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         int going_down;
 173         struct cm_port *port[0];
 174 };
 175
 176 struct cm_av {
 177         struct cm_port *port;
 178         union ib_gid dgid;
 179         struct ib_ah_attr ah_attr;
 180         u16 pkey_index;
 181         u8 timeout;
 182         u8  valid;
 183         u8  smac[ETH_ALEN];
 184 };
 185
 186 struct cm_work {
 187         struct delayed_work work;
 188         struct list_head list;
 189         struct cm_port *port;
 190         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 191         __be32 local_id;                        /* Established / timewait */
 192         __be32 remote_id;
 193         struct ib_cm_event cm_event;
 194         struct ib_sa_path_rec path[0];
 195 };
 196
 197 struct cm_timewait_info {
 198         struct cm_work work;                    /* Must be first. */
 199         struct list_head list;
 200         struct rb_node remote_qp_node;
 201         struct rb_node remote_id_node;
 202         __be64 remote_ca_guid;
 203         __be32 remote_qpn;
 204         u8 inserted_remote_qp;
 205         u8 inserted_remote_id;
 206 };
 207
 208 struct cm_id_private {
 209         struct ib_cm_id id;
 210
 211         struct rb_node service_node;
 212         struct rb_node sidr_id_node;
 213         spinlock_t lock;        /* Do not acquire inside cm.lock */
 214         struct completion comp;
 215         atomic_t refcount;
 216
 217         struct ib_mad_send_buf *msg;
 218         struct cm_timewait_info *timewait_info;
 219         /* todo: use alternate port on send failure */
 220         struct cm_av av;
 221         struct cm_av alt_av;
 222         struct ib_cm_compare_data *compare_data;
 223
 224         void *private_data;
 225         __be64 tid;
 226         __be32 local_qpn;
 227         __be32 remote_qpn;
 228         enum ib_qp_type qp_type;
 229         __be32 sq_psn;
 230         __be32 rq_psn;
 231         int timeout_ms;
 232         enum ib_mtu path_mtu;
 233         __be16 pkey;
 234         u8 private_data_len;
 235         u8 max_cm_retries;
 236         u8 peer_to_peer;
 237         u8 responder_resources;
 238         u8 initiator_depth;
 239         u8 retry_count;
 240         u8 rnr_retry_count;
 241         u8 service_timeout;
 242         u8 target_ack_delay;
 243
 244         struct list_head work_list;
 245         atomic_t work_count;
 246 };
 247
 248 static void cm_work_handler(struct work_struct *work);
 249
 250 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 251 {
 252         if (atomic_dec_and_test(&cm_id_priv->refcount))
 253                 complete(&cm_id_priv->comp);
 254 }
 255
 256 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 257                         struct ib_mad_send_buf **msg)
 258 {
 259         struct ib_mad_agent *mad_agent;
 260         struct ib_mad_send_buf *m;
 261         struct ib_ah *ah;
 262
 263         mad_agent = cm_id_priv->av.port->mad_agent;
 264         ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
 265         if (IS_ERR(ah))
 266                 return PTR_ERR(ah);
 267
 268         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 269                                cm_id_priv->av.pkey_index,
 270                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 271                                GFP_ATOMIC,
 272                                IB_MGMT_BASE_VERSION);
 273         if (IS_ERR(m)) {
 274                 ib_destroy_ah(ah);
 275                 return PTR_ERR(m);
 276         }
 277
 278         /* Timeout set by caller if response is expected. */
 279         m->ah = ah;
 280         m->retries = cm_id_priv->max_cm_retries;
 281
 282         atomic_inc(&cm_id_priv->refcount);
 283         m->context[0] = cm_id_priv;
 284         *msg = m;
 285         return 0;
 286 }
 287
 288 static int cm_alloc_response_msg(struct cm_port *port,
 289                                  struct ib_mad_recv_wc *mad_recv_wc,
 290                                  struct ib_mad_send_buf **msg)
 291 {
 292         struct ib_mad_send_buf *m;
 293         struct ib_ah *ah;
 294
 295         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 296                                   mad_recv_wc->recv_buf.grh, port->port_num);
 297         if (IS_ERR(ah))
 298                 return PTR_ERR(ah);
 299
 300         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 301                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 302                                GFP_ATOMIC,
 303                                IB_MGMT_BASE_VERSION);
 304         if (IS_ERR(m)) {
 305                 ib_destroy_ah(ah);
 306                 return PTR_ERR(m);
 307         }
 308         m->ah = ah;
 309         *msg = m;
 310         return 0;
 311 }
 312
 313 static void cm_free_msg(struct ib_mad_send_buf *msg)
 314 {
 315         ib_destroy_ah(msg->ah);
 316         if (msg->context[0])
 317                 cm_deref_id(msg->context[0]);
 318         ib_free_send_mad(msg);
 319 }
 320
 321 static void * cm_copy_private_data(const void *private_data,
 322                                    u8 private_data_len)
 323 {
 324         void *data;
 325
 326         if (!private_data || !private_data_len)
 327                 return NULL;
 328
 329         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 330         if (!data)
 331                 return ERR_PTR(-ENOMEM);
 332
 333         return data;
 334 }
 335
 336 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 337                                  void *private_data, u8 private_data_len)
 338 {
 339         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 340                 kfree(cm_id_priv->private_data);
 341
 342         cm_id_priv->private_data = private_data;
 343         cm_id_priv->private_data_len = private_data_len;
 344 }
 345
 346 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 347                                     struct ib_grh *grh, struct cm_av *av)
 348 {
 349         av->port = port;
 350         av->pkey_index = wc->pkey_index;
 351         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 352                            grh, &av->ah_attr);
 353 }
 354
 355 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
 356 {
 357         struct cm_device *cm_dev;
 358         struct cm_port *port = NULL;
 359         unsigned long flags;
 360         int ret;
 361         u8 p;
 362
 363         read_lock_irqsave(&cm.device_lock, flags);
 364         list_for_each_entry(cm_dev, &cm.device_list, list) {
 365                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 366                                         &p, NULL)) {
 367                         port = cm_dev->port[p-1];
 368                         break;
 369                 }
 370         }
 371         read_unlock_irqrestore(&cm.device_lock, flags);
 372
 373         if (!port)
 374                 return -EINVAL;
 375
 376         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 377                                   be16_to_cpu(path->pkey), &av->pkey_index);
 378         if (ret)
 379                 return ret;
 380
 381         av->port = port;
 382         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 383                              &av->ah_attr);
 384         av->timeout = path->packet_life_time + 1;
 385         memcpy(av->smac, path->smac, sizeof(av->smac));
 386
 387         av->valid = 1;
 388         return 0;
 389 }
 390
 391 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 392 {
 393         unsigned long flags;
 394         int id;
 395
 396         idr_preload(GFP_KERNEL);
 397         spin_lock_irqsave(&cm.lock, flags);
 398
 399         id = idr_alloc_cyclic(&cm.local_id_table, cm_id_priv, 0, 0, GFP_NOWAIT);
 400
 401         spin_unlock_irqrestore(&cm.lock, flags);
 402         idr_preload_end();
 403
 404         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 405         return id < 0 ? id : 0;
 406 }
 407
 408 static void cm_free_id(__be32 local_id)
 409 {
 410         spin_lock_irq(&cm.lock);
 411         idr_remove(&cm.local_id_table,
 412                    (__force int) (local_id ^ cm.random_id_operand));
 413         spin_unlock_irq(&cm.lock);
 414 }
 415
 416 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 417 {
 418         struct cm_id_private *cm_id_priv;
 419
 420         cm_id_priv = idr_find(&cm.local_id_table,
 421                               (__force int) (local_id ^ cm.random_id_operand));
 422         if (cm_id_priv) {
 423                 if (cm_id_priv->id.remote_id == remote_id)
 424                         atomic_inc(&cm_id_priv->refcount);
 425                 else
 426                         cm_id_priv = NULL;
 427         }
 428
 429         return cm_id_priv;
 430 }
 431
 432 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 433 {
 434         struct cm_id_private *cm_id_priv;
 435
 436         spin_lock_irq(&cm.lock);
 437         cm_id_priv = cm_get_id(local_id, remote_id);
 438         spin_unlock_irq(&cm.lock);
 439
 440         return cm_id_priv;
 441 }
 442
 443 static void cm_mask_copy(u32 *dst, const u32 *src, const u32 *mask)
 444 {
 445         int i;
 446
 447         for (i = 0; i < IB_CM_COMPARE_SIZE; i++)
 448                 dst[i] = src[i] & mask[i];
 449 }
 450
 451 static int cm_compare_data(struct ib_cm_compare_data *src_data,
 452                            struct ib_cm_compare_data *dst_data)
 453 {
 454         u32 src[IB_CM_COMPARE_SIZE];
 455         u32 dst[IB_CM_COMPARE_SIZE];
 456
 457         if (!src_data || !dst_data)
 458                 return 0;
 459
 460         cm_mask_copy(src, src_data->data, dst_data->mask);
 461         cm_mask_copy(dst, dst_data->data, src_data->mask);
 462         return memcmp(src, dst, sizeof(src));
 463 }
 464
 465 static int cm_compare_private_data(u32 *private_data,
 466                                    struct ib_cm_compare_data *dst_data)
 467 {
 468         u32 src[IB_CM_COMPARE_SIZE];
 469
 470         if (!dst_data)
 471                 return 0;
 472
 473         cm_mask_copy(src, private_data, dst_data->mask);
 474         return memcmp(src, dst_data->data, sizeof(src));
 475 }
 476
 477 /*
 478  * Trivial helpers to strip endian annotation and compare; the
 479  * endianness doesn't actually matter since we just need a stable
 480  * order for the RB tree.
