_limit;
2910         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2911         path->reversible = 1;
2912         path->pkey = cm_id_priv->pkey;
2913         path->sl = cm_lap_get_sl(lap_msg);
2914         path->mtu_selector = IB_SA_EQ;
2915         path->mtu = cm_id_priv->path_mtu;
2916         path->rate_selector = IB_SA_EQ;
2917         path->rate = cm_lap_get_packet_rate(lap_msg);
2918         path->packet_life_time_selector = IB_SA_EQ;
2919         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2920         path->packet_life_time -= (path->packet_life_time > 0);
2921 }
2922
2923 static int cm_lap_handler(struct cm_work *work)
2924 {
2925         struct cm_id_private *cm_id_priv;
2926         struct cm_lap_msg *lap_msg;
2927         struct ib_cm_lap_event_param *param;
2928         struct ib_mad_send_buf *msg = NULL;
2929         int ret;
2930
2931         /* todo: verify LAP request and send reject APR if invalid. */
2932         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2933         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2934                                    lap_msg->local_comm_id);
2935         if (!cm_id_priv)
2936                 return -EINVAL;
2937
2938         param = &work->cm_event.param.lap_rcvd;
2939         param->alternate_path = &work->path[0];
2940         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2941         work->cm_event.private_data = &lap_msg->private_data;
2942
2943         spin_lock_irq(&cm_id_priv->lock);
2944         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2945                 goto unlock;
2946
2947         switch (cm_id_priv->id.lap_state) {
2948         case IB_CM_LAP_UNINIT:
2949         case IB_CM_LAP_IDLE:
2950                 break;
2951         case IB_CM_MRA_LAP_SENT:
2952                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2953                                 counter[CM_LAP_COUNTER]);
2954                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2955                         goto unlock;
2956
2957                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2958                               CM_MSG_RESPONSE_OTHER,
2959                               cm_id_priv->service_timeout,
2960                               cm_id_priv->private_data,
2961                               cm_id_priv->private_data_len);
2962                 spin_unlock_irq(&cm_id_priv->lock);
2963
2964                 if (ib_post_send_mad(msg, NULL))
2965                         cm_free_msg(msg);
2966                 goto deref;
2967         case IB_CM_LAP_RCVD:
2968                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2969                                 counter[CM_LAP_COUNTER]);
2970                 goto unlock;
2971         default:
2972                 goto unlock;
2973         }
2974
2975         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2976         cm_id_priv->tid = lap_msg->hdr.tid;
2977         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2978                                 work->mad_recv_wc->recv_buf.grh,
2979                                 &cm_id_priv->av);
2980         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av,
2981                            cm_id_priv);
2982         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2983         if (!ret)
2984                 list_add_tail(&work->list, &cm_id_priv->work_list);
2985         spin_unlock_irq(&cm_id_priv->lock);
2986
2987         if (ret)
2988                 cm_process_work(cm_id_priv, work);
2989         else
2990                 cm_deref_id(cm_id_priv);
2991         return 0;
2992
2993 unlock: spin_unlock_irq(&cm_id_priv->lock);
2994 deref:  cm_deref_id(cm_id_priv);
2995         return -EINVAL;
2996 }
2997
2998 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2999                           struct cm_id_private *cm_id_priv,
3000                           enum ib_cm_apr_status status,
3001                           void *info,
3002                           u8 info_length,
3003                           const void *private_data,
3004                           u8 private_data_len)
3005 {
3006         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
3007         apr_msg->local_comm_id = cm_id_priv->id.local_id;
3008         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
3009         apr_msg->ap_status = (u8) status;
3010
3011         if (info && info_length) {
3012                 apr_msg->info_length = info_length;
3013                 memcpy(apr_msg->info, info, info_length);
3014         }
3015
3016         if (private_data && private_data_len)
3017                 memcpy(apr_msg->private_data, private_data, private_data_len);
3018 }
3019
3020 int ib_send_cm_apr(struct ib_cm_id *cm_id,
3021                    enum ib_cm_apr_status status,
3022                    void *info,
3023                    u8 info_length,
3024                    const void *private_data,
3025                    u8 private_data_len)
3026 {
3027         struct cm_id_private *cm_id_priv;
3028         struct ib_mad_send_buf *msg;
3029         unsigned long flags;
3030         int ret;
3031
3032         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
3033             (info && info_length > IB_CM_APR_INFO_LENGTH))
3034                 return -EINVAL;
3035
3036         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3037         spin_lock_irqsave(&cm_id_priv->lock, flags);
3038         if (cm_id->state != IB_CM_ESTABLISHED ||
3039             (cm_id->lap_state != IB_CM_LAP_RCVD &&
3040              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
3041                 ret = -EINVAL;
3042                 goto out;
3043         }
3044
3045         ret = cm_alloc_msg(cm_id_priv, &msg);
3046         if (ret)
3047                 goto out;
3048
3049         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
3050                       info, info_length, private_data, private_data_len);
3051         ret = ib_post_send_mad(msg, NULL);
3052         if (ret) {
3053                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3054                 cm_free_msg(msg);
3055                 return ret;
3056         }
3057
3058         cm_id->lap_state = IB_CM_LAP_IDLE;
3059 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3060         return ret;
3061 }
3062 EXPORT_SYMBOL(ib_send_cm_apr);
3063
