drivers/infiniband/core/cm.c

   1 /*
   2  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/dma-mapping.h>
  38 #include <linux/device.h>
  39 #include <linux/module.h>
  40 #include <linux/err.h>
  41 #include <linux/idr.h>
  42 #include <linux/interrupt.h>
  43 #include <linux/random.h>
  44 #include <linux/rbtree.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/slab.h>
  47 #include <linux/sysfs.h>
  48 #include <linux/workqueue.h>
  49 #include <linux/kdev_t.h>
  50 #include <linux/etherdevice.h>
  51
  52 #include <rdma/ib_cache.h>
  53 #include <rdma/ib_cm.h>
  54 #include "cm_msgs.h"
  55
  56 MODULE_AUTHOR("Sean Hefty");
  57 MODULE_DESCRIPTION("InfiniBand CM");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59
  60 static void cm_add_one(struct ib_device *device);
  61 static void cm_remove_one(struct ib_device *device, void *client_data);
  62
  63 static struct ib_client cm_client = {
  64         .name   = "cm",
  65         .add    = cm_add_one,
  66         .remove = cm_remove_one
  67 };
  68
  69 static struct ib_cm {
  70         spinlock_t lock;
  71         struct list_head device_list;
  72         rwlock_t device_lock;
  73         struct rb_root listen_service_table;
  74         u64 listen_service_id;
  75         /* struct rb_root peer_service_table; todo: fix peer to peer */
  76         struct rb_root remote_qp_table;
  77         struct rb_root remote_id_table;
  78         struct rb_root remote_sidr_table;
  79         struct idr local_id_table;
  80         __be32 random_id_operand;
  81         struct list_head timewait_list;
  82         struct workqueue_struct *wq;
  83         /* Sync on cm change port state */
  84         spinlock_t state_lock;
  85 } cm;
  86
  87 /* Counter indexes ordered by attribute ID */
  88 enum {
  89         CM_REQ_COUNTER,
  90         CM_MRA_COUNTER,
  91         CM_REJ_COUNTER,
  92         CM_REP_COUNTER,
  93         CM_RTU_COUNTER,
  94         CM_DREQ_COUNTER,
  95         CM_DREP_COUNTER,
  96         CM_SIDR_REQ_COUNTER,
  97         CM_SIDR_REP_COUNTER,
  98         CM_LAP_COUNTER,
  99         CM_APR_COUNTER,
 100         CM_ATTR_COUNT,
 101         CM_ATTR_ID_OFFSET = 0x0010,
 102 };
 103
 104 enum {
 105         CM_XMIT,
 106         CM_XMIT_RETRIES,
 107         CM_RECV,
 108         CM_RECV_DUPLICATES,
 109         CM_COUNTER_GROUPS
 110 };
 111
 112 static char const counter_group_names[CM_COUNTER_GROUPS]
 113                                      [sizeof("cm_rx_duplicates")] = {
 114         "cm_tx_msgs", "cm_tx_retries",
 115         "cm_rx_msgs", "cm_rx_duplicates"
 116 };
 117
 118 struct cm_counter_group {
 119         struct kobject obj;
 120         atomic_long_t counter[CM_ATTR_COUNT];
 121 };
 122
 123 struct cm_counter_attribute {
 124         struct attribute attr;
 125         int index;
 126 };
 127
 128 #define CM_COUNTER_ATTR(_name, _index) \
 129 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 130         .attr = { .name = __stringify(_name), .mode = 0444 }, \
 131         .index = _index \
 132 }
 133
 134 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 135 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 136 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 137 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 138 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 139 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 140 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 141 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 142 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 143 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 144 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 145
 146 static struct attribute *cm_counter_default_attrs[] = {
 147         &cm_req_counter_attr.attr,
 148         &cm_mra_counter_attr.attr,
 149         &cm_rej_counter_attr.attr,
 150         &cm_rep_counter_attr.attr,
 151         &cm_rtu_counter_attr.attr,
 152         &cm_dreq_counter_attr.attr,
 153         &cm_drep_counter_attr.attr,
 154         &cm_sidr_req_counter_attr.attr,
 155         &cm_sidr_rep_counter_attr.attr,
 156         &cm_lap_counter_attr.attr,
 157         &cm_apr_counter_attr.attr,
 158         NULL
 159 };
 160
 161 struct cm_port {
 162         struct cm_device *cm_dev;
 163         struct ib_mad_agent *mad_agent;
 164         struct kobject port_obj;
 165         u8 port_num;
 166         struct list_head cm_priv_prim_list;
 167         struct list_head cm_priv_altr_list;
 168         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 169 };
 170
 171 struct cm_device {
 172         struct list_head list;
 173         struct ib_device *ib_device;
 174         struct device *device;
 175         u8 ack_delay;
 176         int going_down;
 177         struct cm_port *port[0];
 178 };
 179
 180 struct cm_av {
 181         struct cm_port *port;
 182         union ib_gid dgid;
 183         struct ib_ah_attr ah_attr;
 184         u16 pkey_index;
 185         u8 timeout;
 186 };
 187
 188 struct cm_work {
 189         struct delayed_work work;
 190         struct list_head list;
 191         struct cm_port *port;
 192         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 193         __be32 local_id;                        /* Established / timewait */
 194         __be32 remote_id;
 195         struct ib_cm_event cm_event;
 196         struct ib_sa_path_rec path[0];
 197 };
 198
 199 struct cm_timewait_info {
 200         struct cm_work work;                    /* Must be first. */
 201         struct list_head list;
 202         struct rb_node remote_qp_node;
 203         struct rb_node remote_id_node;
 204         __be64 remote_ca_guid;
 205         __be32 remote_qpn;
 206         u8 inserted_remote_qp;
 207         u8 inserted_remote_id;
 208 };
 209
 210 struct cm_id_private {
 211         struct ib_cm_id id;
 212
 213         struct rb_node service_node;
 214         struct rb_node sidr_id_node;
 215         spinlock_t lock;        /* Do not acquire inside cm.lock */
 216         struct completion comp;
 217         atomic_t refcount;
 218         /* Number of clients sharing this ib_cm_id. Only valid for listeners.
 219          * Protected by the cm.lock spinlock. */
 220         int listen_sharecount;
 221
 222         struct ib_mad_send_buf *msg;
 223         struct cm_timewait_info *timewait_info;
 224         /* todo: use alternate port on send failure */
 225         struct cm_av av;
 226         struct cm_av alt_av;
 227
 228         void *private_data;
 229         __be64 tid;
 230         __be32 local_qpn;
 231         __be32 remote_qpn;
 232         enum ib_qp_type qp_type;
 233         __be32 sq_psn;
 234         __be32 rq_psn;
 235         int timeout_ms;
 236         enum ib_mtu path_mtu;
 237         __be16 pkey;
 238         u8 private_data_len;
 239         u8 max_cm_retries;
 240         u8 peer_to_peer;
 241         u8 responder_resources;
 242         u8 initiator_depth;
 243         u8 retry_count;
 244         u8 rnr_retry_count;
 245         u8 service_timeout;
 246         u8 target_ack_delay;
 247
 248         struct list_head prim_list;
 249         struct list_head altr_list;
 250         /* Indicates that the send port mad is registered and av is set */
 251         int prim_send_port_not_ready;
 252         int altr_send_port_not_ready;
 253
 254         struct list_head work_list;
 255         atomic_t work_count;
 256 };
 257
 258 static void cm_work_handler(struct work_struct *work);
 259
 260 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 261 {
 262         if (atomic_dec_and_test(&cm_id_priv->refcount))
 263                 complete(&cm_id_priv->comp);
 264 }
 265
 266 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 267                         struct ib_mad_send_buf **msg)
 268 {
 269         struct ib_mad_agent *mad_agent;
 270         struct ib_mad_send_buf *m;
 271         struct ib_ah *ah;
 272         struct cm_av *av;
 273         unsigned long flags, flags2;
 274         int ret = 0;
 275
 276         /* don't let the port to be released till the agent is down */
 277         spin_lock_irqsave(&cm.state_lock, flags2);
 278         spin_lock_irqsave(&cm.lock, flags);
 279         if (!cm_id_priv->prim_send_port_not_ready)
 280                 av = &cm_id_priv->av;
 281         else if (!cm_id_priv->altr_send_port_not_ready &&
 282                  (cm_id_priv->alt_av.port))
 283                 av = &cm_id_priv->alt_av;
 284         else {
 285                 pr_info("%s: not valid CM id\n", __func__);
 286                 ret = -ENODEV;
 287                 spin_unlock_irqrestore(&cm.lock, flags);
 288                 goto out;
 289         }
 290         spin_unlock_irqrestore(&cm.lock, flags);
 291         /* Make sure the port haven't released the mad yet */
 292         mad_agent = cm_id_priv->av.port->mad_agent;
 293         if (!mad_agent) {
 294                 pr_info("%s: not a valid MAD agent\n", __func__);
 295                 ret = -ENODEV;
 296                 goto out;
 297         }
 298         ah = ib_create_ah(mad_agent->qp->pd, &av->ah_attr);
 299         if (IS_ERR(ah)) {
 300                 ret = PTR_ERR(ah);
 301                 goto out;
 302         }
 303
 304         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 305                                av->pkey_index,
 306                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 307                                GFP_ATOMIC,
 308                                IB_MGMT_BASE_VERSION);
 309         if (IS_ERR(m)) {
 310                 ib_destroy_ah(ah);
 311                 ret = PTR_ERR(m);
 312                 goto out;
 313         }
 314
 315         /* Timeout set by caller if response is expected. */
 316         m->ah = ah;
 317         m->retries = cm_id_priv->max_cm_retries;
 318
 319         atomic_inc(&cm_id_priv->refcount);
 320         m->context[0] = cm_id_priv;
 321         *msg = m;
 322
 323 out:
 324         spin_unlock_irqrestore(&cm.state_lock, flags2);
 325         return ret;
 326 }
 327
 328 static int cm_alloc_response_msg(struct cm_port *port,
 329                                  struct ib_mad_recv_wc *mad_recv_wc,
 330                                  struct ib_mad_send_buf **msg)
 331 {
 332         struct ib_mad_send_buf *m;
 333         struct ib_ah *ah;
 334
 335         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 336                                   mad_recv_wc->recv_buf.grh, port->port_num);
 337         if (IS_ERR(ah))
 338                 return PTR_ERR(ah);
 339
 340         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 341                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 342                                GFP_ATOMIC,
 343                                IB_MGMT_BASE_VERSION);
 344         if (IS_ERR(m)) {
 345                 ib_destroy_ah(ah);
 346                 return PTR_ERR(m);
 347         }
 348         m->ah = ah;
 349         *msg = m;
 350         return 0;
 351 }
 352
 353 static void cm_free_msg(struct ib_mad_send_buf *msg)
 354 {
 355         ib_destroy_ah(msg->ah);
 356         if (msg->context[0])
 357                 cm_deref_id(msg->context[0]);
 358         ib_free_send_mad(msg);
 359 }
 360
 361 static void * cm_copy_private_data(const void *private_data,
 362                                    u8 private_data_len)
 363 {
 364         void *data;
 365
 366         if (!private_data || !private_data_len)
 367                 return NULL;
 368
 369         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 370         if (!data)
 371                 return ERR_PTR(-ENOMEM);
 372
 373         return data;
 374 }
 375
 376 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 377                                  void *private_data, u8 private_data_len)
 378 {
 379         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 380                 kfree(cm_id_priv->private_data);
 381
 382         cm_id_priv->private_data = private_data;
 383         cm_id_priv->private_data_len = private_data_len;
 384 }
 385
 386 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 387                                     struct ib_grh *grh, struct cm_av *av)
 388 {
 389         av->port = port;
 390         av->pkey_index = wc->pkey_index;
 391         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 392                            grh, &av->ah_attr);
 393 }
 394
 395 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av,
 396                               struct cm_id_private *cm_id_priv)
 397 {
 398         struct cm_device *cm_dev;
 399         struct cm_port *port = NULL;
 400         unsigned long flags;
 401         int ret;
 402         u8 p;
 403         struct net_device *ndev = ib_get_ndev_from_path(path);
 404
 405         read_lock_irqsave(&cm.device_lock, flags);
 406         list_for_each_entry(cm_dev, &cm.device_list, list) {
 407                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 408                                         path->gid_type, ndev, &p, NULL)) {
 409                         port = cm_dev->port[p-1];
 410                         break;
 411                 }
 412         }
 413         read_unlock_irqrestore(&cm.device_lock, flags);
 414
 415         if (ndev)
 416                 dev_put(ndev);
 417
 418         if (!port)
 419                 return -EINVAL;
 420
 421         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 422                                   be16_to_cpu(path->pkey), &av->pkey_index);
 423         if (ret)
 424                 return ret;
 425
 426         av->port = port;
 427         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 428                              &av->ah_attr);
 429         av->timeout = path->packet_life_time + 1;
 430
 431         spin_lock_irqsave(&cm.lock, flags);
 432         if (&cm_id_priv->av == av)
 433                 list_add_tail(&cm_id_priv->prim_list, &port->cm_priv_prim_list);
 434         else if (&cm_id_priv->alt_av == av)
 435                 list_add_tail(&cm_id_priv->altr_list, &port->cm_priv_altr_list);
 436         else
 437                 ret = -EINVAL;
 438
 439         spin_unlock_irqrestore(&cm.lock, flags);
 440
 441         return ret;
 442 }
 443
 444 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 445 {
 446         unsigned long flags;
 447         int id;
 448
 449         idr_preload(GFP_KERNEL);
 450         spin_lock_irqsave(&cm.lock, flags);
 451
 452         id = idr_alloc_cyclic(&cm.local_id_table, cm_id_priv, 0, 0, GFP_NOWAIT);
 453
 454         spin_unlock_irqrestore(&cm.lock, flags);
 455         idr_preload_end();
 456
 457         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 458         return id < 0 ? id : 0;
 459 }
 460
 461 static void cm_free_id(__be32 local_id)
 462 {
 463         spin_lock_irq(&cm.lock);
 464         idr_remove(&cm.local_id_table,
 465                    (__force int) (local_id ^ cm.random_id_operand));
 466         spin_unlock_irq(&cm.lock);
 467 }
 468
 469 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 470 {
 471         struct cm_id_private *cm_id_priv;
 472
 473         cm_id_priv = idr_find(&cm.local_id_table,
 474                               (__force int) (local_id ^ cm.random_id_operand));
 475         if (cm_id_priv) {
 476                 if (cm_id_priv->id.remote_id == remote_id)
 477                         atomic_inc(&cm_id_priv->refcount);
 478                 else
 479                         cm_id_priv = NULL;
 480         }
 481
 482         return cm_id_priv;
 483 }
 484
 485 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 486 {
 487         struct cm_id_private *cm_id_priv;
 488
 489         spin_lock_irq(&cm.lock);
 490         cm_id_priv = cm_get_id(local_id, remote_id);
 491         spin_unlock_irq(&cm.lock);
 492
 493         return cm_id_priv;
 494 }
 495
 496 /*
 497  * Trivial helpers to strip endian annotation and compare; the
 498  * endianness doesn't actually matter since we just need a stable
 499  * order for the RB tree.