 481  */
 482 static int be32_lt(__be32 a, __be32 b)
 483 {
 484         return (__force u32) a < (__force u32) b;
 485 }
 486
 487 static int be32_gt(__be32 a, __be32 b)
 488 {
 489         return (__force u32) a > (__force u32) b;
 490 }
 491
 492 static int be64_lt(__be64 a, __be64 b)
 493 {
 494         return (__force u64) a < (__force u64) b;
 495 }
 496
 497 static int be64_gt(__be64 a, __be64 b)
 498 {
 499         return (__force u64) a > (__force u64) b;
 500 }
 501
 502 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 503 {
 504         struct rb_node **link = &cm.listen_service_table.rb_node;
 505         struct rb_node *parent = NULL;
 506         struct cm_id_private *cur_cm_id_priv;
 507         __be64 service_id = cm_id_priv->id.service_id;
 508         __be64 service_mask = cm_id_priv->id.service_mask;
 509         int data_cmp;
 510
 511         while (*link) {
 512                 parent = *link;
 513                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 514                                           service_node);
 515                 data_cmp = cm_compare_data(cm_id_priv->compare_data,
 516                                            cur_cm_id_priv->compare_data);
 517                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 518                     (service_mask & cur_cm_id_priv->id.service_id) &&
 519                     (cm_id_priv->id.device == cur_cm_id_priv->id.device) &&
 520                     !data_cmp)
 521                         return cur_cm_id_priv;
 522
 523                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 524                         link = &(*link)->rb_left;
 525                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 526                         link = &(*link)->rb_right;
 527                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 528                         link = &(*link)->rb_left;
 529                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 530                         link = &(*link)->rb_right;
 531                 else if (data_cmp < 0)
 532                         link = &(*link)->rb_left;
 533                 else
 534                         link = &(*link)->rb_right;
 535         }
 536         rb_link_node(&cm_id_priv->service_node, parent, link);
 537         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 538         return NULL;
 539 }
 540
 541 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 542                                              __be64 service_id,
 543                                              u32 *private_data)
 544 {
 545         struct rb_node *node = cm.listen_service_table.rb_node;
 546         struct cm_id_private *cm_id_priv;
 547         int data_cmp;
 548
 549         while (node) {
 550                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 551                 data_cmp = cm_compare_private_data(private_data,
 552                                                    cm_id_priv->compare_data);
 553                 if ((cm_id_priv->id.service_mask & service_id) ==
 554                      cm_id_priv->id.service_id &&
 555                     (cm_id_priv->id.device == device) && !data_cmp)
 556                         return cm_id_priv;
 557
 558                 if (device < cm_id_priv->id.device)
 559                         node = node->rb_left;
 560                 else if (device > cm_id_priv->id.device)
 561                         node = node->rb_right;
 562                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 563                         node = node->rb_left;
 564                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 565                         node = node->rb_right;
 566                 else if (data_cmp < 0)
 567                         node = node->rb_left;
 568                 else
 569                         node = node->rb_right;
 570         }
 571         return NULL;
 572 }
 573
 574 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 575                                                      *timewait_info)
 576 {
 577         struct rb_node **link = &cm.remote_id_table.rb_node;
 578         struct rb_node *parent = NULL;
 579         struct cm_timewait_info *cur_timewait_info;
 580         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 581         __be32 remote_id = timewait_info->work.remote_id;
 582
 583         while (*link) {
 584                 parent = *link;
 585                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 586                                              remote_id_node);
 587                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 588                         link = &(*link)->rb_left;
 589                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 590                         link = &(*link)->rb_right;
 591                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 592                         link = &(*link)->rb_left;
 593                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 594                         link = &(*link)->rb_right;
 595                 else
 596                         return cur_timewait_info;
 597         }
 598         timewait_info->inserted_remote_id = 1;
 599         rb_link_node(&timewait_info->remote_id_node, parent, link);
 600         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 601         return NULL;
 602 }
 603
 604 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 605                                                    __be32 remote_id)
 606 {
 607         struct rb_node *node = cm.remote_id_table.rb_node;
 608         struct cm_timewait_info *timewait_info;
 609
 610         while (node) {
 611                 timewait_info = rb_entry(node, struct cm_timewait_info,
 612                                          remote_id_node);
 613                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 614                         node = node->rb_left;
 615                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 616                         node = node->rb_right;
 617                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 618                         node = node->rb_left;
 619                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 620                         node = node->rb_right;
 621                 else
 622                         return timewait_info;
 623         }
 624         return NULL;
 625 }
 626
 627 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 628                                                       *timewait_info)
 629 {
 630         struct rb_node **link = &cm.remote_qp_table.rb_node;
 631         struct rb_node *parent = NULL;
 632         struct cm_timewait_info *cur_timewait_info;
 633         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 634         __be32 remote_qpn = timewait_info->remote_qpn;
 635
 636         while (*link) {
 637                 parent = *link;
 638                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 639                                              remote_qp_node);
 640                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 641                         link = &(*link)->rb_left;
 642                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 643                         link = &(*link)->rb_right;
 644                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 645                         link = &(*link)->rb_left;
 646                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 647                         link = &(*link)->rb_right;
 648                 else
 649                         return cur_timewait_info;
 650         }
 651         timewait_info->inserted_remote_qp = 1;
 652         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 653         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 654         return NULL;
 655 }
 656
 657 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 658                                                     *cm_id_priv)
 659 {
 660         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 661         struct rb_node *parent = NULL;
 662         struct cm_id_private *cur_cm_id_priv;
 663         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 664         __be32 remote_id = cm_id_priv->id.remote_id;
 665
 666         while (*link) {
 667                 parent = *link;
 668                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 669                                           sidr_id_node);
 670                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 671                         link = &(*link)->rb_left;
 672                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 673                         link = &(*link)->rb_right;
 674                 else {
 675                         int cmp;
 676                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 677                                      sizeof *port_gid);
 678                         if (cmp < 0)
 679                                 link = &(*link)->rb_left;
 680                         else if (cmp > 0)
 681                                 link = &(*link)->rb_right;
 682                         else
 683                                 return cur_cm_id_priv;
 684                 }
 685         }
 686         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 687         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 688         return NULL;
 689 }
 690
 691 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 692                                enum ib_cm_sidr_status status)
 693 {
 694         struct ib_cm_sidr_rep_param param;
 695
 696         memset(&param, 0, sizeof param);
 697         param.status = status;
 698         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 699 }
 700
 701 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 702                                  ib_cm_handler cm_handler,
 703                                  void *context)
 704 {
 705         struct cm_id_private *cm_id_priv;
 706         int ret;
 707
 708         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 709         if (!cm_id_priv)
 710                 return ERR_PTR(-ENOMEM);
 711
 712         cm_id_priv->id.state = IB_CM_IDLE;
 713         cm_id_priv->id.device = device;
 714         cm_id_priv->id.cm_handler = cm_handler;
 715         cm_id_priv->id.context = context;
 716         cm_id_priv->id.remote_cm_qpn = 1;
 717         ret = cm_alloc_id(cm_id_priv);
 718         if (ret)
 719                 goto error;
 720
 721         spin_lock_init(&cm_id_priv->lock);
 722         init_completion(&cm_id_priv->comp);
 723         INIT_LIST_HEAD(&cm_id_priv->work_list);
 724         atomic_set(&cm_id_priv->work_count, -1);
 725         atomic_set(&cm_id_priv->refcount, 1);
 726         return &cm_id_priv->id;
 727
 728 error:
 729         kfree(cm_id_priv);
 730         return ERR_PTR(-ENOMEM);
 731 }
 732 EXPORT_SYMBOL(ib_create_cm_id);
 733
 734 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 735 {
 736         struct cm_work *work;
 737
 738         if (list_empty(&cm_id_priv->work_list))
 739                 return NULL;
 740
 741         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 742         list_del(&work->list);
 743         return work;
 744 }
 745
 746 static void cm_free_work(struct cm_work *work)
 747 {
 748         if (work->mad_recv_wc)
 749                 ib_free_recv_mad(work->mad_recv_wc);
 750         kfree(work);
 751 }
 752
 753 static inline int cm_convert_to_ms(int iba_time)
 754 {
 755         /* approximate conversion to ms from 4.096us x 2^iba_time */
 756         return 1 << max(iba_time - 8, 0);
 757 }
 758
 759 /*
 760  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 761  * Because of how ack_timeout is stored, adding one doubles the timeout.
 762  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 763  * increment it (round up) only if the other is within 50%.
 764  */
 765 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 766 {
 767         int ack_timeout = packet_life_time + 1;
 768
 769         if (ack_timeout >= ca_ack_delay)
 770                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 771         else
 772                 ack_timeout = ca_ack_delay +
 773                               (ack_timeout >= (ca_ack_delay - 1));
 774
 775         return min(31, ack_timeout);
 776 }
 777
 778 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 779 {
 780         if (timewait_info->inserted_remote_id) {
 781                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 782                 timewait_info->inserted_remote_id = 0;
 783         }
 784
 785         if (timewait_info->inserted_remote_qp) {
 786                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 787                 timewait_info->inserted_remote_qp = 0;
 788         }
 789 }
 790
 791 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 792 {
 793         struct cm_timewait_info *timewait_info;
 794
 795         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 796         if (!timewait_info)
 797                 return ERR_PTR(-ENOMEM);
 798
 799         timewait_info->work.local_id = local_id;
 800         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 801         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 802         return timewait_info;
 803 }
 804
 805 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 806 {
 807         int wait_time;
 808         unsigned long flags;
 809         struct cm_device *cm_dev;
 810
 811         cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
 812         if (!cm_dev)
 813                 return;
 814
 815         spin_lock_irqsave(&cm.lock, flags);
 816         cm_cleanup_timewait(cm_id_priv->timewait_info);
 817         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 818         spin_unlock_irqrestore(&cm.lock, flags);
 819
 820         /*
 821          * The cm_id could be destroyed by the user before we exit timewait.
 822          * To protect against this, we search for the cm_id after exiting
 823          * timewait before notifying the user that we've exited timewait.
 824          */
 825         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 826         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 827
 828         /* Check if the device started its remove_one */
 829         spin_lock_irq(&cm.lock);
 830         if (!cm_dev->going_down)
 831                 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 832                                    msecs_to_jiffies(wait_time));
 833         spin_unlock_irq(&cm.lock);
 834
 835         cm_id_priv->timewait_info = NULL;
 836 }
 837
 838 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 839 {
 840         unsigned long flags;
 841
 842         cm_id_priv->id.state = IB_CM_IDLE;
 843         if (cm_id_priv->timewait_info) {
 844                 spin_lock_irqsave(&cm.lock, flags);
 845                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 846                 spin_unlock_irqrestore(&cm.lock, flags);
 847                 kfree(cm_id_priv->timewait_info);
 848                 cm_id_priv->timewait_info = NULL;
 849         }
 850 }
 851
 852 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 853 {
 854         struct cm_id_private *cm_id_priv;
 855         struct cm_work *work;
 856
 857         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 858 retest:
 859         spin_lock_irq(&cm_id_priv->lock);
 860         switch (cm_id->state) {
 861         case IB_CM_LISTEN:
 862                 cm_id->state = IB_CM_IDLE;
 863                 spin_unlock_irq(&cm_id_priv->lock);
 864                 spin_lock_irq(&cm.lock);
 865                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 866                 spin_unlock_irq(&cm.lock);
 867                 break;
 868         case IB_CM_SIDR_REQ_SENT:
 869                 cm_id->state = IB_CM_IDLE;
 870                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 871                 spin_unlock_irq(&cm_id_priv->lock);
 872                 break;
 873         case IB_CM_SIDR_REQ_RCVD:
 874                 spin_unlock_irq(&cm_id_priv->lock);
 875                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 876                 break;
 877         case IB_CM_REQ_SENT:
 878         case IB_CM_MRA_REQ_RCVD:
 879                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 880                 spin_unlock_irq(&cm_id_priv->lock);
 881                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 882                                &cm_id_priv->id.device->node_guid,
 883                                sizeof cm_id_priv->id.device->node_guid,
 884                                NULL, 0);
 885                 break;
 886         case IB_CM_REQ_RCVD:
 887                 if (err == -ENOMEM) {
 888                         /* Do not reject to allow future retries. */
 889                         cm_reset_to_idle(cm_id_priv);
 890                         spin_unlock_irq(&cm_id_priv->lock);
 891                 } else {
 892                         spin_unlock_irq(&cm_id_priv->lock);
 893                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 894                                        NULL, 0, NULL, 0);
 895                 }
 896                 break;
 897         case IB_CM_REP_SENT:
 898         case IB_CM_MRA_REP_RCVD:
