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 struct ib_mad_send_buf *cm_alloc_response_msg_no_ah(struct cm_port *port,
 377                                                            struct ib_mad_recv_wc *mad_recv_wc)
 378 {
 379         return ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 380                                   0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 381                                   GFP_ATOMIC,
 382                                   IB_MGMT_BASE_VERSION);
 383 }
 384
 385 static int cm_create_response_msg_ah(struct cm_port *port,
 386                                      struct ib_mad_recv_wc *mad_recv_wc,
 387                                      struct ib_mad_send_buf *msg)
 388 {
 389         struct ib_ah *ah;
 390
 391         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 392                                   mad_recv_wc->recv_buf.grh, port->port_num);
 393         if (IS_ERR(ah))
 394                 return PTR_ERR(ah);
 395
 396         msg->ah = ah;
 397         return 0;
 398 }
 399
 400 static void cm_free_msg(struct ib_mad_send_buf *msg)
 401 {
 402         if (msg->ah)
 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 int cm_alloc_response_msg(struct cm_port *port,
 410                                  struct ib_mad_recv_wc *mad_recv_wc,
 411                                  struct ib_mad_send_buf **msg)
 412 {
 413         struct ib_mad_send_buf *m;
 414         int ret;
 415
 416         m = cm_alloc_response_msg_no_ah(port, mad_recv_wc);
 417         if (IS_ERR(m))
 418                 return PTR_ERR(m);
 419
 420         ret = cm_create_response_msg_ah(port, mad_recv_wc, m);
 421         if (ret) {
 422                 cm_free_msg(m);
 423                 return ret;
 424         }
 425
 426         *msg = m;
 427         return 0;
 428 }
 429
 430 static void * cm_copy_private_data(const void *private_data,
 431                                    u8 private_data_len)
 432 {
 433         void *data;
 434
 435         if (!private_data || !private_data_len)
 436                 return NULL;
 437
 438         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 439         if (!data)
 440                 return ERR_PTR(-ENOMEM);
 441
 442         return data;
 443 }
 444
 445 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 446                                  void *private_data, u8 private_data_len)
 447 {
 448         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 449                 kfree(cm_id_priv->private_data);
 450
 451         cm_id_priv->private_data = private_data;
 452         cm_id_priv->private_data_len = private_data_len;
 453 }
 454
 455 static int cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 456                                    struct ib_grh *grh, struct cm_av *av)
 457 {
 458         av->port = port;
 459         av->pkey_index = wc->pkey_index;
 460         return ib_init_ah_attr_from_wc(port->cm_dev->ib_device,
 461                                        port->port_num, wc,
 462                                        grh, &av->ah_attr);
 463 }
 464
 465 static int add_cm_id_to_port_list(struct cm_id_private *cm_id_priv,
 466                                   struct cm_av *av,
 467                                   struct cm_port *port)
 468 {
 469         unsigned long flags;
 470         int ret = 0;
 471
 472         spin_lock_irqsave(&cm.lock, flags);
 473
 474         if (&cm_id_priv->av == av)
 475                 list_add_tail(&cm_id_priv->prim_list, &port->cm_priv_prim_list);
 476         else if (&cm_id_priv->alt_av == av)
 477                 list_add_tail(&cm_id_priv->altr_list, &port->cm_priv_altr_list);
 478         else
 479                 ret = -EINVAL;
 480
 481         spin_unlock_irqrestore(&cm.lock, flags);
 482         return ret;
 483 }
 484
 485 static struct cm_port *get_cm_port_from_path(struct sa_path_rec *path)
 486 {
 487         struct cm_device *cm_dev;
 488         struct cm_port *port = NULL;
 489         unsigned long flags;
 490         u8 p;
 491         struct net_device *ndev = ib_get_ndev_from_path(path);
 492
 493         read_lock_irqsave(&cm.device_lock, flags);
 494         list_for_each_entry(cm_dev, &cm.device_list, list) {
 495                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 496                                         sa_conv_pathrec_to_gid_type(path),
 497                                         ndev, &p, NULL)) {
 498                         port = cm_dev->port[p - 1];
 499                         break;
 500                 }
 501         }
 502         read_unlock_irqrestore(&cm.device_lock, flags);
 503
 504         if (ndev)
 505                 dev_put(ndev);
 506         return port;
 507 }
 508
 509 static int cm_init_av_by_path(struct sa_path_rec *path, struct cm_av *av,
 510                               struct cm_id_private *cm_id_priv)
 511 {
 512         struct cm_device *cm_dev;
 513         struct cm_port *port;
 514         int ret;
 515
 516         port = get_cm_port_from_path(path);
 517         if (!port)
 518                 return -EINVAL;
 519         cm_dev = port->cm_dev;
 520
 521         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 522                                   be16_to_cpu(path->pkey), &av->pkey_index);
 523         if (ret)
 524                 return ret;
 525
 526         av->port = port;
 527         ret = ib_init_ah_attr_from_path(cm_dev->ib_device, port->port_num, path,
 528                                         &av->ah_attr);
 529         if (ret)
 530                 return ret;
 531
 532         av->timeout = path->packet_life_time + 1;
 533
 534         ret = add_cm_id_to_port_list(cm_id_priv, av, port);
 535         return ret;
 536 }
 537
 538 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 539 {
 540         unsigned long flags;
 541         int id;
 542
 543         idr_preload(GFP_KERNEL);
 544         spin_lock_irqsave(&cm.lock, flags);
 545
 546         id = idr_alloc_cyclic(&cm.local_id_table, cm_id_priv, 0, 0, GFP_NOWAIT);
 547
 548         spin_unlock_irqrestore(&cm.lock, flags);
 549         idr_preload_end();
 550
 551         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 552         return id < 0 ? id : 0;
 553 }
 554
 555 static void cm_free_id(__be32 local_id)
 556 {
 557         spin_lock_irq(&cm.lock);
 558         idr_remove(&cm.local_id_table,
 559                    (__force int) (local_id ^ cm.random_id_operand));
 560         spin_unlock_irq(&cm.lock);
 561 }
 562
 563 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 564 {
 565         struct cm_id_private *cm_id_priv;
 566
 567         cm_id_priv = idr_find(&cm.local_id_table,
 568                               (__force int) (local_id ^ cm.random_id_operand));
 569         if (cm_id_priv) {
 570                 if (cm_id_priv->id.remote_id == remote_id)
 571                         atomic_inc(&cm_id_priv->refcount);
 572                 else
 573                         cm_id_priv = NULL;
 574         }
 575
 576         return cm_id_priv;
 577 }
 578
 579 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 580 {
 581         struct cm_id_private *cm_id_priv;
 582
 583         spin_lock_irq(&cm.lock);
 584         cm_id_priv = cm_get_id(local_id, remote_id);
 585         spin_unlock_irq(&cm.lock);
 586
 587         return cm_id_priv;
 588 }
 589
 590 /*
 591  * Trivial helpers to strip endian annotation and compare; the
 592  * endianness doesn't actually matter since we just need a stable
 593  * order for the RB tree.
 594  */
 595 static int be32_lt(__be32 a, __be32 b)
 596 {
 597         return (__force u32) a < (__force u32) b;
 598 }
 599
 600 static int be32_gt(__be32 a, __be32 b)
 601 {
 602         return (__force u32) a > (__force u32) b;
 603 }
 604
 605 static int be64_lt(__be64 a, __be64 b)
 606 {
 607         return (__force u64) a < (__force u64) b;
 608 }
 609
 610 static int be64_gt(__be64 a, __be64 b)
 611 {
 612         return (__force u64) a > (__force u64) b;
 613 }
 614
 615 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 616 {
 617         struct rb_node **link = &cm.listen_service_table.rb_node;
 618         struct rb_node *parent = NULL;
 619         struct cm_id_private *cur_cm_id_priv;
 620         __be64 service_id = cm_id_priv->id.service_id;
 621         __be64 service_mask = cm_id_priv->id.service_mask;
 622
 623         while (*link) {
 624                 parent = *link;
 625                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 626                                           service_node);
 627                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 628                     (service_mask & cur_cm_id_priv->id.service_id) &&
 629                     (cm_id_priv->id.device == cur_cm_id_priv->id.device))
 630                         return cur_cm_id_priv;
 631
 632                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 633                         link = &(*link)->rb_left;
 634                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 635                         link = &(*link)->rb_right;
 636                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 637                         link = &(*link)->rb_left;
 638                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 639                         link = &(*link)->rb_right;
 640                 else
 641                         link = &(*link)->rb_right;
 642         }
 643         rb_link_node(&cm_id_priv->service_node, parent, link);
 644         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 645         return NULL;
 646 }
 647
 648 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 649                                              __be64 service_id)
 650 {
 651         struct rb_node *node = cm.listen_service_table.rb_node;
 652         struct cm_id_private *cm_id_priv;
 653
 654         while (node) {
 655                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 656                 if ((cm_id_priv->id.service_mask & service_id) ==
 657                      cm_id_priv->id.service_id &&
 658                     (cm_id_priv->id.device == device))
 659                         return cm_id_priv;
 660
 661                 if (device < cm_id_priv->id.device)
 662                         node = node->rb_left;
 663                 else if (device > cm_id_priv->id.device)
 664                         node = node->rb_right;
 665                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 666                         node = node->rb_left;
 667                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 668                         node = node->rb_right;
 669                 else
 670                         node = node->rb_right;
 671         }
 672         return NULL;
 673 }
 674
 675 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 676                                                      *timewait_info)
 677 {
 678         struct rb_node **link = &cm.remote_id_table.rb_node;
 679         struct rb_node *parent = NULL;
 680         struct cm_timewait_info *cur_timewait_info;
 681         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 682         __be32 remote_id = timewait_info->work.remote_id;
 683
 684         while (*link) {
 685                 parent = *link;
 686                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 687                                              remote_id_node);
 688                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 689                         link = &(*link)->rb_left;
 690                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 691                         link = &(*link)->rb_right;
 692                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 693                         link = &(*link)->rb_left;
 694                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 695                         link = &(*link)->rb_right;
 696                 else
 697                         return cur_timewait_info;
 698         }
 699         timewait_info->inserted_remote_id = 1;
 700         rb_link_node(&timewait_info->remote_id_node, parent, link);
 701         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 702         return NULL;
 703 }
 704
 705 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 706                                                    __be32 remote_id)
 707 {
 708         struct rb_node *node = cm.remote_id_table.rb_node;
 709         struct cm_timewait_info *timewait_info;
 710
 711         while (node) {
 712                 timewait_info = rb_entry(node, struct cm_timewait_info,
 713                                          remote_id_node);
 714                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 715                         node = node->rb_left;
 716                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 717                         node = node->rb_right;
 718                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 719                         node = node->rb_left;
 720                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 721                         node = node->rb_right;
 722                 else
 723                         return timewait_info;
 724         }
 725         return NULL;
 726 }
 727
 728 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 729                                                       *timewait_info)
 730 {
 731         struct rb_node **link = &cm.remote_qp_table.rb_node;
 732         struct rb_node *parent = NULL;
 733         struct cm_timewait_info *cur_timewait_info;
 734         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 735         __be32 remote_qpn = timewait_info->remote_qpn;
 736
 737         while (*link) {
 738                 parent = *link;
 739                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 740                                              remote_qp_node);
 741                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 742                         link = &(*link)->rb_left;
 743                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 744                         link = &(*link)->rb_right;
 745                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 746                         link = &(*link)->rb_left;
 747                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 748                         link = &(*link)->rb_right;
 749                 else
 750                         return cur_timewait_info;
 751         }
 752         timewait_info->inserted_remote_qp = 1;
 753         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 754         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 755         return NULL;
 756 }
 757
 758 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 759                                                     *cm_id_priv)
 760 {
 761         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 762         struct rb_node *parent = NULL;
 763         struct cm_id_private *cur_cm_id_priv;
 764         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 765         __be32 remote_id = cm_id_priv->id.remote_id;
 766
 767         while (*link) {
 768                 parent = *link;
 769                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 770                                           sidr_id_node);
 771                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 772                         link = &(*link)->rb_left;
 773                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 774                         link = &(*link)->rb_right;
 775                 else {
 776                         int cmp;
 777                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 778                                      sizeof *port_gid);
 779                         if (cmp < 0)
 780                                 link = &(*link)->rb_left;
 781                         else if (cmp > 0)
 782                                 link = &(*link)->rb_right;
 783                         else
 784                                 return cur_cm_id_priv;
 785                 }
 786         }
 787         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 788         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 789         return NULL;
 790 }
 791
 792 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 793                                enum ib_cm_sidr_status status)
 794 {
 795         struct ib_cm_sidr_rep_param param;
 796
 797         memset(&param, 0, sizeof param);
 798         param.status = status;
 799         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 800 }
 801
 802 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 803                                  ib_cm_handler cm_handler,
 804                                  void *context)
 805 {
 806         struct cm_id_private *cm_id_priv;
 807         int ret;
 808
 809         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 810         if (!cm_id_priv)
 811                 return ERR_PTR(-ENOMEM);
 812
 813         cm_id_priv->id.state = IB_CM_IDLE;
 814         cm_id_priv->id.device = device;
 815         cm_id_priv->id.cm_handler = cm_handler;
 816         cm_id_priv->id.context = context;
 817         cm_id_priv->id.remote_cm_qpn = 1;
 818         ret = cm_alloc_id(cm_id_priv);
 819         if (ret)
 820                 goto error;
 821
 822         spin_lock_init(&cm_id_priv->lock);
 823         init_completion(&cm_id_priv->comp);
 824         INIT_LIST_HEAD(&cm_id_priv->work_list);
 825         INIT_LIST_HEAD(&cm_id_priv->prim_list);
 826         INIT_LIST_HEAD(&cm_id_priv->altr_list);
 827         atomic_set(&cm_id_priv->work_count, -1);
 828         atomic_set(&cm_id_priv->refcount, 1);
 829         return &cm_id_priv->id;
 830
 831 error:
 832         kfree(cm_id_priv);
 833         return ERR_PTR(-ENOMEM);
 834 }
 835 EXPORT_SYMBOL(ib_create_cm_id);
 836
 837 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 838 {
 839         struct cm_work *work;
 840
 841         if (list_empty(&cm_id_priv->work_list))
 842                 return NULL;
 843
 844         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 845         list_del(&work->list);
 846         return work;
 847 }
 848
 849 static void cm_free_work(struct cm_work *work)
 850 {
 851         if (work->mad_recv_wc)
 852                 ib_free_recv_mad(work->mad_recv_wc);
 853         kfree(work);
 854 }
 855
 856 static inline int cm_convert_to_ms(int iba_time)
 857 {
 858         /* approximate conversion to ms from 4.096us x 2^iba_time */
 859         return 1 << max(iba_time - 8, 0);
 860 }
 861
 862 /*
 863  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 864  * Because of how ack_timeout is stored, adding one doubles the timeout.
 865  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 866  * increment it (round up) only if the other is within 50%.
 867  */
 868 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 869 {
 870         int ack_timeout = packet_life_time + 1;
 871
 872         if (ack_timeout >= ca_ack_delay)
 873                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 874         else
 875                 ack_timeout = ca_ack_delay +
 876                               (ack_timeout >= (ca_ack_delay - 1));
 877
 878         return min(31, ack_timeout);
 879 }
 880
 881 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 882 {
 883         if (timewait_info->inserted_remote_id) {
 884                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 885                 timewait_info->inserted_remote_id = 0;
 886         }
 887
 888         if (timewait_info->inserted_remote_qp) {
 889                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 890                 timewait_info->inserted_remote_qp = 0;
 891         }
 892 }
 893
 894 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 895 {
 896         struct cm_timewait_info *timewait_info;
 897
 898         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 899         if (!timewait_info)
 900                 return ERR_PTR(-ENOMEM);
 901
 902         timewait_info->work.local_id = local_id;
 903         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 904         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 905         return timewait_info;
 906 }
 907
 908 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 909 {
 910         int wait_time;
 911         unsigned long flags;
 912         struct cm_device *cm_dev;
 913
 914         cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
 915         if (!cm_dev)
 916                 return;
 917
 918         spin_lock_irqsave(&cm.lock, flags);
 919         cm_cleanup_timewait(cm_id_priv->timewait_info);
 920         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 921         spin_unlock_irqrestore(&cm.lock, flags);
 922
 923         /*
 924          * The cm_id could be destroyed by the user before we exit timewait.
 925          * To protect against this, we search for the cm_id after exiting
 926          * timewait before notifying the user that we've exited timewait.
