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 };
 183
 184 struct cm_work {
 185         struct delayed_work work;
 186         struct list_head list;
 187         struct cm_port *port;
 188         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 189         __be32 local_id;                        /* Established / timewait */
 190         __be32 remote_id;
 191         struct ib_cm_event cm_event;
 192         struct ib_sa_path_rec path[0];
 193 };
 194
 195 struct cm_timewait_info {
 196         struct cm_work work;                    /* Must be first. */
 197         struct list_head list;
 198         struct rb_node remote_qp_node;
 199         struct rb_node remote_id_node;
 200         __be64 remote_ca_guid;
 201         __be32 remote_qpn;
 202         u8 inserted_remote_qp;
 203         u8 inserted_remote_id;
 204 };
 205
 206 struct cm_id_private {
 207         struct ib_cm_id id;
 208
 209         struct rb_node service_node;
 210         struct rb_node sidr_id_node;
 211         spinlock_t lock;        /* Do not acquire inside cm.lock */
 212         struct completion comp;
 213         atomic_t refcount;
 214         /* Number of clients sharing this ib_cm_id. Only valid for listeners.
 215          * Protected by the cm.lock spinlock. */
 216         int listen_sharecount;
 217
 218         struct ib_mad_send_buf *msg;
 219         struct cm_timewait_info *timewait_info;
 220         /* todo: use alternate port on send failure */
 221         struct cm_av av;
 222         struct cm_av alt_av;
 223
 224         void *private_data;
 225         __be64 tid;
 226         __be32 local_qpn;
 227         __be32 remote_qpn;
 228         enum ib_qp_type qp_type;
 229         __be32 sq_psn;
 230         __be32 rq_psn;
 231         int timeout_ms;
 232         enum ib_mtu path_mtu;
 233         __be16 pkey;
 234         u8 private_data_len;
 235         u8 max_cm_retries;
 236         u8 peer_to_peer;
 237         u8 responder_resources;
 238         u8 initiator_depth;
 239         u8 retry_count;
 240         u8 rnr_retry_count;
 241         u8 service_timeout;
 242         u8 target_ack_delay;
 243
 244         struct list_head work_list;
 245         atomic_t work_count;
 246 };
 247
 248 static void cm_work_handler(struct work_struct *work);
 249
 250 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 251 {
 252         if (atomic_dec_and_test(&cm_id_priv->refcount))
 253                 complete(&cm_id_priv->comp);
 254 }
 255
 256 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 257                         struct ib_mad_send_buf **msg)
 258 {
 259         struct ib_mad_agent *mad_agent;
 260         struct ib_mad_send_buf *m;
 261         struct ib_ah *ah;
 262
 263         mad_agent = cm_id_priv->av.port->mad_agent;
 264         ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
 265         if (IS_ERR(ah))
 266                 return PTR_ERR(ah);
 267
 268         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 269                                cm_id_priv->av.pkey_index,
 270                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 271                                GFP_ATOMIC,
 272                                IB_MGMT_BASE_VERSION);
 273         if (IS_ERR(m)) {
 274                 ib_destroy_ah(ah);
 275                 return PTR_ERR(m);
 276         }
 277
 278         /* Timeout set by caller if response is expected. */
 279         m->ah = ah;
 280         m->retries = cm_id_priv->max_cm_retries;
 281
 282         atomic_inc(&cm_id_priv->refcount);
 283         m->context[0] = cm_id_priv;
 284         *msg = m;
 285         return 0;
 286 }
 287
 288 static int cm_alloc_response_msg(struct cm_port *port,
 289                                  struct ib_mad_recv_wc *mad_recv_wc,
 290                                  struct ib_mad_send_buf **msg)
 291 {
 292         struct ib_mad_send_buf *m;
 293         struct ib_ah *ah;
 294
 295         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 296                                   mad_recv_wc->recv_buf.grh, port->port_num);
 297         if (IS_ERR(ah))
 298                 return PTR_ERR(ah);
 299
 300         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 301                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 302                                GFP_ATOMIC,
 303                                IB_MGMT_BASE_VERSION);
 304         if (IS_ERR(m)) {
 305                 ib_destroy_ah(ah);
 306                 return PTR_ERR(m);
 307         }
 308         m->ah = ah;
 309         *msg = m;
 310         return 0;
 311 }
 312
 313 static void cm_free_msg(struct ib_mad_send_buf *msg)
 314 {
 315         ib_destroy_ah(msg->ah);
 316         if (msg->context[0])
 317                 cm_deref_id(msg->context[0]);
 318         ib_free_send_mad(msg);
 319 }
 320
 321 static void * cm_copy_private_data(const void *private_data,
 322                                    u8 private_data_len)
 323 {
 324         void *data;
 325
 326         if (!private_data || !private_data_len)
 327                 return NULL;
 328
 329         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 330         if (!data)
 331                 return ERR_PTR(-ENOMEM);
 332
 333         return data;
 334 }
 335
 336 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 337                                  void *private_data, u8 private_data_len)
 338 {
 339         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 340                 kfree(cm_id_priv->private_data);
 341
 342         cm_id_priv->private_data = private_data;
 343         cm_id_priv->private_data_len = private_data_len;
 344 }
 345
 346 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 347                                     struct ib_grh *grh, struct cm_av *av)
 348 {
 349         av->port = port;
 350         av->pkey_index = wc->pkey_index;
 351         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 352                            grh, &av->ah_attr);
 353 }
 354
 355 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
 356 {
 357         struct cm_device *cm_dev;
 358         struct cm_port *port = NULL;
 359         unsigned long flags;
 360         int ret;
 361         u8 p;
 362         struct net_device *ndev = ib_get_ndev_from_path(path);
 363
 364         read_lock_irqsave(&cm.device_lock, flags);
 365         list_for_each_entry(cm_dev, &cm.device_list, list) {
 366                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 367                                         path->gid_type, ndev, &p, NULL)) {
 368                         port = cm_dev->port[p-1];
 369                         break;
 370                 }
 371         }
 372         read_unlock_irqrestore(&cm.device_lock, flags);
 373
 374         if (ndev)
 375                 dev_put(ndev);
 376
 377         if (!port)
 378                 return -EINVAL;
 379
 380         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 381                                   be16_to_cpu(path->pkey), &av->pkey_index);
 382         if (ret)
 383                 return ret;
 384
 385         av->port = port;
 386         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 387                              &av->ah_attr);
 388         av->timeout = path->packet_life_time + 1;
 389
 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         union ib_gid gid;
1604         struct ib_gid_attr gid_attr;
1605         int ret;
1606
1607         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1608
1609         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1610         if (IS_ERR(cm_id))
1611                 return PTR_ERR(cm_id);
1612
1613         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1614         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1615         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1616                                 work->mad_recv_wc->recv_buf.grh,
1617                                 &cm_id_priv->av);
1618         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1619                                                             id.local_id);
1620         if (IS_ERR(cm_id_priv->timewait_info)) {
1621                 ret = PTR_ERR(cm_id_priv->timewait_info);
1622                 goto destroy;
1623         }
1624         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1625         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1626         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1627
1628         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1629         if (!listen_cm_id_priv) {
1630                 ret = -EINVAL;
1631                 kfree(cm_id_priv->timewait_info);
1632                 goto destroy;
1633         }
1634
1635         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1636         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1637         cm_id_priv->id.service_id = req_msg->service_id;
1638         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1639
1640         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1641         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1642
1643         memcpy(work->path[0].dmac, cm_id_priv->av.ah_attr.dmac, ETH_ALEN);
1644         ret = ib_get_cached_gid(work->port->cm_dev->ib_device,
1645                                 work->port->port_num,
1646                                 cm_id_priv->av.ah_attr.grh.sgid_index,
1647                                 &gid, &gid_attr);
1648         if (!ret) {
1649                 if (gid_attr.ndev) {
1650                         work->path[0].ifindex = gid_attr.ndev->ifindex;
1651                         work->path[0].net = dev_net(gid_attr.ndev);
1652                         dev_put(gid_attr.ndev);
1653                 }
1654                 work->path[0].gid_type = gid_attr.gid_type;
1655                 ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
1656         }
1657         if (ret) {
1658                 int err = ib_get_cached_gid(work->port->cm_dev->ib_device,
1659                                             work->port->port_num, 0,
1660                                             &work->path[0].sgid,
1661                                             &gid_attr);
1662                 if (!err && gid_attr.ndev) {
1663                         work->path[0].ifindex = gid_attr.ndev->ifindex;
1664                         work->path[0].net = dev_net(gid_attr.ndev);
1665                         dev_put(gid_attr.ndev);
1666                 }
1667                 work->path[0].gid_type = gid_attr.gid_type;
1668                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1669                                &work->path[0].sgid, sizeof work->path[0].sgid,
1670                                NULL, 0);
1671                 goto rejected;
1672         }
1673         if (req_msg->alt_local_lid) {
1674                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
1675                 if (ret) {
1676                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1677                                        &work->path[0].sgid,
1678                                        sizeof work->path[0].sgid, NULL, 0);
1679                         goto rejected;
1680                 }
1681         }
1682         cm_id_priv->tid = req_msg->hdr.tid;
1683         cm_id_priv->timeout_ms = cm_convert_to_ms(
1684                                         cm_req_get_local_resp_timeout(req_msg));
1685         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1686         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1687         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1688         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1689         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1690         cm_id_priv->pkey = req_msg->pkey;
1691         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1692         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1693         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1694         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1695
1696         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1697         cm_process_work(cm_id_priv, work);
1698         cm_deref_id(listen_cm_id_priv);
1699         return 0;
1700
1701 rejected:
1702         atomic_dec(&cm_id_priv->refcount);
1703         cm_deref_id(listen_cm_id_priv);
1704 destroy:
1705         ib_destroy_cm_id(cm_id);
1706         return ret;