 899                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 900                 /* Fall through */
 901         case IB_CM_MRA_REQ_SENT:
 902         case IB_CM_REP_RCVD:
 903         case IB_CM_MRA_REP_SENT:
 904                 spin_unlock_irq(&cm_id_priv->lock);
 905                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 906                                NULL, 0, NULL, 0);
 907                 break;
 908         case IB_CM_ESTABLISHED:
 909                 spin_unlock_irq(&cm_id_priv->lock);
 910                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT)
 911                         break;
 912                 ib_send_cm_dreq(cm_id, NULL, 0);
 913                 goto retest;
 914         case IB_CM_DREQ_SENT:
 915                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 916                 cm_enter_timewait(cm_id_priv);
 917                 spin_unlock_irq(&cm_id_priv->lock);
 918                 break;
 919         case IB_CM_DREQ_RCVD:
 920                 spin_unlock_irq(&cm_id_priv->lock);
 921                 ib_send_cm_drep(cm_id, NULL, 0);
 922                 break;
 923         default:
 924                 spin_unlock_irq(&cm_id_priv->lock);
 925                 break;
 926         }
 927
 928         cm_free_id(cm_id->local_id);
 929         cm_deref_id(cm_id_priv);
 930         wait_for_completion(&cm_id_priv->comp);
 931         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
 932                 cm_free_work(work);
 933         kfree(cm_id_priv->compare_data);
 934         kfree(cm_id_priv->private_data);
 935         kfree(cm_id_priv);
 936 }
 937
 938 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
 939 {
 940         cm_destroy_id(cm_id, 0);
 941 }
 942 EXPORT_SYMBOL(ib_destroy_cm_id);
 943
 944 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask,
 945                  struct ib_cm_compare_data *compare_data)
 946 {
 947         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
 948         unsigned long flags;
 949         int ret = 0;
 950
 951         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
 952         service_id &= service_mask;
 953         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
 954             (service_id != IB_CM_ASSIGN_SERVICE_ID))
 955                 return -EINVAL;
 956
 957         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 958         if (cm_id->state != IB_CM_IDLE)
 959                 return -EINVAL;
 960
 961         if (compare_data) {
 962                 cm_id_priv->compare_data = kzalloc(sizeof *compare_data,
 963                                                    GFP_KERNEL);
 964                 if (!cm_id_priv->compare_data)
 965                         return -ENOMEM;
 966                 cm_mask_copy(cm_id_priv->compare_data->data,
 967                              compare_data->data, compare_data->mask);
 968                 memcpy(cm_id_priv->compare_data->mask, compare_data->mask,
 969                        sizeof(compare_data->mask));
 970         }
 971
 972         cm_id->state = IB_CM_LISTEN;
 973
 974         spin_lock_irqsave(&cm.lock, flags);
 975         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
 976                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
 977                 cm_id->service_mask = ~cpu_to_be64(0);
 978         } else {
 979                 cm_id->service_id = service_id;
 980                 cm_id->service_mask = service_mask;
 981         }
 982         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
 983         spin_unlock_irqrestore(&cm.lock, flags);
 984
 985         if (cur_cm_id_priv) {
 986                 cm_id->state = IB_CM_IDLE;
 987                 kfree(cm_id_priv->compare_data);
 988                 cm_id_priv->compare_data = NULL;
 989                 ret = -EBUSY;
 990         }
 991         return ret;
 992 }
 993 EXPORT_SYMBOL(ib_cm_listen);
 994
 995 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
 996                           enum cm_msg_sequence msg_seq)
 997 {
 998         u64 hi_tid, low_tid;
 999
1000         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1001         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
1002                           (msg_seq << 30));
1003         return cpu_to_be64(hi_tid | low_tid);
1004 }
1005
1006 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1007                               __be16 attr_id, __be64 tid)
1008 {
1009         hdr->base_version  = IB_MGMT_BASE_VERSION;
1010         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1011         hdr->class_version = IB_CM_CLASS_VERSION;
1012         hdr->method        = IB_MGMT_METHOD_SEND;
1013         hdr->attr_id       = attr_id;
1014         hdr->tid           = tid;
1015 }
1016
1017 static void cm_format_req(struct cm_req_msg *req_msg,
1018                           struct cm_id_private *cm_id_priv,
1019                           struct ib_cm_req_param *param)
1020 {
1021         struct ib_sa_path_rec *pri_path = param->primary_path;
1022         struct ib_sa_path_rec *alt_path = param->alternate_path;
1023
1024         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1025                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1026
1027         req_msg->local_comm_id = cm_id_priv->id.local_id;
1028         req_msg->service_id = param->service_id;
1029         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1030         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1031         cm_req_set_init_depth(req_msg, param->initiator_depth);
1032         cm_req_set_remote_resp_timeout(req_msg,
1033                                        param->remote_cm_response_timeout);
1034         cm_req_set_qp_type(req_msg, param->qp_type);
1035         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1036         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1037         cm_req_set_local_resp_timeout(req_msg,
1038                                       param->local_cm_response_timeout);
1039         req_msg->pkey = param->primary_path->pkey;
1040         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1041         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1042
1043         if (param->qp_type != IB_QPT_XRC_INI) {
1044                 cm_req_set_resp_res(req_msg, param->responder_resources);
1045                 cm_req_set_retry_count(req_msg, param->retry_count);
1046                 cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1047                 cm_req_set_srq(req_msg, param->srq);
1048         }
1049
1050         if (pri_path->hop_limit <= 1) {
1051                 req_msg->primary_local_lid = pri_path->slid;
1052                 req_msg->primary_remote_lid = pri_path->dlid;
1053         } else {
1054                 /* Work-around until there's a way to obtain remote LID info */
1055                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1056                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1057         }
1058         req_msg->primary_local_gid = pri_path->sgid;
1059         req_msg->primary_remote_gid = pri_path->dgid;
1060         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1061         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1062         req_msg->primary_traffic_class = pri_path->traffic_class;
1063         req_msg->primary_hop_limit = pri_path->hop_limit;
1064         cm_req_set_primary_sl(req_msg, pri_path->sl);
1065         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1066         cm_req_set_primary_local_ack_timeout(req_msg,
1067                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1068                                pri_path->packet_life_time));
1069
1070         if (alt_path) {
1071                 if (alt_path->hop_limit <= 1) {
1072                         req_msg->alt_local_lid = alt_path->slid;
1073                         req_msg->alt_remote_lid = alt_path->dlid;
1074                 } else {
1075                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1076                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1077                 }
1078                 req_msg->alt_local_gid = alt_path->sgid;
1079                 req_msg->alt_remote_gid = alt_path->dgid;
1080                 cm_req_set_alt_flow_label(req_msg,
1081                                           alt_path->flow_label);
1082                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1083                 req_msg->alt_traffic_class = alt_path->traffic_class;
1084                 req_msg->alt_hop_limit = alt_path->hop_limit;
1085                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1086                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1087                 cm_req_set_alt_local_ack_timeout(req_msg,
1088                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1089                                        alt_path->packet_life_time));
1090         }
1091
1092         if (param->private_data && param->private_data_len)
1093                 memcpy(req_msg->private_data, param->private_data,
1094                        param->private_data_len);
1095 }
1096
1097 static int cm_validate_req_param(struct ib_cm_req_param *param)
1098 {
1099         /* peer-to-peer not supported */
1100         if (param->peer_to_peer)
1101                 return -EINVAL;
1102
1103         if (!param->primary_path)
1104                 return -EINVAL;
1105
1106         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1107             param->qp_type != IB_QPT_XRC_INI)
1108                 return -EINVAL;
1109
1110         if (param->private_data &&
1111             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1112                 return -EINVAL;
1113
1114         if (param->alternate_path &&
1115             (param->alternate_path->pkey != param->primary_path->pkey ||
1116              param->alternate_path->mtu != param->primary_path->mtu))
1117                 return -EINVAL;
1118
1119         return 0;
1120 }
1121
1122 int ib_send_cm_req(struct ib_cm_id *cm_id,
1123                    struct ib_cm_req_param *param)
1124 {
1125         struct cm_id_private *cm_id_priv;
1126         struct cm_req_msg *req_msg;
1127         unsigned long flags;
1128         int ret;
1129
1130         ret = cm_validate_req_param(param);
1131         if (ret)
1132                 return ret;
1133
1134         /* Verify that we're not in timewait. */
1135         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1136         spin_lock_irqsave(&cm_id_priv->lock, flags);
1137         if (cm_id->state != IB_CM_IDLE) {
1138                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1139                 ret = -EINVAL;
1140                 goto out;
1141         }
1142         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1143
1144         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1145                                                             id.local_id);
1146         if (IS_ERR(cm_id_priv->timewait_info)) {
1147                 ret = PTR_ERR(cm_id_priv->timewait_info);
1148                 goto out;
1149         }
1150
1151         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av);
1152         if (ret)
1153                 goto error1;
1154         if (param->alternate_path) {
1155                 ret = cm_init_av_by_path(param->alternate_path,
1156                                          &cm_id_priv->alt_av);
1157                 if (ret)
1158                         goto error1;
1159         }
1160         cm_id->service_id = param->service_id;
1161         cm_id->service_mask = ~cpu_to_be64(0);
1162         cm_id_priv->timeout_ms = cm_convert_to_ms(
1163                                     param->primary_path->packet_life_time) * 2 +
1164                                  cm_convert_to_ms(
1165                                     param->remote_cm_response_timeout);
1166         cm_id_priv->max_cm_retries = param->max_cm_retries;
1167         cm_id_priv->initiator_depth = param->initiator_depth;
1168         cm_id_priv->responder_resources = param->responder_resources;
1169         cm_id_priv->retry_count = param->retry_count;
1170         cm_id_priv->path_mtu = param->primary_path->mtu;
1171         cm_id_priv->pkey = param->primary_path->pkey;
1172         cm_id_priv->qp_type = param->qp_type;
1173
1174         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1175         if (ret)
1176                 goto error1;
1177
1178         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1179         cm_format_req(req_msg, cm_id_priv, param);
1180         cm_id_priv->tid = req_msg->hdr.tid;
1181         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1182         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1183
1184         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1185         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1186
1187         spin_lock_irqsave(&cm_id_priv->lock, flags);
1188         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1189         if (ret) {
1190                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1191                 goto error2;
1192         }
1193         BUG_ON(cm_id->state != IB_CM_IDLE);
1194         cm_id->state = IB_CM_REQ_SENT;
1195         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1196         return 0;
1197
1198 error2: cm_free_msg(cm_id_priv->msg);
1199 error1: kfree(cm_id_priv->timewait_info);
1200 out:    return ret;
1201 }
1202 EXPORT_SYMBOL(ib_send_cm_req);
1203
1204 static int cm_issue_rej(struct cm_port *port,
1205                         struct ib_mad_recv_wc *mad_recv_wc,
1206                         enum ib_cm_rej_reason reason,
1207                         enum cm_msg_response msg_rejected,
1208                         void *ari, u8 ari_length)
1209 {
1210         struct ib_mad_send_buf *msg = NULL;
1211         struct cm_rej_msg *rej_msg, *rcv_msg;
1212         int ret;
1213
1214         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1215         if (ret)
1216                 return ret;
1217
1218         /* We just need common CM header information.  Cast to any message. */
1219         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1220         rej_msg = (struct cm_rej_msg *) msg->mad;
1221
1222         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1223         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1224         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1225         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1226         rej_msg->reason = cpu_to_be16(reason);
1227
1228         if (ari && ari_length) {
1229                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1230                 memcpy(rej_msg->ari, ari, ari_length);
1231         }
1232
1233         ret = ib_post_send_mad(msg, NULL);
1234         if (ret)
1235                 cm_free_msg(msg);
1236
1237         return ret;
1238 }
1239
1240 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1241                                     __be32 local_qpn, __be32 remote_qpn)
1242 {
1243         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1244                 ((local_ca_guid == remote_ca_guid) &&
1245                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1246 }
1247
1248 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1249                                             struct ib_sa_path_rec *primary_path,
1250                                             struct ib_sa_path_rec *alt_path)
1251 {
1252         memset(primary_path, 0, sizeof *primary_path);
1253         primary_path->dgid = req_msg->primary_local_gid;
1254         primary_path->sgid = req_msg->primary_remote_gid;
1255         primary_path->dlid = req_msg->primary_local_lid;
1256         primary_path->slid = req_msg->primary_remote_lid;
1257         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1258         primary_path->hop_limit = req_msg->primary_hop_limit;
1259         primary_path->traffic_class = req_msg->primary_traffic_class;
1260         primary_path->reversible = 1;
1261         primary_path->pkey = req_msg->pkey;
1262         primary_path->sl = cm_req_get_primary_sl(req_msg);
1263         primary_path->mtu_selector = IB_SA_EQ;
1264         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1265         primary_path->rate_selector = IB_SA_EQ;
1266         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1267         primary_path->packet_life_time_selector = IB_SA_EQ;
1268         primary_path->packet_life_time =
1269                 cm_req_get_primary_local_ack_timeout(req_msg);
1270         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1271
1272         if (req_msg->alt_local_lid) {
1273                 memset(alt_path, 0, sizeof *alt_path);
1274                 alt_path->dgid = req_msg->alt_local_gid;
1275                 alt_path->sgid = req_msg->alt_remote_gid;
1276                 alt_path->dlid = req_msg->alt_local_lid;
1277                 alt_path->slid = req_msg->alt_remote_lid;
1278                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1279                 alt_path->hop_limit = req_msg->alt_hop_limit;
1280                 alt_path->traffic_class = req_msg->alt_traffic_class;
1281                 alt_path->reversible = 1;
1282                 alt_path->pkey = req_msg->pkey;
1283                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1284                 alt_path->mtu_selector = IB_SA_EQ;