 927          */
 928         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 929         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 930
 931         /* Check if the device started its remove_one */
 932         spin_lock_irqsave(&cm.lock, flags);
 933         if (!cm_dev->going_down)
 934                 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 935                                    msecs_to_jiffies(wait_time));
 936         spin_unlock_irqrestore(&cm.lock, flags);
 937
 938         cm_id_priv->timewait_info = NULL;
 939 }
 940
 941 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 942 {
 943         unsigned long flags;
 944
 945         cm_id_priv->id.state = IB_CM_IDLE;
 946         if (cm_id_priv->timewait_info) {
 947                 spin_lock_irqsave(&cm.lock, flags);
 948                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 949                 spin_unlock_irqrestore(&cm.lock, flags);
 950                 kfree(cm_id_priv->timewait_info);
 951                 cm_id_priv->timewait_info = NULL;
 952         }
 953 }
 954
 955 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 956 {
 957         struct cm_id_private *cm_id_priv;
 958         struct cm_work *work;
 959
 960         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 961 retest:
 962         spin_lock_irq(&cm_id_priv->lock);
 963         switch (cm_id->state) {
 964         case IB_CM_LISTEN:
 965                 spin_unlock_irq(&cm_id_priv->lock);
 966
 967                 spin_lock_irq(&cm.lock);
 968                 if (--cm_id_priv->listen_sharecount > 0) {
 969                         /* The id is still shared. */
 970                         cm_deref_id(cm_id_priv);
 971                         spin_unlock_irq(&cm.lock);
 972                         return;
 973                 }
 974                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 975                 spin_unlock_irq(&cm.lock);
 976                 break;
 977         case IB_CM_SIDR_REQ_SENT:
 978                 cm_id->state = IB_CM_IDLE;
 979                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 980                 spin_unlock_irq(&cm_id_priv->lock);
 981                 break;
 982         case IB_CM_SIDR_REQ_RCVD:
 983                 spin_unlock_irq(&cm_id_priv->lock);
 984                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 985                 spin_lock_irq(&cm.lock);
 986                 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node))
 987                         rb_erase(&cm_id_priv->sidr_id_node,
 988                                  &cm.remote_sidr_table);
 989                 spin_unlock_irq(&cm.lock);
 990                 break;
 991         case IB_CM_REQ_SENT:
 992         case IB_CM_MRA_REQ_RCVD:
 993                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 994                 spin_unlock_irq(&cm_id_priv->lock);
 995                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 996                                &cm_id_priv->id.device->node_guid,
 997                                sizeof cm_id_priv->id.device->node_guid,
 998                                NULL, 0);
 999                 break;
1000         case IB_CM_REQ_RCVD:
1001                 if (err == -ENOMEM) {
1002                         /* Do not reject to allow future retries. */
1003                         cm_reset_to_idle(cm_id_priv);
1004                         spin_unlock_irq(&cm_id_priv->lock);
1005                 } else {
1006                         spin_unlock_irq(&cm_id_priv->lock);
1007                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1008                                        NULL, 0, NULL, 0);
1009                 }
1010                 break;
1011         case IB_CM_REP_SENT:
1012         case IB_CM_MRA_REP_RCVD:
1013                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1014                 /* Fall through */
1015         case IB_CM_MRA_REQ_SENT:
1016         case IB_CM_REP_RCVD:
1017         case IB_CM_MRA_REP_SENT:
1018                 spin_unlock_irq(&cm_id_priv->lock);
1019                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1020                                NULL, 0, NULL, 0);
1021                 break;
1022         case IB_CM_ESTABLISHED:
1023                 spin_unlock_irq(&cm_id_priv->lock);
1024                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT)
1025                         break;
1026                 ib_send_cm_dreq(cm_id, NULL, 0);
1027                 goto retest;
1028         case IB_CM_DREQ_SENT:
1029                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1030                 cm_enter_timewait(cm_id_priv);
1031                 spin_unlock_irq(&cm_id_priv->lock);
1032                 break;
1033         case IB_CM_DREQ_RCVD:
1034                 spin_unlock_irq(&cm_id_priv->lock);
1035                 ib_send_cm_drep(cm_id, NULL, 0);
1036                 break;
1037         default:
1038                 spin_unlock_irq(&cm_id_priv->lock);
1039                 break;
1040         }
1041
1042         spin_lock_irq(&cm.lock);
1043         if (!list_empty(&cm_id_priv->altr_list) &&
1044             (!cm_id_priv->altr_send_port_not_ready))
1045                 list_del(&cm_id_priv->altr_list);
1046         if (!list_empty(&cm_id_priv->prim_list) &&
1047             (!cm_id_priv->prim_send_port_not_ready))
1048                 list_del(&cm_id_priv->prim_list);
1049         spin_unlock_irq(&cm.lock);
1050
1051         cm_free_id(cm_id->local_id);
1052         cm_deref_id(cm_id_priv);
1053         wait_for_completion(&cm_id_priv->comp);
1054         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
1055                 cm_free_work(work);
1056         kfree(cm_id_priv->private_data);
1057         kfree(cm_id_priv);
1058 }
1059
1060 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
1061 {
1062         cm_destroy_id(cm_id, 0);
1063 }
1064 EXPORT_SYMBOL(ib_destroy_cm_id);
1065
1066 /**
1067  * __ib_cm_listen - Initiates listening on the specified service ID for
1068  *   connection and service ID resolution requests.
1069  * @cm_id: Connection identifier associated with the listen request.
1070  * @service_id: Service identifier matched against incoming connection
1071  *   and service ID resolution requests.  The service ID should be specified
1072  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1073  *   assign a service ID to the caller.
1074  * @service_mask: Mask applied to service ID used to listen across a
1075  *   range of service IDs.  If set to 0, the service ID is matched
1076  *   exactly.  This parameter is ignored if %service_id is set to
1077  *   IB_CM_ASSIGN_SERVICE_ID.
1078  */
1079 static int __ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id,
1080                           __be64 service_mask)
1081 {
1082         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
1083         int ret = 0;
1084
1085         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
1086         service_id &= service_mask;
1087         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
1088             (service_id != IB_CM_ASSIGN_SERVICE_ID))
1089                 return -EINVAL;
1090
1091         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1092         if (cm_id->state != IB_CM_IDLE)
1093                 return -EINVAL;
1094
1095         cm_id->state = IB_CM_LISTEN;
1096         ++cm_id_priv->listen_sharecount;
1097
1098         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
1099                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
1100                 cm_id->service_mask = ~cpu_to_be64(0);
1101         } else {
1102                 cm_id->service_id = service_id;
1103                 cm_id->service_mask = service_mask;
1104         }
1105         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
1106
1107         if (cur_cm_id_priv) {
1108                 cm_id->state = IB_CM_IDLE;
1109                 --cm_id_priv->listen_sharecount;
1110                 ret = -EBUSY;
1111         }
1112         return ret;
1113 }
1114
1115 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask)
1116 {
1117         unsigned long flags;
1118         int ret;
1119
1120         spin_lock_irqsave(&cm.lock, flags);
1121         ret = __ib_cm_listen(cm_id, service_id, service_mask);
1122         spin_unlock_irqrestore(&cm.lock, flags);
1123
1124         return ret;
1125 }
1126 EXPORT_SYMBOL(ib_cm_listen);
1127
1128 /**
1129  * Create a new listening ib_cm_id and listen on the given service ID.
1130  *
1131  * If there's an existing ID listening on that same device and service ID,
1132  * return it.
1133  *
1134  * @device: Device associated with the cm_id.  All related communication will
1135  * be associated with the specified device.
1136  * @cm_handler: Callback invoked to notify the user of CM events.
1137  * @service_id: Service identifier matched against incoming connection
1138  *   and service ID resolution requests.  The service ID should be specified
1139  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1140  *   assign a service ID to the caller.
1141  *
1142  * Callers should call ib_destroy_cm_id when done with the listener ID.
1143  */
1144 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1145                                      ib_cm_handler cm_handler,
1146                                      __be64 service_id)
1147 {
1148         struct cm_id_private *cm_id_priv;
1149         struct ib_cm_id *cm_id;
1150         unsigned long flags;
1151         int err = 0;
1152
1153         /* Create an ID in advance, since the creation may sleep */
1154         cm_id = ib_create_cm_id(device, cm_handler, NULL);
1155         if (IS_ERR(cm_id))
1156                 return cm_id;
1157
1158         spin_lock_irqsave(&cm.lock, flags);
1159
1160         if (service_id == IB_CM_ASSIGN_SERVICE_ID)
1161                 goto new_id;
1162
1163         /* Find an existing ID */
1164         cm_id_priv = cm_find_listen(device, service_id);
1165         if (cm_id_priv) {
1166                 if (cm_id->cm_handler != cm_handler || cm_id->context) {
1167                         /* Sharing an ib_cm_id with different handlers is not
1168                          * supported */
1169                         spin_unlock_irqrestore(&cm.lock, flags);
1170                         return ERR_PTR(-EINVAL);
1171                 }
1172                 atomic_inc(&cm_id_priv->refcount);
1173                 ++cm_id_priv->listen_sharecount;
1174                 spin_unlock_irqrestore(&cm.lock, flags);
1175
1176                 ib_destroy_cm_id(cm_id);
1177                 cm_id = &cm_id_priv->id;
1178                 return cm_id;
1179         }
1180
1181 new_id:
1182         /* Use newly created ID */
1183         err = __ib_cm_listen(cm_id, service_id, 0);
1184
1185         spin_unlock_irqrestore(&cm.lock, flags);
1186
1187         if (err) {
1188                 ib_destroy_cm_id(cm_id);
1189                 return ERR_PTR(err);
1190         }
1191         return cm_id;
1192 }
1193 EXPORT_SYMBOL(ib_cm_insert_listen);
1194
1195 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
1196                           enum cm_msg_sequence msg_seq)
1197 {
1198         u64 hi_tid, low_tid;
1199
1200         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1201         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
1202                           (msg_seq << 30));
1203         return cpu_to_be64(hi_tid | low_tid);
1204 }
1205
1206 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1207                               __be16 attr_id, __be64 tid)
1208 {
1209         hdr->base_version  = IB_MGMT_BASE_VERSION;
1210         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1211         hdr->class_version = IB_CM_CLASS_VERSION;
1212         hdr->method        = IB_MGMT_METHOD_SEND;
1213         hdr->attr_id       = attr_id;
1214         hdr->tid           = tid;
1215 }
1216
1217 static void cm_format_req(struct cm_req_msg *req_msg,
1218                           struct cm_id_private *cm_id_priv,
1219                           struct ib_cm_req_param *param)
1220 {
1221         struct sa_path_rec *pri_path = param->primary_path;
1222         struct sa_path_rec *alt_path = param->alternate_path;
1223         bool pri_ext = false;
1224
1225         if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA)
1226                 pri_ext = opa_is_extended_lid(pri_path->opa.dlid,
1227                                               pri_path->opa.slid);
1228
1229         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1230                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1231
1232         req_msg->local_comm_id = cm_id_priv->id.local_id;
1233         req_msg->service_id = param->service_id;
1234         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1235         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1236         cm_req_set_init_depth(req_msg, param->initiator_depth);
1237         cm_req_set_remote_resp_timeout(req_msg,
1238                                        param->remote_cm_response_timeout);
1239         cm_req_set_qp_type(req_msg, param->qp_type);
1240         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1241         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1242         cm_req_set_local_resp_timeout(req_msg,
1243                                       param->local_cm_response_timeout);
1244         req_msg->pkey = param->primary_path->pkey;
1245         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1246         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1247
1248         if (param->qp_type != IB_QPT_XRC_INI) {
1249                 cm_req_set_resp_res(req_msg, param->responder_resources);
1250                 cm_req_set_retry_count(req_msg, param->retry_count);
1251                 cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1252                 cm_req_set_srq(req_msg, param->srq);
1253         }
1254
1255         req_msg->primary_local_gid = pri_path->sgid;
1256         req_msg->primary_remote_gid = pri_path->dgid;
1257         if (pri_ext) {
1258                 req_msg->primary_local_gid.global.interface_id
1259                         = OPA_MAKE_ID(be32_to_cpu(pri_path->opa.slid));
1260                 req_msg->primary_remote_gid.global.interface_id
1261                         = OPA_MAKE_ID(be32_to_cpu(pri_path->opa.dlid));
1262         }
1263         if (pri_path->hop_limit <= 1) {
1264                 req_msg->primary_local_lid = pri_ext ? 0 :
1265                         htons(ntohl(sa_path_get_slid(pri_path)));
1266                 req_msg->primary_remote_lid = pri_ext ? 0 :
1267                         htons(ntohl(sa_path_get_dlid(pri_path)));
1268         } else {
1269                 /* Work-around until there's a way to obtain remote LID info */
1270                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1271                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1272         }
1273         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1274         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1275         req_msg->primary_traffic_class = pri_path->traffic_class;
1276         req_msg->primary_hop_limit = pri_path->hop_limit;
1277         cm_req_set_primary_sl(req_msg, pri_path->sl);
1278         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1279         cm_req_set_primary_local_ack_timeout(req_msg,
1280                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1281                                pri_path->packet_life_time));
1282
1283         if (alt_path) {
1284                 bool alt_ext = false;
1285
1286                 if (alt_path->rec_type == SA_PATH_REC_TYPE_OPA)
1287                         alt_ext = opa_is_extended_lid(alt_path->opa.dlid,
1288                                                       alt_path->opa.slid);
1289
1290                 req_msg->alt_local_gid = alt_path->sgid;
1291                 req_msg->alt_remote_gid = alt_path->dgid;
1292                 if (alt_ext) {
1293                         req_msg->alt_local_gid.global.interface_id
1294                                 = OPA_MAKE_ID(be32_to_cpu(alt_path->opa.slid));
1295                         req_msg->alt_remote_gid.global.interface_id
1296                                 = OPA_MAKE_ID(be32_to_cpu(alt_path->opa.dlid));
1297                 }
1298                 if (alt_path->hop_limit <= 1) {
1299                         req_msg->alt_local_lid = alt_ext ? 0 :
1300                                 htons(ntohl(sa_path_get_slid(alt_path)));
1301                         req_msg->alt_remote_lid = alt_ext ? 0 :
1302                                 htons(ntohl(sa_path_get_dlid(alt_path)));
1303                 } else {
1304                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1305                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1306                 }
1307                 cm_req_set_alt_flow_label(req_msg,
1308                                           alt_path->flow_label);
1309                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1310                 req_msg->alt_traffic_class = alt_path->traffic_class;
1311                 req_msg->alt_hop_limit = alt_path->hop_limit;
1312                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1313                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1314                 cm_req_set_alt_local_ack_timeout(req_msg,
1315                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1316                                        alt_path->packet_life_time));
1317         }
1318
1319         if (param->private_data && param->private_data_len)
1320                 memcpy(req_msg->private_data, param->private_data,
1321                        param->private_data_len);
1322 }
1323
1324 static int cm_validate_req_param(struct ib_cm_req_param *param)
1325 {
1326         /* peer-to-peer not supported */
1327         if (param->peer_to_peer)
1328                 return -EINVAL;
1329
1330         if (!param->primary_path)
1331                 return -EINVAL;
1332
1333         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1334             param->qp_type != IB_QPT_XRC_INI)
1335                 return -EINVAL;
1336
1337         if (param->private_data &&
1338             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1339                 return -EINVAL;
1340
1341         if (param->alternate_path &&
1342             (param->alternate_path->pkey != param->primary_path->pkey ||
1343              param->alternate_path->mtu != param->primary_path->mtu))
1344                 return -EINVAL;
1345
1346         return 0;
1347 }
1348
1349 int ib_send_cm_req(struct ib_cm_id *cm_id,
1350                    struct ib_cm_req_param *param)
1351 {
1352         struct cm_id_private *cm_id_priv;
1353         struct cm_req_msg *req_msg;
1354         unsigned long flags;
1355         int ret;
1356
1357         ret = cm_validate_req_param(param);
1358         if (ret)
1359                 return ret;
1360
1361         /* Verify that we're not in timewait. */
1362         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1363         spin_lock_irqsave(&cm_id_priv->lock, flags);
1364         if (cm_id->state != IB_CM_IDLE) {
1365                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1366                 ret = -EINVAL;
1367                 goto out;
1368         }
1369         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1370
1371         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1372                                                             id.local_id);
1373         if (IS_ERR(cm_id_priv->timewait_info)) {
1374                 ret = PTR_ERR(cm_id_priv->timewait_info);
1375                 goto out;
1376         }
1377
1378         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av,
1379                                  cm_id_priv);
1380         if (ret)
1381                 goto error1;
1382         if (param->alternate_path) {
1383                 ret = cm_init_av_by_path(param->alternate_path,
1384                                          &cm_id_priv->alt_av, cm_id_priv);
1385                 if (ret)
1386                         goto error1;
1387         }