1707 }
1708
1709 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1710                           struct cm_id_private *cm_id_priv,
1711                           struct ib_cm_rep_param *param)
1712 {
1713         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1714         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1715         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1716         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1717         rep_msg->resp_resources = param->responder_resources;
1718         cm_rep_set_target_ack_delay(rep_msg,
1719                                     cm_id_priv->av.port->cm_dev->ack_delay);
1720         cm_rep_set_failover(rep_msg, param->failover_accepted);
1721         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1722         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1723
1724         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1725                 rep_msg->initiator_depth = param->initiator_depth;
1726                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1727                 cm_rep_set_srq(rep_msg, param->srq);
1728                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1729         } else {
1730                 cm_rep_set_srq(rep_msg, 1);
1731                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1732         }
1733
1734         if (param->private_data && param->private_data_len)
1735                 memcpy(rep_msg->private_data, param->private_data,
1736                        param->private_data_len);
1737 }
1738
1739 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1740                    struct ib_cm_rep_param *param)
1741 {
1742         struct cm_id_private *cm_id_priv;
1743         struct ib_mad_send_buf *msg;
1744         struct cm_rep_msg *rep_msg;
1745         unsigned long flags;
1746         int ret;
1747
1748         if (param->private_data &&
1749             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1750                 return -EINVAL;
1751
1752         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1753         spin_lock_irqsave(&cm_id_priv->lock, flags);
1754         if (cm_id->state != IB_CM_REQ_RCVD &&
1755             cm_id->state != IB_CM_MRA_REQ_SENT) {
1756                 ret = -EINVAL;
1757                 goto out;
1758         }
1759
1760         ret = cm_alloc_msg(cm_id_priv, &msg);
1761         if (ret)
1762                 goto out;
1763
1764         rep_msg = (struct cm_rep_msg *) msg->mad;
1765         cm_format_rep(rep_msg, cm_id_priv, param);
1766         msg->timeout_ms = cm_id_priv->timeout_ms;
1767         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1768
1769         ret = ib_post_send_mad(msg, NULL);
1770         if (ret) {
1771                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1772                 cm_free_msg(msg);
1773                 return ret;
1774         }
1775
1776         cm_id->state = IB_CM_REP_SENT;
1777         cm_id_priv->msg = msg;
1778         cm_id_priv->initiator_depth = param->initiator_depth;
1779         cm_id_priv->responder_resources = param->responder_resources;
1780         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1781         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
1782
1783 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1784         return ret;
1785 }
1786 EXPORT_SYMBOL(ib_send_cm_rep);
1787
1788 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1789                           struct cm_id_private *cm_id_priv,
1790                           const void *private_data,
1791                           u8 private_data_len)
1792 {
1793         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1794         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1795         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1796
1797         if (private_data && private_data_len)
1798                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1799 }
1800
1801 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1802                    const void *private_data,
1803                    u8 private_data_len)
1804 {
1805         struct cm_id_private *cm_id_priv;
1806         struct ib_mad_send_buf *msg;
1807         unsigned long flags;
1808         void *data;
1809         int ret;
1810
1811         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1812                 return -EINVAL;
1813
1814         data = cm_copy_private_data(private_data, private_data_len);
1815         if (IS_ERR(data))
1816                 return PTR_ERR(data);
1817
1818         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1819         spin_lock_irqsave(&cm_id_priv->lock, flags);
1820         if (cm_id->state != IB_CM_REP_RCVD &&
1821             cm_id->state != IB_CM_MRA_REP_SENT) {
1822                 ret = -EINVAL;
1823                 goto error;
1824         }
1825
1826         ret = cm_alloc_msg(cm_id_priv, &msg);
1827         if (ret)
1828                 goto error;
1829
1830         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1831                       private_data, private_data_len);
1832
1833         ret = ib_post_send_mad(msg, NULL);
1834         if (ret) {
1835                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1836                 cm_free_msg(msg);
1837                 kfree(data);
1838                 return ret;
1839         }
1840
1841         cm_id->state = IB_CM_ESTABLISHED;
1842         cm_set_private_data(cm_id_priv, data, private_data_len);
1843         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1844         return 0;
1845
1846 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1847         kfree(data);
1848         return ret;
1849 }
1850 EXPORT_SYMBOL(ib_send_cm_rtu);
1851
1852 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
1853 {
1854         struct cm_rep_msg *rep_msg;
1855         struct ib_cm_rep_event_param *param;
1856
1857         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1858         param = &work->cm_event.param.rep_rcvd;
1859         param->remote_ca_guid = rep_msg->local_ca_guid;
1860         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1861         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
1862         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1863         param->responder_resources = rep_msg->initiator_depth;
1864         param->initiator_depth = rep_msg->resp_resources;
1865         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1866         param->failover_accepted = cm_rep_get_failover(rep_msg);
1867         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1868         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1869         param->srq = cm_rep_get_srq(rep_msg);
1870         work->cm_event.private_data = &rep_msg->private_data;
1871 }
1872
1873 static void cm_dup_rep_handler(struct cm_work *work)
1874 {
1875         struct cm_id_private *cm_id_priv;
1876         struct cm_rep_msg *rep_msg;
1877         struct ib_mad_send_buf *msg = NULL;
1878         int ret;
1879
1880         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1881         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1882                                    rep_msg->local_comm_id);
1883         if (!cm_id_priv)
1884                 return;
1885
1886         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1887                         counter[CM_REP_COUNTER]);
1888         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1889         if (ret)
1890                 goto deref;
1891
1892         spin_lock_irq(&cm_id_priv->lock);
1893         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1894                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1895                               cm_id_priv->private_data,
1896                               cm_id_priv->private_data_len);
1897         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1898                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1899                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1900                               cm_id_priv->private_data,
1901                               cm_id_priv->private_data_len);
1902         else
1903                 goto unlock;
1904         spin_unlock_irq(&cm_id_priv->lock);
1905
1906         ret = ib_post_send_mad(msg, NULL);
1907         if (ret)
1908                 goto free;
1909         goto deref;
1910
1911 unlock: spin_unlock_irq(&cm_id_priv->lock);
1912 free:   cm_free_msg(msg);
1913 deref:  cm_deref_id(cm_id_priv);
1914 }
1915
1916 static int cm_rep_handler(struct cm_work *work)
1917 {
1918         struct cm_id_private *cm_id_priv;
1919         struct cm_rep_msg *rep_msg;
1920         int ret;
1921
1922         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1923         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1924         if (!cm_id_priv) {
1925                 cm_dup_rep_handler(work);
1926                 return -EINVAL;
1927         }
1928
1929         cm_format_rep_event(work, cm_id_priv->qp_type);
1930
1931         spin_lock_irq(&cm_id_priv->lock);
1932         switch (cm_id_priv->id.state) {
1933         case IB_CM_REQ_SENT:
1934         case IB_CM_MRA_REQ_RCVD:
1935                 break;
1936         default:
1937                 spin_unlock_irq(&cm_id_priv->lock);
1938                 ret = -EINVAL;
1939                 goto error;
1940         }
1941
1942         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1943         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1944         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
1945
1946         spin_lock(&cm.lock);
1947         /* Check for duplicate REP. */
1948         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1949                 spin_unlock(&cm.lock);
1950                 spin_unlock_irq(&cm_id_priv->lock);
1951                 ret = -EINVAL;
1952                 goto error;
1953         }
1954         /* Check for a stale connection. */
1955         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
1956                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
1957                          &cm.remote_id_table);
1958                 cm_id_priv->timewait_info->inserted_remote_id = 0;
1959                 spin_unlock(&cm.lock);
1960                 spin_unlock_irq(&cm_id_priv->lock);
1961                 cm_issue_rej(work->port, work->mad_recv_wc,
1962                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
1963                              NULL, 0);
1964                 ret = -EINVAL;
1965                 goto error;
1966         }
1967         spin_unlock(&cm.lock);
1968
1969         cm_id_priv->id.state = IB_CM_REP_RCVD;
1970         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
1971         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
1972         cm_id_priv->initiator_depth = rep_msg->resp_resources;
1973         cm_id_priv->responder_resources = rep_msg->initiator_depth;
1974         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
1975         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1976         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1977         cm_id_priv->av.timeout =
1978                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1979                                        cm_id_priv->av.timeout - 1);
1980         cm_id_priv->alt_av.timeout =
1981                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1982                                        cm_id_priv->alt_av.timeout - 1);
1983
1984         /* todo: handle peer_to_peer */
1985
1986         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1987         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1988         if (!ret)
1989                 list_add_tail(&work->list, &cm_id_priv->work_list);
1990         spin_unlock_irq(&cm_id_priv->lock);
1991
1992         if (ret)
1993                 cm_process_work(cm_id_priv, work);
1994         else
1995                 cm_deref_id(cm_id_priv);
1996         return 0;
1997
1998 error:
1999         cm_deref_id(cm_id_priv);
2000         return ret;
2001 }
2002
2003 static int cm_establish_handler(struct cm_work *work)
2004 {
2005         struct cm_id_private *cm_id_priv;
2006         int ret;
2007
2008         /* See comment in cm_establish about lookup. */
2009         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2010         if (!cm_id_priv)