1285                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1286                 alt_path->rate_selector = IB_SA_EQ;
1287                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1288                 alt_path->packet_life_time_selector = IB_SA_EQ;
1289                 alt_path->packet_life_time =
1290                         cm_req_get_alt_local_ack_timeout(req_msg);
1291                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1292         }
1293 }
1294
1295 static void cm_format_req_event(struct cm_work *work,
1296                                 struct cm_id_private *cm_id_priv,
1297                                 struct ib_cm_id *listen_id)
1298 {
1299         struct cm_req_msg *req_msg;
1300         struct ib_cm_req_event_param *param;
1301
1302         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1303         param = &work->cm_event.param.req_rcvd;
1304         param->listen_id = listen_id;
1305         param->port = cm_id_priv->av.port->port_num;
1306         param->primary_path = &work->path[0];
1307         if (req_msg->alt_local_lid)
1308                 param->alternate_path = &work->path[1];
1309         else
1310                 param->alternate_path = NULL;
1311         param->remote_ca_guid = req_msg->local_ca_guid;
1312         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1313         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1314         param->qp_type = cm_req_get_qp_type(req_msg);
1315         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1316         param->responder_resources = cm_req_get_init_depth(req_msg);
1317         param->initiator_depth = cm_req_get_resp_res(req_msg);
1318         param->local_cm_response_timeout =
1319                                         cm_req_get_remote_resp_timeout(req_msg);
1320         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1321         param->remote_cm_response_timeout =
1322                                         cm_req_get_local_resp_timeout(req_msg);
1323         param->retry_count = cm_req_get_retry_count(req_msg);
1324         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1325         param->srq = cm_req_get_srq(req_msg);
1326         work->cm_event.private_data = &req_msg->private_data;
1327 }
1328
1329 static void cm_process_work(struct cm_id_private *cm_id_priv,
1330                             struct cm_work *work)
1331 {
1332         int ret;
1333
1334         /* We will typically only have the current event to report. */
1335         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1336         cm_free_work(work);
1337
1338         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1339                 spin_lock_irq(&cm_id_priv->lock);
1340                 work = cm_dequeue_work(cm_id_priv);
1341                 spin_unlock_irq(&cm_id_priv->lock);
1342                 BUG_ON(!work);
1343                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1344                                                 &work->cm_event);
1345                 cm_free_work(work);
1346         }
1347         cm_deref_id(cm_id_priv);
1348         if (ret)
1349                 cm_destroy_id(&cm_id_priv->id, ret);
1350 }
1351
1352 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1353                           struct cm_id_private *cm_id_priv,
1354                           enum cm_msg_response msg_mraed, u8 service_timeout,
1355                           const void *private_data, u8 private_data_len)
1356 {
1357         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1358         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1359         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1360         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1361         cm_mra_set_service_timeout(mra_msg, service_timeout);
1362
1363         if (private_data && private_data_len)
1364                 memcpy(mra_msg->private_data, private_data, private_data_len);
1365 }
1366
1367 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1368                           struct cm_id_private *cm_id_priv,
1369                           enum ib_cm_rej_reason reason,
1370                           void *ari,
1371                           u8 ari_length,
1372                           const void *private_data,
1373                           u8 private_data_len)
1374 {
1375         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1376         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1377
1378         switch(cm_id_priv->id.state) {
1379         case IB_CM_REQ_RCVD:
1380                 rej_msg->local_comm_id = 0;
1381                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1382                 break;
1383         case IB_CM_MRA_REQ_SENT:
1384                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1385                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1386                 break;
1387         case IB_CM_REP_RCVD:
1388         case IB_CM_MRA_REP_SENT:
1389                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1390                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1391                 break;
1392         default:
1393                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1394                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1395                 break;
1396         }
1397
1398         rej_msg->reason = cpu_to_be16(reason);
1399         if (ari && ari_length) {
1400                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1401                 memcpy(rej_msg->ari, ari, ari_length);
1402         }
1403
1404         if (private_data && private_data_len)
1405                 memcpy(rej_msg->private_data, private_data, private_data_len);
1406 }
1407
1408 static void cm_dup_req_handler(struct cm_work *work,
1409                                struct cm_id_private *cm_id_priv)
1410 {
1411         struct ib_mad_send_buf *msg = NULL;
1412         int ret;
1413
1414         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1415                         counter[CM_REQ_COUNTER]);
1416
1417         /* Quick state check to discard duplicate REQs. */
1418         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1419                 return;
1420
1421         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1422         if (ret)
1423                 return;
1424
1425         spin_lock_irq(&cm_id_priv->lock);
1426         switch (cm_id_priv->id.state) {
1427         case IB_CM_MRA_REQ_SENT:
1428                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1429                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1430                               cm_id_priv->private_data,
1431                               cm_id_priv->private_data_len);
1432                 break;
1433         case IB_CM_TIMEWAIT:
1434                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1435                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1436                 break;
1437         default:
1438                 goto unlock;
1439         }
1440         spin_unlock_irq(&cm_id_priv->lock);
1441
1442         ret = ib_post_send_mad(msg, NULL);
1443         if (ret)
1444                 goto free;
1445         return;
1446
1447 unlock: spin_unlock_irq(&cm_id_priv->lock);
1448 free:   cm_free_msg(msg);
1449 }
1450
1451 static struct cm_id_private * cm_match_req(struct cm_work *work,
1452                                            struct cm_id_private *cm_id_priv)
1453 {
1454         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1455         struct cm_timewait_info *timewait_info;
1456         struct cm_req_msg *req_msg;
1457
1458         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1459
1460         /* Check for possible duplicate REQ. */
1461         spin_lock_irq(&cm.lock);
1462         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1463         if (timewait_info) {
1464                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1465                                            timewait_info->work.remote_id);
1466                 spin_unlock_irq(&cm.lock);
1467                 if (cur_cm_id_priv) {
1468                         cm_dup_req_handler(work, cur_cm_id_priv);
1469                         cm_deref_id(cur_cm_id_priv);
1470                 }
1471                 return NULL;
1472         }
1473
1474         /* Check for stale connections. */
1475         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1476         if (timewait_info) {
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_STALE_CONN, CM_MSG_RESPONSE_REQ,
1481                              NULL, 0);
1482                 return NULL;
1483         }
1484
1485         /* Find matching listen request. */
1486         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1487                                            req_msg->service_id,
1488                                            req_msg->private_data);
1489         if (!listen_cm_id_priv) {
1490                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1491                 spin_unlock_irq(&cm.lock);
1492                 cm_issue_rej(work->port, work->mad_recv_wc,
1493                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1494                              NULL, 0);
1495                 goto out;
1496         }
1497         atomic_inc(&listen_cm_id_priv->refcount);
1498         atomic_inc(&cm_id_priv->refcount);
1499         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1500         atomic_inc(&cm_id_priv->work_count);
1501         spin_unlock_irq(&cm.lock);
1502 out:
1503         return listen_cm_id_priv;
1504 }
1505
1506 /*
1507  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1508  * we need to override the LID/SL data in the REQ with the LID information
1509  * in the work completion.
1510  */
1511 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1512 {
1513         if (!cm_req_get_primary_subnet_local(req_msg)) {
1514                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1515                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1516                         cm_req_set_primary_sl(req_msg, wc->sl);
1517                 }
1518
1519                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1520                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1521         }
1522
1523         if (!cm_req_get_alt_subnet_local(req_msg)) {
1524                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1525                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1526                         cm_req_set_alt_sl(req_msg, wc->sl);
1527                 }
1528
1529                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1530                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1531         }
1532 }
1533
1534 static int cm_req_handler(struct cm_work *work)
1535 {
1536         struct ib_cm_id *cm_id;
1537         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1538         struct cm_req_msg *req_msg;
1539         int ret;
1540
1541         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1542
1543         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1544         if (IS_ERR(cm_id))
1545                 return PTR_ERR(cm_id);
1546
1547         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1548         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1549         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1550                                 work->mad_recv_wc->recv_buf.grh,
1551                                 &cm_id_priv->av);
1552         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1553                                                             id.local_id);
1554         if (IS_ERR(cm_id_priv->timewait_info)) {
1555                 ret = PTR_ERR(cm_id_priv->timewait_info);
1556                 goto destroy;
1557         }
1558         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1559         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1560         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1561
1562         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1563         if (!listen_cm_id_priv) {
1564                 ret = -EINVAL;
1565                 kfree(cm_id_priv->timewait_info);
1566                 goto destroy;
1567         }
1568
1569         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1570         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1571         cm_id_priv->id.service_id = req_msg->service_id;
1572         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1573
1574         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1575         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1576
1577         memcpy(work->path[0].dmac, cm_id_priv->av.ah_attr.dmac, ETH_ALEN);
1578         work->path[0].vlan_id = cm_id_priv->av.ah_attr.vlan_id;
1579         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
1580         if (ret) {
1581                 ib_get_cached_gid(work->port->cm_dev->ib_device,
1582                                   work->port->port_num, 0, &work->path[0].sgid);
1583                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1584                                &work->path[0].sgid, sizeof work->path[0].sgid,
1585                                NULL, 0);
1586                 goto rejected;
1587         }
1588         if (req_msg->alt_local_lid) {
1589                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
1590                 if (ret) {
1591                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1592                                        &work->path[0].sgid,
1593                                        sizeof work->path[0].sgid, NULL, 0);
1594                         goto rejected;
1595                 }
1596         }
1597         cm_id_priv->tid = req_msg->hdr.tid;
1598         cm_id_priv->timeout_ms = cm_convert_to_ms(
1599                                         cm_req_get_local_resp_timeout(req_msg));
1600         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1601         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1602         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1603         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1604         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1605         cm_id_priv->pkey = req_msg->pkey;
1606         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1607         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1608         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1609         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1610
1611         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1612         cm_process_work(cm_id_priv, work);
1613         cm_deref_id(listen_cm_id_priv);
1614         return 0;
1615
1616 rejected:
1617         atomic_dec(&cm_id_priv->refcount);
1618         cm_deref_id(listen_cm_id_priv);
1619 destroy:
1620         ib_destroy_cm_id(cm_id);
1621         return ret;
1622 }
1623
1624 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1625                           struct cm_id_private *cm_id_priv,
1626                           struct ib_cm_rep_param *param)
1627 {
1628         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1629         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1630         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1631         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1632         rep_msg->resp_resources = param->responder_resources;
1633         cm_rep_set_target_ack_delay(rep_msg,
1634                                     cm_id_priv->av.port->cm_dev->ack_delay);
1635         cm_rep_set_failover(rep_msg, param->failover_accepted);
1636         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1637         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1638
1639         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1640                 rep_msg->initiator_depth = param->initiator_depth;
1641                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1642                 cm_rep_set_srq(rep_msg, param->srq);
1643                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1644         } else {
1645                 cm_rep_set_srq(rep_msg, 1);
1646                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1647         }
1648
1649         if (param->private_data && param->private_data_len)
1650                 memcpy(rep_msg->private_data, param->private_data,
1651                        param->private_data_len);
1652 }
1653
1654 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1655                    struct ib_cm_rep_param *param)
1656 {
1657         struct cm_id_private *cm_id_priv;
1658         struct ib_mad_send_buf *msg;
1659         struct cm_rep_msg *rep_msg;
1660         unsigned long flags;
1661         int ret;
1662
1663         if (param->private_data &&
1664             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1665                 return -EINVAL;
1666
1667         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1668         spin_lock_irqsave(&cm_id_priv->lock, flags);
1669         if (cm_id->state != IB_CM_REQ_RCVD &&
1670             cm_id->state != IB_CM_MRA_REQ_SENT) {
1671                 ret = -EINVAL;
1672                 goto out;
1673         }
1674
1675         ret = cm_alloc_msg(cm_id_priv, &msg);
1676         if (ret)