1388         cm_id->service_id = param->service_id;
1389         cm_id->service_mask = ~cpu_to_be64(0);
1390         cm_id_priv->timeout_ms = cm_convert_to_ms(
1391                                     param->primary_path->packet_life_time) * 2 +
1392                                  cm_convert_to_ms(
1393                                     param->remote_cm_response_timeout);
1394         cm_id_priv->max_cm_retries = param->max_cm_retries;
1395         cm_id_priv->initiator_depth = param->initiator_depth;
1396         cm_id_priv->responder_resources = param->responder_resources;
1397         cm_id_priv->retry_count = param->retry_count;
1398         cm_id_priv->path_mtu = param->primary_path->mtu;
1399         cm_id_priv->pkey = param->primary_path->pkey;
1400         cm_id_priv->qp_type = param->qp_type;
1401
1402         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1403         if (ret)
1404                 goto error1;
1405
1406         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1407         cm_format_req(req_msg, cm_id_priv, param);
1408         cm_id_priv->tid = req_msg->hdr.tid;
1409         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1410         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1411
1412         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1413         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1414
1415         spin_lock_irqsave(&cm_id_priv->lock, flags);
1416         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1417         if (ret) {
1418                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1419                 goto error2;
1420         }
1421         BUG_ON(cm_id->state != IB_CM_IDLE);
1422         cm_id->state = IB_CM_REQ_SENT;
1423         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1424         return 0;
1425
1426 error2: cm_free_msg(cm_id_priv->msg);
1427 error1: kfree(cm_id_priv->timewait_info);
1428 out:    return ret;
1429 }
1430 EXPORT_SYMBOL(ib_send_cm_req);
1431
1432 static int cm_issue_rej(struct cm_port *port,
1433                         struct ib_mad_recv_wc *mad_recv_wc,
1434                         enum ib_cm_rej_reason reason,
1435                         enum cm_msg_response msg_rejected,
1436                         void *ari, u8 ari_length)
1437 {
1438         struct ib_mad_send_buf *msg = NULL;
1439         struct cm_rej_msg *rej_msg, *rcv_msg;
1440         int ret;
1441
1442         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1443         if (ret)
1444                 return ret;
1445
1446         /* We just need common CM header information.  Cast to any message. */
1447         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1448         rej_msg = (struct cm_rej_msg *) msg->mad;
1449
1450         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1451         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1452         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1453         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1454         rej_msg->reason = cpu_to_be16(reason);
1455
1456         if (ari && ari_length) {
1457                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1458                 memcpy(rej_msg->ari, ari, ari_length);
1459         }
1460
1461         ret = ib_post_send_mad(msg, NULL);
1462         if (ret)
1463                 cm_free_msg(msg);
1464
1465         return ret;
1466 }
1467
1468 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1469                                     __be32 local_qpn, __be32 remote_qpn)
1470 {
1471         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1472                 ((local_ca_guid == remote_ca_guid) &&
1473                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1474 }
1475
1476 static bool cm_req_has_alt_path(struct cm_req_msg *req_msg)
1477 {
1478         return ((req_msg->alt_local_lid) ||
1479                 (ib_is_opa_gid(&req_msg->alt_local_gid)));
1480 }
1481
1482 static void cm_path_set_rec_type(struct ib_device *ib_device, u8 port_num,
1483                                  struct sa_path_rec *path, union ib_gid *gid)
1484 {
1485         if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num))
1486                 path->rec_type = SA_PATH_REC_TYPE_OPA;
1487         else
1488                 path->rec_type = SA_PATH_REC_TYPE_IB;
1489 }
1490
1491 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
1492                                         struct sa_path_rec *primary_path,
1493                                         struct sa_path_rec *alt_path)
1494 {
1495         u32 lid;
1496
1497         if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1498                 sa_path_set_dlid(primary_path,
1499                                  ntohs(req_msg->primary_local_lid));
1500                 sa_path_set_slid(primary_path,
1501                                  ntohs(req_msg->primary_remote_lid));
1502         } else {
1503                 lid = opa_get_lid_from_gid(&req_msg->primary_local_gid);
1504                 sa_path_set_dlid(primary_path, lid);
1505
1506                 lid = opa_get_lid_from_gid(&req_msg->primary_remote_gid);
1507                 sa_path_set_slid(primary_path, lid);
1508         }
1509
1510         if (!cm_req_has_alt_path(req_msg))
1511                 return;
1512
1513         if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1514                 sa_path_set_dlid(alt_path, ntohs(req_msg->alt_local_lid));
1515                 sa_path_set_slid(alt_path, ntohs(req_msg->alt_remote_lid));
1516         } else {
1517                 lid = opa_get_lid_from_gid(&req_msg->alt_local_gid);
1518                 sa_path_set_dlid(alt_path, lid);
1519
1520                 lid = opa_get_lid_from_gid(&req_msg->alt_remote_gid);
1521                 sa_path_set_slid(alt_path, lid);
1522         }
1523 }
1524
1525 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1526                                      struct sa_path_rec *primary_path,
1527                                      struct sa_path_rec *alt_path)
1528 {
1529         primary_path->dgid = req_msg->primary_local_gid;
1530         primary_path->sgid = req_msg->primary_remote_gid;
1531         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1532         primary_path->hop_limit = req_msg->primary_hop_limit;
1533         primary_path->traffic_class = req_msg->primary_traffic_class;
1534         primary_path->reversible = 1;
1535         primary_path->pkey = req_msg->pkey;
1536         primary_path->sl = cm_req_get_primary_sl(req_msg);
1537         primary_path->mtu_selector = IB_SA_EQ;
1538         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1539         primary_path->rate_selector = IB_SA_EQ;
1540         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1541         primary_path->packet_life_time_selector = IB_SA_EQ;
1542         primary_path->packet_life_time =
1543                 cm_req_get_primary_local_ack_timeout(req_msg);
1544         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1545         primary_path->service_id = req_msg->service_id;
1546         if (sa_path_is_roce(primary_path))
1547                 primary_path->roce.route_resolved = false;
1548
1549         if (cm_req_has_alt_path(req_msg)) {
1550                 alt_path->dgid = req_msg->alt_local_gid;
1551                 alt_path->sgid = req_msg->alt_remote_gid;
1552                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1553                 alt_path->hop_limit = req_msg->alt_hop_limit;
1554                 alt_path->traffic_class = req_msg->alt_traffic_class;
1555                 alt_path->reversible = 1;
1556                 alt_path->pkey = req_msg->pkey;
1557                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1558                 alt_path->mtu_selector = IB_SA_EQ;
1559                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1560                 alt_path->rate_selector = IB_SA_EQ;
1561                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1562                 alt_path->packet_life_time_selector = IB_SA_EQ;
1563                 alt_path->packet_life_time =
1564                         cm_req_get_alt_local_ack_timeout(req_msg);
1565                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1566                 alt_path->service_id = req_msg->service_id;
1567
1568                 if (sa_path_is_roce(alt_path))
1569                         alt_path->roce.route_resolved = false;
1570         }
1571         cm_format_path_lid_from_req(req_msg, primary_path, alt_path);
1572 }
1573
1574 static u16 cm_get_bth_pkey(struct cm_work *work)
1575 {
1576         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1577         u8 port_num = work->port->port_num;
1578         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1579         u16 pkey;
1580         int ret;
1581
1582         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1583         if (ret) {
1584                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1585                                      port_num, pkey_index, ret);
1586                 return 0;
1587         }
1588
1589         return pkey;
1590 }
1591
1592 /**
1593  * Convert OPA SGID to IB SGID
1594  * ULPs (such as IPoIB) do not understand OPA GIDs and will
1595  * reject them as the local_gid will not match the sgid. Therefore,
1596  * change the pathrec's SGID to an IB SGID.
1597  *
1598  * @work: Work completion
1599  * @path: Path record
1600  */
1601 static void cm_opa_to_ib_sgid(struct cm_work *work,
1602                               struct sa_path_rec *path)
1603 {
1604         struct ib_device *dev = work->port->cm_dev->ib_device;
1605         u8 port_num = work->port->port_num;
1606
1607         if (rdma_cap_opa_ah(dev, port_num) &&
1608             (ib_is_opa_gid(&path->sgid))) {
1609                 union ib_gid sgid;
1610
1611                 if (ib_get_cached_gid(dev, port_num, 0, &sgid, NULL)) {
1612                         dev_warn(&dev->dev,
1613                                  "Error updating sgid in CM request\n");
1614                         return;
1615                 }
1616
1617                 path->sgid = sgid;
1618         }
1619 }
1620
1621 static void cm_format_req_event(struct cm_work *work,
1622                                 struct cm_id_private *cm_id_priv,
1623                                 struct ib_cm_id *listen_id)
1624 {
1625         struct cm_req_msg *req_msg;
1626         struct ib_cm_req_event_param *param;
1627
1628         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1629         param = &work->cm_event.param.req_rcvd;
1630         param->listen_id = listen_id;
1631         param->bth_pkey = cm_get_bth_pkey(work);
1632         param->port = cm_id_priv->av.port->port_num;
1633         param->primary_path = &work->path[0];
1634         cm_opa_to_ib_sgid(work, param->primary_path);
1635         if (cm_req_has_alt_path(req_msg)) {
1636                 param->alternate_path = &work->path[1];
1637                 cm_opa_to_ib_sgid(work, param->alternate_path);
1638         } else {
1639                 param->alternate_path = NULL;
1640         }
1641         param->remote_ca_guid = req_msg->local_ca_guid;
1642         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1643         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1644         param->qp_type = cm_req_get_qp_type(req_msg);
1645         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1646         param->responder_resources = cm_req_get_init_depth(req_msg);
1647         param->initiator_depth = cm_req_get_resp_res(req_msg);
1648         param->local_cm_response_timeout =
1649                                         cm_req_get_remote_resp_timeout(req_msg);
1650         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1651         param->remote_cm_response_timeout =
1652                                         cm_req_get_local_resp_timeout(req_msg);
1653         param->retry_count = cm_req_get_retry_count(req_msg);
1654         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1655         param->srq = cm_req_get_srq(req_msg);
1656         work->cm_event.private_data = &req_msg->private_data;
1657 }
1658
1659 static void cm_process_work(struct cm_id_private *cm_id_priv,
1660                             struct cm_work *work)
1661 {
1662         int ret;
1663
1664         /* We will typically only have the current event to report. */
1665         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1666         cm_free_work(work);
1667
1668         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1669                 spin_lock_irq(&cm_id_priv->lock);
1670                 work = cm_dequeue_work(cm_id_priv);
1671                 spin_unlock_irq(&cm_id_priv->lock);
1672                 BUG_ON(!work);
1673                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1674                                                 &work->cm_event);
1675                 cm_free_work(work);
1676         }
1677         cm_deref_id(cm_id_priv);
1678         if (ret)
1679                 cm_destroy_id(&cm_id_priv->id, ret);
1680 }
1681
1682 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1683                           struct cm_id_private *cm_id_priv,
1684                           enum cm_msg_response msg_mraed, u8 service_timeout,
1685                           const void *private_data, u8 private_data_len)
1686 {
1687         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1688         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1689         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1690         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1691         cm_mra_set_service_timeout(mra_msg, service_timeout);
1692
1693         if (private_data && private_data_len)
1694                 memcpy(mra_msg->private_data, private_data, private_data_len);
1695 }
1696
1697 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1698                           struct cm_id_private *cm_id_priv,
1699                           enum ib_cm_rej_reason reason,
1700                           void *ari,
1701                           u8 ari_length,
1702                           const void *private_data,
1703                           u8 private_data_len)
1704 {
1705         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1706         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1707
1708         switch(cm_id_priv->id.state) {
1709         case IB_CM_REQ_RCVD:
1710                 rej_msg->local_comm_id = 0;
1711                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1712                 break;
1713         case IB_CM_MRA_REQ_SENT:
1714                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1715                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1716                 break;
1717         case IB_CM_REP_RCVD:
1718         case IB_CM_MRA_REP_SENT:
1719                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1720                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1721                 break;
1722         default:
1723                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1724                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1725                 break;
1726         }
1727
1728         rej_msg->reason = cpu_to_be16(reason);
1729         if (ari && ari_length) {
1730                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1731                 memcpy(rej_msg->ari, ari, ari_length);
1732         }
1733
1734         if (private_data && private_data_len)
1735                 memcpy(rej_msg->private_data, private_data, private_data_len);
1736 }
1737
1738 static void cm_dup_req_handler(struct cm_work *work,
1739                                struct cm_id_private *cm_id_priv)
1740 {
1741         struct ib_mad_send_buf *msg = NULL;
1742         int ret;
1743
1744         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1745                         counter[CM_REQ_COUNTER]);
1746
1747         /* Quick state check to discard duplicate REQs. */
1748         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1749                 return;
1750
1751         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1752         if (ret)
1753                 return;
1754
1755         spin_lock_irq(&cm_id_priv->lock);
1756         switch (cm_id_priv->id.state) {
1757         case IB_CM_MRA_REQ_SENT:
1758                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1759                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1760                               cm_id_priv->private_data,
1761                               cm_id_priv->private_data_len);
1762                 break;
1763         case IB_CM_TIMEWAIT:
1764                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1765                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1766                 break;
1767         default:
1768                 goto unlock;
1769         }
1770         spin_unlock_irq(&cm_id_priv->lock);
1771
1772         ret = ib_post_send_mad(msg, NULL);
1773         if (ret)
1774                 goto free;
1775         return;
1776
1777 unlock: spin_unlock_irq(&cm_id_priv->lock);
1778 free:   cm_free_msg(msg);
1779 }
1780
1781 static struct cm_id_private * cm_match_req(struct cm_work *work,
1782                                            struct cm_id_private *cm_id_priv)
1783 {
1784         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1785         struct cm_timewait_info *timewait_info;
1786         struct cm_req_msg *req_msg;
1787         struct ib_cm_id *cm_id;
1788
1789         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1790
1791         /* Check for possible duplicate REQ. */
1792         spin_lock_irq(&cm.lock);
1793         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1794         if (timewait_info) {
1795                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1796                                            timewait_info->work.remote_id);
1797                 spin_unlock_irq(&cm.lock);
1798                 if (cur_cm_id_priv) {
1799                         cm_dup_req_handler(work, cur_cm_id_priv);
1800                         cm_deref_id(cur_cm_id_priv);
1801                 }
1802                 return NULL;
1803         }
1804
1805         /* Check for stale connections. */
1806         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1807         if (timewait_info) {
1808                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1809                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1810                                            timewait_info->work.remote_id);
1811
1812                 spin_unlock_irq(&cm.lock);
1813                 cm_issue_rej(work->port, work->mad_recv_wc,
1814                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1815                              NULL, 0);
1816                 if (cur_cm_id_priv) {
1817                         cm_id = &cur_cm_id_priv->id;
1818                         ib_send_cm_dreq(cm_id, NULL, 0);
1819                         cm_deref_id(cur_cm_id_priv);
1820                 }
1821                 return NULL;
1822         }
1823
1824         /* Find matching listen request. */
1825         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1826                                            req_msg->service_id);
1827         if (!listen_cm_id_priv) {
1828                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1829                 spin_unlock_irq(&cm.lock);
1830                 cm_issue_rej(work->port, work->mad_recv_wc,
1831                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1832                              NULL, 0);
1833                 goto out;
1834         }
1835         atomic_inc(&listen_cm_id_priv->refcount);
1836         atomic_inc(&cm_id_priv->refcount);
1837         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1838         atomic_inc(&cm_id_priv->work_count);
1839         spin_unlock_irq(&cm.lock);
1840 out:
1841         return listen_cm_id_priv;
1842 }
1843
1844 /*
1845  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1846  * we need to override the LID/SL data in the REQ with the LID information
1847  * in the work completion.
