drivers/infiniband/core/cm.c

   1 /*
   2  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/dma-mapping.h>
  38 #include <linux/device.h>
  39 #include <linux/module.h>
  40 #include <linux/err.h>
  41 #include <linux/idr.h>
  42 #include <linux/interrupt.h>
  43 #include <linux/random.h>
  44 #include <linux/rbtree.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/slab.h>
  47 #include <linux/sysfs.h>
  48 #include <linux/workqueue.h>
  49 #include <linux/kdev_t.h>
  50 #include <linux/etherdevice.h>
  51
  52 #include <rdma/ib_cache.h>
  53 #include <rdma/ib_cm.h>
  54 #include "cm_msgs.h"
  55
  56 MODULE_AUTHOR("Sean Hefty");
  57 MODULE_DESCRIPTION("InfiniBand CM");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59
  60 static const char * const ibcm_rej_reason_strs[] = {
  61         [IB_CM_REJ_NO_QP]                       = "no QP",
  62         [IB_CM_REJ_NO_EEC]                      = "no EEC",
  63         [IB_CM_REJ_NO_RESOURCES]                = "no resources",
  64         [IB_CM_REJ_TIMEOUT]                     = "timeout",
  65         [IB_CM_REJ_UNSUPPORTED]                 = "unsupported",
  66         [IB_CM_REJ_INVALID_COMM_ID]             = "invalid comm ID",
  67         [IB_CM_REJ_INVALID_COMM_INSTANCE]       = "invalid comm instance",
  68         [IB_CM_REJ_INVALID_SERVICE_ID]          = "invalid service ID",
  69         [IB_CM_REJ_INVALID_TRANSPORT_TYPE]      = "invalid transport type",
  70         [IB_CM_REJ_STALE_CONN]                  = "stale conn",
  71         [IB_CM_REJ_RDC_NOT_EXIST]               = "RDC not exist",
  72         [IB_CM_REJ_INVALID_GID]                 = "invalid GID",
  73         [IB_CM_REJ_INVALID_LID]                 = "invalid LID",
  74         [IB_CM_REJ_INVALID_SL]                  = "invalid SL",
  75         [IB_CM_REJ_INVALID_TRAFFIC_CLASS]       = "invalid traffic class",
  76         [IB_CM_REJ_INVALID_HOP_LIMIT]           = "invalid hop limit",
  77         [IB_CM_REJ_INVALID_PACKET_RATE]         = "invalid packet rate",
  78         [IB_CM_REJ_INVALID_ALT_GID]             = "invalid alt GID",
  79         [IB_CM_REJ_INVALID_ALT_LID]             = "invalid alt LID",
  80         [IB_CM_REJ_INVALID_ALT_SL]              = "invalid alt SL",
  81         [IB_CM_REJ_INVALID_ALT_TRAFFIC_CLASS]   = "invalid alt traffic class",
  82         [IB_CM_REJ_INVALID_ALT_HOP_LIMIT]       = "invalid alt hop limit",
  83         [IB_CM_REJ_INVALID_ALT_PACKET_RATE]     = "invalid alt packet rate",
  84         [IB_CM_REJ_PORT_CM_REDIRECT]            = "port CM redirect",
  85         [IB_CM_REJ_PORT_REDIRECT]               = "port redirect",
  86         [IB_CM_REJ_INVALID_MTU]                 = "invalid MTU",
  87         [IB_CM_REJ_INSUFFICIENT_RESP_RESOURCES] = "insufficient resp resources",
  88         [IB_CM_REJ_CONSUMER_DEFINED]            = "consumer defined",
  89         [IB_CM_REJ_INVALID_RNR_RETRY]           = "invalid RNR retry",
  90         [IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID]     = "duplicate local comm ID",
  91         [IB_CM_REJ_INVALID_CLASS_VERSION]       = "invalid class version",
  92         [IB_CM_REJ_INVALID_FLOW_LABEL]          = "invalid flow label",
  93         [IB_CM_REJ_INVALID_ALT_FLOW_LABEL]      = "invalid alt flow label",
  94 };
  95
  96 const char *__attribute_const__ ibcm_reject_msg(int reason)
  97 {
  98         size_t index = reason;
  99
 100         if (index < ARRAY_SIZE(ibcm_rej_reason_strs) &&
 101             ibcm_rej_reason_strs[index])
 102                 return ibcm_rej_reason_strs[index];
 103         else
 104                 return "unrecognized reason";
 105 }
 106 EXPORT_SYMBOL(ibcm_reject_msg);
 107
 108 static void cm_add_one(struct ib_device *device);
 109 static void cm_remove_one(struct ib_device *device, void *client_data);
 110
 111 static struct ib_client cm_client = {
 112         .name   = "cm",
 113         .add    = cm_add_one,
 114         .remove = cm_remove_one
 115 };
 116
 117 static struct ib_cm {
 118         spinlock_t lock;
 119         struct list_head device_list;
 120         rwlock_t device_lock;
 121         struct rb_root listen_service_table;
 122         u64 listen_service_id;
 123         /* struct rb_root peer_service_table; todo: fix peer to peer */
 124         struct rb_root remote_qp_table;
 125         struct rb_root remote_id_table;
 126         struct rb_root remote_sidr_table;
 127         struct idr local_id_table;
 128         __be32 random_id_operand;
 129         struct list_head timewait_list;
 130         struct workqueue_struct *wq;
 131         /* Sync on cm change port state */
 132         spinlock_t state_lock;
 133 } cm;
 134
 135 /* Counter indexes ordered by attribute ID */
 136 enum {
 137         CM_REQ_COUNTER,
 138         CM_MRA_COUNTER,
 139         CM_REJ_COUNTER,
 140         CM_REP_COUNTER,
 141         CM_RTU_COUNTER,
 142         CM_DREQ_COUNTER,
 143         CM_DREP_COUNTER,
 144         CM_SIDR_REQ_COUNTER,
 145         CM_SIDR_REP_COUNTER,
 146         CM_LAP_COUNTER,
 147         CM_APR_COUNTER,
 148         CM_ATTR_COUNT,
 149         CM_ATTR_ID_OFFSET = 0x0010,
 150 };
 151
 152 enum {
 153         CM_XMIT,
 154         CM_XMIT_RETRIES,
 155         CM_RECV,
 156         CM_RECV_DUPLICATES,
 157         CM_COUNTER_GROUPS
 158 };
 159
 160 static char const counter_group_names[CM_COUNTER_GROUPS]
 161                                      [sizeof("cm_rx_duplicates")] = {
 162         "cm_tx_msgs", "cm_tx_retries",
 163         "cm_rx_msgs", "cm_rx_duplicates"
 164 };
 165
 166 struct cm_counter_group {
 167         struct kobject obj;
 168         atomic_long_t counter[CM_ATTR_COUNT];
 169 };
 170
 171 struct cm_counter_attribute {
 172         struct attribute attr;
 173         int index;
 174 };
 175
 176 #define CM_COUNTER_ATTR(_name, _index) \
 177 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 178         .attr = { .name = __stringify(_name), .mode = 0444 }, \
 179         .index = _index \
 180 }
 181
 182 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 183 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 184 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 185 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 186 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 187 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 188 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 189 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 190 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 191 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 192 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 193
 194 static struct attribute *cm_counter_default_attrs[] = {
 195         &cm_req_counter_attr.attr,
 196         &cm_mra_counter_attr.attr,
 197         &cm_rej_counter_attr.attr,
 198         &cm_rep_counter_attr.attr,
 199         &cm_rtu_counter_attr.attr,
 200         &cm_dreq_counter_attr.attr,
 201         &cm_drep_counter_attr.attr,
 202         &cm_sidr_req_counter_attr.attr,
 203         &cm_sidr_rep_counter_attr.attr,
 204         &cm_lap_counter_attr.attr,
 205         &cm_apr_counter_attr.attr,
 206         NULL
 207 };
 208
 209 struct cm_port {
 210         struct cm_device *cm_dev;
 211         struct ib_mad_agent *mad_agent;
 212         struct kobject port_obj;
 213         u8 port_num;
 214         struct list_head cm_priv_prim_list;
 215         struct list_head cm_priv_altr_list;
 216         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 217 };
 218
 219 struct cm_device {
 220         struct list_head list;
 221         struct ib_device *ib_device;
 222         struct device *device;
 223         u8 ack_delay;
 224         int going_down;
 225         struct cm_port *port[0];
 226 };
 227
 228 struct cm_av {
 229         struct cm_port *port;
 230         union ib_gid dgid;
 231         struct rdma_ah_attr ah_attr;
 232         u16 pkey_index;
 233         u8 timeout;
 234 };
 235
 236 struct cm_work {
 237         struct delayed_work work;
 238         struct list_head list;
 239         struct cm_port *port;
 240         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 241         __be32 local_id;                        /* Established / timewait */
 242         __be32 remote_id;
 243         struct ib_cm_event cm_event;
 244         struct sa_path_rec path[0];
 245 };
 246
 247 struct cm_timewait_info {
 248         struct cm_work work;                    /* Must be first. */
 249         struct list_head list;
 250         struct rb_node remote_qp_node;
 251         struct rb_node remote_id_node;
 252         __be64 remote_ca_guid;
 253         __be32 remote_qpn;
 254         u8 inserted_remote_qp;
 255         u8 inserted_remote_id;
 256 };
 257
 258 struct cm_id_private {
 259         struct ib_cm_id id;
 260
 261         struct rb_node service_node;
 262         struct rb_node sidr_id_node;
 263         spinlock_t lock;        /* Do not acquire inside cm.lock */
 264         struct completion comp;
 265         atomic_t refcount;
 266         /* Number of clients sharing this ib_cm_id. Only valid for listeners.
 267          * Protected by the cm.lock spinlock. */
 268         int listen_sharecount;
 269
 270         struct ib_mad_send_buf *msg;
 271         struct cm_timewait_info *timewait_info;
 272         /* todo: use alternate port on send failure */
 273         struct cm_av av;
 274         struct cm_av alt_av;
 275
 276         void *private_data;
 277         __be64 tid;
 278         __be32 local_qpn;
 279         __be32 remote_qpn;
 280         enum ib_qp_type qp_type;
 281         __be32 sq_psn;
 282         __be32 rq_psn;
 283         int timeout_ms;
 284         enum ib_mtu path_mtu;
 285         __be16 pkey;
 286         u8 private_data_len;
 287         u8 max_cm_retries;
 288         u8 peer_to_peer;
 289         u8 responder_resources;
 290         u8 initiator_depth;
 291         u8 retry_count;
 292         u8 rnr_retry_count;
 293         u8 service_timeout;
 294         u8 target_ack_delay;
 295
 296         struct list_head prim_list;
 297         struct list_head altr_list;
 298         /* Indicates that the send port mad is registered and av is set */
 299         int prim_send_port_not_ready;
 300         int altr_send_port_not_ready;
 301
 302         struct list_head work_list;
 303         atomic_t work_count;
 304 };
 305
 306 static void cm_work_handler(struct work_struct *work);
 307
 308 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 309 {
 310         if (atomic_dec_and_test(&cm_id_priv->refcount))
 311                 complete(&cm_id_priv->comp);
 312 }
 313
 314 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 315                         struct ib_mad_send_buf **msg)
 316 {
 317         struct ib_mad_agent *mad_agent;
 318         struct ib_mad_send_buf *m;
 319         struct ib_ah *ah;
 320         struct cm_av *av;
 321         unsigned long flags, flags2;
 322         int ret = 0;
 323
 324         /* don't let the port to be released till the agent is down */
 325         spin_lock_irqsave(&cm.state_lock, flags2);
 326         spin_lock_irqsave(&cm.lock, flags);
 327         if (!cm_id_priv->prim_send_port_not_ready)
 328                 av = &cm_id_priv->av;
 329         else if (!cm_id_priv->altr_send_port_not_ready &&
 330                  (cm_id_priv->alt_av.port))
 331                 av = &cm_id_priv->alt_av;
 332         else {
 333                 pr_info("%s: not valid CM id\n", __func__);
 334                 ret = -ENODEV;
 335                 spin_unlock_irqrestore(&cm.lock, flags);
 336                 goto out;
 337         }
 338         spin_unlock_irqrestore(&cm.lock, flags);
 339         /* Make sure the port haven't released the mad yet */
 340         mad_agent = cm_id_priv->av.port->mad_agent;
 341         if (!mad_agent) {
 342                 pr_info("%s: not a valid MAD agent\n", __func__);
 343                 ret = -ENODEV;
 344                 goto out;
 345         }
 346         ah = rdma_create_ah(mad_agent->qp->pd, &av->ah_attr);
 347         if (IS_ERR(ah)) {
 348                 ret = PTR_ERR(ah);
 349                 goto out;
 350         }
 351
 352         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 353                                av->pkey_index,
 354                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 355                                GFP_ATOMIC,
 356                                IB_MGMT_BASE_VERSION);
 357         if (IS_ERR(m)) {
 358                 rdma_destroy_ah(ah);
 359                 ret = PTR_ERR(m);
 360                 goto out;
 361         }
 362
 363         /* Timeout set by caller if response is expected. */
 364         m->ah = ah;
 365         m->retries = cm_id_priv->max_cm_retries;
 366
 367         atomic_inc(&cm_id_priv->refcount);
 368         m->context[0] = cm_id_priv;
 369         *msg = m;
 370
 371 out:
 372         spin_unlock_irqrestore(&cm.state_lock, flags2);
 373         return ret;
 374 }
 375
 376 static int cm_alloc_response_msg(struct cm_port *port,
 377                                  struct ib_mad_recv_wc *mad_recv_wc,
 378                                  struct ib_mad_send_buf **msg)
 379 {
 380         struct ib_mad_send_buf *m;
 381         struct ib_ah *ah;
 382
 383         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 384                                   mad_recv_wc->recv_buf.grh, port->port_num);
 385         if (IS_ERR(ah))
 386                 return PTR_ERR(ah);
 387
 388         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 389                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 390                                GFP_ATOMIC,
 391                                IB_MGMT_BASE_VERSION);
 392         if (IS_ERR(m)) {
 393                 rdma_destroy_ah(ah);
 394                 return PTR_ERR(m);
 395         }
 396         m->ah = ah;
 397         *msg = m;
 398         return 0;
 399 }
 400
 401 static void cm_free_msg(struct ib_mad_send_buf *msg)
 402 {
 403         rdma_destroy_ah(msg->ah);
 404         if (msg->context[0])
 405                 cm_deref_id(msg->context[0]);
 406         ib_free_send_mad(msg);
 407 }
 408
 409 static void * cm_copy_private_data(const void *private_data,
 410                                    u8 private_data_len)
 411 {
 412         void *data;
 413
 414         if (!private_data || !private_data_len)
 415                 return NULL;
 416
 417         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 418         if (!data)
 419                 return ERR_PTR(-ENOMEM);
 420
 421         return data;
 422 }
 423
 424 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 425                                  void *private_data, u8 private_data_len)
 426 {
 427         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 428                 kfree(cm_id_priv->private_data);
 429
 430         cm_id_priv->private_data = private_data;
 431         cm_id_priv->private_data_len = private_data_len;
 432 }
 433
 434 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 435                                     struct ib_grh *grh, struct cm_av *av)
 436 {
 437         av->port = port;
 438         av->pkey_index = wc->pkey_index;
 439         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 440                            grh, &av->ah_attr);
 441 }
 442
 443 static int cm_init_av_by_path(struct 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                                         sa_conv_pathrec_to_gid_type(path),
 457                                         ndev, &p, NULL)) {
 458                         port = cm_dev->port[p-1];
 459                         break;
 460                 }
 461         }
 462         read_unlock_irqrestore(&cm.device_lock, flags);
 463
 464         if (ndev)
 465                 dev_put(ndev);
 466
 467         if (!port)
 468                 return -EINVAL;
 469
 470         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 471                                   be16_to_cpu(path->pkey), &av->pkey_index);
 472         if (ret)
 473                 return ret;
 474
 475         av->port = port;
 476         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 477                              &av->ah_attr);
 478         av->timeout = path->packet_life_time + 1;
 479
 480         spin_lock_irqsave(&cm.lock, flags);
 481         if (&cm_id_priv->av == av)
 482                 list_add_tail(&cm_id_priv->prim_list, &port->cm_priv_prim_list);
 483         else if (&cm_id_priv->alt_av == av)
 484                 list_add_tail(&cm_id_priv->altr_list, &port->cm_priv_altr_list);
 485         else
 486                 ret = -EINVAL;
 487
 488         spin_unlock_irqrestore(&cm.lock, flags);
 489
 490         return ret;
 491 }
 492
 493 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 494 {
 495         unsigned long flags;
 496         int id;
 497
 498         idr_preload(GFP_KERNEL);
 499         spin_lock_irqsave(&cm.lock, flags);
 500
 501         id = idr_alloc_cyclic(&cm.local_id_table, cm_id_priv, 0, 0, GFP_NOWAIT);
 502
 503         spin_unlock_irqrestore(&cm.lock, flags);
 504         idr_preload_end();
 505
 506         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 507         return id < 0 ? id : 0;
 508 }
 509
 510 static void cm_free_id(__be32 local_id)
 511 {
 512         spin_lock_irq(&cm.lock);
 513         idr_remove(&cm.local_id_table,
 514                    (__force int) (local_id ^ cm.random_id_operand));
 515         spin_unlock_irq(&cm.lock);
 516 }
 517
 518 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 519 {
 520         struct cm_id_private *cm_id_priv;
 521
 522         cm_id_priv = idr_find(&cm.local_id_table,
 523                               (__force int) (local_id ^ cm.random_id_operand));
 524         if (cm_id_priv) {
 525                 if (cm_id_priv->id.remote_id == remote_id)
 526                         atomic_inc(&cm_id_priv->refcount);
 527                 else
 528                         cm_id_priv = NULL;
 529         }
 530
 531         return cm_id_priv;
 532 }
 533
 534 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 535 {
 536         struct cm_id_private *cm_id_priv;
 537
 538         spin_lock_irq(&cm.lock);
 539         cm_id_priv = cm_get_id(local_id, remote_id);
 540         spin_unlock_irq(&cm.lock);
 541
 542         return cm_id_priv;
 543 }
 544
 545 /*
 546  * Trivial helpers to strip endian annotation and compare; the
 547  * endianness doesn't actually matter since we just need a stable
 548  * order for the RB tree.
