kernel/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                                         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                         return ERR_PTR(-EINVAL);
1077                 }
1078                 atomic_inc(&cm_id_priv->refcount);
1079                 ++cm_id_priv->listen_sharecount;
1080                 spin_unlock_irqrestore(&cm.lock, flags);
1081
1082                 ib_destroy_cm_id(cm_id);
1083                 cm_id = &cm_id_priv->id;
1084                 return cm_id;
1085         }
1086
1087 new_id:
1088         /* Use newly created ID */
1089         err = __ib_cm_listen(cm_id, service_id, 0);
1090
1091         spin_unlock_irqrestore(&cm.lock, flags);
1092
1093         if (err) {
1094                 ib_destroy_cm_id(cm_id);
1095                 return ERR_PTR(err);
1096         }
1097         return cm_id;
1098 }
1099 EXPORT_SYMBOL(ib_cm_insert_listen);
1100
1101 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
1102                           enum cm_msg_sequence msg_seq)
1103 {
1104         u64 hi_tid, low_tid;
1105
1106         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1107         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
1108                           (msg_seq << 30));
1109         return cpu_to_be64(hi_tid | low_tid);
1110 }
1111
1112 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1113                               __be16 attr_id, __be64 tid)
1114 {
1115         hdr->base_version  = IB_MGMT_BASE_VERSION;
1116         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1117         hdr->class_version = IB_CM_CLASS_VERSION;
1118         hdr->method        = IB_MGMT_METHOD_SEND;
1119         hdr->attr_id       = attr_id;
1120         hdr->tid           = tid;
1121 }
1122
1123 static void cm_format_req(struct cm_req_msg *req_msg,
1124                           struct cm_id_private *cm_id_priv,
1125                           struct ib_cm_req_param *param)
1126 {
1127         struct ib_sa_path_rec *pri_path = param->primary_path;
1128         struct ib_sa_path_rec *alt_path = param->alternate_path;
1129
1130         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1131                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1132
1133         req_msg->local_comm_id = cm_id_priv->id.local_id;
1134         req_msg->service_id = param->service_id;
1135         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1136         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1137         cm_req_set_init_depth(req_msg, param->initiator_depth);
1138         cm_req_set_remote_resp_timeout(req_msg,
1139                                        param->remote_cm_response_timeout);
1140         cm_req_set_qp_type(req_msg, param->qp_type);
1141         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1142         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1143         cm_req_set_local_resp_timeout(req_msg,
1144                                       param->local_cm_response_timeout);
1145         req_msg->pkey = param->primary_path->pkey;
1146         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1147         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1148
1149         if (param->qp_type != IB_QPT_XRC_INI) {
1150                 cm_req_set_resp_res(req_msg, param->responder_resources);
1151                 cm_req_set_retry_count(req_msg, param->retry_count);
1152                 cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1153                 cm_req_set_srq(req_msg, param->srq);
1154         }
1155
1156         if (pri_path->hop_limit <= 1) {
1157                 req_msg->primary_local_lid = pri_path->slid;
1158                 req_msg->primary_remote_lid = pri_path->dlid;
1159         } else {
1160                 /* Work-around until there's a way to obtain remote LID info */
1161                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1162                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1163         }
1164         req_msg->primary_local_gid = pri_path->sgid;
1165         req_msg->primary_remote_gid = pri_path->dgid;
1166         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1167         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1168         req_msg->primary_traffic_class = pri_path->traffic_class;
1169         req_msg->primary_hop_limit = pri_path->hop_limit;
1170         cm_req_set_primary_sl(req_msg, pri_path->sl);
1171         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1172         cm_req_set_primary_local_ack_timeout(req_msg,
1173                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1174                                pri_path->packet_life_time));
1175
1176         if (alt_path) {
1177                 if (alt_path->hop_limit <= 1) {
1178                         req_msg->alt_local_lid = alt_path->slid;
1179                         req_msg->alt_remote_lid = alt_path->dlid;
1180                 } else {
1181                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1182                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1183                 }
1184                 req_msg->alt_local_gid = alt_path->sgid;
1185                 req_msg->alt_remote_gid = alt_path->dgid;
1186                 cm_req_set_alt_flow_label(req_msg,
1187                                           alt_path->flow_label);
1188                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1189                 req_msg->alt_traffic_class = alt_path->traffic_class;
1190                 req_msg->alt_hop_limit = alt_path->hop_limit;
1191                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1192                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1193                 cm_req_set_alt_local_ack_timeout(req_msg,
1194                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1195                                        alt_path->packet_life_time));
1196         }
1197
1198         if (param->private_data && param->private_data_len)
1199                 memcpy(req_msg->private_data, param->private_data,
1200                        param->private_data_len);
1201 }
1202
1203 static int cm_validate_req_param(struct ib_cm_req_param *param)
1204 {
1205         /* peer-to-peer not supported */
1206         if (param->peer_to_peer)
1207                 return -EINVAL;
1208
1209         if (!param->primary_path)
1210                 return -EINVAL;
1211
1212         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1213             param->qp_type != IB_QPT_XRC_INI)
1214                 return -EINVAL;
1215
1216         if (param->private_data &&
1217             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1218                 return -EINVAL;
1219
1220         if (param->alternate_path &&
1221             (param->alternate_path->pkey != param->primary_path->pkey ||
1222              param->alternate_path->mtu != param->primary_path->mtu))
1223                 return -EINVAL;
1224
1225         return 0;
1226 }
1227
1228 int ib_send_cm_req(struct ib_cm_id *cm_id,
1229                    struct ib_cm_req_param *param)
1230 {
1231         struct cm_id_private *cm_id_priv;
1232         struct cm_req_msg *req_msg;
1233         unsigned long flags;
1234         int ret;
1235
1236         ret = cm_validate_req_param(param);
1237         if (ret)
1238                 return ret;
1239
1240         /* Verify that we're not in timewait. */
1241         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1242         spin_lock_irqsave(&cm_id_priv->lock, flags);
1243         if (cm_id->state != IB_CM_IDLE) {
1244                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1245                 ret = -EINVAL;
1246                 goto out;
1247         }
1248         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1249
1250         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1251                                                             id.local_id);
1252         if (IS_ERR(cm_id_priv->timewait_info)) {
1253                 ret = PTR_ERR(cm_id_priv->timewait_info);
1254                 goto out;
1255         }
1256
1257         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av,
1258                                  cm_id_priv);
1259         if (ret)
1260                 goto error1;
1261         if (param->alternate_path) {
1262                 ret = cm_init_av_by_path(param->alternate_path,
1263                                          &cm_id_priv->alt_av, cm_id_priv);
1264                 if (ret)
1265                         goto error1;
1266         }
1267         cm_id->service_id = param->service_id;
1268         cm_id->service_mask = ~cpu_to_be64(0);
1269         cm_id_priv->timeout_ms = cm_convert_to_ms(
1270                                     param->primary_path->packet_life_time) * 2 +
1271                                  cm_convert_to_ms(
1272                                     param->remote_cm_response_timeout);
1273         cm_id_priv->max_cm_retries = param->max_cm_retries;
1274         cm_id_priv->initiator_depth = param->initiator_depth;
1275         cm_id_priv->responder_resources = param->responder_resources;
1276         cm_id_priv->retry_count = param->retry_count;
1277         cm_id_priv->path_mtu = param->primary_path->mtu;
1278         cm_id_priv->pkey = param->primary_path->pkey;
1279         cm_id_priv->qp_type = param->qp_type;
1280
1281         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1282         if (ret)
1283                 goto error1;
1284
1285         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1286         cm_format_req(req_msg, cm_id_priv, param);
1287         cm_id_priv->tid = req_msg->hdr.tid;
1288         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1289         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1290
1291         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1292         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1293
1294         spin_lock_irqsave(&cm_id_priv->lock, flags);
1295         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1296         if (ret) {
1297                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1298                 goto error2;
1299         }
1300         BUG_ON(cm_id->state != IB_CM_IDLE);
1301         cm_id->state = IB_CM_REQ_SENT;
1302         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1303         return 0;
1304
1305 error2: cm_free_msg(cm_id_priv->msg);
1306 error1: kfree(cm_id_priv->timewait_info);
1307 out:    return ret;
1308 }
1309 EXPORT_SYMBOL(ib_send_cm_req);
1310
1311 static int cm_issue_rej(struct cm_port *port,
1312                         struct ib_mad_recv_wc *mad_recv_wc,
1313                         enum ib_cm_rej_reason reason,
1314                         enum cm_msg_response msg_rejected,
1315                         void *ari, u8 ari_length)
1316 {
1317         struct ib_mad_send_buf *msg = NULL;
1318         struct cm_rej_msg *rej_msg, *rcv_msg;
1319         int ret;
1320
1321         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1322         if (ret)
1323                 return ret;
1324
1325         /* We just need common CM header information.  Cast to any message. */
1326         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1327         rej_msg = (struct cm_rej_msg *) msg->mad;
1328
1329         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1330         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1331         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1332         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1333         rej_msg->reason = cpu_to_be16(reason);
1334
1335         if (ari && ari_length) {
1336                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1337                 memcpy(rej_msg->ari, ari, ari_length);
1338         }
1339
1340         ret = ib_post_send_mad(msg, NULL);
1341         if (ret)
1342                 cm_free_msg(msg);
1343
1344         return ret;
1345 }
1346
1347 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1348                                     __be32 local_qpn, __be32 remote_qpn)
1349 {
1350         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1351                 ((local_ca_guid == remote_ca_guid) &&
1352                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1353 }
1354
1355 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1356                                             struct ib_sa_path_rec *primary_path,
1357                                             struct ib_sa_path_rec *alt_path)
1358 {
1359         memset(primary_path, 0, sizeof *primary_path);
1360         primary_path->dgid = req_msg->primary_local_gid;
1361         primary_path->sgid = req_msg->primary_remote_gid;
1362         primary_path->dlid = req_msg->primary_local_lid;
1363         primary_path->slid = req_msg->primary_remote_lid;
1364         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1365         primary_path->hop_limit = req_msg->primary_hop_limit;
1366         primary_path->traffic_class = req_msg->primary_traffic_class;
1367         primary_path->reversible = 1;
1368         primary_path->pkey = req_msg->pkey;
1369         primary_path->sl = cm_req_get_primary_sl(req_msg);
1370         primary_path->mtu_selector = IB_SA_EQ;
1371         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1372         primary_path->rate_selector = IB_SA_EQ;
1373         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1374         primary_path->packet_life_time_selector = IB_SA_EQ;
1375         primary_path->packet_life_time =
1376                 cm_req_get_primary_local_ack_timeout(req_msg);
1377         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1378         primary_path->service_id = req_msg->service_id;
1379
1380         if (req_msg->alt_local_lid) {
1381                 memset(alt_path, 0, sizeof *alt_path);
1382                 alt_path->dgid = req_msg->alt_local_gid;
1383                 alt_path->sgid = req_msg->alt_remote_gid;
1384                 alt_path->dlid = req_msg->alt_local_lid;
1385                 alt_path->slid = req_msg->alt_remote_lid;
1386                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1387                 alt_path->hop_limit = req_msg->alt_hop_limit;
1388                 alt_path->traffic_class = req_msg->alt_traffic_class;
1389                 alt_path->reversible = 1;
1390                 alt_path->pkey = req_msg->pkey;
1391                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1392                 alt_path->mtu_selector = IB_SA_EQ;
1393                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1394                 alt_path->rate_selector = IB_SA_EQ;
1395                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1396                 alt_path->packet_life_time_selector = IB_SA_EQ;
1397                 alt_path->packet_life_time =
1398                         cm_req_get_alt_local_ack_timeout(req_msg);
1399                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1400                 alt_path->service_id = req_msg->service_id;
1401         }
1402 }
1403
1404 static u16 cm_get_bth_pkey(struct cm_work *work)
1405 {
1406         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1407         u8 port_num = work->port->port_num;
1408         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1409         u16 pkey;
1410         int ret;
1411
1412         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1413         if (ret) {
1414                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1415                                      port_num, pkey_index, ret);
1416                 return 0;
1417         }
1418
1419         return pkey;
1420 }
1421
1422 static void cm_format_req_event(struct cm_work *work,
1423                                 struct cm_id_private *cm_id_priv,
1424                                 struct ib_cm_id *listen_id)
1425 {
1426         struct cm_req_msg *req_msg;
1427         struct ib_cm_req_event_param *param;
1428
1429         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1430         param = &work->cm_event.param.req_rcvd;
1431         param->listen_id = listen_id;
1432         param->bth_pkey = cm_get_bth_pkey(work);
1433         param->port = cm_id_priv->av.port->port_num;
1434         param->primary_path = &work->path[0];
1435         if (req_msg->alt_local_lid)
1436                 param->alternate_path = &work->path[1];
1437         else
1438                 param->alternate_path = NULL;
1439         param->remote_ca_guid = req_msg->local_ca_guid;
1440         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1441         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1442         param->qp_type = cm_req_get_qp_type(req_msg);
1443         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1444         param->responder_resources = cm_req_get_init_depth(req_msg);
1445         param->initiator_depth = cm_req_get_resp_res(req_msg);