1848  */
1849 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1850 {
1851         if (!cm_req_get_primary_subnet_local(req_msg)) {
1852                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1853                         req_msg->primary_local_lid = ib_lid_be16(wc->slid);
1854                         cm_req_set_primary_sl(req_msg, wc->sl);
1855                 }
1856
1857                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1858                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1859         }
1860
1861         if (!cm_req_get_alt_subnet_local(req_msg)) {
1862                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1863                         req_msg->alt_local_lid = ib_lid_be16(wc->slid);
1864                         cm_req_set_alt_sl(req_msg, wc->sl);
1865                 }
1866
1867                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1868                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1869         }
1870 }
1871
1872 static int cm_req_handler(struct cm_work *work)
1873 {
1874         struct ib_cm_id *cm_id;
1875         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1876         struct cm_req_msg *req_msg;
1877         union ib_gid gid;
1878         struct ib_gid_attr gid_attr;
1879         const struct ib_global_route *grh;
1880         int ret;
1881
1882         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1883
1884         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1885         if (IS_ERR(cm_id))
1886                 return PTR_ERR(cm_id);
1887
1888         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1889         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1890         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1891                                       work->mad_recv_wc->recv_buf.grh,
1892                                       &cm_id_priv->av);
1893         if (ret)
1894                 goto destroy;
1895         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1896                                                             id.local_id);
1897         if (IS_ERR(cm_id_priv->timewait_info)) {
1898                 ret = PTR_ERR(cm_id_priv->timewait_info);
1899                 goto destroy;
1900         }
1901         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1902         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1903         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1904
1905         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1906         if (!listen_cm_id_priv) {
1907                 pr_debug("%s: local_id %d, no listen_cm_id_priv\n", __func__,
1908                          be32_to_cpu(cm_id->local_id));
1909                 ret = -EINVAL;
1910                 goto free_timeinfo;
1911         }
1912
1913         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1914         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1915         cm_id_priv->id.service_id = req_msg->service_id;
1916         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1917
1918         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1919
1920         memset(&work->path[0], 0, sizeof(work->path[0]));
1921         if (cm_req_has_alt_path(req_msg))
1922                 memset(&work->path[1], 0, sizeof(work->path[1]));
1923         grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
1924         ret = ib_get_cached_gid(work->port->cm_dev->ib_device,
1925                                 work->port->port_num,
1926                                 grh->sgid_index,
1927                                 &gid, &gid_attr);
1928         if (ret) {
1929                 ib_send_cm_rej(cm_id, IB_CM_REJ_UNSUPPORTED, NULL, 0, NULL, 0);
1930                 goto rejected;
1931         }
1932
1933         if (gid_attr.ndev) {
1934                 work->path[0].rec_type =
1935                         sa_conv_gid_to_pathrec_type(gid_attr.gid_type);
1936                 sa_path_set_ifindex(&work->path[0],
1937                                     gid_attr.ndev->ifindex);
1938                 sa_path_set_ndev(&work->path[0],
1939                                  dev_net(gid_attr.ndev));
1940                 dev_put(gid_attr.ndev);
1941         } else {
1942                 cm_path_set_rec_type(work->port->cm_dev->ib_device,
1943                                      work->port->port_num,
1944                                      &work->path[0],
1945                                      &req_msg->primary_local_gid);
1946         }
1947         if (cm_req_has_alt_path(req_msg))
1948                 work->path[1].rec_type = work->path[0].rec_type;
1949         cm_format_paths_from_req(req_msg, &work->path[0],
1950                                  &work->path[1]);
1951         if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
1952                 sa_path_set_dmac(&work->path[0],
1953                                  cm_id_priv->av.ah_attr.roce.dmac);
1954         work->path[0].hop_limit = grh->hop_limit;
1955         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av,
1956                                  cm_id_priv);
1957         if (ret) {
1958                 int err;
1959
1960                 err = ib_get_cached_gid(work->port->cm_dev->ib_device,
1961                                         work->port->port_num, 0,
1962                                         &work->path[0].sgid,
1963                                         NULL);
1964                 if (err)
1965                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1966                                        NULL, 0, NULL, 0);
1967                 else
1968                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1969                                        &work->path[0].sgid,
1970                                        sizeof(work->path[0].sgid),
1971                                        NULL, 0);
1972                 goto rejected;
1973         }
1974         if (cm_req_has_alt_path(req_msg)) {
1975                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av,
1976                                          cm_id_priv);
1977                 if (ret) {
1978                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1979                                        &work->path[0].sgid,
1980                                        sizeof(work->path[0].sgid), NULL, 0);
1981                         goto rejected;
1982                 }
1983         }
1984         cm_id_priv->tid = req_msg->hdr.tid;
1985         cm_id_priv->timeout_ms = cm_convert_to_ms(
1986                                         cm_req_get_local_resp_timeout(req_msg));
1987         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1988         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1989         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1990         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1991         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1992         cm_id_priv->pkey = req_msg->pkey;
1993         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1994         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1995         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1996         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1997
1998         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1999         cm_process_work(cm_id_priv, work);
2000         cm_deref_id(listen_cm_id_priv);
2001         return 0;
2002
2003 rejected:
2004         atomic_dec(&cm_id_priv->refcount);
2005         cm_deref_id(listen_cm_id_priv);
2006 free_timeinfo:
2007         kfree(cm_id_priv->timewait_info);
2008 destroy:
2009         ib_destroy_cm_id(cm_id);
2010         return ret;
2011 }
2012
2013 static void cm_format_rep(struct cm_rep_msg *rep_msg,
2014                           struct cm_id_private *cm_id_priv,
2015                           struct ib_cm_rep_param *param)
2016 {
2017         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
2018         rep_msg->local_comm_id = cm_id_priv->id.local_id;
2019         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2020         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
2021         rep_msg->resp_resources = param->responder_resources;
2022         cm_rep_set_target_ack_delay(rep_msg,
2023                                     cm_id_priv->av.port->cm_dev->ack_delay);
2024         cm_rep_set_failover(rep_msg, param->failover_accepted);
2025         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
2026         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
2027
2028         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
2029                 rep_msg->initiator_depth = param->initiator_depth;
2030                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
2031                 cm_rep_set_srq(rep_msg, param->srq);
2032                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
2033         } else {
2034                 cm_rep_set_srq(rep_msg, 1);
2035                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
2036         }
2037
2038         if (param->private_data && param->private_data_len)
2039                 memcpy(rep_msg->private_data, param->private_data,
2040                        param->private_data_len);
2041 }
2042
2043 int ib_send_cm_rep(struct ib_cm_id *cm_id,
2044                    struct ib_cm_rep_param *param)
2045 {
2046         struct cm_id_private *cm_id_priv;
2047         struct ib_mad_send_buf *msg;
2048         struct cm_rep_msg *rep_msg;
2049         unsigned long flags;
2050         int ret;
2051
2052         if (param->private_data &&
2053             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
2054                 return -EINVAL;
2055
2056         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2057         spin_lock_irqsave(&cm_id_priv->lock, flags);
2058         if (cm_id->state != IB_CM_REQ_RCVD &&
2059             cm_id->state != IB_CM_MRA_REQ_SENT) {
2060                 pr_debug("%s: local_comm_id %d, cm_id->state: %d\n", __func__,
2061                          be32_to_cpu(cm_id_priv->id.local_id), cm_id->state);
2062                 ret = -EINVAL;
2063                 goto out;
2064         }
2065
2066         ret = cm_alloc_msg(cm_id_priv, &msg);
2067         if (ret)
2068                 goto out;
2069
2070         rep_msg = (struct cm_rep_msg *) msg->mad;
2071         cm_format_rep(rep_msg, cm_id_priv, param);
2072         msg->timeout_ms = cm_id_priv->timeout_ms;
2073         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
2074
2075         ret = ib_post_send_mad(msg, NULL);
2076         if (ret) {
2077                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2078                 cm_free_msg(msg);
2079                 return ret;
2080         }
2081
2082         cm_id->state = IB_CM_REP_SENT;
2083         cm_id_priv->msg = msg;
2084         cm_id_priv->initiator_depth = param->initiator_depth;
2085         cm_id_priv->responder_resources = param->responder_resources;
2086         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
2087         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
2088
2089 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2090         return ret;
2091 }
2092 EXPORT_SYMBOL(ib_send_cm_rep);
2093
2094 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
2095                           struct cm_id_private *cm_id_priv,
2096                           const void *private_data,
2097                           u8 private_data_len)
2098 {
2099         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
2100         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
2101         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
2102
2103         if (private_data && private_data_len)
2104                 memcpy(rtu_msg->private_data, private_data, private_data_len);
2105 }
2106
2107 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
2108                    const void *private_data,
2109                    u8 private_data_len)
2110 {
2111         struct cm_id_private *cm_id_priv;
2112         struct ib_mad_send_buf *msg;
2113         unsigned long flags;
2114         void *data;
2115         int ret;
2116
2117         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
2118                 return -EINVAL;
2119
2120         data = cm_copy_private_data(private_data, private_data_len);
2121         if (IS_ERR(data))
2122                 return PTR_ERR(data);
2123
2124         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2125         spin_lock_irqsave(&cm_id_priv->lock, flags);
2126         if (cm_id->state != IB_CM_REP_RCVD &&
2127             cm_id->state != IB_CM_MRA_REP_SENT) {
2128                 pr_debug("%s: local_id %d, cm_id->state %d\n", __func__,
2129                          be32_to_cpu(cm_id->local_id), cm_id->state);
2130                 ret = -EINVAL;
2131                 goto error;
2132         }
2133
2134         ret = cm_alloc_msg(cm_id_priv, &msg);
2135         if (ret)
2136                 goto error;
2137
2138         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2139                       private_data, private_data_len);
2140
2141         ret = ib_post_send_mad(msg, NULL);
2142         if (ret) {
2143                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2144                 cm_free_msg(msg);
2145                 kfree(data);
2146                 return ret;
2147         }
2148
2149         cm_id->state = IB_CM_ESTABLISHED;
2150         cm_set_private_data(cm_id_priv, data, private_data_len);
2151         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2152         return 0;
2153
2154 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2155         kfree(data);
2156         return ret;
2157 }
2158 EXPORT_SYMBOL(ib_send_cm_rtu);
2159
2160 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
2161 {
2162         struct cm_rep_msg *rep_msg;
2163         struct ib_cm_rep_event_param *param;
2164
2165         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2166         param = &work->cm_event.param.rep_rcvd;
2167         param->remote_ca_guid = rep_msg->local_ca_guid;
2168         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
2169         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
2170         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
2171         param->responder_resources = rep_msg->initiator_depth;
2172         param->initiator_depth = rep_msg->resp_resources;
2173         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2174         param->failover_accepted = cm_rep_get_failover(rep_msg);
2175         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
2176         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2177         param->srq = cm_rep_get_srq(rep_msg);
2178         work->cm_event.private_data = &rep_msg->private_data;
2179 }
2180
2181 static void cm_dup_rep_handler(struct cm_work *work)
2182 {
2183         struct cm_id_private *cm_id_priv;
2184         struct cm_rep_msg *rep_msg;
2185         struct ib_mad_send_buf *msg = NULL;
2186         int ret;
2187
2188         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2189         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
2190                                    rep_msg->local_comm_id);
2191         if (!cm_id_priv)
2192                 return;
2193
2194         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2195                         counter[CM_REP_COUNTER]);
2196         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2197         if (ret)
2198                 goto deref;
2199
2200         spin_lock_irq(&cm_id_priv->lock);
2201         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2202                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2203                               cm_id_priv->private_data,
2204                               cm_id_priv->private_data_len);
2205         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2206                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2207                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2208                               cm_id_priv->private_data,
2209                               cm_id_priv->private_data_len);
2210         else
2211                 goto unlock;
2212         spin_unlock_irq(&cm_id_priv->lock);
2213
2214         ret = ib_post_send_mad(msg, NULL);
2215         if (ret)
2216                 goto free;
2217         goto deref;
2218
2219 unlock: spin_unlock_irq(&cm_id_priv->lock);
2220 free:   cm_free_msg(msg);
2221 deref:  cm_deref_id(cm_id_priv);
2222 }
2223
2224 static int cm_rep_handler(struct cm_work *work)
2225 {
2226         struct cm_id_private *cm_id_priv;
2227         struct cm_rep_msg *rep_msg;
2228         int ret;
2229         struct cm_id_private *cur_cm_id_priv;
2230         struct ib_cm_id *cm_id;
2231         struct cm_timewait_info *timewait_info;
2232
2233         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2234         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
2235         if (!cm_id_priv) {
2236                 cm_dup_rep_handler(work);
2237                 pr_debug("%s: remote_comm_id %d, no cm_id_priv\n", __func__,
2238                          be32_to_cpu(rep_msg->remote_comm_id));
2239                 return -EINVAL;
2240         }
2241
2242         cm_format_rep_event(work, cm_id_priv->qp_type);
2243
2244         spin_lock_irq(&cm_id_priv->lock);
2245         switch (cm_id_priv->id.state) {
2246         case IB_CM_REQ_SENT:
2247         case IB_CM_MRA_REQ_RCVD:
2248                 break;
2249         default:
2250                 spin_unlock_irq(&cm_id_priv->lock);
2251                 ret = -EINVAL;
2252                 pr_debug("%s: cm_id_priv->id.state: %d, local_comm_id %d, remote_comm_id %d\n",
2253                          __func__, cm_id_priv->id.state,
2254                          be32_to_cpu(rep_msg->local_comm_id),
2255                          be32_to_cpu(rep_msg->remote_comm_id));
2256                 goto error;
2257         }
2258
2259         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
2260         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
2261         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2262
2263         spin_lock(&cm.lock);
2264         /* Check for duplicate REP. */
2265         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2266                 spin_unlock(&cm.lock);
2267                 spin_unlock_irq(&cm_id_priv->lock);
2268                 ret = -EINVAL;
2269                 pr_debug("%s: Failed to insert remote id %d\n", __func__,
2270                          be32_to_cpu(rep_msg->remote_comm_id));
2271                 goto error;
2272         }
2273         /* Check for a stale connection. */
2274         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2275         if (timewait_info) {
2276                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
2277                          &cm.remote_id_table);
2278                 cm_id_priv->timewait_info->inserted_remote_id = 0;
2279                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
2280                                            timewait_info->work.remote_id);
2281
2282                 spin_unlock(&cm.lock);
2283                 spin_unlock_irq(&cm_id_priv->lock);
2284                 cm_issue_rej(work->port, work->mad_recv_wc,
2285                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2286                              NULL, 0);
2287                 ret = -EINVAL;
2288                 pr_debug("%s: Stale connection. local_comm_id %d, remote_comm_id %d\n",
2289                          __func__, be32_to_cpu(rep_msg->local_comm_id),
2290                          be32_to_cpu(rep_msg->remote_comm_id));
2291
2292                 if (cur_cm_id_priv) {
2293                         cm_id = &cur_cm_id_priv->id;
2294                         ib_send_cm_dreq(cm_id, NULL, 0);
2295                         cm_deref_id(cur_cm_id_priv);
2296                 }
2297
2298                 goto error;
2299         }
2300         spin_unlock(&cm.lock);
2301
2302         cm_id_priv->id.state = IB_CM_REP_RCVD;
2303         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
2304         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2305         cm_id_priv->initiator_depth = rep_msg->resp_resources;
2306         cm_id_priv->responder_resources = rep_msg->initiator_depth;
2307         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
2308         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2309         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2310         cm_id_priv->av.timeout =
2311                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2312                                        cm_id_priv->av.timeout - 1);
2313         cm_id_priv->alt_av.timeout =
2314                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2315                                        cm_id_priv->alt_av.timeout - 1);
2316
2317         /* todo: handle peer_to_peer */
2318
2319         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2320         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2321         if (!ret)
2322                 list_add_tail(&work->list, &cm_id_priv->work_list);
2323         spin_unlock_irq(&cm_id_priv->lock);
2324
2325         if (ret)
2326                 cm_process_work(cm_id_priv, work);
2327         else
2328                 cm_deref_id(cm_id_priv);
2329         return 0;
2330
2331 error:
2332         cm_deref_id(cm_id_priv);
2333         return ret;
2334 }
2335
2336 static int cm_establish_handler(struct cm_work *work)
2337 {
2338         struct cm_id_private *cm_id_priv;
2339         int ret;
2340
2341         /* See comment in cm_establish about lookup. */
2342         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2343         if (!cm_id_priv)
2344                 return -EINVAL;
2345
2346         spin_lock_irq(&cm_id_priv->lock);
2347         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2348                 spin_unlock_irq(&cm_id_priv->lock);
2349                 goto out;
2350         }
2351
2352         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2353         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2354         if (!ret)
2355                 list_add_tail(&work->list, &cm_id_priv->work_list);
2356         spin_unlock_irq(&cm_id_priv->lock);
2357
2358         if (ret)
2359                 cm_process_work(cm_id_priv, work);
2360         else
2361                 cm_deref_id(cm_id_priv);
2362         return 0;
2363 out:
2364         cm_deref_id(cm_id_priv);
2365         return -EINVAL;
2366 }
2367
2368 static int cm_rtu_handler(struct cm_work *work)
2369 {
2370         struct cm_id_private *cm_id_priv;
2371         struct cm_rtu_msg *rtu_msg;
2372         int ret;
2373
2374         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2375         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
2376                                    rtu_msg->local_comm_id);
2377         if (!cm_id_priv)
2378                 return -EINVAL;
2379
2380         work->cm_event.private_data = &rtu_msg->private_data;
2381
2382         spin_lock_irq(&cm_id_priv->lock);
2383         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2384             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2385                 spin_unlock_irq(&cm_id_priv->lock);
2386                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2387                                 counter[CM_RTU_COUNTER]);
2388                 goto out;
2389         }
2390         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2391
2392         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2393         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2394         if (!ret)
2395                 list_add_tail(&work->list, &cm_id_priv->work_list);
2396         spin_unlock_irq(&cm_id_priv->lock);
2397
2398         if (ret)
2399                 cm_process_work(cm_id_priv, work);
2400         else
2401                 cm_deref_id(cm_id_priv);
2402         return 0;