 500  */
 501 static int be32_lt(__be32 a, __be32 b)
 502 {
 503         return (__force u32) a < (__force u32) b;
 504 }
 505
 506 static int be32_gt(__be32 a, __be32 b)
 507 {
 508         return (__force u32) a > (__force u32) b;
 509 }
 510
 511 static int be64_lt(__be64 a, __be64 b)
 512 {
 513         return (__force u64) a < (__force u64) b;
 514 }
 515
 516 static int be64_gt(__be64 a, __be64 b)
 517 {
 518         return (__force u64) a > (__force u64) b;
 519 }
 520
 521 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 522 {
 523         struct rb_node **link = &cm.listen_service_table.rb_node;
 524         struct rb_node *parent = NULL;
 525         struct cm_id_private *cur_cm_id_priv;
 526         __be64 service_id = cm_id_priv->id.service_id;
 527         __be64 service_mask = cm_id_priv->id.service_mask;
 528
 529         while (*link) {
 530                 parent = *link;
 531                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 532                                           service_node);
 533                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 534                     (service_mask & cur_cm_id_priv->id.service_id) &&
 535                     (cm_id_priv->id.device == cur_cm_id_priv->id.device))
 536                         return cur_cm_id_priv;
 537
 538                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 539                         link = &(*link)->rb_left;
 540                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 541                         link = &(*link)->rb_right;
 542                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 543                         link = &(*link)->rb_left;
 544                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 545                         link = &(*link)->rb_right;
 546                 else
 547                         link = &(*link)->rb_right;
 548         }
 549         rb_link_node(&cm_id_priv->service_node, parent, link);
 550         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 551         return NULL;
 552 }
 553
 554 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 555                                              __be64 service_id)
 556 {
 557         struct rb_node *node = cm.listen_service_table.rb_node;
 558         struct cm_id_private *cm_id_priv;
 559
 560         while (node) {
 561                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 562                 if ((cm_id_priv->id.service_mask & service_id) ==
 563                      cm_id_priv->id.service_id &&
 564                     (cm_id_priv->id.device == device))
 565                         return cm_id_priv;
 566
 567                 if (device < cm_id_priv->id.device)
 568                         node = node->rb_left;
 569                 else if (device > cm_id_priv->id.device)
 570                         node = node->rb_right;
 571                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 572                         node = node->rb_left;
 573                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 574                         node = node->rb_right;
 575                 else
 576                         node = node->rb_right;
 577         }
 578         return NULL;
 579 }
 580
 581 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 582                                                      *timewait_info)
 583 {
 584         struct rb_node **link = &cm.remote_id_table.rb_node;
 585         struct rb_node *parent = NULL;
 586         struct cm_timewait_info *cur_timewait_info;
 587         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 588         __be32 remote_id = timewait_info->work.remote_id;
 589
 590         while (*link) {
 591                 parent = *link;
 592                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 593                                              remote_id_node);
 594                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 595                         link = &(*link)->rb_left;
 596                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 597                         link = &(*link)->rb_right;
 598                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 599                         link = &(*link)->rb_left;
 600                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 601                         link = &(*link)->rb_right;
 602                 else
 603                         return cur_timewait_info;
 604         }
 605         timewait_info->inserted_remote_id = 1;
 606         rb_link_node(&timewait_info->remote_id_node, parent, link);
 607         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 608         return NULL;
 609 }
 610
 611 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 612                                                    __be32 remote_id)
 613 {
 614         struct rb_node *node = cm.remote_id_table.rb_node;
 615         struct cm_timewait_info *timewait_info;
 616
 617         while (node) {
 618                 timewait_info = rb_entry(node, struct cm_timewait_info,
 619                                          remote_id_node);
 620                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 621                         node = node->rb_left;
 622                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 623                         node = node->rb_right;
 624                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 625                         node = node->rb_left;
 626                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 627                         node = node->rb_right;
 628                 else
 629                         return timewait_info;
 630         }
 631         return NULL;
 632 }
 633
 634 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 635                                                       *timewait_info)
 636 {
 637         struct rb_node **link = &cm.remote_qp_table.rb_node;
 638         struct rb_node *parent = NULL;
 639         struct cm_timewait_info *cur_timewait_info;
 640         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 641         __be32 remote_qpn = timewait_info->remote_qpn;
 642
 643         while (*link) {
 644                 parent = *link;
 645                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 646                                              remote_qp_node);
 647                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 648                         link = &(*link)->rb_left;
 649                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 650                         link = &(*link)->rb_right;
 651                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 652                         link = &(*link)->rb_left;
 653                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 654                         link = &(*link)->rb_right;
 655                 else
 656                         return cur_timewait_info;
 657         }
 658         timewait_info->inserted_remote_qp = 1;
 659         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 660         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 661         return NULL;
 662 }
 663
 664 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 665                                                     *cm_id_priv)
 666 {
 667         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 668         struct rb_node *parent = NULL;
 669         struct cm_id_private *cur_cm_id_priv;
 670         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 671         __be32 remote_id = cm_id_priv->id.remote_id;
 672
 673         while (*link) {
 674                 parent = *link;
 675                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 676                                           sidr_id_node);
 677                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 678                         link = &(*link)->rb_left;
 679                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 680                         link = &(*link)->rb_right;
 681                 else {
 682                         int cmp;
 683                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 684                                      sizeof *port_gid);
 685                         if (cmp < 0)
 686                                 link = &(*link)->rb_left;
 687                         else if (cmp > 0)
 688                                 link = &(*link)->rb_right;
 689                         else
 690                                 return cur_cm_id_priv;
 691                 }
 692         }
 693         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 694         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 695         return NULL;
 696 }
 697
 698 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 699                                enum ib_cm_sidr_status status)
 700 {
 701         struct ib_cm_sidr_rep_param param;
 702
 703         memset(&param, 0, sizeof param);
 704         param.status = status;
 705         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 706 }
 707
 708 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 709                                  ib_cm_handler cm_handler,
 710                                  void *context)
 711 {
 712         struct cm_id_private *cm_id_priv;
 713         int ret;
 714
 715         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 716         if (!cm_id_priv)
 717                 return ERR_PTR(-ENOMEM);
 718
 719         cm_id_priv->id.state = IB_CM_IDLE;
 720         cm_id_priv->id.device = device;
 721         cm_id_priv->id.cm_handler = cm_handler;
 722         cm_id_priv->id.context = context;
 723         cm_id_priv->id.remote_cm_qpn = 1;
 724         ret = cm_alloc_id(cm_id_priv);
 725         if (ret)
 726                 goto error;
 727
 728         spin_lock_init(&cm_id_priv->lock);
 729         init_completion(&cm_id_priv->comp);
 730         INIT_LIST_HEAD(&cm_id_priv->work_list);
 731         INIT_LIST_HEAD(&cm_id_priv->prim_list);
 732         INIT_LIST_HEAD(&cm_id_priv->altr_list);
 733         atomic_set(&cm_id_priv->work_count, -1);
 734         atomic_set(&cm_id_priv->refcount, 1);
 735         return &cm_id_priv->id;
 736
 737 error:
 738         kfree(cm_id_priv);
 739         return ERR_PTR(-ENOMEM);
 740 }
 741 EXPORT_SYMBOL(ib_create_cm_id);
 742
 743 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 744 {
 745         struct cm_work *work;
 746
 747         if (list_empty(&cm_id_priv->work_list))
 748                 return NULL;
 749
 750         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 751         list_del(&work->list);
 752         return work;
 753 }
 754
 755 static void cm_free_work(struct cm_work *work)
 756 {
 757         if (work->mad_recv_wc)
 758                 ib_free_recv_mad(work->mad_recv_wc);
 759         kfree(work);
 760 }
 761
 762 static inline int cm_convert_to_ms(int iba_time)
 763 {
 764         /* approximate conversion to ms from 4.096us x 2^iba_time */
 765         return 1 << max(iba_time - 8, 0);
 766 }
 767
 768 /*
 769  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 770  * Because of how ack_timeout is stored, adding one doubles the timeout.
 771  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 772  * increment it (round up) only if the other is within 50%.
 773  */
 774 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 775 {
 776         int ack_timeout = packet_life_time + 1;
 777
 778         if (ack_timeout >= ca_ack_delay)
 779                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 780         else
 781                 ack_timeout = ca_ack_delay +
 782                               (ack_timeout >= (ca_ack_delay - 1));
 783
 784         return min(31, ack_timeout);
 785 }
 786
 787 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 788 {
 789         if (timewait_info->inserted_remote_id) {
 790                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 791                 timewait_info->inserted_remote_id = 0;
 792         }
 793
 794         if (timewait_info->inserted_remote_qp) {
 795                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 796                 timewait_info->inserted_remote_qp = 0;
 797         }
 798 }
 799
 800 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 801 {
 802         struct cm_timewait_info *timewait_info;
 803
 804         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 805         if (!timewait_info)
 806                 return ERR_PTR(-ENOMEM);
 807
 808         timewait_info->work.local_id = local_id;
 809         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 810         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 811         return timewait_info;
 812 }
 813
 814 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 815 {
 816         int wait_time;
 817         unsigned long flags;
 818         struct cm_device *cm_dev;
 819
 820         cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
 821         if (!cm_dev)
 822                 return;
 823
 824         spin_lock_irqsave(&cm.lock, flags);
 825         cm_cleanup_timewait(cm_id_priv->timewait_info);
 826         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 827         spin_unlock_irqrestore(&cm.lock, flags);
 828
 829         /*
 830          * The cm_id could be destroyed by the user before we exit timewait.
 831          * To protect against this, we search for the cm_id after exiting
 832          * timewait before notifying the user that we've exited timewait.
 833          */
 834         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 835         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 836
 837         /* Check if the device started its remove_one */
 838         spin_lock_irqsave(&cm.lock, flags);
 839         if (!cm_dev->going_down)
 840                 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 841                                    msecs_to_jiffies(wait_time));
 842         spin_unlock_irqrestore(&cm.lock, flags);
 843
 844         cm_id_priv->timewait_info = NULL;
 845 }
 846
 847 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 848 {
 849         unsigned long flags;
 850
 851         cm_id_priv->id.state = IB_CM_IDLE;
 852         if (cm_id_priv->timewait_info) {
 853                 spin_lock_irqsave(&cm.lock, flags);
 854                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 855                 spin_unlock_irqrestore(&cm.lock, flags);
 856                 kfree(cm_id_priv->timewait_info);
 857                 cm_id_priv->timewait_info = NULL;
 858         }
 859 }
 860
 861 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 862 {
 863         struct cm_id_private *cm_id_priv;
 864         struct cm_work *work;
 865
 866         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 867 retest:
 868         spin_lock_irq(&cm_id_priv->lock);
 869         switch (cm_id->state) {
 870         case IB_CM_LISTEN:
 871                 spin_unlock_irq(&cm_id_priv->lock);
 872
 873                 spin_lock_irq(&cm.lock);
 874                 if (--cm_id_priv->listen_sharecount > 0) {
 875                         /* The id is still shared. */
 876                         cm_deref_id(cm_id_priv);
 877                         spin_unlock_irq(&cm.lock);
 878                         return;
 879                 }
 880                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 881                 spin_unlock_irq(&cm.lock);
 882                 break;
 883         case IB_CM_SIDR_REQ_SENT:
 884                 cm_id->state = IB_CM_IDLE;
 885                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 886                 spin_unlock_irq(&cm_id_priv->lock);
 887                 break;
 888         case IB_CM_SIDR_REQ_RCVD:
 889                 spin_unlock_irq(&cm_id_priv->lock);
 890                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 891                 spin_lock_irq(&cm.lock);
 892                 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node))
 893                         rb_erase(&cm_id_priv->sidr_id_node,
 894                                  &cm.remote_sidr_table);
 895                 spin_unlock_irq(&cm.lock);
 896                 break;
 897         case IB_CM_REQ_SENT:
 898         case IB_CM_MRA_REQ_RCVD:
 899                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 900                 spin_unlock_irq(&cm_id_priv->lock);
 901                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 902                                &cm_id_priv->id.device->node_guid,
 903                                sizeof cm_id_priv->id.device->node_guid,
 904                                NULL, 0);
 905                 break;
 906         case IB_CM_REQ_RCVD:
 907                 if (err == -ENOMEM) {
 908                         /* Do not reject to allow future retries. */
 909                         cm_reset_to_idle(cm_id_priv);
 910                         spin_unlock_irq(&cm_id_priv->lock);
 911                 } else {
 912                         spin_unlock_irq(&cm_id_priv->lock);
 913                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 914                                        NULL, 0, NULL, 0);
 915                 }
 916                 break;
 917         case IB_CM_REP_SENT:
 918         case IB_CM_MRA_REP_RCVD:
 919                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 920                 /* Fall through */
 921         case IB_CM_MRA_REQ_SENT:
 922         case IB_CM_REP_RCVD:
 923         case IB_CM_MRA_REP_SENT:
 924                 spin_unlock_irq(&cm_id_priv->lock);
 925                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 926                                NULL, 0, NULL, 0);
 927                 break;
 928         case IB_CM_ESTABLISHED:
 929                 spin_unlock_irq(&cm_id_priv->lock);
 930                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT)
 931                         break;
 932                 ib_send_cm_dreq(cm_id, NULL, 0);
 933                 goto retest;
 934         case IB_CM_DREQ_SENT:
 935                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 936                 cm_enter_timewait(cm_id_priv);
 937                 spin_unlock_irq(&cm_id_priv->lock);
 938                 break;
 939         case IB_CM_DREQ_RCVD:
 940                 spin_unlock_irq(&cm_id_priv->lock);
 941                 ib_send_cm_drep(cm_id, NULL, 0);
 942                 break;
 943         default:
 944                 spin_unlock_irq(&cm_id_priv->lock);
 945                 break;
 946         }
 947
 948         spin_lock_irq(&cm.lock);
 949         if (!list_empty(&cm_id_priv->altr_list) &&
 950             (!cm_id_priv->altr_send_port_not_ready))
 951                 list_del(&cm_id_priv->altr_list);
 952         if (!list_empty(&cm_id_priv->prim_list) &&
 953             (!cm_id_priv->prim_send_port_not_ready))
 954                 list_del(&cm_id_priv->prim_list);
 955         spin_unlock_irq(&cm.lock);
 956
 957         cm_free_id(cm_id->local_id);
 958         cm_deref_id(cm_id_priv);
 959         wait_for_completion(&cm_id_priv->comp);
 960         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
 961                 cm_free_work(work);
 962         kfree(cm_id_priv->private_data);
 963         kfree(cm_id_priv);
 964 }
 965
 966 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
 967 {
 968         cm_destroy_id(cm_id, 0);
 969 }
 970 EXPORT_SYMBOL(ib_destroy_cm_id);
 971
 972 /**
 973  * __ib_cm_listen - Initiates listening on the specified service ID for
 974  *   connection and service ID resolution requests.
 975  * @cm_id: Connection identifier associated with the listen request.
 976  * @service_id: Service identifier matched against incoming connection
 977  *   and service ID resolution requests.  The service ID should be specified
 978  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
 979  *   assign a service ID to the caller.
 980  * @service_mask: Mask applied to service ID used to listen across a
 981  *   range of service IDs.  If set to 0, the service ID is matched
 982  *   exactly.  This parameter is ignored if %service_id is set to
 983  *   IB_CM_ASSIGN_SERVICE_ID.
 984  */
 985 static int __ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id,
 986                           __be64 service_mask)
 987 {
 988         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
 989         int ret = 0;
 990
 991         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
 992         service_id &= service_mask;
 993         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
 994             (service_id != IB_CM_ASSIGN_SERVICE_ID))
 995                 return -EINVAL;
 996
 997         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 998         if (cm_id->state != IB_CM_IDLE)
 999                 return -EINVAL;
1000
1001         cm_id->state = IB_CM_LISTEN;
1002         ++cm_id_priv->listen_sharecount;
1003
1004         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
1005                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
1006                 cm_id->service_mask = ~cpu_to_be64(0);
1007         } else {
1008                 cm_id->service_id = service_id;
1009                 cm_id->service_mask = service_mask;
1010         }
1011         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
1012
1013         if (cur_cm_id_priv) {
1014                 cm_id->state = IB_CM_IDLE;
1015                 --cm_id_priv->listen_sharecount;
1016                 ret = -EBUSY;
1017         }
1018         return ret;
1019 }
1020
1021 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask)
1022 {
1023         unsigned long flags;
1024         int ret;
1025
1026         spin_lock_irqsave(&cm.lock, flags);
1027         ret = __ib_cm_listen(cm_id, service_id, service_mask);
1028         spin_unlock_irqrestore(&cm.lock, flags);
1029
1030         return ret;
1031 }
1032 EXPORT_SYMBOL(ib_cm_listen);
1033
1034 /**
1035  * Create a new listening ib_cm_id and listen on the given service ID.
1036  *
1037  * If there's an existing ID listening on that same device and service ID,
1038  * return it.
1039  *
1040  * @device: Device associated with the cm_id.  All related communication will
1041  * be associated with the specified device.
1042  * @cm_handler: Callback invoked to notify the user of CM events.
1043  * @service_id: Service identifier matched against incoming connection
1044  *   and service ID resolution requests.  The service ID should be specified
1045  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1046  *   assign a service ID to the caller.