1446         param->local_cm_response_timeout =
1447                                         cm_req_get_remote_resp_timeout(req_msg);
1448         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1449         param->remote_cm_response_timeout =
1450                                         cm_req_get_local_resp_timeout(req_msg);
1451         param->retry_count = cm_req_get_retry_count(req_msg);
1452         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1453         param->srq = cm_req_get_srq(req_msg);
1454         work->cm_event.private_data = &req_msg->private_data;
1455 }
1456
1457 static void cm_process_work(struct cm_id_private *cm_id_priv,
1458                             struct cm_work *work)
1459 {
1460         int ret;
1461
1462         /* We will typically only have the current event to report. */
1463         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1464         cm_free_work(work);
1465
1466         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1467                 spin_lock_irq(&cm_id_priv->lock);
1468                 work = cm_dequeue_work(cm_id_priv);
1469                 spin_unlock_irq(&cm_id_priv->lock);
1470                 BUG_ON(!work);
1471                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1472                                                 &work->cm_event);
1473                 cm_free_work(work);
1474         }
1475         cm_deref_id(cm_id_priv);
1476         if (ret)
1477                 cm_destroy_id(&cm_id_priv->id, ret);
1478 }
1479
1480 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1481                           struct cm_id_private *cm_id_priv,
1482                           enum cm_msg_response msg_mraed, u8 service_timeout,
1483                           const void *private_data, u8 private_data_len)
1484 {
1485         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1486         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1487         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1488         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1489         cm_mra_set_service_timeout(mra_msg, service_timeout);
1490
1491         if (private_data && private_data_len)
1492                 memcpy(mra_msg->private_data, private_data, private_data_len);
1493 }
1494
1495 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1496                           struct cm_id_private *cm_id_priv,
1497                           enum ib_cm_rej_reason reason,
1498                           void *ari,
1499                           u8 ari_length,
1500                           const void *private_data,
1501                           u8 private_data_len)
1502 {
1503         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1504         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1505
1506         switch(cm_id_priv->id.state) {
1507         case IB_CM_REQ_RCVD:
1508                 rej_msg->local_comm_id = 0;
1509                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1510                 break;
1511         case IB_CM_MRA_REQ_SENT:
1512                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1513                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1514                 break;
1515         case IB_CM_REP_RCVD:
1516         case IB_CM_MRA_REP_SENT:
1517                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1518                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1519                 break;
1520         default:
1521                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1522                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1523                 break;
1524         }
1525
1526         rej_msg->reason = cpu_to_be16(reason);
1527         if (ari && ari_length) {
1528                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1529                 memcpy(rej_msg->ari, ari, ari_length);
1530         }
1531
1532         if (private_data && private_data_len)
1533                 memcpy(rej_msg->private_data, private_data, private_data_len);
1534 }
1535
1536 static void cm_dup_req_handler(struct cm_work *work,
1537                                struct cm_id_private *cm_id_priv)
1538 {
1539         struct ib_mad_send_buf *msg = NULL;
1540         int ret;
1541
1542         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1543                         counter[CM_REQ_COUNTER]);
1544
1545         /* Quick state check to discard duplicate REQs. */
1546         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1547                 return;
1548
1549         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1550         if (ret)
1551                 return;
1552
1553         spin_lock_irq(&cm_id_priv->lock);
1554         switch (cm_id_priv->id.state) {
1555         case IB_CM_MRA_REQ_SENT:
1556                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1557                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1558                               cm_id_priv->private_data,
1559                               cm_id_priv->private_data_len);
1560                 break;
1561         case IB_CM_TIMEWAIT:
1562                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1563                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1564                 break;
1565         default:
1566                 goto unlock;
1567         }
1568         spin_unlock_irq(&cm_id_priv->lock);
1569
1570         ret = ib_post_send_mad(msg, NULL);
1571         if (ret)
1572                 goto free;
1573         return;
1574
1575 unlock: spin_unlock_irq(&cm_id_priv->lock);
1576 free:   cm_free_msg(msg);
1577 }
1578
1579 static struct cm_id_private * cm_match_req(struct cm_work *work,
1580                                            struct cm_id_private *cm_id_priv)
1581 {
1582         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1583         struct cm_timewait_info *timewait_info;
1584         struct cm_req_msg *req_msg;
1585
1586         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1587
1588         /* Check for possible duplicate REQ. */
1589         spin_lock_irq(&cm.lock);
1590         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1591         if (timewait_info) {
1592                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1593                                            timewait_info->work.remote_id);
1594                 spin_unlock_irq(&cm.lock);
1595                 if (cur_cm_id_priv) {
1596                         cm_dup_req_handler(work, cur_cm_id_priv);
1597                         cm_deref_id(cur_cm_id_priv);
1598                 }
1599                 return NULL;
1600         }
1601
1602         /* Check for stale connections. */
1603         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1604         if (timewait_info) {
1605                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1606                 spin_unlock_irq(&cm.lock);
1607                 cm_issue_rej(work->port, work->mad_recv_wc,
1608                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1609                              NULL, 0);
1610                 return NULL;
1611         }
1612
1613         /* Find matching listen request. */
1614         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1615                                            req_msg->service_id);
1616         if (!listen_cm_id_priv) {
1617                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1618                 spin_unlock_irq(&cm.lock);
1619                 cm_issue_rej(work->port, work->mad_recv_wc,
1620                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1621                              NULL, 0);
1622                 goto out;
1623         }
1624         atomic_inc(&listen_cm_id_priv->refcount);
1625         atomic_inc(&cm_id_priv->refcount);
1626         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1627         atomic_inc(&cm_id_priv->work_count);
1628         spin_unlock_irq(&cm.lock);
1629 out:
1630         return listen_cm_id_priv;
1631 }
1632
1633 /*
1634  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1635  * we need to override the LID/SL data in the REQ with the LID information
1636  * in the work completion.
1637  */
1638 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1639 {
1640         if (!cm_req_get_primary_subnet_local(req_msg)) {
1641                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1642                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1643                         cm_req_set_primary_sl(req_msg, wc->sl);
1644                 }
1645
1646                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1647                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1648         }
1649
1650         if (!cm_req_get_alt_subnet_local(req_msg)) {
1651                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1652                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1653                         cm_req_set_alt_sl(req_msg, wc->sl);
1654                 }
1655
1656                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1657                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1658         }
1659 }
1660
1661 static int cm_req_handler(struct cm_work *work)
1662 {
1663         struct ib_cm_id *cm_id;
1664         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1665         struct cm_req_msg *req_msg;
1666         int ret;
1667
1668         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1669
1670         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1671         if (IS_ERR(cm_id))
1672                 return PTR_ERR(cm_id);
1673
1674         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1675         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1676         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1677                                 work->mad_recv_wc->recv_buf.grh,
1678                                 &cm_id_priv->av);
1679         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1680                                                             id.local_id);
1681         if (IS_ERR(cm_id_priv->timewait_info)) {
1682                 ret = PTR_ERR(cm_id_priv->timewait_info);
1683                 goto destroy;
1684         }
1685         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1686         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1687         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1688
1689         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1690         if (!listen_cm_id_priv) {
1691                 ret = -EINVAL;
1692                 kfree(cm_id_priv->timewait_info);
1693                 goto destroy;
1694         }
1695
1696         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1697         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1698         cm_id_priv->id.service_id = req_msg->service_id;
1699         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1700
1701         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1702         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1703
1704         memcpy(work->path[0].dmac, cm_id_priv->av.ah_attr.dmac, ETH_ALEN);
1705         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av,
1706                                  cm_id_priv);
1707         if (ret) {
1708                 ib_get_cached_gid(work->port->cm_dev->ib_device,
1709                                   work->port->port_num, 0, &work->path[0].sgid,
1710                                   NULL);
1711                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1712                                &work->path[0].sgid, sizeof work->path[0].sgid,
1713                                NULL, 0);
1714                 goto rejected;
1715         }
1716         if (req_msg->alt_local_lid) {
1717                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av,
1718                                          cm_id_priv);
1719                 if (ret) {
1720                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1721                                        &work->path[0].sgid,
1722                                        sizeof work->path[0].sgid, NULL, 0);
1723                         goto rejected;
1724                 }
1725         }
1726         cm_id_priv->tid = req_msg->hdr.tid;
1727         cm_id_priv->timeout_ms = cm_convert_to_ms(
1728                                         cm_req_get_local_resp_timeout(req_msg));
1729         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1730         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1731         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1732         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1733         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1734         cm_id_priv->pkey = req_msg->pkey;
1735         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1736         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1737         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1738         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1739
1740         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1741         cm_process_work(cm_id_priv, work);
1742         cm_deref_id(listen_cm_id_priv);
1743         return 0;
1744
1745 rejected:
1746         atomic_dec(&cm_id_priv->refcount);
1747         cm_deref_id(listen_cm_id_priv);
1748 destroy:
1749         ib_destroy_cm_id(cm_id);
1750         return ret;
1751 }
1752
1753 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1754                           struct cm_id_private *cm_id_priv,
1755                           struct ib_cm_rep_param *param)
1756 {
1757         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1758         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1759         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1760         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1761         rep_msg->resp_resources = param->responder_resources;
1762         cm_rep_set_target_ack_delay(rep_msg,
1763                                     cm_id_priv->av.port->cm_dev->ack_delay);
1764         cm_rep_set_failover(rep_msg, param->failover_accepted);
1765         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1766         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1767
1768         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1769                 rep_msg->initiator_depth = param->initiator_depth;
1770                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1771                 cm_rep_set_srq(rep_msg, param->srq);
1772                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1773         } else {
1774                 cm_rep_set_srq(rep_msg, 1);
1775                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1776         }
1777
1778         if (param->private_data && param->private_data_len)
1779                 memcpy(rep_msg->private_data, param->private_data,
1780                        param->private_data_len);
1781 }
1782
1783 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1784                    struct ib_cm_rep_param *param)
1785 {
1786         struct cm_id_private *cm_id_priv;
1787         struct ib_mad_send_buf *msg;
1788         struct cm_rep_msg *rep_msg;
1789         unsigned long flags;
1790         int ret;
1791
1792         if (param->private_data &&
1793             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1794                 return -EINVAL;
1795
1796         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1797         spin_lock_irqsave(&cm_id_priv->lock, flags);
1798         if (cm_id->state != IB_CM_REQ_RCVD &&
1799             cm_id->state != IB_CM_MRA_REQ_SENT) {
1800                 ret = -EINVAL;
1801                 goto out;
1802         }
1803
1804         ret = cm_alloc_msg(cm_id_priv, &msg);
1805         if (ret)
1806                 goto out;
1807
1808         rep_msg = (struct cm_rep_msg *) msg->mad;
1809         cm_format_rep(rep_msg, cm_id_priv, param);
1810         msg->timeout_ms = cm_id_priv->timeout_ms;
1811         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1812
1813         ret = ib_post_send_mad(msg, NULL);
1814         if (ret) {
1815                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1816                 cm_free_msg(msg);
1817                 return ret;
1818         }
1819
1820         cm_id->state = IB_CM_REP_SENT;
1821         cm_id_priv->msg = msg;
1822         cm_id_priv->initiator_depth = param->initiator_depth;
1823         cm_id_priv->responder_resources = param->responder_resources;
1824         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1825         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
1826
1827 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1828         return ret;
1829 }
1830 EXPORT_SYMBOL(ib_send_cm_rep);