 549  */
 550 static int be32_lt(__be32 a, __be32 b)
 551 {
 552         return (__force u32) a < (__force u32) b;
 553 }
 554
 555 static int be32_gt(__be32 a, __be32 b)
 556 {
 557         return (__force u32) a > (__force u32) b;
 558 }
 559
 560 static int be64_lt(__be64 a, __be64 b)
 561 {
 562         return (__force u64) a < (__force u64) b;
 563 }
 564
 565 static int be64_gt(__be64 a, __be64 b)
 566 {
 567         return (__force u64) a > (__force u64) b;
 568 }
 569
 570 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 571 {
 572         struct rb_node **link = &cm.listen_service_table.rb_node;
 573         struct rb_node *parent = NULL;
 574         struct cm_id_private *cur_cm_id_priv;
 575         __be64 service_id = cm_id_priv->id.service_id;
 576         __be64 service_mask = cm_id_priv->id.service_mask;
 577
 578         while (*link) {
 579                 parent = *link;
 580                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 581                                           service_node);
 582                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 583                     (service_mask & cur_cm_id_priv->id.service_id) &&
 584                     (cm_id_priv->id.device == cur_cm_id_priv->id.device))
 585                         return cur_cm_id_priv;
 586
 587                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 588                         link = &(*link)->rb_left;
 589                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 590                         link = &(*link)->rb_right;
 591                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 592                         link = &(*link)->rb_left;
 593                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 594                         link = &(*link)->rb_right;
 595                 else
 596                         link = &(*link)->rb_right;
 597         }
 598         rb_link_node(&cm_id_priv->service_node, parent, link);
 599         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 600         return NULL;
 601 }
 602
 603 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 604                                              __be64 service_id)
 605 {
 606         struct rb_node *node = cm.listen_service_table.rb_node;
 607         struct cm_id_private *cm_id_priv;
 608
 609         while (node) {
 610                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 611                 if ((cm_id_priv->id.service_mask & service_id) ==
 612                      cm_id_priv->id.service_id &&
 613                     (cm_id_priv->id.device == device))
 614                         return cm_id_priv;
 615
 616                 if (device < cm_id_priv->id.device)
 617                         node = node->rb_left;
 618                 else if (device > cm_id_priv->id.device)
 619                         node = node->rb_right;
 620                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 621                         node = node->rb_left;
 622                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 623                         node = node->rb_right;
 624                 else
 625                         node = node->rb_right;
 626         }
 627         return NULL;
 628 }
 629
 630 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 631                                                      *timewait_info)
 632 {
 633         struct rb_node **link = &cm.remote_id_table.rb_node;
 634         struct rb_node *parent = NULL;
 635         struct cm_timewait_info *cur_timewait_info;
 636         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 637         __be32 remote_id = timewait_info->work.remote_id;
 638
 639         while (*link) {
 640                 parent = *link;
 641                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 642                                              remote_id_node);
 643                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 644                         link = &(*link)->rb_left;
 645                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 646                         link = &(*link)->rb_right;
 647                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 648                         link = &(*link)->rb_left;
 649                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 650                         link = &(*link)->rb_right;
 651                 else
 652                         return cur_timewait_info;
 653         }
 654         timewait_info->inserted_remote_id = 1;
 655         rb_link_node(&timewait_info->remote_id_node, parent, link);
 656         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 657         return NULL;
 658 }
 659
 660 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 661                                                    __be32 remote_id)
 662 {
 663         struct rb_node *node = cm.remote_id_table.rb_node;
 664         struct cm_timewait_info *timewait_info;
 665
 666         while (node) {
 667                 timewait_info = rb_entry(node, struct cm_timewait_info,
 668                                          remote_id_node);
 669                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 670                         node = node->rb_left;
 671                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 672                         node = node->rb_right;
 673                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 674                         node = node->rb_left;
 675                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 676                         node = node->rb_right;
 677                 else
 678                         return timewait_info;
 679         }
 680         return NULL;
 681 }
 682
 683 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 684                                                       *timewait_info)
 685 {
 686         struct rb_node **link = &cm.remote_qp_table.rb_node;
 687         struct rb_node *parent = NULL;
 688         struct cm_timewait_info *cur_timewait_info;
 689         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 690         __be32 remote_qpn = timewait_info->remote_qpn;
 691
 692         while (*link) {
 693                 parent = *link;
 694                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 695                                              remote_qp_node);
 696                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 697                         link = &(*link)->rb_left;
 698                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 699                         link = &(*link)->rb_right;
 700                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 701                         link = &(*link)->rb_left;
 702                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 703                         link = &(*link)->rb_right;
 704                 else
 705                         return cur_timewait_info;
 706         }
 707         timewait_info->inserted_remote_qp = 1;
 708         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 709         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 710         return NULL;
 711 }
 712
 713 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 714                                                     *cm_id_priv)
 715 {
 716         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 717         struct rb_node *parent = NULL;
 718         struct cm_id_private *cur_cm_id_priv;
 719         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 720         __be32 remote_id = cm_id_priv->id.remote_id;
 721
 722         while (*link) {
 723                 parent = *link;
 724                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 725                                           sidr_id_node);
 726                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 727                         link = &(*link)->rb_left;
 728                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 729                         link = &(*link)->rb_right;
 730                 else {
 731                         int cmp;
 732                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 733                                      sizeof *port_gid);
 734                         if (cmp < 0)
 735                                 link = &(*link)->rb_left;
 736                         else if (cmp > 0)
 737                                 link = &(*link)->rb_right;
 738                         else
 739                                 return cur_cm_id_priv;
 740                 }
 741         }
 742         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 743         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 744         return NULL;
 745 }
 746
 747 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 748                                enum ib_cm_sidr_status status)
 749 {
 750         struct ib_cm_sidr_rep_param param;
 751
 752         memset(&param, 0, sizeof param);
 753         param.status = status;
 754         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 755 }
 756
 757 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 758                                  ib_cm_handler cm_handler,
 759                                  void *context)
 760 {
 761         struct cm_id_private *cm_id_priv;
 762         int ret;
 763
 764         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 765         if (!cm_id_priv)
 766                 return ERR_PTR(-ENOMEM);
 767
 768         cm_id_priv->id.state = IB_CM_IDLE;
 769         cm_id_priv->id.device = device;
 770         cm_id_priv->id.cm_handler = cm_handler;
 771         cm_id_priv->id.context = context;
 772         cm_id_priv->id.remote_cm_qpn = 1;
 773         ret = cm_alloc_id(cm_id_priv);
 774         if (ret)
 775                 goto error;
 776
 777         spin_lock_init(&cm_id_priv->lock);
 778         init_completion(&cm_id_priv->comp);
 779         INIT_LIST_HEAD(&cm_id_priv->work_list);
 780         INIT_LIST_HEAD(&cm_id_priv->prim_list);
 781         INIT_LIST_HEAD(&cm_id_priv->altr_list);
 782         atomic_set(&cm_id_priv->work_count, -1);
 783         atomic_set(&cm_id_priv->refcount, 1);
 784         return &cm_id_priv->id;
 785
 786 error:
 787         kfree(cm_id_priv);
 788         return ERR_PTR(-ENOMEM);
 789 }
 790 EXPORT_SYMBOL(ib_create_cm_id);
 791
 792 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 793 {
 794         struct cm_work *work;
 795
 796         if (list_empty(&cm_id_priv->work_list))
 797                 return NULL;
 798
 799         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 800         list_del(&work->list);
 801         return work;
 802 }
 803
 804 static void cm_free_work(struct cm_work *work)
 805 {
 806         if (work->mad_recv_wc)
 807                 ib_free_recv_mad(work->mad_recv_wc);
 808         kfree(work);
 809 }
 810
 811 static inline int cm_convert_to_ms(int iba_time)
 812 {
 813         /* approximate conversion to ms from 4.096us x 2^iba_time */
 814         return 1 << max(iba_time - 8, 0);
 815 }
 816
 817 /*
 818  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 819  * Because of how ack_timeout is stored, adding one doubles the timeout.
 820  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 821  * increment it (round up) only if the other is within 50%.
 822  */
 823 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 824 {
 825         int ack_timeout = packet_life_time + 1;
 826
 827         if (ack_timeout >= ca_ack_delay)
 828                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 829         else
 830                 ack_timeout = ca_ack_delay +
 831                               (ack_timeout >= (ca_ack_delay - 1));
 832
 833         return min(31, ack_timeout);
 834 }
 835
 836 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 837 {
 838         if (timewait_info->inserted_remote_id) {
 839                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 840                 timewait_info->inserted_remote_id = 0;
 841         }
 842
 843         if (timewait_info->inserted_remote_qp) {
 844                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 845                 timewait_info->inserted_remote_qp = 0;
 846         }
 847 }
 848
 849 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 850 {
 851         struct cm_timewait_info *timewait_info;
 852
 853         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 854         if (!timewait_info)
 855                 return ERR_PTR(-ENOMEM);
 856
 857         timewait_info->work.local_id = local_id;
 858         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 859         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 860         return timewait_info;
 861 }
 862
 863 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 864 {
 865         int wait_time;
 866         unsigned long flags;
 867         struct cm_device *cm_dev;
 868
 869         cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
 870         if (!cm_dev)
 871                 return;
 872
 873         spin_lock_irqsave(&cm.lock, flags);
 874         cm_cleanup_timewait(cm_id_priv->timewait_info);
 875         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 876         spin_unlock_irqrestore(&cm.lock, flags);
 877
 878         /*
 879          * The cm_id could be destroyed by the user before we exit timewait.
 880          * To protect against this, we search for the cm_id after exiting
 881          * timewait before notifying the user that we've exited timewait.
 882          */
 883         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 884         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 885
 886         /* Check if the device started its remove_one */
 887         spin_lock_irqsave(&cm.lock, flags);
 888         if (!cm_dev->going_down)
 889                 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 890                                    msecs_to_jiffies(wait_time));
 891         spin_unlock_irqrestore(&cm.lock, flags);
 892
 893         cm_id_priv->timewait_info = NULL;
 894 }
 895
 896 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 897 {
 898         unsigned long flags;
 899
 900         cm_id_priv->id.state = IB_CM_IDLE;
 901         if (cm_id_priv->timewait_info) {
 902                 spin_lock_irqsave(&cm.lock, flags);
 903                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 904                 spin_unlock_irqrestore(&cm.lock, flags);
 905                 kfree(cm_id_priv->timewait_info);
 906                 cm_id_priv->timewait_info = NULL;
 907         }
 908 }
 909
 910 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 911 {
 912         struct cm_id_private *cm_id_priv;
 913         struct cm_work *work;
 914
 915         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 916 retest:
 917         spin_lock_irq(&cm_id_priv->lock);
 918         switch (cm_id->state) {
 919         case IB_CM_LISTEN:
 920                 spin_unlock_irq(&cm_id_priv->lock);
 921
 922                 spin_lock_irq(&cm.lock);
 923                 if (--cm_id_priv->listen_sharecount > 0) {
 924                         /* The id is still shared. */
 925                         cm_deref_id(cm_id_priv);
 926                         spin_unlock_irq(&cm.lock);
 927                         return;
 928                 }
 929                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 930                 spin_unlock_irq(&cm.lock);
 931                 break;
 932         case IB_CM_SIDR_REQ_SENT:
 933                 cm_id->state = IB_CM_IDLE;
 934                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 935                 spin_unlock_irq(&cm_id_priv->lock);
 936                 break;
 937         case IB_CM_SIDR_REQ_RCVD:
 938                 spin_unlock_irq(&cm_id_priv->lock);
 939                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 940                 spin_lock_irq(&cm.lock);
 941                 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node))
 942                         rb_erase(&cm_id_priv->sidr_id_node,
 943                                  &cm.remote_sidr_table);
 944                 spin_unlock_irq(&cm.lock);
 945                 break;
 946         case IB_CM_REQ_SENT:
 947         case IB_CM_MRA_REQ_RCVD:
 948                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 949                 spin_unlock_irq(&cm_id_priv->lock);
 950                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 951                                &cm_id_priv->id.device->node_guid,
 952                                sizeof cm_id_priv->id.device->node_guid,
 953                                NULL, 0);
 954                 break;
 955         case IB_CM_REQ_RCVD:
 956                 if (err == -ENOMEM) {
 957                         /* Do not reject to allow future retries. */
 958                         cm_reset_to_idle(cm_id_priv);
 959                         spin_unlock_irq(&cm_id_priv->lock);
 960                 } else {
 961                         spin_unlock_irq(&cm_id_priv->lock);
 962                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 963                                        NULL, 0, NULL, 0);
 964                 }
 965                 break;
 966         case IB_CM_REP_SENT:
 967         case IB_CM_MRA_REP_RCVD:
 968                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 969                 /* Fall through */
 970         case IB_CM_MRA_REQ_SENT:
 971         case IB_CM_REP_RCVD:
 972         case IB_CM_MRA_REP_SENT:
 973                 spin_unlock_irq(&cm_id_priv->lock);
 974                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 975                                NULL, 0, NULL, 0);
 976                 break;
 977         case IB_CM_ESTABLISHED:
 978                 spin_unlock_irq(&cm_id_priv->lock);
 979                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT)
 980                         break;
 981                 ib_send_cm_dreq(cm_id, NULL, 0);
 982                 goto retest;
 983         case IB_CM_DREQ_SENT:
 984                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 985                 cm_enter_timewait(cm_id_priv);
 986                 spin_unlock_irq(&cm_id_priv->lock);
 987                 break;
 988         case IB_CM_DREQ_RCVD:
 989                 spin_unlock_irq(&cm_id_priv->lock);
 990                 ib_send_cm_drep(cm_id, NULL, 0);
 991                 break;
 992         default:
 993                 spin_unlock_irq(&cm_id_priv->lock);
 994                 break;
 995         }
 996
 997         spin_lock_irq(&cm.lock);
 998         if (!list_empty(&cm_id_priv->altr_list) &&
 999             (!cm_id_priv->altr_send_port_not_ready))
1000                 list_del(&cm_id_priv->altr_list);
1001         if (!list_empty(&cm_id_priv->prim_list) &&
1002             (!cm_id_priv->prim_send_port_not_ready))
1003                 list_del(&cm_id_priv->prim_list);
1004         spin_unlock_irq(&cm.lock);
1005
1006         cm_free_id(cm_id->local_id);
1007         cm_deref_id(cm_id_priv);
1008         wait_for_completion(&cm_id_priv->comp);
1009         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
1010                 cm_free_work(work);
1011         kfree(cm_id_priv->private_data);
1012         kfree(cm_id_priv);
1013 }
1014
1015 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
1016 {
1017         cm_destroy_id(cm_id, 0);
1018 }
1019 EXPORT_SYMBOL(ib_destroy_cm_id);
1020
1021 /**
1022  * __ib_cm_listen - Initiates listening on the specified service ID for
1023  *   connection and service ID resolution requests.
1024  * @cm_id: Connection identifier associated with the listen request.
1025  * @service_id: Service identifier matched against incoming connection
1026  *   and service ID resolution requests.  The service ID should be specified
1027  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1028  *   assign a service ID to the caller.
1029  * @service_mask: Mask applied to service ID used to listen across a
1030  *   range of service IDs.  If set to 0, the service ID is matched
1031  *   exactly.  This parameter is ignored if %service_id is set to
1032  *   IB_CM_ASSIGN_SERVICE_ID.
