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 ib_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 = ib_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                 ib_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                 ib_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         ib_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         int ret;
1726
1727         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1728
1729         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1730         if (IS_ERR(cm_id))
1731                 return PTR_ERR(cm_id);
1732
1733         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1734         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1735         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1736                                 work->mad_recv_wc->recv_buf.grh,
1737                                 &cm_id_priv->av);
1738         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1739                                                             id.local_id);
1740         if (IS_ERR(cm_id_priv->timewait_info)) {
1741                 ret = PTR_ERR(cm_id_priv->timewait_info);
1742                 goto destroy;
1743         }
1744         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1745         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1746         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1747
1748         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1749         if (!listen_cm_id_priv) {
1750                 ret = -EINVAL;
1751                 kfree(cm_id_priv->timewait_info);
1752                 goto destroy;
1753         }
1754
1755         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1756         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1757         cm_id_priv->id.service_id = req_msg->service_id;
1758         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1759
1760         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1761         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1762
1763         memcpy(work->path[0].dmac, cm_id_priv->av.ah_attr.dmac, ETH_ALEN);
1764         work->path[0].hop_limit = cm_id_priv->av.ah_attr.grh.hop_limit;
1765         ret = ib_get_cached_gid(work->port->cm_dev->ib_device,
1766                                 work->port->port_num,
1767                                 cm_id_priv->av.ah_attr.grh.sgid_index,
1768                                 &gid, &gid_attr);
1769         if (!ret) {
1770                 if (gid_attr.ndev) {
1771                         work->path[0].ifindex = gid_attr.ndev->ifindex;
1772                         work->path[0].net = dev_net(gid_attr.ndev);
1773                         dev_put(gid_attr.ndev);
1774                 }
1775                 work->path[0].gid_type = gid_attr.gid_type;
1776                 ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av,
1777                                          cm_id_priv);
1778         }
1779         if (ret) {
1780                 int err = ib_get_cached_gid(work->port->cm_dev->ib_device,
1781                                             work->port->port_num, 0,
1782                                             &work->path[0].sgid,
1783                                             &gid_attr);
1784                 if (!err && gid_attr.ndev) {
1785                         work->path[0].ifindex = gid_attr.ndev->ifindex;
1786                         work->path[0].net = dev_net(gid_attr.ndev);
1787                         dev_put(gid_attr.ndev);
1788                 }
1789                 work->path[0].gid_type = gid_attr.gid_type;
1790                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1791                                &work->path[0].sgid, sizeof work->path[0].sgid,
1792                                NULL, 0);
1793                 goto rejected;
1794         }
1795         if (req_msg->alt_local_lid) {
1796                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av,
1797                                          cm_id_priv);
1798                 if (ret) {
1799                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1800                                        &work->path[0].sgid,
1801                                        sizeof work->path[0].sgid, NULL, 0);
1802                         goto rejected;
1803                 }
1804         }
1805         cm_id_priv->tid = req_msg->hdr.tid;
1806         cm_id_priv->timeout_ms = cm_convert_to_ms(
1807                                         cm_req_get_local_resp_timeout(req_msg));
1808         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1809         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1810         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1811         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1812         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1813         cm_id_priv->pkey = req_msg->pkey;
1814         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1815         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1816         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1817         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1818
1819         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1820         cm_process_work(cm_id_priv, work);
1821         cm_deref_id(listen_cm_id_priv);
1822         return 0;
1823
1824 rejected:
1825         atomic_dec(&cm_id_priv->refcount);
1826         cm_deref_id(listen_cm_id_priv);
1827 destroy:
1828         ib_destroy_cm_id(cm_id);
1829         return ret;
1830 }
1831
1832 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1833                           struct cm_id_private *cm_id_priv,
1834                           struct ib_cm_rep_param *param)
1835 {
1836         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1837         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1838         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1839         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1840         rep_msg->resp_resources = param->responder_resources;
1841         cm_rep_set_target_ack_delay(rep_msg,
1842                                     cm_id_priv->av.port->cm_dev->ack_delay);
1843         cm_rep_set_failover(rep_msg, param->failover_accepted);
1844         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1845         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1846
1847         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1848                 rep_msg->initiator_depth = param->initiator_depth;
1849                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1850                 cm_rep_set_srq(rep_msg, param->srq);
1851                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1852         } else {
1853                 cm_rep_set_srq(rep_msg, 1);
1854                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1855         }
1856
1857         if (param->private_data && param->private_data_len)
1858                 memcpy(rep_msg->private_data, param->private_data,
1859                        param->private_data_len);
1860 }
1861
1862 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1863                    struct ib_cm_rep_param *param)
1864 {
1865         struct cm_id_private *cm_id_priv;
1866         struct ib_mad_send_buf *msg;
1867         struct cm_rep_msg *rep_msg;
1868         unsigned long flags;
1869         int ret;
1870
1871         if (param->private_data &&
1872             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1873                 return -EINVAL;
1874
1875         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1876         spin_lock_irqsave(&cm_id_priv->lock, flags);
1877         if (cm_id->state != IB_CM_REQ_RCVD &&
1878             cm_id->state != IB_CM_MRA_REQ_SENT) {
1879                 ret = -EINVAL;
1880                 goto out;
1881         }
1882
1883         ret = cm_alloc_msg(cm_id_priv, &msg);
1884         if (ret)
1885                 goto out;
1886
1887         rep_msg = (struct cm_rep_msg *) msg->mad;
1888         cm_format_rep(rep_msg, cm_id_priv, param);
1889         msg->timeout_ms = cm_id_priv->timeout_ms;
1890         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1891
1892         ret = ib_post_send_mad(msg, NULL);
1893         if (ret) {
1894                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1895                 cm_free_msg(msg);
1896                 return ret;
1897         }
1898
1899         cm_id->state = IB_CM_REP_SENT;
1900         cm_id_priv->msg = msg;
1901         cm_id_priv->initiator_depth = param->initiator_depth;
1902         cm_id_priv->responder_resources = param->responder_resources;
1903         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1904         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
1905
1906 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1907         return ret;
1908 }
1909 EXPORT_SYMBOL(ib_send_cm_rep);
1910
1911 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1912                           struct cm_id_private *cm_id_priv,
1913                           const void *private_data,
1914                           u8 private_data_len)
1915 {
1916         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1917         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1918         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1919
1920         if (private_data && private_data_len)
1921                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1922 }
1923
1924 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1925                    const void *private_data,
1926                    u8 private_data_len)
1927 {
1928         struct cm_id_private *cm_id_priv;
1929         struct ib_mad_send_buf *msg;
1930         unsigned long flags;
1931         void *data;
1932         int ret;
1933
1934         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1935                 return -EINVAL;
1936
1937         data = cm_copy_private_data(private_data, private_data_len);
1938         if (IS_ERR(data))
1939                 return PTR_ERR(data);
1940
1941         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1942         spin_lock_irqsave(&cm_id_priv->lock, flags);
1943         if (cm_id->state != IB_CM_REP_RCVD &&
1944             cm_id->state != IB_CM_MRA_REP_SENT) {
1945                 ret = -EINVAL;
1946                 goto error;
1947         }
1948
1949         ret = cm_alloc_msg(cm_id_priv, &msg);
1950         if (ret)
1951                 goto error;
1952
1953         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1954                       private_data, private_data_len);
1955
1956         ret = ib_post_send_mad(msg, NULL);
1957         if (ret) {
1958                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1959                 cm_free_msg(msg);
1960                 kfree(data);
1961                 return ret;
1962         }
1963
1964         cm_id->state = IB_CM_ESTABLISHED;
1965         cm_set_private_data(cm_id_priv, data, private_data_len);
1966         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1967         return 0;
1968
1969 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1970         kfree(data);
1971         return ret;
1972 }
1973 EXPORT_SYMBOL(ib_send_cm_rtu);
1974
1975 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
1976 {
1977         struct cm_rep_msg *rep_msg;
1978         struct ib_cm_rep_event_param *param;
1979
1980         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1981         param = &work->cm_event.param.rep_rcvd;
1982         param->remote_ca_guid = rep_msg->local_ca_guid;
1983         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1984         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
1985         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1986         param->responder_resources = rep_msg->initiator_depth;
1987         param->initiator_depth = rep_msg->resp_resources;
1988         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1989         param->failover_accepted = cm_rep_get_failover(rep_msg);
1990         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1991         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1992         param->srq = cm_rep_get_srq(rep_msg);
1993         work->cm_event.private_data = &rep_msg->private_data;
1994 }
1995
1996 static void cm_dup_rep_handler(struct cm_work *work)
1997 {
1998         struct cm_id_private *cm_id_priv;
1999         struct cm_rep_msg *rep_msg;
2000         struct ib_mad_send_buf *msg = NULL;
2001         int ret;
2002
2003         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2004         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