1831
1832 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1833                           struct cm_id_private *cm_id_priv,
1834                           const void *private_data,
1835                           u8 private_data_len)
1836 {
1837         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1838         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1839         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1840
1841         if (private_data && private_data_len)
1842                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1843 }
1844
1845 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1846                    const void *private_data,
1847                    u8 private_data_len)
1848 {
1849         struct cm_id_private *cm_id_priv;
1850         struct ib_mad_send_buf *msg;
1851         unsigned long flags;
1852         void *data;
1853         int ret;
1854
1855         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1856                 return -EINVAL;
1857
1858         data = cm_copy_private_data(private_data, private_data_len);
1859         if (IS_ERR(data))
1860                 return PTR_ERR(data);
1861
1862         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1863         spin_lock_irqsave(&cm_id_priv->lock, flags);
1864         if (cm_id->state != IB_CM_REP_RCVD &&
1865             cm_id->state != IB_CM_MRA_REP_SENT) {
1866                 ret = -EINVAL;
1867                 goto error;
1868         }
1869
1870         ret = cm_alloc_msg(cm_id_priv, &msg);
1871         if (ret)
1872                 goto error;
1873
1874         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1875                       private_data, private_data_len);
1876
1877         ret = ib_post_send_mad(msg, NULL);
1878         if (ret) {
1879                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1880                 cm_free_msg(msg);
1881                 kfree(data);
1882                 return ret;
1883         }
1884
1885         cm_id->state = IB_CM_ESTABLISHED;
1886         cm_set_private_data(cm_id_priv, data, private_data_len);
1887         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1888         return 0;
1889
1890 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1891         kfree(data);
1892         return ret;
1893 }
1894 EXPORT_SYMBOL(ib_send_cm_rtu);
1895
1896 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
1897 {
1898         struct cm_rep_msg *rep_msg;
1899         struct ib_cm_rep_event_param *param;
1900
1901         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1902         param = &work->cm_event.param.rep_rcvd;
1903         param->remote_ca_guid = rep_msg->local_ca_guid;
1904         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1905         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
1906         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1907         param->responder_resources = rep_msg->initiator_depth;
1908         param->initiator_depth = rep_msg->resp_resources;
1909         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1910         param->failover_accepted = cm_rep_get_failover(rep_msg);
1911         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1912         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1913         param->srq = cm_rep_get_srq(rep_msg);
1914         work->cm_event.private_data = &rep_msg->private_data;
1915 }
1916
1917 static void cm_dup_rep_handler(struct cm_work *work)
1918 {
1919         struct cm_id_private *cm_id_priv;
1920         struct cm_rep_msg *rep_msg;
1921         struct ib_mad_send_buf *msg = NULL;
1922         int ret;
1923
1924         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1925         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1926                                    rep_msg->local_comm_id);
1927         if (!cm_id_priv)
1928                 return;
1929
1930         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1931                         counter[CM_REP_COUNTER]);
1932         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1933         if (ret)
1934                 goto deref;
1935
1936         spin_lock_irq(&cm_id_priv->lock);
1937         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1938                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1939                               cm_id_priv->private_data,
1940                               cm_id_priv->private_data_len);
1941         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1942                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1943                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1944                               cm_id_priv->private_data,
1945                               cm_id_priv->private_data_len);
1946         else
1947                 goto unlock;
1948         spin_unlock_irq(&cm_id_priv->lock);
1949
1950         ret = ib_post_send_mad(msg, NULL);
1951         if (ret)
1952                 goto free;
1953         goto deref;
1954
1955 unlock: spin_unlock_irq(&cm_id_priv->lock);
1956 free:   cm_free_msg(msg);
1957 deref:  cm_deref_id(cm_id_priv);
1958 }
1959
1960 static int cm_rep_handler(struct cm_work *work)
1961 {
1962         struct cm_id_private *cm_id_priv;
1963         struct cm_rep_msg *rep_msg;
1964         int ret;
1965
1966         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1967         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1968         if (!cm_id_priv) {
1969                 cm_dup_rep_handler(work);
1970                 return -EINVAL;
1971         }
1972
1973         cm_format_rep_event(work, cm_id_priv->qp_type);
1974
1975         spin_lock_irq(&cm_id_priv->lock);
1976         switch (cm_id_priv->id.state) {
1977         case IB_CM_REQ_SENT:
1978         case IB_CM_MRA_REQ_RCVD:
1979                 break;
1980         default:
1981                 spin_unlock_irq(&cm_id_priv->lock);
1982                 ret = -EINVAL;
1983                 goto error;
1984         }
1985
1986         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1987         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1988         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
1989
1990         spin_lock(&cm.lock);
1991         /* Check for duplicate REP. */
1992         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1993                 spin_unlock(&cm.lock);
1994                 spin_unlock_irq(&cm_id_priv->lock);
1995                 ret = -EINVAL;
1996                 goto error;
1997         }
1998         /* Check for a stale connection. */
1999         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
2000                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
2001                          &cm.remote_id_table);
2002                 cm_id_priv->timewait_info->inserted_remote_id = 0;
2003                 spin_unlock(&cm.lock);
2004                 spin_unlock_irq(&cm_id_priv->lock);
2005                 cm_issue_rej(work->port, work->mad_recv_wc,
2006                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2007                              NULL, 0);
2008                 ret = -EINVAL;
2009                 goto error;
2010         }
2011         spin_unlock(&cm.lock);
2012
2013         cm_id_priv->id.state = IB_CM_REP_RCVD;
2014         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
2015         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2016         cm_id_priv->initiator_depth = rep_msg->resp_resources;
2017         cm_id_priv->responder_resources = rep_msg->initiator_depth;
2018         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
2019         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2020         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2021         cm_id_priv->av.timeout =
2022                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2023                                        cm_id_priv->av.timeout - 1);
2024         cm_id_priv->alt_av.timeout =
2025                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2026                                        cm_id_priv->alt_av.timeout - 1);
2027
2028         /* todo: handle peer_to_peer */
2029
2030         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2031         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2032         if (!ret)
2033                 list_add_tail(&work->list, &cm_id_priv->work_list);
2034         spin_unlock_irq(&cm_id_priv->lock);
2035
2036         if (ret)
2037                 cm_process_work(cm_id_priv, work);
2038         else
2039                 cm_deref_id(cm_id_priv);
2040         return 0;
2041
2042 error:
2043         cm_deref_id(cm_id_priv);
2044         return ret;
2045 }
2046
2047 static int cm_establish_handler(struct cm_work *work)
2048 {
2049         struct cm_id_private *cm_id_priv;
2050         int ret;
2051
2052         /* See comment in cm_establish about lookup. */
2053         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2054         if (!cm_id_priv)
2055                 return -EINVAL;
2056
2057         spin_lock_irq(&cm_id_priv->lock);
2058         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2059                 spin_unlock_irq(&cm_id_priv->lock);
2060                 goto out;
2061         }
2062
2063         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2064         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2065         if (!ret)
2066                 list_add_tail(&work->list, &cm_id_priv->work_list);
2067         spin_unlock_irq(&cm_id_priv->lock);
2068
2069         if (ret)
2070                 cm_process_work(cm_id_priv, work);
2071         else
2072                 cm_deref_id(cm_id_priv);
2073         return 0;
2074 out:
2075         cm_deref_id(cm_id_priv);
2076         return -EINVAL;
2077 }
2078
2079 static int cm_rtu_handler(struct cm_work *work)
2080 {
2081         struct cm_id_private *cm_id_priv;
2082         struct cm_rtu_msg *rtu_msg;
2083         int ret;
2084
2085         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2086         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
2087                                    rtu_msg->local_comm_id);
2088         if (!cm_id_priv)
2089                 return -EINVAL;
2090
2091         work->cm_event.private_data = &rtu_msg->private_data;
2092
2093         spin_lock_irq(&cm_id_priv->lock);
2094         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2095             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2096                 spin_unlock_irq(&cm_id_priv->lock);
2097                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2098                                 counter[CM_RTU_COUNTER]);
2099                 goto out;
2100         }
2101         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2102
2103         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2104         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2105         if (!ret)
2106                 list_add_tail(&work->list, &cm_id_priv->work_list);
2107         spin_unlock_irq(&cm_id_priv->lock);
2108
2109         if (ret)
2110                 cm_process_work(cm_id_priv, work);
2111         else
2112                 cm_deref_id(cm_id_priv);
2113         return 0;
2114 out:
2115         cm_deref_id(cm_id_priv);
2116         return -EINVAL;
2117 }
2118
2119 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2120                           struct cm_id_private *cm_id_priv,
2121                           const void *private_data,
2122                           u8 private_data_len)
2123 {
2124         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2125                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
2126         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
2127         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
2128         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2129
2130         if (private_data && private_data_len)
2131                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2132 }
2133
2134 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2135                     const void *private_data,
2136                     u8 private_data_len)
2137 {
2138         struct cm_id_private *cm_id_priv;
2139         struct ib_mad_send_buf *msg;
2140         unsigned long flags;
2141         int ret;
2142
2143         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2144                 return -EINVAL;
2145
2146         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2147         spin_lock_irqsave(&cm_id_priv->lock, flags);
2148         if (cm_id->state != IB_CM_ESTABLISHED) {
2149                 ret = -EINVAL;
2150                 goto out;
2151         }
2152
2153         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2154             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2155                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2156
2157         ret = cm_alloc_msg(cm_id_priv, &msg);
2158         if (ret) {
2159                 cm_enter_timewait(cm_id_priv);
2160                 goto out;
2161         }
2162
2163         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2164                        private_data, private_data_len);
2165         msg->timeout_ms = cm_id_priv->timeout_ms;
2166         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2167
2168         ret = ib_post_send_mad(msg, NULL);
2169         if (ret) {
2170                 cm_enter_timewait(cm_id_priv);
2171                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2172                 cm_free_msg(msg);
2173                 return ret;
2174         }
2175
2176         cm_id->state = IB_CM_DREQ_SENT;
2177         cm_id_priv->msg = msg;
2178 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2179         return ret;
2180 }
2181 EXPORT_SYMBOL(ib_send_cm_dreq);
2182
2183 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2184                           struct cm_id_private *cm_id_priv,
2185                           const void *private_data,
2186                           u8 private_data_len)
2187 {
2188         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2189         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2190         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2191
2192         if (private_data && private_data_len)
2193                 memcpy(drep_msg->private_data, private_data, private_data_len);
2194 }
2195
2196 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2197                     const void *private_data,
2198                     u8 private_data_len)
2199 {
2200         struct cm_id_private *cm_id_priv;
2201         struct ib_mad_send_buf *msg;
2202         unsigned long flags;
2203         void *data;
2204         int ret;
2205
2206         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2207                 return -EINVAL;
2208
2209         data = cm_copy_private_data(private_data, private_data_len);
2210         if (IS_ERR(data))
2211                 return PTR_ERR(data);
2212
2213         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2214         spin_lock_irqsave(&cm_id_priv->lock, flags);
2215         if (cm_id->state != IB_CM_DREQ_RCVD) {
2216                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2217                 kfree(data);
2218                 return -EINVAL;
2219         }
2220
2221         cm_set_private_data(cm_id_priv, data, private_data_len);
2222         cm_enter_timewait(cm_id_priv);
2223
2224         ret = cm_alloc_msg(cm_id_priv, &msg);
2225         if (ret)
2226                 goto out;
2227
2228         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2229                        private_data, private_data_len);
2230
2231         ret = ib_post_send_mad(msg, NULL);
2232         if (ret) {
2233                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2234                 cm_free_msg(msg);
2235                 return ret;
2236         }
2237
2238 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2239         return ret;
2240 }
2241 EXPORT_SYMBOL(ib_send_cm_drep);
2242
2243 static int cm_issue_drep(struct cm_port *port,
2244                          struct ib_mad_recv_wc *mad_recv_wc)
2245 {
2246         struct ib_mad_send_buf *msg = NULL;
2247         struct cm_dreq_msg *dreq_msg;
2248         struct cm_drep_msg *drep_msg;
2249         int ret;
2250
2251         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2252         if (ret)
2253                 return ret;
2254
2255         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2256         drep_msg = (struct cm_drep_msg *) msg->mad;
2257
2258         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2259         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2260         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2261
2262         ret = ib_post_send_mad(msg, NULL);
2263         if (ret)
2264                 cm_free_msg(msg);
2265