1033  */
1034 static int __ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id,
1035                           __be64 service_mask)
1036 {
1037         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
1038         int ret = 0;
1039
1040         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
1041         service_id &= service_mask;
1042         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
1043             (service_id != IB_CM_ASSIGN_SERVICE_ID))
1044                 return -EINVAL;
1045
1046         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1047         if (cm_id->state != IB_CM_IDLE)
1048                 return -EINVAL;
1049
1050         cm_id->state = IB_CM_LISTEN;
1051         ++cm_id_priv->listen_sharecount;
1052
1053         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
1054                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
1055                 cm_id->service_mask = ~cpu_to_be64(0);
1056         } else {
1057                 cm_id->service_id = service_id;
1058                 cm_id->service_mask = service_mask;
1059         }
1060         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
1061
1062         if (cur_cm_id_priv) {
1063                 cm_id->state = IB_CM_IDLE;
1064                 --cm_id_priv->listen_sharecount;
1065                 ret = -EBUSY;
1066         }
1067         return ret;
1068 }
1069
1070 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask)
1071 {
1072         unsigned long flags;
1073         int ret;
1074
1075         spin_lock_irqsave(&cm.lock, flags);
1076         ret = __ib_cm_listen(cm_id, service_id, service_mask);
1077         spin_unlock_irqrestore(&cm.lock, flags);
1078
1079         return ret;
1080 }
1081 EXPORT_SYMBOL(ib_cm_listen);
1082
1083 /**
1084  * Create a new listening ib_cm_id and listen on the given service ID.
1085  *
1086  * If there's an existing ID listening on that same device and service ID,
1087  * return it.
1088  *
1089  * @device: Device associated with the cm_id.  All related communication will
1090  * be associated with the specified device.
1091  * @cm_handler: Callback invoked to notify the user of CM events.
1092  * @service_id: Service identifier matched against incoming connection
1093  *   and service ID resolution requests.  The service ID should be specified
1094  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1095  *   assign a service ID to the caller.
1096  *
1097  * Callers should call ib_destroy_cm_id when done with the listener ID.
1098  */
1099 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1100                                      ib_cm_handler cm_handler,
1101                                      __be64 service_id)
1102 {
1103         struct cm_id_private *cm_id_priv;
1104         struct ib_cm_id *cm_id;
1105         unsigned long flags;
1106         int err = 0;
1107
1108         /* Create an ID in advance, since the creation may sleep */
1109         cm_id = ib_create_cm_id(device, cm_handler, NULL);
1110         if (IS_ERR(cm_id))
1111                 return cm_id;
1112
1113         spin_lock_irqsave(&cm.lock, flags);
1114
1115         if (service_id == IB_CM_ASSIGN_SERVICE_ID)
1116                 goto new_id;
1117
1118         /* Find an existing ID */
1119         cm_id_priv = cm_find_listen(device, service_id);
1120         if (cm_id_priv) {
1121                 if (cm_id->cm_handler != cm_handler || cm_id->context) {
1122                         /* Sharing an ib_cm_id with different handlers is not
1123                          * supported */
1124                         spin_unlock_irqrestore(&cm.lock, flags);
1125                         return ERR_PTR(-EINVAL);
1126                 }
1127                 atomic_inc(&cm_id_priv->refcount);
1128                 ++cm_id_priv->listen_sharecount;
1129                 spin_unlock_irqrestore(&cm.lock, flags);
1130
1131                 ib_destroy_cm_id(cm_id);
1132                 cm_id = &cm_id_priv->id;
1133                 return cm_id;
1134         }
1135
1136 new_id:
1137         /* Use newly created ID */
1138         err = __ib_cm_listen(cm_id, service_id, 0);
1139
1140         spin_unlock_irqrestore(&cm.lock, flags);
1141
1142         if (err) {
1143                 ib_destroy_cm_id(cm_id);
1144                 return ERR_PTR(err);
1145         }
1146         return cm_id;
1147 }
1148 EXPORT_SYMBOL(ib_cm_insert_listen);
1149
1150 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
1151                           enum cm_msg_sequence msg_seq)
1152 {
1153         u64 hi_tid, low_tid;
1154
1155         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1156         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
1157                           (msg_seq << 30));
1158         return cpu_to_be64(hi_tid | low_tid);
1159 }
1160
1161 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1162                               __be16 attr_id, __be64 tid)
1163 {
1164         hdr->base_version  = IB_MGMT_BASE_VERSION;
1165         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1166         hdr->class_version = IB_CM_CLASS_VERSION;
1167         hdr->method        = IB_MGMT_METHOD_SEND;
1168         hdr->attr_id       = attr_id;
1169         hdr->tid           = tid;
1170 }
1171
1172 static void cm_format_req(struct cm_req_msg *req_msg,
1173                           struct cm_id_private *cm_id_priv,
1174                           struct ib_cm_req_param *param)
1175 {
1176         struct sa_path_rec *pri_path = param->primary_path;
1177         struct sa_path_rec *alt_path = param->alternate_path;
1178
1179         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1180                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1181
1182         req_msg->local_comm_id = cm_id_priv->id.local_id;
1183         req_msg->service_id = param->service_id;
1184         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1185         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1186         cm_req_set_init_depth(req_msg, param->initiator_depth);
1187         cm_req_set_remote_resp_timeout(req_msg,
1188                                        param->remote_cm_response_timeout);
1189         cm_req_set_qp_type(req_msg, param->qp_type);
1190         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1191         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1192         cm_req_set_local_resp_timeout(req_msg,
1193                                       param->local_cm_response_timeout);
1194         req_msg->pkey = param->primary_path->pkey;
1195         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1196         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1197
1198         if (param->qp_type != IB_QPT_XRC_INI) {
1199                 cm_req_set_resp_res(req_msg, param->responder_resources);
1200                 cm_req_set_retry_count(req_msg, param->retry_count);
1201                 cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1202                 cm_req_set_srq(req_msg, param->srq);
1203         }
1204
1205         if (pri_path->hop_limit <= 1) {
1206                 req_msg->primary_local_lid =
1207                         htons(ntohl(sa_path_get_slid(pri_path)));
1208                 req_msg->primary_remote_lid =
1209                         htons(ntohl(sa_path_get_dlid(pri_path)));
1210         } else {
1211                 /* Work-around until there's a way to obtain remote LID info */
1212                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1213                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1214         }
1215         req_msg->primary_local_gid = pri_path->sgid;
1216         req_msg->primary_remote_gid = pri_path->dgid;
1217         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1218         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1219         req_msg->primary_traffic_class = pri_path->traffic_class;
1220         req_msg->primary_hop_limit = pri_path->hop_limit;
1221         cm_req_set_primary_sl(req_msg, pri_path->sl);
1222         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1223         cm_req_set_primary_local_ack_timeout(req_msg,
1224                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1225                                pri_path->packet_life_time));
1226
1227         if (alt_path) {
1228                 if (alt_path->hop_limit <= 1) {
1229                         req_msg->alt_local_lid =
1230                                 htons(ntohl(sa_path_get_slid(alt_path)));
1231                         req_msg->alt_remote_lid =
1232                                 htons(ntohl(sa_path_get_dlid(alt_path)));
1233                 } else {
1234                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1235                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1236                 }
1237                 req_msg->alt_local_gid = alt_path->sgid;
1238                 req_msg->alt_remote_gid = alt_path->dgid;
1239                 cm_req_set_alt_flow_label(req_msg,
1240                                           alt_path->flow_label);
1241                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1242                 req_msg->alt_traffic_class = alt_path->traffic_class;
1243                 req_msg->alt_hop_limit = alt_path->hop_limit;
1244                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1245                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1246                 cm_req_set_alt_local_ack_timeout(req_msg,
1247                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1248                                        alt_path->packet_life_time));
1249         }
1250
1251         if (param->private_data && param->private_data_len)
1252                 memcpy(req_msg->private_data, param->private_data,
1253                        param->private_data_len);
1254 }
1255
1256 static int cm_validate_req_param(struct ib_cm_req_param *param)
1257 {
1258         /* peer-to-peer not supported */
1259         if (param->peer_to_peer)
1260                 return -EINVAL;
1261
1262         if (!param->primary_path)
1263                 return -EINVAL;
1264
1265         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1266             param->qp_type != IB_QPT_XRC_INI)
1267                 return -EINVAL;
1268
1269         if (param->private_data &&
1270             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1271                 return -EINVAL;
1272
1273         if (param->alternate_path &&
1274             (param->alternate_path->pkey != param->primary_path->pkey ||
1275              param->alternate_path->mtu != param->primary_path->mtu))
1276                 return -EINVAL;
1277
1278         return 0;
1279 }
1280
1281 int ib_send_cm_req(struct ib_cm_id *cm_id,
1282                    struct ib_cm_req_param *param)
1283 {
1284         struct cm_id_private *cm_id_priv;
1285         struct cm_req_msg *req_msg;
1286         unsigned long flags;
1287         int ret;
1288
1289         ret = cm_validate_req_param(param);
1290         if (ret)
1291                 return ret;
1292
1293         /* Verify that we're not in timewait. */
1294         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1295         spin_lock_irqsave(&cm_id_priv->lock, flags);
1296         if (cm_id->state != IB_CM_IDLE) {
1297                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1298                 ret = -EINVAL;
1299                 goto out;
1300         }
1301         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1302
1303         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1304                                                             id.local_id);
1305         if (IS_ERR(cm_id_priv->timewait_info)) {
1306                 ret = PTR_ERR(cm_id_priv->timewait_info);
1307                 goto out;
1308         }
1309
1310         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av,
1311                                  cm_id_priv);
1312         if (ret)
1313                 goto error1;
1314         if (param->alternate_path) {
1315                 ret = cm_init_av_by_path(param->alternate_path,
1316                                          &cm_id_priv->alt_av, cm_id_priv);
1317                 if (ret)
1318                         goto error1;
1319         }
1320         cm_id->service_id = param->service_id;
1321         cm_id->service_mask = ~cpu_to_be64(0);
1322         cm_id_priv->timeout_ms = cm_convert_to_ms(
1323                                     param->primary_path->packet_life_time) * 2 +
1324                                  cm_convert_to_ms(
1325                                     param->remote_cm_response_timeout);
1326         cm_id_priv->max_cm_retries = param->max_cm_retries;
1327         cm_id_priv->initiator_depth = param->initiator_depth;
1328         cm_id_priv->responder_resources = param->responder_resources;
1329         cm_id_priv->retry_count = param->retry_count;
1330         cm_id_priv->path_mtu = param->primary_path->mtu;
1331         cm_id_priv->pkey = param->primary_path->pkey;
1332         cm_id_priv->qp_type = param->qp_type;
1333
1334         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1335         if (ret)
1336                 goto error1;
1337
1338         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1339         cm_format_req(req_msg, cm_id_priv, param);
1340         cm_id_priv->tid = req_msg->hdr.tid;
1341         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1342         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1343
1344         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1345         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1346
1347         spin_lock_irqsave(&cm_id_priv->lock, flags);
1348         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1349         if (ret) {
1350                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1351                 goto error2;
1352         }
1353         BUG_ON(cm_id->state != IB_CM_IDLE);
1354         cm_id->state = IB_CM_REQ_SENT;
1355         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1356         return 0;
1357
1358 error2: cm_free_msg(cm_id_priv->msg);
1359 error1: kfree(cm_id_priv->timewait_info);
1360 out:    return ret;
1361 }
1362 EXPORT_SYMBOL(ib_send_cm_req);
1363
1364 static int cm_issue_rej(struct cm_port *port,
1365                         struct ib_mad_recv_wc *mad_recv_wc,
1366                         enum ib_cm_rej_reason reason,
1367                         enum cm_msg_response msg_rejected,
1368                         void *ari, u8 ari_length)
1369 {
1370         struct ib_mad_send_buf *msg = NULL;
1371         struct cm_rej_msg *rej_msg, *rcv_msg;
1372         int ret;
1373
1374         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1375         if (ret)
1376                 return ret;
1377
1378         /* We just need common CM header information.  Cast to any message. */
1379         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1380         rej_msg = (struct cm_rej_msg *) msg->mad;
1381
1382         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1383         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1384         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1385         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1386         rej_msg->reason = cpu_to_be16(reason);
1387
1388         if (ari && ari_length) {
1389                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1390                 memcpy(rej_msg->ari, ari, ari_length);
1391         }
1392
1393         ret = ib_post_send_mad(msg, NULL);
1394         if (ret)
1395                 cm_free_msg(msg);
1396
1397         return ret;
1398 }
1399
1400 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1401                                     __be32 local_qpn, __be32 remote_qpn)
1402 {
1403         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1404                 ((local_ca_guid == remote_ca_guid) &&
1405                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1406 }
1407
1408 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1409                                      struct sa_path_rec *primary_path,
1410                                      struct sa_path_rec *alt_path)
1411 {
1412         primary_path->dgid = req_msg->primary_local_gid;
1413         primary_path->sgid = req_msg->primary_remote_gid;
1414         sa_path_set_dlid(primary_path,
1415                          htonl(ntohs(req_msg->primary_local_lid)));
1416         sa_path_set_slid(primary_path,
1417                          htonl(ntohs(req_msg->primary_remote_lid)));
1418         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1419         primary_path->hop_limit = req_msg->primary_hop_limit;
1420         primary_path->traffic_class = req_msg->primary_traffic_class;
1421         primary_path->reversible = 1;
1422         primary_path->pkey = req_msg->pkey;
1423         primary_path->sl = cm_req_get_primary_sl(req_msg);
1424         primary_path->mtu_selector = IB_SA_EQ;
1425         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1426         primary_path->rate_selector = IB_SA_EQ;
1427         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1428         primary_path->packet_life_time_selector = IB_SA_EQ;
1429         primary_path->packet_life_time =
1430                 cm_req_get_primary_local_ack_timeout(req_msg);
1431         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1432         primary_path->service_id = req_msg->service_id;
1433
1434         if (req_msg->alt_local_lid) {
1435                 alt_path->dgid = req_msg->alt_local_gid;
1436                 alt_path->sgid = req_msg->alt_remote_gid;
1437                 sa_path_set_dlid(alt_path,
1438                                  htonl(ntohs(req_msg->alt_local_lid)));
1439                 sa_path_set_slid(alt_path,
1440                                  htonl(ntohs(req_msg->alt_remote_lid)));
1441                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1442                 alt_path->hop_limit = req_msg->alt_hop_limit;
1443                 alt_path->traffic_class = req_msg->alt_traffic_class;
1444                 alt_path->reversible = 1;
1445                 alt_path->pkey = req_msg->pkey;
1446                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1447                 alt_path->mtu_selector = IB_SA_EQ;
1448                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1449                 alt_path->rate_selector = IB_SA_EQ;
1450                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1451                 alt_path->packet_life_time_selector = IB_SA_EQ;
1452                 alt_path->packet_life_time =
1453                         cm_req_get_alt_local_ack_timeout(req_msg);
1454                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1455                 alt_path->service_id = req_msg->service_id;
1456         }
1457 }
1458
1459 static u16 cm_get_bth_pkey(struct cm_work *work)
1460 {
1461         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1462         u8 port_num = work->port->port_num;
1463         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1464         u16 pkey;
1465         int ret;
1466
1467         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1468         if (ret) {
1469                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1470                                      port_num, pkey_index, ret);
1471                 return 0;
1472         }
1473
1474         return pkey;
1475 }
1476
1477 static void cm_format_req_event(struct cm_work *work,
1478                                 struct cm_id_private *cm_id_priv,
1479                                 struct ib_cm_id *listen_id)
1480 {
1481         struct cm_req_msg *req_msg;
1482         struct ib_cm_req_event_param *param;
1483
1484         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1485         param = &work->cm_event.param.req_rcvd;
1486         param->listen_id = listen_id;
1487         param->bth_pkey = cm_get_bth_pkey(work);
1488         param->port = cm_id_priv->av.port->port_num;
1489         param->primary_path = &work->path[0];
1490         if (req_msg->alt_local_lid)
1491                 param->alternate_path = &work->path[1];
1492         else
1493                 param->alternate_path = NULL;
1494         param->remote_ca_guid = req_msg->local_ca_guid;
1495         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1496         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1497         param->qp_type = cm_req_get_qp_type(req_msg);
1498         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1499         param->responder_resources = cm_req_get_init_depth(req_msg);
1500         param->initiator_depth = cm_req_get_resp_res(req_msg);
1501         param->local_cm_response_timeout =
1502                                         cm_req_get_remote_resp_timeout(req_msg);
1503         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1504         param->remote_cm_response_timeout =
1505                                         cm_req_get_local_resp_timeout(req_msg);
1506         param->retry_count = cm_req_get_retry_count(req_msg);
1507         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1508         param->srq = cm_req_get_srq(req_msg);
1509         work->cm_event.private_data = &req_msg->private_data;
1510 }
1511
1512 static void cm_process_work(struct cm_id_private *cm_id_priv,
1513                             struct cm_work *work)
1514 {
1515         int ret;
1516
1517         /* We will typically only have the current event to report. */
1518         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1519         cm_free_work(work);
1520
1521         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1522                 spin_lock_irq(&cm_id_priv->lock);
1523                 work = cm_dequeue_work(cm_id_priv);
1524                 spin_unlock_irq(&cm_id_priv->lock);
1525                 BUG_ON(!work);
1526                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1527                                                 &work->cm_event);
1528                 cm_free_work(work);
1529         }
1530         cm_deref_id(cm_id_priv);
1531         if (ret)
1532                 cm_destroy_id(&cm_id_priv->id, ret);
1533 }
1534
1535 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1536                           struct cm_id_private *cm_id_priv,
1537                           enum cm_msg_response msg_mraed, u8 service_timeout,
1538                           const void *private_data, u8 private_data_len)