2011                 return -EINVAL;
2012
2013         spin_lock_irq(&cm_id_priv->lock);
2014         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2015                 spin_unlock_irq(&cm_id_priv->lock);
2016                 goto out;
2017         }
2018
2019         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2020         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2021         if (!ret)
2022                 list_add_tail(&work->list, &cm_id_priv->work_list);
2023         spin_unlock_irq(&cm_id_priv->lock);
2024
2025         if (ret)
2026                 cm_process_work(cm_id_priv, work);
2027         else
2028                 cm_deref_id(cm_id_priv);
2029         return 0;
2030 out:
2031         cm_deref_id(cm_id_priv);
2032         return -EINVAL;
2033 }
2034
2035 static int cm_rtu_handler(struct cm_work *work)
2036 {
2037         struct cm_id_private *cm_id_priv;
2038         struct cm_rtu_msg *rtu_msg;
2039         int ret;
2040
2041         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2042         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
2043                                    rtu_msg->local_comm_id);
2044         if (!cm_id_priv)
2045                 return -EINVAL;
2046
2047         work->cm_event.private_data = &rtu_msg->private_data;
2048
2049         spin_lock_irq(&cm_id_priv->lock);
2050         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2051             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2052                 spin_unlock_irq(&cm_id_priv->lock);
2053                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2054                                 counter[CM_RTU_COUNTER]);
2055                 goto out;
2056         }
2057         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2058
2059         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2060         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2061         if (!ret)
2062                 list_add_tail(&work->list, &cm_id_priv->work_list);
2063         spin_unlock_irq(&cm_id_priv->lock);
2064
2065         if (ret)
2066                 cm_process_work(cm_id_priv, work);
2067         else
2068                 cm_deref_id(cm_id_priv);
2069         return 0;
2070 out:
2071         cm_deref_id(cm_id_priv);
2072         return -EINVAL;
2073 }
2074
2075 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2076                           struct cm_id_private *cm_id_priv,
2077                           const void *private_data,
2078                           u8 private_data_len)
2079 {
2080         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2081                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
2082         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
2083         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
2084         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2085
2086         if (private_data && private_data_len)
2087                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2088 }
2089
2090 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2091                     const void *private_data,
2092                     u8 private_data_len)
2093 {
2094         struct cm_id_private *cm_id_priv;
2095         struct ib_mad_send_buf *msg;
2096         unsigned long flags;
2097         int ret;
2098
2099         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2100                 return -EINVAL;
2101
2102         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2103         spin_lock_irqsave(&cm_id_priv->lock, flags);
2104         if (cm_id->state != IB_CM_ESTABLISHED) {
2105                 ret = -EINVAL;
2106                 goto out;
2107         }
2108
2109         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2110             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2111                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2112
2113         ret = cm_alloc_msg(cm_id_priv, &msg);
2114         if (ret) {
2115                 cm_enter_timewait(cm_id_priv);
2116                 goto out;
2117         }
2118
2119         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2120                        private_data, private_data_len);
2121         msg->timeout_ms = cm_id_priv->timeout_ms;
2122         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2123
2124         ret = ib_post_send_mad(msg, NULL);
2125         if (ret) {
2126                 cm_enter_timewait(cm_id_priv);
2127                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2128                 cm_free_msg(msg);
2129                 return ret;
2130         }
2131
2132         cm_id->state = IB_CM_DREQ_SENT;
2133         cm_id_priv->msg = msg;
2134 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2135         return ret;
2136 }
2137 EXPORT_SYMBOL(ib_send_cm_dreq);
2138
2139 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2140                           struct cm_id_private *cm_id_priv,
2141                           const void *private_data,
2142                           u8 private_data_len)
2143 {
2144         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2145         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2146         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2147
2148         if (private_data && private_data_len)
2149                 memcpy(drep_msg->private_data, private_data, private_data_len);
2150 }
2151
2152 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2153                     const void *private_data,
2154                     u8 private_data_len)
2155 {
2156         struct cm_id_private *cm_id_priv;
2157         struct ib_mad_send_buf *msg;
2158         unsigned long flags;
2159         void *data;
2160         int ret;
2161
2162         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2163                 return -EINVAL;
2164
2165         data = cm_copy_private_data(private_data, private_data_len);
2166         if (IS_ERR(data))
2167                 return PTR_ERR(data);
2168
2169         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2170         spin_lock_irqsave(&cm_id_priv->lock, flags);
2171         if (cm_id->state != IB_CM_DREQ_RCVD) {
2172                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2173                 kfree(data);
2174                 return -EINVAL;
2175         }
2176
2177         cm_set_private_data(cm_id_priv, data, private_data_len);
2178         cm_enter_timewait(cm_id_priv);
2179
2180         ret = cm_alloc_msg(cm_id_priv, &msg);
2181         if (ret)
2182                 goto out;
2183
2184         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2185                        private_data, private_data_len);
2186
2187         ret = ib_post_send_mad(msg, NULL);
2188         if (ret) {
2189                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2190                 cm_free_msg(msg);
2191                 return ret;
2192         }
2193
2194 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2195         return ret;
2196 }
2197 EXPORT_SYMBOL(ib_send_cm_drep);
2198
2199 static int cm_issue_drep(struct cm_port *port,
2200                          struct ib_mad_recv_wc *mad_recv_wc)
2201 {
2202         struct ib_mad_send_buf *msg = NULL;
2203         struct cm_dreq_msg *dreq_msg;
2204         struct cm_drep_msg *drep_msg;
2205         int ret;
2206
2207         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2208         if (ret)
2209                 return ret;
2210
2211         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2212         drep_msg = (struct cm_drep_msg *) msg->mad;
2213
2214         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2215         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2216         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2217
2218         ret = ib_post_send_mad(msg, NULL);
2219         if (ret)
2220                 cm_free_msg(msg);
2221
2222         return ret;
2223 }
2224
2225 static int cm_dreq_handler(struct cm_work *work)
2226 {
2227         struct cm_id_private *cm_id_priv;
2228         struct cm_dreq_msg *dreq_msg;
2229         struct ib_mad_send_buf *msg = NULL;
2230         int ret;
2231
2232         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2233         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2234                                    dreq_msg->local_comm_id);
2235         if (!cm_id_priv) {
2236                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2237                                 counter[CM_DREQ_COUNTER]);
2238                 cm_issue_drep(work->port, work->mad_recv_wc);
2239                 return -EINVAL;
2240         }
2241
2242         work->cm_event.private_data = &dreq_msg->private_data;
2243
2244         spin_lock_irq(&cm_id_priv->lock);
2245         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2246                 goto unlock;
2247
2248         switch (cm_id_priv->id.state) {
2249         case IB_CM_REP_SENT:
2250         case IB_CM_DREQ_SENT:
2251                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2252                 break;
2253         case IB_CM_ESTABLISHED:
2254                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2255                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2256                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2257                 break;
2258         case IB_CM_MRA_REP_RCVD:
2259                 break;
2260         case IB_CM_TIMEWAIT:
2261                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2262                                 counter[CM_DREQ_COUNTER]);
2263                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2264                         goto unlock;
2265
2266                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2267                                cm_id_priv->private_data,
2268                                cm_id_priv->private_data_len);
2269                 spin_unlock_irq(&cm_id_priv->lock);
2270
2271                 if (ib_post_send_mad(msg, NULL))
2272                         cm_free_msg(msg);
2273                 goto deref;
2274         case IB_CM_DREQ_RCVD:
2275                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2276                                 counter[CM_DREQ_COUNTER]);
2277                 goto unlock;
2278         default:
2279                 goto unlock;
2280         }
2281         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2282         cm_id_priv->tid = dreq_msg->hdr.tid;
2283         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2284         if (!ret)
2285                 list_add_tail(&work->list, &cm_id_priv->work_list);
2286         spin_unlock_irq(&cm_id_priv->lock);
2287
2288         if (ret)
2289                 cm_process_work(cm_id_priv, work);
2290         else
2291                 cm_deref_id(cm_id_priv);
2292         return 0;
2293
2294 unlock: spin_unlock_irq(&cm_id_priv->lock);
2295 deref:  cm_deref_id(cm_id_priv);
2296         return -EINVAL;
2297 }
2298
2299 static int cm_drep_handler(struct cm_work *work)
2300 {
2301         struct cm_id_private *cm_id_priv;
2302         struct cm_drep_msg *drep_msg;
2303         int ret;
2304
2305         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2306         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2307                                    drep_msg->local_comm_id);
2308         if (!cm_id_priv)
2309                 return -EINVAL;
2310
2311         work->cm_event.private_data = &drep_msg->private_data;
2312
2313         spin_lock_irq(&cm_id_priv->lock);
2314         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2315             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2316                 spin_unlock_irq(&cm_id_priv->lock);
2317                 goto out;
2318         }
2319         cm_enter_timewait(cm_id_priv);
2320
2321         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2322         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2323         if (!ret)
2324                 list_add_tail(&work->list, &cm_id_priv->work_list);
2325         spin_unlock_irq(&cm_id_priv->lock);
2326
2327         if (ret)
2328                 cm_process_work(cm_id_priv, work);
2329         else
2330                 cm_deref_id(cm_id_priv);
2331         return 0;
2332 out:
2333         cm_deref_id(cm_id_priv);
2334         return -EINVAL;
2335 }
2336