2266         return ret;
2267 }
2268
2269 static int cm_dreq_handler(struct cm_work *work)
2270 {
2271         struct cm_id_private *cm_id_priv;
2272         struct cm_dreq_msg *dreq_msg;
2273         struct ib_mad_send_buf *msg = NULL;
2274         int ret;
2275
2276         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2277         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2278                                    dreq_msg->local_comm_id);
2279         if (!cm_id_priv) {
2280                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2281                                 counter[CM_DREQ_COUNTER]);
2282                 cm_issue_drep(work->port, work->mad_recv_wc);
2283                 return -EINVAL;
2284         }
2285
2286         work->cm_event.private_data = &dreq_msg->private_data;
2287
2288         spin_lock_irq(&cm_id_priv->lock);
2289         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2290                 goto unlock;
2291
2292         switch (cm_id_priv->id.state) {
2293         case IB_CM_REP_SENT:
2294         case IB_CM_DREQ_SENT:
2295                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2296                 break;
2297         case IB_CM_ESTABLISHED:
2298                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2299                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2300                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2301                 break;
2302         case IB_CM_MRA_REP_RCVD:
2303                 break;
2304         case IB_CM_TIMEWAIT:
2305                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2306                                 counter[CM_DREQ_COUNTER]);
2307                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2308                         goto unlock;
2309
2310                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2311                                cm_id_priv->private_data,
2312                                cm_id_priv->private_data_len);
2313                 spin_unlock_irq(&cm_id_priv->lock);
2314
2315                 if (ib_post_send_mad(msg, NULL))
2316                         cm_free_msg(msg);
2317                 goto deref;
2318         case IB_CM_DREQ_RCVD:
2319                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2320                                 counter[CM_DREQ_COUNTER]);
2321                 goto unlock;
2322         default:
2323                 goto unlock;
2324         }
2325         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2326         cm_id_priv->tid = dreq_msg->hdr.tid;
2327         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2328         if (!ret)
2329                 list_add_tail(&work->list, &cm_id_priv->work_list);
2330         spin_unlock_irq(&cm_id_priv->lock);
2331
2332         if (ret)
2333                 cm_process_work(cm_id_priv, work);
2334         else
2335                 cm_deref_id(cm_id_priv);
2336         return 0;
2337
2338 unlock: spin_unlock_irq(&cm_id_priv->lock);
2339 deref:  cm_deref_id(cm_id_priv);
2340         return -EINVAL;
2341 }
2342
2343 static int cm_drep_handler(struct cm_work *work)
2344 {
2345         struct cm_id_private *cm_id_priv;
2346         struct cm_drep_msg *drep_msg;
2347         int ret;
2348
2349         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2350         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2351                                    drep_msg->local_comm_id);
2352         if (!cm_id_priv)
2353                 return -EINVAL;
2354
2355         work->cm_event.private_data = &drep_msg->private_data;
2356
2357         spin_lock_irq(&cm_id_priv->lock);
2358         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2359             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2360                 spin_unlock_irq(&cm_id_priv->lock);
2361                 goto out;
2362         }
2363         cm_enter_timewait(cm_id_priv);
2364
2365         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2366         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2367         if (!ret)
2368                 list_add_tail(&work->list, &cm_id_priv->work_list);
2369         spin_unlock_irq(&cm_id_priv->lock);
2370
2371         if (ret)
2372                 cm_process_work(cm_id_priv, work);
2373         else
2374                 cm_deref_id(cm_id_priv);
2375         return 0;
2376 out:
2377         cm_deref_id(cm_id_priv);
2378         return -EINVAL;
2379 }
2380
2381 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2382                    enum ib_cm_rej_reason reason,
2383                    void *ari,
2384                    u8 ari_length,
2385                    const void *private_data,
2386                    u8 private_data_len)
2387 {
2388         struct cm_id_private *cm_id_priv;
2389         struct ib_mad_send_buf *msg;
2390         unsigned long flags;
2391         int ret;
2392
2393         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2394             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2395                 return -EINVAL;
2396
2397         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2398
2399         spin_lock_irqsave(&cm_id_priv->lock, flags);
2400         switch (cm_id->state) {
2401         case IB_CM_REQ_SENT:
2402         case IB_CM_MRA_REQ_RCVD:
2403         case IB_CM_REQ_RCVD:
2404         case IB_CM_MRA_REQ_SENT:
2405         case IB_CM_REP_RCVD:
2406         case IB_CM_MRA_REP_SENT:
2407                 ret = cm_alloc_msg(cm_id_priv, &msg);
2408                 if (!ret)
2409                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2410                                       cm_id_priv, reason, ari, ari_length,
2411                                       private_data, private_data_len);
2412
2413                 cm_reset_to_idle(cm_id_priv);
2414                 break;
2415         case IB_CM_REP_SENT:
2416         case IB_CM_MRA_REP_RCVD:
2417                 ret = cm_alloc_msg(cm_id_priv, &msg);
2418                 if (!ret)
2419                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2420                                       cm_id_priv, reason, ari, ari_length,
2421                                       private_data, private_data_len);
2422
2423                 cm_enter_timewait(cm_id_priv);
2424                 break;
2425         default:
2426                 ret = -EINVAL;
2427                 goto out;
2428         }
2429
2430         if (ret)
2431                 goto out;
2432
2433         ret = ib_post_send_mad(msg, NULL);
2434         if (ret)
2435                 cm_free_msg(msg);
2436
2437 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2438         return ret;
2439 }
2440 EXPORT_SYMBOL(ib_send_cm_rej);
2441
2442 static void cm_format_rej_event(struct cm_work *work)
2443 {
2444         struct cm_rej_msg *rej_msg;
2445         struct ib_cm_rej_event_param *param;
2446
2447         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2448         param = &work->cm_event.param.rej_rcvd;
2449         param->ari = rej_msg->ari;
2450         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2451         param->reason = __be16_to_cpu(rej_msg->reason);
2452         work->cm_event.private_data = &rej_msg->private_data;
2453 }
2454
2455 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2456 {
2457         struct cm_timewait_info *timewait_info;
2458         struct cm_id_private *cm_id_priv;
2459         __be32 remote_id;
2460
2461         remote_id = rej_msg->local_comm_id;
2462
2463         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2464                 spin_lock_irq(&cm.lock);
2465                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2466                                                   remote_id);
2467                 if (!timewait_info) {
2468                         spin_unlock_irq(&cm.lock);
2469                         return NULL;
2470                 }
2471                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2472                                       (timewait_info->work.local_id ^
2473                                        cm.random_id_operand));
2474                 if (cm_id_priv) {
2475                         if (cm_id_priv->id.remote_id == remote_id)
2476                                 atomic_inc(&cm_id_priv->refcount);
2477                         else
2478                                 cm_id_priv = NULL;
2479                 }
2480                 spin_unlock_irq(&cm.lock);
2481         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2482                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2483         else
2484                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2485
2486         return cm_id_priv;
2487 }
2488
2489 static int cm_rej_handler(struct cm_work *work)
2490 {
2491         struct cm_id_private *cm_id_priv;
2492         struct cm_rej_msg *rej_msg;
2493         int ret;
2494
2495         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2496         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2497         if (!cm_id_priv)
2498                 return -EINVAL;
2499
2500         cm_format_rej_event(work);
2501
2502         spin_lock_irq(&cm_id_priv->lock);
2503         switch (cm_id_priv->id.state) {
2504         case IB_CM_REQ_SENT:
2505         case IB_CM_MRA_REQ_RCVD:
2506         case IB_CM_REP_SENT:
2507         case IB_CM_MRA_REP_RCVD:
2508                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2509                 /* fall through */
2510         case IB_CM_REQ_RCVD:
2511         case IB_CM_MRA_REQ_SENT:
2512                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2513                         cm_enter_timewait(cm_id_priv);
2514                 else
2515                         cm_reset_to_idle(cm_id_priv);
2516                 break;
2517         case IB_CM_DREQ_SENT:
2518                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2519                 /* fall through */
2520         case IB_CM_REP_RCVD:
2521         case IB_CM_MRA_REP_SENT:
2522                 cm_enter_timewait(cm_id_priv);
2523                 break;
2524         case IB_CM_ESTABLISHED:
2525                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2526                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2527                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2528                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2529                                               cm_id_priv->msg);
2530                         cm_enter_timewait(cm_id_priv);
2531                         break;
2532                 }
2533                 /* fall through */
2534         default:
2535                 spin_unlock_irq(&cm_id_priv->lock);
2536                 ret = -EINVAL;
2537                 goto out;
2538         }
2539
2540         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2541         if (!ret)
2542                 list_add_tail(&work->list, &cm_id_priv->work_list);
2543         spin_unlock_irq(&cm_id_priv->lock);
2544
2545         if (ret)
2546                 cm_process_work(cm_id_priv, work);
2547         else
2548                 cm_deref_id(cm_id_priv);
2549         return 0;
2550 out:
2551         cm_deref_id(cm_id_priv);
2552         return -EINVAL;
2553 }
2554
2555 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2556                    u8 service_timeout,
2557                    const void *private_data,
2558                    u8 private_data_len)
2559 {
2560         struct cm_id_private *cm_id_priv;
2561         struct ib_mad_send_buf *msg;
2562         enum ib_cm_state cm_state;
2563         enum ib_cm_lap_state lap_state;
2564         enum cm_msg_response msg_response;
2565         void *data;
2566         unsigned long flags;
2567         int ret;
2568
2569         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2570                 return -EINVAL;
2571
2572         data = cm_copy_private_data(private_data, private_data_len);
2573         if (IS_ERR(data))
2574                 return PTR_ERR(data);
2575
2576         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2577
2578         spin_lock_irqsave(&cm_id_priv->lock, flags);
2579         switch(cm_id_priv->id.state) {
2580         case IB_CM_REQ_RCVD:
2581                 cm_state = IB_CM_MRA_REQ_SENT;
2582                 lap_state = cm_id->lap_state;
2583                 msg_response = CM_MSG_RESPONSE_REQ;
2584                 break;
2585         case IB_CM_REP_RCVD:
2586                 cm_state = IB_CM_MRA_REP_SENT;
2587                 lap_state = cm_id->lap_state;
2588                 msg_response = CM_MSG_RESPONSE_REP;
2589                 break;
2590         case IB_CM_ESTABLISHED:
2591                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2592                         cm_state = cm_id->state;
2593                         lap_state = IB_CM_MRA_LAP_SENT;
2594                         msg_response = CM_MSG_RESPONSE_OTHER;
2595                         break;
2596                 }
2597         default:
2598                 ret = -EINVAL;
2599                 goto error1;
2600         }
2601
2602         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2603                 ret = cm_alloc_msg(cm_id_priv, &msg);
2604                 if (ret)
2605                         goto error1;
2606
2607                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2608                               msg_response, service_timeout,
2609                               private_data, private_data_len);
2610                 ret = ib_post_send_mad(msg, NULL);
2611                 if (ret)
2612                         goto error2;
2613         }
2614
2615         cm_id->state = cm_state;
2616         cm_id->lap_state = lap_state;
2617         cm_id_priv->service_timeout = service_timeout;
2618         cm_set_private_data(cm_id_priv, data, private_data_len);
2619         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2620         return 0;
2621
2622 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2623         kfree(data);
2624         return ret;
2625
2626 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2627         kfree(data);
2628         cm_free_msg(msg);
2629         return ret;
2630 }
2631 EXPORT_SYMBOL(ib_send_cm_mra);
2632
2633 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2634 {
2635         switch (cm_mra_get_msg_mraed(mra_msg)) {
2636         case CM_MSG_RESPONSE_REQ:
2637                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2638         case CM_MSG_RESPONSE_REP:
2639         case CM_MSG_RESPONSE_OTHER:
2640                 return cm_acquire_id(mra_msg->remote_comm_id,
2641                                      mra_msg->local_comm_id);
2642         default:
2643                 return NULL;
2644         }
2645 }
2646
2647 static int cm_mra_handler(struct cm_work *work)
2648 {
2649         struct cm_id_private *cm_id_priv;
2650         struct cm_mra_msg *mra_msg;
2651         int timeout, ret;
2652
2653         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2654         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2655         if (!cm_id_priv)
2656                 return -EINVAL;
2657
2658         work->cm_event.private_data = &mra_msg->private_data;
2659         work->cm_event.param.mra_rcvd.service_timeout =
2660                                         cm_mra_get_service_timeout(mra_msg);
2661         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2662                   cm_convert_to_ms(cm_id_priv->av.timeout);
2663
2664         spin_lock_irq(&cm_id_priv->lock);
2665         switch (cm_id_priv->id.state) {
2666         case IB_CM_REQ_SENT:
2667                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2668                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2669                                   cm_id_priv->msg, timeout))
2670                         goto out;
2671                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2672                 break;
2673         case IB_CM_REP_SENT:
2674                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2675                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2676                                   cm_id_priv->msg, timeout))
2677                         goto out;
2678                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2679                 break;
2680         case IB_CM_ESTABLISHED:
2681                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2682                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2683                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2684                                   cm_id_priv->msg, timeout)) {
2685                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2686                                 atomic_long_inc(&work->port->
2687                                                 counter_group[CM_RECV_DUPLICATES].