1677                 goto out;
1678
1679         rep_msg = (struct cm_rep_msg *) msg->mad;
1680         cm_format_rep(rep_msg, cm_id_priv, param);
1681         msg->timeout_ms = cm_id_priv->timeout_ms;
1682         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1683
1684         ret = ib_post_send_mad(msg, NULL);
1685         if (ret) {
1686                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1687                 cm_free_msg(msg);
1688                 return ret;
1689         }
1690
1691         cm_id->state = IB_CM_REP_SENT;
1692         cm_id_priv->msg = msg;
1693         cm_id_priv->initiator_depth = param->initiator_depth;
1694         cm_id_priv->responder_resources = param->responder_resources;
1695         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1696         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
1697
1698 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1699         return ret;
1700 }
1701 EXPORT_SYMBOL(ib_send_cm_rep);
1702
1703 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1704                           struct cm_id_private *cm_id_priv,
1705                           const void *private_data,
1706                           u8 private_data_len)
1707 {
1708         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1709         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1710         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1711
1712         if (private_data && private_data_len)
1713                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1714 }
1715
1716 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1717                    const void *private_data,
1718                    u8 private_data_len)
1719 {
1720         struct cm_id_private *cm_id_priv;
1721         struct ib_mad_send_buf *msg;
1722         unsigned long flags;
1723         void *data;
1724         int ret;
1725
1726         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1727                 return -EINVAL;
1728
1729         data = cm_copy_private_data(private_data, private_data_len);
1730         if (IS_ERR(data))
1731                 return PTR_ERR(data);
1732
1733         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1734         spin_lock_irqsave(&cm_id_priv->lock, flags);
1735         if (cm_id->state != IB_CM_REP_RCVD &&
1736             cm_id->state != IB_CM_MRA_REP_SENT) {
1737                 ret = -EINVAL;
1738                 goto error;
1739         }
1740
1741         ret = cm_alloc_msg(cm_id_priv, &msg);
1742         if (ret)
1743                 goto error;
1744
1745         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1746                       private_data, private_data_len);
1747
1748         ret = ib_post_send_mad(msg, NULL);
1749         if (ret) {
1750                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1751                 cm_free_msg(msg);
1752                 kfree(data);
1753                 return ret;
1754         }
1755
1756         cm_id->state = IB_CM_ESTABLISHED;
1757         cm_set_private_data(cm_id_priv, data, private_data_len);
1758         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1759         return 0;
1760
1761 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1762         kfree(data);
1763         return ret;
1764 }
1765 EXPORT_SYMBOL(ib_send_cm_rtu);
1766
1767 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
1768 {
1769         struct cm_rep_msg *rep_msg;
1770         struct ib_cm_rep_event_param *param;
1771
1772         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1773         param = &work->cm_event.param.rep_rcvd;
1774         param->remote_ca_guid = rep_msg->local_ca_guid;
1775         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1776         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
1777         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1778         param->responder_resources = rep_msg->initiator_depth;
1779         param->initiator_depth = rep_msg->resp_resources;
1780         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1781         param->failover_accepted = cm_rep_get_failover(rep_msg);
1782         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1783         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1784         param->srq = cm_rep_get_srq(rep_msg);
1785         work->cm_event.private_data = &rep_msg->private_data;
1786 }
1787
1788 static void cm_dup_rep_handler(struct cm_work *work)
1789 {
1790         struct cm_id_private *cm_id_priv;
1791         struct cm_rep_msg *rep_msg;
1792         struct ib_mad_send_buf *msg = NULL;
1793         int ret;
1794
1795         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1796         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1797                                    rep_msg->local_comm_id);
1798         if (!cm_id_priv)
1799                 return;
1800
1801         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1802                         counter[CM_REP_COUNTER]);
1803         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1804         if (ret)
1805                 goto deref;
1806
1807         spin_lock_irq(&cm_id_priv->lock);
1808         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1809                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1810                               cm_id_priv->private_data,
1811                               cm_id_priv->private_data_len);
1812         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1813                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1814                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1815                               cm_id_priv->private_data,
1816                               cm_id_priv->private_data_len);
1817         else
1818                 goto unlock;
1819         spin_unlock_irq(&cm_id_priv->lock);
1820
1821         ret = ib_post_send_mad(msg, NULL);
1822         if (ret)
1823                 goto free;
1824         goto deref;
1825
1826 unlock: spin_unlock_irq(&cm_id_priv->lock);
1827 free:   cm_free_msg(msg);
1828 deref:  cm_deref_id(cm_id_priv);
1829 }
1830
1831 static int cm_rep_handler(struct cm_work *work)
1832 {
1833         struct cm_id_private *cm_id_priv;
1834         struct cm_rep_msg *rep_msg;
1835         int ret;
1836
1837         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1838         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1839         if (!cm_id_priv) {
1840                 cm_dup_rep_handler(work);
1841                 return -EINVAL;
1842         }
1843
1844         cm_format_rep_event(work, cm_id_priv->qp_type);
1845
1846         spin_lock_irq(&cm_id_priv->lock);
1847         switch (cm_id_priv->id.state) {
1848         case IB_CM_REQ_SENT:
1849         case IB_CM_MRA_REQ_RCVD:
1850                 break;
1851         default:
1852                 spin_unlock_irq(&cm_id_priv->lock);
1853                 ret = -EINVAL;
1854                 goto error;
1855         }
1856
1857         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1858         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1859         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
1860
1861         spin_lock(&cm.lock);
1862         /* Check for duplicate REP. */
1863         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1864                 spin_unlock(&cm.lock);
1865                 spin_unlock_irq(&cm_id_priv->lock);
1866                 ret = -EINVAL;
1867                 goto error;
1868         }
1869         /* Check for a stale connection. */
1870         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
1871                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
1872                          &cm.remote_id_table);
1873                 cm_id_priv->timewait_info->inserted_remote_id = 0;
1874                 spin_unlock(&cm.lock);
1875                 spin_unlock_irq(&cm_id_priv->lock);
1876                 cm_issue_rej(work->port, work->mad_recv_wc,
1877                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
1878                              NULL, 0);
1879                 ret = -EINVAL;
1880                 goto error;
1881         }
1882         spin_unlock(&cm.lock);
1883
1884         cm_id_priv->id.state = IB_CM_REP_RCVD;
1885         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
1886         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
1887         cm_id_priv->initiator_depth = rep_msg->resp_resources;
1888         cm_id_priv->responder_resources = rep_msg->initiator_depth;
1889         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
1890         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1891         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1892         cm_id_priv->av.timeout =
1893                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1894                                        cm_id_priv->av.timeout - 1);
1895         cm_id_priv->alt_av.timeout =
1896                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1897                                        cm_id_priv->alt_av.timeout - 1);
1898
1899         /* todo: handle peer_to_peer */
1900
1901         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1902         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1903         if (!ret)
1904                 list_add_tail(&work->list, &cm_id_priv->work_list);
1905         spin_unlock_irq(&cm_id_priv->lock);
1906
1907         if (ret)
1908                 cm_process_work(cm_id_priv, work);
1909         else
1910                 cm_deref_id(cm_id_priv);
1911         return 0;
1912
1913 error:
1914         cm_deref_id(cm_id_priv);
1915         return ret;
1916 }
1917
1918 static int cm_establish_handler(struct cm_work *work)
1919 {
1920         struct cm_id_private *cm_id_priv;
1921         int ret;
1922
1923         /* See comment in cm_establish about lookup. */
1924         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
1925         if (!cm_id_priv)
1926                 return -EINVAL;
1927
1928         spin_lock_irq(&cm_id_priv->lock);
1929         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
1930                 spin_unlock_irq(&cm_id_priv->lock);
1931                 goto out;
1932         }
1933
1934         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1935         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1936         if (!ret)
1937                 list_add_tail(&work->list, &cm_id_priv->work_list);
1938         spin_unlock_irq(&cm_id_priv->lock);
1939
1940         if (ret)
1941                 cm_process_work(cm_id_priv, work);
1942         else
1943                 cm_deref_id(cm_id_priv);
1944         return 0;
1945 out:
1946         cm_deref_id(cm_id_priv);
1947         return -EINVAL;
1948 }
1949
1950 static int cm_rtu_handler(struct cm_work *work)
1951 {
1952         struct cm_id_private *cm_id_priv;
1953         struct cm_rtu_msg *rtu_msg;
1954         int ret;
1955
1956         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
1957         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
1958                                    rtu_msg->local_comm_id);
1959         if (!cm_id_priv)
1960                 return -EINVAL;
1961
1962         work->cm_event.private_data = &rtu_msg->private_data;
1963
1964         spin_lock_irq(&cm_id_priv->lock);
1965         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
1966             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
1967                 spin_unlock_irq(&cm_id_priv->lock);
1968                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1969                                 counter[CM_RTU_COUNTER]);
1970                 goto out;
1971         }
1972         cm_id_priv->id.state = IB_CM_ESTABLISHED;
1973
1974         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1975         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1976         if (!ret)
1977                 list_add_tail(&work->list, &cm_id_priv->work_list);
1978         spin_unlock_irq(&cm_id_priv->lock);
1979
1980         if (ret)
1981                 cm_process_work(cm_id_priv, work);
1982         else
1983                 cm_deref_id(cm_id_priv);
1984         return 0;
1985 out:
1986         cm_deref_id(cm_id_priv);
1987         return -EINVAL;
1988 }
1989
1990 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
1991                           struct cm_id_private *cm_id_priv,
1992                           const void *private_data,
1993                           u8 private_data_len)
1994 {
1995         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
1996                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
1997         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
1998         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
1999         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2000
2001         if (private_data && private_data_len)
2002                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2003 }
2004
2005 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2006                     const void *private_data,
2007                     u8 private_data_len)
2008 {
2009         struct cm_id_private *cm_id_priv;
2010         struct ib_mad_send_buf *msg;
2011         unsigned long flags;
2012         int ret;
2013
2014         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2015                 return -EINVAL;
2016
2017         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2018         spin_lock_irqsave(&cm_id_priv->lock, flags);
2019         if (cm_id->state != IB_CM_ESTABLISHED) {
2020                 ret = -EINVAL;
2021                 goto out;
2022         }
2023
2024         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2025             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2026                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2027
2028         ret = cm_alloc_msg(cm_id_priv, &msg);
2029         if (ret) {
2030                 cm_enter_timewait(cm_id_priv);
2031                 goto out;
2032         }
2033
2034         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2035                        private_data, private_data_len);
2036         msg->timeout_ms = cm_id_priv->timeout_ms;
2037         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2038
2039         ret = ib_post_send_mad(msg, NULL);
2040         if (ret) {
2041                 cm_enter_timewait(cm_id_priv);
2042                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2043                 cm_free_msg(msg);
2044                 return ret;
2045         }
2046
2047         cm_id->state = IB_CM_DREQ_SENT;
2048         cm_id_priv->msg = msg;
2049 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2050         return ret;
2051 }
2052 EXPORT_SYMBOL(ib_send_cm_dreq);
2053
2054 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2055                           struct cm_id_private *cm_id_priv,
2056                           const void *private_data,
2057                           u8 private_data_len)
2058 {
2059         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2060         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2061         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2062
2063         if (private_data && private_data_len)
2064                 memcpy(drep_msg->private_data, private_data, private_data_len);
2065 }
2066
2067 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2068                     const void *private_data,
2069                     u8 private_data_len)
2070 {
2071         struct cm_id_private *cm_id_priv;
2072         struct ib_mad_send_buf *msg;
2073         unsigned long flags;
2074         void *data;
2075         int ret;
2076
2077         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2078                 return -EINVAL;
2079
2080         data = cm_copy_private_data(private_data, private_data_len);
2081         if (IS_ERR(data))
2082                 return PTR_ERR(data);
2083
2084         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2085         spin_lock_irqsave(&cm_id_priv->lock, flags);
2086         if (cm_id->state != IB_CM_DREQ_RCVD) {
2087                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2088                 kfree(data);
2089                 return -EINVAL;
2090         }
2091
2092         cm_set_private_data(cm_id_priv, data, private_data_len);
2093         cm_enter_timewait(cm_id_priv);
2094
2095         ret = cm_alloc_msg(cm_id_priv, &msg);
2096         if (ret)
2097                 goto out;
2098
2099         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2100                        private_data, private_data_len);
2101
2102         ret = ib_post_send_mad(msg, NULL);
2103         if (ret) {
2104                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2105                 cm_free_msg(msg);
2106                 return ret;
2107         }
2108
2109 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2110         return ret;
2111 }
2112 EXPORT_SYMBOL(ib_send_cm_drep);
2113
2114 static int cm_issue_drep(struct cm_port *port,
2115                          struct ib_mad_recv_wc *mad_recv_wc)
2116 {
2117         struct ib_mad_send_buf *msg = NULL;
2118         struct cm_dreq_msg *dreq_msg;
2119         struct cm_drep_msg *drep_msg;
2120         int ret;
2121
2122         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2123         if (ret)