2337 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2338                    enum ib_cm_rej_reason reason,
2339                    void *ari,
2340                    u8 ari_length,
2341                    const void *private_data,
2342                    u8 private_data_len)
2343 {
2344         struct cm_id_private *cm_id_priv;
2345         struct ib_mad_send_buf *msg;
2346         unsigned long flags;
2347         int ret;
2348
2349         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2350             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2351                 return -EINVAL;
2352
2353         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2354
2355         spin_lock_irqsave(&cm_id_priv->lock, flags);
2356         switch (cm_id->state) {
2357         case IB_CM_REQ_SENT:
2358         case IB_CM_MRA_REQ_RCVD:
2359         case IB_CM_REQ_RCVD:
2360         case IB_CM_MRA_REQ_SENT:
2361         case IB_CM_REP_RCVD:
2362         case IB_CM_MRA_REP_SENT:
2363                 ret = cm_alloc_msg(cm_id_priv, &msg);
2364                 if (!ret)
2365                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2366                                       cm_id_priv, reason, ari, ari_length,
2367                                       private_data, private_data_len);
2368
2369                 cm_reset_to_idle(cm_id_priv);
2370                 break;
2371         case IB_CM_REP_SENT:
2372         case IB_CM_MRA_REP_RCVD:
2373                 ret = cm_alloc_msg(cm_id_priv, &msg);
2374                 if (!ret)
2375                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2376                                       cm_id_priv, reason, ari, ari_length,
2377                                       private_data, private_data_len);
2378
2379                 cm_enter_timewait(cm_id_priv);
2380                 break;
2381         default:
2382                 ret = -EINVAL;
2383                 goto out;
2384         }
2385
2386         if (ret)
2387                 goto out;
2388
2389         ret = ib_post_send_mad(msg, NULL);
2390         if (ret)
2391                 cm_free_msg(msg);
2392
2393 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2394         return ret;
2395 }
2396 EXPORT_SYMBOL(ib_send_cm_rej);
2397
2398 static void cm_format_rej_event(struct cm_work *work)
2399 {
2400         struct cm_rej_msg *rej_msg;
2401         struct ib_cm_rej_event_param *param;
2402
2403         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2404         param = &work->cm_event.param.rej_rcvd;
2405         param->ari = rej_msg->ari;
2406         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2407         param->reason = __be16_to_cpu(rej_msg->reason);
2408         work->cm_event.private_data = &rej_msg->private_data;
2409 }
2410
2411 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2412 {
2413         struct cm_timewait_info *timewait_info;
2414         struct cm_id_private *cm_id_priv;
2415         __be32 remote_id;
2416
2417         remote_id = rej_msg->local_comm_id;
2418
2419         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2420                 spin_lock_irq(&cm.lock);
2421                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2422                                                   remote_id);
2423                 if (!timewait_info) {
2424                         spin_unlock_irq(&cm.lock);
2425                         return NULL;
2426                 }
2427                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2428                                       (timewait_info->work.local_id ^
2429                                        cm.random_id_operand));
2430                 if (cm_id_priv) {
2431                         if (cm_id_priv->id.remote_id == remote_id)
2432                                 atomic_inc(&cm_id_priv->refcount);
2433                         else
2434                                 cm_id_priv = NULL;
2435                 }
2436                 spin_unlock_irq(&cm.lock);
2437         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2438                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2439         else
2440                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2441
2442         return cm_id_priv;
2443 }
2444
2445 static int cm_rej_handler(struct cm_work *work)
2446 {
2447         struct cm_id_private *cm_id_priv;
2448         struct cm_rej_msg *rej_msg;
2449         int ret;
2450
2451         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2452         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2453         if (!cm_id_priv)
2454                 return -EINVAL;
2455
2456         cm_format_rej_event(work);
2457
2458         spin_lock_irq(&cm_id_priv->lock);
2459         switch (cm_id_priv->id.state) {
2460         case IB_CM_REQ_SENT:
2461         case IB_CM_MRA_REQ_RCVD:
2462         case IB_CM_REP_SENT:
2463         case IB_CM_MRA_REP_RCVD:
2464                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2465                 /* fall through */
2466         case IB_CM_REQ_RCVD:
2467         case IB_CM_MRA_REQ_SENT:
2468                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2469                         cm_enter_timewait(cm_id_priv);
2470                 else
2471                         cm_reset_to_idle(cm_id_priv);
2472                 break;
2473         case IB_CM_DREQ_SENT:
2474                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2475                 /* fall through */
2476         case IB_CM_REP_RCVD:
2477         case IB_CM_MRA_REP_SENT:
2478                 cm_enter_timewait(cm_id_priv);
2479                 break;
2480         case IB_CM_ESTABLISHED:
2481                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2482                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2483                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2484                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2485                                               cm_id_priv->msg);
2486                         cm_enter_timewait(cm_id_priv);
2487                         break;
2488                 }
2489                 /* fall through */
2490         default:
2491                 spin_unlock_irq(&cm_id_priv->lock);
2492                 ret = -EINVAL;
2493                 goto out;
2494         }
2495
2496         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2497         if (!ret)
2498                 list_add_tail(&work->list, &cm_id_priv->work_list);
2499         spin_unlock_irq(&cm_id_priv->lock);
2500
2501         if (ret)
2502                 cm_process_work(cm_id_priv, work);
2503         else
2504                 cm_deref_id(cm_id_priv);
2505         return 0;
2506 out:
2507         cm_deref_id(cm_id_priv);
2508         return -EINVAL;
2509 }
2510
2511 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2512                    u8 service_timeout,
2513                    const void *private_data,
2514                    u8 private_data_len)
2515 {
2516         struct cm_id_private *cm_id_priv;
2517         struct ib_mad_send_buf *msg;
2518         enum ib_cm_state cm_state;
2519         enum ib_cm_lap_state lap_state;
2520         enum cm_msg_response msg_response;
2521         void *data;
2522         unsigned long flags;
2523         int ret;
2524
2525         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2526                 return -EINVAL;
2527
2528         data = cm_copy_private_data(private_data, private_data_len);
2529         if (IS_ERR(data))
2530                 return PTR_ERR(data);
2531
2532         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2533
2534         spin_lock_irqsave(&cm_id_priv->lock, flags);
2535         switch(cm_id_priv->id.state) {
2536         case IB_CM_REQ_RCVD:
2537                 cm_state = IB_CM_MRA_REQ_SENT;
2538                 lap_state = cm_id->lap_state;
2539                 msg_response = CM_MSG_RESPONSE_REQ;
2540                 break;
2541         case IB_CM_REP_RCVD:
2542                 cm_state = IB_CM_MRA_REP_SENT;
2543                 lap_state = cm_id->lap_state;
2544                 msg_response = CM_MSG_RESPONSE_REP;
2545                 break;
2546         case IB_CM_ESTABLISHED:
2547                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2548                         cm_state = cm_id->state;
2549                         lap_state = IB_CM_MRA_LAP_SENT;
2550                         msg_response = CM_MSG_RESPONSE_OTHER;
2551                         break;
2552                 }
2553         default:
2554                 ret = -EINVAL;
2555                 goto error1;
2556         }
2557
2558         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2559                 ret = cm_alloc_msg(cm_id_priv, &msg);
2560                 if (ret)
2561                         goto error1;
2562
2563                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2564                               msg_response, service_timeout,
2565                               private_data, private_data_len);
2566                 ret = ib_post_send_mad(msg, NULL);
2567                 if (ret)
2568                         goto error2;
2569         }
2570
2571         cm_id->state = cm_state;
2572         cm_id->lap_state = lap_state;
2573         cm_id_priv->service_timeout = service_timeout;
2574         cm_set_private_data(cm_id_priv, data, private_data_len);
2575         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2576         return 0;
2577
2578 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2579         kfree(data);
2580         return ret;
2581
2582 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2583         kfree(data);
2584         cm_free_msg(msg);
2585         return ret;
2586 }
2587 EXPORT_SYMBOL(ib_send_cm_mra);
2588
2589 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2590 {
2591         switch (cm_mra_get_msg_mraed(mra_msg)) {
2592         case CM_MSG_RESPONSE_REQ:
2593                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2594         case CM_MSG_RESPONSE_REP:
2595         case CM_MSG_RESPONSE_OTHER:
2596                 return cm_acquire_id(mra_msg->remote_comm_id,
2597                                      mra_msg->local_comm_id);
2598         default:
2599                 return NULL;
2600         }
2601 }
2602
2603 static int cm_mra_handler(struct cm_work *work)
2604 {
2605         struct cm_id_private *cm_id_priv;
2606         struct cm_mra_msg *mra_msg;
2607         int timeout, ret;
2608
2609         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2610         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2611         if (!cm_id_priv)
2612                 return -EINVAL;
2613
2614         work->cm_event.private_data = &mra_msg->private_data;
2615         work->cm_event.param.mra_rcvd.service_timeout =
2616                                         cm_mra_get_service_timeout(mra_msg);
2617         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2618                   cm_convert_to_ms(cm_id_priv->av.timeout);
2619
2620         spin_lock_irq(&cm_id_priv->lock);
2621         switch (cm_id_priv->id.state) {
2622         case IB_CM_REQ_SENT:
2623                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2624                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2625                                   cm_id_priv->msg, timeout))
2626                         goto out;
2627                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2628                 break;
2629         case IB_CM_REP_SENT:
2630                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2631                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2632                                   cm_id_priv->msg, timeout))
2633                         goto out;
2634                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2635                 break;
2636         case IB_CM_ESTABLISHED:
2637                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2638                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2639                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2640                                   cm_id_priv->msg, timeout)) {
2641                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2642                                 atomic_long_inc(&work->port->
2643                                                 counter_group[CM_RECV_DUPLICATES].