2403 out:
2404         cm_deref_id(cm_id_priv);
2405         return -EINVAL;
2406 }
2407
2408 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2409                           struct cm_id_private *cm_id_priv,
2410                           const void *private_data,
2411                           u8 private_data_len)
2412 {
2413         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2414                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
2415         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
2416         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
2417         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2418
2419         if (private_data && private_data_len)
2420                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2421 }
2422
2423 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2424                     const void *private_data,
2425                     u8 private_data_len)
2426 {
2427         struct cm_id_private *cm_id_priv;
2428         struct ib_mad_send_buf *msg;
2429         unsigned long flags;
2430         int ret;
2431
2432         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2433                 return -EINVAL;
2434
2435         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2436         spin_lock_irqsave(&cm_id_priv->lock, flags);
2437         if (cm_id->state != IB_CM_ESTABLISHED) {
2438                 pr_debug("%s: local_id %d, cm_id->state: %d\n", __func__,
2439                          be32_to_cpu(cm_id->local_id), cm_id->state);
2440                 ret = -EINVAL;
2441                 goto out;
2442         }
2443
2444         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2445             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2446                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2447
2448         ret = cm_alloc_msg(cm_id_priv, &msg);
2449         if (ret) {
2450                 cm_enter_timewait(cm_id_priv);
2451                 goto out;
2452         }
2453
2454         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2455                        private_data, private_data_len);
2456         msg->timeout_ms = cm_id_priv->timeout_ms;
2457         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2458
2459         ret = ib_post_send_mad(msg, NULL);
2460         if (ret) {
2461                 cm_enter_timewait(cm_id_priv);
2462                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2463                 cm_free_msg(msg);
2464                 return ret;
2465         }
2466
2467         cm_id->state = IB_CM_DREQ_SENT;
2468         cm_id_priv->msg = msg;
2469 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2470         return ret;
2471 }
2472 EXPORT_SYMBOL(ib_send_cm_dreq);
2473
2474 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2475                           struct cm_id_private *cm_id_priv,
2476                           const void *private_data,
2477                           u8 private_data_len)
2478 {
2479         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2480         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2481         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2482
2483         if (private_data && private_data_len)
2484                 memcpy(drep_msg->private_data, private_data, private_data_len);
2485 }
2486
2487 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2488                     const void *private_data,
2489                     u8 private_data_len)
2490 {
2491         struct cm_id_private *cm_id_priv;
2492         struct ib_mad_send_buf *msg;
2493         unsigned long flags;
2494         void *data;
2495         int ret;
2496
2497         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2498                 return -EINVAL;
2499
2500         data = cm_copy_private_data(private_data, private_data_len);
2501         if (IS_ERR(data))
2502                 return PTR_ERR(data);
2503
2504         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2505         spin_lock_irqsave(&cm_id_priv->lock, flags);
2506         if (cm_id->state != IB_CM_DREQ_RCVD) {
2507                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2508                 kfree(data);
2509                 pr_debug("%s: local_id %d, cm_idcm_id->state(%d) != IB_CM_DREQ_RCVD\n",
2510                          __func__, be32_to_cpu(cm_id->local_id), cm_id->state);
2511                 return -EINVAL;
2512         }
2513
2514         cm_set_private_data(cm_id_priv, data, private_data_len);
2515         cm_enter_timewait(cm_id_priv);
2516
2517         ret = cm_alloc_msg(cm_id_priv, &msg);
2518         if (ret)
2519                 goto out;
2520
2521         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2522                        private_data, private_data_len);
2523
2524         ret = ib_post_send_mad(msg, NULL);
2525         if (ret) {
2526                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2527                 cm_free_msg(msg);
2528                 return ret;
2529         }
2530
2531 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2532         return ret;
2533 }
2534 EXPORT_SYMBOL(ib_send_cm_drep);
2535
2536 static int cm_issue_drep(struct cm_port *port,
2537                          struct ib_mad_recv_wc *mad_recv_wc)
2538 {
2539         struct ib_mad_send_buf *msg = NULL;
2540         struct cm_dreq_msg *dreq_msg;
2541         struct cm_drep_msg *drep_msg;
2542         int ret;
2543
2544         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2545         if (ret)
2546                 return ret;
2547
2548         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2549         drep_msg = (struct cm_drep_msg *) msg->mad;
2550
2551         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2552         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2553         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2554
2555         ret = ib_post_send_mad(msg, NULL);
2556         if (ret)
2557                 cm_free_msg(msg);
2558
2559         return ret;
2560 }
2561
2562 static int cm_dreq_handler(struct cm_work *work)
2563 {
2564         struct cm_id_private *cm_id_priv;
2565         struct cm_dreq_msg *dreq_msg;
2566         struct ib_mad_send_buf *msg = NULL;
2567         int ret;
2568
2569         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2570         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2571                                    dreq_msg->local_comm_id);
2572         if (!cm_id_priv) {
2573                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2574                                 counter[CM_DREQ_COUNTER]);
2575                 cm_issue_drep(work->port, work->mad_recv_wc);
2576                 pr_debug("%s: no cm_id_priv, local_comm_id %d, remote_comm_id %d\n",
2577                          __func__, be32_to_cpu(dreq_msg->local_comm_id),
2578                          be32_to_cpu(dreq_msg->remote_comm_id));
2579                 return -EINVAL;
2580         }
2581
2582         work->cm_event.private_data = &dreq_msg->private_data;
2583
2584         spin_lock_irq(&cm_id_priv->lock);
2585         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2586                 goto unlock;
2587
2588         switch (cm_id_priv->id.state) {
2589         case IB_CM_REP_SENT:
2590         case IB_CM_DREQ_SENT:
2591                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2592                 break;
2593         case IB_CM_ESTABLISHED:
2594                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2595                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2596                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2597                 break;
2598         case IB_CM_MRA_REP_RCVD:
2599                 break;
2600         case IB_CM_TIMEWAIT:
2601                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2602                                 counter[CM_DREQ_COUNTER]);
2603                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
2604                 if (IS_ERR(msg))
2605                         goto unlock;
2606
2607                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2608                                cm_id_priv->private_data,
2609                                cm_id_priv->private_data_len);
2610                 spin_unlock_irq(&cm_id_priv->lock);
2611
2612                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
2613                     ib_post_send_mad(msg, NULL))
2614                         cm_free_msg(msg);
2615                 goto deref;
2616         case IB_CM_DREQ_RCVD:
2617                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2618                                 counter[CM_DREQ_COUNTER]);
2619                 goto unlock;
2620         default:
2621                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
2622                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
2623                          cm_id_priv->id.state);
2624                 goto unlock;
2625         }
2626         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2627         cm_id_priv->tid = dreq_msg->hdr.tid;
2628         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2629         if (!ret)
2630                 list_add_tail(&work->list, &cm_id_priv->work_list);
2631         spin_unlock_irq(&cm_id_priv->lock);
2632
2633         if (ret)
2634                 cm_process_work(cm_id_priv, work);
2635         else
2636                 cm_deref_id(cm_id_priv);
2637         return 0;
2638
2639 unlock: spin_unlock_irq(&cm_id_priv->lock);
2640 deref:  cm_deref_id(cm_id_priv);
2641         return -EINVAL;
2642 }
2643
2644 static int cm_drep_handler(struct cm_work *work)
2645 {
2646         struct cm_id_private *cm_id_priv;
2647         struct cm_drep_msg *drep_msg;
2648         int ret;
2649
2650         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2651         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2652                                    drep_msg->local_comm_id);
2653         if (!cm_id_priv)
2654                 return -EINVAL;
2655
2656         work->cm_event.private_data = &drep_msg->private_data;
2657
2658         spin_lock_irq(&cm_id_priv->lock);
2659         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2660             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2661                 spin_unlock_irq(&cm_id_priv->lock);
2662                 goto out;
2663         }
2664         cm_enter_timewait(cm_id_priv);
2665
2666         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2667         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2668         if (!ret)
2669                 list_add_tail(&work->list, &cm_id_priv->work_list);
2670         spin_unlock_irq(&cm_id_priv->lock);
2671
2672         if (ret)
2673                 cm_process_work(cm_id_priv, work);
2674         else
2675                 cm_deref_id(cm_id_priv);
2676         return 0;
2677 out:
2678         cm_deref_id(cm_id_priv);
2679         return -EINVAL;
2680 }
2681
2682 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2683                    enum ib_cm_rej_reason reason,
2684                    void *ari,
2685                    u8 ari_length,
2686                    const void *private_data,
2687                    u8 private_data_len)
2688 {
2689         struct cm_id_private *cm_id_priv;
2690         struct ib_mad_send_buf *msg;
2691         unsigned long flags;
2692         int ret;
2693
2694         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2695             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2696                 return -EINVAL;
2697
2698         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2699
2700         spin_lock_irqsave(&cm_id_priv->lock, flags);
2701         switch (cm_id->state) {
2702         case IB_CM_REQ_SENT:
2703         case IB_CM_MRA_REQ_RCVD:
2704         case IB_CM_REQ_RCVD:
2705         case IB_CM_MRA_REQ_SENT:
2706         case IB_CM_REP_RCVD:
2707         case IB_CM_MRA_REP_SENT:
2708                 ret = cm_alloc_msg(cm_id_priv, &msg);
2709                 if (!ret)
2710                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2711                                       cm_id_priv, reason, ari, ari_length,
2712                                       private_data, private_data_len);
2713
2714                 cm_reset_to_idle(cm_id_priv);
2715                 break;
2716         case IB_CM_REP_SENT:
2717         case IB_CM_MRA_REP_RCVD:
2718                 ret = cm_alloc_msg(cm_id_priv, &msg);
2719                 if (!ret)
2720                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2721                                       cm_id_priv, reason, ari, ari_length,
2722                                       private_data, private_data_len);
2723
2724                 cm_enter_timewait(cm_id_priv);
2725                 break;
2726         default:
2727                 pr_debug("%s: local_id %d, cm_id->state: %d\n", __func__,
2728                          be32_to_cpu(cm_id_priv->id.local_id), cm_id->state);
2729                 ret = -EINVAL;
2730                 goto out;
2731         }
2732
2733         if (ret)
2734                 goto out;
2735
2736         ret = ib_post_send_mad(msg, NULL);
2737         if (ret)
2738                 cm_free_msg(msg);
2739
2740 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2741         return ret;
2742 }
2743 EXPORT_SYMBOL(ib_send_cm_rej);
2744
2745 static void cm_format_rej_event(struct cm_work *work)
2746 {
2747         struct cm_rej_msg *rej_msg;
2748         struct ib_cm_rej_event_param *param;
2749
2750         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2751         param = &work->cm_event.param.rej_rcvd;
2752         param->ari = rej_msg->ari;
2753         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2754         param->reason = __be16_to_cpu(rej_msg->reason);
2755         work->cm_event.private_data = &rej_msg->private_data;
2756 }
2757
2758 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2759 {
2760         struct cm_timewait_info *timewait_info;
2761         struct cm_id_private *cm_id_priv;
2762         __be32 remote_id;
2763
2764         remote_id = rej_msg->local_comm_id;
2765
2766         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2767                 spin_lock_irq(&cm.lock);
2768                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2769                                                   remote_id);
2770                 if (!timewait_info) {
2771                         spin_unlock_irq(&cm.lock);
2772                         return NULL;
2773                 }
2774                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2775                                       (timewait_info->work.local_id ^
2776                                        cm.random_id_operand));
2777                 if (cm_id_priv) {
2778                         if (cm_id_priv->id.remote_id == remote_id)
2779                                 atomic_inc(&cm_id_priv->refcount);
2780                         else
2781                                 cm_id_priv = NULL;
2782                 }
2783                 spin_unlock_irq(&cm.lock);
2784         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2785                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2786         else
2787                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2788
2789         return cm_id_priv;
2790 }
2791
2792 static int cm_rej_handler(struct cm_work *work)
2793 {
2794         struct cm_id_private *cm_id_priv;
2795         struct cm_rej_msg *rej_msg;
2796         int ret;
2797
2798         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2799         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2800         if (!cm_id_priv)
2801                 return -EINVAL;
2802
2803         cm_format_rej_event(work);
2804
2805         spin_lock_irq(&cm_id_priv->lock);
2806         switch (cm_id_priv->id.state) {
2807         case IB_CM_REQ_SENT:
2808         case IB_CM_MRA_REQ_RCVD:
2809         case IB_CM_REP_SENT:
2810         case IB_CM_MRA_REP_RCVD:
2811                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2812                 /* fall through */
2813         case IB_CM_REQ_RCVD:
2814         case IB_CM_MRA_REQ_SENT:
2815                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2816                         cm_enter_timewait(cm_id_priv);
2817                 else
2818                         cm_reset_to_idle(cm_id_priv);
2819                 break;
2820         case IB_CM_DREQ_SENT:
2821                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2822                 /* fall through */
2823         case IB_CM_REP_RCVD:
2824         case IB_CM_MRA_REP_SENT:
2825                 cm_enter_timewait(cm_id_priv);
2826                 break;
2827         case IB_CM_ESTABLISHED:
2828                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2829                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2830                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2831                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2832                                               cm_id_priv->msg);
2833                         cm_enter_timewait(cm_id_priv);
2834                         break;
2835                 }
2836                 /* fall through */
2837         default:
2838                 spin_unlock_irq(&cm_id_priv->lock);
2839                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
2840                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
2841                          cm_id_priv->id.state);
2842                 ret = -EINVAL;
2843                 goto out;
2844         }
2845
2846         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2847         if (!ret)
2848                 list_add_tail(&work->list, &cm_id_priv->work_list);
2849         spin_unlock_irq(&cm_id_priv->lock);
2850
2851         if (ret)
2852                 cm_process_work(cm_id_priv, work);
2853         else
2854                 cm_deref_id(cm_id_priv);
2855         return 0;
2856 out:
2857         cm_deref_id(cm_id_priv);
2858         return -EINVAL;
2859 }
2860
2861 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2862                    u8 service_timeout,
2863                    const void *private_data,
2864                    u8 private_data_len)
2865 {
2866         struct cm_id_private *cm_id_priv;
2867         struct ib_mad_send_buf *msg;
2868         enum ib_cm_state cm_state;
2869         enum ib_cm_lap_state lap_state;
2870         enum cm_msg_response msg_response;
2871         void *data;
2872         unsigned long flags;
2873         int ret;
2874
2875         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2876                 return -EINVAL;
2877
2878         data = cm_copy_private_data(private_data, private_data_len);
2879         if (IS_ERR(data))
2880                 return PTR_ERR(data);
2881
2882         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2883
2884         spin_lock_irqsave(&cm_id_priv->lock, flags);
2885         switch(cm_id_priv->id.state) {
2886         case IB_CM_REQ_RCVD:
2887                 cm_state = IB_CM_MRA_REQ_SENT;
2888                 lap_state = cm_id->lap_state;
2889                 msg_response = CM_MSG_RESPONSE_REQ;
2890                 break;
2891         case IB_CM_REP_RCVD:
2892                 cm_state = IB_CM_MRA_REP_SENT;
2893                 lap_state = cm_id->lap_state;
2894                 msg_response = CM_MSG_RESPONSE_REP;
2895                 break;
2896         case IB_CM_ESTABLISHED:
2897                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2898                         cm_state = cm_id->state;
2899                         lap_state = IB_CM_MRA_LAP_SENT;
2900                         msg_response = CM_MSG_RESPONSE_OTHER;
2901                         break;
2902                 }
2903                 /* fall through */
2904         default:
2905                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
2906                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
2907                          cm_id_priv->id.state);
2908                 ret = -EINVAL;
2909                 goto error1;
2910         }
2911
2912         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2913                 ret = cm_alloc_msg(cm_id_priv, &msg);
2914                 if (ret)
2915                         goto error1;
2916
2917                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2918                               msg_response, service_timeout,
2919                               private_data, private_data_len);
2920                 ret = ib_post_send_mad(msg, NULL);
2921                 if (ret)
2922                         goto error2;
2923         }
2924
2925         cm_id->state = cm_state;
2926         cm_id->lap_state = lap_state;
2927         cm_id_priv->service_timeout = service_timeout;
2928         cm_set_private_data(cm_id_priv, data, private_data_len);
2929         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2930         return 0;
2931
2932 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2933         kfree(data);
2934         return ret;
2935
2936 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2937         kfree(data);
2938         cm_free_msg(msg);
2939         return ret;
2940 }
2941 EXPORT_SYMBOL(ib_send_cm_mra);
2942
2943 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2944 {
2945         switch (cm_mra_get_msg_mraed(mra_msg)) {
2946         case CM_MSG_RESPONSE_REQ:
2947                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2948         case CM_MSG_RESPONSE_REP:
2949         case CM_MSG_RESPONSE_OTHER:
2950                 return cm_acquire_id(mra_msg->remote_comm_id,
2951                                      mra_msg->local_comm_id);
2952         default:
2953                 return NULL;
2954         }
2955 }
2956
2957 static int cm_mra_handler(struct cm_work *work)
2958 {
2959         struct cm_id_private *cm_id_priv;
2960         struct cm_mra_msg *mra_msg;
2961         int timeout, ret;
2962
2963         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2964         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2965         if (!cm_id_priv)
2966                 return -EINVAL;
2967
2968         work->cm_event.private_data = &mra_msg->private_data;
2969         work->cm_event.param.mra_rcvd.service_timeout =
2970                                         cm_mra_get_service_timeout(mra_msg);
2971         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2972                   cm_convert_to_ms(cm_id_priv->av.timeout);
2973
2974         spin_lock_irq(&cm_id_priv->lock);
2975         switch (cm_id_priv->id.state) {
2976         case IB_CM_REQ_SENT:
2977                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2978                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2979                                   cm_id_priv->msg, timeout))
2980                         goto out;
2981                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2982                 break;
2983         case IB_CM_REP_SENT:
2984                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2985                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2986                                   cm_id_priv->msg, timeout))
2987                         goto out;
2988                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2989                 break;
2990         case IB_CM_ESTABLISHED:
2991                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2992                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2993                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2994                                   cm_id_priv->msg, timeout)) {
2995                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2996                                 atomic_long_inc(&work->port->
2997                                                 counter_group[CM_RECV_DUPLICATES].