2005                                    rep_msg->local_comm_id);
2006         if (!cm_id_priv)
2007                 return;
2008
2009         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2010                         counter[CM_REP_COUNTER]);
2011         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2012         if (ret)
2013                 goto deref;
2014
2015         spin_lock_irq(&cm_id_priv->lock);
2016         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2017                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2018                               cm_id_priv->private_data,
2019                               cm_id_priv->private_data_len);
2020         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2021                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2022                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2023                               cm_id_priv->private_data,
2024                               cm_id_priv->private_data_len);
2025         else
2026                 goto unlock;
2027         spin_unlock_irq(&cm_id_priv->lock);
2028
2029         ret = ib_post_send_mad(msg, NULL);
2030         if (ret)
2031                 goto free;
2032         goto deref;
2033
2034 unlock: spin_unlock_irq(&cm_id_priv->lock);
2035 free:   cm_free_msg(msg);
2036 deref:  cm_deref_id(cm_id_priv);
2037 }
2038
2039 static int cm_rep_handler(struct cm_work *work)
2040 {
2041         struct cm_id_private *cm_id_priv;
2042         struct cm_rep_msg *rep_msg;
2043         int ret;
2044         struct cm_id_private *cur_cm_id_priv;
2045         struct ib_cm_id *cm_id;
2046         struct cm_timewait_info *timewait_info;
2047
2048         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2049         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
2050         if (!cm_id_priv) {
2051                 cm_dup_rep_handler(work);
2052                 return -EINVAL;
2053         }
2054
2055         cm_format_rep_event(work, cm_id_priv->qp_type);
2056
2057         spin_lock_irq(&cm_id_priv->lock);
2058         switch (cm_id_priv->id.state) {
2059         case IB_CM_REQ_SENT:
2060         case IB_CM_MRA_REQ_RCVD:
2061                 break;
2062         default:
2063                 spin_unlock_irq(&cm_id_priv->lock);
2064                 ret = -EINVAL;
2065                 goto error;
2066         }
2067
2068         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
2069         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
2070         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2071
2072         spin_lock(&cm.lock);
2073         /* Check for duplicate REP. */
2074         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2075                 spin_unlock(&cm.lock);
2076                 spin_unlock_irq(&cm_id_priv->lock);
2077                 ret = -EINVAL;
2078                 goto error;
2079         }
2080         /* Check for a stale connection. */
2081         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2082         if (timewait_info) {
2083                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
2084                          &cm.remote_id_table);
2085                 cm_id_priv->timewait_info->inserted_remote_id = 0;
2086                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
2087                                            timewait_info->work.remote_id);
2088
2089                 spin_unlock(&cm.lock);
2090                 spin_unlock_irq(&cm_id_priv->lock);
2091                 cm_issue_rej(work->port, work->mad_recv_wc,
2092                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2093                              NULL, 0);
2094                 ret = -EINVAL;
2095                 if (cur_cm_id_priv) {
2096                         cm_id = &cur_cm_id_priv->id;
2097                         ib_send_cm_dreq(cm_id, NULL, 0);
2098                         cm_deref_id(cur_cm_id_priv);
2099                 }
2100
2101                 goto error;
2102         }
2103         spin_unlock(&cm.lock);
2104
2105         cm_id_priv->id.state = IB_CM_REP_RCVD;
2106         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
2107         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2108         cm_id_priv->initiator_depth = rep_msg->resp_resources;
2109         cm_id_priv->responder_resources = rep_msg->initiator_depth;
2110         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
2111         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2112         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2113         cm_id_priv->av.timeout =
2114                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2115                                        cm_id_priv->av.timeout - 1);
2116         cm_id_priv->alt_av.timeout =
2117                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2118                                        cm_id_priv->alt_av.timeout - 1);
2119
2120         /* todo: handle peer_to_peer */
2121
2122         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2123         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2124         if (!ret)
2125                 list_add_tail(&work->list, &cm_id_priv->work_list);
2126         spin_unlock_irq(&cm_id_priv->lock);
2127
2128         if (ret)
2129                 cm_process_work(cm_id_priv, work);
2130         else
2131                 cm_deref_id(cm_id_priv);
2132         return 0;
2133
2134 error:
2135         cm_deref_id(cm_id_priv);
2136         return ret;
2137 }
2138
2139 static int cm_establish_handler(struct cm_work *work)
2140 {
2141         struct cm_id_private *cm_id_priv;
2142         int ret;
2143
2144         /* See comment in cm_establish about lookup. */
2145         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2146         if (!cm_id_priv)
2147                 return -EINVAL;
2148
2149         spin_lock_irq(&cm_id_priv->lock);
2150         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2151                 spin_unlock_irq(&cm_id_priv->lock);
2152                 goto out;
2153         }
2154
2155         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2156         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2157         if (!ret)
2158                 list_add_tail(&work->list, &cm_id_priv->work_list);
2159         spin_unlock_irq(&cm_id_priv->lock);
2160
2161         if (ret)
2162                 cm_process_work(cm_id_priv, work);
2163         else
2164                 cm_deref_id(cm_id_priv);
2165         return 0;
2166 out:
2167         cm_deref_id(cm_id_priv);
2168         return -EINVAL;
2169 }
2170
2171 static int cm_rtu_handler(struct cm_work *work)
2172 {
2173         struct cm_id_private *cm_id_priv;
2174         struct cm_rtu_msg *rtu_msg;
2175         int ret;
2176
2177         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2178         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
2179                                    rtu_msg->local_comm_id);
2180         if (!cm_id_priv)
2181                 return -EINVAL;
2182
2183         work->cm_event.private_data = &rtu_msg->private_data;
2184
2185         spin_lock_irq(&cm_id_priv->lock);
2186         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2187             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2188                 spin_unlock_irq(&cm_id_priv->lock);
2189                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2190                                 counter[CM_RTU_COUNTER]);
2191                 goto out;
2192         }
2193         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2194
2195         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2196         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2197         if (!ret)
2198                 list_add_tail(&work->list, &cm_id_priv->work_list);
2199         spin_unlock_irq(&cm_id_priv->lock);
2200
2201         if (ret)
2202                 cm_process_work(cm_id_priv, work);
2203         else
2204                 cm_deref_id(cm_id_priv);
2205         return 0;
2206 out:
2207         cm_deref_id(cm_id_priv);
2208         return -EINVAL;
2209 }
2210
2211 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2212                           struct cm_id_private *cm_id_priv,
2213                           const void *private_data,
2214                           u8 private_data_len)
2215 {
2216         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2217                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
2218         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
2219         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
2220         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2221
2222         if (private_data && private_data_len)
2223                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2224 }
2225
2226 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2227                     const void *private_data,
2228                     u8 private_data_len)
2229 {
2230         struct cm_id_private *cm_id_priv;
2231         struct ib_mad_send_buf *msg;
2232         unsigned long flags;
2233         int ret;
2234
2235         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2236                 return -EINVAL;
2237
2238         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2239         spin_lock_irqsave(&cm_id_priv->lock, flags);
2240         if (cm_id->state != IB_CM_ESTABLISHED) {
2241                 ret = -EINVAL;
2242                 goto out;
2243         }
2244
2245         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2246             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2247                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2248
2249         ret = cm_alloc_msg(cm_id_priv, &msg);
2250         if (ret) {
2251                 cm_enter_timewait(cm_id_priv);
2252                 goto out;
2253         }
2254
2255         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2256                        private_data, private_data_len);
2257         msg->timeout_ms = cm_id_priv->timeout_ms;
2258         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2259
2260         ret = ib_post_send_mad(msg, NULL);
2261         if (ret) {
2262                 cm_enter_timewait(cm_id_priv);
2263                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2264                 cm_free_msg(msg);
2265                 return ret;
2266         }
2267
2268         cm_id->state = IB_CM_DREQ_SENT;
2269         cm_id_priv->msg = msg;
2270 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2271         return ret;
2272 }
2273 EXPORT_SYMBOL(ib_send_cm_dreq);
2274
2275 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2276                           struct cm_id_private *cm_id_priv,
2277                           const void *private_data,
2278                           u8 private_data_len)
2279 {
2280         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2281         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2282         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2283
2284         if (private_data && private_data_len)
2285                 memcpy(drep_msg->private_data, private_data, private_data_len);
2286 }
2287
2288 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2289                     const void *private_data,
2290                     u8 private_data_len)
2291 {
2292         struct cm_id_private *cm_id_priv;
2293         struct ib_mad_send_buf *msg;
2294         unsigned long flags;
2295         void *data;
2296         int ret;
2297
2298         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2299                 return -EINVAL;
2300
2301         data = cm_copy_private_data(private_data, private_data_len);
2302         if (IS_ERR(data))
2303                 return PTR_ERR(data);
2304
2305         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2306         spin_lock_irqsave(&cm_id_priv->lock, flags);
2307         if (cm_id->state != IB_CM_DREQ_RCVD) {
2308                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2309                 kfree(data);
2310                 return -EINVAL;
2311         }
2312
2313         cm_set_private_data(cm_id_priv, data, private_data_len);
2314         cm_enter_timewait(cm_id_priv);
2315
2316         ret = cm_alloc_msg(cm_id_priv, &msg);
2317         if (ret)
2318                 goto out;
2319
2320         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2321                        private_data, private_data_len);
2322
2323         ret = ib_post_send_mad(msg, NULL);
2324         if (ret) {