2644                                                 counter[CM_MRA_COUNTER]);
2645                         goto out;
2646                 }
2647                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2648                 break;
2649         case IB_CM_MRA_REQ_RCVD:
2650         case IB_CM_MRA_REP_RCVD:
2651                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2652                                 counter[CM_MRA_COUNTER]);
2653                 /* fall through */
2654         default:
2655                 goto out;
2656         }
2657
2658         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2659                                       cm_id_priv->id.state;
2660         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2661         if (!ret)
2662                 list_add_tail(&work->list, &cm_id_priv->work_list);
2663         spin_unlock_irq(&cm_id_priv->lock);
2664
2665         if (ret)
2666                 cm_process_work(cm_id_priv, work);
2667         else
2668                 cm_deref_id(cm_id_priv);
2669         return 0;
2670 out:
2671         spin_unlock_irq(&cm_id_priv->lock);
2672         cm_deref_id(cm_id_priv);
2673         return -EINVAL;
2674 }
2675
2676 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2677                           struct cm_id_private *cm_id_priv,
2678                           struct ib_sa_path_rec *alternate_path,
2679                           const void *private_data,
2680                           u8 private_data_len)
2681 {
2682         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2683                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2684         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2685         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2686         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2687         /* todo: need remote CM response timeout */
2688         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2689         lap_msg->alt_local_lid = alternate_path->slid;
2690         lap_msg->alt_remote_lid = alternate_path->dlid;
2691         lap_msg->alt_local_gid = alternate_path->sgid;
2692         lap_msg->alt_remote_gid = alternate_path->dgid;
2693         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2694         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2695         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2696         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2697         cm_lap_set_sl(lap_msg, alternate_path->sl);
2698         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2699         cm_lap_set_local_ack_timeout(lap_msg,
2700                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2701                                alternate_path->packet_life_time));
2702
2703         if (private_data && private_data_len)
2704                 memcpy(lap_msg->private_data, private_data, private_data_len);
2705 }
2706
2707 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2708                    struct ib_sa_path_rec *alternate_path,
2709                    const void *private_data,
2710                    u8 private_data_len)
2711 {
2712         struct cm_id_private *cm_id_priv;
2713         struct ib_mad_send_buf *msg;
2714         unsigned long flags;
2715         int ret;
2716
2717         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2718                 return -EINVAL;
2719
2720         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2721         spin_lock_irqsave(&cm_id_priv->lock, flags);
2722         if (cm_id->state != IB_CM_ESTABLISHED ||
2723             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2724              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2725                 ret = -EINVAL;
2726                 goto out;
2727         }
2728
2729         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
2730         if (ret)
2731                 goto out;
2732         cm_id_priv->alt_av.timeout =
2733                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2734                                        cm_id_priv->alt_av.timeout - 1);
2735
2736         ret = cm_alloc_msg(cm_id_priv, &msg);
2737         if (ret)
2738                 goto out;
2739
2740         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2741                       alternate_path, private_data, private_data_len);
2742         msg->timeout_ms = cm_id_priv->timeout_ms;
2743         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2744
2745         ret = ib_post_send_mad(msg, NULL);
2746         if (ret) {
2747                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2748                 cm_free_msg(msg);
2749                 return ret;
2750         }
2751
2752         cm_id->lap_state = IB_CM_LAP_SENT;
2753         cm_id_priv->msg = msg;
2754
2755 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2756         return ret;
2757 }
2758 EXPORT_SYMBOL(ib_send_cm_lap);
2759
2760 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2761                                     struct ib_sa_path_rec *path,
2762                                     struct cm_lap_msg *lap_msg)
2763 {
2764         memset(path, 0, sizeof *path);
2765         path->dgid = lap_msg->alt_local_gid;
2766         path->sgid = lap_msg->alt_remote_gid;
2767         path->dlid = lap_msg->alt_local_lid;
2768         path->slid = lap_msg->alt_remote_lid;
2769         path->flow_label = cm_lap_get_flow_label(lap_msg);
2770         path->hop_limit = lap_msg->alt_hop_limit;
2771         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2772         path->reversible = 1;
2773         path->pkey = cm_id_priv->pkey;
2774         path->sl = cm_lap_get_sl(lap_msg);
2775         path->mtu_selector = IB_SA_EQ;
2776         path->mtu = cm_id_priv->path_mtu;
2777         path->rate_selector = IB_SA_EQ;
2778         path->rate = cm_lap_get_packet_rate(lap_msg);
2779         path->packet_life_time_selector = IB_SA_EQ;
2780         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2781         path->packet_life_time -= (path->packet_life_time > 0);
2782 }
2783
2784 static int cm_lap_handler(struct cm_work *work)
2785 {
2786         struct cm_id_private *cm_id_priv;
2787         struct cm_lap_msg *lap_msg;
2788         struct ib_cm_lap_event_param *param;
2789         struct ib_mad_send_buf *msg = NULL;
2790         int ret;
2791
2792         /* todo: verify LAP request and send reject APR if invalid. */
2793         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2794         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2795                                    lap_msg->local_comm_id);
2796         if (!cm_id_priv)
2797                 return -EINVAL;
2798
2799         param = &work->cm_event.param.lap_rcvd;
2800         param->alternate_path = &work->path[0];
2801         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2802         work->cm_event.private_data = &lap_msg->private_data;
2803
2804         spin_lock_irq(&cm_id_priv->lock);
2805         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2806                 goto unlock;
2807
2808         switch (cm_id_priv->id.lap_state) {
2809         case IB_CM_LAP_UNINIT:
2810         case IB_CM_LAP_IDLE:
2811                 break;
2812         case IB_CM_MRA_LAP_SENT:
2813                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2814                                 counter[CM_LAP_COUNTER]);
2815                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2816                         goto unlock;
2817
2818                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2819                               CM_MSG_RESPONSE_OTHER,
2820                               cm_id_priv->service_timeout,
2821                               cm_id_priv->private_data,
2822                               cm_id_priv->private_data_len);
2823                 spin_unlock_irq(&cm_id_priv->lock);
2824
2825                 if (ib_post_send_mad(msg, NULL))
2826                         cm_free_msg(msg);
2827                 goto deref;
2828         case IB_CM_LAP_RCVD:
2829                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2830                                 counter[CM_LAP_COUNTER]);
2831                 goto unlock;
2832         default:
2833                 goto unlock;
2834         }
2835
2836         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2837         cm_id_priv->tid = lap_msg->hdr.tid;
2838         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2839                                 work->mad_recv_wc->recv_buf.grh,
2840                                 &cm_id_priv->av);
2841         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av);
2842         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2843         if (!ret)
2844                 list_add_tail(&work->list, &cm_id_priv->work_list);
2845         spin_unlock_irq(&cm_id_priv->lock);
2846
2847         if (ret)
2848                 cm_process_work(cm_id_priv, work);
2849         else
2850                 cm_deref_id(cm_id_priv);
2851         return 0;
2852
2853 unlock: spin_unlock_irq(&cm_id_priv->lock);
2854 deref:  cm_deref_id(cm_id_priv);
2855         return -EINVAL;
2856 }
2857
2858 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2859                           struct cm_id_private *cm_id_priv,
2860                           enum ib_cm_apr_status status,
2861                           void *info,
2862                           u8 info_length,
2863                           const void *private_data,
2864                           u8 private_data_len)
2865 {
2866         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2867         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2868         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2869         apr_msg->ap_status = (u8) status;
2870
2871         if (info && info_length) {
2872                 apr_msg->info_length = info_length;
2873                 memcpy(apr_msg->info, info, info_length);
2874         }
2875
2876         if (private_data && private_data_len)
2877                 memcpy(apr_msg->private_data, private_data, private_data_len);
2878 }
2879
2880 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2881                    enum ib_cm_apr_status status,
2882                    void *info,
2883                    u8 info_length,
2884                    const void *private_data,
2885                    u8 private_data_len)
2886 {
2887         struct cm_id_private *cm_id_priv;
2888         struct ib_mad_send_buf *msg;
2889         unsigned long flags;
2890         int ret;
2891
2892         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2893             (info && info_length > IB_CM_APR_INFO_LENGTH))
2894                 return -EINVAL;
2895
2896         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2897         spin_lock_irqsave(&cm_id_priv->lock, flags);
2898         if (cm_id->state != IB_CM_ESTABLISHED ||
2899             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2900              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2901                 ret = -EINVAL;
2902                 goto out;
2903         }
2904
2905         ret = cm_alloc_msg(cm_id_priv, &msg);
2906         if (ret)
2907                 goto out;
2908
2909         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2910                       info, info_length, private_data, private_data_len);
2911         ret = ib_post_send_mad(msg, NULL);
2912         if (ret) {
2913                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2914                 cm_free_msg(msg);
2915                 return ret;
2916         }
2917
2918         cm_id->lap_state = IB_CM_LAP_IDLE;
2919 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2920         return ret;
2921 }
2922 EXPORT_SYMBOL(ib_send_cm_apr);
2923
2924 static int cm_apr_handler(struct cm_work *work)
2925 {
2926         struct cm_id_private *cm_id_priv;
2927         struct cm_apr_msg *apr_msg;
2928         int ret;
2929
2930         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2931         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2932                                    apr_msg->local_comm_id);
2933         if (!cm_id_priv)
2934                 return -EINVAL; /* Unmatched reply. */
2935
2936         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2937         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2938         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2939         work->cm_event.private_data = &apr_msg->private_data;
2940
2941         spin_lock_irq(&cm_id_priv->lock);
2942         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2943             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2944              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2945                 spin_unlock_irq(&cm_id_priv->lock);
2946                 goto out;
2947         }
2948         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2949         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2950         cm_id_priv->msg = NULL;
2951
2952         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2953         if (!ret)
2954                 list_add_tail(&work->list, &cm_id_priv->work_list);
2955         spin_unlock_irq(&cm_id_priv->lock);
2956
2957         if (ret)
2958                 cm_process_work(cm_id_priv, work);
2959         else
2960                 cm_deref_id(cm_id_priv);
2961         return 0;
2962 out:
2963         cm_deref_id(cm_id_priv);
2964         return -EINVAL;
2965 }
2966
2967 static int cm_timewait_handler(struct cm_work *work)
2968 {
2969         struct cm_timewait_info *timewait_info;
2970         struct cm_id_private *cm_id_priv;
2971         int ret;
2972
2973         timewait_info = (struct cm_timewait_info *)work;
2974         spin_lock_irq(&cm.lock);
2975         list_del(&timewait_info->list);
2976         spin_unlock_irq(&cm.lock);
2977
2978         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2979                                    timewait_info->work.remote_id);
2980         if (!cm_id_priv)
2981                 return -EINVAL;
2982
2983         spin_lock_irq(&cm_id_priv->lock);
2984         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2985             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2986                 spin_unlock_irq(&cm_id_priv->lock);
