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