2325                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2326                 cm_free_msg(msg);
2327                 return ret;
2328         }
2329
2330 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2331         return ret;
2332 }
2333 EXPORT_SYMBOL(ib_send_cm_drep);
2334
2335 static int cm_issue_drep(struct cm_port *port,
2336                          struct ib_mad_recv_wc *mad_recv_wc)
2337 {
2338         struct ib_mad_send_buf *msg = NULL;
2339         struct cm_dreq_msg *dreq_msg;
2340         struct cm_drep_msg *drep_msg;
2341         int ret;
2342
2343         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2344         if (ret)
2345                 return ret;
2346
2347         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2348         drep_msg = (struct cm_drep_msg *) msg->mad;
2349
2350         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2351         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2352         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2353
2354         ret = ib_post_send_mad(msg, NULL);
2355         if (ret)
2356                 cm_free_msg(msg);
2357
2358         return ret;
2359 }
2360
2361 static int cm_dreq_handler(struct cm_work *work)
2362 {
2363         struct cm_id_private *cm_id_priv;
2364         struct cm_dreq_msg *dreq_msg;
2365         struct ib_mad_send_buf *msg = NULL;
2366         int ret;
2367
2368         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2369         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2370                                    dreq_msg->local_comm_id);
2371         if (!cm_id_priv) {
2372                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2373                                 counter[CM_DREQ_COUNTER]);
2374                 cm_issue_drep(work->port, work->mad_recv_wc);
2375                 return -EINVAL;
2376         }
2377
2378         work->cm_event.private_data = &dreq_msg->private_data;
2379
2380         spin_lock_irq(&cm_id_priv->lock);
2381         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2382                 goto unlock;
2383
2384         switch (cm_id_priv->id.state) {
2385         case IB_CM_REP_SENT:
2386         case IB_CM_DREQ_SENT:
2387                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2388                 break;
2389         case IB_CM_ESTABLISHED:
2390                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2391                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2392                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2393                 break;
2394         case IB_CM_MRA_REP_RCVD:
2395                 break;
2396         case IB_CM_TIMEWAIT:
2397                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2398                                 counter[CM_DREQ_COUNTER]);
2399                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2400                         goto unlock;
2401
2402                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2403                                cm_id_priv->private_data,
2404                                cm_id_priv->private_data_len);
2405                 spin_unlock_irq(&cm_id_priv->lock);
2406
2407                 if (ib_post_send_mad(msg, NULL))
2408                         cm_free_msg(msg);
2409                 goto deref;
2410         case IB_CM_DREQ_RCVD:
2411                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2412                                 counter[CM_DREQ_COUNTER]);
2413                 goto unlock;
2414         default:
2415                 goto unlock;
2416         }
2417         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2418         cm_id_priv->tid = dreq_msg->hdr.tid;
2419         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2420         if (!ret)
2421                 list_add_tail(&work->list, &cm_id_priv->work_list);
2422         spin_unlock_irq(&cm_id_priv->lock);
2423
2424         if (ret)
2425                 cm_process_work(cm_id_priv, work);
2426         else
2427                 cm_deref_id(cm_id_priv);
2428         return 0;
2429
2430 unlock: spin_unlock_irq(&cm_id_priv->lock);
2431 deref:  cm_deref_id(cm_id_priv);
2432         return -EINVAL;
2433 }
2434
2435 static int cm_drep_handler(struct cm_work *work)
2436 {
2437         struct cm_id_private *cm_id_priv;
2438         struct cm_drep_msg *drep_msg;
2439         int ret;
2440
2441         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2442         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2443                                    drep_msg->local_comm_id);
2444         if (!cm_id_priv)
2445                 return -EINVAL;
2446
2447         work->cm_event.private_data = &drep_msg->private_data;
2448
2449         spin_lock_irq(&cm_id_priv->lock);
2450         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2451             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2452                 spin_unlock_irq(&cm_id_priv->lock);
2453                 goto out;
2454         }
2455         cm_enter_timewait(cm_id_priv);
2456
2457         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2458         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2459         if (!ret)
2460                 list_add_tail(&work->list, &cm_id_priv->work_list);
2461         spin_unlock_irq(&cm_id_priv->lock);
2462
2463         if (ret)
2464                 cm_process_work(cm_id_priv, work);
2465         else
2466                 cm_deref_id(cm_id_priv);
2467         return 0;
2468 out:
2469         cm_deref_id(cm_id_priv);
2470         return -EINVAL;
2471 }
2472
2473 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2474                    enum ib_cm_rej_reason reason,
2475                    void *ari,
2476                    u8 ari_length,
2477                    const void *private_data,
2478                    u8 private_data_len)
2479 {
2480         struct cm_id_private *cm_id_priv;
2481         struct ib_mad_send_buf *msg;
2482         unsigned long flags;
2483         int ret;
2484
2485         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2486             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2487                 return -EINVAL;
2488
2489         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2490
2491         spin_lock_irqsave(&cm_id_priv->lock, flags);
2492         switch (cm_id->state) {
2493         case IB_CM_REQ_SENT:
2494         case IB_CM_MRA_REQ_RCVD:
2495         case IB_CM_REQ_RCVD:
2496         case IB_CM_MRA_REQ_SENT:
2497         case IB_CM_REP_RCVD:
2498         case IB_CM_MRA_REP_SENT:
2499                 ret = cm_alloc_msg(cm_id_priv, &msg);
2500                 if (!ret)
2501                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2502                                       cm_id_priv, reason, ari, ari_length,
2503                                       private_data, private_data_len);
2504
2505                 cm_reset_to_idle(cm_id_priv);
2506                 break;
2507         case IB_CM_REP_SENT:
2508         case IB_CM_MRA_REP_RCVD:
2509                 ret = cm_alloc_msg(cm_id_priv, &msg);
2510                 if (!ret)
2511                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2512                                       cm_id_priv, reason, ari, ari_length,
2513                                       private_data, private_data_len);
2514
2515                 cm_enter_timewait(cm_id_priv);
2516                 break;
2517         default:
2518                 ret = -EINVAL;
2519                 goto out;
2520         }
2521
2522         if (ret)
2523                 goto out;
2524
2525         ret = ib_post_send_mad(msg, NULL);
2526         if (ret)
2527                 cm_free_msg(msg);
2528
2529 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2530         return ret;
2531 }
2532 EXPORT_SYMBOL(ib_send_cm_rej);
2533
2534 static void cm_format_rej_event(struct cm_work *work)
2535 {
2536         struct cm_rej_msg *rej_msg;
2537         struct ib_cm_rej_event_param *param;
2538
2539         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2540         param = &work->cm_event.param.rej_rcvd;
2541         param->ari = rej_msg->ari;
2542         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2543         param->reason = __be16_to_cpu(rej_msg->reason);
2544         work->cm_event.private_data = &rej_msg->private_data;
2545 }
2546
2547 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2548 {
2549         struct cm_timewait_info *timewait_info;
2550         struct cm_id_private *cm_id_priv;
2551         __be32 remote_id;
2552
2553         remote_id = rej_msg->local_comm_id;
2554
2555         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2556                 spin_lock_irq(&cm.lock);
2557                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2558                                                   remote_id);
2559                 if (!timewait_info) {
2560                         spin_unlock_irq(&cm.lock);
2561                         return NULL;
2562                 }
2563                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2564                                       (timewait_info->work.local_id ^
2565                                        cm.random_id_operand));
2566                 if (cm_id_priv) {
2567                         if (cm_id_priv->id.remote_id == remote_id)
2568                                 atomic_inc(&cm_id_priv->refcount);
2569                         else
2570                                 cm_id_priv = NULL;
2571                 }
2572                 spin_unlock_irq(&cm.lock);
2573         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2574                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2575         else
2576                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2577
2578         return cm_id_priv;
2579 }
2580
2581 static int cm_rej_handler(struct cm_work *work)
2582 {
2583         struct cm_id_private *cm_id_priv;
2584         struct cm_rej_msg *rej_msg;
2585         int ret;
2586
2587         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2588         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2589         if (!cm_id_priv)
2590                 return -EINVAL;
2591
2592         cm_format_rej_event(work);
2593
2594         spin_lock_irq(&cm_id_priv->lock);
2595         switch (cm_id_priv->id.state) {
2596         case IB_CM_REQ_SENT:
2597         case IB_CM_MRA_REQ_RCVD:
2598         case IB_CM_REP_SENT:
2599         case IB_CM_MRA_REP_RCVD:
2600                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2601                 /* fall through */
2602         case IB_CM_REQ_RCVD:
2603         case IB_CM_MRA_REQ_SENT:
2604                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2605                         cm_enter_timewait(cm_id_priv);
2606                 else
2607                         cm_reset_to_idle(cm_id_priv);
2608                 break;
2609         case IB_CM_DREQ_SENT:
2610                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2611                 /* fall through */
2612         case IB_CM_REP_RCVD:
2613         case IB_CM_MRA_REP_SENT:
2614                 cm_enter_timewait(cm_id_priv);
2615                 break;
2616         case IB_CM_ESTABLISHED:
2617                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2618                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2619                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2620                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2621                                               cm_id_priv->msg);
2622                         cm_enter_timewait(cm_id_priv);
2623                         break;
2624                 }
2625                 /* fall through */
2626         default:
2627                 spin_unlock_irq(&cm_id_priv->lock);
2628                 ret = -EINVAL;
2629                 goto out;
2630         }
2631
2632         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2633         if (!ret)
2634                 list_add_tail(&work->list, &cm_id_priv->work_list);
2635         spin_unlock_irq(&cm_id_priv->lock);
2636
2637         if (ret)
2638                 cm_process_work(cm_id_priv, work);
2639         else
2640                 cm_deref_id(cm_id_priv);
2641         return 0;
2642 out:
2643         cm_deref_id(cm_id_priv);
2644         return -EINVAL;
2645 }
2646
2647 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2648                    u8 service_timeout,
2649                    const void *private_data,
2650                    u8 private_data_len)
2651 {
2652         struct cm_id_private *cm_id_priv;
2653         struct ib_mad_send_buf *msg;