2987                 goto out;
2988         }
2989         cm_id_priv->id.state = IB_CM_IDLE;
2990         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2991         if (!ret)
2992                 list_add_tail(&work->list, &cm_id_priv->work_list);
2993         spin_unlock_irq(&cm_id_priv->lock);
2994
2995         if (ret)
2996                 cm_process_work(cm_id_priv, work);
2997         else
2998                 cm_deref_id(cm_id_priv);
2999         return 0;
3000 out:
3001         cm_deref_id(cm_id_priv);
3002         return -EINVAL;
3003 }
3004
3005 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3006                                struct cm_id_private *cm_id_priv,
3007                                struct ib_cm_sidr_req_param *param)
3008 {
3009         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3010                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
3011         sidr_req_msg->request_id = cm_id_priv->id.local_id;
3012         sidr_req_msg->pkey = param->path->pkey;
3013         sidr_req_msg->service_id = param->service_id;
3014
3015         if (param->private_data && param->private_data_len)
3016                 memcpy(sidr_req_msg->private_data, param->private_data,
3017                        param->private_data_len);
3018 }
3019
3020 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3021                         struct ib_cm_sidr_req_param *param)
3022 {
3023         struct cm_id_private *cm_id_priv;
3024         struct ib_mad_send_buf *msg;
3025         unsigned long flags;
3026         int ret;
3027
3028         if (!param->path || (param->private_data &&
3029              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3030                 return -EINVAL;
3031
3032         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3033         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
3034         if (ret)
3035                 goto out;
3036
3037         cm_id->service_id = param->service_id;
3038         cm_id->service_mask = ~cpu_to_be64(0);
3039         cm_id_priv->timeout_ms = param->timeout_ms;
3040         cm_id_priv->max_cm_retries = param->max_cm_retries;
3041         ret = cm_alloc_msg(cm_id_priv, &msg);
3042         if (ret)
3043                 goto out;
3044
3045         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3046                            param);
3047         msg->timeout_ms = cm_id_priv->timeout_ms;
3048         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3049
3050         spin_lock_irqsave(&cm_id_priv->lock, flags);
3051         if (cm_id->state == IB_CM_IDLE)
3052                 ret = ib_post_send_mad(msg, NULL);
3053         else
3054                 ret = -EINVAL;
3055
3056         if (ret) {
3057                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3058                 cm_free_msg(msg);
3059                 goto out;
3060         }
3061         cm_id->state = IB_CM_SIDR_REQ_SENT;
3062         cm_id_priv->msg = msg;
3063         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3064 out:
3065         return ret;
3066 }
3067 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3068
3069 static void cm_format_sidr_req_event(struct cm_work *work,
3070                                      struct ib_cm_id *listen_id)
3071 {
3072         struct cm_sidr_req_msg *sidr_req_msg;
3073         struct ib_cm_sidr_req_event_param *param;
3074
3075         sidr_req_msg = (struct cm_sidr_req_msg *)
3076                                 work->mad_recv_wc->recv_buf.mad;
3077         param = &work->cm_event.param.sidr_req_rcvd;
3078         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
3079         param->listen_id = listen_id;
3080         param->service_id = sidr_req_msg->service_id;
3081         param->bth_pkey = cm_get_bth_pkey(work);
3082         param->port = work->port->port_num;
3083         work->cm_event.private_data = &sidr_req_msg->private_data;
3084 }
3085
3086 static int cm_sidr_req_handler(struct cm_work *work)
3087 {
3088         struct ib_cm_id *cm_id;
3089         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3090         struct cm_sidr_req_msg *sidr_req_msg;
3091         struct ib_wc *wc;
3092
3093         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3094         if (IS_ERR(cm_id))
3095                 return PTR_ERR(cm_id);
3096         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3097
3098         /* Record SGID/SLID and request ID for lookup. */
3099         sidr_req_msg = (struct cm_sidr_req_msg *)
3100                                 work->mad_recv_wc->recv_buf.mad;
3101         wc = work->mad_recv_wc->wc;
3102         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3103         cm_id_priv->av.dgid.global.interface_id = 0;
3104         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3105                                 work->mad_recv_wc->recv_buf.grh,
3106                                 &cm_id_priv->av);
3107         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3108         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3109         atomic_inc(&cm_id_priv->work_count);
3110
3111         spin_lock_irq(&cm.lock);
3112         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3113         if (cur_cm_id_priv) {
3114                 spin_unlock_irq(&cm.lock);
3115                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3116                                 counter[CM_SIDR_REQ_COUNTER]);
3117                 goto out; /* Duplicate message. */
3118         }
3119         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3120         cur_cm_id_priv = cm_find_listen(cm_id->device,
3121                                         sidr_req_msg->service_id);
3122         if (!cur_cm_id_priv) {
3123                 spin_unlock_irq(&cm.lock);
3124                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3125                 goto out; /* No match. */
3126         }
3127         atomic_inc(&cur_cm_id_priv->refcount);
3128         atomic_inc(&cm_id_priv->refcount);
3129         spin_unlock_irq(&cm.lock);
3130
3131         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3132         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3133         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3134         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3135
3136         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3137         cm_process_work(cm_id_priv, work);
3138         cm_deref_id(cur_cm_id_priv);
3139         return 0;
3140 out:
3141         ib_destroy_cm_id(&cm_id_priv->id);
3142         return -EINVAL;
3143 }
3144
3145 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3146                                struct cm_id_private *cm_id_priv,
3147                                struct ib_cm_sidr_rep_param *param)
3148 {
3149         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3150                           cm_id_priv->tid);
3151         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3152         sidr_rep_msg->status = param->status;
3153         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3154         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3155         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3156
3157         if (param->info && param->info_length)
3158                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3159
3160         if (param->private_data && param->private_data_len)
3161                 memcpy(sidr_rep_msg->private_data, param->private_data,
3162                        param->private_data_len);
3163 }
3164
3165 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3166                         struct ib_cm_sidr_rep_param *param)
3167 {
3168         struct cm_id_private *cm_id_priv;
3169         struct ib_mad_send_buf *msg;
3170         unsigned long flags;
3171         int ret;
3172
3173         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3174             (param->private_data &&
3175              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3176                 return -EINVAL;
3177
3178         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3179         spin_lock_irqsave(&cm_id_priv->lock, flags);
3180         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3181                 ret = -EINVAL;
3182                 goto error;
3183         }
3184
3185         ret = cm_alloc_msg(cm_id_priv, &msg);
3186         if (ret)
3187                 goto error;
3188
3189         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3190                            param);
3191         ret = ib_post_send_mad(msg, NULL);
3192         if (ret) {
3193                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3194                 cm_free_msg(msg);
3195                 return ret;
3196         }
3197         cm_id->state = IB_CM_IDLE;
3198         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3199
3200         spin_lock_irqsave(&cm.lock, flags);
3201         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3202                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3203                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3204         }
3205         spin_unlock_irqrestore(&cm.lock, flags);
3206         return 0;
3207
3208 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3209         return ret;
3210 }
3211 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3212
3213 static void cm_format_sidr_rep_event(struct cm_work *work)
3214 {
3215         struct cm_sidr_rep_msg *sidr_rep_msg;
3216         struct ib_cm_sidr_rep_event_param *param;
3217
3218         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3219                                 work->mad_recv_wc->recv_buf.mad;
3220         param = &work->cm_event.param.sidr_rep_rcvd;
3221         param->status = sidr_rep_msg->status;
3222         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3223         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3224         param->info = &sidr_rep_msg->info;
3225         param->info_len = sidr_rep_msg->info_length;
3226         work->cm_event.private_data = &sidr_rep_msg->private_data;
3227 }
3228
3229 static int cm_sidr_rep_handler(struct cm_work *work)
3230 {
3231         struct cm_sidr_rep_msg *sidr_rep_msg;
3232         struct cm_id_private *cm_id_priv;
3233
3234         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3235                                 work->mad_recv_wc->recv_buf.mad;
3236         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3237         if (!cm_id_priv)
3238                 return -EINVAL; /* Unmatched reply. */
3239
3240         spin_lock_irq(&cm_id_priv->lock);
3241         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3242                 spin_unlock_irq(&cm_id_priv->lock);
3243                 goto out;
3244         }
3245         cm_id_priv->id.state = IB_CM_IDLE;
3246         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3247         spin_unlock_irq(&cm_id_priv->lock);
3248
3249         cm_format_sidr_rep_event(work);
3250         cm_process_work(cm_id_priv, work);
3251         return 0;
3252 out:
3253         cm_deref_id(cm_id_priv);
3254         return -EINVAL;
3255 }
3256
3257 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3258                                   enum ib_wc_status wc_status)
3259 {
3260         struct cm_id_private *cm_id_priv;
3261         struct ib_cm_event cm_event;
3262         enum ib_cm_state state;
3263         int ret;
3264
3265         memset(&cm_event, 0, sizeof cm_event);
3266         cm_id_priv = msg->context[0];
3267
3268         /* Discard old sends or ones without a response. */
3269         spin_lock_irq(&cm_id_priv->lock);
3270         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3271         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3272                 goto discard;
3273
3274         switch (state) {
3275         case IB_CM_REQ_SENT:
3276         case IB_CM_MRA_REQ_RCVD:
3277                 cm_reset_to_idle(cm_id_priv);
3278                 cm_event.event = IB_CM_REQ_ERROR;
3279                 break;
3280         case IB_CM_REP_SENT:
3281         case IB_CM_MRA_REP_RCVD:
3282                 cm_reset_to_idle(cm_id_priv);
3283                 cm_event.event = IB_CM_REP_ERROR;
3284                 break;
3285         case IB_CM_DREQ_SENT:
3286                 cm_enter_timewait(cm_id_priv);
3287                 cm_event.event = IB_CM_DREQ_ERROR;
3288                 break;
3289         case IB_CM_SIDR_REQ_SENT:
3290                 cm_id_priv->id.state = IB_CM_IDLE;
3291                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3292                 break;
3293         default:
3294                 goto discard;
3295         }
3296         spin_unlock_irq(&cm_id_priv->lock);
3297         cm_event.param.send_status = wc_status;
3298
3299         /* No other events can occur on the cm_id at this point. */
3300         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3301         cm_free_msg(msg);
3302         if (ret)
3303                 ib_destroy_cm_id(&cm_id_priv->id);
3304         return;
3305 discard:
3306         spin_unlock_irq(&cm_id_priv->lock);
3307         cm_free_msg(msg);
3308 }
3309
3310 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3311                             struct ib_mad_send_wc *mad_send_wc)
3312 {
3313         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3314         struct cm_port *port;
3315         u16 attr_index;
3316
3317         port = mad_agent->context;
3318         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3319                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3320
3321         /*
3322          * If the send was in response to a received message (context[0] is not
3323          * set to a cm_id), and is not a REJ, then it is a send that was
3324          * manually retried.
3325          */
3326         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3327                 msg->retries = 1;
3328
3329         atomic_long_add(1 + msg->retries,
3330                         &port->counter_group[CM_XMIT].counter[attr_index]);
3331         if (msg->retries)
3332                 atomic_long_add(msg->retries,
3333                                 &port->counter_group[CM_XMIT_RETRIES].