2124                 return ret;
2125
2126         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2127         drep_msg = (struct cm_drep_msg *) msg->mad;
2128
2129         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2130         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2131         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2132
2133         ret = ib_post_send_mad(msg, NULL);
2134         if (ret)
2135                 cm_free_msg(msg);
2136
2137         return ret;
2138 }
2139
2140 static int cm_dreq_handler(struct cm_work *work)
2141 {
2142         struct cm_id_private *cm_id_priv;
2143         struct cm_dreq_msg *dreq_msg;
2144         struct ib_mad_send_buf *msg = NULL;
2145         int ret;
2146
2147         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2148         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2149                                    dreq_msg->local_comm_id);
2150         if (!cm_id_priv) {
2151                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2152                                 counter[CM_DREQ_COUNTER]);
2153                 cm_issue_drep(work->port, work->mad_recv_wc);
2154                 return -EINVAL;
2155         }
2156
2157         work->cm_event.private_data = &dreq_msg->private_data;
2158
2159         spin_lock_irq(&cm_id_priv->lock);
2160         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2161                 goto unlock;
2162
2163         switch (cm_id_priv->id.state) {
2164         case IB_CM_REP_SENT:
2165         case IB_CM_DREQ_SENT:
2166                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2167                 break;
2168         case IB_CM_ESTABLISHED:
2169                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2170                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2171                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2172                 break;
2173         case IB_CM_MRA_REP_RCVD:
2174                 break;
2175         case IB_CM_TIMEWAIT:
2176                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2177                                 counter[CM_DREQ_COUNTER]);
2178                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2179                         goto unlock;
2180
2181                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2182                                cm_id_priv->private_data,
2183                                cm_id_priv->private_data_len);
2184                 spin_unlock_irq(&cm_id_priv->lock);
2185
2186                 if (ib_post_send_mad(msg, NULL))
2187                         cm_free_msg(msg);
2188                 goto deref;
2189         case IB_CM_DREQ_RCVD:
2190                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2191                                 counter[CM_DREQ_COUNTER]);
2192                 goto unlock;
2193         default:
2194                 goto unlock;
2195         }
2196         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2197         cm_id_priv->tid = dreq_msg->hdr.tid;
2198         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2199         if (!ret)
2200                 list_add_tail(&work->list, &cm_id_priv->work_list);
2201         spin_unlock_irq(&cm_id_priv->lock);
2202
2203         if (ret)
2204                 cm_process_work(cm_id_priv, work);
2205         else
2206                 cm_deref_id(cm_id_priv);
2207         return 0;
2208
2209 unlock: spin_unlock_irq(&cm_id_priv->lock);
2210 deref:  cm_deref_id(cm_id_priv);
2211         return -EINVAL;
2212 }
2213
2214 static int cm_drep_handler(struct cm_work *work)
2215 {
2216         struct cm_id_private *cm_id_priv;
2217         struct cm_drep_msg *drep_msg;
2218         int ret;
2219
2220         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2221         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2222                                    drep_msg->local_comm_id);
2223         if (!cm_id_priv)
2224                 return -EINVAL;
2225
2226         work->cm_event.private_data = &drep_msg->private_data;
2227
2228         spin_lock_irq(&cm_id_priv->lock);
2229         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2230             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2231                 spin_unlock_irq(&cm_id_priv->lock);
2232                 goto out;
2233         }
2234         cm_enter_timewait(cm_id_priv);
2235
2236         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2237         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2238         if (!ret)
2239                 list_add_tail(&work->list, &cm_id_priv->work_list);
2240         spin_unlock_irq(&cm_id_priv->lock);
2241
2242         if (ret)
2243                 cm_process_work(cm_id_priv, work);
2244         else
2245                 cm_deref_id(cm_id_priv);
2246         return 0;
2247 out:
2248         cm_deref_id(cm_id_priv);
2249         return -EINVAL;
2250 }
2251
2252 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2253                    enum ib_cm_rej_reason reason,
2254                    void *ari,
2255                    u8 ari_length,
2256                    const void *private_data,
2257                    u8 private_data_len)
2258 {
2259         struct cm_id_private *cm_id_priv;
2260         struct ib_mad_send_buf *msg;
2261         unsigned long flags;
2262         int ret;
2263
2264         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2265             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2266                 return -EINVAL;
2267
2268         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2269
2270         spin_lock_irqsave(&cm_id_priv->lock, flags);
2271         switch (cm_id->state) {
2272         case IB_CM_REQ_SENT:
2273         case IB_CM_MRA_REQ_RCVD:
2274         case IB_CM_REQ_RCVD:
2275         case IB_CM_MRA_REQ_SENT:
2276         case IB_CM_REP_RCVD:
2277         case IB_CM_MRA_REP_SENT:
2278                 ret = cm_alloc_msg(cm_id_priv, &msg);
2279                 if (!ret)
2280                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2281                                       cm_id_priv, reason, ari, ari_length,
2282                                       private_data, private_data_len);
2283
2284                 cm_reset_to_idle(cm_id_priv);
2285                 break;
2286         case IB_CM_REP_SENT:
2287         case IB_CM_MRA_REP_RCVD:
2288                 ret = cm_alloc_msg(cm_id_priv, &msg);
2289                 if (!ret)
2290                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2291                                       cm_id_priv, reason, ari, ari_length,
2292                                       private_data, private_data_len);
2293
2294                 cm_enter_timewait(cm_id_priv);
2295                 break;
2296         default:
2297                 ret = -EINVAL;
2298                 goto out;
2299         }
2300
2301         if (ret)
2302                 goto out;
2303
2304         ret = ib_post_send_mad(msg, NULL);
2305         if (ret)
2306                 cm_free_msg(msg);
2307
2308 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2309         return ret;
2310 }
2311 EXPORT_SYMBOL(ib_send_cm_rej);
2312
2313 static void cm_format_rej_event(struct cm_work *work)
2314 {
2315         struct cm_rej_msg *rej_msg;
2316         struct ib_cm_rej_event_param *param;
2317
2318         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2319         param = &work->cm_event.param.rej_rcvd;
2320         param->ari = rej_msg->ari;
2321         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2322         param->reason = __be16_to_cpu(rej_msg->reason);
2323         work->cm_event.private_data = &rej_msg->private_data;
2324 }
2325
2326 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2327 {
2328         struct cm_timewait_info *timewait_info;
2329         struct cm_id_private *cm_id_priv;
2330         __be32 remote_id;
2331
2332         remote_id = rej_msg->local_comm_id;
2333
2334         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2335                 spin_lock_irq(&cm.lock);
2336                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2337                                                   remote_id);
2338                 if (!timewait_info) {
2339                         spin_unlock_irq(&cm.lock);
2340                         return NULL;
2341                 }
2342                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2343                                       (timewait_info->work.local_id ^
2344                                        cm.random_id_operand));
2345                 if (cm_id_priv) {
2346                         if (cm_id_priv->id.remote_id == remote_id)
2347                                 atomic_inc(&cm_id_priv->refcount);
2348                         else
2349                                 cm_id_priv = NULL;
2350                 }
2351                 spin_unlock_irq(&cm.lock);
2352         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2353                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2354         else
2355                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2356
2357         return cm_id_priv;
2358 }
2359
2360 static int cm_rej_handler(struct cm_work *work)
2361 {
2362         struct cm_id_private *cm_id_priv;
2363         struct cm_rej_msg *rej_msg;
2364         int ret;
2365
2366         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2367         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2368         if (!cm_id_priv)
2369                 return -EINVAL;
2370
2371         cm_format_rej_event(work);
2372
2373         spin_lock_irq(&cm_id_priv->lock);
2374         switch (cm_id_priv->id.state) {
2375         case IB_CM_REQ_SENT:
2376         case IB_CM_MRA_REQ_RCVD:
2377         case IB_CM_REP_SENT:
2378         case IB_CM_MRA_REP_RCVD:
2379                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2380                 /* fall through */
2381         case IB_CM_REQ_RCVD:
2382         case IB_CM_MRA_REQ_SENT:
2383                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2384                         cm_enter_timewait(cm_id_priv);
2385                 else
2386                         cm_reset_to_idle(cm_id_priv);
2387                 break;
2388         case IB_CM_DREQ_SENT:
2389                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2390                 /* fall through */
2391         case IB_CM_REP_RCVD:
2392         case IB_CM_MRA_REP_SENT:
2393                 cm_enter_timewait(cm_id_priv);
2394                 break;
2395         case IB_CM_ESTABLISHED:
2396                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2397                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2398                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2399                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2400                                               cm_id_priv->msg);
2401                         cm_enter_timewait(cm_id_priv);
2402                         break;
2403                 }
2404                 /* fall through */
2405         default:
2406                 spin_unlock_irq(&cm_id_priv->lock);
2407                 ret = -EINVAL;
2408                 goto out;
2409         }
2410
2411         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2412         if (!ret)
2413                 list_add_tail(&work->list, &cm_id_priv->work_list);
2414         spin_unlock_irq(&cm_id_priv->lock);
2415
2416         if (ret)
2417                 cm_process_work(cm_id_priv, work);
2418         else
2419                 cm_deref_id(cm_id_priv);
2420         return 0;
2421 out:
2422         cm_deref_id(cm_id_priv);
2423         return -EINVAL;
2424 }
2425
2426 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2427                    u8 service_timeout,
2428                    const void *private_data,
2429                    u8 private_data_len)
2430 {
2431         struct cm_id_private *cm_id_priv;
2432         struct ib_mad_send_buf *msg;
2433         enum ib_cm_state cm_state;
2434         enum ib_cm_lap_state lap_state;
2435         enum cm_msg_response msg_response;
2436         void *data;
2437         unsigned long flags;
2438         int ret;
2439
2440         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2441                 return -EINVAL;
2442
2443         data = cm_copy_private_data(private_data, private_data_len);
2444         if (IS_ERR(data))
2445                 return PTR_ERR(data);
2446
2447         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2448
2449         spin_lock_irqsave(&cm_id_priv->lock, flags);
2450         switch(cm_id_priv->id.state) {
2451         case IB_CM_REQ_RCVD:
2452                 cm_state = IB_CM_MRA_REQ_SENT;
2453                 lap_state = cm_id->lap_state;
2454                 msg_response = CM_MSG_RESPONSE_REQ;
2455                 break;
2456         case IB_CM_REP_RCVD:
2457                 cm_state = IB_CM_MRA_REP_SENT;
2458                 lap_state = cm_id->lap_state;
2459                 msg_response = CM_MSG_RESPONSE_REP;
2460                 break;
2461         case IB_CM_ESTABLISHED:
2462                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2463                         cm_state = cm_id->state;
2464                         lap_state = IB_CM_MRA_LAP_SENT;
2465                         msg_response = CM_MSG_RESPONSE_OTHER;
2466                         break;
2467                 }
2468         default:
2469                 ret = -EINVAL;
2470                 goto error1;
2471         }
2472
2473         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2474                 ret = cm_alloc_msg(cm_id_priv, &msg);
2475                 if (ret)
2476                         goto error1;
2477
2478                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2479                               msg_response, service_timeout,
2480                               private_data, private_data_len);
2481                 ret = ib_post_send_mad(msg, NULL);
2482                 if (ret)
2483                         goto error2;
2484         }
2485
2486         cm_id->state = cm_state;
2487         cm_id->lap_state = lap_state;
2488         cm_id_priv->service_timeout = service_timeout;
2489         cm_set_private_data(cm_id_priv, data, private_data_len);
2490         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2491         return 0;
2492
2493 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2494         kfree(data);
2495         return ret;
2496
2497 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2498         kfree(data);
2499         cm_free_msg(msg);
2500         return ret;
2501 }
2502 EXPORT_SYMBOL(ib_send_cm_mra);
2503
2504 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2505 {
2506         switch (cm_mra_get_msg_mraed(mra_msg)) {
2507         case CM_MSG_RESPONSE_REQ:
2508                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2509         case CM_MSG_RESPONSE_REP:
2510         case CM_MSG_RESPONSE_OTHER:
2511                 return cm_acquire_id(mra_msg->remote_comm_id,
2512                                      mra_msg->local_comm_id);
2513         default:
2514                 return NULL;
2515         }
2516 }
2517
2518 static int cm_mra_handler(struct cm_work *work)
2519 {
2520         struct cm_id_private *cm_id_priv;
2521         struct cm_mra_msg *mra_msg;
2522         int timeout, ret;
2523
2524         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2525         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2526         if (!cm_id_priv)
2527                 return -EINVAL;
2528
2529         work->cm_event.private_data = &mra_msg->private_data;
2530         work->cm_event.param.mra_rcvd.service_timeout =
2531                                         cm_mra_get_service_timeout(mra_msg);
2532         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2533                   cm_convert_to_ms(cm_id_priv->av.timeout);
2534
2535         spin_lock_irq(&cm_id_priv->lock);
2536         switch (cm_id_priv->id.state) {
2537         case IB_CM_REQ_SENT:
2538                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2539                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2540                                   cm_id_priv->msg, timeout))
2541                         goto out;
2542                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2543                 break;
2544         case IB_CM_REP_SENT:
2545                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2546                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2547                                   cm_id_priv->msg, timeout))
2548                         goto out;
2549                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2550                 break;
2551         case IB_CM_ESTABLISHED:
2552                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2553                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2554                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2555                                   cm_id_priv->msg, timeout)) {
2556                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2557                                 atomic_long_inc(&work->port->
2558                                                 counter_group[CM_RECV_DUPLICATES].
