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