3064 static int cm_apr_handler(struct cm_work *work)
3065 {
3066         struct cm_id_private *cm_id_priv;
3067         struct cm_apr_msg *apr_msg;
3068         int ret;
3069
3070         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3071         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
3072                                    apr_msg->local_comm_id);
3073         if (!cm_id_priv)
3074                 return -EINVAL; /* Unmatched reply. */
3075
3076         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
3077         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
3078         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
3079         work->cm_event.private_data = &apr_msg->private_data;
3080
3081         spin_lock_irq(&cm_id_priv->lock);
3082         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3083             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3084              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3085                 spin_unlock_irq(&cm_id_priv->lock);
3086                 goto out;
3087         }
3088         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3089         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3090         cm_id_priv->msg = NULL;
3091
3092         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3093         if (!ret)
3094                 list_add_tail(&work->list, &cm_id_priv->work_list);
3095         spin_unlock_irq(&cm_id_priv->lock);
3096
3097         if (ret)
3098                 cm_process_work(cm_id_priv, work);
3099         else
3100                 cm_deref_id(cm_id_priv);
3101         return 0;
3102 out:
3103         cm_deref_id(cm_id_priv);
3104         return -EINVAL;
3105 }
3106
3107 static int cm_timewait_handler(struct cm_work *work)
3108 {
3109         struct cm_timewait_info *timewait_info;
3110         struct cm_id_private *cm_id_priv;
3111         int ret;
3112
3113         timewait_info = (struct cm_timewait_info *)work;
3114         spin_lock_irq(&cm.lock);
3115         list_del(&timewait_info->list);
3116         spin_unlock_irq(&cm.lock);
3117
3118         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3119                                    timewait_info->work.remote_id);
3120         if (!cm_id_priv)
3121                 return -EINVAL;
3122
3123         spin_lock_irq(&cm_id_priv->lock);
3124         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3125             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3126                 spin_unlock_irq(&cm_id_priv->lock);
3127                 goto out;
3128         }
3129         cm_id_priv->id.state = IB_CM_IDLE;
3130         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3131         if (!ret)
3132                 list_add_tail(&work->list, &cm_id_priv->work_list);
3133         spin_unlock_irq(&cm_id_priv->lock);
3134
3135         if (ret)
3136                 cm_process_work(cm_id_priv, work);
3137         else
3138                 cm_deref_id(cm_id_priv);
3139         return 0;
3140 out:
3141         cm_deref_id(cm_id_priv);
3142         return -EINVAL;
3143 }
3144
3145 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3146                                struct cm_id_private *cm_id_priv,
3147                                struct ib_cm_sidr_req_param *param)
3148 {
3149         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3150                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
3151         sidr_req_msg->request_id = cm_id_priv->id.local_id;
3152         sidr_req_msg->pkey = param->path->pkey;
3153         sidr_req_msg->service_id = param->service_id;
3154
3155         if (param->private_data && param->private_data_len)
3156                 memcpy(sidr_req_msg->private_data, param->private_data,
3157                        param->private_data_len);
3158 }
3159
3160 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3161                         struct ib_cm_sidr_req_param *param)
3162 {
3163         struct cm_id_private *cm_id_priv;
3164         struct ib_mad_send_buf *msg;
3165         unsigned long flags;
3166         int ret;
3167
3168         if (!param->path || (param->private_data &&
3169              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3170                 return -EINVAL;
3171
3172         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3173         ret = cm_init_av_by_path(param->path, &cm_id_priv->av, cm_id_priv);
3174         if (ret)
3175                 goto out;
3176
3177         cm_id->service_id = param->service_id;
3178         cm_id->service_mask = ~cpu_to_be64(0);
3179         cm_id_priv->timeout_ms = param->timeout_ms;
3180         cm_id_priv->max_cm_retries = param->max_cm_retries;
3181         ret = cm_alloc_msg(cm_id_priv, &msg);
3182         if (ret)
3183                 goto out;
3184
3185         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3186                            param);
3187         msg->timeout_ms = cm_id_priv->timeout_ms;
3188         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3189
3190         spin_lock_irqsave(&cm_id_priv->lock, flags);
3191         if (cm_id->state == IB_CM_IDLE)
3192                 ret = ib_post_send_mad(msg, NULL);
3193         else
3194                 ret = -EINVAL;
3195
3196         if (ret) {
3197                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3198                 cm_free_msg(msg);
3199                 goto out;
3200         }
3201         cm_id->state = IB_CM_SIDR_REQ_SENT;
3202         cm_id_priv->msg = msg;
3203         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3204 out:
3205         return ret;
3206 }
3207 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3208
3209 static void cm_format_sidr_req_event(struct cm_work *work,
3210                                      struct ib_cm_id *listen_id)
3211 {
3212         struct cm_sidr_req_msg *sidr_req_msg;
3213         struct ib_cm_sidr_req_event_param *param;
3214
3215         sidr_req_msg = (struct cm_sidr_req_msg *)
3216                                 work->mad_recv_wc->recv_buf.mad;
3217         param = &work->cm_event.param.sidr_req_rcvd;
3218         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
3219         param->listen_id = listen_id;
3220         param->service_id = sidr_req_msg->service_id;
3221         param->bth_pkey = cm_get_bth_pkey(work);
3222         param->port = work->port->port_num;