2559                                                 counter[CM_MRA_COUNTER]);
2560                         goto out;
2561                 }
2562                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2563                 break;
2564         case IB_CM_MRA_REQ_RCVD:
2565         case IB_CM_MRA_REP_RCVD:
2566                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2567                                 counter[CM_MRA_COUNTER]);
2568                 /* fall through */
2569         default:
2570                 goto out;
2571         }
2572
2573         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2574                                       cm_id_priv->id.state;
2575         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2576         if (!ret)
2577                 list_add_tail(&work->list, &cm_id_priv->work_list);
2578         spin_unlock_irq(&cm_id_priv->lock);
2579
2580         if (ret)
2581                 cm_process_work(cm_id_priv, work);
2582         else
2583                 cm_deref_id(cm_id_priv);
2584         return 0;
2585 out:
2586         spin_unlock_irq(&cm_id_priv->lock);
2587         cm_deref_id(cm_id_priv);
2588         return -EINVAL;
2589 }
2590
2591 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2592                           struct cm_id_private *cm_id_priv,
2593                           struct ib_sa_path_rec *alternate_path,
2594                           const void *private_data,
2595                           u8 private_data_len)
2596 {
2597         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2598                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2599         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2600         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2601         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2602         /* todo: need remote CM response timeout */
2603         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2604         lap_msg->alt_local_lid = alternate_path->slid;
2605         lap_msg->alt_remote_lid = alternate_path->dlid;
2606         lap_msg->alt_local_gid = alternate_path->sgid;
2607         lap_msg->alt_remote_gid = alternate_path->dgid;
2608         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2609         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2610         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2611         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2612         cm_lap_set_sl(lap_msg, alternate_path->sl);
2613         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2614         cm_lap_set_local_ack_timeout(lap_msg,
2615                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2616                                alternate_path->packet_life_time));
2617
2618         if (private_data && private_data_len)
2619                 memcpy(lap_msg->private_data, private_data, private_data_len);
2620 }
2621
2622 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2623                    struct ib_sa_path_rec *alternate_path,
2624                    const void *private_data,
2625                    u8 private_data_len)
2626 {
2627         struct cm_id_private *cm_id_priv;
2628         struct ib_mad_send_buf *msg;
2629         unsigned long flags;
2630         int ret;
2631
2632         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2633                 return -EINVAL;
2634
2635         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2636         spin_lock_irqsave(&cm_id_priv->lock, flags);
2637         if (cm_id->state != IB_CM_ESTABLISHED ||
2638             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2639              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2640                 ret = -EINVAL;
2641                 goto out;
2642         }
2643
2644         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
2645         if (ret)
2646                 goto out;
2647         cm_id_priv->alt_av.timeout =
2648                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2649                                        cm_id_priv->alt_av.timeout - 1);
2650
2651         ret = cm_alloc_msg(cm_id_priv, &msg);
2652         if (ret)
2653                 goto out;
2654
2655         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2656                       alternate_path, private_data, private_data_len);
2657         msg->timeout_ms = cm_id_priv->timeout_ms;
2658         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2659
2660         ret = ib_post_send_mad(msg, NULL);
2661         if (ret) {
2662                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2663                 cm_free_msg(msg);
2664                 return ret;
2665         }
2666
2667         cm_id->lap_state = IB_CM_LAP_SENT;
2668         cm_id_priv->msg = msg;
2669
2670 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2671         return ret;
2672 }
2673 EXPORT_SYMBOL(ib_send_cm_lap);
2674
2675 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2676                                     struct ib_sa_path_rec *path,
2677                                     struct cm_lap_msg *lap_msg)
2678 {
2679         memset(path, 0, sizeof *path);
2680         path->dgid = lap_msg->alt_local_gid;
2681         path->sgid = lap_msg->alt_remote_gid;
2682         path->dlid = lap_msg->alt_local_lid;
2683         path->slid = lap_msg->alt_remote_lid;
2684         path->flow_label = cm_lap_get_flow_label(lap_msg);
2685         path->hop_limit = lap_msg->alt_hop_limit;
2686         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2687         path->reversible = 1;
2688         path->pkey = cm_id_priv->pkey;
2689         path->sl = cm_lap_get_sl(lap_msg);
2690         path->mtu_selector = IB_SA_EQ;
2691         path->mtu = cm_id_priv->path_mtu;
2692         path->rate_selector = IB_SA_EQ;
2693         path->rate = cm_lap_get_packet_rate(lap_msg);
2694         path->packet_life_time_selector = IB_SA_EQ;
2695         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2696         path->packet_life_time -= (path->packet_life_time > 0);
2697 }
2698
2699 static int cm_lap_handler(struct cm_work *work)
2700 {
2701         struct cm_id_private *cm_id_priv;
2702         struct cm_lap_msg *lap_msg;
2703         struct ib_cm_lap_event_param *param;
2704         struct ib_mad_send_buf *msg = NULL;
2705         int ret;
2706
2707         /* todo: verify LAP request and send reject APR if invalid. */
2708         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2709         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2710                                    lap_msg->local_comm_id);
2711         if (!cm_id_priv)
2712                 return -EINVAL;
2713
2714         param = &work->cm_event.param.lap_rcvd;
2715         param->alternate_path = &work->path[0];
2716         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2717         work->cm_event.private_data = &lap_msg->private_data;
2718
2719         spin_lock_irq(&cm_id_priv->lock);
2720         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2721                 goto unlock;
2722
2723         switch (cm_id_priv->id.lap_state) {
2724         case IB_CM_LAP_UNINIT:
2725         case IB_CM_LAP_IDLE:
2726                 break;
2727         case IB_CM_MRA_LAP_SENT:
2728                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2729                                 counter[CM_LAP_COUNTER]);
2730                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2731                         goto unlock;
2732
2733                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2734                               CM_MSG_RESPONSE_OTHER,
2735                               cm_id_priv->service_timeout,
2736                               cm_id_priv->private_data,
2737                               cm_id_priv->private_data_len);
2738                 spin_unlock_irq(&cm_id_priv->lock);
2739
2740                 if (ib_post_send_mad(msg, NULL))
2741                         cm_free_msg(msg);
2742                 goto deref;
2743         case IB_CM_LAP_RCVD:
2744                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2745                                 counter[CM_LAP_COUNTER]);
2746                 goto unlock;
2747         default:
2748                 goto unlock;
2749         }
2750
2751         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2752         cm_id_priv->tid = lap_msg->hdr.tid;
2753         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2754                                 work->mad_recv_wc->recv_buf.grh,
2755                                 &cm_id_priv->av);
2756         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av);
2757         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2758         if (!ret)
2759                 list_add_tail(&work->list, &cm_id_priv->work_list);
2760         spin_unlock_irq(&cm_id_priv->lock);
2761
2762         if (ret)
2763                 cm_process_work(cm_id_priv, work);
2764         else
2765                 cm_deref_id(cm_id_priv);
2766         return 0;
2767
2768 unlock: spin_unlock_irq(&cm_id_priv->lock);
2769 deref:  cm_deref_id(cm_id_priv);
2770         return -EINVAL;
2771 }
2772
2773 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2774                           struct cm_id_private *cm_id_priv,
2775                           enum ib_cm_apr_status status,
2776                           void *info,
2777                           u8 info_length,
2778                           const void *private_data,
2779                           u8 private_data_len)
2780 {
2781         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2782         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2783         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2784         apr_msg->ap_status = (u8) status;
2785
2786         if (info && info_length) {
2787                 apr_msg->info_length = info_length;
2788                 memcpy(apr_msg->info, info, info_length);
2789         }
2790
2791         if (private_data && private_data_len)
2792                 memcpy(apr_msg->private_data, private_data, private_data_len);
2793 }
2794
2795 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2796                    enum ib_cm_apr_status status,
2797                    void *info,
2798                    u8 info_length,
2799                    const void *private_data,
2800                    u8 private_data_len)
2801 {
2802         struct cm_id_private *cm_id_priv;
2803         struct ib_mad_send_buf *msg;
2804         unsigned long flags;
2805         int ret;
2806
2807         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2808             (info && info_length > IB_CM_APR_INFO_LENGTH))
2809                 return -EINVAL;
2810
2811         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2812         spin_lock_irqsave(&cm_id_priv->lock, flags);
2813         if (cm_id->state != IB_CM_ESTABLISHED ||
2814             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2815              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2816                 ret = -EINVAL;
2817                 goto out;
2818         }
2819
2820         ret = cm_alloc_msg(cm_id_priv, &msg);
2821         if (ret)
2822                 goto out;
2823
2824         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2825                       info, info_length, private_data, private_data_len);
2826         ret = ib_post_send_mad(msg, NULL);
2827         if (ret) {
2828                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2829                 cm_free_msg(msg);
2830                 return ret;
2831         }
2832
2833         cm_id->lap_state = IB_CM_LAP_IDLE;
2834 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2835         return ret;
2836 }
2837 EXPORT_SYMBOL(ib_send_cm_apr);
2838
2839 static int cm_apr_handler(struct cm_work *work)
2840 {
2841         struct cm_id_private *cm_id_priv;
2842         struct cm_apr_msg *apr_msg;
2843         int ret;
2844
2845         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2846         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2847                                    apr_msg->local_comm_id);
2848         if (!cm_id_priv)
2849                 return -EINVAL; /* Unmatched reply. */
2850
2851         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2852         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2853         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2854         work->cm_event.private_data = &apr_msg->private_data;
2855
2856         spin_lock_irq(&cm_id_priv->lock);
2857         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2858             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2859              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2860                 spin_unlock_irq(&cm_id_priv->lock);
2861                 goto out;
2862         }
2863         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2864         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2865         cm_id_priv->msg = NULL;
2866
2867         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2868         if (!ret)
2869                 list_add_tail(&work->list, &cm_id_priv->work_list);
2870         spin_unlock_irq(&cm_id_priv->lock);
2871
2872         if (ret)
2873                 cm_process_work(cm_id_priv, work);
2874         else
2875                 cm_deref_id(cm_id_priv);
2876         return 0;
2877 out:
2878         cm_deref_id(cm_id_priv);
2879         return -EINVAL;
2880 }
2881
2882 static int cm_timewait_handler(struct cm_work *work)
2883 {
2884         struct cm_timewait_info *timewait_info;
2885         struct cm_id_private *cm_id_priv;
2886         int ret;
2887
2888         timewait_info = (struct cm_timewait_info *)work;
2889         spin_lock_irq(&cm.lock);
2890         list_del(&timewait_info->list);
2891         spin_unlock_irq(&cm.lock);
2892
2893         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2894                                    timewait_info->work.remote_id);
2895         if (!cm_id_priv)
2896                 return -EINVAL;
2897
2898         spin_lock_irq(&cm_id_priv->lock);
2899         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2900             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2901                 spin_unlock_irq(&cm_id_priv->lock);
2902                 goto out;
2903         }
2904         cm_id_priv->id.state = IB_CM_IDLE;
2905         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2906         if (!ret)
2907                 list_add_tail(&work->list, &cm_id_priv->work_list);
2908         spin_unlock_irq(&cm_id_priv->lock);
2909
2910         if (ret)
2911                 cm_process_work(cm_id_priv, work);
2912         else
2913                 cm_deref_id(cm_id_priv);
2914         return 0;
2915 out:
2916         cm_deref_id(cm_id_priv);
2917         return -EINVAL;
2918 }
2919
2920 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
2921                                struct cm_id_private *cm_id_priv,
2922                                struct ib_cm_sidr_req_param *param)
2923 {
2924         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
2925                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
2926         sidr_req_msg->request_id = cm_id_priv->id.local_id;
2927         sidr_req_msg->pkey = param->path->pkey;
2928         sidr_req_msg->service_id = param->service_id;
2929
2930         if (param->private_data && param->private_data_len)
2931                 memcpy(sidr_req_msg->private_data, param->private_data,
2932                        param->private_data_len);
2933 }
2934
2935 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
2936                         struct ib_cm_sidr_req_param *param)
2937 {
2938         struct cm_id_private *cm_id_priv;
2939         struct ib_mad_send_buf *msg;
2940         unsigned long flags;
2941         int ret;
2942
2943         if (!param->path || (param->private_data &&
2944              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
2945                 return -EINVAL;
2946
2947         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2948         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
2949         if (ret)
2950                 goto out;
2951
2952         cm_id->service_id = param->service_id;
2953         cm_id->service_mask = ~cpu_to_be64(0);
2954         cm_id_priv->timeout_ms = param->timeout_ms;
2955         cm_id_priv->max_cm_retries = param->max_cm_retries;
2956         ret = cm_alloc_msg(cm_id_priv, &msg);
2957         if (ret)
2958                 goto out;
2959
2960         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
2961                            param);
2962         msg->timeout_ms = cm_id_priv->timeout_ms;
2963         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
2964
2965         spin_lock_irqsave(&cm_id_priv->lock, flags);
2966         if (cm_id->state == IB_CM_IDLE)
2967                 ret = ib_post_send_mad(msg, NULL);
2968         else
2969                 ret = -EINVAL;
2970
2971         if (ret) {
2972                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2973                 cm_free_msg(msg);
2974                 goto out;
2975         }
2976         cm_id->state = IB_CM_SIDR_REQ_SENT;
2977         cm_id_priv->msg = msg;
2978         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2979 out:
2980         return ret;
2981 }
2982 EXPORT_SYMBOL(ib_send_cm_sidr_req);
2983
2984 static void cm_format_sidr_req_event(struct cm_work *work,
2985                                      struct ib_cm_id *listen_id)
2986 {
2987         struct cm_sidr_req_msg *sidr_req_msg;
2988         struct ib_cm_sidr_req_event_param *param;
2989
2990         sidr_req_msg = (struct cm_sidr_req_msg *)
2991                                 work->mad_recv_wc->recv_buf.mad;
2992         param = &work->cm_event.param.sidr_req_rcvd;
2993         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
2994         param->listen_id = listen_id;
2995         param->port = work->port->port_num;
2996         work->cm_event.private_data = &sidr_req_msg->private_data;
2997 }
2998
2999 static int cm_sidr_req_handler(struct cm_work *work)
3000 {
3001         struct ib_cm_id *cm_id;