2998                                                 counter[CM_MRA_COUNTER]);
2999                         goto out;
3000                 }
3001                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
3002                 break;
3003         case IB_CM_MRA_REQ_RCVD:
3004         case IB_CM_MRA_REP_RCVD:
3005                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3006                                 counter[CM_MRA_COUNTER]);
3007                 /* fall through */
3008         default:
3009                 pr_debug("%s local_id %d, cm_id_priv->id.state: %d\n",
3010                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
3011                          cm_id_priv->id.state);
3012                 goto out;
3013         }
3014
3015         cm_id_priv->msg->context[1] = (void *) (unsigned long)
3016                                       cm_id_priv->id.state;
3017         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3018         if (!ret)
3019                 list_add_tail(&work->list, &cm_id_priv->work_list);
3020         spin_unlock_irq(&cm_id_priv->lock);
3021
3022         if (ret)
3023                 cm_process_work(cm_id_priv, work);
3024         else
3025                 cm_deref_id(cm_id_priv);
3026         return 0;
3027 out:
3028         spin_unlock_irq(&cm_id_priv->lock);
3029         cm_deref_id(cm_id_priv);
3030         return -EINVAL;
3031 }
3032
3033 static void cm_format_lap(struct cm_lap_msg *lap_msg,
3034                           struct cm_id_private *cm_id_priv,
3035                           struct sa_path_rec *alternate_path,
3036                           const void *private_data,
3037                           u8 private_data_len)
3038 {
3039         bool alt_ext = false;
3040
3041         if (alternate_path->rec_type == SA_PATH_REC_TYPE_OPA)
3042                 alt_ext = opa_is_extended_lid(alternate_path->opa.dlid,
3043                                               alternate_path->opa.slid);
3044         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
3045                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
3046         lap_msg->local_comm_id = cm_id_priv->id.local_id;
3047         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
3048         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
3049         /* todo: need remote CM response timeout */
3050         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
3051         lap_msg->alt_local_lid =
3052                 htons(ntohl(sa_path_get_slid(alternate_path)));
3053         lap_msg->alt_remote_lid =
3054                 htons(ntohl(sa_path_get_dlid(alternate_path)));
3055         lap_msg->alt_local_gid = alternate_path->sgid;
3056         lap_msg->alt_remote_gid = alternate_path->dgid;
3057         if (alt_ext) {
3058                 lap_msg->alt_local_gid.global.interface_id
3059                         = OPA_MAKE_ID(be32_to_cpu(alternate_path->opa.slid));
3060                 lap_msg->alt_remote_gid.global.interface_id
3061                         = OPA_MAKE_ID(be32_to_cpu(alternate_path->opa.dlid));
3062         }
3063         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
3064         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
3065         lap_msg->alt_hop_limit = alternate_path->hop_limit;
3066         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
3067         cm_lap_set_sl(lap_msg, alternate_path->sl);
3068         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
3069         cm_lap_set_local_ack_timeout(lap_msg,
3070                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
3071                                alternate_path->packet_life_time));
3072
3073         if (private_data && private_data_len)
3074                 memcpy(lap_msg->private_data, private_data, private_data_len);
3075 }
3076
3077 int ib_send_cm_lap(struct ib_cm_id *cm_id,
3078                    struct sa_path_rec *alternate_path,
3079                    const void *private_data,
3080                    u8 private_data_len)
3081 {
3082         struct cm_id_private *cm_id_priv;
3083         struct ib_mad_send_buf *msg;
3084         unsigned long flags;
3085         int ret;
3086
3087         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
3088                 return -EINVAL;
3089
3090         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3091         spin_lock_irqsave(&cm_id_priv->lock, flags);
3092         if (cm_id->state != IB_CM_ESTABLISHED ||
3093             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
3094              cm_id->lap_state != IB_CM_LAP_IDLE)) {
3095                 ret = -EINVAL;
3096                 goto out;
3097         }
3098
3099         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av,
3100                                  cm_id_priv);
3101         if (ret)
3102                 goto out;
3103         cm_id_priv->alt_av.timeout =
3104                         cm_ack_timeout(cm_id_priv->target_ack_delay,
3105                                        cm_id_priv->alt_av.timeout - 1);
3106
3107         ret = cm_alloc_msg(cm_id_priv, &msg);
3108         if (ret)
3109                 goto out;
3110
3111         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
3112                       alternate_path, private_data, private_data_len);
3113         msg->timeout_ms = cm_id_priv->timeout_ms;
3114         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
3115
3116         ret = ib_post_send_mad(msg, NULL);
3117         if (ret) {
3118                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3119                 cm_free_msg(msg);
3120                 return ret;
3121         }
3122
3123         cm_id->lap_state = IB_CM_LAP_SENT;
3124         cm_id_priv->msg = msg;
3125
3126 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3127         return ret;
3128 }
3129 EXPORT_SYMBOL(ib_send_cm_lap);
3130
3131 static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg,
3132                                         struct sa_path_rec *path)
3133 {
3134         u32 lid;
3135
3136         if (path->rec_type != SA_PATH_REC_TYPE_OPA) {
3137                 sa_path_set_dlid(path, ntohs(lap_msg->alt_local_lid));
3138                 sa_path_set_slid(path, ntohs(lap_msg->alt_remote_lid));
3139         } else {
3140                 lid = opa_get_lid_from_gid(&lap_msg->alt_local_gid);
3141                 sa_path_set_dlid(path, lid);
3142
3143                 lid = opa_get_lid_from_gid(&lap_msg->alt_remote_gid);
3144                 sa_path_set_slid(path, lid);
3145         }
3146 }
3147
3148 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
3149                                     struct sa_path_rec *path,
3150                                     struct cm_lap_msg *lap_msg)
3151 {
3152         path->dgid = lap_msg->alt_local_gid;
3153         path->sgid = lap_msg->alt_remote_gid;
3154         path->flow_label = cm_lap_get_flow_label(lap_msg);
3155         path->hop_limit = lap_msg->alt_hop_limit;
3156         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
3157         path->reversible = 1;
3158         path->pkey = cm_id_priv->pkey;
3159         path->sl = cm_lap_get_sl(lap_msg);
3160         path->mtu_selector = IB_SA_EQ;
3161         path->mtu = cm_id_priv->path_mtu;
3162         path->rate_selector = IB_SA_EQ;
3163         path->rate = cm_lap_get_packet_rate(lap_msg);
3164         path->packet_life_time_selector = IB_SA_EQ;
3165         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
3166         path->packet_life_time -= (path->packet_life_time > 0);
3167         cm_format_path_lid_from_lap(lap_msg, path);
3168 }
3169
3170 static int cm_lap_handler(struct cm_work *work)
3171 {
3172         struct cm_id_private *cm_id_priv;
3173         struct cm_lap_msg *lap_msg;
3174         struct ib_cm_lap_event_param *param;
3175         struct ib_mad_send_buf *msg = NULL;
3176         int ret;
3177
3178         /* Currently Alternate path messages are not supported for
3179          * RoCE link layer.
3180          */
3181         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3182                                work->port->port_num))
3183                 return -EINVAL;
3184
3185         /* todo: verify LAP request and send reject APR if invalid. */
3186         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
3187         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
3188                                    lap_msg->local_comm_id);
3189         if (!cm_id_priv)
3190                 return -EINVAL;
3191
3192         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3193                                       work->mad_recv_wc->recv_buf.grh,
3194                                       &cm_id_priv->av);
3195         if (ret)
3196                 goto deref;
3197
3198         param = &work->cm_event.param.lap_rcvd;
3199         memset(&work->path[0], 0, sizeof(work->path[1]));
3200         cm_path_set_rec_type(work->port->cm_dev->ib_device,
3201                              work->port->port_num,
3202                              &work->path[0],
3203                              &lap_msg->alt_local_gid);
3204         param->alternate_path = &work->path[0];
3205         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
3206         work->cm_event.private_data = &lap_msg->private_data;
3207
3208         spin_lock_irq(&cm_id_priv->lock);
3209         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
3210                 goto unlock;
3211
3212         switch (cm_id_priv->id.lap_state) {
3213         case IB_CM_LAP_UNINIT:
3214         case IB_CM_LAP_IDLE:
3215                 break;
3216         case IB_CM_MRA_LAP_SENT:
3217                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3218                                 counter[CM_LAP_COUNTER]);
3219                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
3220                 if (IS_ERR(msg))
3221                         goto unlock;
3222
3223                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3224                               CM_MSG_RESPONSE_OTHER,
3225                               cm_id_priv->service_timeout,
3226                               cm_id_priv->private_data,
3227                               cm_id_priv->private_data_len);
3228                 spin_unlock_irq(&cm_id_priv->lock);
3229
3230                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
3231                     ib_post_send_mad(msg, NULL))
3232                         cm_free_msg(msg);
3233                 goto deref;
3234         case IB_CM_LAP_RCVD:
3235                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3236                                 counter[CM_LAP_COUNTER]);
3237                 goto unlock;
3238         default:
3239                 goto unlock;
3240         }
3241
3242         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
3243         cm_id_priv->tid = lap_msg->hdr.tid;
3244         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av,
3245                            cm_id_priv);
3246         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3247         if (!ret)
3248                 list_add_tail(&work->list, &cm_id_priv->work_list);
3249         spin_unlock_irq(&cm_id_priv->lock);
3250
3251         if (ret)
3252                 cm_process_work(cm_id_priv, work);
3253         else
3254                 cm_deref_id(cm_id_priv);
3255         return 0;
3256
3257 unlock: spin_unlock_irq(&cm_id_priv->lock);
3258 deref:  cm_deref_id(cm_id_priv);
3259         return -EINVAL;
3260 }
3261
3262 static void cm_format_apr(struct cm_apr_msg *apr_msg,
3263                           struct cm_id_private *cm_id_priv,
3264                           enum ib_cm_apr_status status,
3265                           void *info,
3266                           u8 info_length,
3267                           const void *private_data,
3268                           u8 private_data_len)
3269 {
3270         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
3271         apr_msg->local_comm_id = cm_id_priv->id.local_id;
3272         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
3273         apr_msg->ap_status = (u8) status;
3274
3275         if (info && info_length) {
3276                 apr_msg->info_length = info_length;
3277                 memcpy(apr_msg->info, info, info_length);
3278         }
3279
3280         if (private_data && private_data_len)
3281                 memcpy(apr_msg->private_data, private_data, private_data_len);
3282 }
3283
3284 int ib_send_cm_apr(struct ib_cm_id *cm_id,
3285                    enum ib_cm_apr_status status,
3286                    void *info,
3287                    u8 info_length,
3288                    const void *private_data,
3289                    u8 private_data_len)
3290 {
3291         struct cm_id_private *cm_id_priv;
3292         struct ib_mad_send_buf *msg;
3293         unsigned long flags;
3294         int ret;
3295
3296         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
3297             (info && info_length > IB_CM_APR_INFO_LENGTH))
3298                 return -EINVAL;
3299
3300         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3301         spin_lock_irqsave(&cm_id_priv->lock, flags);
3302         if (cm_id->state != IB_CM_ESTABLISHED ||
3303             (cm_id->lap_state != IB_CM_LAP_RCVD &&
3304              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
3305                 ret = -EINVAL;
3306                 goto out;
3307         }
3308
3309         ret = cm_alloc_msg(cm_id_priv, &msg);
3310         if (ret)
3311                 goto out;
3312
3313         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
3314                       info, info_length, private_data, private_data_len);
3315         ret = ib_post_send_mad(msg, NULL);
3316         if (ret) {
3317                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3318                 cm_free_msg(msg);
3319                 return ret;
3320         }
3321
3322         cm_id->lap_state = IB_CM_LAP_IDLE;
3323 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3324         return ret;
3325 }
3326 EXPORT_SYMBOL(ib_send_cm_apr);
3327
3328 static int cm_apr_handler(struct cm_work *work)
3329 {
3330         struct cm_id_private *cm_id_priv;
3331         struct cm_apr_msg *apr_msg;
3332         int ret;
3333
3334         /* Currently Alternate path messages are not supported for
3335          * RoCE link layer.