3223         work->cm_event.private_data = &sidr_req_msg->private_data;
3224 }
3225
3226 static int cm_sidr_req_handler(struct cm_work *work)
3227 {
3228         struct ib_cm_id *cm_id;
3229         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3230         struct cm_sidr_req_msg *sidr_req_msg;
3231         struct ib_wc *wc;
3232
3233         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3234         if (IS_ERR(cm_id))
3235                 return PTR_ERR(cm_id);
3236         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3237
3238         /* Record SGID/SLID and request ID for lookup. */
3239         sidr_req_msg = (struct cm_sidr_req_msg *)
3240                                 work->mad_recv_wc->recv_buf.mad;
3241         wc = work->mad_recv_wc->wc;
3242         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3243         cm_id_priv->av.dgid.global.interface_id = 0;
3244         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3245                                 work->mad_recv_wc->recv_buf.grh,
3246                                 &cm_id_priv->av);
3247         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3248         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3249         atomic_inc(&cm_id_priv->work_count);
3250
3251         spin_lock_irq(&cm.lock);
3252         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3253         if (cur_cm_id_priv) {
3254                 spin_unlock_irq(&cm.lock);
3255                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3256                                 counter[CM_SIDR_REQ_COUNTER]);
3257                 goto out; /* Duplicate message. */
3258         }
3259         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3260         cur_cm_id_priv = cm_find_listen(cm_id->device,
3261                                         sidr_req_msg->service_id);
3262         if (!cur_cm_id_priv) {
3263                 spin_unlock_irq(&cm.lock);
3264                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3265                 goto out; /* No match. */
3266         }
3267         atomic_inc(&cur_cm_id_priv->refcount);
3268         atomic_inc(&cm_id_priv->refcount);
3269         spin_unlock_irq(&cm.lock);
3270
3271         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3272         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3273         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3274         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3275
3276         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3277         cm_process_work(cm_id_priv, work);
3278         cm_deref_id(cur_cm_id_priv);
3279         return 0;
3280 out:
3281         ib_destroy_cm_id(&cm_id_priv->id);
3282         return -EINVAL;
3283 }
3284
3285 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3286                                struct cm_id_private *cm_id_priv,
3287                                struct ib_cm_sidr_rep_param *param)
3288 {
3289         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3290                           cm_id_priv->tid);
3291         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3292         sidr_rep_msg->status = param->status;
3293         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3294         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3295         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3296
3297         if (param->info && param->info_length)
3298                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3299
3300         if (param->private_data && param->private_data_len)
3301                 memcpy(sidr_rep_msg->private_data, param->private_data,
3302                        param->private_data_len);
3303 }
3304
3305 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3306                         struct ib_cm_sidr_rep_param *param)
3307 {
3308         struct cm_id_private *cm_id_priv;
3309         struct ib_mad_send_buf *msg;
3310         unsigned long flags;
3311         int ret;
3312
3313         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3314             (param->private_data &&
3315              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3316                 return -EINVAL;
3317
3318         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3319         spin_lock_irqsave(&cm_id_priv->lock, flags);
3320         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3321                 ret = -EINVAL;
3322                 goto error;
3323         }
3324
3325         ret = cm_alloc_msg(cm_id_priv, &msg);
3326         if (ret)
3327                 goto error;
3328
3329         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3330                            param);
3331         ret = ib_post_send_mad(msg, NULL);
3332         if (ret) {
3333                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3334                 cm_free_msg(msg);
3335                 return ret;
3336         }
3337         cm_id->state = IB_CM_IDLE;
3338         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3339
3340         spin_lock_irqsave(&cm.lock, flags);
3341         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3342                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3343                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3344         }
3345         spin_unlock_irqrestore(&cm.lock, flags);
3346         return 0;
3347
3348 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3349         return ret;
3350 }
3351 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3352
3353 static void cm_format_sidr_rep_event(struct cm_work *work)
3354 {
3355         struct cm_sidr_rep_msg *sidr_rep_msg;
3356         struct ib_cm_sidr_rep_event_param *param;
3357
3358         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3359                                 work->mad_recv_wc->recv_buf.mad;
3360         param = &work->cm_event.param.sidr_rep_rcvd;
3361         param->status = sidr_rep_msg->status;
3362         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3363         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3364         param->info = &sidr_rep_msg->info;
3365         param->info_len = sidr_rep_msg->info_length;
3366         work->cm_event.private_data = &sidr_rep_msg->private_data;
3367 }
3368
3369 static int cm_sidr_rep_handler(struct cm_work *work)