1047  *
1048  * Callers should call ib_destroy_cm_id when done with the listener ID.
1049  */
1050 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1051                                      ib_cm_handler cm_handler,
1052                                      __be64 service_id)
1053 {
1054         struct cm_id_private *cm_id_priv;
1055         struct ib_cm_id *cm_id;
1056         unsigned long flags;
1057         int err = 0;
1058
1059         /* Create an ID in advance, since the creation may sleep */
1060         cm_id = ib_create_cm_id(device, cm_handler, NULL);
1061         if (IS_ERR(cm_id))
1062                 return cm_id;
1063
1064         spin_lock_irqsave(&cm.lock, flags);
1065
1066         if (service_id == IB_CM_ASSIGN_SERVICE_ID)
1067                 goto new_id;
1068
1069         /* Find an existing ID */
1070         cm_id_priv = cm_find_listen(device, service_id);
1071         if (cm_id_priv) {
1072                 if (cm_id->cm_handler != cm_handler || cm_id->context) {
1073                         /* Sharing an ib_cm_id with different handlers is not
1074                          * supported */
1075                         spin_unlock_irqrestore(&cm.lock, flags);
1076                         ib_destroy_cm_id(cm_id);
1077                         return ERR_PTR(-EINVAL);
1078                 }
1079                 atomic_inc(&cm_id_priv->refcount);
1080                 ++cm_id_priv->listen_sharecount;
1081                 spin_unlock_irqrestore(&cm.lock, flags);
1082
1083                 ib_destroy_cm_id(cm_id);
1084                 cm_id = &cm_id_priv->id;
1085                 return cm_id;
1086         }
1087
1088 new_id:
1089         /* Use newly created ID */
1090         err = __ib_cm_listen(cm_id, service_id, 0);
1091
1092         spin_unlock_irqrestore(&cm.lock, flags);
1093
1094         if (err) {
1095                 ib_destroy_cm_id(cm_id);
1096                 return ERR_PTR(err);
1097         }
1098         return cm_id;
1099 }
1100 EXPORT_SYMBOL(ib_cm_insert_listen);
1101
1102 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
1103                           enum cm_msg_sequence msg_seq)
1104 {
1105         u64 hi_tid, low_tid;
1106
1107         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1108         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
1109                           (msg_seq << 30));
1110         return cpu_to_be64(hi_tid | low_tid);
1111 }
1112
1113 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1114                               __be16 attr_id, __be64 tid)
1115 {
1116         hdr->base_version  = IB_MGMT_BASE_VERSION;
1117         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1118         hdr->class_version = IB_CM_CLASS_VERSION;
1119         hdr->method        = IB_MGMT_METHOD_SEND;
1120         hdr->attr_id       = attr_id;
1121         hdr->tid           = tid;
1122 }
1123
1124 static void cm_format_req(struct cm_req_msg *req_msg,
1125                           struct cm_id_private *cm_id_priv,
1126                           struct ib_cm_req_param *param)
1127 {
1128         struct ib_sa_path_rec *pri_path = param->primary_path;
1129         struct ib_sa_path_rec *alt_path = param->alternate_path;
1130
1131         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1132                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1133
1134         req_msg->local_comm_id = cm_id_priv->id.local_id;
1135         req_msg->service_id = param->service_id;
1136         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1137         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1138         cm_req_set_init_depth(req_msg, param->initiator_depth);
1139         cm_req_set_remote_resp_timeout(req_msg,
1140                                        param->remote_cm_response_timeout);
1141         cm_req_set_qp_type(req_msg, param->qp_type);
1142         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1143         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1144         cm_req_set_local_resp_timeout(req_msg,
1145                                       param->local_cm_response_timeout);
1146         req_msg->pkey = param->primary_path->pkey;
1147         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1148         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1149
1150         if (param->qp_type != IB_QPT_XRC_INI) {
1151                 cm_req_set_resp_res(req_msg, param->responder_resources);
1152                 cm_req_set_retry_count(req_msg, param->retry_count);
1153                 cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1154                 cm_req_set_srq(req_msg, param->srq);
1155         }
1156
1157         if (pri_path->hop_limit <= 1) {
1158                 req_msg->primary_local_lid = pri_path->slid;
1159                 req_msg->primary_remote_lid = pri_path->dlid;
1160         } else {
1161                 /* Work-around until there's a way to obtain remote LID info */
1162                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1163                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1164         }
1165         req_msg->primary_local_gid = pri_path->sgid;
1166         req_msg->primary_remote_gid = pri_path->dgid;
1167         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1168         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1169         req_msg->primary_traffic_class = pri_path->traffic_class;
1170         req_msg->primary_hop_limit = pri_path->hop_limit;
1171         cm_req_set_primary_sl(req_msg, pri_path->sl);
1172         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1173         cm_req_set_primary_local_ack_timeout(req_msg,
1174                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1175                                pri_path->packet_life_time));
1176
1177         if (alt_path) {
1178                 if (alt_path->hop_limit <= 1) {
1179                         req_msg->alt_local_lid = alt_path->slid;
1180                         req_msg->alt_remote_lid = alt_path->dlid;
1181                 } else {
1182                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1183                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1184                 }
1185                 req_msg->alt_local_gid = alt_path->sgid;
1186                 req_msg->alt_remote_gid = alt_path->dgid;
1187                 cm_req_set_alt_flow_label(req_msg,
1188                                           alt_path->flow_label);
1189                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1190                 req_msg->alt_traffic_class = alt_path->traffic_class;
1191                 req_msg->alt_hop_limit = alt_path->hop_limit;
1192                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1193                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1194                 cm_req_set_alt_local_ack_timeout(req_msg,
1195                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1196                                        alt_path->packet_life_time));
1197         }
1198
1199         if (param->private_data && param->private_data_len)
1200                 memcpy(req_msg->private_data, param->private_data,
1201                        param->private_data_len);
1202 }
1203
1204 static int cm_validate_req_param(struct ib_cm_req_param *param)
1205 {
1206         /* peer-to-peer not supported */
1207         if (param->peer_to_peer)
1208                 return -EINVAL;
1209
1210         if (!param->primary_path)
1211                 return -EINVAL;
1212
1213         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1214             param->qp_type != IB_QPT_XRC_INI)
1215                 return -EINVAL;
1216
1217         if (param->private_data &&
1218             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1219                 return -EINVAL;
1220
1221         if (param->alternate_path &&
1222             (param->alternate_path->pkey != param->primary_path->pkey ||
1223              param->alternate_path->mtu != param->primary_path->mtu))
1224                 return -EINVAL;
1225
1226         return 0;
1227 }
1228
1229 int ib_send_cm_req(struct ib_cm_id *cm_id,
1230                    struct ib_cm_req_param *param)
1231 {
1232         struct cm_id_private *cm_id_priv;
1233         struct cm_req_msg *req_msg;
1234         unsigned long flags;
1235         int ret;
1236
1237         ret = cm_validate_req_param(param);
1238         if (ret)
1239                 return ret;
1240
1241         /* Verify that we're not in timewait. */
1242         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1243         spin_lock_irqsave(&cm_id_priv->lock, flags);
1244         if (cm_id->state != IB_CM_IDLE) {
1245                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1246                 ret = -EINVAL;
1247                 goto out;
1248         }
1249         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1250
1251         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1252                                                             id.local_id);
1253         if (IS_ERR(cm_id_priv->timewait_info)) {
1254                 ret = PTR_ERR(cm_id_priv->timewait_info);
1255                 cm_id_priv->timewait_info = NULL;
1256                 goto out;
1257         }
1258
1259         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av,
1260                                  cm_id_priv);
1261         if (ret)
1262                 goto error1;
1263         if (param->alternate_path) {
1264                 ret = cm_init_av_by_path(param->alternate_path,
1265                                          &cm_id_priv->alt_av, cm_id_priv);
1266                 if (ret)
1267                         goto error1;
1268         }
1269         cm_id->service_id = param->service_id;
1270         cm_id->service_mask = ~cpu_to_be64(0);
1271         cm_id_priv->timeout_ms = cm_convert_to_ms(
1272                                     param->primary_path->packet_life_time) * 2 +
1273                                  cm_convert_to_ms(
1274                                     param->remote_cm_response_timeout);
1275         cm_id_priv->max_cm_retries = param->max_cm_retries;
1276         cm_id_priv->initiator_depth = param->initiator_depth;
1277         cm_id_priv->responder_resources = param->responder_resources;
1278         cm_id_priv->retry_count = param->retry_count;
1279         cm_id_priv->path_mtu = param->primary_path->mtu;
1280         cm_id_priv->pkey = param->primary_path->pkey;
1281         cm_id_priv->qp_type = param->qp_type;
1282
1283         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1284         if (ret)
1285                 goto error1;
1286
1287         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1288         cm_format_req(req_msg, cm_id_priv, param);
1289         cm_id_priv->tid = req_msg->hdr.tid;
1290         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1291         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1292
1293         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1294         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1295
1296         spin_lock_irqsave(&cm_id_priv->lock, flags);
1297         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1298         if (ret) {
1299                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1300                 goto error2;
1301         }
1302         BUG_ON(cm_id->state != IB_CM_IDLE);
1303         cm_id->state = IB_CM_REQ_SENT;
1304         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1305         return 0;
1306
1307 error2: cm_free_msg(cm_id_priv->msg);
1308 error1: kfree(cm_id_priv->timewait_info);
1309 out:    return ret;
1310 }
1311 EXPORT_SYMBOL(ib_send_cm_req);
1312
1313 static int cm_issue_rej(struct cm_port *port,
1314                         struct ib_mad_recv_wc *mad_recv_wc,
1315                         enum ib_cm_rej_reason reason,
1316                         enum cm_msg_response msg_rejected,
1317                         void *ari, u8 ari_length)
1318 {
1319         struct ib_mad_send_buf *msg = NULL;
1320         struct cm_rej_msg *rej_msg, *rcv_msg;
1321         int ret;
1322
1323         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1324         if (ret)
1325                 return ret;
1326
1327         /* We just need common CM header information.  Cast to any message. */
1328         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1329         rej_msg = (struct cm_rej_msg *) msg->mad;
1330
1331         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1332         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1333         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1334         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1335         rej_msg->reason = cpu_to_be16(reason);
1336
1337         if (ari && ari_length) {
1338                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1339                 memcpy(rej_msg->ari, ari, ari_length);
1340         }
1341
1342         ret = ib_post_send_mad(msg, NULL);
1343         if (ret)
1344                 cm_free_msg(msg);
1345
1346         return ret;
1347 }
1348
1349 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1350                                     __be32 local_qpn, __be32 remote_qpn)
1351 {
1352         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1353                 ((local_ca_guid == remote_ca_guid) &&
1354                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1355 }
1356
1357 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1358                                             struct ib_sa_path_rec *primary_path,
1359                                             struct ib_sa_path_rec *alt_path)
1360 {
1361         memset(primary_path, 0, sizeof *primary_path);
1362         primary_path->dgid = req_msg->primary_local_gid;
1363         primary_path->sgid = req_msg->primary_remote_gid;
1364         primary_path->dlid = req_msg->primary_local_lid;
1365         primary_path->slid = req_msg->primary_remote_lid;
1366         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1367         primary_path->hop_limit = req_msg->primary_hop_limit;
1368         primary_path->traffic_class = req_msg->primary_traffic_class;
1369         primary_path->reversible = 1;
1370         primary_path->pkey = req_msg->pkey;
1371         primary_path->sl = cm_req_get_primary_sl(req_msg);
1372         primary_path->mtu_selector = IB_SA_EQ;
1373         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1374         primary_path->rate_selector = IB_SA_EQ;
1375         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1376         primary_path->packet_life_time_selector = IB_SA_EQ;
1377         primary_path->packet_life_time =
1378                 cm_req_get_primary_local_ack_timeout(req_msg);
1379         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1380         primary_path->service_id = req_msg->service_id;
1381
1382         if (req_msg->alt_local_lid) {
1383                 memset(alt_path, 0, sizeof *alt_path);
1384                 alt_path->dgid = req_msg->alt_local_gid;
1385                 alt_path->sgid = req_msg->alt_remote_gid;
1386                 alt_path->dlid = req_msg->alt_local_lid;
1387                 alt_path->slid = req_msg->alt_remote_lid;
1388                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1389                 alt_path->hop_limit = req_msg->alt_hop_limit;
1390                 alt_path->traffic_class = req_msg->alt_traffic_class;
1391                 alt_path->reversible = 1;
1392                 alt_path->pkey = req_msg->pkey;
1393                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1394                 alt_path->mtu_selector = IB_SA_EQ;
1395                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1396                 alt_path->rate_selector = IB_SA_EQ;
1397                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1398                 alt_path->packet_life_time_selector = IB_SA_EQ;
1399                 alt_path->packet_life_time =
1400                         cm_req_get_alt_local_ack_timeout(req_msg);
1401                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1402                 alt_path->service_id = req_msg->service_id;
1403         }
1404 }
1405
1406 static u16 cm_get_bth_pkey(struct cm_work *work)
1407 {
1408         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1409         u8 port_num = work->port->port_num;
1410         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1411         u16 pkey;
1412         int ret;
1413
1414         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1415         if (ret) {
1416                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1417                                      port_num, pkey_index, ret);
1418                 return 0;
1419         }
1420
1421         return pkey;
1422 }
1423
1424 static void cm_format_req_event(struct cm_work *work,
1425                                 struct cm_id_private *cm_id_priv,
1426                                 struct ib_cm_id *listen_id)
1427 {
1428         struct cm_req_msg *req_msg;
1429         struct ib_cm_req_event_param *param;
1430
1431         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1432         param = &work->cm_event.param.req_rcvd;
1433         param->listen_id = listen_id;
1434         param->bth_pkey = cm_get_bth_pkey(work);
1435         param->port = cm_id_priv->av.port->port_num;
1436         param->primary_path = &work->path[0];
1437         if (req_msg->alt_local_lid)
1438                 param->alternate_path = &work->path[1];
1439         else
1440                 param->alternate_path = NULL;
1441         param->remote_ca_guid = req_msg->local_ca_guid;
1442         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1443         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1444         param->qp_type = cm_req_get_qp_type(req_msg);
1445         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1446         param->responder_resources = cm_req_get_init_depth(req_msg);
1447         param->initiator_depth = cm_req_get_resp_res(req_msg);
1448         param->local_cm_response_timeout =
1449                                         cm_req_get_remote_resp_timeout(req_msg);
1450         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1451         param->remote_cm_response_timeout =
1452                                         cm_req_get_local_resp_timeout(req_msg);
1453         param->retry_count = cm_req_get_retry_count(req_msg);
1454         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1455         param->srq = cm_req_get_srq(req_msg);
1456         work->cm_event.private_data = &req_msg->private_data;
1457 }
1458
1459 static void cm_process_work(struct cm_id_private *cm_id_priv,
1460                             struct cm_work *work)
1461 {
1462         int ret;
1463
1464         /* We will typically only have the current event to report. */
1465         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1466         cm_free_work(work);
1467
1468         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1469                 spin_lock_irq(&cm_id_priv->lock);
1470                 work = cm_dequeue_work(cm_id_priv);
1471                 spin_unlock_irq(&cm_id_priv->lock);
1472                 BUG_ON(!work);
1473                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1474                                                 &work->cm_event);
1475                 cm_free_work(work);
1476         }
1477         cm_deref_id(cm_id_priv);
1478         if (ret)
1479                 cm_destroy_id(&cm_id_priv->id, ret);
1480 }
1481
1482 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1483                           struct cm_id_private *cm_id_priv,
1484                           enum cm_msg_response msg_mraed, u8 service_timeout,
1485                           const void *private_data, u8 private_data_len)
1486 {
1487         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1488         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1489         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1490         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1491         cm_mra_set_service_timeout(mra_msg, service_timeout);
1492
1493         if (private_data && private_data_len)
1494                 memcpy(mra_msg->private_data, private_data, private_data_len);
1495 }
1496
1497 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1498                           struct cm_id_private *cm_id_priv,
1499                           enum ib_cm_rej_reason reason,
1500                           void *ari,
1501                           u8 ari_length,
1502                           const void *private_data,
1503                           u8 private_data_len)
1504 {
1505         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1506         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1507
1508         switch(cm_id_priv->id.state) {
1509         case IB_CM_REQ_RCVD:
1510                 rej_msg->local_comm_id = 0;
1511                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1512                 break;
1513         case IB_CM_MRA_REQ_SENT:
1514                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1515                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1516                 break;
1517         case IB_CM_REP_RCVD:
1518         case IB_CM_MRA_REP_SENT:
1519                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1520                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1521                 break;
1522         default:
1523                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1524                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1525                 break;
1526         }
1527
1528         rej_msg->reason = cpu_to_be16(reason);
1529         if (ari && ari_length) {
1530                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1531                 memcpy(rej_msg->ari, ari, ari_length);
1532         }
1533
1534         if (private_data && private_data_len)
1535                 memcpy(rej_msg->private_data, private_data, private_data_len);
1536 }
1537
1538 static void cm_dup_req_handler(struct cm_work *work,
1539                                struct cm_id_private *cm_id_priv)
1540 {
1541         struct ib_mad_send_buf *msg = NULL;
1542         int ret;
1543
1544         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1545                         counter[CM_REQ_COUNTER]);
1546
1547         /* Quick state check to discard duplicate REQs. */
1548         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1549                 return;
1550
1551         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1552         if (ret)
1553                 return;
1554
1555         spin_lock_irq(&cm_id_priv->lock);
1556         switch (cm_id_priv->id.state) {
1557         case IB_CM_MRA_REQ_SENT:
1558                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1559                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1560                               cm_id_priv->private_data,
1561                               cm_id_priv->private_data_len);
1562                 break;
1563         case IB_CM_TIMEWAIT:
1564                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1565                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1566                 break;
1567         default:
1568                 goto unlock;
1569         }
1570         spin_unlock_irq(&cm_id_priv->lock);
1571
1572         ret = ib_post_send_mad(msg, NULL);
1573         if (ret)
1574                 goto free;
1575         return;
1576
1577 unlock: spin_unlock_irq(&cm_id_priv->lock);
1578 free:   cm_free_msg(msg);
1579 }
1580
1581 static struct cm_id_private * cm_match_req(struct cm_work *work,
1582                                            struct cm_id_private *cm_id_priv)
1583 {
1584         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1585         struct cm_timewait_info *timewait_info;
1586         struct cm_req_msg *req_msg;
1587
1588         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1589
1590         /* Check for possible duplicate REQ. */
1591         spin_lock_irq(&cm.lock);
1592         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1593         if (timewait_info) {
1594                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1595                                            timewait_info->work.remote_id);
1596                 spin_unlock_irq(&cm.lock);
1597                 if (cur_cm_id_priv) {
1598                         cm_dup_req_handler(work, cur_cm_id_priv);
1599                         cm_deref_id(cur_cm_id_priv);
1600                 }
1601                 return NULL;
1602         }
1603
1604         /* Check for stale connections. */
1605         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1606         if (timewait_info) {
1607                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1608                 spin_unlock_irq(&cm.lock);
1609                 cm_issue_rej(work->port, work->mad_recv_wc,
1610                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1611                              NULL, 0);
1612                 return NULL;
1613         }
1614
1615         /* Find matching listen request. */
1616         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1617                                            req_msg->service_id);
1618         if (!listen_cm_id_priv) {
1619                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1620                 spin_unlock_irq(&cm.lock);
1621                 cm_issue_rej(work->port, work->mad_recv_wc,
1622                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1623                              NULL, 0);
1624                 goto out;
1625         }
1626         atomic_inc(&listen_cm_id_priv->refcount);
1627         atomic_inc(&cm_id_priv->refcount);
1628         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1629         atomic_inc(&cm_id_priv->work_count);
1630         spin_unlock_irq(&cm.lock);
1631 out:
1632         return listen_cm_id_priv;
1633 }
1634
1635 /*
1636  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1637  * we need to override the LID/SL data in the REQ with the LID information
1638  * in the work completion.