3336          */
3337         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3338                                work->port->port_num))
3339                 return -EINVAL;
3340
3341         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3342         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
3343                                    apr_msg->local_comm_id);
3344         if (!cm_id_priv)
3345                 return -EINVAL; /* Unmatched reply. */
3346
3347         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
3348         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
3349         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
3350         work->cm_event.private_data = &apr_msg->private_data;
3351
3352         spin_lock_irq(&cm_id_priv->lock);
3353         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3354             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3355              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3356                 spin_unlock_irq(&cm_id_priv->lock);
3357                 goto out;
3358         }
3359         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3360         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3361         cm_id_priv->msg = NULL;
3362
3363         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3364         if (!ret)
3365                 list_add_tail(&work->list, &cm_id_priv->work_list);
3366         spin_unlock_irq(&cm_id_priv->lock);
3367
3368         if (ret)
3369                 cm_process_work(cm_id_priv, work);
3370         else
3371                 cm_deref_id(cm_id_priv);
3372         return 0;
3373 out:
3374         cm_deref_id(cm_id_priv);
3375         return -EINVAL;
3376 }
3377
3378 static int cm_timewait_handler(struct cm_work *work)
3379 {
3380         struct cm_timewait_info *timewait_info;
3381         struct cm_id_private *cm_id_priv;
3382         int ret;
3383
3384         timewait_info = (struct cm_timewait_info *)work;
3385         spin_lock_irq(&cm.lock);
3386         list_del(&timewait_info->list);
3387         spin_unlock_irq(&cm.lock);
3388
3389         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3390                                    timewait_info->work.remote_id);
3391         if (!cm_id_priv)
3392                 return -EINVAL;
3393
3394         spin_lock_irq(&cm_id_priv->lock);
3395         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3396             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3397                 spin_unlock_irq(&cm_id_priv->lock);
3398                 goto out;
3399         }
3400         cm_id_priv->id.state = IB_CM_IDLE;
3401         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3402         if (!ret)
3403                 list_add_tail(&work->list, &cm_id_priv->work_list);
3404         spin_unlock_irq(&cm_id_priv->lock);
3405
3406         if (ret)
3407                 cm_process_work(cm_id_priv, work);
3408         else
3409                 cm_deref_id(cm_id_priv);
3410         return 0;
3411 out:
3412         cm_deref_id(cm_id_priv);
3413         return -EINVAL;
3414 }
3415
3416 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3417                                struct cm_id_private *cm_id_priv,
3418                                struct ib_cm_sidr_req_param *param)
3419 {
3420         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3421                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
3422         sidr_req_msg->request_id = cm_id_priv->id.local_id;
3423         sidr_req_msg->pkey = param->path->pkey;
3424         sidr_req_msg->service_id = param->service_id;
3425
3426         if (param->private_data && param->private_data_len)
3427                 memcpy(sidr_req_msg->private_data, param->private_data,
3428                        param->private_data_len);
3429 }
3430
3431 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3432                         struct ib_cm_sidr_req_param *param)
3433 {
3434         struct cm_id_private *cm_id_priv;
3435         struct ib_mad_send_buf *msg;
3436         unsigned long flags;
3437         int ret;
3438
3439         if (!param->path || (param->private_data &&
3440              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3441                 return -EINVAL;
3442
3443         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3444         ret = cm_init_av_by_path(param->path, &cm_id_priv->av, cm_id_priv);
3445         if (ret)
3446                 goto out;
3447
3448         cm_id->service_id = param->service_id;
3449         cm_id->service_mask = ~cpu_to_be64(0);
3450         cm_id_priv->timeout_ms = param->timeout_ms;
3451         cm_id_priv->max_cm_retries = param->max_cm_retries;
3452         ret = cm_alloc_msg(cm_id_priv, &msg);
3453         if (ret)
3454                 goto out;
3455
3456         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3457                            param);
3458         msg->timeout_ms = cm_id_priv->timeout_ms;
3459         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3460
3461         spin_lock_irqsave(&cm_id_priv->lock, flags);
3462         if (cm_id->state == IB_CM_IDLE)
3463                 ret = ib_post_send_mad(msg, NULL);
3464         else
3465                 ret = -EINVAL;
3466
3467         if (ret) {
3468                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3469                 cm_free_msg(msg);
3470                 goto out;
3471         }
3472         cm_id->state = IB_CM_SIDR_REQ_SENT;
3473         cm_id_priv->msg = msg;
3474         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3475 out:
3476         return ret;
3477 }
3478 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3479
3480 static void cm_format_sidr_req_event(struct cm_work *work,
3481                                      struct ib_cm_id *listen_id)
3482 {
3483         struct cm_sidr_req_msg *sidr_req_msg;
3484         struct ib_cm_sidr_req_event_param *param;
3485
3486         sidr_req_msg = (struct cm_sidr_req_msg *)
3487                                 work->mad_recv_wc->recv_buf.mad;
3488         param = &work->cm_event.param.sidr_req_rcvd;
3489         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
3490         param->listen_id = listen_id;
3491         param->service_id = sidr_req_msg->service_id;
3492         param->bth_pkey = cm_get_bth_pkey(work);
3493         param->port = work->port->port_num;
3494         work->cm_event.private_data = &sidr_req_msg->private_data;
3495 }
3496
3497 static int cm_sidr_req_handler(struct cm_work *work)
3498 {
3499         struct ib_cm_id *cm_id;
3500         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3501         struct cm_sidr_req_msg *sidr_req_msg;
3502         struct ib_wc *wc;
3503         int ret;
3504
3505         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3506         if (IS_ERR(cm_id))
3507                 return PTR_ERR(cm_id);
3508         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3509
3510         /* Record SGID/SLID and request ID for lookup. */
3511         sidr_req_msg = (struct cm_sidr_req_msg *)
3512                                 work->mad_recv_wc->recv_buf.mad;
3513         wc = work->mad_recv_wc->wc;
3514         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3515         cm_id_priv->av.dgid.global.interface_id = 0;
3516         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3517                                       work->mad_recv_wc->recv_buf.grh,
3518                                       &cm_id_priv->av);
3519         if (ret)
3520                 goto out;
3521
3522         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3523         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3524         atomic_inc(&cm_id_priv->work_count);
3525
3526         spin_lock_irq(&cm.lock);
3527         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3528         if (cur_cm_id_priv) {
3529                 spin_unlock_irq(&cm.lock);
3530                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3531                                 counter[CM_SIDR_REQ_COUNTER]);
3532                 goto out; /* Duplicate message. */
3533         }
3534         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3535         cur_cm_id_priv = cm_find_listen(cm_id->device,
3536                                         sidr_req_msg->service_id);
3537         if (!cur_cm_id_priv) {
3538                 spin_unlock_irq(&cm.lock);
3539                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3540                 goto out; /* No match. */
3541         }
3542         atomic_inc(&cur_cm_id_priv->refcount);
3543         atomic_inc(&cm_id_priv->refcount);
3544         spin_unlock_irq(&cm.lock);
3545
3546         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3547         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3548         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3549         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3550
3551         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3552         cm_process_work(cm_id_priv, work);
3553         cm_deref_id(cur_cm_id_priv);
3554         return 0;
3555 out:
3556         ib_destroy_cm_id(&cm_id_priv->id);
3557         return -EINVAL;
3558 }
3559
3560 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3561                                struct cm_id_private *cm_id_priv,
3562                                struct ib_cm_sidr_rep_param *param)
3563 {
3564         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3565                           cm_id_priv->tid);
3566         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3567         sidr_rep_msg->status = param->status;
3568         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3569         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3570         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3571
3572         if (param->info && param->info_length)
3573                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3574
3575         if (param->private_data && param->private_data_len)
3576                 memcpy(sidr_rep_msg->private_data, param->private_data,
3577                        param->private_data_len);
3578 }
3579
3580 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3581                         struct ib_cm_sidr_rep_param *param)
3582 {
3583         struct cm_id_private *cm_id_priv;
3584         struct ib_mad_send_buf *msg;
3585         unsigned long flags;
3586         int ret;
3587
3588         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3589             (param->private_data &&
3590              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3591                 return -EINVAL;
3592
3593         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3594         spin_lock_irqsave(&cm_id_priv->lock, flags);
3595         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3596                 ret = -EINVAL;
3597                 goto error;
3598         }
3599
3600         ret = cm_alloc_msg(cm_id_priv, &msg);
3601         if (ret)
3602                 goto error;
3603
3604         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3605                            param);
3606         ret = ib_post_send_mad(msg, NULL);
3607         if (ret) {
3608                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3609                 cm_free_msg(msg);
3610                 return ret;
3611         }
3612         cm_id->state = IB_CM_IDLE;
3613         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3614
3615         spin_lock_irqsave(&cm.lock, flags);
3616         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3617                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3618                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3619         }
3620         spin_unlock_irqrestore(&cm.lock, flags);
3621         return 0;
3622
3623 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3624         return ret;
3625 }
3626 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3627
3628 static void cm_format_sidr_rep_event(struct cm_work *work)
3629 {
3630         struct cm_sidr_rep_msg *sidr_rep_msg;
3631         struct ib_cm_sidr_rep_event_param *param;
3632
3633         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3634                                 work->mad_recv_wc->recv_buf.mad;
3635         param = &work->cm_event.param.sidr_rep_rcvd;
3636         param->status = sidr_rep_msg->status;
3637         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3638         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3639         param->info = &sidr_rep_msg->info;
3640         param->info_len = sidr_rep_msg->info_length;
3641         work->cm_event.private_data = &sidr_rep_msg->private_data;
3642 }
3643
3644 static int cm_sidr_rep_handler(struct cm_work *work)
3645 {
3646         struct cm_sidr_rep_msg *sidr_rep_msg;
3647         struct cm_id_private *cm_id_priv;
3648
3649         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3650                                 work->mad_recv_wc->recv_buf.mad;
3651         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3652         if (!cm_id_priv)
3653                 return -EINVAL; /* Unmatched reply. */
3654
3655         spin_lock_irq(&cm_id_priv->lock);
3656         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3657                 spin_unlock_irq(&cm_id_priv->lock);
3658                 goto out;
3659         }
3660         cm_id_priv->id.state = IB_CM_IDLE;
3661         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3662         spin_unlock_irq(&cm_id_priv->lock);
3663
3664         cm_format_sidr_rep_event(work);
3665         cm_process_work(cm_id_priv, work);
3666         return 0;
3667 out:
3668         cm_deref_id(cm_id_priv);
3669         return -EINVAL;
3670 }
3671
3672 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3673                                   enum ib_wc_status wc_status)
3674 {
3675         struct cm_id_private *cm_id_priv;
3676         struct ib_cm_event cm_event;
3677         enum ib_cm_state state;
3678         int ret;
3679
3680         memset(&cm_event, 0, sizeof cm_event);
3681         cm_id_priv = msg->context[0];
3682
3683         /* Discard old sends or ones without a response. */
3684         spin_lock_irq(&cm_id_priv->lock);
3685         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3686         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3687                 goto discard;
3688
3689         pr_debug_ratelimited("CM: failed sending MAD in state %d. (%s)\n",
3690                              state, ib_wc_status_msg(wc_status));
3691         switch (state) {
3692         case IB_CM_REQ_SENT:
3693         case IB_CM_MRA_REQ_RCVD:
3694                 cm_reset_to_idle(cm_id_priv);
3695                 cm_event.event = IB_CM_REQ_ERROR;
3696                 break;
3697         case IB_CM_REP_SENT:
3698         case IB_CM_MRA_REP_RCVD:
3699                 cm_reset_to_idle(cm_id_priv);
3700                 cm_event.event = IB_CM_REP_ERROR;
3701                 break;
3702         case IB_CM_DREQ_SENT:
3703                 cm_enter_timewait(cm_id_priv);
3704                 cm_event.event = IB_CM_DREQ_ERROR;
3705                 break;
3706         case IB_CM_SIDR_REQ_SENT:
3707                 cm_id_priv->id.state = IB_CM_IDLE;
3708                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3709                 break;
3710         default:
3711                 goto discard;
3712         }
3713         spin_unlock_irq(&cm_id_priv->lock);
3714         cm_event.param.send_status = wc_status;
3715
3716         /* No other events can occur on the cm_id at this point. */
3717         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3718         cm_free_msg(msg);
3719         if (ret)
3720                 ib_destroy_cm_id(&cm_id_priv->id);
3721         return;
3722 discard:
3723         spin_unlock_irq(&cm_id_priv->lock);
3724         cm_free_msg(msg);
3725 }
3726
3727 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3728                             struct ib_mad_send_wc *mad_send_wc)
3729 {
3730         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3731         struct cm_port *port;
3732         u16 attr_index;
3733
3734         port = mad_agent->context;
3735         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3736                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3737
3738         /*
3739          * If the send was in response to a received message (context[0] is not
3740          * set to a cm_id), and is not a REJ, then it is a send that was
3741          * manually retried.
3742          */
3743         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3744                 msg->retries = 1;
3745
3746         atomic_long_add(1 + msg->retries,
3747                         &port->counter_group[CM_XMIT].counter[attr_index]);
3748         if (msg->retries)
3749                 atomic_long_add(msg->retries,
3750                                 &port->counter_group[CM_XMIT_RETRIES].
3751                                 counter[attr_index]);
3752
3753         switch (mad_send_wc->status) {
3754         case IB_WC_SUCCESS:
3755         case IB_WC_WR_FLUSH_ERR:
3756                 cm_free_msg(msg);
3757                 break;
3758         default:
3759                 if (msg->context[0] && msg->context[1])
3760                         cm_process_send_error(msg, mad_send_wc->status);
3761                 else
3762                         cm_free_msg(msg);
3763                 break;
3764         }
3765 }
3766
3767 static void cm_work_handler(struct work_struct *_work)
3768 {
3769         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3770         int ret;
3771
3772         switch (work->cm_event.event) {
3773         case IB_CM_REQ_RECEIVED:
3774                 ret = cm_req_handler(work);
3775                 break;
3776         case IB_CM_MRA_RECEIVED:
3777                 ret = cm_mra_handler(work);
3778                 break;
3779         case IB_CM_REJ_RECEIVED:
3780                 ret = cm_rej_handler(work);
3781                 break;
3782         case IB_CM_REP_RECEIVED:
3783                 ret = cm_rep_handler(work);
3784                 break;
3785         case IB_CM_RTU_RECEIVED:
3786                 ret = cm_rtu_handler(work);
3787                 break;
3788         case IB_CM_USER_ESTABLISHED:
3789                 ret = cm_establish_handler(work);
3790                 break;
3791         case IB_CM_DREQ_RECEIVED:
3792                 ret = cm_dreq_handler(work);
3793                 break;
3794         case IB_CM_DREP_RECEIVED:
3795                 ret = cm_drep_handler(work);
3796                 break;
3797         case IB_CM_SIDR_REQ_RECEIVED:
3798                 ret = cm_sidr_req_handler(work);
3799                 break;
3800         case IB_CM_SIDR_REP_RECEIVED:
3801                 ret = cm_sidr_rep_handler(work);
3802                 break;
3803         case IB_CM_LAP_RECEIVED:
3804                 ret = cm_lap_handler(work);
3805                 break;
3806         case IB_CM_APR_RECEIVED:
3807                 ret = cm_apr_handler(work);
3808                 break;
3809         case IB_CM_TIMEWAIT_EXIT:
3810                 ret = cm_timewait_handler(work);
3811                 break;
3812         default:
3813                 pr_debug("cm_event.event: 0x%x\n", work->cm_event.event);
3814                 ret = -EINVAL;
3815                 break;
3816         }
3817         if (ret)
3818                 cm_free_work(work);
3819 }
3820
3821 static int cm_establish(struct ib_cm_id *cm_id)
3822 {
3823         struct cm_id_private *cm_id_priv;
3824         struct cm_work *work;
3825         unsigned long flags;
3826         int ret = 0;
3827         struct cm_device *cm_dev;
3828
3829         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3830         if (!cm_dev)
3831                 return -ENODEV;
3832
3833         work = kmalloc(sizeof *work, GFP_ATOMIC);
3834         if (!work)
3835                 return -ENOMEM;
3836
3837         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3838         spin_lock_irqsave(&cm_id_priv->lock, flags);
3839         switch (cm_id->state)
3840         {
3841         case IB_CM_REP_SENT:
3842         case IB_CM_MRA_REP_RCVD:
3843                 cm_id->state = IB_CM_ESTABLISHED;
3844                 break;
3845         case IB_CM_ESTABLISHED:
3846                 ret = -EISCONN;
3847                 break;
3848         default:
3849                 pr_debug("%s: local_id %d, cm_id->state: %d\n", __func__,
3850                          be32_to_cpu(cm_id->local_id), cm_id->state);
3851                 ret = -EINVAL;
3852                 break;
3853         }
3854         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3855
3856         if (ret) {
3857                 kfree(work);
3858                 goto out;
3859         }
3860
3861         /*
3862          * The CM worker thread may try to destroy the cm_id before it
3863          * can execute this work item.  To prevent potential deadlock,
3864          * we need to find the cm_id once we're in the context of the
3865          * worker thread, rather than holding a reference on it.
3866          */
3867         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3868         work->local_id = cm_id->local_id;
3869         work->remote_id = cm_id->remote_id;
3870         work->mad_recv_wc = NULL;
3871         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3872
3873         /* Check if the device started its remove_one */
3874         spin_lock_irqsave(&cm.lock, flags);
3875         if (!cm_dev->going_down) {
3876                 queue_delayed_work(cm.wq, &work->work, 0);
3877         } else {
3878                 kfree(work);
3879                 ret = -ENODEV;
3880         }
3881         spin_unlock_irqrestore(&cm.lock, flags);
3882
3883 out:
3884         return ret;
3885 }
3886
3887 static int cm_migrate(struct ib_cm_id *cm_id)
3888 {
3889         struct cm_id_private *cm_id_priv;
3890         struct cm_av tmp_av;
3891         unsigned long flags;
3892         int tmp_send_port_not_ready;
3893         int ret = 0;
3894
3895         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3896         spin_lock_irqsave(&cm_id_priv->lock, flags);
3897         if (cm_id->state == IB_CM_ESTABLISHED &&
3898             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3899              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3900                 cm_id->lap_state = IB_CM_LAP_IDLE;
3901                 /* Swap address vector */
3902                 tmp_av = cm_id_priv->av;
3903                 cm_id_priv->av = cm_id_priv->alt_av;
3904                 cm_id_priv->alt_av = tmp_av;
3905                 /* Swap port send ready state */
3906                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3907                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3908                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3909         } else
3910                 ret = -EINVAL;
3911         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3912
3913         return ret;
3914 }
3915
3916 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3917 {
3918         int ret;
3919
3920         switch (event) {
3921         case IB_EVENT_COMM_EST:
3922                 ret = cm_establish(cm_id);
3923                 break;
3924         case IB_EVENT_PATH_MIG:
3925                 ret = cm_migrate(cm_id);
3926                 break;
3927         default:
3928                 ret = -EINVAL;
3929         }
3930         return ret;
3931 }
3932 EXPORT_SYMBOL(ib_cm_notify);
3933
3934 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3935                             struct ib_mad_send_buf *send_buf,
3936                             struct ib_mad_recv_wc *mad_recv_wc)
3937 {
3938         struct cm_port *port = mad_agent->context;
3939         struct cm_work *work;
3940         enum ib_cm_event_type event;
3941         bool alt_path = false;
3942         u16 attr_id;
3943         int paths = 0;
3944         int going_down = 0;
3945
3946         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3947         case CM_REQ_ATTR_ID:
3948                 alt_path = cm_req_has_alt_path((struct cm_req_msg *)
3949                                                 mad_recv_wc->recv_buf.mad);
3950                 paths = 1 + (alt_path != 0);
3951                 event = IB_CM_REQ_RECEIVED;
3952                 break;
3953         case CM_MRA_ATTR_ID:
3954                 event = IB_CM_MRA_RECEIVED;
3955                 break;
3956         case CM_REJ_ATTR_ID:
3957                 event = IB_CM_REJ_RECEIVED;
3958                 break;
3959         case CM_REP_ATTR_ID:
3960                 event = IB_CM_REP_RECEIVED;
3961                 break;
3962         case CM_RTU_ATTR_ID:
3963                 event = IB_CM_RTU_RECEIVED;
3964                 break;
3965         case CM_DREQ_ATTR_ID:
3966                 event = IB_CM_DREQ_RECEIVED;
3967                 break;
3968         case CM_DREP_ATTR_ID:
3969                 event = IB_CM_DREP_RECEIVED;
3970                 break;
3971         case CM_SIDR_REQ_ATTR_ID:
3972                 event = IB_CM_SIDR_REQ_RECEIVED;
3973                 break;
3974         case CM_SIDR_REP_ATTR_ID:
3975                 event = IB_CM_SIDR_REP_RECEIVED;
3976                 break;
3977         case CM_LAP_ATTR_ID:
3978                 paths = 1;
3979                 event = IB_CM_LAP_RECEIVED;
3980                 break;
3981         case CM_APR_ATTR_ID:
3982                 event = IB_CM_APR_RECEIVED;
3983                 break;
3984         default:
3985                 ib_free_recv_mad(mad_recv_wc);
3986                 return;
3987         }
3988
3989         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3990         atomic_long_inc(&port->counter_group[CM_RECV].