3370 {
3371         struct cm_sidr_rep_msg *sidr_rep_msg;
3372         struct cm_id_private *cm_id_priv;
3373
3374         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3375                                 work->mad_recv_wc->recv_buf.mad;
3376         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3377         if (!cm_id_priv)
3378                 return -EINVAL; /* Unmatched reply. */
3379
3380         spin_lock_irq(&cm_id_priv->lock);
3381         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3382                 spin_unlock_irq(&cm_id_priv->lock);
3383                 goto out;
3384         }
3385         cm_id_priv->id.state = IB_CM_IDLE;
3386         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3387         spin_unlock_irq(&cm_id_priv->lock);
3388
3389         cm_format_sidr_rep_event(work);
3390         cm_process_work(cm_id_priv, work);
3391         return 0;
3392 out:
3393         cm_deref_id(cm_id_priv);
3394         return -EINVAL;
3395 }
3396
3397 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3398                                   enum ib_wc_status wc_status)
3399 {
3400         struct cm_id_private *cm_id_priv;
3401         struct ib_cm_event cm_event;
3402         enum ib_cm_state state;
3403         int ret;
3404
3405         memset(&cm_event, 0, sizeof cm_event);
3406         cm_id_priv = msg->context[0];
3407
3408         /* Discard old sends or ones without a response. */
3409         spin_lock_irq(&cm_id_priv->lock);
3410         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3411         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3412                 goto discard;
3413
3414         pr_debug_ratelimited("CM: failed sending MAD in state %d. (%s)\n",
3415                              state, ib_wc_status_msg(wc_status));
3416         switch (state) {
3417         case IB_CM_REQ_SENT:
3418         case IB_CM_MRA_REQ_RCVD:
3419                 cm_reset_to_idle(cm_id_priv);
3420                 cm_event.event = IB_CM_REQ_ERROR;
3421                 break;
3422         case IB_CM_REP_SENT:
3423         case IB_CM_MRA_REP_RCVD:
3424                 cm_reset_to_idle(cm_id_priv);
3425                 cm_event.event = IB_CM_REP_ERROR;
3426                 break;
3427         case IB_CM_DREQ_SENT:
3428                 cm_enter_timewait(cm_id_priv);
3429                 cm_event.event = IB_CM_DREQ_ERROR;
3430                 break;
3431         case IB_CM_SIDR_REQ_SENT:
3432                 cm_id_priv->id.state = IB_CM_IDLE;
3433                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3434                 break;
3435         default:
3436                 goto discard;
3437         }
3438         spin_unlock_irq(&cm_id_priv->lock);
3439         cm_event.param.send_status = wc_status;
3440
3441         /* No other events can occur on the cm_id at this point. */
3442         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3443         cm_free_msg(msg);
3444         if (ret)
3445                 ib_destroy_cm_id(&cm_id_priv->id);
3446         return;
3447 discard:
3448         spin_unlock_irq(&cm_id_priv->lock);
3449         cm_free_msg(msg);
3450 }
3451
3452 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3453                             struct ib_mad_send_wc *mad_send_wc)
3454 {
3455         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3456         struct cm_port *port;
3457         u16 attr_index;
3458
3459         port = mad_agent->context;
3460         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3461                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3462
3463         /*
3464          * If the send was in response to a received message (context[0] is not
3465          * set to a cm_id), and is not a REJ, then it is a send that was
3466          * manually retried.
3467          */
3468         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3469                 msg->retries = 1;
3470
3471         atomic_long_add(1 + msg->retries,
3472                         &port->counter_group[CM_XMIT].counter[attr_index]);
3473         if (msg->retries)
3474                 atomic_long_add(msg->retries,
3475                                 &port->counter_group[CM_XMIT_RETRIES].
3476                                 counter[attr_index]);
3477
3478         switch (mad_send_wc->status) {
3479         case IB_WC_SUCCESS:
3480         case IB_WC_WR_FLUSH_ERR:
3481                 cm_free_msg(msg);
3482                 break;
3483         default:
3484                 if (msg->context[0] && msg->context[1])
3485                         cm_process_send_error(msg, mad_send_wc->status);
3486                 else
3487                         cm_free_msg(msg);
3488                 break;
3489         }
3490 }
3491
3492 static void cm_work_handler(struct work_struct *_work)
3493 {
3494         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3495         int ret;
3496
3497         switch (work->cm_event.event) {
3498         case IB_CM_REQ_RECEIVED:
3499                 ret = cm_req_handler(work);
3500                 break;
3501         case IB_CM_MRA_RECEIVED:
3502                 ret = cm_mra_handler(work);
3503                 break;
3504         case IB_CM_REJ_RECEIVED:
3505                 ret = cm_rej_handler(work);
3506                 break;
3507         case IB_CM_REP_RECEIVED:
3508                 ret = cm_rep_handler(work);
3509                 break;
3510         case IB_CM_RTU_RECEIVED:
3511                 ret = cm_rtu_handler(work);
3512                 break;
3513         case IB_CM_USER_ESTABLISHED:
3514                 ret = cm_establish_handler(work);
3515                 break;
3516         case IB_CM_DREQ_RECEIVED:
3517                 ret = cm_dreq_handler(work);
3518                 break;
3519         case IB_CM_DREP_RECEIVED:
3520                 ret = cm_drep_handler(work);
3521                 break;
3522         case IB_CM_SIDR_REQ_RECEIVED:
3523                 ret = cm_sidr_req_handler(work);
3524                 break;
3525         case IB_CM_SIDR_REP_RECEIVED:
3526                 ret = cm_sidr_rep_handler(work);
3527                 break;
3528         case IB_CM_LAP_RECEIVED:
3529                 ret = cm_lap_handler(work);
3530                 break;
3531         case IB_CM_APR_RECEIVED:
3532                 ret = cm_apr_handler(work);
3533                 break;
3534         case IB_CM_TIMEWAIT_EXIT:
3535                 ret = cm_timewait_handler(work);
3536                 break;
3537         default:
3538                 ret = -EINVAL;
3539                 break;
3540         }
3541         if (ret)
3542                 cm_free_work(work);
3543 }
3544
3545 static int cm_establish(struct ib_cm_id *cm_id)
3546 {
3547         struct cm_id_private *cm_id_priv;
3548         struct cm_work *work;
3549         unsigned long flags;
3550         int ret = 0;
3551         struct cm_device *cm_dev;
3552
3553         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3554         if (!cm_dev)
3555                 return -ENODEV;
3556
3557         work = kmalloc(sizeof *work, GFP_ATOMIC);
3558         if (!work)
3559                 return -ENOMEM;
3560
3561         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3562         spin_lock_irqsave(&cm_id_priv->lock, flags);
3563         switch (cm_id->state)
3564         {
3565         case IB_CM_REP_SENT:
3566         case IB_CM_MRA_REP_RCVD:
3567                 cm_id->state = IB_CM_ESTABLISHED;
3568                 break;
3569         case IB_CM_ESTABLISHED:
3570                 ret = -EISCONN;
3571                 break;
3572         default:
3573                 ret = -EINVAL;
3574                 break;
3575         }
3576         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3577
3578         if (ret) {
3579                 kfree(work);
3580                 goto out;
3581         }
3582
3583         /*
3584          * The CM worker thread may try to destroy the cm_id before it
3585          * can execute this work item.  To prevent potential deadlock,
3586          * we need to find the cm_id once we're in the context of the
3587          * worker thread, rather than holding a reference on it.
3588          */
3589         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3590         work->local_id = cm_id->local_id;
3591         work->remote_id = cm_id->remote_id;
3592         work->mad_recv_wc = NULL;
3593         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3594
3595         /* Check if the device started its remove_one */
3596         spin_lock_irqsave(&cm.lock, flags);
3597         if (!cm_dev->going_down) {
3598                 queue_delayed_work(cm.wq, &work->work, 0);
3599         } else {
3600                 kfree(work);
3601                 ret = -ENODEV;
3602         }
3603         spin_unlock_irqrestore(&cm.lock, flags);
3604
3605 out:
3606         return ret;
3607 }
3608
3609 static int cm_migrate(struct ib_cm_id *cm_id)
3610 {
3611         struct cm_id_private *cm_id_priv;
3612         struct cm_av tmp_av;
3613         unsigned long flags;
3614         int tmp_send_port_not_ready;
3615         int ret = 0;
3616
3617         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3618         spin_lock_irqsave(&cm_id_priv->lock, flags);
3619         if (cm_id->state == IB_CM_ESTABLISHED &&
3620             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3621              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3622                 cm_id->lap_state = IB_CM_LAP_IDLE;
3623                 /* Swap address vector */
3624                 tmp_av = cm_id_priv->av;
3625                 cm_id_priv->av = cm_id_priv->alt_av;
3626                 cm_id_priv->alt_av = tmp_av;
3627                 /* Swap port send ready state */
3628                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3629                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3630                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3631         } else
3632                 ret = -EINVAL;
3633         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3634
3635         return ret;
3636 }
3637
3638 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3639 {
3640         int ret;
3641
3642         switch (event) {
3643         case IB_EVENT_COMM_EST:
3644                 ret = cm_establish(cm_id);
3645                 break;
3646         case IB_EVENT_PATH_MIG:
3647                 ret = cm_migrate(cm_id);
3648                 break;
3649         default:
3650                 ret = -EINVAL;
3651         }
3652         return ret;
3653 }
3654 EXPORT_SYMBOL(ib_cm_notify);
3655
3656 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3657                             struct ib_mad_send_buf *send_buf,
3658                             struct ib_mad_recv_wc *mad_recv_wc)
3659 {
3660         struct cm_port *port = mad_agent->context;
3661         struct cm_work *work;
3662         enum ib_cm_event_type event;
3663         u16 attr_id;
3664         int paths = 0;
3665         int going_down = 0;
3666
3667         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3668         case CM_REQ_ATTR_ID:
3669                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3670                                                     alt_local_lid != 0);
3671                 event = IB_CM_REQ_RECEIVED;
3672                 break;
3673         case CM_MRA_ATTR_ID:
3674                 event = IB_CM_MRA_RECEIVED;
3675                 break;
3676         case CM_REJ_ATTR_ID:
3677                 event = IB_CM_REJ_RECEIVED;
3678                 break;
3679         case CM_REP_ATTR_ID:
3680                 event = IB_CM_REP_RECEIVED;
3681                 break;
3682         case CM_RTU_ATTR_ID:
3683                 event = IB_CM_RTU_RECEIVED;
3684                 break;
3685         case CM_DREQ_ATTR_ID:
3686                 event = IB_CM_DREQ_RECEIVED;
3687                 break;
3688         case CM_DREP_ATTR_ID:
3689                 event = IB_CM_DREP_RECEIVED;
3690                 break;
3691         case CM_SIDR_REQ_ATTR_ID:
3692                 event = IB_CM_SIDR_REQ_RECEIVED;
3693                 break;
3694         case CM_SIDR_REP_ATTR_ID:
3695                 event = IB_CM_SIDR_REP_RECEIVED;
3696                 break;
3697         case CM_LAP_ATTR_ID:
3698                 paths = 1;
3699                 event = IB_CM_LAP_RECEIVED;
3700                 break;
3701         case CM_APR_ATTR_ID:
3702                 event = IB_CM_APR_RECEIVED;
3703                 break;
3704         default:
3705                 ib_free_recv_mad(mad_recv_wc);
3706                 return;
3707         }
3708
3709         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3710         atomic_long_inc(&port->counter_group[CM_RECV].