2688                                                 counter[CM_MRA_COUNTER]);
2689                         goto out;
2690                 }
2691                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2692                 break;
2693         case IB_CM_MRA_REQ_RCVD:
2694         case IB_CM_MRA_REP_RCVD:
2695                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2696                                 counter[CM_MRA_COUNTER]);
2697                 /* fall through */
2698         default:
2699                 goto out;
2700         }
2701
2702         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2703                                       cm_id_priv->id.state;
2704         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2705         if (!ret)
2706                 list_add_tail(&work->list, &cm_id_priv->work_list);
2707         spin_unlock_irq(&cm_id_priv->lock);
2708
2709         if (ret)
2710                 cm_process_work(cm_id_priv, work);
2711         else
2712                 cm_deref_id(cm_id_priv);
2713         return 0;
2714 out:
2715         spin_unlock_irq(&cm_id_priv->lock);
2716         cm_deref_id(cm_id_priv);
2717         return -EINVAL;
2718 }
2719
2720 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2721                           struct cm_id_private *cm_id_priv,
2722                           struct ib_sa_path_rec *alternate_path,
2723                           const void *private_data,
2724                           u8 private_data_len)
2725 {
2726         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2727                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2728         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2729         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2730         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2731         /* todo: need remote CM response timeout */
2732         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2733         lap_msg->alt_local_lid = alternate_path->slid;
2734         lap_msg->alt_remote_lid = alternate_path->dlid;
2735         lap_msg->alt_local_gid = alternate_path->sgid;
2736         lap_msg->alt_remote_gid = alternate_path->dgid;
2737         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2738         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2739         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2740         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2741         cm_lap_set_sl(lap_msg, alternate_path->sl);
2742         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2743         cm_lap_set_local_ack_timeout(lap_msg,
2744                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2745                                alternate_path->packet_life_time));
2746
2747         if (private_data && private_data_len)
2748                 memcpy(lap_msg->private_data, private_data, private_data_len);
2749 }
2750
2751 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2752                    struct ib_sa_path_rec *alternate_path,
2753                    const void *private_data,
2754                    u8 private_data_len)
2755 {
2756         struct cm_id_private *cm_id_priv;
2757         struct ib_mad_send_buf *msg;
2758         unsigned long flags;
2759         int ret;
2760
2761         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2762                 return -EINVAL;
2763
2764         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2765         spin_lock_irqsave(&cm_id_priv->lock, flags);
2766         if (cm_id->state != IB_CM_ESTABLISHED ||
2767             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2768              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2769                 ret = -EINVAL;
2770                 goto out;
2771         }
2772
2773         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av,
2774                                  cm_id_priv);
2775         if (ret)
2776                 goto out;
2777         cm_id_priv->alt_av.timeout =
2778                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2779                                        cm_id_priv->alt_av.timeout - 1);
2780
2781         ret = cm_alloc_msg(cm_id_priv, &msg);
2782         if (ret)
2783                 goto out;
2784
2785         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2786                       alternate_path, private_data, private_data_len);
2787         msg->timeout_ms = cm_id_priv->timeout_ms;
2788         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2789
2790         ret = ib_post_send_mad(msg, NULL);
2791         if (ret) {
2792                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2793                 cm_free_msg(msg);
2794                 return ret;
2795         }
2796
2797         cm_id->lap_state = IB_CM_LAP_SENT;
2798         cm_id_priv->msg = msg;
2799
2800 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2801         return ret;
2802 }
2803 EXPORT_SYMBOL(ib_send_cm_lap);
2804
2805 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2806                                     struct ib_sa_path_rec *path,
2807                                     struct cm_lap_msg *lap_msg)
2808 {
2809         memset(path, 0, sizeof *path);
2810         path->dgid = lap_msg->alt_local_gid;
2811         path->sgid = lap_msg->alt_remote_gid;
2812         path->dlid = lap_msg->alt_local_lid;
2813         path->slid = lap_msg->alt_remote_lid;
2814         path->flow_label = cm_lap_get_flow_label(lap_msg);
2815         path->hop_limit = lap_msg->alt_hop_limit;
2816         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2817         path->reversible = 1;
2818         path->pkey = cm_id_priv->pkey;
2819         path->sl = cm_lap_get_sl(lap_msg);
2820         path->mtu_selector = IB_SA_EQ;
2821         path->mtu = cm_id_priv->path_mtu;
2822         path->rate_selector = IB_SA_EQ;
2823         path->rate = cm_lap_get_packet_rate(lap_msg);
2824         path->packet_life_time_selector = IB_SA_EQ;
2825         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2826         path->packet_life_time -= (path->packet_life_time > 0);
2827 }
2828
2829 static int cm_lap_handler(struct cm_work *work)
2830 {
2831         struct cm_id_private *cm_id_priv;
2832         struct cm_lap_msg *lap_msg;
2833         struct ib_cm_lap_event_param *param;
2834         struct ib_mad_send_buf *msg = NULL;
2835         int ret;
2836
2837         /* todo: verify LAP request and send reject APR if invalid. */
2838         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2839         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2840                                    lap_msg->local_comm_id);
2841         if (!cm_id_priv)
2842                 return -EINVAL;
2843
2844         param = &work->cm_event.param.lap_rcvd;
2845         param->alternate_path = &work->path[0];
2846         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2847         work->cm_event.private_data = &lap_msg->private_data;
2848
2849         spin_lock_irq(&cm_id_priv->lock);
2850         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2851                 goto unlock;
2852
2853         switch (cm_id_priv->id.lap_state) {
2854         case IB_CM_LAP_UNINIT:
2855         case IB_CM_LAP_IDLE:
2856                 break;
2857         case IB_CM_MRA_LAP_SENT:
2858                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2859                                 counter[CM_LAP_COUNTER]);
2860                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2861                         goto unlock;
2862
2863                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2864                               CM_MSG_RESPONSE_OTHER,
2865                               cm_id_priv->service_timeout,
2866                               cm_id_priv->private_data,
2867                               cm_id_priv->private_data_len);
2868                 spin_unlock_irq(&cm_id_priv->lock);
2869
2870                 if (ib_post_send_mad(msg, NULL))
2871                         cm_free_msg(msg);
2872                 goto deref;
2873         case IB_CM_LAP_RCVD:
2874                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2875                                 counter[CM_LAP_COUNTER]);
2876                 goto unlock;
2877         default:
2878                 goto unlock;
2879         }
2880
2881         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2882         cm_id_priv->tid = lap_msg->hdr.tid;
2883         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2884                                 work->mad_recv_wc->recv_buf.grh,
2885                                 &cm_id_priv->av);
2886         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av,
2887                            cm_id_priv);
2888         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2889         if (!ret)
2890                 list_add_tail(&work->list, &cm_id_priv->work_list);
2891         spin_unlock_irq(&cm_id_priv->lock);
2892
2893         if (ret)
2894                 cm_process_work(cm_id_priv, work);
2895         else
2896                 cm_deref_id(cm_id_priv);
2897         return 0;
2898
2899 unlock: spin_unlock_irq(&cm_id_priv->lock);
2900 deref:  cm_deref_id(cm_id_priv);
2901         return -EINVAL;
2902 }
2903
2904 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2905                           struct cm_id_private *cm_id_priv,
2906                           enum ib_cm_apr_status status,
2907                           void *info,
2908                           u8 info_length,
2909                           const void *private_data,
2910                           u8 private_data_len)
2911 {
2912         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2913         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2914         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2915         apr_msg->ap_status = (u8) status;
2916
2917         if (info && info_length) {
2918                 apr_msg->info_length = info_length;
2919                 memcpy(apr_msg->info, info, info_length);
2920         }
2921
2922         if (private_data && private_data_len)
2923                 memcpy(apr_msg->private_data, private_data, private_data_len);
2924 }
2925
2926 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2927                    enum ib_cm_apr_status status,
2928                    void *info,
2929                    u8 info_length,
2930                    const void *private_data,
2931                    u8 private_data_len)
2932 {
2933         struct cm_id_private *cm_id_priv;
2934         struct ib_mad_send_buf *msg;
2935         unsigned long flags;
2936         int ret;
2937
2938         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2939             (info && info_length > IB_CM_APR_INFO_LENGTH))
2940                 return -EINVAL;
2941
2942         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2943         spin_lock_irqsave(&cm_id_priv->lock, flags);
2944         if (cm_id->state != IB_CM_ESTABLISHED ||
2945             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2946              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2947                 ret = -EINVAL;
2948                 goto out;
2949         }
2950
2951         ret = cm_alloc_msg(cm_id_priv, &msg);
2952         if (ret)
2953                 goto out;
2954
2955         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2956                       info, info_length, private_data, private_data_len);
2957         ret = ib_post_send_mad(msg, NULL);
2958         if (ret) {
2959                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2960                 cm_free_msg(msg);
2961                 return ret;
2962         }
2963
2964         cm_id->lap_state = IB_CM_LAP_IDLE;
2965 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2966         return ret;
2967 }
2968 EXPORT_SYMBOL(ib_send_cm_apr);
2969
2970 static int cm_apr_handler(struct cm_work *work)
2971 {
2972         struct cm_id_private *cm_id_priv;
2973         struct cm_apr_msg *apr_msg;
2974         int ret;
2975
2976         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2977         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2978                                    apr_msg->local_comm_id);
2979         if (!cm_id_priv)
2980                 return -EINVAL; /* Unmatched reply. */
2981
2982         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2983         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2984         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2985         work->cm_event.private_data = &apr_msg->private_data;
2986
2987         spin_lock_irq(&cm_id_priv->lock);
2988         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2989             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2990              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2991                 spin_unlock_irq(&cm_id_priv->lock);
2992                 goto out;
2993         }
2994         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2995         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2996         cm_id_priv->msg = NULL;
2997
2998         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2999         if (!ret)
3000                 list_add_tail(&work->list, &cm_id_priv->work_list);
3001         spin_unlock_irq(&cm_id_priv->lock);
3002
3003         if (ret)
3004                 cm_process_work(cm_id_priv, work);
3005         else
3006                 cm_deref_id(cm_id_priv);
3007         return 0;
3008 out:
3009         cm_deref_id(cm_id_priv);
3010         return -EINVAL;
3011 }
3012
3013 static int cm_timewait_handler(struct cm_work *work)
3014 {
3015         struct cm_timewait_info *timewait_info;
3016         struct cm_id_private *cm_id_priv;
3017         int ret;
3018
3019         timewait_info = (struct cm_timewait_info *)work;
3020         spin_lock_irq(&cm.lock);
3021         list_del(&timewait_info->list);
3022         spin_unlock_irq(&cm.lock);
3023
3024         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3025                                    timewait_info->work.remote_id);