2654         enum ib_cm_state cm_state;
2655         enum ib_cm_lap_state lap_state;
2656         enum cm_msg_response msg_response;
2657         void *data;
2658         unsigned long flags;
2659         int ret;
2660
2661         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2662                 return -EINVAL;
2663
2664         data = cm_copy_private_data(private_data, private_data_len);
2665         if (IS_ERR(data))
2666                 return PTR_ERR(data);
2667
2668         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2669
2670         spin_lock_irqsave(&cm_id_priv->lock, flags);
2671         switch(cm_id_priv->id.state) {
2672         case IB_CM_REQ_RCVD:
2673                 cm_state = IB_CM_MRA_REQ_SENT;
2674                 lap_state = cm_id->lap_state;
2675                 msg_response = CM_MSG_RESPONSE_REQ;
2676                 break;
2677         case IB_CM_REP_RCVD:
2678                 cm_state = IB_CM_MRA_REP_SENT;
2679                 lap_state = cm_id->lap_state;
2680                 msg_response = CM_MSG_RESPONSE_REP;
2681                 break;
2682         case IB_CM_ESTABLISHED:
2683                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2684                         cm_state = cm_id->state;
2685                         lap_state = IB_CM_MRA_LAP_SENT;
2686                         msg_response = CM_MSG_RESPONSE_OTHER;
2687                         break;
2688                 }
2689         default:
2690                 ret = -EINVAL;
2691                 goto error1;
2692         }
2693
2694         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2695                 ret = cm_alloc_msg(cm_id_priv, &msg);
2696                 if (ret)
2697                         goto error1;
2698
2699                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2700                               msg_response, service_timeout,
2701                               private_data, private_data_len);
2702                 ret = ib_post_send_mad(msg, NULL);
2703                 if (ret)
2704                         goto error2;
2705         }
2706
2707         cm_id->state = cm_state;
2708         cm_id->lap_state = lap_state;
2709         cm_id_priv->service_timeout = service_timeout;
2710         cm_set_private_data(cm_id_priv, data, private_data_len);
2711         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2712         return 0;
2713
2714 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2715         kfree(data);
2716         return ret;
2717
2718 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2719         kfree(data);
2720         cm_free_msg(msg);
2721         return ret;
2722 }
2723 EXPORT_SYMBOL(ib_send_cm_mra);
2724
2725 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2726 {
2727         switch (cm_mra_get_msg_mraed(mra_msg)) {
2728         case CM_MSG_RESPONSE_REQ:
2729                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2730         case CM_MSG_RESPONSE_REP:
2731         case CM_MSG_RESPONSE_OTHER:
2732                 return cm_acquire_id(mra_msg->remote_comm_id,
2733                                      mra_msg->local_comm_id);
2734         default:
2735                 return NULL;
2736         }
2737 }
2738
2739 static int cm_mra_handler(struct cm_work *work)
2740 {
2741         struct cm_id_private *cm_id_priv;
2742         struct cm_mra_msg *mra_msg;
2743         int timeout, ret;
2744
2745         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2746         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2747         if (!cm_id_priv)
2748                 return -EINVAL;
2749
2750         work->cm_event.private_data = &mra_msg->private_data;
2751         work->cm_event.param.mra_rcvd.service_timeout =
2752                                         cm_mra_get_service_timeout(mra_msg);
2753         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2754                   cm_convert_to_ms(cm_id_priv->av.timeout);
2755
2756         spin_lock_irq(&cm_id_priv->lock);
2757         switch (cm_id_priv->id.state) {
2758         case IB_CM_REQ_SENT:
2759                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2760                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2761                                   cm_id_priv->msg, timeout))
2762                         goto out;
2763                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2764                 break;
2765         case IB_CM_REP_SENT:
2766                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2767                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2768                                   cm_id_priv->msg, timeout))
2769                         goto out;
2770                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2771                 break;
2772         case IB_CM_ESTABLISHED:
2773                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2774                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2775                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2776                                   cm_id_priv->msg, timeout)) {
2777                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2778                                 atomic_long_inc(&work->port->
2779                                                 counter_group[CM_RECV_DUPLICATES].
2780                                                 counter[CM_MRA_COUNTER]);
2781                         goto out;
2782                 }
2783                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2784                 break;
2785         case IB_CM_MRA_REQ_RCVD:
2786         case IB_CM_MRA_REP_RCVD:
2787                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2788                                 counter[CM_MRA_COUNTER]);
2789                 /* fall through */
2790         default:
2791                 goto out;
2792         }
2793
2794         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2795                                       cm_id_priv->id.state;
2796         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2797         if (!ret)
2798                 list_add_tail(&work->list, &cm_id_priv->work_list);
2799         spin_unlock_irq(&cm_id_priv->lock);
2800
2801         if (ret)
2802                 cm_process_work(cm_id_priv, work);
2803         else
2804                 cm_deref_id(cm_id_priv);
2805         return 0;
2806 out:
2807         spin_unlock_irq(&cm_id_priv->lock);
2808         cm_deref_id(cm_id_priv);
2809         return -EINVAL;
2810 }
2811
2812 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2813                           struct cm_id_private *cm_id_priv,
2814                           struct ib_sa_path_rec *alternate_path,
2815                           const void *private_data,
2816                           u8 private_data_len)
2817 {
2818         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2819                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2820         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2821         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2822         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2823         /* todo: need remote CM response timeout */
2824         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2825         lap_msg->alt_local_lid = alternate_path->slid;
2826         lap_msg->alt_remote_lid = alternate_path->dlid;
2827         lap_msg->alt_local_gid = alternate_path->sgid;
2828         lap_msg->alt_remote_gid = alternate_path->dgid;
2829         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2830         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2831         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2832         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2833         cm_lap_set_sl(lap_msg, alternate_path->sl);
2834         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2835         cm_lap_set_local_ack_timeout(lap_msg,
2836                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2837                                alternate_path->packet_life_time));
2838
2839         if (private_data && private_data_len)
2840                 memcpy(lap_msg->private_data, private_data, private_data_len);
2841 }
2842
2843 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2844                    struct ib_sa_path_rec *alternate_path,
2845                    const void *private_data,
2846                    u8 private_data_len)
2847 {
2848         struct cm_id_private *cm_id_priv;
2849         struct ib_mad_send_buf *msg;
2850         unsigned long flags;
2851         int ret;
2852
2853         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2854                 return -EINVAL;
2855
2856         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2857         spin_lock_irqsave(&cm_id_priv->lock, flags);
2858         if (cm_id->state != IB_CM_ESTABLISHED ||
2859             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2860              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2861                 ret = -EINVAL;
2862                 goto out;
2863         }
2864
2865         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av,
2866                                  cm_id_priv);
2867         if (ret)
2868                 goto out;
2869         cm_id_priv->alt_av.timeout =
2870                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2871                                        cm_id_priv->alt_av.timeout - 1);
2872
2873         ret = cm_alloc_msg(cm_id_priv, &msg);
2874         if (ret)
2875                 goto out;
2876
2877         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2878                       alternate_path, private_data, private_data_len);
2879         msg->timeout_ms = cm_id_priv->timeout_ms;
2880         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2881
2882         ret = ib_post_send_mad(msg, NULL);
2883         if (ret) {
2884                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2885                 cm_free_msg(msg);
2886                 return ret;
2887         }
2888
2889         cm_id->lap_state = IB_CM_LAP_SENT;
2890         cm_id_priv->msg = msg;
2891
2892 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2893         return ret;
2894 }
2895 EXPORT_SYMBOL(ib_send_cm_lap);
2896
2897 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2898                                     struct ib_sa_path_rec *path,
2899                                     struct cm_lap_msg *lap_msg)
2900 {
2901         memset(path, 0, sizeof *path);
2902         path->dgid = lap_msg->alt_local_gid;
2903         path->sgid = lap_msg->alt_remote_gid;
2904         path->dlid = lap_msg->alt_local_lid;
2905         path->slid = lap_msg->alt_remote_lid;
2906         path->flow_label = cm_lap_get_flow_label(lap_msg);
2907         path->hop_limit = lap_msg->alt_hop_limit;
2908         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2909         path->reversible = 1;
2910         path->pkey = cm_id_priv->pkey;
2911         path->sl = cm_lap_get_sl(lap_msg);
2912         path->mtu_selector = IB_SA_EQ;
2913         path->mtu = cm_id_priv->path_mtu;
2914         path->rate_selector = IB_SA_EQ;
2915         path->rate = cm_lap_get_packet_rate(lap_msg);
2916         path->packet_life_time_selector = IB_SA_EQ;
2917         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2918         path->packet_life_time -= (path->packet_life_time > 0);
2919 }
2920
2921 static int cm_lap_handler(struct cm_work *work)
2922 {
2923         struct cm_id_private *cm_id_priv;
2924         struct cm_lap_msg *lap_msg;
2925         struct ib_cm_lap_event_param *param;
2926         struct ib_mad_send_buf *msg = NULL;
2927         int ret;
2928
2929         /* todo: verify LAP request and send reject APR if invalid. */
2930         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2931         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2932                                    lap_msg->local_comm_id);
2933         if (!cm_id_priv)
2934                 return -EINVAL;
2935
2936         param = &work->cm_event.param.lap_rcvd;
2937         param->alternate_path = &work->path[0];
2938         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2939         work->cm_event.private_data = &lap_msg->private_data;
2940
2941         spin_lock_irq(&cm_id_priv->lock);
2942         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2943                 goto unlock;
2944
2945         switch (cm_id_priv->id.lap_state) {
2946         case IB_CM_LAP_UNINIT:
2947         case IB_CM_LAP_IDLE:
2948                 break;
2949         case IB_CM_MRA_LAP_SENT:
2950                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2951                                 counter[CM_LAP_COUNTER]);
2952                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2953                         goto unlock;
2954
2955                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2956                               CM_MSG_RESPONSE_OTHER,
2957                               cm_id_priv->service_timeout,
2958                               cm_id_priv->private_data,
2959                               cm_id_priv->private_data_len);
2960                 spin_unlock_irq(&cm_id_priv->lock);
2961
2962                 if (ib_post_send_mad(msg, NULL))
2963                         cm_free_msg(msg);
2964                 goto deref;
2965         case IB_CM_LAP_RCVD:
2966                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2967                                 counter[CM_LAP_COUNTER]);
2968                 goto unlock;
2969         default:
2970                 goto unlock;
2971         }
2972
2973         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2974         cm_id_priv->tid = lap_msg->hdr.tid;
2975         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2976                                 work->mad_recv_wc->recv_buf.grh,
2977                                 &cm_id_priv->av);
2978         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av,
2979                            cm_id_priv);
2980         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2981         if (!ret)
2982                 list_add_tail(&work->list, &cm_id_priv->work_list);
2983         spin_unlock_irq(&cm_id_priv->lock);
2984
2985         if (ret)
2986                 cm_process_work(cm_id_priv, work);
2987         else
2988                 cm_deref_id(cm_id_priv);
2989         return 0;
2990
2991 unlock: spin_unlock_irq(&cm_id_priv->lock);
2992 deref:  cm_deref_id(cm_id_priv);
2993         return -EINVAL;
2994 }
2995
2996 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2997                           struct cm_id_private *cm_id_priv,
2998                           enum ib_cm_apr_status status,
2999                           void *info,
3000                           u8 info_length,
3001                           const void *private_data,
3002                           u8 private_data_len)
3003 {
3004         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
3005         apr_msg->local_comm_id = cm_id_priv->id.local_id;
3006         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
3007         apr_msg->ap_status = (u8) status;
3008
3009         if (info && info_length) {
3010                 apr_msg->info_length = info_length;
3011                 memcpy(apr_msg->info, info, info_length);
3012         }
3013
3014         if (private_data && private_data_len)
3015                 memcpy(apr_msg->private_data, private_data, private_data_len);
3016 }
3017
3018 int ib_send_cm_apr(struct ib_cm_id *cm_id,
3019                    enum ib_cm_apr_status status,
3020                    void *info,
3021                    u8 info_length,
3022                    const void *private_data,
3023                    u8 private_data_len)
3024 {
3025         struct cm_id_private *cm_id_priv;
3026         struct ib_mad_send_buf *msg;
3027         unsigned long flags;
3028         int ret;
3029
3030         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
3031             (info && info_length > IB_CM_APR_INFO_LENGTH))
3032                 return -EINVAL;
3033
3034         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3035         spin_lock_irqsave(&cm_id_priv->lock, flags);
3036         if (cm_id->state != IB_CM_ESTABLISHED ||
3037             (cm_id->lap_state != IB_CM_LAP_RCVD &&
3038              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
3039                 ret = -EINVAL;
3040                 goto out;
3041         }
3042
3043         ret = cm_alloc_msg(cm_id_priv, &msg);
3044         if (ret)
3045                 goto out;
3046
3047         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
3048                       info, info_length, private_data, private_data_len);
3049         ret = ib_post_send_mad(msg, NULL);
3050         if (ret) {
3051                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3052                 cm_free_msg(msg);
3053                 return ret;
3054         }
3055
3056         cm_id->lap_state = IB_CM_LAP_IDLE;
3057 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3058         return ret;
3059 }
3060 EXPORT_SYMBOL(ib_send_cm_apr);
3061
3062 static int cm_apr_handler(struct cm_work *work)
3063 {
3064         struct cm_id_private *cm_id_priv;
3065         struct cm_apr_msg *apr_msg;
3066         int ret;
3067
3068         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3069         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
3070                                    apr_msg->local_comm_id);
3071         if (!cm_id_priv)
3072                 return -EINVAL; /* Unmatched reply. */
3073
3074         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
3075         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
3076         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
3077         work->cm_event.private_data = &apr_msg->private_data;
3078
3079         spin_lock_irq(&cm_id_priv->lock);
3080         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3081             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3082              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3083                 spin_unlock_irq(&cm_id_priv->lock);
3084                 goto out;
3085         }
3086         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3087         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3088         cm_id_priv->msg = NULL;
3089
3090         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3091         if (!ret)
3092                 list_add_tail(&work->list, &cm_id_priv->work_list);
3093         spin_unlock_irq(&cm_id_priv->lock);
3094
3095         if (ret)
3096                 cm_process_work(cm_id_priv, work);
3097         else
3098                 cm_deref_id(cm_id_priv);
3099         return 0;
3100 out:
3101         cm_deref_id(cm_id_priv);
3102         return -EINVAL;
3103 }
3104
3105 static int cm_timewait_handler(struct cm_work *work)
3106 {
3107         struct cm_timewait_info *timewait_info;
3108         struct cm_id_private *cm_id_priv;
3109         int ret;
3110
3111         timewait_info = (struct cm_timewait_info *)work;
3112         spin_lock_irq(&cm.lock);
3113         list_del(&timewait_info->list);
3114         spin_unlock_irq(&cm.lock);
3115
3116         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3117                                    timewait_info->work.remote_id);
3118         if (!cm_id_priv)
3119                 return -EINVAL;
3120
3121         spin_lock_irq(&cm_id_priv->lock);
3122         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3123             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3124                 spin_unlock_irq(&cm_id_priv->lock);
3125                 goto out;
3126         }
3127         cm_id_priv->id.state = IB_CM_IDLE;
3128         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3129         if (!ret)
3130                 list_add_tail(&work->list, &cm_id_priv->work_list);
3131         spin_unlock_irq(&cm_id_priv->lock);
3132
3133         if (ret)
3134                 cm_process_work(cm_id_priv, work);
3135         else
3136                 cm_deref_id(cm_id_priv);
3137         return 0;
3138 out:
3139         cm_deref_id(cm_id_priv);
3140         return -EINVAL;
3141 }
3142
3143 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3144                                struct cm_id_private *cm_id_priv,
3145                                struct ib_cm_sidr_req_param *param)
3146 {
3147         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3148                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
3149         sidr_req_msg->request_id = cm_id_priv->id.local_id;
3150         sidr_req_msg->pkey = param->path->pkey;
3151         sidr_req_msg->service_id = param->service_id;
3152
3153         if (param->private_data && param->private_data_len)
3154                 memcpy(sidr_req_msg->private_data, param->private_data,
3155                        param->private_data_len);
3156 }
3157
3158 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3159                         struct ib_cm_sidr_req_param *param)
3160 {
3161         struct cm_id_private *cm_id_priv;
3162         struct ib_mad_send_buf *msg;
3163         unsigned long flags;
3164         int ret;
3165
3166         if (!param->path || (param->private_data &&
3167              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3168                 return -EINVAL;
3169
3170         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3171         ret = cm_init_av_by_path(param->path, &cm_id_priv->av, cm_id_priv);
3172         if (ret)
3173                 goto out;
3174
3175         cm_id->service_id = param->service_id;
3176         cm_id->service_mask = ~cpu_to_be64(0);
3177         cm_id_priv->timeout_ms = param->timeout_ms;
3178         cm_id_priv->max_cm_retries = param->max_cm_retries;
3179         ret = cm_alloc_msg(cm_id_priv, &msg);
3180         if (ret)
3181                 goto out;
3182
3183         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3184                            param);
3185         msg->timeout_ms = cm_id_priv->timeout_ms;
3186         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3187
3188         spin_lock_irqsave(&cm_id_priv->lock, flags);
3189         if (cm_id->state == IB_CM_IDLE)
3190                 ret = ib_post_send_mad(msg, NULL);
3191         else
3192                 ret = -EINVAL;
3193
3194         if (ret) {
3195                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3196                 cm_free_msg(msg);
3197                 goto out;
3198         }
3199         cm_id->state = IB_CM_SIDR_REQ_SENT;
3200         cm_id_priv->msg = msg;
3201         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3202 out:
3203         return ret;
3204 }
3205 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3206
3207 static void cm_format_sidr_req_event(struct cm_work *work,
3208                                      struct ib_cm_id *listen_id)
3209 {
3210         struct cm_sidr_req_msg *sidr_req_msg;
3211         struct ib_cm_sidr_req_event_param *param;
3212
3213         sidr_req_msg = (struct cm_sidr_req_msg *)
3214                                 work->mad_recv_wc->recv_buf.mad;
3215         param = &work->cm_event.param.sidr_req_rcvd;
3216         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
3217         param->listen_id = listen_id;
3218         param->service_id = sidr_req_msg->service_id;
3219         param->bth_pkey = cm_get_bth_pkey(work);
3220         param->port = work->port->port_num;
3221         work->cm_event.private_data = &sidr_req_msg->private_data;
3222 }
3223
3224 static int cm_sidr_req_handler(struct cm_work *work)
3225 {
3226         struct ib_cm_id *cm_id;
3227         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3228         struct cm_sidr_req_msg *sidr_req_msg;
3229         struct ib_wc *wc;
3230
3231         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3232         if (IS_ERR(cm_id))
3233                 return PTR_ERR(cm_id);
3234         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3235
3236         /* Record SGID/SLID and request ID for lookup. */
3237         sidr_req_msg = (struct cm_sidr_req_msg *)
3238                                 work->mad_recv_wc->recv_buf.mad;
3239         wc = work->mad_recv_wc->wc;
3240         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3241         cm_id_priv->av.dgid.global.interface_id = 0;
3242         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3243                                 work->mad_recv_wc->recv_buf.grh,
3244                                 &cm_id_priv->av);
3245         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3246         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3247         atomic_inc(&cm_id_priv->work_count);
3248
3249         spin_lock_irq(&cm.lock);
3250         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3251         if (cur_cm_id_priv) {
3252                 spin_unlock_irq(&cm.lock);
3253                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3254                                 counter[CM_SIDR_REQ_COUNTER]);
3255                 goto out; /* Duplicate message. */
3256         }
3257         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3258         cur_cm_id_priv = cm_find_listen(cm_id->device,