3711                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3712
3713         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3714                        GFP_KERNEL);
3715         if (!work) {
3716                 ib_free_recv_mad(mad_recv_wc);
3717                 return;
3718         }
3719
3720         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3721         work->cm_event.event = event;
3722         work->mad_recv_wc = mad_recv_wc;
3723         work->port = port;
3724
3725         /* Check if the device started its remove_one */
3726         spin_lock_irq(&cm.lock);
3727         if (!port->cm_dev->going_down)
3728                 queue_delayed_work(cm.wq, &work->work, 0);
3729         else
3730                 going_down = 1;
3731         spin_unlock_irq(&cm.lock);
3732
3733         if (going_down) {
3734                 kfree(work);
3735                 ib_free_recv_mad(mad_recv_wc);
3736         }
3737 }
3738
3739 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3740                                 struct ib_qp_attr *qp_attr,
3741                                 int *qp_attr_mask)
3742 {
3743         unsigned long flags;
3744         int ret;
3745
3746         spin_lock_irqsave(&cm_id_priv->lock, flags);
3747         switch (cm_id_priv->id.state) {
3748         case IB_CM_REQ_SENT:
3749         case IB_CM_MRA_REQ_RCVD:
3750         case IB_CM_REQ_RCVD:
3751         case IB_CM_MRA_REQ_SENT:
3752         case IB_CM_REP_RCVD:
3753         case IB_CM_MRA_REP_SENT:
3754         case IB_CM_REP_SENT:
3755         case IB_CM_MRA_REP_RCVD:
3756         case IB_CM_ESTABLISHED:
3757                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3758                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3759                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3760                 if (cm_id_priv->responder_resources)
3761                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3762                                                     IB_ACCESS_REMOTE_ATOMIC;
3763                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3764                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3765                 ret = 0;
3766                 break;
3767         default:
3768                 ret = -EINVAL;
3769                 break;
3770         }
3771         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3772         return ret;
3773 }
3774
3775 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3776                                struct ib_qp_attr *qp_attr,
3777                                int *qp_attr_mask)
3778 {
3779         unsigned long flags;
3780         int ret;
3781
3782         spin_lock_irqsave(&cm_id_priv->lock, flags);
3783         switch (cm_id_priv->id.state) {
3784         case IB_CM_REQ_RCVD:
3785         case IB_CM_MRA_REQ_SENT:
3786         case IB_CM_REP_RCVD:
3787         case IB_CM_MRA_REP_SENT:
3788         case IB_CM_REP_SENT:
3789         case IB_CM_MRA_REP_RCVD:
3790         case IB_CM_ESTABLISHED:
3791                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3792                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3793                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3794                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3795                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3796                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3797                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3798                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3799                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3800                                          IB_QP_MIN_RNR_TIMER;
3801                         qp_attr->max_dest_rd_atomic =
3802                                         cm_id_priv->responder_resources;
3803                         qp_attr->min_rnr_timer = 0;
3804                 }
3805                 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
3806                         *qp_attr_mask |= IB_QP_ALT_PATH;
3807                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3808                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3809                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3810                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3811                 }
3812                 ret = 0;
3813                 break;
3814         default:
3815                 ret = -EINVAL;
3816                 break;
3817         }
3818         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3819         return ret;
3820 }
3821
3822 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3823                                struct ib_qp_attr *qp_attr,
3824                                int *qp_attr_mask)
3825 {
3826         unsigned long flags;
3827         int ret;
3828
3829         spin_lock_irqsave(&cm_id_priv->lock, flags);
3830         switch (cm_id_priv->id.state) {
3831         /* Allow transition to RTS before sending REP */
3832         case IB_CM_REQ_RCVD:
3833         case IB_CM_MRA_REQ_SENT:
3834
3835         case IB_CM_REP_RCVD:
3836         case IB_CM_MRA_REP_SENT:
3837         case IB_CM_REP_SENT:
3838         case IB_CM_MRA_REP_RCVD:
3839         case IB_CM_ESTABLISHED:
3840                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3841                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3842                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3843                         switch (cm_id_priv->qp_type) {
3844                         case IB_QPT_RC:
3845                         case IB_QPT_XRC_INI:
3846                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3847                                                  IB_QP_MAX_QP_RD_ATOMIC;
3848                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3849                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3850                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3851                                 /* fall through */
3852                         case IB_QPT_XRC_TGT:
3853                                 *qp_attr_mask |= IB_QP_TIMEOUT;
3854                                 qp_attr->timeout = cm_id_priv->av.timeout;
3855                                 break;
3856                         default:
3857                                 break;
3858                         }