1539 {
1540         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1541         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1542         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1543         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1544         cm_mra_set_service_timeout(mra_msg, service_timeout);
1545
1546         if (private_data && private_data_len)
1547                 memcpy(mra_msg->private_data, private_data, private_data_len);
1548 }
1549
1550 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1551                           struct cm_id_private *cm_id_priv,
1552                           enum ib_cm_rej_reason reason,
1553                           void *ari,
1554                           u8 ari_length,
1555                           const void *private_data,
1556                           u8 private_data_len)
1557 {
1558         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1559         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1560
1561         switch(cm_id_priv->id.state) {
1562         case IB_CM_REQ_RCVD:
1563                 rej_msg->local_comm_id = 0;
1564                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1565                 break;
1566         case IB_CM_MRA_REQ_SENT:
1567                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1568                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1569                 break;
1570         case IB_CM_REP_RCVD:
1571         case IB_CM_MRA_REP_SENT:
1572                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1573                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1574                 break;
1575         default:
1576                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1577                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1578                 break;
1579         }
1580
1581         rej_msg->reason = cpu_to_be16(reason);
1582         if (ari && ari_length) {
1583                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1584                 memcpy(rej_msg->ari, ari, ari_length);
1585         }
1586
1587         if (private_data && private_data_len)
1588                 memcpy(rej_msg->private_data, private_data, private_data_len);
1589 }
1590
1591 static void cm_dup_req_handler(struct cm_work *work,
1592                                struct cm_id_private *cm_id_priv)
1593 {
1594         struct ib_mad_send_buf *msg = NULL;
1595         int ret;
1596
1597         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1598                         counter[CM_REQ_COUNTER]);
1599
1600         /* Quick state check to discard duplicate REQs. */
1601         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1602                 return;
1603
1604         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1605         if (ret)
1606                 return;
1607
1608         spin_lock_irq(&cm_id_priv->lock);
1609         switch (cm_id_priv->id.state) {
1610         case IB_CM_MRA_REQ_SENT:
1611                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1612                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1613                               cm_id_priv->private_data,
1614                               cm_id_priv->private_data_len);
1615                 break;
1616         case IB_CM_TIMEWAIT:
1617                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1618                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1619                 break;
1620         default:
1621                 goto unlock;
1622         }
1623         spin_unlock_irq(&cm_id_priv->lock);
1624
1625         ret = ib_post_send_mad(msg, NULL);
1626         if (ret)
1627                 goto free;
1628         return;
1629
1630 unlock: spin_unlock_irq(&cm_id_priv->lock);
1631 free:   cm_free_msg(msg);
1632 }
1633
1634 static struct cm_id_private * cm_match_req(struct cm_work *work,
1635                                            struct cm_id_private *cm_id_priv)
1636 {
1637         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1638         struct cm_timewait_info *timewait_info;
1639         struct cm_req_msg *req_msg;
1640         struct ib_cm_id *cm_id;
1641
1642         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1643
1644         /* Check for possible duplicate REQ. */
1645         spin_lock_irq(&cm.lock);
1646         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1647         if (timewait_info) {
1648                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1649                                            timewait_info->work.remote_id);
1650                 spin_unlock_irq(&cm.lock);
1651                 if (cur_cm_id_priv) {
1652                         cm_dup_req_handler(work, cur_cm_id_priv);
1653                         cm_deref_id(cur_cm_id_priv);
1654                 }
1655                 return NULL;
1656         }
1657
1658         /* Check for stale connections. */
1659         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1660         if (timewait_info) {
1661                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1662                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1663                                            timewait_info->work.remote_id);
1664
1665                 spin_unlock_irq(&cm.lock);
1666                 cm_issue_rej(work->port, work->mad_recv_wc,
1667                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1668                              NULL, 0);
1669                 if (cur_cm_id_priv) {
1670                         cm_id = &cur_cm_id_priv->id;
1671                         ib_send_cm_dreq(cm_id, NULL, 0);
1672                         cm_deref_id(cur_cm_id_priv);
1673                 }
1674                 return NULL;
1675         }
1676
1677         /* Find matching listen request. */
1678         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1679                                            req_msg->service_id);
1680         if (!listen_cm_id_priv) {
1681                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1682                 spin_unlock_irq(&cm.lock);
1683                 cm_issue_rej(work->port, work->mad_recv_wc,
1684                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1685                              NULL, 0);
1686                 goto out;
1687         }
1688         atomic_inc(&listen_cm_id_priv->refcount);
1689         atomic_inc(&cm_id_priv->refcount);
1690         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1691         atomic_inc(&cm_id_priv->work_count);
1692         spin_unlock_irq(&cm.lock);
1693 out:
1694         return listen_cm_id_priv;
1695 }
1696
1697 /*
1698  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1699  * we need to override the LID/SL data in the REQ with the LID information
1700  * in the work completion.
1701  */
1702 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1703 {
1704         if (!cm_req_get_primary_subnet_local(req_msg)) {
1705                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1706                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1707                         cm_req_set_primary_sl(req_msg, wc->sl);
1708                 }
1709
1710                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1711                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1712         }
1713
1714         if (!cm_req_get_alt_subnet_local(req_msg)) {
1715                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1716                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1717                         cm_req_set_alt_sl(req_msg, wc->sl);
1718                 }
1719
1720                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1721                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1722         }
1723 }
1724
1725 static int cm_req_handler(struct cm_work *work)
1726 {
1727         struct ib_cm_id *cm_id;
1728         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1729         struct cm_req_msg *req_msg;
1730         union ib_gid gid;
1731         struct ib_gid_attr gid_attr;
1732         const struct ib_global_route *grh;
1733         int ret;
1734
1735         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1736
1737         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1738         if (IS_ERR(cm_id))
1739                 return PTR_ERR(cm_id);
1740
1741         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1742         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1743         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1744                                 work->mad_recv_wc->recv_buf.grh,
1745                                 &cm_id_priv->av);
1746         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1747                                                             id.local_id);
1748         if (IS_ERR(cm_id_priv->timewait_info)) {
1749                 ret = PTR_ERR(cm_id_priv->timewait_info);
1750                 goto destroy;
1751         }
1752         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1753         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1754         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1755
1756         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1757         if (!listen_cm_id_priv) {
1758                 ret = -EINVAL;
1759                 kfree(cm_id_priv->timewait_info);
1760                 goto destroy;
1761         }
1762
1763         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1764         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1765         cm_id_priv->id.service_id = req_msg->service_id;
1766         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1767
1768         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1769
1770         memset(&work->path[0], 0, sizeof(work->path[0]));
1771         memset(&work->path[1], 0, sizeof(work->path[1]));
1772         grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
1773         ret = ib_get_cached_gid(work->port->cm_dev->ib_device,
1774                                 work->port->port_num,
1775                                 grh->sgid_index,
1776                                 &gid, &gid_attr);
1777         if (!ret) {
1778                 if (gid_attr.ndev) {
1779                         work->path[0].rec_type =
1780                                 sa_conv_gid_to_pathrec_type(gid_attr.gid_type);
1781                         sa_path_set_ifindex(&work->path[0],
1782                                             gid_attr.ndev->ifindex);
1783                         sa_path_set_ndev(&work->path[0],
1784                                          dev_net(gid_attr.ndev));
1785                         dev_put(gid_attr.ndev);
1786                 } else {
1787                         work->path[0].rec_type = SA_PATH_REC_TYPE_IB;
1788                 }
1789                 if (req_msg->alt_local_lid)
1790                         work->path[1].rec_type = work->path[0].rec_type;
1791                 cm_format_paths_from_req(req_msg, &work->path[0],
1792                                          &work->path[1]);
1793                 if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
1794                         sa_path_set_dmac(&work->path[0],
1795                                          cm_id_priv->av.ah_attr.roce.dmac);
1796                 work->path[0].hop_limit = grh->hop_limit;
1797                 ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av,
1798                                          cm_id_priv);
1799         }
1800         if (ret) {
1801                 int err = ib_get_cached_gid(work->port->cm_dev->ib_device,
1802                                             work->port->port_num, 0,
1803                                             &work->path[0].sgid,
1804                                             &gid_attr);
1805                 if (!err && gid_attr.ndev) {
1806                         work->path[0].rec_type =
1807                                 sa_conv_gid_to_pathrec_type(gid_attr.gid_type);
1808                         sa_path_set_ifindex(&work->path[0],
1809                                             gid_attr.ndev->ifindex);
1810                         sa_path_set_ndev(&work->path[0],
1811                                          dev_net(gid_attr.ndev));
1812                         dev_put(gid_attr.ndev);
1813                 } else {
1814                         work->path[0].rec_type = SA_PATH_REC_TYPE_IB;
1815                 }
1816                 if (req_msg->alt_local_lid)
1817                         work->path[1].rec_type = work->path[0].rec_type;
1818                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1819                                &work->path[0].sgid, sizeof work->path[0].sgid,
1820                                NULL, 0);
1821                 goto rejected;
1822         }
1823         if (req_msg->alt_local_lid) {
1824                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av,
1825                                          cm_id_priv);
1826                 if (ret) {
1827                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1828                                        &work->path[0].sgid,
1829                                        sizeof work->path[0].sgid, NULL, 0);
1830                         goto rejected;
1831                 }
1832         }
1833         cm_id_priv->tid = req_msg->hdr.tid;
1834         cm_id_priv->timeout_ms = cm_convert_to_ms(
1835                                         cm_req_get_local_resp_timeout(req_msg));
1836         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1837         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1838         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1839         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1840         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1841         cm_id_priv->pkey = req_msg->pkey;
1842         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1843         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1844         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1845         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1846
1847         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1848         cm_process_work(cm_id_priv, work);
1849         cm_deref_id(listen_cm_id_priv);
1850         return 0;
1851
1852 rejected:
1853         atomic_dec(&cm_id_priv->refcount);
1854         cm_deref_id(listen_cm_id_priv);
1855 destroy:
1856         ib_destroy_cm_id(cm_id);
1857         return ret;
1858 }
1859
1860 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1861                           struct cm_id_private *cm_id_priv,
1862                           struct ib_cm_rep_param *param)
1863 {
1864         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1865         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1866         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1867         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1868         rep_msg->resp_resources = param->responder_resources;
1869         cm_rep_set_target_ack_delay(rep_msg,
1870                                     cm_id_priv->av.port->cm_dev->ack_delay);
1871         cm_rep_set_failover(rep_msg, param->failover_accepted);
1872         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1873         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1874
1875         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1876                 rep_msg->initiator_depth = param->initiator_depth;
1877                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1878                 cm_rep_set_srq(rep_msg, param->srq);
1879                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1880         } else {
1881                 cm_rep_set_srq(rep_msg, 1);
1882                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1883         }
1884
1885         if (param->private_data && param->private_data_len)
1886                 memcpy(rep_msg->private_data, param->private_data,
1887                        param->private_data_len);
1888 }
1889
1890 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1891                    struct ib_cm_rep_param *param)
1892 {
1893         struct cm_id_private *cm_id_priv;
1894         struct ib_mad_send_buf *msg;
1895         struct cm_rep_msg *rep_msg;
1896         unsigned long flags;
1897         int ret;
1898
1899         if (param->private_data &&
1900             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1901                 return -EINVAL;
1902
1903         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1904         spin_lock_irqsave(&cm_id_priv->lock, flags);
1905         if (cm_id->state != IB_CM_REQ_RCVD &&
1906             cm_id->state != IB_CM_MRA_REQ_SENT) {
1907                 ret = -EINVAL;
1908                 goto out;
1909         }
1910
1911         ret = cm_alloc_msg(cm_id_priv, &msg);
1912         if (ret)
1913                 goto out;
1914
1915         rep_msg = (struct cm_rep_msg *) msg->mad;
1916         cm_format_rep(rep_msg, cm_id_priv, param);
1917         msg->timeout_ms = cm_id_priv->timeout_ms;
1918         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1919
1920         ret = ib_post_send_mad(msg, NULL);
1921         if (ret) {
1922                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1923                 cm_free_msg(msg);
1924                 return ret;
1925         }
1926
1927         cm_id->state = IB_CM_REP_SENT;
1928         cm_id_priv->msg = msg;
1929         cm_id_priv->initiator_depth = param->initiator_depth;
1930         cm_id_priv->responder_resources = param->responder_resources;
1931         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1932         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
1933
1934 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1935         return ret;
1936 }
1937 EXPORT_SYMBOL(ib_send_cm_rep);
1938
1939 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1940                           struct cm_id_private *cm_id_priv,
1941                           const void *private_data,
1942                           u8 private_data_len)
1943 {
1944         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1945         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1946         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1947
1948         if (private_data && private_data_len)
1949                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1950 }
1951
1952 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1953                    const void *private_data,
1954                    u8 private_data_len)
1955 {
1956         struct cm_id_private *cm_id_priv;
1957         struct ib_mad_send_buf *msg;
1958         unsigned long flags;
1959         void *data;
1960         int ret;
1961
1962         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1963                 return -EINVAL;
1964
1965         data = cm_copy_private_data(private_data, private_data_len);
1966         if (IS_ERR(data))
1967                 return PTR_ERR(data);
1968
1969         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1970         spin_lock_irqsave(&cm_id_priv->lock, flags);
1971         if (cm_id->state != IB_CM_REP_RCVD &&
1972             cm_id->state != IB_CM_MRA_REP_SENT) {
1973                 ret = -EINVAL;
1974                 goto error;
1975         }
1976
1977         ret = cm_alloc_msg(cm_id_priv, &msg);
1978         if (ret)
1979                 goto error;
1980
1981         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1982                       private_data, private_data_len);
1983
1984         ret = ib_post_send_mad(msg, NULL);
1985         if (ret) {
1986                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1987                 cm_free_msg(msg);
1988                 kfree(data);
1989                 return ret;
1990         }
1991
1992         cm_id->state = IB_CM_ESTABLISHED;
1993         cm_set_private_data(cm_id_priv, data, private_data_len);
1994         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1995         return 0;
1996
1997 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1998         kfree(data);
1999         return ret;
2000 }
2001 EXPORT_SYMBOL(ib_send_cm_rtu);
2002
2003 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
2004 {
2005         struct cm_rep_msg *rep_msg;
2006         struct ib_cm_rep_event_param *param;
2007
2008         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2009         param = &work->cm_event.param.rep_rcvd;
2010         param->remote_ca_guid = rep_msg->local_ca_guid;
2011         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
2012         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
2013         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
2014         param->responder_resources = rep_msg->initiator_depth;
2015         param->initiator_depth = rep_msg->resp_resources;
2016         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2017         param->failover_accepted = cm_rep_get_failover(rep_msg);
2018         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
2019         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2020         param->srq = cm_rep_get_srq(rep_msg);
2021         work->cm_event.private_data = &rep_msg->private_data;
2022 }
2023
2024 static void cm_dup_rep_handler(struct cm_work *work)
2025 {
2026         struct cm_id_private *cm_id_priv;
2027         struct cm_rep_msg *rep_msg;
2028         struct ib_mad_send_buf *msg = NULL;
2029         int ret;
2030
2031         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2032         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
2033                                    rep_msg->local_comm_id);
2034         if (!cm_id_priv)
2035                 return;
2036
2037         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2038                         counter[CM_REP_COUNTER]);
2039         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2040         if (ret)
2041                 goto deref;
2042
2043         spin_lock_irq(&cm_id_priv->lock);
2044         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2045                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2046                               cm_id_priv->private_data,
2047                               cm_id_priv->private_data_len);