3026         if (!cm_id_priv)
3027                 return -EINVAL;
3028
3029         spin_lock_irq(&cm_id_priv->lock);
3030         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3031             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3032                 spin_unlock_irq(&cm_id_priv->lock);
3033                 goto out;
3034         }
3035         cm_id_priv->id.state = IB_CM_IDLE;
3036         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3037         if (!ret)
3038                 list_add_tail(&work->list, &cm_id_priv->work_list);
3039         spin_unlock_irq(&cm_id_priv->lock);
3040
3041         if (ret)
3042                 cm_process_work(cm_id_priv, work);
3043         else
3044                 cm_deref_id(cm_id_priv);
3045         return 0;
3046 out:
3047         cm_deref_id(cm_id_priv);
3048         return -EINVAL;
3049 }
3050
3051 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3052                                struct cm_id_private *cm_id_priv,
3053                                struct ib_cm_sidr_req_param *param)
3054 {
3055         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3056                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
3057         sidr_req_msg->request_id = cm_id_priv->id.local_id;
3058         sidr_req_msg->pkey = param->path->pkey;
3059         sidr_req_msg->service_id = param->service_id;
3060
3061         if (param->private_data && param->private_data_len)
3062                 memcpy(sidr_req_msg->private_data, param->private_data,
3063                        param->private_data_len);
3064 }
3065
3066 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3067                         struct ib_cm_sidr_req_param *param)
3068 {
3069         struct cm_id_private *cm_id_priv;
3070         struct ib_mad_send_buf *msg;
3071         unsigned long flags;
3072         int ret;
3073
3074         if (!param->path || (param->private_data &&
3075              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3076                 return -EINVAL;
3077
3078         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3079         ret = cm_init_av_by_path(param->path, &cm_id_priv->av, cm_id_priv);
3080         if (ret)
3081                 goto out;
3082
3083         cm_id->service_id = param->service_id;
3084         cm_id->service_mask = ~cpu_to_be64(0);
3085         cm_id_priv->timeout_ms = param->timeout_ms;
3086         cm_id_priv->max_cm_retries = param->max_cm_retries;
3087         ret = cm_alloc_msg(cm_id_priv, &msg);
3088         if (ret)
3089                 goto out;
3090
3091         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3092                            param);
3093         msg->timeout_ms = cm_id_priv->timeout_ms;
3094         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3095
3096         spin_lock_irqsave(&cm_id_priv->lock, flags);
3097         if (cm_id->state == IB_CM_IDLE)
3098                 ret = ib_post_send_mad(msg, NULL);
3099         else
3100                 ret = -EINVAL;
3101
3102         if (ret) {
3103                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3104                 cm_free_msg(msg);
3105                 goto out;
3106         }
3107         cm_id->state = IB_CM_SIDR_REQ_SENT;
3108         cm_id_priv->msg = msg;
3109         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3110 out:
3111         return ret;
3112 }
3113 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3114
3115 static void cm_format_sidr_req_event(struct cm_work *work,
3116                                      struct ib_cm_id *listen_id)
3117 {
3118         struct cm_sidr_req_msg *sidr_req_msg;
3119         struct ib_cm_sidr_req_event_param *param;
3120
3121         sidr_req_msg = (struct cm_sidr_req_msg *)
3122                                 work->mad_recv_wc->recv_buf.mad;
3123         param = &work->cm_event.param.sidr_req_rcvd;
3124         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
3125         param->listen_id = listen_id;
3126         param->service_id = sidr_req_msg->service_id;
3127         param->bth_pkey = cm_get_bth_pkey(work);
3128         param->port = work->port->port_num;
3129         work->cm_event.private_data = &sidr_req_msg->private_data;
3130 }
3131
3132 static int cm_sidr_req_handler(struct cm_work *work)
3133 {
3134         struct ib_cm_id *cm_id;
3135         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3136         struct cm_sidr_req_msg *sidr_req_msg;
3137         struct ib_wc *wc;
3138
3139         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3140         if (IS_ERR(cm_id))
3141                 return PTR_ERR(cm_id);
3142         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3143
3144         /* Record SGID/SLID and request ID for lookup. */
3145         sidr_req_msg = (struct cm_sidr_req_msg *)
3146                                 work->mad_recv_wc->recv_buf.mad;
3147         wc = work->mad_recv_wc->wc;
3148         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3149         cm_id_priv->av.dgid.global.interface_id = 0;
3150         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3151                                 work->mad_recv_wc->recv_buf.grh,
3152                                 &cm_id_priv->av);
3153         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3154         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3155         atomic_inc(&cm_id_priv->work_count);
3156
3157         spin_lock_irq(&cm.lock);
3158         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3159         if (cur_cm_id_priv) {
3160                 spin_unlock_irq(&cm.lock);
3161                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3162                                 counter[CM_SIDR_REQ_COUNTER]);
3163                 goto out; /* Duplicate message. */
3164         }
3165         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3166         cur_cm_id_priv = cm_find_listen(cm_id->device,
3167                                         sidr_req_msg->service_id);
3168         if (!cur_cm_id_priv) {
3169                 spin_unlock_irq(&cm.lock);
3170                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3171                 goto out; /* No match. */
3172         }
3173         atomic_inc(&cur_cm_id_priv->refcount);
3174         atomic_inc(&cm_id_priv->refcount);
3175         spin_unlock_irq(&cm.lock);
3176
3177         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3178         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3179         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3180         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3181
3182         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3183         cm_process_work(cm_id_priv, work);
3184         cm_deref_id(cur_cm_id_priv);
3185         return 0;
3186 out:
3187         ib_destroy_cm_id(&cm_id_priv->id);
3188         return -EINVAL;
3189 }
3190
3191 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3192                                struct cm_id_private *cm_id_priv,
3193                                struct ib_cm_sidr_rep_param *param)
3194 {
3195         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3196                           cm_id_priv->tid);
3197         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3198         sidr_rep_msg->status = param->status;
3199         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3200         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3201         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3202
3203         if (param->info && param->info_length)
3204                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3205
3206         if (param->private_data && param->private_data_len)
3207                 memcpy(sidr_rep_msg->private_data, param->private_data,
3208                        param->private_data_len);
3209 }
3210
3211 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3212                         struct ib_cm_sidr_rep_param *param)
3213 {
3214         struct cm_id_private *cm_id_priv;
3215         struct ib_mad_send_buf *msg;
3216         unsigned long flags;
3217         int ret;
3218
3219         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3220             (param->private_data &&
3221              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3222                 return -EINVAL;
3223
3224         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3225         spin_lock_irqsave(&cm_id_priv->lock, flags);
3226         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3227                 ret = -EINVAL;
3228                 goto error;
3229         }
3230
3231         ret = cm_alloc_msg(cm_id_priv, &msg);
3232         if (ret)
3233                 goto error;
3234
3235         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3236                            param);
3237         ret = ib_post_send_mad(msg, NULL);
3238         if (ret) {
3239                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3240                 cm_free_msg(msg);
3241                 return ret;
3242         }
3243         cm_id->state = IB_CM_IDLE;
3244         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3245
3246         spin_lock_irqsave(&cm.lock, flags);
3247         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3248                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3249                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3250         }
3251         spin_unlock_irqrestore(&cm.lock, flags);
3252         return 0;
3253
3254 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3255         return ret;
3256 }
3257 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3258
3259 static void cm_format_sidr_rep_event(struct cm_work *work)
3260 {
3261         struct cm_sidr_rep_msg *sidr_rep_msg;
3262         struct ib_cm_sidr_rep_event_param *param;
3263
3264         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3265                                 work->mad_recv_wc->recv_buf.mad;
3266         param = &work->cm_event.param.sidr_rep_rcvd;
3267         param->status = sidr_rep_msg->status;
3268         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3269         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3270         param->info = &sidr_rep_msg->info;
3271         param->info_len = sidr_rep_msg->info_length;
3272         work->cm_event.private_data = &sidr_rep_msg->private_data;
3273 }
3274
3275 static int cm_sidr_rep_handler(struct cm_work *work)
3276 {
3277         struct cm_sidr_rep_msg *sidr_rep_msg;
3278         struct cm_id_private *cm_id_priv;
3279
3280         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3281                                 work->mad_recv_wc->recv_buf.mad;
3282         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3283         if (!cm_id_priv)
3284                 return -EINVAL; /* Unmatched reply. */
3285
3286         spin_lock_irq(&cm_id_priv->lock);
3287         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3288                 spin_unlock_irq(&cm_id_priv->lock);
3289                 goto out;
3290         }
3291         cm_id_priv->id.state = IB_CM_IDLE;
3292         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3293         spin_unlock_irq(&cm_id_priv->lock);
3294
3295         cm_format_sidr_rep_event(work);
3296         cm_process_work(cm_id_priv, work);
3297         return 0;
3298 out:
3299         cm_deref_id(cm_id_priv);
3300         return -EINVAL;
3301 }
3302
3303 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3304                                   enum ib_wc_status wc_status)
3305 {
3306         struct cm_id_private *cm_id_priv;
3307         struct ib_cm_event cm_event;
3308         enum ib_cm_state state;
3309         int ret;
3310
3311         memset(&cm_event, 0, sizeof cm_event);
3312         cm_id_priv = msg->context[0];
3313
3314         /* Discard old sends or ones without a response. */
3315         spin_lock_irq(&cm_id_priv->lock);
3316         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3317         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3318                 goto discard;
3319
3320         switch (state) {
3321         case IB_CM_REQ_SENT:
3322         case IB_CM_MRA_REQ_RCVD:
3323                 cm_reset_to_idle(cm_id_priv);
3324                 cm_event.event = IB_CM_REQ_ERROR;
3325                 break;
3326         case IB_CM_REP_SENT:
3327         case IB_CM_MRA_REP_RCVD:
3328                 cm_reset_to_idle(cm_id_priv);
3329                 cm_event.event = IB_CM_REP_ERROR;
3330                 break;
3331         case IB_CM_DREQ_SENT:
3332                 cm_enter_timewait(cm_id_priv);
3333                 cm_event.event = IB_CM_DREQ_ERROR;
3334                 break;
3335         case IB_CM_SIDR_REQ_SENT:
3336                 cm_id_priv->id.state = IB_CM_IDLE;
3337                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3338                 break;
3339         default:
3340                 goto discard;
3341         }
3342         spin_unlock_irq(&cm_id_priv->lock);
3343         cm_event.param.send_status = wc_status;
3344
3345         /* No other events can occur on the cm_id at this point. */
3346         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3347         cm_free_msg(msg);
3348         if (ret)
3349                 ib_destroy_cm_id(&cm_id_priv->id);
3350         return;
3351 discard:
3352         spin_unlock_irq(&cm_id_priv->lock);
3353         cm_free_msg(msg);
3354 }
3355
3356 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3357                             struct ib_mad_send_wc *mad_send_wc)
3358 {
3359         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3360         struct cm_port *port;
3361         u16 attr_index;
3362
3363         port = mad_agent->context;
3364         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3365                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3366
3367         /*
3368          * If the send was in response to a received message (context[0] is not
3369          * set to a cm_id), and is not a REJ, then it is a send that was
3370          * manually retried.