3859                         if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
3860                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3861                                 qp_attr->path_mig_state = IB_MIG_REARM;
3862                         }
3863                 } else {
3864                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3865                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3866                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3867                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3868                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3869                         qp_attr->path_mig_state = IB_MIG_REARM;
3870                 }
3871                 ret = 0;
3872                 break;
3873         default:
3874                 ret = -EINVAL;
3875                 break;
3876         }
3877         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3878         return ret;
3879 }
3880
3881 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3882                        struct ib_qp_attr *qp_attr,
3883                        int *qp_attr_mask)
3884 {
3885         struct cm_id_private *cm_id_priv;
3886         int ret;
3887
3888         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3889         switch (qp_attr->qp_state) {
3890         case IB_QPS_INIT:
3891                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3892                 break;
3893         case IB_QPS_RTR:
3894                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3895                 break;
3896         case IB_QPS_RTS:
3897                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3898                 break;
3899         default:
3900                 ret = -EINVAL;
3901                 break;
3902         }
3903         return ret;
3904 }
3905 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3906
3907 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3908                                char *buf)
3909 {
3910         struct cm_counter_group *group;
3911         struct cm_counter_attribute *cm_attr;
3912
3913         group = container_of(obj, struct cm_counter_group, obj);
3914         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3915
3916         return sprintf(buf, "%ld\n",
3917                        atomic_long_read(&group->counter[cm_attr->index]));
3918 }
3919
3920 static const struct sysfs_ops cm_counter_ops = {
3921         .show = cm_show_counter
3922 };
3923
3924 static struct kobj_type cm_counter_obj_type = {
3925         .sysfs_ops = &cm_counter_ops,
3926         .default_attrs = cm_counter_default_attrs
3927 };
3928
3929 static void cm_release_port_obj(struct kobject *obj)
3930 {
3931         struct cm_port *cm_port;
3932
3933         cm_port = container_of(obj, struct cm_port, port_obj);
3934         kfree(cm_port);
3935 }
3936
3937 static struct kobj_type cm_port_obj_type = {
3938         .release = cm_release_port_obj
3939 };
3940
3941 static char *cm_devnode(struct device *dev, umode_t *mode)
3942 {
3943         if (mode)
3944                 *mode = 0666;
3945         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3946 }
3947
3948 struct class cm_class = {
3949         .owner   = THIS_MODULE,
3950         .name    = "infiniband_cm",
3951         .devnode = cm_devnode,
3952 };
3953 EXPORT_SYMBOL(cm_class);
3954
3955 static int cm_create_port_fs(struct cm_port *port)
3956 {
3957         int i, ret;
3958
3959         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3960                                    &port->cm_dev->device->kobj,
3961                                    "%d", port->port_num);
3962         if (ret) {
3963                 kfree(port);
3964                 return ret;
3965         }
3966
3967         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3968                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3969                                            &cm_counter_obj_type,
3970                                            &port->port_obj,
3971                                            "%s", counter_group_names[i]);
3972                 if (ret)
3973                         goto error;
3974         }
3975
3976         return 0;
3977
3978 error:
3979         while (i--)
3980                 kobject_put(&port->counter_group[i].obj);
3981         kobject_put(&port->port_obj);
3982         return ret;
3983
3984 }
3985
3986 static void cm_remove_port_fs(struct cm_port *port)
3987 {
3988         int i;
3989
3990         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3991                 kobject_put(&port->counter_group[i].obj);
3992
3993         kobject_put(&port->port_obj);
3994 }
3995
3996 static void cm_add_one(struct ib_device *ib_device)
3997 {
3998         struct cm_device *cm_dev;
3999         struct cm_port *port;
4000         struct ib_mad_reg_req reg_req = {
4001                 .mgmt_class = IB_MGMT_CLASS_CM,
4002                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4003         };
4004         struct ib_port_modify port_modify = {
4005                 .set_port_cap_mask = IB_PORT_CM_SUP
4006         };
4007         unsigned long flags;
4008         int ret;
4009         int count = 0;
4010         u8 i;
4011
4012         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
4013                          ib_device->phys_port_cnt, GFP_KERNEL);
4014         if (!cm_dev)
4015                 return;
4016
4017         cm_dev->ib_device = ib_device;
4018         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4019         cm_dev->going_down = 0;
4020         cm_dev->device = device_create(&cm_class, &ib_device->dev,
4021                                        MKDEV(0, 0), NULL,
4022                                        "%s", ib_device->name);
4023         if (IS_ERR(cm_dev->device)) {
4024                 kfree(cm_dev);
4025                 return;
4026         }
4027
4028         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4029         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4030                 if (!rdma_cap_ib_cm(ib_device, i))
4031                         continue;
4032
4033                 port = kzalloc(sizeof *port, GFP_KERNEL);
4034                 if (!port)
4035                         goto error1;
4036
4037                 cm_dev->port[i-1] = port;
4038                 port->cm_dev = cm_dev;
4039                 port->port_num = i;
4040