3991                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3992
3993         work = kmalloc(sizeof(*work) + sizeof(struct sa_path_rec) * paths,
3994                        GFP_KERNEL);
3995         if (!work) {
3996                 ib_free_recv_mad(mad_recv_wc);
3997                 return;
3998         }
3999
4000         INIT_DELAYED_WORK(&work->work, cm_work_handler);
4001         work->cm_event.event = event;
4002         work->mad_recv_wc = mad_recv_wc;
4003         work->port = port;
4004
4005         /* Check if the device started its remove_one */
4006         spin_lock_irq(&cm.lock);
4007         if (!port->cm_dev->going_down)
4008                 queue_delayed_work(cm.wq, &work->work, 0);
4009         else
4010                 going_down = 1;
4011         spin_unlock_irq(&cm.lock);
4012
4013         if (going_down) {
4014                 kfree(work);
4015                 ib_free_recv_mad(mad_recv_wc);
4016         }
4017 }
4018
4019 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
4020                                 struct ib_qp_attr *qp_attr,
4021                                 int *qp_attr_mask)
4022 {
4023         unsigned long flags;
4024         int ret;
4025
4026         spin_lock_irqsave(&cm_id_priv->lock, flags);
4027         switch (cm_id_priv->id.state) {
4028         case IB_CM_REQ_SENT:
4029         case IB_CM_MRA_REQ_RCVD:
4030         case IB_CM_REQ_RCVD:
4031         case IB_CM_MRA_REQ_SENT:
4032         case IB_CM_REP_RCVD:
4033         case IB_CM_MRA_REP_SENT:
4034         case IB_CM_REP_SENT:
4035         case IB_CM_MRA_REP_RCVD:
4036         case IB_CM_ESTABLISHED:
4037                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
4038                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
4039                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
4040                 if (cm_id_priv->responder_resources)
4041                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
4042                                                     IB_ACCESS_REMOTE_ATOMIC;
4043                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
4044                 qp_attr->port_num = cm_id_priv->av.port->port_num;
4045                 ret = 0;
4046                 break;
4047         default:
4048                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
4049                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
4050                          cm_id_priv->id.state);
4051                 ret = -EINVAL;
4052                 break;
4053         }
4054         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4055         return ret;
4056 }
4057
4058 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
4059                                struct ib_qp_attr *qp_attr,
4060                                int *qp_attr_mask)
4061 {
4062         unsigned long flags;
4063         int ret;
4064
4065         spin_lock_irqsave(&cm_id_priv->lock, flags);
4066         switch (cm_id_priv->id.state) {
4067         case IB_CM_REQ_RCVD:
4068         case IB_CM_MRA_REQ_SENT:
4069         case IB_CM_REP_RCVD:
4070         case IB_CM_MRA_REP_SENT:
4071         case IB_CM_REP_SENT:
4072         case IB_CM_MRA_REP_RCVD:
4073         case IB_CM_ESTABLISHED:
4074                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
4075                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
4076                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
4077                 qp_attr->path_mtu = cm_id_priv->path_mtu;
4078                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
4079                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
4080                 if (cm_id_priv->qp_type == IB_QPT_RC ||
4081                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
4082                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
4083                                          IB_QP_MIN_RNR_TIMER;
4084                         qp_attr->max_dest_rd_atomic =
4085                                         cm_id_priv->responder_resources;
4086                         qp_attr->min_rnr_timer = 0;
4087                 }
4088                 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4089                         *qp_attr_mask |= IB_QP_ALT_PATH;
4090                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4091                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4092                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4093                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4094                 }
4095                 ret = 0;
4096                 break;
4097         default:
4098                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
4099                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
4100                          cm_id_priv->id.state);
4101                 ret = -EINVAL;
4102                 break;
4103         }
4104         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4105         return ret;
4106 }
4107
4108 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
4109                                struct ib_qp_attr *qp_attr,
4110                                int *qp_attr_mask)
4111 {
4112         unsigned long flags;
4113         int ret;
4114
4115         spin_lock_irqsave(&cm_id_priv->lock, flags);
4116         switch (cm_id_priv->id.state) {
4117         /* Allow transition to RTS before sending REP */
4118         case IB_CM_REQ_RCVD:
4119         case IB_CM_MRA_REQ_SENT:
4120
4121         case IB_CM_REP_RCVD:
4122         case IB_CM_MRA_REP_SENT:
4123         case IB_CM_REP_SENT:
4124         case IB_CM_MRA_REP_RCVD:
4125         case IB_CM_ESTABLISHED:
4126                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
4127                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
4128                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
4129                         switch (cm_id_priv->qp_type) {
4130                         case IB_QPT_RC:
4131                         case IB_QPT_XRC_INI:
4132                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4133                                                  IB_QP_MAX_QP_RD_ATOMIC;
4134                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
4135                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
4136                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
4137                                 /* fall through */
4138                         case IB_QPT_XRC_TGT:
4139                                 *qp_attr_mask |= IB_QP_TIMEOUT;
4140                                 qp_attr->timeout = cm_id_priv->av.timeout;
4141                                 break;
4142                         default:
4143                                 break;
4144                         }
4145                         if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4146                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
4147                                 qp_attr->path_mig_state = IB_MIG_REARM;
4148                         }
4149                 } else {
4150                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
4151                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4152                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4153                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4154                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4155                         qp_attr->path_mig_state = IB_MIG_REARM;
4156                 }
4157                 ret = 0;
4158                 break;
4159         default:
4160                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
4161                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
4162                          cm_id_priv->id.state);
4163                 ret = -EINVAL;
4164                 break;
4165         }
4166         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4167         return ret;
4168 }
4169
4170 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
4171                        struct ib_qp_attr *qp_attr,
4172                        int *qp_attr_mask)
4173 {
4174         struct cm_id_private *cm_id_priv;
4175         int ret;
4176
4177         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
4178         switch (qp_attr->qp_state) {
4179         case IB_QPS_INIT:
4180                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
4181                 break;
4182         case IB_QPS_RTR:
4183                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
4184                 break;
4185         case IB_QPS_RTS:
4186                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
4187                 break;
4188         default:
4189                 ret = -EINVAL;
4190                 break;
4191         }
4192         return ret;
4193 }
4194 EXPORT_SYMBOL(ib_cm_init_qp_attr);
4195
4196 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
4197                                char *buf)
4198 {
4199         struct cm_counter_group *group;
4200         struct cm_counter_attribute *cm_attr;
4201
4202         group = container_of(obj, struct cm_counter_group, obj);
4203         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
4204
4205         return sprintf(buf, "%ld\n",
4206                        atomic_long_read(&group->counter[cm_attr->index]));
4207 }
4208
4209 static const struct sysfs_ops cm_counter_ops = {
4210         .show = cm_show_counter
4211 };
4212
4213 static struct kobj_type cm_counter_obj_type = {
4214         .sysfs_ops = &cm_counter_ops,
4215         .default_attrs = cm_counter_default_attrs
4216 };
4217
4218 static void cm_release_port_obj(struct kobject *obj)
4219 {
4220         struct cm_port *cm_port;
4221
4222         cm_port = container_of(obj, struct cm_port, port_obj);
4223         kfree(cm_port);
4224 }
4225
4226 static struct kobj_type cm_port_obj_type = {
4227         .release = cm_release_port_obj
4228 };
4229
4230 static char *cm_devnode(struct device *dev, umode_t *mode)
4231 {
4232         if (mode)
4233                 *mode = 0666;
4234         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
4235 }
4236
4237 struct class cm_class = {
4238         .owner   = THIS_MODULE,
4239         .name    = "infiniband_cm",
4240         .devnode = cm_devnode,
4241 };
4242 EXPORT_SYMBOL(cm_class);
4243
4244 static int cm_create_port_fs(struct cm_port *port)
4245 {
4246         int i, ret;
4247
4248         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
4249                                    &port->cm_dev->device->kobj,
4250                                    "%d", port->port_num);
4251         if (ret) {
4252                 kfree(port);
4253                 return ret;
4254         }
4255
4256         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
4257                 ret = kobject_init_and_add(&port->counter_group[i].obj,
4258                                            &cm_counter_obj_type,
4259                                            &port->port_obj,
4260                                            "%s", counter_group_names[i]);
4261                 if (ret)
4262                         goto error;
4263         }
4264
4265         return 0;
4266
4267 error:
4268         while (i--)
4269                 kobject_put(&port->counter_group[i].obj);
4270         kobject_put(&port->port_obj);
4271         return ret;
4272
4273 }
4274
4275 static void cm_remove_port_fs(struct cm_port *port)
4276 {
4277         int i;
4278
4279         for (i = 0; i < CM_COUNTER_GROUPS; i++)
4280                 kobject_put(&port->counter_group[i].obj);
4281
4282         kobject_put(&port->port_obj);
4283 }
4284
4285 static void cm_add_one(struct ib_device *ib_device)
4286 {
4287         struct cm_device *cm_dev;
4288         struct cm_port *port;
4289         struct ib_mad_reg_req reg_req = {
4290                 .mgmt_class = IB_MGMT_CLASS_CM,
4291                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4292         };
4293         struct ib_port_modify port_modify = {
4294                 .set_port_cap_mask = IB_PORT_CM_SUP
4295         };
4296         unsigned long flags;
4297         int ret;
4298         int count = 0;
4299         u8 i;
4300
4301         cm_dev = kzalloc(struct_size(cm_dev, port, ib_device->phys_port_cnt),
4302                          GFP_KERNEL);
4303         if (!cm_dev)
4304                 return;
4305
4306         cm_dev->ib_device = ib_device;
4307         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4308         cm_dev->going_down = 0;
4309         cm_dev->device = device_create(&cm_class, &ib_device->dev,
4310                                        MKDEV(0, 0), NULL,
4311                                        "%s", ib_device->name);
4312         if (IS_ERR(cm_dev->device)) {
4313                 kfree(cm_dev);
4314                 return;
4315         }
4316
4317         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4318         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4319                 if (!rdma_cap_ib_cm(ib_device, i))
4320                         continue;
4321
4322                 port = kzalloc(sizeof *port, GFP_KERNEL);
4323                 if (!port)
4324                         goto error1;
4325
4326                 cm_dev->port[i-1] = port;
4327                 port->cm_dev = cm_dev;
4328                 port->port_num = i;
4329
4330                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
4331                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
4332
4333                 ret = cm_create_port_fs(port);
4334                 if (ret)
4335                         goto error1;
4336
4337                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4338                                                         IB_QPT_GSI,
4339                                                         &reg_req,
4340                                                         0,
4341                                                         cm_send_handler,
4342                                                         cm_recv_handler,
4343                                                         port,
4344                                                         0);
4345                 if (IS_ERR(port->mad_agent))
4346                         goto error2;
4347
4348                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4349                 if (ret)
4350                         goto error3;
4351
4352                 count++;
4353         }
4354
4355         if (!count)
4356                 goto free;
4357
4358         ib_set_client_data(ib_device, &cm_client, cm_dev);
4359
4360         write_lock_irqsave(&cm.device_lock, flags);
4361         list_add_tail(&cm_dev->list, &cm.device_list);
4362         write_unlock_irqrestore(&cm.device_lock, flags);
4363         return;
4364
4365 error3:
4366         ib_unregister_mad_agent(port->mad_agent);
4367 error2:
4368         cm_remove_port_fs(port);
4369 error1:
4370         port_modify.set_port_cap_mask = 0;
4371         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4372         while (--i) {
4373                 if (!rdma_cap_ib_cm(ib_device, i))
4374                         continue;
4375
4376                 port = cm_dev->port[i-1];
4377                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4378                 ib_unregister_mad_agent(port->mad_agent);
4379                 cm_remove_port_fs(port);
4380         }
4381 free:
4382         device_unregister(cm_dev->device);
4383         kfree(cm_dev);
4384 }
4385
4386 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4387 {
4388         struct cm_device *cm_dev = client_data;
4389         struct cm_port *port;
4390         struct cm_id_private *cm_id_priv;
4391         struct ib_mad_agent *cur_mad_agent;
4392         struct ib_port_modify port_modify = {
4393                 .clr_port_cap_mask = IB_PORT_CM_SUP
4394         };
4395         unsigned long flags;
4396         int i;
4397
4398         if (!cm_dev)
4399                 return;
4400
4401         write_lock_irqsave(&cm.device_lock, flags);
4402         list_del(&cm_dev->list);
4403         write_unlock_irqrestore(&cm.device_lock, flags);
4404
4405         spin_lock_irq(&cm.lock);
4406         cm_dev->going_down = 1;
4407         spin_unlock_irq(&cm.lock);
4408
4409         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4410                 if (!rdma_cap_ib_cm(ib_device, i))
4411                         continue;
4412
4413                 port = cm_dev->port[i-1];
4414                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4415                 /* Mark all the cm_id's as not valid */
4416                 spin_lock_irq(&cm.lock);
4417                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4418                         cm_id_priv->altr_send_port_not_ready = 1;
4419                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4420                         cm_id_priv->prim_send_port_not_ready = 1;
4421                 spin_unlock_irq(&cm.lock);
4422                 /*
4423                  * We flush the queue here after the going_down set, this
4424                  * verify that no new works will be queued in the recv handler,
4425                  * after that we can call the unregister_mad_agent
4426                  */
4427                 flush_workqueue(cm.wq);
4428                 spin_lock_irq(&cm.state_lock);
4429                 cur_mad_agent = port->mad_agent;
4430                 port->mad_agent = NULL;
4431                 spin_unlock_irq(&cm.state_lock);
4432                 ib_unregister_mad_agent(cur_mad_agent);
4433                 cm_remove_port_fs(port);
4434         }
4435
4436         device_unregister(cm_dev->device);
4437         kfree(cm_dev);
4438 }
4439
4440 static int __init ib_cm_init(void)
4441 {
4442         int ret;
4443
4444         memset(&cm, 0, sizeof cm);
4445         INIT_LIST_HEAD(&cm.device_list);
4446         rwlock_init(&cm.device_lock);
4447         spin_lock_init(&cm.lock);
4448         spin_lock_init(&cm.state_lock);
4449         cm.listen_service_table = RB_ROOT;
4450         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4451         cm.remote_id_table = RB_ROOT;
4452         cm.remote_qp_table = RB_ROOT;
4453         cm.remote_sidr_table = RB_ROOT;
4454         idr_init(&cm.local_id_table);
4455         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4456         INIT_LIST_HEAD(&cm.timewait_list);
4457
4458         ret = class_register(&cm_class);
4459         if (ret) {
4460                 ret = -ENOMEM;
4461                 goto error1;
4462         }
4463
4464         cm.wq = alloc_workqueue("ib_cm", 0, 1);
4465         if (!cm.wq) {
4466                 ret = -ENOMEM;
4467                 goto error2;
4468         }
4469
4470         ret = ib_register_client(&cm_client);
4471         if (ret)
4472                 goto error3;
4473
4474         return 0;
4475 error3:
4476         destroy_workqueue(cm.wq);
4477 error2:
4478         class_unregister(&cm_class);
4479 error1:
4480         idr_destroy(&cm.local_id_table);
4481         return ret;
4482 }
4483
4484 static void __exit ib_cm_cleanup(void)
4485 {
4486         struct cm_timewait_info *timewait_info, *tmp;
4487
4488         spin_lock_irq(&cm.lock);
4489         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4490                 cancel_delayed_work(&timewait_info->work.work);
4491         spin_unlock_irq(&cm.lock);
4492
4493         ib_unregister_client(&cm_client);
4494         destroy_workqueue(cm.wq);
4495
4496         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4497                 list_del(&timewait_info->list);
4498                 kfree(timewait_info);
4499         }
4500
4501         class_unregister(&cm_class);
4502         idr_destroy(&cm.local_id_table);
4503 }
4504
4505 module_init(ib_cm_init);
4506 module_exit(ib_cm_cleanup);
4507