1639  */
1640 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1641 {
1642         if (!cm_req_get_primary_subnet_local(req_msg)) {
1643                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1644                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1645                         cm_req_set_primary_sl(req_msg, wc->sl);
1646                 }
1647
1648                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1649                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1650         }
1651
1652         if (!cm_req_get_alt_subnet_local(req_msg)) {
1653                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1654                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1655                         cm_req_set_alt_sl(req_msg, wc->sl);
1656                 }
1657
1658                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1659                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1660         }
1661 }
1662
1663 static int cm_req_handler(struct cm_work *work)
1664 {
1665         struct ib_cm_id *cm_id;
1666         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1667         struct cm_req_msg *req_msg;
1668         union ib_gid gid;
1669         struct ib_gid_attr gid_attr;
1670         int ret;
1671
1672         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1673
1674         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1675         if (IS_ERR(cm_id))
1676                 return PTR_ERR(cm_id);
1677
1678         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1679         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1680         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1681                                 work->mad_recv_wc->recv_buf.grh,
1682                                 &cm_id_priv->av);
1683         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1684                                                             id.local_id);
1685         if (IS_ERR(cm_id_priv->timewait_info)) {
1686                 ret = PTR_ERR(cm_id_priv->timewait_info);
1687                 cm_id_priv->timewait_info = NULL;
1688                 goto destroy;
1689         }
1690         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1691         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1692         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1693
1694         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1695         if (!listen_cm_id_priv) {
1696                 ret = -EINVAL;
1697                 kfree(cm_id_priv->timewait_info);
1698                 goto destroy;
1699         }
1700
1701         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1702         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1703         cm_id_priv->id.service_id = req_msg->service_id;
1704         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1705
1706         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1707         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1708
1709         memcpy(work->path[0].dmac, cm_id_priv->av.ah_attr.dmac, ETH_ALEN);
1710         work->path[0].hop_limit = cm_id_priv->av.ah_attr.grh.hop_limit;
1711         ret = ib_get_cached_gid(work->port->cm_dev->ib_device,
1712                                 work->port->port_num,
1713                                 cm_id_priv->av.ah_attr.grh.sgid_index,
1714                                 &gid, &gid_attr);
1715         if (!ret) {
1716                 if (gid_attr.ndev) {
1717                         work->path[0].ifindex = gid_attr.ndev->ifindex;
1718                         work->path[0].net = dev_net(gid_attr.ndev);
1719                         dev_put(gid_attr.ndev);
1720                 }
1721                 work->path[0].gid_type = gid_attr.gid_type;
1722                 ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av,
1723                                          cm_id_priv);
1724         }
1725         if (ret) {
1726                 int err = ib_get_cached_gid(work->port->cm_dev->ib_device,
1727                                             work->port->port_num, 0,
1728                                             &work->path[0].sgid,
1729                                             &gid_attr);
1730                 if (!err && gid_attr.ndev) {
1731                         work->path[0].ifindex = gid_attr.ndev->ifindex;
1732                         work->path[0].net = dev_net(gid_attr.ndev);
1733                         dev_put(gid_attr.ndev);
1734                 }
1735                 work->path[0].gid_type = gid_attr.gid_type;
1736                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1737                                &work->path[0].sgid, sizeof work->path[0].sgid,
1738                                NULL, 0);
1739                 goto rejected;
1740         }
1741         if (req_msg->alt_local_lid) {
1742                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av,
1743                                          cm_id_priv);
1744                 if (ret) {
1745                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1746                                        &work->path[0].sgid,
1747                                        sizeof work->path[0].sgid, NULL, 0);
1748                         goto rejected;
1749                 }
1750         }
1751         cm_id_priv->tid = req_msg->hdr.tid;
1752         cm_id_priv->timeout_ms = cm_convert_to_ms(
1753                                         cm_req_get_local_resp_timeout(req_msg));
1754         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1755         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1756         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1757         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1758         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1759         cm_id_priv->pkey = req_msg->pkey;
1760         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1761         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1762         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1763         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1764
1765         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1766         cm_process_work(cm_id_priv, work);
1767         cm_deref_id(listen_cm_id_priv);
1768         return 0;
1769
1770 rejected:
1771         atomic_dec(&cm_id_priv->refcount);
1772         cm_deref_id(listen_cm_id_priv);
1773 destroy:
1774         ib_destroy_cm_id(cm_id);
1775         return ret;
1776 }
1777
1778 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1779                           struct cm_id_private *cm_id_priv,
1780                           struct ib_cm_rep_param *param)
1781 {
1782         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1783         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1784         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1785         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1786         rep_msg->resp_resources = param->responder_resources;
1787         cm_rep_set_target_ack_delay(rep_msg,
1788                                     cm_id_priv->av.port->cm_dev->ack_delay);
1789         cm_rep_set_failover(rep_msg, param->failover_accepted);
1790         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1791         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1792
1793         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1794                 rep_msg->initiator_depth = param->initiator_depth;
1795                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1796                 cm_rep_set_srq(rep_msg, param->srq);
1797                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1798         } else {
1799                 cm_rep_set_srq(rep_msg, 1);
1800                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1801         }
1802
1803         if (param->private_data && param->private_data_len)
1804                 memcpy(rep_msg->private_data, param->private_data,
1805                        param->private_data_len);
1806 }
1807
1808 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1809                    struct ib_cm_rep_param *param)
1810 {
1811         struct cm_id_private *cm_id_priv;
1812         struct ib_mad_send_buf *msg;
1813         struct cm_rep_msg *rep_msg;
1814         unsigned long flags;
1815         int ret;
1816
1817         if (param->private_data &&
1818             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1819                 return -EINVAL;
1820
1821         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1822         spin_lock_irqsave(&cm_id_priv->lock, flags);
1823         if (cm_id->state != IB_CM_REQ_RCVD &&
1824             cm_id->state != IB_CM_MRA_REQ_SENT) {
1825                 ret = -EINVAL;
1826                 goto out;
1827         }
1828
1829         ret = cm_alloc_msg(cm_id_priv, &msg);
1830         if (ret)
1831                 goto out;
1832
1833         rep_msg = (struct cm_rep_msg *) msg->mad;
1834         cm_format_rep(rep_msg, cm_id_priv, param);
1835         msg->timeout_ms = cm_id_priv->timeout_ms;
1836         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1837
1838         ret = ib_post_send_mad(msg, NULL);
1839         if (ret) {
1840                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1841                 cm_free_msg(msg);
1842                 return ret;
1843         }
1844
1845         cm_id->state = IB_CM_REP_SENT;
1846         cm_id_priv->msg = msg;
1847         cm_id_priv->initiator_depth = param->initiator_depth;
1848         cm_id_priv->responder_resources = param->responder_resources;
1849         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1850         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
1851
1852 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1853         return ret;
1854 }
1855 EXPORT_SYMBOL(ib_send_cm_rep);
1856
1857 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1858                           struct cm_id_private *cm_id_priv,
1859                           const void *private_data,
1860                           u8 private_data_len)
1861 {
1862         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1863         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1864         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1865
1866         if (private_data && private_data_len)
1867                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1868 }
1869
1870 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1871                    const void *private_data,
1872                    u8 private_data_len)
1873 {
1874         struct cm_id_private *cm_id_priv;
1875         struct ib_mad_send_buf *msg;
1876         unsigned long flags;
1877         void *data;
1878         int ret;
1879
1880         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1881                 return -EINVAL;
1882
1883         data = cm_copy_private_data(private_data, private_data_len);
1884         if (IS_ERR(data))
1885                 return PTR_ERR(data);
1886
1887         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1888         spin_lock_irqsave(&cm_id_priv->lock, flags);
1889         if (cm_id->state != IB_CM_REP_RCVD &&
1890             cm_id->state != IB_CM_MRA_REP_SENT) {
1891                 ret = -EINVAL;
1892                 goto error;
1893         }
1894
1895         ret = cm_alloc_msg(cm_id_priv, &msg);
1896         if (ret)
1897                 goto error;
1898
1899         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1900                       private_data, private_data_len);
1901
1902         ret = ib_post_send_mad(msg, NULL);
1903         if (ret) {
1904                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1905                 cm_free_msg(msg);
1906                 kfree(data);
1907                 return ret;
1908         }
1909
1910         cm_id->state = IB_CM_ESTABLISHED;
1911         cm_set_private_data(cm_id_priv, data, private_data_len);
1912         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1913         return 0;
1914
1915 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1916         kfree(data);
1917         return ret;
1918 }
1919 EXPORT_SYMBOL(ib_send_cm_rtu);
1920
1921 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
1922 {
1923         struct cm_rep_msg *rep_msg;
1924         struct ib_cm_rep_event_param *param;
1925
1926         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1927         param = &work->cm_event.param.rep_rcvd;
1928         param->remote_ca_guid = rep_msg->local_ca_guid;
1929         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1930         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
1931         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1932         param->responder_resources = rep_msg->initiator_depth;
1933         param->initiator_depth = rep_msg->resp_resources;
1934         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1935         param->failover_accepted = cm_rep_get_failover(rep_msg);
1936         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1937         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1938         param->srq = cm_rep_get_srq(rep_msg);
1939         work->cm_event.private_data = &rep_msg->private_data;
1940 }
1941
1942 static void cm_dup_rep_handler(struct cm_work *work)
1943 {
1944         struct cm_id_private *cm_id_priv;
1945         struct cm_rep_msg *rep_msg;
1946         struct ib_mad_send_buf *msg = NULL;
1947         int ret;
1948
1949         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1950         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1951                                    rep_msg->local_comm_id);
1952         if (!cm_id_priv)
1953                 return;
1954
1955         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1956                         counter[CM_REP_COUNTER]);
1957         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1958         if (ret)
1959                 goto deref;
1960
1961         spin_lock_irq(&cm_id_priv->lock);
1962         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1963                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1964                               cm_id_priv->private_data,
1965                               cm_id_priv->private_data_len);
1966         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1967                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1968                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1969                               cm_id_priv->private_data,
1970                               cm_id_priv->private_data_len);
1971         else
1972                 goto unlock;
1973         spin_unlock_irq(&cm_id_priv->lock);
1974
1975         ret = ib_post_send_mad(msg, NULL);
1976         if (ret)
1977                 goto free;
1978         goto deref;
1979
1980 unlock: spin_unlock_irq(&cm_id_priv->lock);
1981 free:   cm_free_msg(msg);
1982 deref:  cm_deref_id(cm_id_priv);
1983 }
1984
1985 static int cm_rep_handler(struct cm_work *work)
1986 {
1987         struct cm_id_private *cm_id_priv;
1988         struct cm_rep_msg *rep_msg;
1989         int ret;
1990
1991         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1992         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1993         if (!cm_id_priv) {
1994                 cm_dup_rep_handler(work);
1995                 return -EINVAL;
1996         }
1997
1998         cm_format_rep_event(work, cm_id_priv->qp_type);
1999
2000         spin_lock_irq(&cm_id_priv->lock);
2001         switch (cm_id_priv->id.state) {
2002         case IB_CM_REQ_SENT:
2003         case IB_CM_MRA_REQ_RCVD:
2004                 break;
2005         default:
2006                 spin_unlock_irq(&cm_id_priv->lock);
2007                 ret = -EINVAL;
2008                 goto error;
2009         }
2010
2011         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
2012         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
2013         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2014
2015         spin_lock(&cm.lock);
2016         /* Check for duplicate REP. */
2017         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2018                 spin_unlock(&cm.lock);
2019                 spin_unlock_irq(&cm_id_priv->lock);
2020                 ret = -EINVAL;
2021                 goto error;
2022         }
2023         /* Check for a stale connection. */
2024         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
2025                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
2026                          &cm.remote_id_table);
2027                 cm_id_priv->timewait_info->inserted_remote_id = 0;
2028                 spin_unlock(&cm.lock);
2029                 spin_unlock_irq(&cm_id_priv->lock);
2030                 cm_issue_rej(work->port, work->mad_recv_wc,
2031                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2032                              NULL, 0);
2033                 ret = -EINVAL;
2034                 goto error;
2035         }
2036         spin_unlock(&cm.lock);
2037
2038         cm_id_priv->id.state = IB_CM_REP_RCVD;
2039         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
2040         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2041         cm_id_priv->initiator_depth = rep_msg->resp_resources;
2042         cm_id_priv->responder_resources = rep_msg->initiator_depth;
2043         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
2044         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2045         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2046         cm_id_priv->av.timeout =
2047                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2048                                        cm_id_priv->av.timeout - 1);
2049         cm_id_priv->alt_av.timeout =
2050                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2051                                        cm_id_priv->alt_av.timeout - 1);
2052
2053         /* todo: handle peer_to_peer */
2054
2055         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2056         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2057         if (!ret)
2058                 list_add_tail(&work->list, &cm_id_priv->work_list);
2059         spin_unlock_irq(&cm_id_priv->lock);
2060
2061         if (ret)
2062                 cm_process_work(cm_id_priv, work);
2063         else
2064                 cm_deref_id(cm_id_priv);
2065         return 0;
2066
2067 error:
2068         cm_deref_id(cm_id_priv);
2069         return ret;
2070 }
2071
2072 static int cm_establish_handler(struct cm_work *work)
2073 {
2074         struct cm_id_private *cm_id_priv;
2075         int ret;
2076
2077         /* See comment in cm_establish about lookup. */
2078         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2079         if (!cm_id_priv)
2080                 return -EINVAL;
2081
2082         spin_lock_irq(&cm_id_priv->lock);
2083         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2084                 spin_unlock_irq(&cm_id_priv->lock);
2085                 goto out;
2086         }
2087
2088         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2089         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2090         if (!ret)
2091                 list_add_tail(&work->list, &cm_id_priv->work_list);
2092         spin_unlock_irq(&cm_id_priv->lock);
2093
2094         if (ret)
2095                 cm_process_work(cm_id_priv, work);
2096         else
2097                 cm_deref_id(cm_id_priv);
2098         return 0;
2099 out:
2100         cm_deref_id(cm_id_priv);
2101         return -EINVAL;
2102 }
2103
2104 static int cm_rtu_handler(struct cm_work *work)
2105 {
2106         struct cm_id_private *cm_id_priv;
2107         struct cm_rtu_msg *rtu_msg;
2108         int ret;
2109
2110         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2111         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
2112                                    rtu_msg->local_comm_id);
2113         if (!cm_id_priv)
2114                 return -EINVAL;
2115
2116         work->cm_event.private_data = &rtu_msg->private_data;
2117
2118         spin_lock_irq(&cm_id_priv->lock);
2119         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2120             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2121                 spin_unlock_irq(&cm_id_priv->lock);
2122                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2123                                 counter[CM_RTU_COUNTER]);
2124                 goto out;
2125         }
2126         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2127
2128         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2129         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2130         if (!ret)
2131                 list_add_tail(&work->list, &cm_id_priv->work_list);
2132         spin_unlock_irq(&cm_id_priv->lock);
2133
2134         if (ret)
2135                 cm_process_work(cm_id_priv, work);
2136         else
2137                 cm_deref_id(cm_id_priv);
2138         return 0;
2139 out:
2140         cm_deref_id(cm_id_priv);
2141         return -EINVAL;
2142 }
2143
2144 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2145                           struct cm_id_private *cm_id_priv,
2146                           const void *private_data,
2147                           u8 private_data_len)
2148 {
2149         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2150                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
2151         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
2152         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
2153         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2154
2155         if (private_data && private_data_len)