3371          */
3372         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3373                 msg->retries = 1;
3374
3375         atomic_long_add(1 + msg->retries,
3376                         &port->counter_group[CM_XMIT].counter[attr_index]);
3377         if (msg->retries)
3378                 atomic_long_add(msg->retries,
3379                                 &port->counter_group[CM_XMIT_RETRIES].
3380                                 counter[attr_index]);
3381
3382         switch (mad_send_wc->status) {
3383         case IB_WC_SUCCESS:
3384         case IB_WC_WR_FLUSH_ERR:
3385                 cm_free_msg(msg);
3386                 break;
3387         default:
3388                 if (msg->context[0] && msg->context[1])
3389                         cm_process_send_error(msg, mad_send_wc->status);
3390                 else
3391                         cm_free_msg(msg);
3392                 break;
3393         }
3394 }
3395
3396 static void cm_work_handler(struct work_struct *_work)
3397 {
3398         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3399         int ret;
3400
3401         switch (work->cm_event.event) {
3402         case IB_CM_REQ_RECEIVED:
3403                 ret = cm_req_handler(work);
3404                 break;
3405         case IB_CM_MRA_RECEIVED:
3406                 ret = cm_mra_handler(work);
3407                 break;
3408         case IB_CM_REJ_RECEIVED:
3409                 ret = cm_rej_handler(work);
3410                 break;
3411         case IB_CM_REP_RECEIVED:
3412                 ret = cm_rep_handler(work);
3413                 break;
3414         case IB_CM_RTU_RECEIVED:
3415                 ret = cm_rtu_handler(work);
3416                 break;
3417         case IB_CM_USER_ESTABLISHED:
3418                 ret = cm_establish_handler(work);
3419                 break;
3420         case IB_CM_DREQ_RECEIVED:
3421                 ret = cm_dreq_handler(work);
3422                 break;
3423         case IB_CM_DREP_RECEIVED:
3424                 ret = cm_drep_handler(work);
3425                 break;
3426         case IB_CM_SIDR_REQ_RECEIVED:
3427                 ret = cm_sidr_req_handler(work);
3428                 break;
3429         case IB_CM_SIDR_REP_RECEIVED:
3430                 ret = cm_sidr_rep_handler(work);
3431                 break;
3432         case IB_CM_LAP_RECEIVED:
3433                 ret = cm_lap_handler(work);
3434                 break;
3435         case IB_CM_APR_RECEIVED:
3436                 ret = cm_apr_handler(work);
3437                 break;
3438         case IB_CM_TIMEWAIT_EXIT:
3439                 ret = cm_timewait_handler(work);
3440                 break;
3441         default:
3442                 ret = -EINVAL;
3443                 break;
3444         }
3445         if (ret)
3446                 cm_free_work(work);
3447 }
3448
3449 static int cm_establish(struct ib_cm_id *cm_id)
3450 {
3451         struct cm_id_private *cm_id_priv;
3452         struct cm_work *work;
3453         unsigned long flags;
3454         int ret = 0;
3455         struct cm_device *cm_dev;
3456
3457         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3458         if (!cm_dev)
3459                 return -ENODEV;
3460
3461         work = kmalloc(sizeof *work, GFP_ATOMIC);
3462         if (!work)
3463                 return -ENOMEM;
3464
3465         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3466         spin_lock_irqsave(&cm_id_priv->lock, flags);
3467         switch (cm_id->state)
3468         {
3469         case IB_CM_REP_SENT:
3470         case IB_CM_MRA_REP_RCVD:
3471                 cm_id->state = IB_CM_ESTABLISHED;
3472                 break;
3473         case IB_CM_ESTABLISHED:
3474                 ret = -EISCONN;
3475                 break;
3476         default:
3477                 ret = -EINVAL;
3478                 break;
3479         }
3480         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3481
3482         if (ret) {
3483                 kfree(work);
3484                 goto out;
3485         }
3486
3487         /*
3488          * The CM worker thread may try to destroy the cm_id before it
3489          * can execute this work item.  To prevent potential deadlock,
3490          * we need to find the cm_id once we're in the context of the
3491          * worker thread, rather than holding a reference on it.
3492          */
3493         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3494         work->local_id = cm_id->local_id;
3495         work->remote_id = cm_id->remote_id;
3496         work->mad_recv_wc = NULL;
3497         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3498
3499         /* Check if the device started its remove_one */
3500         spin_lock_irqsave(&cm.lock, flags);
3501         if (!cm_dev->going_down) {
3502                 queue_delayed_work(cm.wq, &work->work, 0);
3503         } else {
3504                 kfree(work);
3505                 ret = -ENODEV;
3506         }
3507         spin_unlock_irqrestore(&cm.lock, flags);
3508
3509 out:
3510         return ret;
3511 }
3512
3513 static int cm_migrate(struct ib_cm_id *cm_id)
3514 {
3515         struct cm_id_private *cm_id_priv;
3516         struct cm_av tmp_av;
3517         unsigned long flags;
3518         int tmp_send_port_not_ready;
3519         int ret = 0;
3520
3521         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3522         spin_lock_irqsave(&cm_id_priv->lock, flags);
3523         if (cm_id->state == IB_CM_ESTABLISHED &&
3524             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3525              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3526                 cm_id->lap_state = IB_CM_LAP_IDLE;
3527                 /* Swap address vector */
3528                 tmp_av = cm_id_priv->av;
3529                 cm_id_priv->av = cm_id_priv->alt_av;
3530                 cm_id_priv->alt_av = tmp_av;
3531                 /* Swap port send ready state */
3532                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3533                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3534                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3535         } else
3536                 ret = -EINVAL;
3537         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3538
3539         return ret;
3540 }
3541
3542 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3543 {
3544         int ret;
3545
3546         switch (event) {
3547         case IB_EVENT_COMM_EST:
3548                 ret = cm_establish(cm_id);
3549                 break;
3550         case IB_EVENT_PATH_MIG:
3551                 ret = cm_migrate(cm_id);
3552                 break;
3553         default:
3554                 ret = -EINVAL;
3555         }
3556         return ret;
3557 }
3558 EXPORT_SYMBOL(ib_cm_notify);
3559
3560 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3561                             struct ib_mad_recv_wc *mad_recv_wc)
3562 {
3563         struct cm_port *port = mad_agent->context;
3564         struct cm_work *work;
3565         enum ib_cm_event_type event;
3566         u16 attr_id;
3567         int paths = 0;
3568         int going_down = 0;
3569
3570         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3571         case CM_REQ_ATTR_ID:
3572                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3573                                                     alt_local_lid != 0);
3574                 event = IB_CM_REQ_RECEIVED;
3575                 break;
3576         case CM_MRA_ATTR_ID:
3577                 event = IB_CM_MRA_RECEIVED;
3578                 break;
3579         case CM_REJ_ATTR_ID:
3580                 event = IB_CM_REJ_RECEIVED;
3581                 break;
3582         case CM_REP_ATTR_ID:
3583                 event = IB_CM_REP_RECEIVED;
3584                 break;
3585         case CM_RTU_ATTR_ID:
3586                 event = IB_CM_RTU_RECEIVED;
3587                 break;
3588         case CM_DREQ_ATTR_ID:
3589                 event = IB_CM_DREQ_RECEIVED;
3590                 break;
3591         case CM_DREP_ATTR_ID:
3592                 event = IB_CM_DREP_RECEIVED;
3593                 break;
3594         case CM_SIDR_REQ_ATTR_ID:
3595                 event = IB_CM_SIDR_REQ_RECEIVED;
3596                 break;
3597         case CM_SIDR_REP_ATTR_ID:
3598                 event = IB_CM_SIDR_REP_RECEIVED;
3599                 break;
3600         case CM_LAP_ATTR_ID:
3601                 paths = 1;
3602                 event = IB_CM_LAP_RECEIVED;
3603                 break;
3604         case CM_APR_ATTR_ID:
3605                 event = IB_CM_APR_RECEIVED;
3606                 break;
3607         default:
3608                 ib_free_recv_mad(mad_recv_wc);
3609                 return;
3610         }
3611
3612         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3613         atomic_long_inc(&port->counter_group[CM_RECV].
3614                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3615
3616         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3617                        GFP_KERNEL);
3618         if (!work) {
3619                 ib_free_recv_mad(mad_recv_wc);
3620                 return;
3621         }
3622
3623         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3624         work->cm_event.event = event;
3625         work->mad_recv_wc = mad_recv_wc;
3626         work->port = port;
3627
3628         /* Check if the device started its remove_one */
3629         spin_lock_irq(&cm.lock);
3630         if (!port->cm_dev->going_down)
3631                 queue_delayed_work(cm.wq, &work->work, 0);
3632         else
3633                 going_down = 1;
3634         spin_unlock_irq(&cm.lock);
3635
3636         if (going_down) {
3637                 kfree(work);
3638                 ib_free_recv_mad(mad_recv_wc);
3639         }
3640 }
3641
3642 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3643                                 struct ib_qp_attr *qp_attr,
3644                                 int *qp_attr_mask)
3645 {
3646         unsigned long flags;
3647         int ret;
3648
3649         spin_lock_irqsave(&cm_id_priv->lock, flags);
3650         switch (cm_id_priv->id.state) {
3651         case IB_CM_REQ_SENT:
3652         case IB_CM_MRA_REQ_RCVD:
3653         case IB_CM_REQ_RCVD:
3654         case IB_CM_MRA_REQ_SENT:
3655         case IB_CM_REP_RCVD:
3656         case IB_CM_MRA_REP_SENT:
3657         case IB_CM_REP_SENT:
3658         case IB_CM_MRA_REP_RCVD:
3659         case IB_CM_ESTABLISHED:
3660                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3661                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3662                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3663                 if (cm_id_priv->responder_resources)
3664                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3665                                                     IB_ACCESS_REMOTE_ATOMIC;
3666                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3667                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3668                 ret = 0;
3669                 break;
3670         default:
3671                 ret = -EINVAL;
3672                 break;
3673         }
3674         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3675         return ret;
3676 }
3677
3678 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3679                                struct ib_qp_attr *qp_attr,
3680                                int *qp_attr_mask)
3681 {
3682         unsigned long flags;
3683         int ret;
3684
3685         spin_lock_irqsave(&cm_id_priv->lock, flags);
3686         switch (cm_id_priv->id.state) {
3687         case IB_CM_REQ_RCVD:
3688         case IB_CM_MRA_REQ_SENT:
3689         case IB_CM_REP_RCVD:
3690         case IB_CM_MRA_REP_SENT:
3691         case IB_CM_REP_SENT:
3692         case IB_CM_MRA_REP_RCVD:
3693         case IB_CM_ESTABLISHED:
3694                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3695                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3696                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3697                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3698                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3699                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3700                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3701                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3702                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3703                                          IB_QP_MIN_RNR_TIMER;
3704                         qp_attr->max_dest_rd_atomic =
3705                                         cm_id_priv->responder_resources;
3706                         qp_attr->min_rnr_timer = 0;
3707                 }
3708                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3709                         *qp_attr_mask |= IB_QP_ALT_PATH;
3710                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3711                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3712                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3713                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3714                 }
3715                 ret = 0;
3716                 break;
3717         default:
3718                 ret = -EINVAL;
3719                 break;
3720         }
3721         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3722         return ret;
3723 }
3724
3725 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3726                                struct ib_qp_attr *qp_attr,
3727                                int *qp_attr_mask)
3728 {
3729         unsigned long flags;
3730         int ret;
3731
3732         spin_lock_irqsave(&cm_id_priv->lock, flags);
3733         switch (cm_id_priv->id.state) {
3734         /* Allow transition to RTS before sending REP */
3735         case IB_CM_REQ_RCVD:
3736         case IB_CM_MRA_REQ_SENT:
3737
3738         case IB_CM_REP_RCVD:
3739         case IB_CM_MRA_REP_SENT:
3740         case IB_CM_REP_SENT:
3741         case IB_CM_MRA_REP_RCVD:
3742         case IB_CM_ESTABLISHED:
3743                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3744                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3745                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3746                         switch (cm_id_priv->qp_type) {
3747                         case IB_QPT_RC:
3748                         case IB_QPT_XRC_INI:
3749                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3750                                                  IB_QP_MAX_QP_RD_ATOMIC;
3751                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3752                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3753                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3754                                 /* fall through */
3755                         case IB_QPT_XRC_TGT:
3756                                 *qp_attr_mask |= IB_QP_TIMEOUT;
3757                                 qp_attr->timeout = cm_id_priv->av.timeout;
3758                                 break;
3759                         default:
3760                                 break;
3761                         }
3762                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3763                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3764                                 qp_attr->path_mig_state = IB_MIG_REARM;
3765                         }
3766                 } else {
3767                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3768                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3769                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3770                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3771                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3772                         qp_attr->path_mig_state = IB_MIG_REARM;
3773                 }
3774                 ret = 0;
3775                 break;
3776         default:
3777                 ret = -EINVAL;
3778                 break;
3779         }
3780         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3781         return ret;
3782 }
3783
3784 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3785                        struct ib_qp_attr *qp_attr,
3786                        int *qp_attr_mask)
3787 {
3788         struct cm_id_private *cm_id_priv;
3789         int ret;
3790
3791         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3792         switch (qp_attr->qp_state) {
3793         case IB_QPS_INIT:
3794                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3795                 break;
3796         case IB_QPS_RTR:
3797                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3798                 break;
3799         case IB_QPS_RTS:
3800                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3801                 break;
3802         default:
3803                 ret = -EINVAL;
3804                 break;
3805         }
3806         return ret;
3807 }
3808 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3809
3810 static void cm_get_ack_delay(struct cm_device *cm_dev)
3811 {
3812         struct ib_device_attr attr;
3813
3814         if (ib_query_device(cm_dev->ib_device, &attr))
3815                 cm_dev->ack_delay = 0; /* acks will rely on packet life time */
3816         else
3817                 cm_dev->ack_delay = attr.local_ca_ack_delay;
3818 }
3819
3820 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3821                                char *buf)
3822 {
3823         struct cm_counter_group *group;
3824         struct cm_counter_attribute *cm_attr;
3825
3826         group = container_of(obj, struct cm_counter_group, obj);
3827         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3828
3829         return sprintf(buf, "%ld\n",
3830                        atomic_long_read(&group->counter[cm_attr->index]));
3831 }
3832
3833 static const struct sysfs_ops cm_counter_ops = {
3834         .show = cm_show_counter
3835 };
3836
3837 static struct kobj_type cm_counter_obj_type = {
3838         .sysfs_ops = &cm_counter_ops,
3839         .default_attrs = cm_counter_default_attrs
3840 };
3841
3842 static void cm_release_port_obj(struct kobject *obj)
3843 {
3844         struct cm_port *cm_port;
3845
3846         cm_port = container_of(obj, struct cm_port, port_obj);
3847         kfree(cm_port);
3848 }
3849
3850 static struct kobj_type cm_port_obj_type = {
3851         .release = cm_release_port_obj
3852 };
3853
3854 static char *cm_devnode(struct device *dev, umode_t *mode)
3855 {
3856         if (mode)
3857                 *mode = 0666;
3858         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3859 }
3860
3861 struct class cm_class = {
3862         .owner   = THIS_MODULE,
3863         .name    = "infiniband_cm",
3864         .devnode = cm_devnode,
3865 };
3866 EXPORT_SYMBOL(cm_class);
3867
3868 static int cm_create_port_fs(struct cm_port *port)
3869 {
3870         int i, ret;
3871
3872         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3873                                    &port->cm_dev->device->kobj,
3874                                    "%d", port->port_num);
3875         if (ret) {
3876                 kfree(port);
3877                 return ret;
3878         }
3879
3880         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3881                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3882                                            &cm_counter_obj_type,
3883                                            &port->port_obj,
3884                                            "%s", counter_group_names[i]);
3885                 if (ret)
3886                         goto error;
3887         }
3888
3889         return 0;
3890
3891 error:
3892         while (i--)
3893                 kobject_put(&port->counter_group[i].obj);
3894         kobject_put(&port->port_obj);
3895         return ret;
3896
3897 }
3898
3899 static void cm_remove_port_fs(struct cm_port *port)
3900 {
3901         int i;
3902
3903         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3904                 kobject_put(&port->counter_group[i].obj);
3905
3906         kobject_put(&port->port_obj);
3907 }
3908
3909 static void cm_add_one(struct ib_device *ib_device)
3910 {
3911         struct cm_device *cm_dev;
3912         struct cm_port *port;
3913         struct ib_mad_reg_req reg_req = {
3914                 .mgmt_class = IB_MGMT_CLASS_CM,
3915                 .mgmt_class_version = IB_CM_CLASS_VERSION,
3916         };
3917         struct ib_port_modify port_modify = {
3918                 .set_port_cap_mask = IB_PORT_CM_SUP
3919         };
3920         unsigned long flags;
3921         int ret;
3922         int count = 0;
3923         u8 i;
3924
3925         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3926                          ib_device->phys_port_cnt, GFP_KERNEL);
3927         if (!cm_dev)
3928                 return;
3929
3930         cm_dev->ib_device = ib_device;
3931         cm_get_ack_delay(cm_dev);
3932         cm_dev->going_down = 0;
3933         cm_dev->device = device_create(&cm_class, &ib_device->dev,
3934                                        MKDEV(0, 0), NULL,
3935                                        "%s", ib_device->name);
3936         if (IS_ERR(cm_dev->device)) {
3937                 kfree(cm_dev);
3938                 return;
3939         }
3940
3941         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3942         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3943                 if (!rdma_cap_ib_cm(ib_device, i))
3944                         continue;
3945
3946                 port = kzalloc(sizeof *port, GFP_KERNEL);
3947                 if (!port)
3948                         goto error1;
3949
3950                 cm_dev->port[i-1] = port;
3951                 port->cm_dev = cm_dev;
3952                 port->port_num = i;
3953
3954                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
3955                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
3956
3957                 ret = cm_create_port_fs(port);
3958                 if (ret)
3959                         goto error1;
3960
3961                 port->mad_agent = ib_register_mad_agent(ib_device, i,
3962                                                         IB_QPT_GSI,
3963                                                         &reg_req,
3964                                                         0,
3965                                                         cm_send_handler,
3966                                                         cm_recv_handler,
3967                                                         port,
3968                                                         0);
3969                 if (IS_ERR(port->mad_agent))
3970                         goto error2;
3971
3972                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
3973                 if (ret)
3974                         goto error3;
3975
3976                 count++;
3977         }
3978
3979         if (!count)
3980                 goto free;
3981
3982         ib_set_client_data(ib_device, &cm_client, cm_dev);
3983
3984         write_lock_irqsave(&cm.device_lock, flags);
3985         list_add_tail(&cm_dev->list, &cm.device_list);
3986         write_unlock_irqrestore(&cm.device_lock, flags);
3987         return;
3988
3989 error3:
3990         ib_unregister_mad_agent(port->mad_agent);
3991 error2:
3992         cm_remove_port_fs(port);
3993 error1:
3994         port_modify.set_port_cap_mask = 0;
3995         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3996         while (--i) {
3997                 if (!rdma_cap_ib_cm(ib_device, i))
3998                         continue;
3999
4000                 port = cm_dev->port[i-1];
4001                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4002                 ib_unregister_mad_agent(port->mad_agent);
4003                 cm_remove_port_fs(port);
4004         }
4005 free:
4006         device_unregister(cm_dev->device);
4007         kfree(cm_dev);
4008 }
4009
4010 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4011 {
4012         struct cm_device *cm_dev = client_data;
4013         struct cm_port *port;
4014         struct cm_id_private *cm_id_priv;
4015         struct ib_mad_agent *cur_mad_agent;
4016         struct ib_port_modify port_modify = {
4017                 .clr_port_cap_mask = IB_PORT_CM_SUP
4018         };
4019         unsigned long flags;
4020         int i;
4021
4022         if (!cm_dev)
4023                 return;
4024
4025         write_lock_irqsave(&cm.device_lock, flags);
4026         list_del(&cm_dev->list);
4027         write_unlock_irqrestore(&cm.device_lock, flags);
4028
4029         spin_lock_irq(&cm.lock);
4030         cm_dev->going_down = 1;
4031         spin_unlock_irq(&cm.lock);
4032
4033         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4034                 if (!rdma_cap_ib_cm(ib_device, i))
4035                         continue;
4036
4037                 port = cm_dev->port[i-1];
4038                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4039                 /* Mark all the cm_id's as not valid */
4040                 spin_lock_irq(&cm.lock);
4041                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4042                         cm_id_priv->altr_send_port_not_ready = 1;
4043                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4044                         cm_id_priv->prim_send_port_not_ready = 1;
4045                 spin_unlock_irq(&cm.lock);
4046                 /*
4047                  * We flush the queue here after the going_down set, this
4048                  * verify that no new works will be queued in the recv handler,
4049                  * after that we can call the unregister_mad_agent
4050                  */
4051                 flush_workqueue(cm.wq);
4052                 spin_lock_irq(&cm.state_lock);
4053                 cur_mad_agent = port->mad_agent;
4054                 port->mad_agent = NULL;
4055                 spin_unlock_irq(&cm.state_lock);
4056                 ib_unregister_mad_agent(cur_mad_agent);
4057                 cm_remove_port_fs(port);
4058         }
4059
4060         device_unregister(cm_dev->device);
4061         kfree(cm_dev);
4062 }
4063
4064 static int __init ib_cm_init(void)
4065 {
4066         int ret;
4067
4068         memset(&cm, 0, sizeof cm);
4069         INIT_LIST_HEAD(&cm.device_list);
4070         rwlock_init(&cm.device_lock);
4071         spin_lock_init(&cm.lock);
4072         spin_lock_init(&cm.state_lock);
4073         cm.listen_service_table = RB_ROOT;
4074         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4075         cm.remote_id_table = RB_ROOT;
4076         cm.remote_qp_table = RB_ROOT;
4077         cm.remote_sidr_table = RB_ROOT;
4078         idr_init(&cm.local_id_table);
4079         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4080         INIT_LIST_HEAD(&cm.timewait_list);
4081
4082         ret = class_register(&cm_class);
4083         if (ret) {
4084                 ret = -ENOMEM;
4085                 goto error1;
4086         }
4087
4088         cm.wq = create_workqueue("ib_cm");
4089         if (!cm.wq) {
4090                 ret = -ENOMEM;
4091                 goto error2;
4092         }
4093
4094         ret = ib_register_client(&cm_client);
4095         if (ret)
4096                 goto error3;
4097
4098         return 0;
4099 error3:
4100         destroy_workqueue(cm.wq);
4101 error2:
4102         class_unregister(&cm_class);
4103 error1:
4104         idr_destroy(&cm.local_id_table);
4105         return ret;
4106 }
4107
4108 static void __exit ib_cm_cleanup(void)
4109 {
4110         struct cm_timewait_info *timewait_info, *tmp;
4111
4112         spin_lock_irq(&cm.lock);
4113         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4114                 cancel_delayed_work(&timewait_info->work.work);
4115         spin_unlock_irq(&cm.lock);
4116
4117         ib_unregister_client(&cm_client);
4118         destroy_workqueue(cm.wq);
4119
4120         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4121                 list_del(&timewait_info->list);
4122                 kfree(timewait_info);
4123         }
4124
4125         class_unregister(&cm_class);
4126         idr_destroy(&cm.local_id_table);
4127 }
4128
4129 module_init(ib_cm_init);
4130 module_exit(ib_cm_cleanup);
4131