2048         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2049                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2050                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2051                               cm_id_priv->private_data,
2052                               cm_id_priv->private_data_len);
2053         else
2054                 goto unlock;
2055         spin_unlock_irq(&cm_id_priv->lock);
2056
2057         ret = ib_post_send_mad(msg, NULL);
2058         if (ret)
2059                 goto free;
2060         goto deref;
2061
2062 unlock: spin_unlock_irq(&cm_id_priv->lock);
2063 free:   cm_free_msg(msg);
2064 deref:  cm_deref_id(cm_id_priv);
2065 }
2066
2067 static int cm_rep_handler(struct cm_work *work)
2068 {
2069         struct cm_id_private *cm_id_priv;
2070         struct cm_rep_msg *rep_msg;
2071         int ret;
2072         struct cm_id_private *cur_cm_id_priv;
2073         struct ib_cm_id *cm_id;
2074         struct cm_timewait_info *timewait_info;
2075
2076         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2077         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
2078         if (!cm_id_priv) {
2079                 cm_dup_rep_handler(work);
2080                 return -EINVAL;
2081         }
2082
2083         cm_format_rep_event(work, cm_id_priv->qp_type);
2084
2085         spin_lock_irq(&cm_id_priv->lock);
2086         switch (cm_id_priv->id.state) {
2087         case IB_CM_REQ_SENT:
2088         case IB_CM_MRA_REQ_RCVD:
2089                 break;
2090         default:
2091                 spin_unlock_irq(&cm_id_priv->lock);
2092                 ret = -EINVAL;
2093                 goto error;
2094         }
2095
2096         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
2097         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
2098         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2099
2100         spin_lock(&cm.lock);
2101         /* Check for duplicate REP. */
2102         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2103                 spin_unlock(&cm.lock);
2104                 spin_unlock_irq(&cm_id_priv->lock);
2105                 ret = -EINVAL;
2106                 goto error;
2107         }
2108         /* Check for a stale connection. */
2109         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2110         if (timewait_info) {
2111                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
2112                          &cm.remote_id_table);
2113                 cm_id_priv->timewait_info->inserted_remote_id = 0;
2114                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
2115                                            timewait_info->work.remote_id);
2116
2117                 spin_unlock(&cm.lock);
2118                 spin_unlock_irq(&cm_id_priv->lock);
2119                 cm_issue_rej(work->port, work->mad_recv_wc,
2120                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2121                              NULL, 0);
2122                 ret = -EINVAL;
2123                 if (cur_cm_id_priv) {
2124                         cm_id = &cur_cm_id_priv->id;
2125                         ib_send_cm_dreq(cm_id, NULL, 0);
2126                         cm_deref_id(cur_cm_id_priv);
2127                 }
2128
2129                 goto error;
2130         }
2131         spin_unlock(&cm.lock);
2132
2133         cm_id_priv->id.state = IB_CM_REP_RCVD;
2134         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
2135         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2136         cm_id_priv->initiator_depth = rep_msg->resp_resources;
2137         cm_id_priv->responder_resources = rep_msg->initiator_depth;
2138         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
2139         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2140         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2141         cm_id_priv->av.timeout =
2142                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2143                                        cm_id_priv->av.timeout - 1);
2144         cm_id_priv->alt_av.timeout =
2145                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2146                                        cm_id_priv->alt_av.timeout - 1);
2147
2148         /* todo: handle peer_to_peer */
2149
2150         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2151         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2152         if (!ret)
2153                 list_add_tail(&work->list, &cm_id_priv->work_list);
2154         spin_unlock_irq(&cm_id_priv->lock);
2155
2156         if (ret)
2157                 cm_process_work(cm_id_priv, work);
2158         else
2159                 cm_deref_id(cm_id_priv);
2160         return 0;
2161
2162 error:
2163         cm_deref_id(cm_id_priv);
2164         return ret;
2165 }
2166
2167 static int cm_establish_handler(struct cm_work *work)
2168 {
2169         struct cm_id_private *cm_id_priv;
2170         int ret;
2171
2172         /* See comment in cm_establish about lookup. */
2173         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2174         if (!cm_id_priv)
2175                 return -EINVAL;
2176
2177         spin_lock_irq(&cm_id_priv->lock);
2178         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2179                 spin_unlock_irq(&cm_id_priv->lock);
2180                 goto out;
2181         }
2182
2183         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2184         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2185         if (!ret)
2186                 list_add_tail(&work->list, &cm_id_priv->work_list);
2187         spin_unlock_irq(&cm_id_priv->lock);
2188
2189         if (ret)
2190                 cm_process_work(cm_id_priv, work);
2191         else
2192                 cm_deref_id(cm_id_priv);
2193         return 0;
2194 out:
2195         cm_deref_id(cm_id_priv);
2196         return -EINVAL;
2197 }
2198
2199 static int cm_rtu_handler(struct cm_work *work)
2200 {
2201         struct cm_id_private *cm_id_priv;
2202         struct cm_rtu_msg *rtu_msg;
2203         int ret;
2204
2205         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2206         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
2207                                    rtu_msg->local_comm_id);
2208         if (!cm_id_priv)
2209                 return -EINVAL;
2210
2211         work->cm_event.private_data = &rtu_msg->private_data;
2212
2213         spin_lock_irq(&cm_id_priv->lock);
2214         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2215             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2216                 spin_unlock_irq(&cm_id_priv->lock);
2217                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2218                                 counter[CM_RTU_COUNTER]);
2219                 goto out;
2220         }
2221         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2222
2223         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2224         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2225         if (!ret)
2226                 list_add_tail(&work->list, &cm_id_priv->work_list);
2227         spin_unlock_irq(&cm_id_priv->lock);
2228
2229         if (ret)
2230                 cm_process_work(cm_id_priv, work);
2231         else
2232                 cm_deref_id(cm_id_priv);
2233         return 0;
2234 out:
2235         cm_deref_id(cm_id_priv);
2236         return -EINVAL;
2237 }
2238
2239 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2240                           struct cm_id_private *cm_id_priv,
2241                           const void *private_data,
2242                           u8 private_data_len)
2243 {
2244         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2245                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
2246         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
2247         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
2248         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2249
2250         if (private_data && private_data_len)
2251                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2252 }
2253
2254 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2255                     const void *private_data,
2256                     u8 private_data_len)
2257 {
2258         struct cm_id_private *cm_id_priv;
2259         struct ib_mad_send_buf *msg;
2260         unsigned long flags;
2261         int ret;
2262
2263         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2264                 return -EINVAL;
2265
2266         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2267         spin_lock_irqsave(&cm_id_priv->lock, flags);
2268         if (cm_id->state != IB_CM_ESTABLISHED) {
2269                 ret = -EINVAL;
2270                 goto out;
2271         }
2272
2273         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2274             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2275                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2276
2277         ret = cm_alloc_msg(cm_id_priv, &msg);
2278         if (ret) {
2279                 cm_enter_timewait(cm_id_priv);
2280                 goto out;
2281         }
2282
2283         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2284                        private_data, private_data_len);
2285         msg->timeout_ms = cm_id_priv->timeout_ms;
2286         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2287
2288         ret = ib_post_send_mad(msg, NULL);
2289         if (ret) {
2290                 cm_enter_timewait(cm_id_priv);
2291                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2292                 cm_free_msg(msg);
2293                 return ret;
2294         }
2295
2296         cm_id->state = IB_CM_DREQ_SENT;
2297         cm_id_priv->msg = msg;
2298 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2299         return ret;
2300 }
2301 EXPORT_SYMBOL(ib_send_cm_dreq);
2302
2303 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2304                           struct cm_id_private *cm_id_priv,
2305                           const void *private_data,
2306                           u8 private_data_len)
2307 {
2308         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2309         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2310         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2311
2312         if (private_data && private_data_len)
2313                 memcpy(drep_msg->private_data, private_data, private_data_len);
2314 }
2315
2316 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2317                     const void *private_data,
2318                     u8 private_data_len)
2319 {
2320         struct cm_id_private *cm_id_priv;
2321         struct ib_mad_send_buf *msg;
2322         unsigned long flags;
2323         void *data;
2324         int ret;
2325
2326         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2327                 return -EINVAL;
2328
2329         data = cm_copy_private_data(private_data, private_data_len);
2330         if (IS_ERR(data))
2331                 return PTR_ERR(data);
2332
2333         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2334         spin_lock_irqsave(&cm_id_priv->lock, flags);
2335         if (cm_id->state != IB_CM_DREQ_RCVD) {
2336                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2337                 kfree(data);
2338                 return -EINVAL;
2339         }
2340
2341         cm_set_private_data(cm_id_priv, data, private_data_len);
2342         cm_enter_timewait(cm_id_priv);
2343
2344         ret = cm_alloc_msg(cm_id_priv, &msg);
2345         if (ret)
2346                 goto out;
2347
2348         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2349                        private_data, private_data_len);
2350
2351         ret = ib_post_send_mad(msg, NULL);
2352         if (ret) {
2353                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2354                 cm_free_msg(msg);
2355                 return ret;
2356         }
2357
2358 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2359         return ret;
2360 }
2361 EXPORT_SYMBOL(ib_send_cm_drep);
2362
2363 static int cm_issue_drep(struct cm_port *port,
2364                          struct ib_mad_recv_wc *mad_recv_wc)
2365 {
2366         struct ib_mad_send_buf *msg = NULL;
2367         struct cm_dreq_msg *dreq_msg;
2368         struct cm_drep_msg *drep_msg;
2369         int ret;
2370
2371         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2372         if (ret)
2373                 return ret;
2374
2375         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2376         drep_msg = (struct cm_drep_msg *) msg->mad;
2377
2378         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2379         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2380         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2381
2382         ret = ib_post_send_mad(msg, NULL);
2383         if (ret)
2384                 cm_free_msg(msg);
2385
2386         return ret;
2387 }
2388
2389 static int cm_dreq_handler(struct cm_work *work)
2390 {
2391         struct cm_id_private *cm_id_priv;
2392         struct cm_dreq_msg *dreq_msg;
2393         struct ib_mad_send_buf *msg = NULL;
2394         int ret;
2395
2396         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2397         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2398                                    dreq_msg->local_comm_id);
2399         if (!cm_id_priv) {
2400                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2401                                 counter[CM_DREQ_COUNTER]);
2402                 cm_issue_drep(work->port, work->mad_recv_wc);
2403                 return -EINVAL;
2404         }
2405
2406         work->cm_event.private_data = &dreq_msg->private_data;
2407
2408         spin_lock_irq(&cm_id_priv->lock);
2409         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2410                 goto unlock;
2411
2412         switch (cm_id_priv->id.state) {
2413         case IB_CM_REP_SENT:
2414         case IB_CM_DREQ_SENT:
2415                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2416                 break;
2417         case IB_CM_ESTABLISHED:
2418                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2419                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2420                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2421                 break;
2422         case IB_CM_MRA_REP_RCVD:
2423                 break;
2424         case IB_CM_TIMEWAIT:
2425                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2426                                 counter[CM_DREQ_COUNTER]);
2427                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2428                         goto unlock;
2429
2430                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2431                                cm_id_priv->private_data,
2432                                cm_id_priv->private_data_len);
2433                 spin_unlock_irq(&cm_id_priv->lock);
2434
2435                 if (ib_post_send_mad(msg, NULL))
2436                         cm_free_msg(msg);
2437                 goto deref;
2438         case IB_CM_DREQ_RCVD:
2439                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2440                                 counter[CM_DREQ_COUNTER]);
2441                 goto unlock;
2442         default:
2443                 goto unlock;
2444         }
2445         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2446         cm_id_priv->tid = dreq_msg->hdr.tid;
2447         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2448         if (!ret)
2449                 list_add_tail(&work->list, &cm_id_priv->work_list);
2450         spin_unlock_irq(&cm_id_priv->lock);
2451
2452         if (ret)
2453                 cm_process_work(cm_id_priv, work);
2454         else
2455                 cm_deref_id(cm_id_priv);
2456         return 0;
2457
2458 unlock: spin_unlock_irq(&cm_id_priv->lock);
2459 deref:  cm_deref_id(cm_id_priv);
2460         return -EINVAL;
2461 }
2462
2463 static int cm_drep_handler(struct cm_work *work)
2464 {
2465         struct cm_id_private *cm_id_priv;
2466         struct cm_drep_msg *drep_msg;
2467         int ret;
2468
2469         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2470         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2471                                    drep_msg->local_comm_id);
2472         if (!cm_id_priv)
2473                 return -EINVAL;
2474
2475         work->cm_event.private_data = &drep_msg->private_data;
2476
2477         spin_lock_irq(&cm_id_priv->lock);
2478         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2479             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2480                 spin_unlock_irq(&cm_id_priv->lock);
2481                 goto out;
2482         }
2483         cm_enter_timewait(cm_id_priv);
2484
2485         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2486         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2487         if (!ret)
2488                 list_add_tail(&work->list, &cm_id_priv->work_list);
2489         spin_unlock_irq(&cm_id_priv->lock);
2490
2491         if (ret)
2492                 cm_process_work(cm_id_priv, work);
2493         else
2494                 cm_deref_id(cm_id_priv);
2495         return 0;
2496 out:
2497         cm_deref_id(cm_id_priv);
2498         return -EINVAL;
2499 }
2500
2501 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2502                    enum ib_cm_rej_reason reason,
2503                    void *ari,
2504                    u8 ari_length,
2505                    const void *private_data,
2506                    u8 private_data_len)
2507 {
2508         struct cm_id_private *cm_id_priv;
2509         struct ib_mad_send_buf *msg;
2510         unsigned long flags;
2511         int ret;
2512
2513         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2514             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2515                 return -EINVAL;
2516
2517         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2518
2519         spin_lock_irqsave(&cm_id_priv->lock, flags);
2520         switch (cm_id->state) {
2521         case IB_CM_REQ_SENT:
2522         case IB_CM_MRA_REQ_RCVD:
2523         case IB_CM_REQ_RCVD:
2524         case IB_CM_MRA_REQ_SENT:
2525         case IB_CM_REP_RCVD:
2526         case IB_CM_MRA_REP_SENT:
2527                 ret = cm_alloc_msg(cm_id_priv, &msg);
2528                 if (!ret)
2529                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2530                                       cm_id_priv, reason, ari, ari_length,
2531                                       private_data, private_data_len);
2532
2533                 cm_reset_to_idle(cm_id_priv);
2534                 break;
2535         case IB_CM_REP_SENT:
2536         case IB_CM_MRA_REP_RCVD:
2537                 ret = cm_alloc_msg(cm_id_priv, &msg);
2538                 if (!ret)
2539                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2540                                       cm_id_priv, reason, ari, ari_length,
2541                                       private_data, private_data_len);
2542
2543                 cm_enter_timewait(cm_id_priv);
2544                 break;
2545         default:
2546                 ret = -EINVAL;
2547                 goto out;
2548         }
2549
2550         if (ret)
2551                 goto out;
2552
2553         ret = ib_post_send_mad(msg, NULL);
2554         if (ret)
2555                 cm_free_msg(msg);
2556
2557 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2558         return ret;
2559 }
2560 EXPORT_SYMBOL(ib_send_cm_rej);
2561
2562 static void cm_format_rej_event(struct cm_work *work)
2563 {
2564         struct cm_rej_msg *rej_msg;
2565         struct ib_cm_rej_event_param *param;
2566
2567         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2568         param = &work->cm_event.param.rej_rcvd;
2569         param->ari = rej_msg->ari;
2570         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2571         param->reason = __be16_to_cpu(rej_msg->reason);
2572         work->cm_event.private_data = &rej_msg->private_data;
2573 }
2574
2575 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2576 {
2577         struct cm_timewait_info *timewait_info;
2578         struct cm_id_private *cm_id_priv;
2579         __be32 remote_id;
2580
2581         remote_id = rej_msg->local_comm_id;
2582
2583         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2584                 spin_lock_irq(&cm.lock);
2585                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2586                                                   remote_id);
2587                 if (!timewait_info) {
2588                         spin_unlock_irq(&cm.lock);
2589                         return NULL;
2590                 }
2591                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2592                                       (timewait_info->work.local_id ^
2593                                        cm.random_id_operand));
2594                 if (cm_id_priv) {
2595                         if (cm_id_priv->id.remote_id == remote_id)
2596                                 atomic_inc(&cm_id_priv->refcount);
2597                         else
2598                                 cm_id_priv = NULL;
2599                 }
2600                 spin_unlock_irq(&cm.lock);
2601         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2602                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2603         else
2604                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2605
2606         return cm_id_priv;
2607 }
2608
2609 static int cm_rej_handler(struct cm_work *work)
2610 {
2611         struct cm_id_private *cm_id_priv;
2612         struct cm_rej_msg *rej_msg;
2613         int ret;
2614
2615         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2616         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2617         if (!cm_id_priv)
2618                 return -EINVAL;
2619
2620         cm_format_rej_event(work);
2621
2622         spin_lock_irq(&cm_id_priv->lock);
2623         switch (cm_id_priv->id.state) {
2624         case IB_CM_REQ_SENT:
2625         case IB_CM_MRA_REQ_RCVD:
2626         case IB_CM_REP_SENT:
2627         case IB_CM_MRA_REP_RCVD:
2628                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2629                 /* fall through */
2630         case IB_CM_REQ_RCVD:
2631         case IB_CM_MRA_REQ_SENT:
2632                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2633                         cm_enter_timewait(cm_id_priv);
2634                 else
2635                         cm_reset_to_idle(cm_id_priv);
2636                 break;
2637         case IB_CM_DREQ_SENT:
2638                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2639                 /* fall through */
2640         case IB_CM_REP_RCVD:
2641         case IB_CM_MRA_REP_SENT:
2642                 cm_enter_timewait(cm_id_priv);
2643                 break;
2644         case IB_CM_ESTABLISHED:
2645                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2646                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2647                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2648                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2649                                               cm_id_priv->msg);
2650                         cm_enter_timewait(cm_id_priv);
2651                         break;
2652                 }
2653                 /* fall through */
2654         default:
2655                 spin_unlock_irq(&cm_id_priv->lock);
2656                 ret = -EINVAL;
2657                 goto out;
2658         }
2659
2660         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2661         if (!ret)
2662                 list_add_tail(&work->list, &cm_id_priv->work_list);
2663         spin_unlock_irq(&cm_id_priv->lock);
2664
2665         if (ret)
2666                 cm_process_work(cm_id_priv, work);
2667         else
2668                 cm_deref_id(cm_id_priv);
2669         return 0;
2670 out:
2671         cm_deref_id(cm_id_priv);
2672         return -EINVAL;
2673 }
2674
2675 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2676                    u8 service_timeout,
2677                    const void *private_data,
2678                    u8 private_data_len)
2679 {
2680         struct cm_id_private *cm_id_priv;
2681         struct ib_mad_send_buf *msg;
2682         enum ib_cm_state cm_state;
2683         enum ib_cm_lap_state lap_state;
2684         enum cm_msg_response msg_response;
2685         void *data;
2686         unsigned long flags;
2687         int ret;
2688
2689         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2690                 return -EINVAL;
2691
2692         data = cm_copy_private_data(private_data, private_data_len);
2693         if (IS_ERR(data))
2694                 return PTR_ERR(data);
2695
2696         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2697
2698         spin_lock_irqsave(&cm_id_priv->lock, flags);
2699         switch(cm_id_priv->id.state) {
2700         case IB_CM_REQ_RCVD:
2701                 cm_state = IB_CM_MRA_REQ_SENT;
2702                 lap_state = cm_id->lap_state;
2703                 msg_response = CM_MSG_RESPONSE_REQ;
2704                 break;
2705         case IB_CM_REP_RCVD:
2706                 cm_state = IB_CM_MRA_REP_SENT;
2707                 lap_state = cm_id->lap_state;
2708                 msg_response = CM_MSG_RESPONSE_REP;
2709                 break;
2710         case IB_CM_ESTABLISHED:
2711                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2712                         cm_state = cm_id->state;
2713                         lap_state = IB_CM_MRA_LAP_SENT;
2714                         msg_response = CM_MSG_RESPONSE_OTHER;
2715                         break;
2716                 }
2717         default:
2718                 ret = -EINVAL;
2719                 goto error1;
2720         }
2721
2722         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2723                 ret = cm_alloc_msg(cm_id_priv, &msg);
2724                 if (ret)
2725                         goto error1;
2726
2727                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2728                               msg_response, service_timeout,
2729                               private_data, private_data_len);
2730                 ret = ib_post_send_mad(msg, NULL);
2731                 if (ret)
2732                         goto error2;
2733         }
2734
2735         cm_id->state = cm_state;
2736         cm_id->lap_state = lap_state;
2737         cm_id_priv->service_timeout = service_timeout;
2738         cm_set_private_data(cm_id_priv, data, private_data_len);
2739         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2740         return 0;
2741
2742 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2743         kfree(data);
2744         return ret;
2745
2746 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2747         kfree(data);
2748         cm_free_msg(msg);
2749         return ret;
2750 }
2751 EXPORT_SYMBOL(ib_send_cm_mra);
2752
2753 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2754 {
2755         switch (cm_mra_get_msg_mraed(mra_msg)) {
2756         case CM_MSG_RESPONSE_REQ:
2757                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2758         case CM_MSG_RESPONSE_REP:
2759         case CM_MSG_RESPONSE_OTHER:
2760                 return cm_acquire_id(mra_msg->remote_comm_id,
2761                                      mra_msg->local_comm_id);
2762         default:
2763                 return NULL;
2764         }
2765 }
2766
2767 static int cm_mra_handler(struct cm_work *work)
2768 {
2769         struct cm_id_private *cm_id_priv;
2770         struct cm_mra_msg *mra_msg;
2771         int timeout, ret;
2772
2773         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2774         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2775         if (!cm_id_priv)
2776                 return -EINVAL;
2777
2778         work->cm_event.private_data = &mra_msg->private_data;
2779         work->cm_event.param.mra_rcvd.service_timeout =
2780                                         cm_mra_get_service_timeout(mra_msg);
2781         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2782                   cm_convert_to_ms(cm_id_priv->av.timeout);
2783
2784         spin_lock_irq(&cm_id_priv->lock);
2785         switch (cm_id_priv->id.state) {
2786         case IB_CM_REQ_SENT:
2787                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2788                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2789                                   cm_id_priv->msg, timeout))
2790                         goto out;
2791                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2792                 break;
2793         case IB_CM_REP_SENT:
2794                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2795                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2796                                   cm_id_priv->msg, timeout))
2797                         goto out;
2798                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2799                 break;
2800         case IB_CM_ESTABLISHED:
2801                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2802                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2803                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2804                                   cm_id_priv->msg, timeout)) {
2805                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2806                                 atomic_long_inc(&work->port->
2807                                                 counter_group[CM_RECV_DUPLICATES].
2808                                                 counter[CM_MRA_COUNTER]);
2809                         goto out;
2810                 }
2811                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2812                 break;
2813         case IB_CM_MRA_REQ_RCVD:
2814         case IB_CM_MRA_REP_RCVD:
2815                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2816                                 counter[CM_MRA_COUNTER]);
2817                 /* fall through */
2818         default:
2819                 goto out;
2820         }
2821
2822         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2823                                       cm_id_priv->id.state;
2824         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2825         if (!ret)
2826                 list_add_tail(&work->list, &cm_id_priv->work_list);
2827         spin_unlock_irq(&cm_id_priv->lock);
2828
2829         if (ret)
2830                 cm_process_work(cm_id_priv, work);
2831         else
2832                 cm_deref_id(cm_id_priv);
2833         return 0;
2834 out:
2835         spin_unlock_irq(&cm_id_priv->lock);
2836         cm_deref_id(cm_id_priv);
2837         return -EINVAL;
2838 }
2839
2840 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2841                           struct cm_id_private *cm_id_priv,
2842                           struct sa_path_rec *alternate_path,
2843                           const void *private_data,
2844                           u8 private_data_len)
2845 {
2846         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2847                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2848         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2849         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2850         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2851         /* todo: need remote CM response timeout */
2852         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2853         lap_msg->alt_local_lid =
2854                 htons(ntohl(sa_path_get_slid(alternate_path)));
2855         lap_msg->alt_remote_lid =
2856                 htons(ntohl(sa_path_get_dlid(alternate_path)));
2857         lap_msg->alt_local_gid = alternate_path->sgid;
2858         lap_msg->alt_remote_gid = alternate_path->dgid;
2859         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2860         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2861         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2862         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2863         cm_lap_set_sl(lap_msg, alternate_path->sl);
2864         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2865         cm_lap_set_local_ack_timeout(lap_msg,
2866                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2867                                alternate_path->packet_life_time));
2868
2869         if (private_data && private_data_len)
2870                 memcpy(lap_msg->private_data, private_data, private_data_len);
2871 }
2872
2873 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2874                    struct sa_path_rec *alternate_path,
2875                    const void *private_data,
2876                    u8 private_data_len)
2877 {
2878         struct cm_id_private *cm_id_priv;
2879         struct ib_mad_send_buf *msg;
2880         unsigned long flags;
2881         int ret;
2882
2883         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2884                 return -EINVAL;
2885
2886         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2887         spin_lock_irqsave(&cm_id_priv->lock, flags);
2888         if (cm_id->state != IB_CM_ESTABLISHED ||
2889             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2890              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2891                 ret = -EINVAL;
2892                 goto out;
2893         }
2894
2895         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av,
2896                                  cm_id_priv);
2897         if (ret)
2898                 goto out;
2899         cm_id_priv->alt_av.timeout =
2900                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2901                                        cm_id_priv->alt_av.timeout - 1);
2902
2903         ret = cm_alloc_msg(cm_id_priv, &msg);
2904         if (ret)
2905                 goto out;
2906
2907         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2908                       alternate_path, private_data, private_data_len);
2909         msg->timeout_ms = cm_id_priv->timeout_ms;
2910         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2911
2912         ret = ib_post_send_mad(msg, NULL);
2913         if (ret) {
2914                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2915                 cm_free_msg(msg);
2916                 return ret;
2917         }
2918
2919         cm_id->lap_state = IB_CM_LAP_SENT;
2920         cm_id_priv->msg = msg;
2921
2922 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2923         return ret;
2924 }
2925 EXPORT_SYMBOL(ib_send_cm_lap);
2926
2927 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2928                                     struct sa_path_rec *path,
2929                                     struct cm_lap_msg *lap_msg)
2930 {
2931         memset(path, 0, sizeof *path);
2932         path->rec_type = SA_PATH_REC_TYPE_IB;
2933         path->dgid = lap_msg->alt_local_gid;
2934         path->sgid = lap_msg->alt_remote_gid;
2935         sa_path_set_dlid(path, htonl(ntohs(lap_msg->alt_local_lid)));
2936         sa_path_set_slid(path, htonl(ntohs(lap_msg->alt_remote_lid)));
2937         path->flow_label = cm_lap_get_flow_label(lap_msg);
2938         path->hop_limit = lap_msg->alt_hop_limit;
2939         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2940         path->reversible = 1;
2941         path->pkey = cm_id_priv->pkey;
2942         path->sl = cm_lap_get_sl(lap_msg);
2943         path->mtu_selector = IB_SA_EQ;
2944         path->mtu = cm_id_priv->path_mtu;
2945         path->rate_selector = IB_SA_EQ;
2946         path->rate = cm_lap_get_packet_rate(lap_msg);
2947         path->packet_life_time_selector = IB_SA_EQ;
2948         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2949         path->packet_life_time -= (path->packet_life_time > 0);
2950 }
2951
2952 static int cm_lap_handler(struct cm_work *work)
2953 {
2954         struct cm_id_private *cm_id_priv;
2955         struct cm_lap_msg *lap_msg;
2956         struct ib_cm_lap_event_param *param;
2957         struct ib_mad_send_buf *msg = NULL;
2958         int ret;
2959
2960         /* todo: verify LAP request and send reject APR if invalid. */
2961         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2962         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2963                                    lap_msg->local_comm_id);
2964         if (!cm_id_priv)
2965                 return -EINVAL;
2966
2967         param = &work->cm_event.param.lap_rcvd;
2968         param->alternate_path = &work->path[0];
2969         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2970         work->cm_event.private_data = &lap_msg->private_data;
2971
2972         spin_lock_irq(&cm_id_priv->lock);
2973         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2974                 goto unlock;
2975
2976         switch (cm_id_priv->id.lap_state) {
2977         case IB_CM_LAP_UNINIT:
2978         case IB_CM_LAP_IDLE:
2979                 break;
2980         case IB_CM_MRA_LAP_SENT:
2981                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2982                                 counter[CM_LAP_COUNTER]);
2983                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2984                         goto unlock;
2985
2986                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2987                               CM_MSG_RESPONSE_OTHER,
2988                               cm_id_priv->service_timeout,
2989                               cm_id_priv->private_data,
2990                               cm_id_priv->private_data_len);
2991                 spin_unlock_irq(&cm_id_priv->lock);
2992
2993                 if (ib_post_send_mad(msg, NULL))
2994                         cm_free_msg(msg);
2995                 goto deref;
2996         case IB_CM_LAP_RCVD:
2997                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2998                                 counter[CM_LAP_COUNTER]);
2999                 goto unlock;
3000         default:
3001                 goto unlock;
3002         }
3003
3004         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
3005         cm_id_priv->tid = lap_msg->hdr.tid;
3006         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3007                                 work->mad_recv_wc->recv_buf.grh,
3008                                 &cm_id_priv->av);
3009         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av,
3010                            cm_id_priv);
3011         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3012         if (!ret)
3013                 list_add_tail(&work->list, &cm_id_priv->work_list);
3014         spin_unlock_irq(&cm_id_priv->lock);
3015
3016         if (ret)
3017                 cm_process_work(cm_id_priv, work);
3018         else
3019                 cm_deref_id(cm_id_priv);
3020         return 0;
3021
3022 unlock: spin_unlock_irq(&cm_id_priv->lock);
3023 deref:  cm_deref_id(cm_id_priv);
3024         return -EINVAL;
3025 }
3026
3027 static void cm_format_apr(struct cm_apr_msg *apr_msg,
3028                           struct cm_id_private *cm_id_priv,
3029                           enum ib_cm_apr_status status,
3030                           void *info,
3031                           u8 info_length,
3032                           const void *private_data,
3033                           u8 private_data_len)
3034 {
3035         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
3036         apr_msg->local_comm_id = cm_id_priv->id.local_id;
3037         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
3038         apr_msg->ap_status = (u8) status;
3039
3040         if (info && info_length) {
3041                 apr_msg->info_length = info_length;
3042                 memcpy(apr_msg->info, info, info_length);
3043         }
3044
3045         if (private_data && private_data_len)
3046                 memcpy(apr_msg->private_data, private_data, private_data_len);
3047 }
3048
3049 int ib_send_cm_apr(struct ib_cm_id *cm_id,
3050                    enum ib_cm_apr_status status,
3051                    void *info,
3052                    u8 info_length,
3053                    const void *private_data,
3054                    u8 private_data_len)
3055 {
3056         struct cm_id_private *cm_id_priv;
3057         struct ib_mad_send_buf *msg;
3058         unsigned long flags;
3059         int ret;
3060
3061         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
3062             (info && info_length > IB_CM_APR_INFO_LENGTH))
3063                 return -EINVAL;
3064
3065         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3066         spin_lock_irqsave(&cm_id_priv->lock, flags);
3067         if (cm_id->state != IB_CM_ESTABLISHED ||
3068             (cm_id->lap_state != IB_CM_LAP_RCVD &&
3069              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
3070                 ret = -EINVAL;
3071                 goto out;
3072         }
3073
3074         ret = cm_alloc_msg(cm_id_priv, &msg);
3075         if (ret)
3076                 goto out;
3077
3078         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
3079                       info, info_length, private_data, private_data_len);
3080         ret = ib_post_send_mad(msg, NULL);
3081         if (ret) {
3082                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3083                 cm_free_msg(msg);
3084                 return ret;
3085         }
3086
3087         cm_id->lap_state = IB_CM_LAP_IDLE;
3088 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3089         return ret;
3090 }
3091 EXPORT_SYMBOL(ib_send_cm_apr);
3092
3093 static int cm_apr_handler(struct cm_work *work)
3094 {
3095         struct cm_id_private *cm_id_priv;
3096         struct cm_apr_msg *apr_msg;
3097         int ret;
3098
3099         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3100         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
3101                                    apr_msg->local_comm_id);
3102         if (!cm_id_priv)
3103                 return -EINVAL; /* Unmatched reply. */
3104
3105         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
3106         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
3107         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
3108         work->cm_event.private_data = &apr_msg->private_data;
3109
3110         spin_lock_irq(&cm_id_priv->lock);
3111         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3112             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3113              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3114                 spin_unlock_irq(&cm_id_priv->lock);
3115                 goto out;
3116         }
3117         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3118         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3119         cm_id_priv->msg = NULL;
3120
3121         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3122         if (!ret)
3123                 list_add_tail(&work->list, &cm_id_priv->work_list);
3124         spin_unlock_irq(&cm_id_priv->lock);
3125
3126         if (ret)
3127                 cm_process_work(cm_id_priv, work);
3128         else
3129                 cm_deref_id(cm_id_priv);
3130         return 0;
3131 out:
3132         cm_deref_id(cm_id_priv);
3133         return -EINVAL;
3134 }
3135
3136 static int cm_timewait_handler(struct cm_work *work)
3137 {
3138         struct cm_timewait_info *timewait_info;
3139         struct cm_id_private *cm_id_priv;
3140         int ret;
3141
3142         timewait_info = (struct cm_timewait_info *)work;
3143         spin_lock_irq(&cm.lock);
3144         list_del(&timewait_info->list);
3145         spin_unlock_irq(&cm.lock);
3146
3147         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3148                                    timewait_info->work.remote_id);
3149         if (!cm_id_priv)
3150                 return -EINVAL;
3151
3152         spin_lock_irq(&cm_id_priv->lock);
3153         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3154             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3155                 spin_unlock_irq(&cm_id_priv->lock);
3156                 goto out;
3157         }
3158         cm_id_priv->id.state = IB_CM_IDLE;
3159         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3160         if (!ret)
3161                 list_add_tail(&work->list, &cm_id_priv->work_list);
3162         spin_unlock_irq(&cm_id_priv->lock);
3163
3164         if (ret)
3165                 cm_process_work(cm_id_priv, work);
3166         else