2156                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2157 }
2158
2159 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2160                     const void *private_data,
2161                     u8 private_data_len)
2162 {
2163         struct cm_id_private *cm_id_priv;
2164         struct ib_mad_send_buf *msg;
2165         unsigned long flags;
2166         int ret;
2167
2168         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2169                 return -EINVAL;
2170
2171         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2172         spin_lock_irqsave(&cm_id_priv->lock, flags);
2173         if (cm_id->state != IB_CM_ESTABLISHED) {
2174                 ret = -EINVAL;
2175                 goto out;
2176         }
2177
2178         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2179             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2180                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2181
2182         ret = cm_alloc_msg(cm_id_priv, &msg);
2183         if (ret) {
2184                 cm_enter_timewait(cm_id_priv);
2185                 goto out;
2186         }
2187
2188         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2189                        private_data, private_data_len);
2190         msg->timeout_ms = cm_id_priv->timeout_ms;
2191         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2192
2193         ret = ib_post_send_mad(msg, NULL);
2194         if (ret) {
2195                 cm_enter_timewait(cm_id_priv);
2196                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2197                 cm_free_msg(msg);
2198                 return ret;
2199         }
2200
2201         cm_id->state = IB_CM_DREQ_SENT;
2202         cm_id_priv->msg = msg;
2203 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2204         return ret;
2205 }
2206 EXPORT_SYMBOL(ib_send_cm_dreq);
2207
2208 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2209                           struct cm_id_private *cm_id_priv,
2210                           const void *private_data,
2211                           u8 private_data_len)
2212 {
2213         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2214         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2215         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2216
2217         if (private_data && private_data_len)
2218                 memcpy(drep_msg->private_data, private_data, private_data_len);
2219 }
2220
2221 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2222                     const void *private_data,
2223                     u8 private_data_len)
2224 {
2225         struct cm_id_private *cm_id_priv;
2226         struct ib_mad_send_buf *msg;
2227         unsigned long flags;
2228         void *data;
2229         int ret;
2230
2231         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2232                 return -EINVAL;
2233
2234         data = cm_copy_private_data(private_data, private_data_len);
2235         if (IS_ERR(data))
2236                 return PTR_ERR(data);
2237
2238         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2239         spin_lock_irqsave(&cm_id_priv->lock, flags);
2240         if (cm_id->state != IB_CM_DREQ_RCVD) {
2241                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2242                 kfree(data);
2243                 return -EINVAL;
2244         }
2245
2246         cm_set_private_data(cm_id_priv, data, private_data_len);
2247         cm_enter_timewait(cm_id_priv);
2248
2249         ret = cm_alloc_msg(cm_id_priv, &msg);
2250         if (ret)
2251                 goto out;
2252
2253         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2254                        private_data, private_data_len);
2255
2256         ret = ib_post_send_mad(msg, NULL);
2257         if (ret) {
2258                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2259                 cm_free_msg(msg);
2260                 return ret;
2261         }
2262
2263 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2264         return ret;
2265 }
2266 EXPORT_SYMBOL(ib_send_cm_drep);
2267
2268 static int cm_issue_drep(struct cm_port *port,
2269                          struct ib_mad_recv_wc *mad_recv_wc)
2270 {
2271         struct ib_mad_send_buf *msg = NULL;
2272         struct cm_dreq_msg *dreq_msg;
2273         struct cm_drep_msg *drep_msg;
2274         int ret;
2275
2276         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2277         if (ret)
2278                 return ret;
2279
2280         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2281         drep_msg = (struct cm_drep_msg *) msg->mad;
2282
2283         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2284         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2285         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2286
2287         ret = ib_post_send_mad(msg, NULL);
2288         if (ret)
2289                 cm_free_msg(msg);
2290
2291         return ret;
2292 }
2293
2294 static int cm_dreq_handler(struct cm_work *work)
2295 {
2296         struct cm_id_private *cm_id_priv;
2297         struct cm_dreq_msg *dreq_msg;
2298         struct ib_mad_send_buf *msg = NULL;
2299         int ret;
2300
2301         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2302         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2303                                    dreq_msg->local_comm_id);
2304         if (!cm_id_priv) {
2305                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2306                                 counter[CM_DREQ_COUNTER]);
2307                 cm_issue_drep(work->port, work->mad_recv_wc);
2308                 return -EINVAL;
2309         }
2310
2311         work->cm_event.private_data = &dreq_msg->private_data;
2312
2313         spin_lock_irq(&cm_id_priv->lock);
2314         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2315                 goto unlock;
2316
2317         switch (cm_id_priv->id.state) {
2318         case IB_CM_REP_SENT:
2319         case IB_CM_DREQ_SENT:
2320                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2321                 break;
2322         case IB_CM_ESTABLISHED:
2323                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2324                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2325                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2326                 break;
2327         case IB_CM_MRA_REP_RCVD:
2328                 break;
2329         case IB_CM_TIMEWAIT:
2330                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2331                                 counter[CM_DREQ_COUNTER]);
2332                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2333                         goto unlock;
2334
2335                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2336                                cm_id_priv->private_data,
2337                                cm_id_priv->private_data_len);
2338                 spin_unlock_irq(&cm_id_priv->lock);
2339
2340                 if (ib_post_send_mad(msg, NULL))
2341                         cm_free_msg(msg);
2342                 goto deref;
2343         case IB_CM_DREQ_RCVD:
2344                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2345                                 counter[CM_DREQ_COUNTER]);
2346                 goto unlock;
2347         default:
2348                 goto unlock;
2349         }
2350         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2351         cm_id_priv->tid = dreq_msg->hdr.tid;
2352         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2353         if (!ret)
2354                 list_add_tail(&work->list, &cm_id_priv->work_list);
2355         spin_unlock_irq(&cm_id_priv->lock);
2356
2357         if (ret)
2358                 cm_process_work(cm_id_priv, work);
2359         else
2360                 cm_deref_id(cm_id_priv);
2361         return 0;
2362
2363 unlock: spin_unlock_irq(&cm_id_priv->lock);
2364 deref:  cm_deref_id(cm_id_priv);
2365         return -EINVAL;
2366 }
2367
2368 static int cm_drep_handler(struct cm_work *work)
2369 {
2370         struct cm_id_private *cm_id_priv;
2371         struct cm_drep_msg *drep_msg;
2372         int ret;
2373
2374         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2375         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2376                                    drep_msg->local_comm_id);
2377         if (!cm_id_priv)
2378                 return -EINVAL;
2379
2380         work->cm_event.private_data = &drep_msg->private_data;
2381
2382         spin_lock_irq(&cm_id_priv->lock);
2383         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2384             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2385                 spin_unlock_irq(&cm_id_priv->lock);
2386                 goto out;
2387         }
2388         cm_enter_timewait(cm_id_priv);
2389
2390         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2391         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2392         if (!ret)
2393                 list_add_tail(&work->list, &cm_id_priv->work_list);
2394         spin_unlock_irq(&cm_id_priv->lock);
2395
2396         if (ret)
2397                 cm_process_work(cm_id_priv, work);
2398         else
2399                 cm_deref_id(cm_id_priv);
2400         return 0;
2401 out:
2402         cm_deref_id(cm_id_priv);
2403         return -EINVAL;
2404 }
2405
2406 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2407                    enum ib_cm_rej_reason reason,
2408                    void *ari,
2409                    u8 ari_length,
2410                    const void *private_data,
2411                    u8 private_data_len)
2412 {
2413         struct cm_id_private *cm_id_priv;
2414         struct ib_mad_send_buf *msg;
2415         unsigned long flags;
2416         int ret;
2417
2418         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2419             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2420                 return -EINVAL;
2421
2422         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2423
2424         spin_lock_irqsave(&cm_id_priv->lock, flags);
2425         switch (cm_id->state) {
2426         case IB_CM_REQ_SENT:
2427         case IB_CM_MRA_REQ_RCVD:
2428         case IB_CM_REQ_RCVD:
2429         case IB_CM_MRA_REQ_SENT:
2430         case IB_CM_REP_RCVD:
2431         case IB_CM_MRA_REP_SENT:
2432                 ret = cm_alloc_msg(cm_id_priv, &msg);
2433                 if (!ret)
2434                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2435                                       cm_id_priv, reason, ari, ari_length,
2436                                       private_data, private_data_len);
2437
2438                 cm_reset_to_idle(cm_id_priv);
2439                 break;
2440         case IB_CM_REP_SENT:
2441         case IB_CM_MRA_REP_RCVD:
2442                 ret = cm_alloc_msg(cm_id_priv, &msg);
2443                 if (!ret)
2444                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2445                                       cm_id_priv, reason, ari, ari_length,
2446                                       private_data, private_data_len);
2447
2448                 cm_enter_timewait(cm_id_priv);
2449                 break;
2450         default:
2451                 ret = -EINVAL;
2452                 goto out;
2453         }
2454
2455         if (ret)
2456                 goto out;
2457
2458         ret = ib_post_send_mad(msg, NULL);
2459         if (ret)
2460                 cm_free_msg(msg);
2461
2462 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2463         return ret;
2464 }
2465 EXPORT_SYMBOL(ib_send_cm_rej);
2466
2467 static void cm_format_rej_event(struct cm_work *work)
2468 {
2469         struct cm_rej_msg *rej_msg;
2470         struct ib_cm_rej_event_param *param;
2471
2472         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2473         param = &work->cm_event.param.rej_rcvd;
2474         param->ari = rej_msg->ari;
2475         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2476         param->reason = __be16_to_cpu(rej_msg->reason);
2477         work->cm_event.private_data = &rej_msg->private_data;
2478 }
2479
2480 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2481 {
2482         struct cm_timewait_info *timewait_info;
2483         struct cm_id_private *cm_id_priv;
2484         __be32 remote_id;
2485
2486         remote_id = rej_msg->local_comm_id;
2487
2488         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2489                 spin_lock_irq(&cm.lock);
2490                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2491                                                   remote_id);
2492                 if (!timewait_info) {
2493                         spin_unlock_irq(&cm.lock);
2494                         return NULL;
2495                 }
2496                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2497                                       (timewait_info->work.local_id ^
2498                                        cm.random_id_operand));
2499                 if (cm_id_priv) {
2500                         if (cm_id_priv->id.remote_id == remote_id)
2501                                 atomic_inc(&cm_id_priv->refcount);
2502                         else
2503                                 cm_id_priv = NULL;
2504                 }
2505                 spin_unlock_irq(&cm.lock);
2506         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2507                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2508         else
2509                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2510
2511         return cm_id_priv;
2512 }
2513
2514 static int cm_rej_handler(struct cm_work *work)
2515 {
2516         struct cm_id_private *cm_id_priv;
2517         struct cm_rej_msg *rej_msg;
2518         int ret;
2519
2520         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2521         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2522         if (!cm_id_priv)
2523                 return -EINVAL;
2524
2525         cm_format_rej_event(work);
2526
2527         spin_lock_irq(&cm_id_priv->lock);
2528         switch (cm_id_priv->id.state) {
2529         case IB_CM_REQ_SENT:
2530         case IB_CM_MRA_REQ_RCVD:
2531         case IB_CM_REP_SENT:
2532         case IB_CM_MRA_REP_RCVD:
2533                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2534                 /* fall through */
2535         case IB_CM_REQ_RCVD:
2536         case IB_CM_MRA_REQ_SENT:
2537                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2538                         cm_enter_timewait(cm_id_priv);
2539                 else
2540                         cm_reset_to_idle(cm_id_priv);
2541                 break;
2542         case IB_CM_DREQ_SENT:
2543                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2544                 /* fall through */
2545         case IB_CM_REP_RCVD:
2546         case IB_CM_MRA_REP_SENT:
2547                 cm_enter_timewait(cm_id_priv);
2548                 break;
2549         case IB_CM_ESTABLISHED:
2550                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2551                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2552                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2553                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2554                                               cm_id_priv->msg);
2555                         cm_enter_timewait(cm_id_priv);
2556                         break;
2557                 }
2558                 /* fall through */
2559         default:
2560                 spin_unlock_irq(&cm_id_priv->lock);
2561                 ret = -EINVAL;
2562                 goto out;
2563         }
2564
2565         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2566         if (!ret)
2567                 list_add_tail(&work->list, &cm_id_priv->work_list);
2568         spin_unlock_irq(&cm_id_priv->lock);
2569
2570         if (ret)
2571                 cm_process_work(cm_id_priv, work);
2572         else
2573                 cm_deref_id(cm_id_priv);
2574         return 0;
2575 out:
2576         cm_deref_id(cm_id_priv);
2577         return -EINVAL;
2578 }
2579
2580 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2581                    u8 service_timeout,
2582                    const void *private_data,
2583                    u8 private_data_len)
2584 {
2585         struct cm_id_private *cm_id_priv;
2586         struct ib_mad_send_buf *msg;
2587         enum ib_cm_state cm_state;
2588         enum ib_cm_lap_state lap_state;
2589         enum cm_msg_response msg_response;
2590         void *data;
2591         unsigned long flags;
2592         int ret;
2593
2594         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2595                 return -EINVAL;
2596
2597         data = cm_copy_private_data(private_data, private_data_len);
2598         if (IS_ERR(data))
2599                 return PTR_ERR(data);
2600
2601         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2602
2603         spin_lock_irqsave(&cm_id_priv->lock, flags);
2604         switch(cm_id_priv->id.state) {
2605         case IB_CM_REQ_RCVD:
2606                 cm_state = IB_CM_MRA_REQ_SENT;
2607                 lap_state = cm_id->lap_state;
2608                 msg_response = CM_MSG_RESPONSE_REQ;
2609                 break;
2610         case IB_CM_REP_RCVD:
2611                 cm_state = IB_CM_MRA_REP_SENT;
2612                 lap_state = cm_id->lap_state;
2613                 msg_response = CM_MSG_RESPONSE_REP;
2614                 break;
2615         case IB_CM_ESTABLISHED:
2616                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2617                         cm_state = cm_id->state;
2618                         lap_state = IB_CM_MRA_LAP_SENT;
2619                         msg_response = CM_MSG_RESPONSE_OTHER;
2620                         break;
2621                 }
2622         default:
2623                 ret = -EINVAL;
2624                 goto error1;
2625         }
2626
2627         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2628                 ret = cm_alloc_msg(cm_id_priv, &msg);
2629                 if (ret)
2630                         goto error1;
2631
2632                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2633                               msg_response, service_timeout,
2634                               private_data, private_data_len);
2635                 ret = ib_post_send_mad(msg, NULL);
2636                 if (ret)
2637                         goto error2;
2638         }
2639
2640         cm_id->state = cm_state;
2641         cm_id->lap_state = lap_state;
2642         cm_id_priv->service_timeout = service_timeout;
2643         cm_set_private_data(cm_id_priv, data, private_data_len);
2644         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2645         return 0;
2646
2647 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2648         kfree(data);
2649         return ret;
2650
2651 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2652         kfree(data);
2653         cm_free_msg(msg);
2654         return ret;
2655 }
2656 EXPORT_SYMBOL(ib_send_cm_mra);
2657
2658 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2659 {
2660         switch (cm_mra_get_msg_mraed(mra_msg)) {
2661         case CM_MSG_RESPONSE_REQ:
2662                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2663         case CM_MSG_RESPONSE_REP:
2664         case CM_MSG_RESPONSE_OTHER:
2665                 return cm_acquire_id(mra_msg->remote_comm_id,
2666                                      mra_msg->local_comm_id);
2667         default:
2668                 return NULL;
2669         }
2670 }
2671
2672 static int cm_mra_handler(struct cm_work *work)
2673 {
2674         struct cm_id_private *cm_id_priv;
2675         struct cm_mra_msg *mra_msg;
2676         int timeout, ret;
2677
2678         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2679         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2680         if (!cm_id_priv)
2681                 return -EINVAL;
2682
2683         work->cm_event.private_data = &mra_msg->private_data;
2684         work->cm_event.param.mra_rcvd.service_timeout =
2685                                         cm_mra_get_service_timeout(mra_msg);
2686         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2687                   cm_convert_to_ms(cm_id_priv->av.timeout);
2688
2689         spin_lock_irq(&cm_id_priv->lock);
2690         switch (cm_id_priv->id.state) {
2691         case IB_CM_REQ_SENT:
2692                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2693                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2694                                   cm_id_priv->msg, timeout))
2695                         goto out;
2696                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2697                 break;
2698         case IB_CM_REP_SENT:
2699                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2700                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2701                                   cm_id_priv->msg, timeout))
2702                         goto out;
2703                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2704                 break;
2705         case IB_CM_ESTABLISHED:
2706                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2707                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2708                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2709                                   cm_id_priv->msg, timeout)) {
2710                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2711                                 atomic_long_inc(&work->port->
2712                                                 counter_group[CM_RECV_DUPLICATES].