4041                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
4042                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
4043
4044                 ret = cm_create_port_fs(port);
4045                 if (ret)
4046                         goto error1;
4047
4048                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4049                                                         IB_QPT_GSI,
4050                                                         &reg_req,
4051                                                         0,
4052                                                         cm_send_handler,
4053                                                         cm_recv_handler,
4054                                                         port,
4055                                                         0);
4056                 if (IS_ERR(port->mad_agent))
4057                         goto error2;
4058
4059                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4060                 if (ret)
4061                         goto error3;
4062
4063                 count++;
4064         }
4065
4066         if (!count)
4067                 goto free;
4068
4069         ib_set_client_data(ib_device, &cm_client, cm_dev);
4070
4071         write_lock_irqsave(&cm.device_lock, flags);
4072         list_add_tail(&cm_dev->list, &cm.device_list);
4073         write_unlock_irqrestore(&cm.device_lock, flags);
4074         return;
4075
4076 error3:
4077         ib_unregister_mad_agent(port->mad_agent);
4078 error2:
4079         cm_remove_port_fs(port);
4080 error1:
4081         port_modify.set_port_cap_mask = 0;
4082         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4083         while (--i) {
4084                 if (!rdma_cap_ib_cm(ib_device, i))
4085                         continue;
4086
4087                 port = cm_dev->port[i-1];
4088                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4089                 ib_unregister_mad_agent(port->mad_agent);
4090                 cm_remove_port_fs(port);
4091         }
4092 free:
4093         device_unregister(cm_dev->device);
4094         kfree(cm_dev);
4095 }
4096
4097 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4098 {
4099         struct cm_device *cm_dev = client_data;
4100         struct cm_port *port;
4101         struct cm_id_private *cm_id_priv;
4102         struct ib_mad_agent *cur_mad_agent;
4103         struct ib_port_modify port_modify = {
4104                 .clr_port_cap_mask = IB_PORT_CM_SUP
4105         };
4106         unsigned long flags;
4107         int i;
4108
4109         if (!cm_dev)
4110                 return;
4111
4112         write_lock_irqsave(&cm.device_lock, flags);
4113         list_del(&cm_dev->list);
4114         write_unlock_irqrestore(&cm.device_lock, flags);
4115
4116         spin_lock_irq(&cm.lock);
4117         cm_dev->going_down = 1;
4118         spin_unlock_irq(&cm.lock);
4119
4120         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4121                 if (!rdma_cap_ib_cm(ib_device, i))
4122                         continue;
4123
4124                 port = cm_dev->port[i-1];
4125                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4126                 /* Mark all the cm_id's as not valid */
4127                 spin_lock_irq(&cm.lock);
4128                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4129                         cm_id_priv->altr_send_port_not_ready = 1;
4130                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4131                         cm_id_priv->prim_send_port_not_ready = 1;
4132                 spin_unlock_irq(&cm.lock);
4133                 /*
4134                  * We flush the queue here after the going_down set, this
4135                  * verify that no new works will be queued in the recv handler,
4136                  * after that we can call the unregister_mad_agent
4137                  */
4138                 flush_workqueue(cm.wq);
4139                 spin_lock_irq(&cm.state_lock);
4140                 cur_mad_agent = port->mad_agent;
4141                 port->mad_agent = NULL;
4142                 spin_unlock_irq(&cm.state_lock);
4143                 ib_unregister_mad_agent(cur_mad_agent);
4144                 cm_remove_port_fs(port);
4145         }
4146
4147         device_unregister(cm_dev->device);
4148         kfree(cm_dev);
4149 }
4150
4151 static int __init ib_cm_init(void)
4152 {
4153         int ret;
4154
4155         memset(&cm, 0, sizeof cm);
4156         INIT_LIST_HEAD(&cm.device_list);
4157         rwlock_init(&cm.device_lock);
4158         spin_lock_init(&cm.lock);
4159         spin_lock_init(&cm.state_lock);
4160         cm.listen_service_table = RB_ROOT;
4161         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4162         cm.remote_id_table = RB_ROOT;
4163         cm.remote_qp_table = RB_ROOT;
4164         cm.remote_sidr_table = RB_ROOT;
4165         idr_init(&cm.local_id_table);
4166         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4167         INIT_LIST_HEAD(&cm.timewait_list);
4168
4169         ret = class_register(&cm_class);
4170         if (ret) {
4171                 ret = -ENOMEM;
4172                 goto error1;
4173         }
4174
4175         cm.wq = create_workqueue("ib_cm");
4176         if (!cm.wq) {
4177                 ret = -ENOMEM;
4178                 goto error2;
4179         }
4180
4181         ret = ib_register_client(&cm_client);
4182         if (ret)
4183                 goto error3;
4184
4185         return 0;
4186 error3:
4187         destroy_workqueue(cm.wq);
4188 error2:
4189         class_unregister(&cm_class);
4190 error1:
4191         idr_destroy(&cm.local_id_table);
4192         return ret;
4193 }
4194
4195 static void __exit ib_cm_cleanup(void)
4196 {
4197         struct cm_timewait_info *timewait_info, *tmp;
4198
4199         spin_lock_irq(&cm.lock);
4200         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4201                 cancel_delayed_work(&timewait_info->work.work);
4202         spin_unlock_irq(&cm.lock);
4203
4204         ib_unregister_client(&cm_client);
4205         destroy_workqueue(cm.wq);
4206
4207         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4208                 list_del(&timewait_info->list);
4209                 kfree(timewait_info);
4210         }
4211
4212         class_unregister(&cm_class);
4213         idr_destroy(&cm.local_id_table);
4214 }
4215
4216 module_init(ib_cm_init);
4217 module_exit(ib_cm_cleanup);
4218