3259                                         sidr_req_msg->service_id);
3260         if (!cur_cm_id_priv) {
3261                 spin_unlock_irq(&cm.lock);
3262                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3263                 goto out; /* No match. */
3264         }
3265         atomic_inc(&cur_cm_id_priv->refcount);
3266         atomic_inc(&cm_id_priv->refcount);
3267         spin_unlock_irq(&cm.lock);
3268
3269         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3270         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3271         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3272         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3273
3274         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3275         cm_process_work(cm_id_priv, work);
3276         cm_deref_id(cur_cm_id_priv);
3277         return 0;
3278 out:
3279         ib_destroy_cm_id(&cm_id_priv->id);
3280         return -EINVAL;
3281 }
3282
3283 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3284                                struct cm_id_private *cm_id_priv,
3285                                struct ib_cm_sidr_rep_param *param)
3286 {
3287         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3288                           cm_id_priv->tid);
3289         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3290         sidr_rep_msg->status = param->status;
3291         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3292         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3293         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3294
3295         if (param->info && param->info_length)
3296                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3297
3298         if (param->private_data && param->private_data_len)
3299                 memcpy(sidr_rep_msg->private_data, param->private_data,
3300                        param->private_data_len);
3301 }
3302
3303 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3304                         struct ib_cm_sidr_rep_param *param)
3305 {
3306         struct cm_id_private *cm_id_priv;
3307         struct ib_mad_send_buf *msg;
3308         unsigned long flags;
3309         int ret;
3310
3311         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3312             (param->private_data &&
3313              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3314                 return -EINVAL;
3315
3316         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3317         spin_lock_irqsave(&cm_id_priv->lock, flags);
3318         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3319                 ret = -EINVAL;
3320                 goto error;
3321         }
3322
3323         ret = cm_alloc_msg(cm_id_priv, &msg);
3324         if (ret)
3325                 goto error;
3326
3327         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3328                            param);
3329         ret = ib_post_send_mad(msg, NULL);
3330         if (ret) {
3331                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3332                 cm_free_msg(msg);
3333                 return ret;
3334         }
3335         cm_id->state = IB_CM_IDLE;
3336         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3337
3338         spin_lock_irqsave(&cm.lock, flags);
3339         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3340                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3341                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3342         }
3343         spin_unlock_irqrestore(&cm.lock, flags);
3344         return 0;
3345
3346 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3347         return ret;
3348 }
3349 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3350
3351 static void cm_format_sidr_rep_event(struct cm_work *work)
3352 {
3353         struct cm_sidr_rep_msg *sidr_rep_msg;
3354         struct ib_cm_sidr_rep_event_param *param;
3355
3356         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3357                                 work->mad_recv_wc->recv_buf.mad;
3358         param = &work->cm_event.param.sidr_rep_rcvd;
3359         param->status = sidr_rep_msg->status;
3360         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3361         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3362         param->info = &sidr_rep_msg->info;
3363         param->info_len = sidr_rep_msg->info_length;
3364         work->cm_event.private_data = &sidr_rep_msg->private_data;
3365 }
3366
3367 static int cm_sidr_rep_handler(struct cm_work *work)
3368 {
3369         struct cm_sidr_rep_msg *sidr_rep_msg;
3370         struct cm_id_private *cm_id_priv;
3371
3372         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3373                                 work->mad_recv_wc->recv_buf.mad;
3374         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3375         if (!cm_id_priv)
3376                 return -EINVAL; /* Unmatched reply. */
3377
3378         spin_lock_irq(&cm_id_priv->lock);
3379         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3380                 spin_unlock_irq(&cm_id_priv->lock);
3381                 goto out;
3382         }
3383         cm_id_priv->id.state = IB_CM_IDLE;
3384         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3385         spin_unlock_irq(&cm_id_priv->lock);
3386
3387         cm_format_sidr_rep_event(work);
3388         cm_process_work(cm_id_priv, work);
3389         return 0;
3390 out:
3391         cm_deref_id(cm_id_priv);
3392         return -EINVAL;
3393 }
3394
3395 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3396                                   enum ib_wc_status wc_status)
3397 {
3398         struct cm_id_private *cm_id_priv;
3399         struct ib_cm_event cm_event;
3400         enum ib_cm_state state;
3401         int ret;
3402
3403         memset(&cm_event, 0, sizeof cm_event);
3404         cm_id_priv = msg->context[0];
3405
3406         /* Discard old sends or ones without a response. */
3407         spin_lock_irq(&cm_id_priv->lock);
3408         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3409         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3410                 goto discard;
3411
3412         pr_debug_ratelimited("CM: failed sending MAD in state %d. (%s)\n",
3413                              state, ib_wc_status_msg(wc_status));
3414         switch (state) {
3415         case IB_CM_REQ_SENT:
3416         case IB_CM_MRA_REQ_RCVD:
3417                 cm_reset_to_idle(cm_id_priv);
3418                 cm_event.event = IB_CM_REQ_ERROR;
3419                 break;
3420         case IB_CM_REP_SENT:
3421         case IB_CM_MRA_REP_RCVD:
3422                 cm_reset_to_idle(cm_id_priv);
3423                 cm_event.event = IB_CM_REP_ERROR;
3424                 break;
3425         case IB_CM_DREQ_SENT:
3426                 cm_enter_timewait(cm_id_priv);
3427                 cm_event.event = IB_CM_DREQ_ERROR;
3428                 break;
3429         case IB_CM_SIDR_REQ_SENT:
3430                 cm_id_priv->id.state = IB_CM_IDLE;
3431                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3432                 break;
3433         default:
3434                 goto discard;
3435         }
3436         spin_unlock_irq(&cm_id_priv->lock);
3437         cm_event.param.send_status = wc_status;
3438
3439         /* No other events can occur on the cm_id at this point. */
3440         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3441         cm_free_msg(msg);
3442         if (ret)
3443                 ib_destroy_cm_id(&cm_id_priv->id);
3444         return;
3445 discard:
3446         spin_unlock_irq(&cm_id_priv->lock);
3447         cm_free_msg(msg);
3448 }
3449
3450 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3451                             struct ib_mad_send_wc *mad_send_wc)
3452 {
3453         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3454         struct cm_port *port;
3455         u16 attr_index;
3456
3457         port = mad_agent->context;
3458         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3459                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3460
3461         /*
3462          * If the send was in response to a received message (context[0] is not
3463          * set to a cm_id), and is not a REJ, then it is a send that was
3464          * manually retried.
3465          */
3466         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3467                 msg->retries = 1;
3468
3469         atomic_long_add(1 + msg->retries,
3470                         &port->counter_group[CM_XMIT].counter[attr_index]);
3471         if (msg->retries)
3472                 atomic_long_add(msg->retries,
3473                                 &port->counter_group[CM_XMIT_RETRIES].
3474                                 counter[attr_index]);
3475
3476         switch (mad_send_wc->status) {
3477         case IB_WC_SUCCESS:
3478         case IB_WC_WR_FLUSH_ERR:
3479                 cm_free_msg(msg);
3480                 break;
3481         default:
3482                 if (msg->context[0] && msg->context[1])
3483                         cm_process_send_error(msg, mad_send_wc->status);
3484                 else
3485                         cm_free_msg(msg);
3486                 break;
3487         }
3488 }
3489
3490 static void cm_work_handler(struct work_struct *_work)
3491 {
3492         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3493         int ret;
3494
3495         switch (work->cm_event.event) {
3496         case IB_CM_REQ_RECEIVED:
3497                 ret = cm_req_handler(work);
3498                 break;
3499         case IB_CM_MRA_RECEIVED:
3500                 ret = cm_mra_handler(work);
3501                 break;
3502         case IB_CM_REJ_RECEIVED:
3503                 ret = cm_rej_handler(work);
3504                 break;
3505         case IB_CM_REP_RECEIVED:
3506                 ret = cm_rep_handler(work);
3507                 break;
3508         case IB_CM_RTU_RECEIVED:
3509                 ret = cm_rtu_handler(work);
3510                 break;
3511         case IB_CM_USER_ESTABLISHED:
3512                 ret = cm_establish_handler(work);
3513                 break;
3514         case IB_CM_DREQ_RECEIVED:
3515                 ret = cm_dreq_handler(work);
3516                 break;
3517         case IB_CM_DREP_RECEIVED:
3518                 ret = cm_drep_handler(work);
3519                 break;
3520         case IB_CM_SIDR_REQ_RECEIVED:
3521                 ret = cm_sidr_req_handler(work);
3522                 break;
3523         case IB_CM_SIDR_REP_RECEIVED:
3524                 ret = cm_sidr_rep_handler(work);
3525                 break;
3526         case IB_CM_LAP_RECEIVED:
3527                 ret = cm_lap_handler(work);
3528                 break;
3529         case IB_CM_APR_RECEIVED:
3530                 ret = cm_apr_handler(work);
3531                 break;
3532         case IB_CM_TIMEWAIT_EXIT:
3533                 ret = cm_timewait_handler(work);
3534                 break;
3535         default:
3536                 ret = -EINVAL;
3537                 break;
3538         }
3539         if (ret)
3540                 cm_free_work(work);
3541 }
3542
3543 static int cm_establish(struct ib_cm_id *cm_id)
3544 {
3545         struct cm_id_private *cm_id_priv;
3546         struct cm_work *work;
3547         unsigned long flags;
3548         int ret = 0;
3549         struct cm_device *cm_dev;
3550
3551         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3552         if (!cm_dev)
3553                 return -ENODEV;
3554
3555         work = kmalloc(sizeof *work, GFP_ATOMIC);
3556         if (!work)
3557                 return -ENOMEM;
3558
3559         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3560         spin_lock_irqsave(&cm_id_priv->lock, flags);
3561         switch (cm_id->state)
3562         {
3563         case IB_CM_REP_SENT:
3564         case IB_CM_MRA_REP_RCVD:
3565                 cm_id->state = IB_CM_ESTABLISHED;
3566                 break;
3567         case IB_CM_ESTABLISHED:
3568                 ret = -EISCONN;
3569                 break;
3570         default:
3571                 ret = -EINVAL;
3572                 break;
3573         }
3574         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3575
3576         if (ret) {
3577                 kfree(work);
3578                 goto out;
3579         }
3580
3581         /*
3582          * The CM worker thread may try to destroy the cm_id before it
3583          * can execute this work item.  To prevent potential deadlock,
3584          * we need to find the cm_id once we're in the context of the
3585          * worker thread, rather than holding a reference on it.