2713                                                 counter[CM_MRA_COUNTER]);
2714                         goto out;
2715                 }
2716                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2717                 break;
2718         case IB_CM_MRA_REQ_RCVD:
2719         case IB_CM_MRA_REP_RCVD:
2720                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2721                                 counter[CM_MRA_COUNTER]);
2722                 /* fall through */
2723         default:
2724                 goto out;
2725         }
2726
2727         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2728                                       cm_id_priv->id.state;
2729         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2730         if (!ret)
2731                 list_add_tail(&work->list, &cm_id_priv->work_list);
2732         spin_unlock_irq(&cm_id_priv->lock);
2733
2734         if (ret)
2735                 cm_process_work(cm_id_priv, work);
2736         else
2737                 cm_deref_id(cm_id_priv);
2738         return 0;
2739 out:
2740         spin_unlock_irq(&cm_id_priv->lock);
2741         cm_deref_id(cm_id_priv);
2742         return -EINVAL;
2743 }
2744
2745 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2746                           struct cm_id_private *cm_id_priv,
2747                           struct ib_sa_path_rec *alternate_path,
2748                           const void *private_data,
2749                           u8 private_data_len)
2750 {
2751         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2752                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2753         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2754         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2755         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2756         /* todo: need remote CM response timeout */
2757         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2758         lap_msg->alt_local_lid = alternate_path->slid;
2759         lap_msg->alt_remote_lid = alternate_path->dlid;
2760         lap_msg->alt_local_gid = alternate_path->sgid;
2761         lap_msg->alt_remote_gid = alternate_path->dgid;
2762         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2763         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2764         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2765         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2766         cm_lap_set_sl(lap_msg, alternate_path->sl);
2767         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2768         cm_lap_set_local_ack_timeout(lap_msg,
2769                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2770                                alternate_path->packet_life_time));
2771
2772         if (private_data && private_data_len)
2773                 memcpy(lap_msg->private_data, private_data, private_data_len);
2774 }
2775
2776 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2777                    struct ib_sa_path_rec *alternate_path,
2778                    const void *private_data,
2779                    u8 private_data_len)
2780 {
2781         struct cm_id_private *cm_id_priv;
2782         struct ib_mad_send_buf *msg;
2783         unsigned long flags;
2784         int ret;
2785
2786         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2787                 return -EINVAL;
2788
2789         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2790         spin_lock_irqsave(&cm_id_priv->lock, flags);
2791         if (cm_id->state != IB_CM_ESTABLISHED ||
2792             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2793              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2794                 ret = -EINVAL;
2795                 goto out;
2796         }
2797
2798         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av,
2799                                  cm_id_priv);
2800         if (ret)
2801                 goto out;
2802         cm_id_priv->alt_av.timeout =
2803                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2804                                        cm_id_priv->alt_av.timeout - 1);
2805
2806         ret = cm_alloc_msg(cm_id_priv, &msg);
2807         if (ret)
2808                 goto out;
2809
2810         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2811                       alternate_path, private_data, private_data_len);
2812         msg->timeout_ms = cm_id_priv->timeout_ms;
2813         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2814
2815         ret = ib_post_send_mad(msg, NULL);
2816         if (ret) {
2817                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2818                 cm_free_msg(msg);
2819                 return ret;
2820         }
2821
2822         cm_id->lap_state = IB_CM_LAP_SENT;
2823         cm_id_priv->msg = msg;
2824
2825 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2826         return ret;
2827 }
2828 EXPORT_SYMBOL(ib_send_cm_lap);
2829
2830 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2831                                     struct ib_sa_path_rec *path,
2832                                     struct cm_lap_msg *lap_msg)
2833 {
2834         memset(path, 0, sizeof *path);
2835         path->dgid = lap_msg->alt_local_gid;
2836         path->sgid = lap_msg->alt_remote_gid;
2837         path->dlid = lap_msg->alt_local_lid;
2838         path->slid = lap_msg->alt_remote_lid;
2839         path->flow_label = cm_lap_get_flow_label(lap_msg);
2840         path->hop_limit = lap_msg->alt_hop_limit;
2841         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2842         path->reversible = 1;
2843         path->pkey = cm_id_priv->pkey;
2844         path->sl = cm_lap_get_sl(lap_msg);
2845         path->mtu_selector = IB_SA_EQ;
2846         path->mtu = cm_id_priv->path_mtu;
2847         path->rate_selector = IB_SA_EQ;
2848         path->rate = cm_lap_get_packet_rate(lap_msg);
2849         path->packet_life_time_selector = IB_SA_EQ;
2850         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2851         path->packet_life_time -= (path->packet_life_time > 0);
2852 }
2853
2854 static int cm_lap_handler(struct cm_work *work)
2855 {
2856         struct cm_id_private *cm_id_priv;
2857         struct cm_lap_msg *lap_msg;
2858         struct ib_cm_lap_event_param *param;
2859         struct ib_mad_send_buf *msg = NULL;
2860         int ret;
2861
2862         /* todo: verify LAP request and send reject APR if invalid. */
2863         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2864         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2865                                    lap_msg->local_comm_id);
2866         if (!cm_id_priv)
2867                 return -EINVAL;
2868
2869         param = &work->cm_event.param.lap_rcvd;
2870         param->alternate_path = &work->path[0];
2871         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2872         work->cm_event.private_data = &lap_msg->private_data;
2873
2874         spin_lock_irq(&cm_id_priv->lock);
2875         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2876                 goto unlock;
2877
2878         switch (cm_id_priv->id.lap_state) {
2879         case IB_CM_LAP_UNINIT:
2880         case IB_CM_LAP_IDLE:
2881                 break;
2882         case IB_CM_MRA_LAP_SENT:
2883                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2884                                 counter[CM_LAP_COUNTER]);
2885                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2886                         goto unlock;
2887
2888                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2889                               CM_MSG_RESPONSE_OTHER,
2890                               cm_id_priv->service_timeout,
2891                               cm_id_priv->private_data,
2892                               cm_id_priv->private_data_len);
2893                 spin_unlock_irq(&cm_id_priv->lock);
2894
2895                 if (ib_post_send_mad(msg, NULL))
2896                         cm_free_msg(msg);
2897                 goto deref;
2898         case IB_CM_LAP_RCVD:
2899                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2900                                 counter[CM_LAP_COUNTER]);
2901                 goto unlock;
2902         default:
2903                 goto unlock;
2904         }
2905
2906         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2907         cm_id_priv->tid = lap_msg->hdr.tid;
2908         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2909                                 work->mad_recv_wc->recv_buf.grh,
2910                                 &cm_id_priv->av);
2911         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av,
2912                            cm_id_priv);
2913         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2914         if (!ret)
2915                 list_add_tail(&work->list, &cm_id_priv->work_list);
2916         spin_unlock_irq(&cm_id_priv->lock);
2917
2918         if (ret)
2919                 cm_process_work(cm_id_priv, work);
2920         else
2921                 cm_deref_id(cm_id_priv);
2922         return 0;
2923
2924 unlock: spin_unlock_irq(&cm_id_priv->lock);
2925 deref:  cm_deref_id(cm_id_priv);
2926         return -EINVAL;
2927 }
2928
2929 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2930                           struct cm_id_private *cm_id_priv,
2931                           enum ib_cm_apr_status status,
2932                           void *info,
2933                           u8 info_length,
2934                           const void *private_data,
2935                           u8 private_data_len)
2936 {
2937         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2938         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2939         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2940         apr_msg->ap_status = (u8) status;
2941
2942         if (info && info_length) {
2943                 apr_msg->info_length = info_length;
2944                 memcpy(apr_msg->info, info, info_length);
2945         }
2946
2947         if (private_data && private_data_len)
2948                 memcpy(apr_msg->private_data, private_data, private_data_len);
2949 }
2950
2951 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2952                    enum ib_cm_apr_status status,
2953                    void *info,
2954                    u8 info_length,
2955                    const void *private_data,
2956                    u8 private_data_len)
2957 {
2958         struct cm_id_private *cm_id_priv;
2959         struct ib_mad_send_buf *msg;
2960         unsigned long flags;
2961         int ret;
2962
2963         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2964             (info && info_length > IB_CM_APR_INFO_LENGTH))
2965                 return -EINVAL;
2966
2967         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2968         spin_lock_irqsave(&cm_id_priv->lock, flags);
2969         if (cm_id->state != IB_CM_ESTABLISHED ||
2970             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2971              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2972                 ret = -EINVAL;
2973                 goto out;
2974         }
2975
2976         ret = cm_alloc_msg(cm_id_priv, &msg);
2977         if (ret)
2978                 goto out;
2979
2980         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2981                       info, info_length, private_data, private_data_len);
2982         ret = ib_post_send_mad(msg, NULL);
2983         if (ret) {
2984                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2985                 cm_free_msg(msg);
2986                 return ret;
2987         }
2988
2989         cm_id->lap_state = IB_CM_LAP_IDLE;
2990 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2991         return ret;
2992 }
2993 EXPORT_SYMBOL(ib_send_cm_apr);
2994
2995 static int cm_apr_handler(struct cm_work *work)
2996 {
2997         struct cm_id_private *cm_id_priv;
2998         struct cm_apr_msg *apr_msg;
2999         int ret;
3000
3001         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3002         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
3003                                    apr_msg->local_comm_id);
3004         if (!cm_id_priv)
3005                 return -EINVAL; /* Unmatched reply. */
3006
3007         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
3008         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
3009         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
3010         work->cm_event.private_data = &apr_msg->private_data;
3011
3012         spin_lock_irq(&cm_id_priv->lock);
3013         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3014             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3015              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3016                 spin_unlock_irq(&cm_id_priv->lock);
3017                 goto out;
3018         }
3019         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3020         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3021         cm_id_priv->msg = NULL;
3022
3023         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3024         if (!ret)
3025                 list_add_tail(&work->list, &cm_id_priv->work_list);
3026         spin_unlock_irq(&cm_id_priv->lock);
3027
3028         if (ret)
3029                 cm_process_work(cm_id_priv, work);
3030         else
3031                 cm_deref_id(cm_id_priv);
3032         return 0;
3033 out:
3034         cm_deref_id(cm_id_priv);
3035         return -EINVAL;
3036 }
3037
3038 static int cm_timewait_handler(struct cm_work *work)
3039 {
3040         struct cm_timewait_info *timewait_info;
3041         struct cm_id_private *cm_id_priv;
3042         int ret;
3043
3044         timewait_info = (struct cm_timewait_info *)work;
3045         spin_lock_irq(&cm.lock);
3046         list_del(&timewait_info->list);
3047         spin_unlock_irq(&cm.lock);
3048
3049         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3050                                    timewait_info->work.remote_id);
3051         if (!cm_id_priv)
3052                 return -EINVAL;
3053
3054         spin_lock_irq(&cm_id_priv->lock);
3055         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3056             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3057                 spin_unlock_irq(&cm_id_priv->lock);
3058                 goto out;
3059         }
3060         cm_id_priv->id.state = IB_CM_IDLE;
3061         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3062         if (!ret)
3063                 list_add_tail(&work->list, &cm_id_priv->work_list);
3064         spin_unlock_irq(&cm_id_priv->lock);
3065
3066         if (ret)
3067                 cm_process_work(cm_id_priv, work);
3068         else
3069                 cm_deref_id(cm_id_priv);
3070         return 0;
3071 out:
3072         cm_deref_id(cm_id_priv);
3073         return -EINVAL;
3074 }
3075
3076 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3077                                struct cm_id_private *cm_id_priv,
3078                                struct ib_cm_sidr_req_param *param)
3079 {
3080         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3081                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
3082         sidr_req_msg->request_id = cm_id_priv->id.local_id;
3083         sidr_req_msg->pkey = param->path->pkey;
3084         sidr_req_msg->service_id = param->service_id;
3085
3086         if (param->private_data && param->private_data_len)
3087                 memcpy(sidr_req_msg->private_data, param->private_data,
3088                        param->private_data_len);
3089 }
3090
3091 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3092                         struct ib_cm_sidr_req_param *param)
3093 {
3094         struct cm_id_private *cm_id_priv;
3095         struct ib_mad_send_buf *msg;
3096         unsigned long flags;
3097         int ret;
3098
3099         if (!param->path || (param->private_data &&
3100              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3101                 return -EINVAL;
3102
3103         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3104         ret = cm_init_av_by_path(param->path, &cm_id_priv->av, cm_id_priv);
3105         if (ret)
3106                 goto out;
3107
3108         cm_id->service_id = param->service_id;
3109         cm_id->service_mask = ~cpu_to_be64(0);
3110         cm_id_priv->timeout_ms = param->timeout_ms;
3111         cm_id_priv->max_cm_retries = param->max_cm_retries;
3112         ret = cm_alloc_msg(cm_id_priv, &msg);
3113         if (ret)
3114                 goto out;
3115
3116         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3117                            param);
3118         msg->timeout_ms = cm_id_priv->timeout_ms;
3119         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3120
3121         spin_lock_irqsave(&cm_id_priv->lock, flags);
3122         if (cm_id->state == IB_CM_IDLE)
3123                 ret = ib_post_send_mad(msg, NULL);
3124         else
3125                 ret = -EINVAL;
3126
3127         if (ret) {
3128                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3129                 cm_free_msg(msg);
3130                 goto out;
3131         }
3132         cm_id->state = IB_CM_SIDR_REQ_SENT;
3133         cm_id_priv->msg = msg;
3134         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3135 out:
3136         return ret;
3137 }
3138 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3139
3140 static void cm_format_sidr_req_event(struct cm_work *work,
3141                                      struct ib_cm_id *listen_id)
3142 {
3143         struct cm_sidr_req_msg *sidr_req_msg;
3144         struct ib_cm_sidr_req_event_param *param;
3145
3146         sidr_req_msg = (struct cm_sidr_req_msg *)
3147                                 work->mad_recv_wc->recv_buf.mad;
3148         param = &work->cm_event.param.sidr_req_rcvd;
3149         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
3150         param->listen_id = listen_id;
3151         param->service_id = sidr_req_msg->service_id;
3152         param->bth_pkey = cm_get_bth_pkey(work);
3153         param->port = work->port->port_num;
3154         work->cm_event.private_data = &sidr_req_msg->private_data;
3155 }
3156
3157 static int cm_sidr_req_handler(struct cm_work *work)
3158 {
3159         struct ib_cm_id *cm_id;
3160         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3161         struct cm_sidr_req_msg *sidr_req_msg;