3167                 cm_deref_id(cm_id_priv);
3168         return 0;
3169 out:
3170         cm_deref_id(cm_id_priv);
3171         return -EINVAL;
3172 }
3173
3174 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3175                                struct cm_id_private *cm_id_priv,
3176                                struct ib_cm_sidr_req_param *param)
3177 {
3178         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3179                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
3180         sidr_req_msg->request_id = cm_id_priv->id.local_id;
3181         sidr_req_msg->pkey = param->path->pkey;
3182         sidr_req_msg->service_id = param->service_id;
3183
3184         if (param->private_data && param->private_data_len)
3185                 memcpy(sidr_req_msg->private_data, param->private_data,
3186                        param->private_data_len);
3187 }
3188
3189 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3190                         struct ib_cm_sidr_req_param *param)
3191 {
3192         struct cm_id_private *cm_id_priv;
3193         struct ib_mad_send_buf *msg;
3194         unsigned long flags;
3195         int ret;
3196
3197         if (!param->path || (param->private_data &&
3198              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3199                 return -EINVAL;
3200
3201         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3202         ret = cm_init_av_by_path(param->path, &cm_id_priv->av, cm_id_priv);
3203         if (ret)
3204                 goto out;
3205
3206         cm_id->service_id = param->service_id;
3207         cm_id->service_mask = ~cpu_to_be64(0);
3208         cm_id_priv->timeout_ms = param->timeout_ms;
3209         cm_id_priv->max_cm_retries = param->max_cm_retries;
3210         ret = cm_alloc_msg(cm_id_priv, &msg);
3211         if (ret)
3212                 goto out;
3213
3214         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3215                            param);
3216         msg->timeout_ms = cm_id_priv->timeout_ms;
3217         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3218
3219         spin_lock_irqsave(&cm_id_priv->lock, flags);
3220         if (cm_id->state == IB_CM_IDLE)
3221                 ret = ib_post_send_mad(msg, NULL);
3222         else
3223                 ret = -EINVAL;
3224
3225         if (ret) {
3226                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3227                 cm_free_msg(msg);
3228                 goto out;
3229         }
3230         cm_id->state = IB_CM_SIDR_REQ_SENT;
3231         cm_id_priv->msg = msg;
3232         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3233 out:
3234         return ret;
3235 }
3236 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3237
3238 static void cm_format_sidr_req_event(struct cm_work *work,
3239                                      struct ib_cm_id *listen_id)
3240 {
3241         struct cm_sidr_req_msg *sidr_req_msg;
3242         struct ib_cm_sidr_req_event_param *param;
3243
3244         sidr_req_msg = (struct cm_sidr_req_msg *)
3245                                 work->mad_recv_wc->recv_buf.mad;
3246         param = &work->cm_event.param.sidr_req_rcvd;
3247         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
3248         param->listen_id = listen_id;
3249         param->service_id = sidr_req_msg->service_id;
3250         param->bth_pkey = cm_get_bth_pkey(work);
3251         param->port = work->port->port_num;
3252         work->cm_event.private_data = &sidr_req_msg->private_data;
3253 }
3254
3255 static int cm_sidr_req_handler(struct cm_work *work)
3256 {
3257         struct ib_cm_id *cm_id;
3258         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3259         struct cm_sidr_req_msg *sidr_req_msg;
3260         struct ib_wc *wc;
3261
3262         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3263         if (IS_ERR(cm_id))
3264                 return PTR_ERR(cm_id);
3265         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3266
3267         /* Record SGID/SLID and request ID for lookup. */
3268         sidr_req_msg = (struct cm_sidr_req_msg *)
3269                                 work->mad_recv_wc->recv_buf.mad;
3270         wc = work->mad_recv_wc->wc;
3271         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3272         cm_id_priv->av.dgid.global.interface_id = 0;
3273         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3274                                 work->mad_recv_wc->recv_buf.grh,
3275                                 &cm_id_priv->av);
3276         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3277         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3278         atomic_inc(&cm_id_priv->work_count);
3279
3280         spin_lock_irq(&cm.lock);
3281         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3282         if (cur_cm_id_priv) {
3283                 spin_unlock_irq(&cm.lock);
3284                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3285                                 counter[CM_SIDR_REQ_COUNTER]);
3286                 goto out; /* Duplicate message. */
3287         }
3288         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3289         cur_cm_id_priv = cm_find_listen(cm_id->device,
3290                                         sidr_req_msg->service_id);
3291         if (!cur_cm_id_priv) {
3292                 spin_unlock_irq(&cm.lock);
3293                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3294                 goto out; /* No match. */
3295         }
3296         atomic_inc(&cur_cm_id_priv->refcount);
3297         atomic_inc(&cm_id_priv->refcount);
3298         spin_unlock_irq(&cm.lock);
3299
3300         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3301         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3302         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3303         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3304
3305         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3306         cm_process_work(cm_id_priv, work);
3307         cm_deref_id(cur_cm_id_priv);
3308         return 0;
3309 out:
3310         ib_destroy_cm_id(&cm_id_priv->id);
3311         return -EINVAL;
3312 }
3313
3314 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3315                                struct cm_id_private *cm_id_priv,
3316                                struct ib_cm_sidr_rep_param *param)
3317 {
3318         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3319                           cm_id_priv->tid);
3320         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3321         sidr_rep_msg->status = param->status;
3322         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3323         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3324         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3325
3326         if (param->info && param->info_length)
3327                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3328
3329         if (param->private_data && param->private_data_len)
3330                 memcpy(sidr_rep_msg->private_data, param->private_data,
3331                        param->private_data_len);
3332 }
3333
3334 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3335                         struct ib_cm_sidr_rep_param *param)
3336 {
3337         struct cm_id_private *cm_id_priv;
3338         struct ib_mad_send_buf *msg;
3339         unsigned long flags;
3340         int ret;
3341
3342         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3343             (param->private_data &&
3344              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3345                 return -EINVAL;
3346
3347         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3348         spin_lock_irqsave(&cm_id_priv->lock, flags);
3349         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3350                 ret = -EINVAL;
3351                 goto error;
3352         }
3353
3354         ret = cm_alloc_msg(cm_id_priv, &msg);
3355         if (ret)
3356                 goto error;
3357
3358         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3359                            param);
3360         ret = ib_post_send_mad(msg, NULL);
3361         if (ret) {
3362                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3363                 cm_free_msg(msg);
3364                 return ret;
3365         }
3366         cm_id->state = IB_CM_IDLE;
3367         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3368
3369         spin_lock_irqsave(&cm.lock, flags);
3370         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3371                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3372                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3373         }
3374         spin_unlock_irqrestore(&cm.lock, flags);
3375         return 0;
3376
3377 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3378         return ret;
3379 }
3380 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3381
3382 static void cm_format_sidr_rep_event(struct cm_work *work)
3383 {
3384         struct cm_sidr_rep_msg *sidr_rep_msg;
3385         struct ib_cm_sidr_rep_event_param *param;
3386
3387         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3388                                 work->mad_recv_wc->recv_buf.mad;
3389         param = &work->cm_event.param.sidr_rep_rcvd;
3390         param->status = sidr_rep_msg->status;
3391         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3392         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3393         param->info = &sidr_rep_msg->info;
3394         param->info_len = sidr_rep_msg->info_length;
3395         work->cm_event.private_data = &sidr_rep_msg->private_data;
3396 }
3397
3398 static int cm_sidr_rep_handler(struct cm_work *work)
3399 {
3400         struct cm_sidr_rep_msg *sidr_rep_msg;
3401         struct cm_id_private *cm_id_priv;
3402
3403         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3404                                 work->mad_recv_wc->recv_buf.mad;
3405         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3406         if (!cm_id_priv)
3407                 return -EINVAL; /* Unmatched reply. */
3408
3409         spin_lock_irq(&cm_id_priv->lock);
3410         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3411                 spin_unlock_irq(&cm_id_priv->lock);
3412                 goto out;
3413         }
3414         cm_id_priv->id.state = IB_CM_IDLE;
3415         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3416         spin_unlock_irq(&cm_id_priv->lock);
3417
3418         cm_format_sidr_rep_event(work);
3419         cm_process_work(cm_id_priv, work);
3420         return 0;
3421 out:
3422         cm_deref_id(cm_id_priv);
3423         return -EINVAL;
3424 }
3425
3426 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3427                                   enum ib_wc_status wc_status)
3428 {
3429         struct cm_id_private *cm_id_priv;
3430         struct ib_cm_event cm_event;
3431         enum ib_cm_state state;
3432         int ret;
3433
3434         memset(&cm_event, 0, sizeof cm_event);
3435         cm_id_priv = msg->context[0];
3436
3437         /* Discard old sends or ones without a response. */
3438         spin_lock_irq(&cm_id_priv->lock);
3439         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3440         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3441                 goto discard;
3442
3443         pr_debug_ratelimited("CM: failed sending MAD in state %d. (%s)\n",
3444                              state, ib_wc_status_msg(wc_status));
3445         switch (state) {
3446         case IB_CM_REQ_SENT:
3447         case IB_CM_MRA_REQ_RCVD:
3448                 cm_reset_to_idle(cm_id_priv);
3449                 cm_event.event = IB_CM_REQ_ERROR;
3450                 break;
3451         case IB_CM_REP_SENT:
3452         case IB_CM_MRA_REP_RCVD:
3453                 cm_reset_to_idle(cm_id_priv);
3454                 cm_event.event = IB_CM_REP_ERROR;
3455                 break;
3456         case IB_CM_DREQ_SENT:
3457                 cm_enter_timewait(cm_id_priv);
3458                 cm_event.event = IB_CM_DREQ_ERROR;
3459                 break;
3460         case IB_CM_SIDR_REQ_SENT:
3461                 cm_id_priv->id.state = IB_CM_IDLE;
3462                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3463                 break;
3464         default:
3465                 goto discard;
3466         }
3467         spin_unlock_irq(&cm_id_priv->lock);
3468         cm_event.param.send_status = wc_status;
3469
3470         /* No other events can occur on the cm_id at this point. */
3471         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3472         cm_free_msg(msg);
3473         if (ret)
3474                 ib_destroy_cm_id(&cm_id_priv->id);
3475         return;
3476 discard:
3477         spin_unlock_irq(&cm_id_priv->lock);
3478         cm_free_msg(msg);
3479 }
3480
3481 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3482                             struct ib_mad_send_wc *mad_send_wc)
3483 {
3484         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3485         struct cm_port *port;
3486         u16 attr_index;
3487
3488         port = mad_agent->context;
3489         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3490                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3491
3492         /*
3493          * If the send was in response to a received message (context[0] is not
3494          * set to a cm_id), and is not a REJ, then it is a send that was
3495          * manually retried.
3496          */
3497         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3498                 msg->retries = 1;
3499
3500         atomic_long_add(1 + msg->retries,
3501                         &port->counter_group[CM_XMIT].counter[attr_index]);
3502         if (msg->retries)
3503                 atomic_long_add(msg->retries,
3504                                 &port->counter_group[CM_XMIT_RETRIES].
3505                                 counter[attr_index]);
3506
3507         switch (mad_send_wc->status) {
3508         case IB_WC_SUCCESS:
3509         case IB_WC_WR_FLUSH_ERR:
3510                 cm_free_msg(msg);
3511                 break;
3512         default:
3513                 if (msg->context[0] && msg->context[1])
3514                         cm_process_send_error(msg, mad_send_wc->status);
3515                 else
3516                         cm_free_msg(msg);
3517                 break;
3518         }
3519 }
3520
3521 static void cm_work_handler(struct work_struct *_work)
3522 {
3523         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3524         int ret;
3525
3526         switch (work->cm_event.event) {
3527         case IB_CM_REQ_RECEIVED:
3528                 ret = cm_req_handler(work);
3529                 break;
3530         case IB_CM_MRA_RECEIVED:
3531                 ret = cm_mra_handler(work);
3532                 break;
3533         case IB_CM_REJ_RECEIVED:
3534                 ret = cm_rej_handler(work);
3535                 break;
3536         case IB_CM_REP_RECEIVED:
3537                 ret = cm_rep_handler(work);
3538                 break;
3539         case IB_CM_RTU_RECEIVED:
3540                 ret = cm_rtu_handler(work);
3541                 break;
3542         case IB_CM_USER_ESTABLISHED:
3543                 ret = cm_establish_handler(work);
3544                 break;
3545         case IB_CM_DREQ_RECEIVED:
3546                 ret = cm_dreq_handler(work);
3547                 break;
3548         case IB_CM_DREP_RECEIVED:
3549                 ret = cm_drep_handler(work);
3550                 break;
3551         case IB_CM_SIDR_REQ_RECEIVED:
3552                 ret = cm_sidr_req_handler(work);
3553                 break;
3554         case IB_CM_SIDR_REP_RECEIVED:
3555                 ret = cm_sidr_rep_handler(work);
3556                 break;
3557         case IB_CM_LAP_RECEIVED:
3558                 ret = cm_lap_handler(work);
3559                 break;
3560         case IB_CM_APR_RECEIVED:
3561                 ret = cm_apr_handler(work);
3562                 break;
3563         case IB_CM_TIMEWAIT_EXIT:
3564                 ret = cm_timewait_handler(work);
3565                 break;
3566         default:
3567                 ret = -EINVAL;
3568                 break;
3569         }
3570         if (ret)
3571                 cm_free_work(work);
3572 }
3573
3574 static int cm_establish(struct ib_cm_id *cm_id)
3575 {
3576         struct cm_id_private *cm_id_priv;
3577         struct cm_work *work;
3578         unsigned long flags;
3579         int ret = 0;
3580         struct cm_device *cm_dev;
3581
3582         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3583         if (!cm_dev)
3584                 return -ENODEV;
3585
3586         work = kmalloc(sizeof *work, GFP_ATOMIC);
3587         if (!work)
3588                 return -ENOMEM;
3589
3590         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3591         spin_lock_irqsave(&cm_id_priv->lock, flags);
3592         switch (cm_id->state)
3593         {
3594         case IB_CM_REP_SENT:
3595         case IB_CM_MRA_REP_RCVD:
3596                 cm_id->state = IB_CM_ESTABLISHED;
3597                 break;
3598         case IB_CM_ESTABLISHED:
3599                 ret = -EISCONN;
3600                 break;
3601         default:
3602                 ret = -EINVAL;
3603                 break;
3604         }
3605         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3606
3607         if (ret) {
3608                 kfree(work);
3609                 goto out;
3610         }
3611
3612         /*
3613          * The CM worker thread may try to destroy the cm_id before it
3614          * can execute this work item.  To prevent potential deadlock,
3615          * we need to find the cm_id once we're in the context of the
3616          * worker thread, rather than holding a reference on it.
3617          */
3618         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3619         work->local_id = cm_id->local_id;
3620         work->remote_id = cm_id->remote_id;
3621         work->mad_recv_wc = NULL;
3622         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3623
3624         /* Check if the device started its remove_one */
3625         spin_lock_irqsave(&cm.lock, flags);
3626         if (!cm_dev->going_down) {
3627                 queue_delayed_work(cm.wq, &work->work, 0);
3628         } else {
3629                 kfree(work);
3630                 ret = -ENODEV;
3631         }
3632         spin_unlock_irqrestore(&cm.lock, flags);
3633
3634 out:
3635         return ret;
3636 }
3637
3638 static int cm_migrate(struct ib_cm_id *cm_id)
3639 {
3640         struct cm_id_private *cm_id_priv;
3641         struct cm_av tmp_av;
3642         unsigned long flags;
3643         int tmp_send_port_not_ready;
3644         int ret = 0;
3645
3646         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3647         spin_lock_irqsave(&cm_id_priv->lock, flags);
3648         if (cm_id->state == IB_CM_ESTABLISHED &&
3649             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3650              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3651                 cm_id->lap_state = IB_CM_LAP_IDLE;
3652                 /* Swap address vector */
3653                 tmp_av = cm_id_priv->av;
3654                 cm_id_priv->av = cm_id_priv->alt_av;
3655                 cm_id_priv->alt_av = tmp_av;
3656                 /* Swap port send ready state */
3657                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3658                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3659                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3660         } else
3661                 ret = -EINVAL;
3662         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3663
3664         return ret;
3665 }
3666
3667 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3668 {
3669         int ret;
3670
3671         switch (event) {
3672         case IB_EVENT_COMM_EST:
3673                 ret = cm_establish(cm_id);
3674                 break;
3675         case IB_EVENT_PATH_MIG:
3676                 ret = cm_migrate(cm_id);
3677                 break;
3678         default:
3679                 ret = -EINVAL;
3680         }
3681         return ret;
3682 }
3683 EXPORT_SYMBOL(ib_cm_notify);
3684
3685 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3686                             struct ib_mad_send_buf *send_buf,
3687                             struct ib_mad_recv_wc *mad_recv_wc)
3688 {
3689         struct cm_port *port = mad_agent->context;
3690         struct cm_work *work;
3691         enum ib_cm_event_type event;
3692         u16 attr_id;
3693         int paths = 0;
3694         int going_down = 0;
3695
3696         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3697         case CM_REQ_ATTR_ID:
3698                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3699                                                     alt_local_lid != 0);
3700                 event = IB_CM_REQ_RECEIVED;
3701                 break;
3702         case CM_MRA_ATTR_ID:
3703                 event = IB_CM_MRA_RECEIVED;
3704                 break;
3705         case CM_REJ_ATTR_ID:
3706                 event = IB_CM_REJ_RECEIVED;
3707                 break;
3708         case CM_REP_ATTR_ID:
3709                 event = IB_CM_REP_RECEIVED;
3710                 break;
3711         case CM_RTU_ATTR_ID:
3712                 event = IB_CM_RTU_RECEIVED;
3713                 break;
3714         case CM_DREQ_ATTR_ID:
3715                 event = IB_CM_DREQ_RECEIVED;
3716                 break;
3717         case CM_DREP_ATTR_ID:
3718                 event = IB_CM_DREP_RECEIVED;
3719                 break;
3720         case CM_SIDR_REQ_ATTR_ID:
3721                 event = IB_CM_SIDR_REQ_RECEIVED;
3722                 break;
3723         case CM_SIDR_REP_ATTR_ID:
3724                 event = IB_CM_SIDR_REP_RECEIVED;
3725                 break;
3726         case CM_LAP_ATTR_ID:
3727                 paths = 1;
3728                 event = IB_CM_LAP_RECEIVED;
3729                 break;
3730         case CM_APR_ATTR_ID:
3731                 event = IB_CM_APR_RECEIVED;
3732                 break;
3733         default:
3734                 ib_free_recv_mad(mad_recv_wc);
3735                 return;
3736         }
3737
3738         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3739         atomic_long_inc(&port->counter_group[CM_RECV].