3586          */
3587         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3588         work->local_id = cm_id->local_id;
3589         work->remote_id = cm_id->remote_id;
3590         work->mad_recv_wc = NULL;
3591         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3592
3593         /* Check if the device started its remove_one */
3594         spin_lock_irqsave(&cm.lock, flags);
3595         if (!cm_dev->going_down) {
3596                 queue_delayed_work(cm.wq, &work->work, 0);
3597         } else {
3598                 kfree(work);
3599                 ret = -ENODEV;
3600         }
3601         spin_unlock_irqrestore(&cm.lock, flags);
3602
3603 out:
3604         return ret;
3605 }
3606
3607 static int cm_migrate(struct ib_cm_id *cm_id)
3608 {
3609         struct cm_id_private *cm_id_priv;
3610         struct cm_av tmp_av;
3611         unsigned long flags;
3612         int tmp_send_port_not_ready;
3613         int ret = 0;
3614
3615         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3616         spin_lock_irqsave(&cm_id_priv->lock, flags);
3617         if (cm_id->state == IB_CM_ESTABLISHED &&
3618             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3619              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3620                 cm_id->lap_state = IB_CM_LAP_IDLE;
3621                 /* Swap address vector */
3622                 tmp_av = cm_id_priv->av;
3623                 cm_id_priv->av = cm_id_priv->alt_av;
3624                 cm_id_priv->alt_av = tmp_av;
3625                 /* Swap port send ready state */
3626                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3627                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3628                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3629         } else
3630                 ret = -EINVAL;
3631         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3632
3633         return ret;
3634 }
3635
3636 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3637 {
3638         int ret;
3639
3640         switch (event) {
3641         case IB_EVENT_COMM_EST:
3642                 ret = cm_establish(cm_id);
3643                 break;
3644         case IB_EVENT_PATH_MIG:
3645                 ret = cm_migrate(cm_id);
3646                 break;
3647         default:
3648                 ret = -EINVAL;
3649         }
3650         return ret;
3651 }
3652 EXPORT_SYMBOL(ib_cm_notify);
3653
3654 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3655                             struct ib_mad_send_buf *send_buf,
3656                             struct ib_mad_recv_wc *mad_recv_wc)
3657 {
3658         struct cm_port *port = mad_agent->context;
3659         struct cm_work *work;
3660         enum ib_cm_event_type event;
3661         u16 attr_id;
3662         int paths = 0;
3663         int going_down = 0;
3664
3665         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3666         case CM_REQ_ATTR_ID:
3667                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3668                                                     alt_local_lid != 0);
3669                 event = IB_CM_REQ_RECEIVED;
3670                 break;
3671         case CM_MRA_ATTR_ID:
3672                 event = IB_CM_MRA_RECEIVED;
3673                 break;
3674         case CM_REJ_ATTR_ID:
3675                 event = IB_CM_REJ_RECEIVED;
3676                 break;
3677         case CM_REP_ATTR_ID:
3678                 event = IB_CM_REP_RECEIVED;
3679                 break;
3680         case CM_RTU_ATTR_ID:
3681                 event = IB_CM_RTU_RECEIVED;
3682                 break;
3683         case CM_DREQ_ATTR_ID:
3684                 event = IB_CM_DREQ_RECEIVED;
3685                 break;
3686         case CM_DREP_ATTR_ID:
3687                 event = IB_CM_DREP_RECEIVED;
3688                 break;
3689         case CM_SIDR_REQ_ATTR_ID:
3690                 event = IB_CM_SIDR_REQ_RECEIVED;
3691                 break;
3692         case CM_SIDR_REP_ATTR_ID:
3693                 event = IB_CM_SIDR_REP_RECEIVED;
3694                 break;
3695         case CM_LAP_ATTR_ID:
3696                 paths = 1;
3697                 event = IB_CM_LAP_RECEIVED;
3698                 break;
3699         case CM_APR_ATTR_ID:
3700                 event = IB_CM_APR_RECEIVED;
3701                 break;
3702         default:
3703                 ib_free_recv_mad(mad_recv_wc);
3704                 return;
3705         }
3706
3707         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3708         atomic_long_inc(&port->counter_group[CM_RECV].
3709                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3710
3711         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3712                        GFP_KERNEL);
3713         if (!work) {
3714                 ib_free_recv_mad(mad_recv_wc);
3715                 return;
3716         }
3717
3718         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3719         work->cm_event.event = event;
3720         work->mad_recv_wc = mad_recv_wc;
3721         work->port = port;
3722
3723         /* Check if the device started its remove_one */
3724         spin_lock_irq(&cm.lock);
3725         if (!port->cm_dev->going_down)
3726                 queue_delayed_work(cm.wq, &work->work, 0);
3727         else
3728                 going_down = 1;
3729         spin_unlock_irq(&cm.lock);
3730
3731         if (going_down) {
3732                 kfree(work);
3733                 ib_free_recv_mad(mad_recv_wc);
3734         }
3735 }
3736
3737 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3738                                 struct ib_qp_attr *qp_attr,
3739                                 int *qp_attr_mask)
3740 {
3741         unsigned long flags;
3742         int ret;
3743
3744         spin_lock_irqsave(&cm_id_priv->lock, flags);
3745         switch (cm_id_priv->id.state) {
3746         case IB_CM_REQ_SENT:
3747         case IB_CM_MRA_REQ_RCVD:
3748         case IB_CM_REQ_RCVD:
3749         case IB_CM_MRA_REQ_SENT:
3750         case IB_CM_REP_RCVD:
3751         case IB_CM_MRA_REP_SENT:
3752         case IB_CM_REP_SENT:
3753         case IB_CM_MRA_REP_RCVD:
3754         case IB_CM_ESTABLISHED:
3755                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3756                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3757                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3758                 if (cm_id_priv->responder_resources)
3759                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3760                                                     IB_ACCESS_REMOTE_ATOMIC;
3761                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3762                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3763                 ret = 0;
3764                 break;
3765         default:
3766                 ret = -EINVAL;
3767                 break;
3768         }
3769         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3770         return ret;
3771 }
3772
3773 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3774                                struct ib_qp_attr *qp_attr,
3775                                int *qp_attr_mask)
3776 {
3777         unsigned long flags;
3778         int ret;
3779
3780         spin_lock_irqsave(&cm_id_priv->lock, flags);
3781         switch (cm_id_priv->id.state) {
3782         case IB_CM_REQ_RCVD:
3783         case IB_CM_MRA_REQ_SENT:
3784         case IB_CM_REP_RCVD:
3785         case IB_CM_MRA_REP_SENT:
3786         case IB_CM_REP_SENT:
3787         case IB_CM_MRA_REP_RCVD:
3788         case IB_CM_ESTABLISHED:
3789                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3790                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3791                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3792                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3793                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3794                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3795                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3796                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3797                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3798                                          IB_QP_MIN_RNR_TIMER;
3799                         qp_attr->max_dest_rd_atomic =
3800                                         cm_id_priv->responder_resources;
3801                         qp_attr->min_rnr_timer = 0;
3802                 }
3803                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3804                         *qp_attr_mask |= IB_QP_ALT_PATH;
3805                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3806                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3807                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3808                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3809                 }
3810                 ret = 0;
3811                 break;
3812         default:
3813                 ret = -EINVAL;
3814                 break;
3815         }
3816         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3817         return ret;
3818 }
3819
3820 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3821                                struct ib_qp_attr *qp_attr,
3822                                int *qp_attr_mask)
3823 {
3824         unsigned long flags;
3825         int ret;
3826
3827         spin_lock_irqsave(&cm_id_priv->lock, flags);
3828         switch (cm_id_priv->id.state) {
3829         /* Allow transition to RTS before sending REP */
3830         case IB_CM_REQ_RCVD:
3831         case IB_CM_MRA_REQ_SENT:
3832
3833         case IB_CM_REP_RCVD:
3834         case IB_CM_MRA_REP_SENT:
3835         case IB_CM_REP_SENT:
3836         case IB_CM_MRA_REP_RCVD:
3837         case IB_CM_ESTABLISHED:
3838                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3839                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3840                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3841                         switch (cm_id_priv->qp_type) {
3842                         case IB_QPT_RC:
3843                         case IB_QPT_XRC_INI:
3844                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3845                                                  IB_QP_MAX_QP_RD_ATOMIC;
3846                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3847                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3848                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3849                                 /* fall through */
3850                         case IB_QPT_XRC_TGT:
3851                                 *qp_attr_mask |= IB_QP_TIMEOUT;
3852                                 qp_attr->timeout = cm_id_priv->av.timeout;
3853                                 break;
3854                         default:
3855                                 break;
3856                         }
3857                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3858                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3859                                 qp_attr->path_mig_state = IB_MIG_REARM;
3860                         }
3861                 } else {
3862                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3863                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3864                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3865                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3866                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3867                         qp_attr->path_mig_state = IB_MIG_REARM;
3868                 }
3869                 ret = 0;
3870                 break;
3871         default:
3872                 ret = -EINVAL;
3873                 break;
3874         }
3875         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3876         return ret;
