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