3334                                 counter[attr_index]);
3335
3336         switch (mad_send_wc->status) {
3337         case IB_WC_SUCCESS:
3338         case IB_WC_WR_FLUSH_ERR:
3339                 cm_free_msg(msg);
3340                 break;
3341         default:
3342                 if (msg->context[0] && msg->context[1])
3343                         cm_process_send_error(msg, mad_send_wc->status);
3344                 else
3345                         cm_free_msg(msg);
3346                 break;
3347         }
3348 }
3349
3350 static void cm_work_handler(struct work_struct *_work)
3351 {
3352         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3353         int ret;
3354
3355         switch (work->cm_event.event) {
3356         case IB_CM_REQ_RECEIVED:
3357                 ret = cm_req_handler(work);
3358                 break;
3359         case IB_CM_MRA_RECEIVED:
3360                 ret = cm_mra_handler(work);
3361                 break;
3362         case IB_CM_REJ_RECEIVED:
3363                 ret = cm_rej_handler(work);
3364                 break;
3365         case IB_CM_REP_RECEIVED:
3366                 ret = cm_rep_handler(work);
3367                 break;
3368         case IB_CM_RTU_RECEIVED:
3369                 ret = cm_rtu_handler(work);
3370                 break;
3371         case IB_CM_USER_ESTABLISHED:
3372                 ret = cm_establish_handler(work);
3373                 break;
3374         case IB_CM_DREQ_RECEIVED:
3375                 ret = cm_dreq_handler(work);
3376                 break;
3377         case IB_CM_DREP_RECEIVED:
3378                 ret = cm_drep_handler(work);
3379                 break;
3380         case IB_CM_SIDR_REQ_RECEIVED:
3381                 ret = cm_sidr_req_handler(work);
3382                 break;
3383         case IB_CM_SIDR_REP_RECEIVED:
3384                 ret = cm_sidr_rep_handler(work);
3385                 break;
3386         case IB_CM_LAP_RECEIVED:
3387                 ret = cm_lap_handler(work);
3388                 break;
3389         case IB_CM_APR_RECEIVED:
3390                 ret = cm_apr_handler(work);
3391                 break;
3392         case IB_CM_TIMEWAIT_EXIT:
3393                 ret = cm_timewait_handler(work);
3394                 break;
3395         default:
3396                 ret = -EINVAL;
3397                 break;
3398         }
3399         if (ret)
3400                 cm_free_work(work);
3401 }
3402
3403 static int cm_establish(struct ib_cm_id *cm_id)
3404 {
3405         struct cm_id_private *cm_id_priv;
3406         struct cm_work *work;
3407         unsigned long flags;
3408         int ret = 0;
3409         struct cm_device *cm_dev;
3410
3411         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3412         if (!cm_dev)
3413                 return -ENODEV;
3414
3415         work = kmalloc(sizeof *work, GFP_ATOMIC);
3416         if (!work)
3417                 return -ENOMEM;
3418
3419         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3420         spin_lock_irqsave(&cm_id_priv->lock, flags);
3421         switch (cm_id->state)
3422         {
3423         case IB_CM_REP_SENT:
3424         case IB_CM_MRA_REP_RCVD:
3425                 cm_id->state = IB_CM_ESTABLISHED;
3426                 break;
3427         case IB_CM_ESTABLISHED:
3428                 ret = -EISCONN;
3429                 break;
3430         default:
3431                 ret = -EINVAL;
3432                 break;
3433         }
3434         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3435
3436         if (ret) {
3437                 kfree(work);
3438                 goto out;
3439         }
3440
3441         /*
3442          * The CM worker thread may try to destroy the cm_id before it
3443          * can execute this work item.  To prevent potential deadlock,
3444          * we need to find the cm_id once we're in the context of the
3445          * worker thread, rather than holding a reference on it.
3446          */
3447         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3448         work->local_id = cm_id->local_id;
3449         work->remote_id = cm_id->remote_id;
3450         work->mad_recv_wc = NULL;
3451         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3452
3453         /* Check if the device started its remove_one */
3454         spin_lock_irq(&cm.lock);
3455         if (!cm_dev->going_down) {
3456                 queue_delayed_work(cm.wq, &work->work, 0);
3457         } else {
3458                 kfree(work);
3459                 ret = -ENODEV;
3460         }
3461         spin_unlock_irq(&cm.lock);
3462
3463 out:
3464         return ret;
3465 }
3466
3467 static int cm_migrate(struct ib_cm_id *cm_id)
3468 {
3469         struct cm_id_private *cm_id_priv;
3470         unsigned long flags;
3471         int ret = 0;
3472
3473         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3474         spin_lock_irqsave(&cm_id_priv->lock, flags);
3475         if (cm_id->state == IB_CM_ESTABLISHED &&
3476             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3477              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3478                 cm_id->lap_state = IB_CM_LAP_IDLE;
3479                 cm_id_priv->av = cm_id_priv->alt_av;
3480         } else
3481                 ret = -EINVAL;
3482         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3483
3484         return ret;
3485 }
3486
3487 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3488 {
3489         int ret;
3490
3491         switch (event) {
3492         case IB_EVENT_COMM_EST:
3493                 ret = cm_establish(cm_id);
3494                 break;
3495         case IB_EVENT_PATH_MIG:
3496                 ret = cm_migrate(cm_id);
3497                 break;
3498         default:
3499                 ret = -EINVAL;
3500         }
3501         return ret;
3502 }
3503 EXPORT_SYMBOL(ib_cm_notify);
3504
3505 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3506                             struct ib_mad_recv_wc *mad_recv_wc)
3507 {
3508         struct cm_port *port = mad_agent->context;
3509         struct cm_work *work;
3510         enum ib_cm_event_type event;
3511         u16 attr_id;
3512         int paths = 0;
3513         int going_down = 0;
3514
3515         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3516         case CM_REQ_ATTR_ID:
3517                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3518                                                     alt_local_lid != 0);
3519                 event = IB_CM_REQ_RECEIVED;
3520                 break;
3521         case CM_MRA_ATTR_ID:
3522                 event = IB_CM_MRA_RECEIVED;
3523                 break;
3524         case CM_REJ_ATTR_ID:
3525                 event = IB_CM_REJ_RECEIVED;
3526                 break;
3527         case CM_REP_ATTR_ID:
3528                 event = IB_CM_REP_RECEIVED;
3529                 break;
3530         case CM_RTU_ATTR_ID:
3531                 event = IB_CM_RTU_RECEIVED;
3532                 break;
3533         case CM_DREQ_ATTR_ID:
3534                 event = IB_CM_DREQ_RECEIVED;
3535                 break;
3536         case CM_DREP_ATTR_ID:
3537                 event = IB_CM_DREP_RECEIVED;
3538                 break;
3539         case CM_SIDR_REQ_ATTR_ID:
3540                 event = IB_CM_SIDR_REQ_RECEIVED;
3541                 break;
3542         case CM_SIDR_REP_ATTR_ID:
3543                 event = IB_CM_SIDR_REP_RECEIVED;
3544                 break;
3545         case CM_LAP_ATTR_ID:
3546                 paths = 1;
3547                 event = IB_CM_LAP_RECEIVED;
3548                 break;
3549         case CM_APR_ATTR_ID:
3550                 event = IB_CM_APR_RECEIVED;
3551                 break;
3552         default:
3553                 ib_free_recv_mad(mad_recv_wc);
3554                 return;
3555         }
3556
3557         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3558         atomic_long_inc(&port->counter_group[CM_RECV].
3559                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3560
3561         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3562                        GFP_KERNEL);
3563         if (!work) {
3564                 ib_free_recv_mad(mad_recv_wc);
3565                 return;
3566         }
3567
3568         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3569         work->cm_event.event = event;
3570         work->mad_recv_wc = mad_recv_wc;
3571         work->port = port;
3572
3573         /* Check if the device started its remove_one */
3574         spin_lock_irq(&cm.lock);
3575         if (!port->cm_dev->going_down)
3576                 queue_delayed_work(cm.wq, &work->work, 0);
3577         else
3578                 going_down = 1;
3579         spin_unlock_irq(&cm.lock);
3580
3581         if (going_down) {
3582                 kfree(work);
3583                 ib_free_recv_mad(mad_recv_wc);
3584         }
3585 }
3586
3587 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3588                                 struct ib_qp_attr *qp_attr,
3589                                 int *qp_attr_mask)
3590 {
3591         unsigned long flags;
3592         int ret;
3593
3594         spin_lock_irqsave(&cm_id_priv->lock, flags);
3595         switch (cm_id_priv->id.state) {
3596         case IB_CM_REQ_SENT:
3597         case IB_CM_MRA_REQ_RCVD:
3598         case IB_CM_REQ_RCVD:
3599         case IB_CM_MRA_REQ_SENT:
3600         case IB_CM_REP_RCVD:
3601         case IB_CM_MRA_REP_SENT:
3602         case IB_CM_REP_SENT:
3603         case IB_CM_MRA_REP_RCVD:
3604         case IB_CM_ESTABLISHED:
3605                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3606                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3607                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3608                 if (cm_id_priv->responder_resources)
3609                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3610                                                     IB_ACCESS_REMOTE_ATOMIC;
3611                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3612                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3613                 ret = 0;
3614                 break;
3615         default:
3616                 ret = -EINVAL;
3617                 break;
3618         }
3619         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3620         return ret;
3621 }
3622
3623 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3624                                struct ib_qp_attr *qp_attr,
3625                                int *qp_attr_mask)
3626 {
3627         unsigned long flags;
3628         int ret;
3629
3630         spin_lock_irqsave(&cm_id_priv->lock, flags);
3631         switch (cm_id_priv->id.state) {
3632         case IB_CM_REQ_RCVD:
3633         case IB_CM_MRA_REQ_SENT:
3634         case IB_CM_REP_RCVD:
3635         case IB_CM_MRA_REP_SENT:
3636         case IB_CM_REP_SENT:
3637         case IB_CM_MRA_REP_RCVD:
3638         case IB_CM_ESTABLISHED:
3639                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3640                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3641                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3642                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3643                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3644                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3645                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3646                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3647                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3648                                          IB_QP_MIN_RNR_TIMER;
3649                         qp_attr->max_dest_rd_atomic =
3650                                         cm_id_priv->responder_resources;
3651                         qp_attr->min_rnr_timer = 0;
3652                 }
3653                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3654                         *qp_attr_mask |= IB_QP_ALT_PATH;
3655                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3656                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3657                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3658                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3659                 }
3660                 ret = 0;
3661                 break;
3662         default:
3663                 ret = -EINVAL;
3664                 break;
3665         }
3666         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3667         return ret;
3668 }
3669
3670 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3671                                struct ib_qp_attr *qp_attr,
3672                                int *qp_attr_mask)
3673 {
3674         unsigned long flags;
3675         int ret;
3676
3677         spin_lock_irqsave(&cm_id_priv->lock, flags);
3678         switch (cm_id_priv->id.state) {
3679         /* Allow transition to RTS before sending REP */
3680         case IB_CM_REQ_RCVD:
3681         case IB_CM_MRA_REQ_SENT:
3682
3683         case IB_CM_REP_RCVD:
3684         case IB_CM_MRA_REP_SENT:
3685         case IB_CM_REP_SENT:
3686         case IB_CM_MRA_REP_RCVD:
3687         case IB_CM_ESTABLISHED:
3688                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3689                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3690                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3691                         switch (cm_id_priv->qp_type) {
3692                         case IB_QPT_RC:
3693                         case IB_QPT_XRC_INI:
3694                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3695                                                  IB_QP_MAX_QP_RD_ATOMIC;
3696                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3697                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3698                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3699                                 /* fall through */
3700                         case IB_QPT_XRC_TGT:
3701                                 *qp_attr_mask |= IB_QP_TIMEOUT;
3702                                 qp_attr->timeout = cm_id_priv->av.timeout;
3703                                 break;
3704                         default:
3705                                 break;
3706                         }
3707                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3708                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3709                                 qp_attr->path_mig_state = IB_MIG_REARM;
3710                         }
3711                 } else {
3712                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3713                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3714                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3715                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3716                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3717                         qp_attr->path_mig_state = IB_MIG_REARM;
3718                 }
3719                 ret = 0;
3720                 break;
3721         default:
3722                 ret = -EINVAL;
3723                 break;
3724         }
3725         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3726         return ret;
3727 }
3728
3729 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3730                        struct ib_qp_attr *qp_attr,
3731                        int *qp_attr_mask)
3732 {
3733         struct cm_id_private *cm_id_priv;
3734         int ret;
3735
3736         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3737         switch (qp_attr->qp_state) {
3738         case IB_QPS_INIT:
3739                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3740                 break;
3741         case IB_QPS_RTR:
3742                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3743                 break;
3744         case IB_QPS_RTS:
3745                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3746                 break;
3747         default:
3748                 ret = -EINVAL;
3749                 break;
3750         }
3751         return ret;
3752 }
3753 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3754
3755 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3756                                char *buf)
3757 {
3758         struct cm_counter_group *group;
3759         struct cm_counter_attribute *cm_attr;
3760
3761         group = container_of(obj, struct cm_counter_group, obj);
3762         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3763
3764         return sprintf(buf, "%ld\n",
3765                        atomic_long_read(&group->counter[cm_attr->index]));
3766 }
3767
3768 static const struct sysfs_ops cm_counter_ops = {
3769         .show = cm_show_counter
3770 };
3771
3772 static struct kobj_type cm_counter_obj_type = {
3773         .sysfs_ops = &cm_counter_ops,
3774         .default_attrs = cm_counter_default_attrs
3775 };
3776
3777 static void cm_release_port_obj(struct kobject *obj)
3778 {
3779         struct cm_port *cm_port;
3780
3781         cm_port = container_of(obj, struct cm_port, port_obj);
3782         kfree(cm_port);
3783 }
3784
3785 static struct kobj_type cm_port_obj_type = {
3786         .release = cm_release_port_obj
3787 };
3788
3789 static char *cm_devnode(struct device *dev, umode_t *mode)
3790 {
3791         if (mode)
3792                 *mode = 0666;
3793         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3794 }
3795
3796 struct class cm_class = {
3797         .owner   = THIS_MODULE,
3798         .name    = "infiniband_cm",
3799         .devnode = cm_devnode,
3800 };
3801 EXPORT_SYMBOL(cm_class);
3802
3803 static int cm_create_port_fs(struct cm_port *port)
3804 {
3805         int i, ret;
3806
3807         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3808                                    &port->cm_dev->device->kobj,
3809                                    "%d", port->port_num);
3810         if (ret) {
3811                 kfree(port);
3812                 return ret;
3813         }
3814
3815         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3816                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3817                                            &cm_counter_obj_type,
3818                                            &port->port_obj,
3819                                            "%s", counter_group_names[i]);
3820                 if (ret)
3821                         goto error;
3822         }
3823
3824         return 0;
3825
3826 error:
3827         while (i--)
3828                 kobject_put(&port->counter_group[i].obj);
3829         kobject_put(&port->port_obj);
3830         return ret;
3831
3832 }
3833
3834 static void cm_remove_port_fs(struct cm_port *port)
3835 {
3836         int i;
3837
3838         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3839                 kobject_put(&port->counter_group[i].obj);
3840
3841         kobject_put(&port->port_obj);
3842 }
3843
3844 static void cm_add_one(struct ib_device *ib_device)
3845 {
3846         struct cm_device *cm_dev;
3847         struct cm_port *port;
3848         struct ib_mad_reg_req reg_req = {
3849                 .mgmt_class = IB_MGMT_CLASS_CM,
3850                 .mgmt_class_version = IB_CM_CLASS_VERSION,
3851         };
3852         struct ib_port_modify port_modify = {
3853                 .set_port_cap_mask = IB_PORT_CM_SUP
3854         };
3855         unsigned long flags;
3856         int ret;
3857         int count = 0;
3858         u8 i;
3859
3860         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3861                          ib_device->phys_port_cnt, GFP_KERNEL);
3862         if (!cm_dev)
3863                 return;
3864
3865         cm_dev->ib_device = ib_device;
3866         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
3867         cm_dev->going_down = 0;
3868         cm_dev->device = device_create(&cm_class, &ib_device->dev,
3869                                        MKDEV(0, 0), NULL,
3870                                        "%s", ib_device->name);
3871         if (IS_ERR(cm_dev->device)) {
3872                 kfree(cm_dev);
3873                 return;
3874         }
3875
3876         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3877         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3878                 if (!rdma_cap_ib_cm(ib_device, i))
3879                         continue;
3880
3881                 port = kzalloc(sizeof *port, GFP_KERNEL);
3882                 if (!port)
3883                         goto error1;
3884
3885                 cm_dev->port[i-1] = port;
3886                 port->cm_dev = cm_dev;
3887                 port->port_num = i;
3888
3889                 ret = cm_create_port_fs(port);
3890                 if (ret)
3891                         goto error1;
3892
3893                 port->mad_agent = ib_register_mad_agent(ib_device, i,
3894                                                         IB_QPT_GSI,
3895                                                         &reg_req,
3896                                                         0,
3897                                                         cm_send_handler,
3898                                                         cm_recv_handler,
3899                                                         port,
3900                                                         0);
3901                 if (IS_ERR(port->mad_agent))
3902                         goto error2;
3903
3904                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
3905                 if (ret)
3906                         goto error3;
3907
3908                 count++;
3909         }
3910
3911         if (!count)
3912                 goto free;
3913
3914         ib_set_client_data(ib_device, &cm_client, cm_dev);
3915
3916         write_lock_irqsave(&cm.device_lock, flags);
3917         list_add_tail(&cm_dev->list, &cm.device_list);
3918         write_unlock_irqrestore(&cm.device_lock, flags);
3919         return;
3920
3921 error3:
3922         ib_unregister_mad_agent(port->mad_agent);
3923 error2:
3924         cm_remove_port_fs(port);
3925 error1:
3926         port_modify.set_port_cap_mask = 0;
3927         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3928         while (--i) {
3929                 if (!rdma_cap_ib_cm(ib_device, i))
3930                         continue;
3931
3932                 port = cm_dev->port[i-1];
3933                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3934                 ib_unregister_mad_agent(port->mad_agent);
3935                 cm_remove_port_fs(port);
3936         }
3937 free:
3938         device_unregister(cm_dev->device);
3939         kfree(cm_dev);
3940 }
3941
3942 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
3943 {
3944         struct cm_device *cm_dev = client_data;
3945         struct cm_port *port;
3946         struct ib_port_modify port_modify = {
3947                 .clr_port_cap_mask = IB_PORT_CM_SUP
3948         };
3949         unsigned long flags;
3950         int i;
3951
3952         if (!cm_dev)
3953                 return;
3954
3955         write_lock_irqsave(&cm.device_lock, flags);
3956         list_del(&cm_dev->list);
3957         write_unlock_irqrestore(&cm.device_lock, flags);
3958
3959         spin_lock_irq(&cm.lock);
3960         cm_dev->going_down = 1;
3961         spin_unlock_irq(&cm.lock);
3962
3963         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3964                 if (!rdma_cap_ib_cm(ib_device, i))
3965                         continue;
3966
3967                 port = cm_dev->port[i-1];
3968                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3969                 /*
3970                  * We flush the queue here after the going_down set, this
3971                  * verify that no new works will be queued in the recv handler,
3972                  * after that we can call the unregister_mad_agent
3973                  */
3974                 flush_workqueue(cm.wq);
3975                 ib_unregister_mad_agent(port->mad_agent);
3976                 cm_remove_port_fs(port);
3977         }
3978         device_unregister(cm_dev->device);
3979         kfree(cm_dev);
3980 }
3981
3982 static int __init ib_cm_init(void)
3983 {
3984         int ret;
3985
3986         memset(&cm, 0, sizeof cm);
3987         INIT_LIST_HEAD(&cm.device_list);
3988         rwlock_init(&cm.device_lock);
3989         spin_lock_init(&cm.lock);
3990         cm.listen_service_table = RB_ROOT;
3991         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3992         cm.remote_id_table = RB_ROOT;
3993         cm.remote_qp_table = RB_ROOT;
3994         cm.remote_sidr_table = RB_ROOT;
3995         idr_init(&cm.local_id_table);
3996         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
3997         INIT_LIST_HEAD(&cm.timewait_list);
3998
3999         ret = class_register(&cm_class);
4000         if (ret) {
4001                 ret = -ENOMEM;
4002                 goto error1;
4003         }
4004
4005         cm.wq = create_workqueue("ib_cm");
4006         if (!cm.wq) {
4007                 ret = -ENOMEM;
4008                 goto error2;
4009         }
4010
4011         ret = ib_register_client(&cm_client);
4012         if (ret)
4013                 goto error3;
4014
4015         return 0;
4016 error3:
4017         destroy_workqueue(cm.wq);
4018 error2:
4019         class_unregister(&cm_class);
4020 error1:
4021         idr_destroy(&cm.local_id_table);
4022         return ret;
4023 }
4024
4025 static void __exit ib_cm_cleanup(void)
4026 {
4027         struct cm_timewait_info *timewait_info, *tmp;
4028
4029         spin_lock_irq(&cm.lock);
4030         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4031                 cancel_delayed_work(&timewait_info->work.work);
4032         spin_unlock_irq(&cm.lock);
4033
4034         ib_unregister_client(&cm_client);
4035         destroy_workqueue(cm.wq);
4036
4037         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4038                 list_del(&timewait_info->list);
4039                 kfree(timewait_info);
4040         }
4041
4042         class_unregister(&cm_class);
4043         idr_destroy(&cm.local_id_table);
4044 }
4045
4046 module_init(ib_cm_init);
4047 module_exit(ib_cm_cleanup);
4048