3877 }
3878
3879 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3880                        struct ib_qp_attr *qp_attr,
3881                        int *qp_attr_mask)
3882 {
3883         struct cm_id_private *cm_id_priv;
3884         int ret;
3885
3886         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3887         switch (qp_attr->qp_state) {
3888         case IB_QPS_INIT:
3889                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3890                 break;
3891         case IB_QPS_RTR:
3892                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3893                 break;
3894         case IB_QPS_RTS:
3895                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3896                 break;
3897         default:
3898                 ret = -EINVAL;
3899                 break;
3900         }
3901         return ret;
3902 }
3903 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3904
3905 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3906                                char *buf)
3907 {
3908         struct cm_counter_group *group;
3909         struct cm_counter_attribute *cm_attr;
3910
3911         group = container_of(obj, struct cm_counter_group, obj);
3912         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3913
3914         return sprintf(buf, "%ld\n",
3915                        atomic_long_read(&group->counter[cm_attr->index]));
3916 }
3917
3918 static const struct sysfs_ops cm_counter_ops = {
3919         .show = cm_show_counter
3920 };
3921
3922 static struct kobj_type cm_counter_obj_type = {
3923         .sysfs_ops = &cm_counter_ops,
3924         .default_attrs = cm_counter_default_attrs
3925 };
3926
3927 static void cm_release_port_obj(struct kobject *obj)
3928 {
3929         struct cm_port *cm_port;
3930
3931         cm_port = container_of(obj, struct cm_port, port_obj);
3932         kfree(cm_port);
3933 }
3934
3935 static struct kobj_type cm_port_obj_type = {
3936         .release = cm_release_port_obj
3937 };
3938
3939 static char *cm_devnode(struct device *dev, umode_t *mode)
3940 {
3941         if (mode)
3942                 *mode = 0666;
3943         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3944 }
3945
3946 struct class cm_class = {
3947         .owner   = THIS_MODULE,
3948         .name    = "infiniband_cm",
3949         .devnode = cm_devnode,
3950 };
3951 EXPORT_SYMBOL(cm_class);
3952
3953 static int cm_create_port_fs(struct cm_port *port)
3954 {
3955         int i, ret;
3956
3957         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3958                                    &port->cm_dev->device->kobj,
3959                                    "%d", port->port_num);
3960         if (ret) {
3961                 kfree(port);
3962                 return ret;
3963         }
3964
3965         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3966                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3967                                            &cm_counter_obj_type,
3968                                            &port->port_obj,
3969                                            "%s", counter_group_names[i]);
3970                 if (ret)
3971                         goto error;
3972         }
3973
3974         return 0;
3975
3976 error:
3977         while (i--)
3978                 kobject_put(&port->counter_group[i].obj);
3979         kobject_put(&port->port_obj);
3980         return ret;
3981
3982 }
3983
3984 static void cm_remove_port_fs(struct cm_port *port)
3985 {
3986         int i;
3987
3988         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3989                 kobject_put(&port->counter_group[i].obj);
3990
3991         kobject_put(&port->port_obj);
3992 }
3993
3994 static void cm_add_one(struct ib_device *ib_device)
3995 {
3996         struct cm_device *cm_dev;
3997         struct cm_port *port;
3998         struct ib_mad_reg_req reg_req = {
3999                 .mgmt_class = IB_MGMT_CLASS_CM,
4000                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4001         };
4002         struct ib_port_modify port_modify = {
4003                 .set_port_cap_mask = IB_PORT_CM_SUP
4004         };
4005         unsigned long flags;
4006         int ret;
4007         int count = 0;
4008         u8 i;
4009
4010         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
4011                          ib_device->phys_port_cnt, GFP_KERNEL);
4012         if (!cm_dev)
4013                 return;
4014
4015         cm_dev->ib_device = ib_device;
4016         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4017         cm_dev->going_down = 0;
4018         cm_dev->device = device_create(&cm_class, &ib_device->dev,
4019                                        MKDEV(0, 0), NULL,
4020                                        "%s", ib_device->name);
4021         if (IS_ERR(cm_dev->device)) {
4022                 kfree(cm_dev);
4023                 return;
4024         }
4025
4026         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4027         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4028                 if (!rdma_cap_ib_cm(ib_device, i))
4029                         continue;
4030
4031                 port = kzalloc(sizeof *port, GFP_KERNEL);
4032                 if (!port)
4033                         goto error1;
4034
4035                 cm_dev->port[i-1] = port;
4036                 port->cm_dev = cm_dev;
4037                 port->port_num = i;
4038
4039                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
4040                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
4041
4042                 ret = cm_create_port_fs(port);
4043                 if (ret)
4044                         goto error1;
4045
4046                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4047                                                         IB_QPT_GSI,
4048                                                         &reg_req,
4049                                                         0,
4050                                                         cm_send_handler,
4051                                                         cm_recv_handler,
4052                                                         port,
4053                                                         0);
4054                 if (IS_ERR(port->mad_agent))
4055                         goto error2;
4056
4057                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4058                 if (ret)
4059                         goto error3;
4060
4061                 count++;
4062         }
4063
4064         if (!count)
4065                 goto free;
4066
4067         ib_set_client_data(ib_device, &cm_client, cm_dev);
4068
4069         write_lock_irqsave(&cm.device_lock, flags);
4070         list_add_tail(&cm_dev->list, &cm.device_list);
4071         write_unlock_irqrestore(&cm.device_lock, flags);
4072         return;
4073
4074 error3:
4075         ib_unregister_mad_agent(port->mad_agent);
4076 error2:
4077         cm_remove_port_fs(port);
4078 error1:
4079         port_modify.set_port_cap_mask = 0;
4080         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4081         while (--i) {
4082                 if (!rdma_cap_ib_cm(ib_device, i))
4083                         continue;
4084
4085                 port = cm_dev->port[i-1];
4086                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4087                 ib_unregister_mad_agent(port->mad_agent);
4088                 cm_remove_port_fs(port);
4089         }
4090 free:
4091         device_unregister(cm_dev->device);
4092         kfree(cm_dev);
4093 }
4094
4095 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4096 {
4097         struct cm_device *cm_dev = client_data;
4098         struct cm_port *port;
4099         struct cm_id_private *cm_id_priv;
4100         struct ib_mad_agent *cur_mad_agent;
4101         struct ib_port_modify port_modify = {
4102                 .clr_port_cap_mask = IB_PORT_CM_SUP
4103         };
4104         unsigned long flags;
4105         int i;
4106
4107         if (!cm_dev)
4108                 return;
4109
4110         write_lock_irqsave(&cm.device_lock, flags);
4111         list_del(&cm_dev->list);
4112         write_unlock_irqrestore(&cm.device_lock, flags);
4113
4114         spin_lock_irq(&cm.lock);
4115         cm_dev->going_down = 1;
4116         spin_unlock_irq(&cm.lock);
4117
4118         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4119                 if (!rdma_cap_ib_cm(ib_device, i))
4120                         continue;
4121
4122                 port = cm_dev->port[i-1];
4123                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4124                 /* Mark all the cm_id's as not valid */
4125                 spin_lock_irq(&cm.lock);
4126                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4127                         cm_id_priv->altr_send_port_not_ready = 1;
4128                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4129                         cm_id_priv->prim_send_port_not_ready = 1;
4130                 spin_unlock_irq(&cm.lock);
4131                 /*
4132                  * We flush the queue here after the going_down set, this
4133                  * verify that no new works will be queued in the recv handler,
4134                  * after that we can call the unregister_mad_agent
4135                  */
4136                 flush_workqueue(cm.wq);
4137                 spin_lock_irq(&cm.state_lock);
4138                 cur_mad_agent = port->mad_agent;
4139                 port->mad_agent = NULL;
4140                 spin_unlock_irq(&cm.state_lock);
4141                 ib_unregister_mad_agent(cur_mad_agent);
4142                 cm_remove_port_fs(port);
4143         }
4144
4145         device_unregister(cm_dev->device);
4146         kfree(cm_dev);
4147 }
4148
4149 static int __init ib_cm_init(void)
4150 {
4151         int ret;
4152
4153         memset(&cm, 0, sizeof cm);
4154         INIT_LIST_HEAD(&cm.device_list);
4155         rwlock_init(&cm.device_lock);
4156         spin_lock_init(&cm.lock);
4157         spin_lock_init(&cm.state_lock);
4158         cm.listen_service_table = RB_ROOT;
4159         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4160         cm.remote_id_table = RB_ROOT;
4161         cm.remote_qp_table = RB_ROOT;
4162         cm.remote_sidr_table = RB_ROOT;
4163         idr_init(&cm.local_id_table);
4164         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4165         INIT_LIST_HEAD(&cm.timewait_list);
4166
4167         ret = class_register(&cm_class);
4168         if (ret) {
4169                 ret = -ENOMEM;
4170                 goto error1;
4171         }
4172
4173         cm.wq = create_workqueue("ib_cm");
4174         if (!cm.wq) {
4175                 ret = -ENOMEM;
4176                 goto error2;
4177         }
4178
4179         ret = ib_register_client(&cm_client);
4180         if (ret)
4181                 goto error3;
4182
4183         return 0;
4184 error3:
4185         destroy_workqueue(cm.wq);
4186 error2:
4187         class_unregister(&cm_class);
4188 error1:
4189         idr_destroy(&cm.local_id_table);
4190         return ret;
4191 }
4192
4193 static void __exit ib_cm_cleanup(void)
4194 {
4195         struct cm_timewait_info *timewait_info, *tmp;
4196
4197         spin_lock_irq(&cm.lock);
4198         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4199                 cancel_delayed_work(&timewait_info->work.work);
4200         spin_unlock_irq(&cm.lock);
4201
4202         ib_unregister_client(&cm_client);
4203         destroy_workqueue(cm.wq);
4204
4205         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4206                 list_del(&timewait_info->list);
4207                 kfree(timewait_info);
4208         }
4209
4210         class_unregister(&cm_class);
4211         idr_destroy(&cm.local_id_table);
4212 }
4213
4214 module_init(ib_cm_init);
4215 module_exit(ib_cm_cleanup);
4216