3162         struct ib_wc *wc;
3163
3164         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3165         if (IS_ERR(cm_id))
3166                 return PTR_ERR(cm_id);
3167         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3168
3169         /* Record SGID/SLID and request ID for lookup. */
3170         sidr_req_msg = (struct cm_sidr_req_msg *)
3171                                 work->mad_recv_wc->recv_buf.mad;
3172         wc = work->mad_recv_wc->wc;
3173         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3174         cm_id_priv->av.dgid.global.interface_id = 0;
3175         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3176                                 work->mad_recv_wc->recv_buf.grh,
3177                                 &cm_id_priv->av);
3178         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3179         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3180         atomic_inc(&cm_id_priv->work_count);
3181
3182         spin_lock_irq(&cm.lock);
3183         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3184         if (cur_cm_id_priv) {
3185                 spin_unlock_irq(&cm.lock);
3186                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3187                                 counter[CM_SIDR_REQ_COUNTER]);
3188                 goto out; /* Duplicate message. */
3189         }
3190         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3191         cur_cm_id_priv = cm_find_listen(cm_id->device,
3192                                         sidr_req_msg->service_id);
3193         if (!cur_cm_id_priv) {
3194                 spin_unlock_irq(&cm.lock);
3195                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3196                 goto out; /* No match. */
3197         }
3198         atomic_inc(&cur_cm_id_priv->refcount);
3199         atomic_inc(&cm_id_priv->refcount);
3200         spin_unlock_irq(&cm.lock);
3201
3202         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3203         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3204         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3205         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3206
3207         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3208         cm_process_work(cm_id_priv, work);
3209         cm_deref_id(cur_cm_id_priv);
3210         return 0;
3211 out:
3212         ib_destroy_cm_id(&cm_id_priv->id);
3213         return -EINVAL;
3214 }
3215
3216 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3217                                struct cm_id_private *cm_id_priv,
3218                                struct ib_cm_sidr_rep_param *param)
3219 {
3220         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3221                           cm_id_priv->tid);
3222         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3223         sidr_rep_msg->status = param->status;
3224         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3225         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3226         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3227
3228         if (param->info && param->info_length)
3229                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3230
3231         if (param->private_data && param->private_data_len)
3232                 memcpy(sidr_rep_msg->private_data, param->private_data,
3233                        param->private_data_len);
3234 }
3235
3236 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3237                         struct ib_cm_sidr_rep_param *param)
3238 {
3239         struct cm_id_private *cm_id_priv;
3240         struct ib_mad_send_buf *msg;
3241         unsigned long flags;
3242         int ret;
3243
3244         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3245             (param->private_data &&
3246              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3247                 return -EINVAL;
3248
3249         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3250         spin_lock_irqsave(&cm_id_priv->lock, flags);
3251         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3252                 ret = -EINVAL;
3253                 goto error;
3254         }
3255
3256         ret = cm_alloc_msg(cm_id_priv, &msg);
3257         if (ret)
3258                 goto error;
3259
3260         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3261                            param);
3262         ret = ib_post_send_mad(msg, NULL);
3263         if (ret) {
3264                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3265                 cm_free_msg(msg);
3266                 return ret;
3267         }
3268         cm_id->state = IB_CM_IDLE;
3269         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3270
3271         spin_lock_irqsave(&cm.lock, flags);
3272         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3273                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3274                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3275         }
3276         spin_unlock_irqrestore(&cm.lock, flags);
3277         return 0;
3278
3279 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3280         return ret;
3281 }
3282 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3283
3284 static void cm_format_sidr_rep_event(struct cm_work *work)
3285 {
3286         struct cm_sidr_rep_msg *sidr_rep_msg;
3287         struct ib_cm_sidr_rep_event_param *param;
3288
3289         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3290                                 work->mad_recv_wc->recv_buf.mad;
3291         param = &work->cm_event.param.sidr_rep_rcvd;
3292         param->status = sidr_rep_msg->status;
3293         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3294         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3295         param->info = &sidr_rep_msg->info;
3296         param->info_len = sidr_rep_msg->info_length;
3297         work->cm_event.private_data = &sidr_rep_msg->private_data;
3298 }
3299
3300 static int cm_sidr_rep_handler(struct cm_work *work)
3301 {
3302         struct cm_sidr_rep_msg *sidr_rep_msg;
3303         struct cm_id_private *cm_id_priv;
3304
3305         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3306                                 work->mad_recv_wc->recv_buf.mad;
3307         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3308         if (!cm_id_priv)
3309                 return -EINVAL; /* Unmatched reply. */
3310
3311         spin_lock_irq(&cm_id_priv->lock);
3312         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3313                 spin_unlock_irq(&cm_id_priv->lock);
3314                 goto out;
3315         }
3316         cm_id_priv->id.state = IB_CM_IDLE;
3317         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3318         spin_unlock_irq(&cm_id_priv->lock);
3319
3320         cm_format_sidr_rep_event(work);
3321         cm_process_work(cm_id_priv, work);
3322         return 0;
3323 out:
3324         cm_deref_id(cm_id_priv);
3325         return -EINVAL;
3326 }
3327
3328 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3329                                   enum ib_wc_status wc_status)
3330 {
3331         struct cm_id_private *cm_id_priv;
3332         struct ib_cm_event cm_event;
3333         enum ib_cm_state state;
3334         int ret;
3335
3336         memset(&cm_event, 0, sizeof cm_event);
3337         cm_id_priv = msg->context[0];
3338
3339         /* Discard old sends or ones without a response. */
3340         spin_lock_irq(&cm_id_priv->lock);
3341         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3342         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3343                 goto discard;
3344
3345         switch (state) {
3346         case IB_CM_REQ_SENT:
3347         case IB_CM_MRA_REQ_RCVD:
3348                 cm_reset_to_idle(cm_id_priv);
3349                 cm_event.event = IB_CM_REQ_ERROR;
3350                 break;
3351         case IB_CM_REP_SENT:
3352         case IB_CM_MRA_REP_RCVD:
3353                 cm_reset_to_idle(cm_id_priv);
3354                 cm_event.event = IB_CM_REP_ERROR;
3355                 break;
3356         case IB_CM_DREQ_SENT:
3357                 cm_enter_timewait(cm_id_priv);
3358                 cm_event.event = IB_CM_DREQ_ERROR;
3359                 break;
3360         case IB_CM_SIDR_REQ_SENT:
3361                 cm_id_priv->id.state = IB_CM_IDLE;
3362                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3363                 break;
3364         default:
3365                 goto discard;
3366         }
3367         spin_unlock_irq(&cm_id_priv->lock);
3368         cm_event.param.send_status = wc_status;
3369
3370         /* No other events can occur on the cm_id at this point. */
3371         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3372         cm_free_msg(msg);
3373         if (ret)
3374                 ib_destroy_cm_id(&cm_id_priv->id);
3375         return;
3376 discard:
3377         spin_unlock_irq(&cm_id_priv->lock);
3378         cm_free_msg(msg);
3379 }
3380
3381 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3382                             struct ib_mad_send_wc *mad_send_wc)
3383 {
3384         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3385         struct cm_port *port;
3386         u16 attr_index;
3387
3388         port = mad_agent->context;
3389         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3390                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3391
3392         /*
3393          * If the send was in response to a received message (context[0] is not
3394          * set to a cm_id), and is not a REJ, then it is a send that was
3395          * manually retried.
3396          */
3397         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3398                 msg->retries = 1;
3399
3400         atomic_long_add(1 + msg->retries,
3401                         &port->counter_group[CM_XMIT].counter[attr_index]);
3402         if (msg->retries)
3403                 atomic_long_add(msg->retries,
3404                                 &port->counter_group[CM_XMIT_RETRIES].
3405                                 counter[attr_index]);
3406
3407         switch (mad_send_wc->status) {
3408         case IB_WC_SUCCESS:
3409         case IB_WC_WR_FLUSH_ERR:
3410                 cm_free_msg(msg);
3411                 break;
3412         default:
3413                 if (msg->context[0] && msg->context[1])
3414                         cm_process_send_error(msg, mad_send_wc->status);
3415                 else
3416                         cm_free_msg(msg);
3417                 break;
3418         }
3419 }
3420
3421 static void cm_work_handler(struct work_struct *_work)
3422 {
3423         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3424         int ret;
3425
3426         switch (work->cm_event.event) {
3427         case IB_CM_REQ_RECEIVED:
3428                 ret = cm_req_handler(work);
3429                 break;
3430         case IB_CM_MRA_RECEIVED:
3431                 ret = cm_mra_handler(work);
3432                 break;
3433         case IB_CM_REJ_RECEIVED:
3434                 ret = cm_rej_handler(work);
3435                 break;
3436         case IB_CM_REP_RECEIVED:
3437                 ret = cm_rep_handler(work);
3438                 break;
3439         case IB_CM_RTU_RECEIVED:
3440                 ret = cm_rtu_handler(work);
3441                 break;
3442         case IB_CM_USER_ESTABLISHED:
3443                 ret = cm_establish_handler(work);
3444                 break;
3445         case IB_CM_DREQ_RECEIVED:
3446                 ret = cm_dreq_handler(work);
3447                 break;
3448         case IB_CM_DREP_RECEIVED:
3449                 ret = cm_drep_handler(work);
3450                 break;
3451         case IB_CM_SIDR_REQ_RECEIVED:
3452                 ret = cm_sidr_req_handler(work);
3453                 break;
3454         case IB_CM_SIDR_REP_RECEIVED:
3455                 ret = cm_sidr_rep_handler(work);
3456                 break;
3457         case IB_CM_LAP_RECEIVED:
3458                 ret = cm_lap_handler(work);
3459                 break;
3460         case IB_CM_APR_RECEIVED:
3461                 ret = cm_apr_handler(work);
3462                 break;
3463         case IB_CM_TIMEWAIT_EXIT:
3464                 ret = cm_timewait_handler(work);
3465                 break;
3466         default:
3467                 ret = -EINVAL;
3468                 break;
3469         }
3470         if (ret)
3471                 cm_free_work(work);
3472 }
3473
3474 static int cm_establish(struct ib_cm_id *cm_id)
3475 {
3476         struct cm_id_private *cm_id_priv;
3477         struct cm_work *work;
3478         unsigned long flags;
3479         int ret = 0;
3480         struct cm_device *cm_dev;
3481
3482         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3483         if (!cm_dev)
3484                 return -ENODEV;
3485
3486         work = kmalloc(sizeof *work, GFP_ATOMIC);
3487         if (!work)
3488                 return -ENOMEM;
3489
3490         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3491         spin_lock_irqsave(&cm_id_priv->lock, flags);
3492         switch (cm_id->state)
3493         {
3494         case IB_CM_REP_SENT:
3495         case IB_CM_MRA_REP_RCVD:
3496                 cm_id->state = IB_CM_ESTABLISHED;
3497                 break;
3498         case IB_CM_ESTABLISHED:
3499                 ret = -EISCONN;
3500                 break;
3501         default:
3502                 ret = -EINVAL;
3503                 break;
3504         }
3505         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3506
3507         if (ret) {
3508                 kfree(work);
3509                 goto out;
3510         }
3511
3512         /*
3513          * The CM worker thread may try to destroy the cm_id before it
3514          * can execute this work item.  To prevent potential deadlock,
3515          * we need to find the cm_id once we're in the context of the
3516          * worker thread, rather than holding a reference on it.
3517          */
3518         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3519         work->local_id = cm_id->local_id;
3520         work->remote_id = cm_id->remote_id;
3521         work->mad_recv_wc = NULL;
3522         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3523
3524         /* Check if the device started its remove_one */
3525         spin_lock_irqsave(&cm.lock, flags);
3526         if (!cm_dev->going_down) {
3527                 queue_delayed_work(cm.wq, &work->work, 0);
3528         } else {
3529                 kfree(work);
3530                 ret = -ENODEV;
3531         }
3532         spin_unlock_irqrestore(&cm.lock, flags);
3533
3534 out:
3535         return ret;
3536 }
3537
3538 static int cm_migrate(struct ib_cm_id *cm_id)
3539 {
3540         struct cm_id_private *cm_id_priv;
3541         struct cm_av tmp_av;
3542         unsigned long flags;
3543         int tmp_send_port_not_ready;
3544         int ret = 0;
3545
3546         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3547         spin_lock_irqsave(&cm_id_priv->lock, flags);
3548         if (cm_id->state == IB_CM_ESTABLISHED &&
3549             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3550              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3551                 cm_id->lap_state = IB_CM_LAP_IDLE;
3552                 /* Swap address vector */
3553                 tmp_av = cm_id_priv->av;
3554                 cm_id_priv->av = cm_id_priv->alt_av;
3555                 cm_id_priv->alt_av = tmp_av;
3556                 /* Swap port send ready state */
3557                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3558                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3559                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3560         } else
3561                 ret = -EINVAL;
3562         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3563
3564         return ret;
3565 }
3566
3567 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3568 {
3569         int ret;
3570
3571         switch (event) {
3572         case IB_EVENT_COMM_EST:
3573                 ret = cm_establish(cm_id);
3574                 break;
3575         case IB_EVENT_PATH_MIG:
3576                 ret = cm_migrate(cm_id);
3577                 break;
3578         default:
3579                 ret = -EINVAL;
3580         }
3581         return ret;
3582 }
3583 EXPORT_SYMBOL(ib_cm_notify);
3584
3585 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3586                             struct ib_mad_send_buf *send_buf,
3587                             struct ib_mad_recv_wc *mad_recv_wc)
3588 {
3589         struct cm_port *port = mad_agent->context;
3590         struct cm_work *work;
3591         enum ib_cm_event_type event;
3592         u16 attr_id;
3593         int paths = 0;
3594         int going_down = 0;
3595
3596         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3597         case CM_REQ_ATTR_ID:
3598                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3599                                                     alt_local_lid != 0);
3600                 event = IB_CM_REQ_RECEIVED;
3601                 break;
3602         case CM_MRA_ATTR_ID:
3603                 event = IB_CM_MRA_RECEIVED;
3604                 break;
3605         case CM_REJ_ATTR_ID:
3606                 event = IB_CM_REJ_RECEIVED;
3607                 break;
3608         case CM_REP_ATTR_ID:
3609                 event = IB_CM_REP_RECEIVED;
3610                 break;
3611         case CM_RTU_ATTR_ID:
3612                 event = IB_CM_RTU_RECEIVED;
3613                 break;
3614         case CM_DREQ_ATTR_ID:
3615                 event = IB_CM_DREQ_RECEIVED;
3616                 break;
3617         case CM_DREP_ATTR_ID:
3618                 event = IB_CM_DREP_RECEIVED;
3619                 break;
3620         case CM_SIDR_REQ_ATTR_ID:
3621                 event = IB_CM_SIDR_REQ_RECEIVED;
3622                 break;
3623         case CM_SIDR_REP_ATTR_ID:
3624                 event = IB_CM_SIDR_REP_RECEIVED;
3625                 break;
3626         case CM_LAP_ATTR_ID:
3627                 paths = 1;
3628                 event = IB_CM_LAP_RECEIVED;
3629                 break;
3630         case CM_APR_ATTR_ID:
3631                 event = IB_CM_APR_RECEIVED;
3632                 break;
3633         default:
3634                 ib_free_recv_mad(mad_recv_wc);
3635                 return;
3636         }
3637
3638         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3639         atomic_long_inc(&port->counter_group[CM_RECV].
