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