3740                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3741
3742         work = kmalloc(sizeof(*work) + sizeof(struct sa_path_rec) * paths,
3743                        GFP_KERNEL);
3744         if (!work) {
3745                 ib_free_recv_mad(mad_recv_wc);
3746                 return;
3747         }
3748
3749         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3750         work->cm_event.event = event;
3751         work->mad_recv_wc = mad_recv_wc;
3752         work->port = port;
3753
3754         /* Check if the device started its remove_one */
3755         spin_lock_irq(&cm.lock);
3756         if (!port->cm_dev->going_down)
3757                 queue_delayed_work(cm.wq, &work->work, 0);
3758         else
3759                 going_down = 1;
3760         spin_unlock_irq(&cm.lock);
3761
3762         if (going_down) {
3763                 kfree(work);
3764                 ib_free_recv_mad(mad_recv_wc);
3765         }
3766 }
3767
3768 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3769                                 struct ib_qp_attr *qp_attr,
3770                                 int *qp_attr_mask)
3771 {
3772         unsigned long flags;
3773         int ret;
3774
3775         spin_lock_irqsave(&cm_id_priv->lock, flags);
3776         switch (cm_id_priv->id.state) {
3777         case IB_CM_REQ_SENT:
3778         case IB_CM_MRA_REQ_RCVD:
3779         case IB_CM_REQ_RCVD:
3780         case IB_CM_MRA_REQ_SENT:
3781         case IB_CM_REP_RCVD:
3782         case IB_CM_MRA_REP_SENT:
3783         case IB_CM_REP_SENT:
3784         case IB_CM_MRA_REP_RCVD:
3785         case IB_CM_ESTABLISHED:
3786                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3787                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3788                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3789                 if (cm_id_priv->responder_resources)
3790                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3791                                                     IB_ACCESS_REMOTE_ATOMIC;
3792                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3793                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3794                 ret = 0;
3795                 break;
3796         default:
3797                 ret = -EINVAL;
3798                 break;
3799         }
3800         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3801         return ret;
3802 }
3803
3804 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3805                                struct ib_qp_attr *qp_attr,
3806                                int *qp_attr_mask)
3807 {
3808         unsigned long flags;
3809         int ret;
3810
3811         spin_lock_irqsave(&cm_id_priv->lock, flags);
3812         switch (cm_id_priv->id.state) {
3813         case IB_CM_REQ_RCVD:
3814         case IB_CM_MRA_REQ_SENT:
3815         case IB_CM_REP_RCVD:
3816         case IB_CM_MRA_REP_SENT:
3817         case IB_CM_REP_SENT:
3818         case IB_CM_MRA_REP_RCVD:
3819         case IB_CM_ESTABLISHED:
3820                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3821                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3822                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3823                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3824                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3825                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3826                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3827                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3828                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3829                                          IB_QP_MIN_RNR_TIMER;
3830                         qp_attr->max_dest_rd_atomic =
3831                                         cm_id_priv->responder_resources;
3832                         qp_attr->min_rnr_timer = 0;
3833                 }
3834                 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
3835                         *qp_attr_mask |= IB_QP_ALT_PATH;
3836                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3837                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3838                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3839                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3840                 }
3841                 ret = 0;
3842                 break;
3843         default:
3844                 ret = -EINVAL;
3845                 break;
3846         }
3847         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3848         return ret;
3849 }
3850
3851 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3852                                struct ib_qp_attr *qp_attr,
3853                                int *qp_attr_mask)
3854 {
3855         unsigned long flags;
3856         int ret;
3857
3858         spin_lock_irqsave(&cm_id_priv->lock, flags);
3859         switch (cm_id_priv->id.state) {
3860         /* Allow transition to RTS before sending REP */
3861         case IB_CM_REQ_RCVD:
3862         case IB_CM_MRA_REQ_SENT:
3863
3864         case IB_CM_REP_RCVD:
3865         case IB_CM_MRA_REP_SENT:
3866         case IB_CM_REP_SENT:
3867         case IB_CM_MRA_REP_RCVD:
3868         case IB_CM_ESTABLISHED:
3869                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3870                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3871                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3872                         switch (cm_id_priv->qp_type) {
3873                         case IB_QPT_RC:
3874                         case IB_QPT_XRC_INI:
3875                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3876                                                  IB_QP_MAX_QP_RD_ATOMIC;
3877                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3878                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3879                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3880                                 /* fall through */
3881                         case IB_QPT_XRC_TGT:
3882                                 *qp_attr_mask |= IB_QP_TIMEOUT;
3883                                 qp_attr->timeout = cm_id_priv->av.timeout;
3884                                 break;
3885                         default:
3886                                 break;
3887                         }
3888                         if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
3889                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3890                                 qp_attr->path_mig_state = IB_MIG_REARM;
3891                         }
3892                 } else {
3893                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3894                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3895                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3896                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3897                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3898                         qp_attr->path_mig_state = IB_MIG_REARM;
3899                 }
3900                 ret = 0;
3901                 break;
3902         default:
3903                 ret = -EINVAL;
3904                 break;
3905         }
3906         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3907         return ret;
3908 }
3909
3910 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3911                        struct ib_qp_attr *qp_attr,
3912                        int *qp_attr_mask)
3913 {
3914         struct cm_id_private *cm_id_priv;
3915         int ret;
3916
3917         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3918         switch (qp_attr->qp_state) {
3919         case IB_QPS_INIT:
3920                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3921                 break;
3922         case IB_QPS_RTR:
3923                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3924                 break;
3925         case IB_QPS_RTS:
3926                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3927                 break;
3928         default:
3929                 ret = -EINVAL;
3930                 break;
3931         }
3932         return ret;
3933 }
3934 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3935
3936 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3937                                char *buf)
3938 {
3939         struct cm_counter_group *group;
3940         struct cm_counter_attribute *cm_attr;
3941
3942         group = container_of(obj, struct cm_counter_group, obj);
3943         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3944
3945         return sprintf(buf, "%ld\n",
3946                        atomic_long_read(&group->counter[cm_attr->index]));
3947 }
3948
3949 static const struct sysfs_ops cm_counter_ops = {
3950         .show = cm_show_counter
3951 };
3952
3953 static struct kobj_type cm_counter_obj_type = {
3954         .sysfs_ops = &cm_counter_ops,
3955         .default_attrs = cm_counter_default_attrs
3956 };
3957
3958 static void cm_release_port_obj(struct kobject *obj)
3959 {
3960         struct cm_port *cm_port;
3961
3962         cm_port = container_of(obj, struct cm_port, port_obj);
3963         kfree(cm_port);
3964 }
3965
3966 static struct kobj_type cm_port_obj_type = {
3967         .release = cm_release_port_obj
3968 };
3969
3970 static char *cm_devnode(struct device *dev, umode_t *mode)
3971 {
3972         if (mode)
3973                 *mode = 0666;
3974         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3975 }
3976
3977 struct class cm_class = {
3978         .owner   = THIS_MODULE,
3979         .name    = "infiniband_cm",
3980         .devnode = cm_devnode,
3981 };
3982 EXPORT_SYMBOL(cm_class);
3983
3984 static int cm_create_port_fs(struct cm_port *port)
3985 {
3986         int i, ret;
3987
3988         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3989                                    &port->cm_dev->device->kobj,
3990                                    "%d", port->port_num);
3991         if (ret) {
3992                 kfree(port);
3993                 return ret;
3994         }
3995
3996         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3997                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3998                                            &cm_counter_obj_type,
3999                                            &port->port_obj,
4000                                            "%s", counter_group_names[i]);
4001                 if (ret)
4002                         goto error;
4003         }
4004
4005         return 0;
4006
4007 error:
4008         while (i--)
4009                 kobject_put(&port->counter_group[i].obj);
4010         kobject_put(&port->port_obj);
4011         return ret;
4012
4013 }
4014
4015 static void cm_remove_port_fs(struct cm_port *port)
4016 {
4017         int i;
4018
4019         for (i = 0; i < CM_COUNTER_GROUPS; i++)
4020                 kobject_put(&port->counter_group[i].obj);
4021
4022         kobject_put(&port->port_obj);
4023 }
4024
4025 static void cm_add_one(struct ib_device *ib_device)
4026 {
4027         struct cm_device *cm_dev;
4028         struct cm_port *port;
4029         struct ib_mad_reg_req reg_req = {
4030                 .mgmt_class = IB_MGMT_CLASS_CM,
4031                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4032         };
4033         struct ib_port_modify port_modify = {
4034                 .set_port_cap_mask = IB_PORT_CM_SUP
4035         };
4036         unsigned long flags;
4037         int ret;
4038         int count = 0;
4039         u8 i;
4040
4041         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
4042                          ib_device->phys_port_cnt, GFP_KERNEL);
4043         if (!cm_dev)
4044                 return;
4045
4046         cm_dev->ib_device = ib_device;
4047         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4048         cm_dev->going_down = 0;
4049         cm_dev->device = device_create(&cm_class, &ib_device->dev,
4050                                        MKDEV(0, 0), NULL,
4051                                        "%s", ib_device->name);
4052         if (IS_ERR(cm_dev->device)) {
4053                 kfree(cm_dev);
4054                 return;
4055         }
4056
4057         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4058         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4059                 if (!rdma_cap_ib_cm(ib_device, i))
4060                         continue;
4061
4062                 port = kzalloc(sizeof *port, GFP_KERNEL);
4063                 if (!port)
4064                         goto error1;
4065
4066                 cm_dev->port[i-1] = port;
4067                 port->cm_dev = cm_dev;
4068                 port->port_num = i;
4069
4070                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
4071                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
4072
4073                 ret = cm_create_port_fs(port);
4074                 if (ret)
4075                         goto error1;
4076
4077                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4078                                                         IB_QPT_GSI,
4079                                                         &reg_req,
4080                                                         0,
4081                                                         cm_send_handler,
4082                                                         cm_recv_handler,
4083                                                         port,
4084                                                         0);
4085                 if (IS_ERR(port->mad_agent))
4086                         goto error2;
4087
4088                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4089                 if (ret)
4090                         goto error3;
4091
4092                 count++;
4093         }
4094
4095         if (!count)
4096                 goto free;
4097
4098         ib_set_client_data(ib_device, &cm_client, cm_dev);
4099
4100         write_lock_irqsave(&cm.device_lock, flags);
4101         list_add_tail(&cm_dev->list, &cm.device_list);
4102         write_unlock_irqrestore(&cm.device_lock, flags);
4103         return;
4104
4105 error3:
4106         ib_unregister_mad_agent(port->mad_agent);
4107 error2:
4108         cm_remove_port_fs(port);
4109 error1:
4110         port_modify.set_port_cap_mask = 0;
4111         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4112         while (--i) {
4113                 if (!rdma_cap_ib_cm(ib_device, i))
4114                         continue;
4115
4116                 port = cm_dev->port[i-1];
4117                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4118                 ib_unregister_mad_agent(port->mad_agent);
4119                 cm_remove_port_fs(port);
4120         }
4121 free:
4122         device_unregister(cm_dev->device);
4123         kfree(cm_dev);
4124 }
4125
4126 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4127 {
4128         struct cm_device *cm_dev = client_data;
4129         struct cm_port *port;
4130         struct cm_id_private *cm_id_priv;
4131         struct ib_mad_agent *cur_mad_agent;
4132         struct ib_port_modify port_modify = {
4133                 .clr_port_cap_mask = IB_PORT_CM_SUP
4134         };
4135         unsigned long flags;
4136         int i;
4137
4138         if (!cm_dev)
4139                 return;
4140
4141         write_lock_irqsave(&cm.device_lock, flags);
4142         list_del(&cm_dev->list);
4143         write_unlock_irqrestore(&cm.device_lock, flags);
4144
4145         spin_lock_irq(&cm.lock);
4146         cm_dev->going_down = 1;
4147         spin_unlock_irq(&cm.lock);
4148
4149         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4150                 if (!rdma_cap_ib_cm(ib_device, i))
4151                         continue;
4152
4153                 port = cm_dev->port[i-1];
4154                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4155                 /* Mark all the cm_id's as not valid */
4156                 spin_lock_irq(&cm.lock);
4157                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4158                         cm_id_priv->altr_send_port_not_ready = 1;
4159                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4160                         cm_id_priv->prim_send_port_not_ready = 1;
4161                 spin_unlock_irq(&cm.lock);
4162                 /*
4163                  * We flush the queue here after the going_down set, this
4164                  * verify that no new works will be queued in the recv handler,
4165                  * after that we can call the unregister_mad_agent
4166                  */
4167                 flush_workqueue(cm.wq);
4168                 spin_lock_irq(&cm.state_lock);
4169                 cur_mad_agent = port->mad_agent;
4170                 port->mad_agent = NULL;
4171                 spin_unlock_irq(&cm.state_lock);
4172                 ib_unregister_mad_agent(cur_mad_agent);
4173                 cm_remove_port_fs(port);
4174         }
4175
4176         device_unregister(cm_dev->device);
4177         kfree(cm_dev);
4178 }
4179
4180 static int __init ib_cm_init(void)
4181 {
4182         int ret;
4183
4184         memset(&cm, 0, sizeof cm);
4185         INIT_LIST_HEAD(&cm.device_list);
4186         rwlock_init(&cm.device_lock);
4187         spin_lock_init(&cm.lock);
4188         spin_lock_init(&cm.state_lock);
4189         cm.listen_service_table = RB_ROOT;
4190         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4191         cm.remote_id_table = RB_ROOT;
4192         cm.remote_qp_table = RB_ROOT;
4193         cm.remote_sidr_table = RB_ROOT;
4194         idr_init(&cm.local_id_table);
4195         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4196         INIT_LIST_HEAD(&cm.timewait_list);
4197
4198         ret = class_register(&cm_class);
4199         if (ret) {
4200                 ret = -ENOMEM;
4201                 goto error1;
4202         }
4203
4204         cm.wq = create_workqueue("ib_cm");
4205         if (!cm.wq) {
4206                 ret = -ENOMEM;
4207                 goto error2;
4208         }
4209
4210         ret = ib_register_client(&cm_client);
4211         if (ret)
4212                 goto error3;
4213
4214         return 0;
4215 error3:
4216         destroy_workqueue(cm.wq);
4217 error2:
4218         class_unregister(&cm_class);
4219 error1:
4220         idr_destroy(&cm.local_id_table);
4221         return ret;
4222 }
4223
4224 static void __exit ib_cm_cleanup(void)
4225 {
4226         struct cm_timewait_info *timewait_info, *tmp;
4227
4228         spin_lock_irq(&cm.lock);
4229         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4230                 cancel_delayed_work(&timewait_info->work.work);
4231         spin_unlock_irq(&cm.lock);
4232
4233         ib_unregister_client(&cm_client);
4234         destroy_workqueue(cm.wq);
4235
4236         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4237                 list_del(&timewait_info->list);
4238                 kfree(timewait_info);
4239         }
4240
4241         class_unregister(&cm_class);
4242         idr_destroy(&cm.local_id_table);
4243 }
4244
4245 module_init(ib_cm_init);
4246 module_exit(ib_cm_cleanup);
4247