3640                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3641
3642         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3643                        GFP_KERNEL);
3644         if (!work) {
3645                 ib_free_recv_mad(mad_recv_wc);
3646                 return;
3647         }
3648
3649         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3650         work->cm_event.event = event;
3651         work->mad_recv_wc = mad_recv_wc;
3652         work->port = port;
3653
3654         /* Check if the device started its remove_one */
3655         spin_lock_irq(&cm.lock);
3656         if (!port->cm_dev->going_down)
3657                 queue_delayed_work(cm.wq, &work->work, 0);
3658         else
3659                 going_down = 1;
3660         spin_unlock_irq(&cm.lock);
3661
3662         if (going_down) {
3663                 kfree(work);
3664                 ib_free_recv_mad(mad_recv_wc);
3665         }
3666 }
3667
3668 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3669                                 struct ib_qp_attr *qp_attr,
3670                                 int *qp_attr_mask)
3671 {
3672         unsigned long flags;
3673         int ret;
3674
3675         spin_lock_irqsave(&cm_id_priv->lock, flags);
3676         switch (cm_id_priv->id.state) {
3677         case IB_CM_REQ_SENT:
3678         case IB_CM_MRA_REQ_RCVD:
3679         case IB_CM_REQ_RCVD:
3680         case IB_CM_MRA_REQ_SENT:
3681         case IB_CM_REP_RCVD:
3682         case IB_CM_MRA_REP_SENT:
3683         case IB_CM_REP_SENT:
3684         case IB_CM_MRA_REP_RCVD:
3685         case IB_CM_ESTABLISHED:
3686                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3687                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3688                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3689                 if (cm_id_priv->responder_resources)
3690                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3691                                                     IB_ACCESS_REMOTE_ATOMIC;
3692                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3693                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3694                 ret = 0;
3695                 break;
3696         default:
3697                 ret = -EINVAL;
3698                 break;
3699         }
3700         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3701         return ret;
3702 }
3703
3704 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3705                                struct ib_qp_attr *qp_attr,
3706                                int *qp_attr_mask)
3707 {
3708         unsigned long flags;
3709         int ret;
3710
3711         spin_lock_irqsave(&cm_id_priv->lock, flags);
3712         switch (cm_id_priv->id.state) {
3713         case IB_CM_REQ_RCVD:
3714         case IB_CM_MRA_REQ_SENT:
3715         case IB_CM_REP_RCVD:
3716         case IB_CM_MRA_REP_SENT:
3717         case IB_CM_REP_SENT:
3718         case IB_CM_MRA_REP_RCVD:
3719         case IB_CM_ESTABLISHED:
3720                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3721                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3722                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3723                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3724                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3725                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3726                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3727                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3728                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3729                                          IB_QP_MIN_RNR_TIMER;
3730                         qp_attr->max_dest_rd_atomic =
3731                                         cm_id_priv->responder_resources;
3732                         qp_attr->min_rnr_timer = 0;
3733                 }
3734                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3735                         *qp_attr_mask |= IB_QP_ALT_PATH;
3736                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3737                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3738                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3739                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3740                 }
3741                 ret = 0;
3742                 break;
3743         default:
3744                 ret = -EINVAL;
3745                 break;
3746         }
3747         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3748         return ret;
3749 }
3750
3751 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3752                                struct ib_qp_attr *qp_attr,
3753                                int *qp_attr_mask)
3754 {
3755         unsigned long flags;
3756         int ret;
3757
3758         spin_lock_irqsave(&cm_id_priv->lock, flags);
3759         switch (cm_id_priv->id.state) {
3760         /* Allow transition to RTS before sending REP */
3761         case IB_CM_REQ_RCVD:
3762         case IB_CM_MRA_REQ_SENT:
3763
3764         case IB_CM_REP_RCVD:
3765         case IB_CM_MRA_REP_SENT:
3766         case IB_CM_REP_SENT:
3767         case IB_CM_MRA_REP_RCVD:
3768         case IB_CM_ESTABLISHED:
3769                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3770                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3771                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3772                         switch (cm_id_priv->qp_type) {
3773                         case IB_QPT_RC:
3774                         case IB_QPT_XRC_INI:
3775                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3776                                                  IB_QP_MAX_QP_RD_ATOMIC;
3777                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3778                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3779                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3780                                 /* fall through */
3781                         case IB_QPT_XRC_TGT:
3782                                 *qp_attr_mask |= IB_QP_TIMEOUT;
3783                                 qp_attr->timeout = cm_id_priv->av.timeout;
3784                                 break;
3785                         default:
3786                                 break;
3787                         }
3788                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3789                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3790                                 qp_attr->path_mig_state = IB_MIG_REARM;
3791                         }
3792                 } else {
3793                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3794                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3795                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3796                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3797                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3798                         qp_attr->path_mig_state = IB_MIG_REARM;
3799                 }
3800                 ret = 0;
3801                 break;
3802         default:
3803                 ret = -EINVAL;
3804                 break;
3805         }
3806         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3807         return ret;
3808 }
3809
3810 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3811                        struct ib_qp_attr *qp_attr,
3812                        int *qp_attr_mask)
3813 {
3814         struct cm_id_private *cm_id_priv;
3815         int ret;
3816
3817         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3818         switch (qp_attr->qp_state) {
3819         case IB_QPS_INIT:
3820                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3821                 break;
3822         case IB_QPS_RTR:
3823                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3824                 break;
3825         case IB_QPS_RTS:
3826                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3827                 break;
3828         default:
3829                 ret = -EINVAL;
3830                 break;
3831         }
3832         return ret;
3833 }
3834 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3835
3836 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3837                                char *buf)
3838 {
3839         struct cm_counter_group *group;
3840         struct cm_counter_attribute *cm_attr;
3841
3842         group = container_of(obj, struct cm_counter_group, obj);
3843         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3844
3845         return sprintf(buf, "%ld\n",
3846                        atomic_long_read(&group->counter[cm_attr->index]));
3847 }
3848
3849 static const struct sysfs_ops cm_counter_ops = {
3850         .show = cm_show_counter
3851 };
3852
3853 static struct kobj_type cm_counter_obj_type = {
3854         .sysfs_ops = &cm_counter_ops,
3855         .default_attrs = cm_counter_default_attrs
3856 };
3857
3858 static void cm_release_port_obj(struct kobject *obj)
3859 {
3860         struct cm_port *cm_port;
3861
3862         cm_port = container_of(obj, struct cm_port, port_obj);
3863         kfree(cm_port);
3864 }
3865
3866 static struct kobj_type cm_port_obj_type = {
3867         .release = cm_release_port_obj
3868 };
3869
3870 static char *cm_devnode(struct device *dev, umode_t *mode)
3871 {
3872         if (mode)
3873                 *mode = 0666;
3874         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3875 }
3876
3877 struct class cm_class = {
3878         .owner   = THIS_MODULE,
3879         .name    = "infiniband_cm",
3880         .devnode = cm_devnode,
3881 };
3882 EXPORT_SYMBOL(cm_class);
3883
3884 static int cm_create_port_fs(struct cm_port *port)
3885 {
3886         int i, ret;
3887
3888         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3889                                    &port->cm_dev->device->kobj,
3890                                    "%d", port->port_num);
3891         if (ret) {
3892                 kfree(port);
3893                 return ret;
3894         }
3895
3896         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3897                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3898                                            &cm_counter_obj_type,
3899                                            &port->port_obj,
3900                                            "%s", counter_group_names[i]);
3901                 if (ret)
3902                         goto error;
3903         }
3904
3905         return 0;
3906
3907 error:
3908         while (i--)
3909                 kobject_put(&port->counter_group[i].obj);
3910         kobject_put(&port->port_obj);
3911         return ret;
3912
3913 }
3914
3915 static void cm_remove_port_fs(struct cm_port *port)
3916 {
3917         int i;
3918
3919         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3920                 kobject_put(&port->counter_group[i].obj);
3921
3922         kobject_put(&port->port_obj);
3923 }
3924
3925 static void cm_add_one(struct ib_device *ib_device)
3926 {
3927         struct cm_device *cm_dev;
3928         struct cm_port *port;
3929         struct ib_mad_reg_req reg_req = {
3930                 .mgmt_class = IB_MGMT_CLASS_CM,
3931                 .mgmt_class_version = IB_CM_CLASS_VERSION,
3932         };
3933         struct ib_port_modify port_modify = {
3934                 .set_port_cap_mask = IB_PORT_CM_SUP
3935         };
3936         unsigned long flags;
3937         int ret;
3938         int count = 0;
3939         u8 i;
3940
3941         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3942                          ib_device->phys_port_cnt, GFP_KERNEL);
3943         if (!cm_dev)
3944                 return;
3945
3946         cm_dev->ib_device = ib_device;
3947         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
3948         cm_dev->going_down = 0;
3949         cm_dev->device = device_create(&cm_class, &ib_device->dev,
3950                                        MKDEV(0, 0), NULL,
3951                                        "%s", ib_device->name);
3952         if (IS_ERR(cm_dev->device)) {
3953                 kfree(cm_dev);
3954                 return;
3955         }
3956
3957         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3958         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3959                 if (!rdma_cap_ib_cm(ib_device, i))
3960                         continue;
3961
3962                 port = kzalloc(sizeof *port, GFP_KERNEL);
3963                 if (!port)
3964                         goto error1;
3965
3966                 cm_dev->port[i-1] = port;
3967                 port->cm_dev = cm_dev;
3968                 port->port_num = i;
3969
3970                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
3971                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
3972
3973                 ret = cm_create_port_fs(port);
3974                 if (ret)
3975                         goto error1;
3976
3977                 port->mad_agent = ib_register_mad_agent(ib_device, i,
3978                                                         IB_QPT_GSI,
3979                                                         &reg_req,
3980                                                         0,
3981                                                         cm_send_handler,
3982                                                         cm_recv_handler,
3983                                                         port,
3984                                                         0);
3985                 if (IS_ERR(port->mad_agent))
3986                         goto error2;
3987
3988                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
3989                 if (ret)
3990                         goto error3;
3991
3992                 count++;
3993         }
3994
3995         if (!count)
3996                 goto free;
3997
3998         ib_set_client_data(ib_device, &cm_client, cm_dev);
3999
4000         write_lock_irqsave(&cm.device_lock, flags);
4001         list_add_tail(&cm_dev->list, &cm.device_list);
4002         write_unlock_irqrestore(&cm.device_lock, flags);
4003         return;
4004
4005 error3:
4006         ib_unregister_mad_agent(port->mad_agent);
4007 error2:
4008         cm_remove_port_fs(port);
4009 error1:
4010         port_modify.set_port_cap_mask = 0;
4011         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4012         while (--i) {
4013                 if (!rdma_cap_ib_cm(ib_device, i))
4014                         continue;
4015
4016                 port = cm_dev->port[i-1];
4017                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4018                 ib_unregister_mad_agent(port->mad_agent);
4019                 cm_remove_port_fs(port);
4020         }
4021 free:
4022         device_unregister(cm_dev->device);
4023         kfree(cm_dev);
4024 }
4025
4026 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4027 {
4028         struct cm_device *cm_dev = client_data;
4029         struct cm_port *port;
4030         struct cm_id_private *cm_id_priv;
4031         struct ib_mad_agent *cur_mad_agent;
4032         struct ib_port_modify port_modify = {
4033                 .clr_port_cap_mask = IB_PORT_CM_SUP
4034         };
4035         unsigned long flags;
4036         int i;
4037
4038         if (!cm_dev)
4039                 return;
4040
4041         write_lock_irqsave(&cm.device_lock, flags);
4042         list_del(&cm_dev->list);
4043         write_unlock_irqrestore(&cm.device_lock, flags);
4044
4045         spin_lock_irq(&cm.lock);
4046         cm_dev->going_down = 1;
4047         spin_unlock_irq(&cm.lock);
4048
4049         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4050                 if (!rdma_cap_ib_cm(ib_device, i))
4051                         continue;
4052
4053                 port = cm_dev->port[i-1];
4054                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4055                 /* Mark all the cm_id's as not valid */
4056                 spin_lock_irq(&cm.lock);
4057                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4058                         cm_id_priv->altr_send_port_not_ready = 1;
4059                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4060                         cm_id_priv->prim_send_port_not_ready = 1;
4061                 spin_unlock_irq(&cm.lock);
4062                 /*
4063                  * We flush the queue here after the going_down set, this
4064                  * verify that no new works will be queued in the recv handler,
4065                  * after that we can call the unregister_mad_agent
4066                  */
4067                 flush_workqueue(cm.wq);
4068                 spin_lock_irq(&cm.state_lock);
4069                 cur_mad_agent = port->mad_agent;
4070                 port->mad_agent = NULL;
4071                 spin_unlock_irq(&cm.state_lock);
4072                 ib_unregister_mad_agent(cur_mad_agent);
4073                 cm_remove_port_fs(port);
4074         }
4075
4076         device_unregister(cm_dev->device);
4077         kfree(cm_dev);
4078 }
4079
4080 static int __init ib_cm_init(void)
4081 {
4082         int ret;
4083
4084         memset(&cm, 0, sizeof cm);
4085         INIT_LIST_HEAD(&cm.device_list);
4086         rwlock_init(&cm.device_lock);
4087         spin_lock_init(&cm.lock);
4088         spin_lock_init(&cm.state_lock);
4089         cm.listen_service_table = RB_ROOT;
4090         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4091         cm.remote_id_table = RB_ROOT;
4092         cm.remote_qp_table = RB_ROOT;
4093         cm.remote_sidr_table = RB_ROOT;
4094         idr_init(&cm.local_id_table);
4095         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4096         INIT_LIST_HEAD(&cm.timewait_list);
4097
4098         ret = class_register(&cm_class);
4099         if (ret) {
4100                 ret = -ENOMEM;
4101                 goto error1;
4102         }
4103
4104         cm.wq = create_workqueue("ib_cm");
4105         if (!cm.wq) {
4106                 ret = -ENOMEM;
4107                 goto error2;
4108         }
4109
4110         ret = ib_register_client(&cm_client);
4111         if (ret)
4112                 goto error3;
4113
4114         return 0;
4115 error3:
4116         destroy_workqueue(cm.wq);
4117 error2:
4118         class_unregister(&cm_class);
4119 error1:
4120         idr_destroy(&cm.local_id_table);
4121         return ret;
4122 }
4123
4124 static void __exit ib_cm_cleanup(void)
4125 {
4126         struct cm_timewait_info *timewait_info, *tmp;
4127
4128         spin_lock_irq(&cm.lock);
4129         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4130                 cancel_delayed_work(&timewait_info->work.work);
4131         spin_unlock_irq(&cm.lock);
4132
4133         ib_unregister_client(&cm_client);
4134         destroy_workqueue(cm.wq);
4135
4136         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4137                 list_del(&timewait_info->list);
4138                 kfree(timewait_info);
4139         }
4140
4141         class_unregister(&cm_class);
4142         idr_destroy(&cm.local_id_table);
4143 }
4144
4145 module_init(ib_cm_init);
4146 module_exit(ib_cm_cleanup);
4147