3002         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3003         struct cm_sidr_req_msg *sidr_req_msg;
3004         struct ib_wc *wc;
3005
3006         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3007         if (IS_ERR(cm_id))
3008                 return PTR_ERR(cm_id);
3009         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3010
3011         /* Record SGID/SLID and request ID for lookup. */
3012         sidr_req_msg = (struct cm_sidr_req_msg *)
3013                                 work->mad_recv_wc->recv_buf.mad;
3014         wc = work->mad_recv_wc->wc;
3015         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3016         cm_id_priv->av.dgid.global.interface_id = 0;
3017         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3018                                 work->mad_recv_wc->recv_buf.grh,
3019                                 &cm_id_priv->av);
3020         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3021         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3022         atomic_inc(&cm_id_priv->work_count);
3023
3024         spin_lock_irq(&cm.lock);
3025         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3026         if (cur_cm_id_priv) {
3027                 spin_unlock_irq(&cm.lock);
3028                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3029                                 counter[CM_SIDR_REQ_COUNTER]);
3030                 goto out; /* Duplicate message. */
3031         }
3032         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3033         cur_cm_id_priv = cm_find_listen(cm_id->device,
3034                                         sidr_req_msg->service_id,
3035                                         sidr_req_msg->private_data);
3036         if (!cur_cm_id_priv) {
3037                 spin_unlock_irq(&cm.lock);
3038                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3039                 goto out; /* No match. */
3040         }
3041         atomic_inc(&cur_cm_id_priv->refcount);
3042         atomic_inc(&cm_id_priv->refcount);
3043         spin_unlock_irq(&cm.lock);
3044
3045         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3046         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3047         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3048         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3049
3050         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3051         cm_process_work(cm_id_priv, work);
3052         cm_deref_id(cur_cm_id_priv);
3053         return 0;
3054 out:
3055         ib_destroy_cm_id(&cm_id_priv->id);
3056         return -EINVAL;
3057 }
3058
3059 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3060                                struct cm_id_private *cm_id_priv,
3061                                struct ib_cm_sidr_rep_param *param)
3062 {
3063         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3064                           cm_id_priv->tid);
3065         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3066         sidr_rep_msg->status = param->status;
3067         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3068         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3069         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3070
3071         if (param->info && param->info_length)
3072                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3073
3074         if (param->private_data && param->private_data_len)
3075                 memcpy(sidr_rep_msg->private_data, param->private_data,
3076                        param->private_data_len);
3077 }
3078
3079 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3080                         struct ib_cm_sidr_rep_param *param)
3081 {
3082         struct cm_id_private *cm_id_priv;
3083         struct ib_mad_send_buf *msg;
3084         unsigned long flags;
3085         int ret;
3086
3087         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3088             (param->private_data &&
3089              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3090                 return -EINVAL;
3091
3092         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3093         spin_lock_irqsave(&cm_id_priv->lock, flags);
3094         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3095                 ret = -EINVAL;
3096                 goto error;
3097         }
3098
3099         ret = cm_alloc_msg(cm_id_priv, &msg);
3100         if (ret)
3101                 goto error;
3102
3103         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3104                            param);
3105         ret = ib_post_send_mad(msg, NULL);
3106         if (ret) {
3107                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3108                 cm_free_msg(msg);
3109                 return ret;
3110         }
3111         cm_id->state = IB_CM_IDLE;
3112         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3113
3114         spin_lock_irqsave(&cm.lock, flags);
3115         rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3116         spin_unlock_irqrestore(&cm.lock, flags);
3117         return 0;
3118
3119 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3120         return ret;
3121 }
3122 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3123
3124 static void cm_format_sidr_rep_event(struct cm_work *work)
3125 {
3126         struct cm_sidr_rep_msg *sidr_rep_msg;
3127         struct ib_cm_sidr_rep_event_param *param;
3128
3129         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3130                                 work->mad_recv_wc->recv_buf.mad;
3131         param = &work->cm_event.param.sidr_rep_rcvd;
3132         param->status = sidr_rep_msg->status;
3133         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3134         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3135         param->info = &sidr_rep_msg->info;
3136         param->info_len = sidr_rep_msg->info_length;
3137         work->cm_event.private_data = &sidr_rep_msg->private_data;
3138 }
3139
3140 static int cm_sidr_rep_handler(struct cm_work *work)
3141 {
3142         struct cm_sidr_rep_msg *sidr_rep_msg;
3143         struct cm_id_private *cm_id_priv;
3144
3145         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3146                                 work->mad_recv_wc->recv_buf.mad;
3147         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3148         if (!cm_id_priv)
3149                 return -EINVAL; /* Unmatched reply. */
3150
3151         spin_lock_irq(&cm_id_priv->lock);
3152         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3153                 spin_unlock_irq(&cm_id_priv->lock);
3154                 goto out;
3155         }
3156         cm_id_priv->id.state = IB_CM_IDLE;
3157         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3158         spin_unlock_irq(&cm_id_priv->lock);
3159
3160         cm_format_sidr_rep_event(work);
3161         cm_process_work(cm_id_priv, work);
3162         return 0;
3163 out:
3164         cm_deref_id(cm_id_priv);
3165         return -EINVAL;
3166 }
3167
3168 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3169                                   enum ib_wc_status wc_status)
3170 {
3171         struct cm_id_private *cm_id_priv;
3172         struct ib_cm_event cm_event;
3173         enum ib_cm_state state;
3174         int ret;
3175
3176         memset(&cm_event, 0, sizeof cm_event);
3177         cm_id_priv = msg->context[0];
3178
3179         /* Discard old sends or ones without a response. */
3180         spin_lock_irq(&cm_id_priv->lock);
3181         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3182         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3183                 goto discard;
3184
3185         switch (state) {
3186         case IB_CM_REQ_SENT:
3187         case IB_CM_MRA_REQ_RCVD:
3188                 cm_reset_to_idle(cm_id_priv);
3189                 cm_event.event = IB_CM_REQ_ERROR;
3190                 break;
3191         case IB_CM_REP_SENT:
3192         case IB_CM_MRA_REP_RCVD:
3193                 cm_reset_to_idle(cm_id_priv);
3194                 cm_event.event = IB_CM_REP_ERROR;
3195                 break;
3196         case IB_CM_DREQ_SENT:
3197                 cm_enter_timewait(cm_id_priv);
3198                 cm_event.event = IB_CM_DREQ_ERROR;
3199                 break;
3200         case IB_CM_SIDR_REQ_SENT:
3201                 cm_id_priv->id.state = IB_CM_IDLE;
3202                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3203                 break;
3204         default:
3205                 goto discard;
3206         }
3207         spin_unlock_irq(&cm_id_priv->lock);
3208         cm_event.param.send_status = wc_status;
3209
3210         /* No other events can occur on the cm_id at this point. */
3211         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3212         cm_free_msg(msg);
3213         if (ret)
3214                 ib_destroy_cm_id(&cm_id_priv->id);
3215         return;
3216 discard:
3217         spin_unlock_irq(&cm_id_priv->lock);
3218         cm_free_msg(msg);
3219 }
3220
3221 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3222                             struct ib_mad_send_wc *mad_send_wc)
3223 {
3224         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3225         struct cm_port *port;
3226         u16 attr_index;
3227
3228         port = mad_agent->context;
3229         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3230                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3231
3232         /*
3233          * If the send was in response to a received message (context[0] is not
3234          * set to a cm_id), and is not a REJ, then it is a send that was
3235          * manually retried.
3236          */
3237         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3238                 msg->retries = 1;
3239
3240         atomic_long_add(1 + msg->retries,
3241                         &port->counter_group[CM_XMIT].counter[attr_index]);
3242         if (msg->retries)
3243                 atomic_long_add(msg->retries,
3244                                 &port->counter_group[CM_XMIT_RETRIES].
3245                                 counter[attr_index]);
3246
3247         switch (mad_send_wc->status) {
3248         case IB_WC_SUCCESS:
3249         case IB_WC_WR_FLUSH_ERR:
3250                 cm_free_msg(msg);
3251                 break;
3252         default:
3253                 if (msg->context[0] && msg->context[1])
3254                         cm_process_send_error(msg, mad_send_wc->status);
3255                 else
3256                         cm_free_msg(msg);
3257                 break;
3258         }
3259 }
3260
3261 static void cm_work_handler(struct work_struct *_work)
3262 {
3263         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3264         int ret;
3265
3266         switch (work->cm_event.event) {
3267         case IB_CM_REQ_RECEIVED:
3268                 ret = cm_req_handler(work);
3269                 break;
3270         case IB_CM_MRA_RECEIVED:
3271                 ret = cm_mra_handler(work);
3272                 break;
3273         case IB_CM_REJ_RECEIVED:
3274                 ret = cm_rej_handler(work);
3275                 break;
3276         case IB_CM_REP_RECEIVED:
3277                 ret = cm_rep_handler(work);
3278                 break;
3279         case IB_CM_RTU_RECEIVED:
3280                 ret = cm_rtu_handler(work);
3281                 break;
3282         case IB_CM_USER_ESTABLISHED:
3283                 ret = cm_establish_handler(work);
3284                 break;
3285         case IB_CM_DREQ_RECEIVED:
3286                 ret = cm_dreq_handler(work);
3287                 break;
3288         case IB_CM_DREP_RECEIVED:
3289                 ret = cm_drep_handler(work);
3290                 break;
3291         case IB_CM_SIDR_REQ_RECEIVED:
3292                 ret = cm_sidr_req_handler(work);
3293                 break;
3294         case IB_CM_SIDR_REP_RECEIVED:
3295                 ret = cm_sidr_rep_handler(work);
3296                 break;
3297         case IB_CM_LAP_RECEIVED:
3298                 ret = cm_lap_handler(work);
3299                 break;
3300         case IB_CM_APR_RECEIVED:
3301                 ret = cm_apr_handler(work);
3302                 break;
3303         case IB_CM_TIMEWAIT_EXIT:
3304                 ret = cm_timewait_handler(work);
3305                 break;
3306         default:
3307                 ret = -EINVAL;
3308                 break;
3309         }
3310         if (ret)
3311                 cm_free_work(work);
3312 }
3313
3314 static int cm_establish(struct ib_cm_id *cm_id)
3315 {
3316         struct cm_id_private *cm_id_priv;
3317         struct cm_work *work;
3318         unsigned long flags;
3319         int ret = 0;
3320         struct cm_device *cm_dev;
3321
3322         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3323         if (!cm_dev)
3324                 return -ENODEV;
3325
3326         work = kmalloc(sizeof *work, GFP_ATOMIC);
3327         if (!work)
3328                 return -ENOMEM;
3329
3330         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3331         spin_lock_irqsave(&cm_id_priv->lock, flags);
3332         switch (cm_id->state)
3333         {
3334         case IB_CM_REP_SENT:
3335         case IB_CM_MRA_REP_RCVD:
3336                 cm_id->state = IB_CM_ESTABLISHED;
3337                 break;
3338         case IB_CM_ESTABLISHED:
3339                 ret = -EISCONN;
3340                 break;
3341         default:
3342                 ret = -EINVAL;
3343                 break;
3344         }
3345         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3346
3347         if (ret) {
3348                 kfree(work);
3349                 goto out;
3350         }
3351
3352         /*
3353          * The CM worker thread may try to destroy the cm_id before it
3354          * can execute this work item.  To prevent potential deadlock,
3355          * we need to find the cm_id once we're in the context of the
3356          * worker thread, rather than holding a reference on it.
3357          */
3358         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3359         work->local_id = cm_id->local_id;
3360         work->remote_id = cm_id->remote_id;
3361         work->mad_recv_wc = NULL;
3362         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3363
3364         /* Check if the device started its remove_one */
3365         spin_lock_irq(&cm.lock);
3366         if (!cm_dev->going_down) {
3367                 queue_delayed_work(cm.wq, &work->work, 0);
3368         } else {
3369                 kfree(work);
3370                 ret = -ENODEV;
3371         }
3372         spin_unlock_irq(&cm.lock);
3373
3374 out:
3375         return ret;
3376 }
3377
3378 static int cm_migrate(struct ib_cm_id *cm_id)
3379 {
3380         struct cm_id_private *cm_id_priv;
3381         unsigned long flags;
3382         int ret = 0;
3383
3384         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3385         spin_lock_irqsave(&cm_id_priv->lock, flags);
3386         if (cm_id->state == IB_CM_ESTABLISHED &&
3387             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3388              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3389                 cm_id->lap_state = IB_CM_LAP_IDLE;
3390                 cm_id_priv->av = cm_id_priv->alt_av;
3391         } else
3392                 ret = -EINVAL;
3393         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3394
3395         return ret;
3396 }
3397
3398 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3399 {
3400         int ret;
3401
3402         switch (event) {
3403         case IB_EVENT_COMM_EST:
3404                 ret = cm_establish(cm_id);
3405                 break;
3406         case IB_EVENT_PATH_MIG:
3407                 ret = cm_migrate(cm_id);
3408                 break;
3409         default:
3410                 ret = -EINVAL;
3411         }
3412         return ret;
3413 }
3414 EXPORT_SYMBOL(ib_cm_notify);
3415
3416 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3417                             struct ib_mad_recv_wc *mad_recv_wc)
3418 {
3419         struct cm_port *port = mad_agent->context;
3420         struct cm_work *work;
3421         enum ib_cm_event_type event;
3422         u16 attr_id;
3423         int paths = 0;
3424         int going_down = 0;
3425
3426         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3427         case CM_REQ_ATTR_ID:
3428                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3429                                                     alt_local_lid != 0);
3430                 event = IB_CM_REQ_RECEIVED;
3431                 break;
3432         case CM_MRA_ATTR_ID:
3433                 event = IB_CM_MRA_RECEIVED;
3434                 break;
3435         case CM_REJ_ATTR_ID:
3436                 event = IB_CM_REJ_RECEIVED;
3437                 break;
3438         case CM_REP_ATTR_ID:
3439                 event = IB_CM_REP_RECEIVED;
3440                 break;
3441         case CM_RTU_ATTR_ID:
3442                 event = IB_CM_RTU_RECEIVED;
3443                 break;
3444         case CM_DREQ_ATTR_ID:
3445                 event = IB_CM_DREQ_RECEIVED;
3446                 break;
3447         case CM_DREP_ATTR_ID:
3448                 event = IB_CM_DREP_RECEIVED;
3449                 break;
3450         case CM_SIDR_REQ_ATTR_ID:
3451                 event = IB_CM_SIDR_REQ_RECEIVED;
3452                 break;
3453         case CM_SIDR_REP_ATTR_ID:
3454                 event = IB_CM_SIDR_REP_RECEIVED;
3455                 break;
3456         case CM_LAP_ATTR_ID:
3457                 paths = 1;
3458                 event = IB_CM_LAP_RECEIVED;
3459                 break;
3460         case CM_APR_ATTR_ID:
3461                 event = IB_CM_APR_RECEIVED;
3462                 break;
3463         default:
3464                 ib_free_recv_mad(mad_recv_wc);
3465                 return;
3466         }
3467
3468         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3469         atomic_long_inc(&port->counter_group[CM_RECV].
3470                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3471
3472         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3473                        GFP_KERNEL);
3474         if (!work) {
3475                 ib_free_recv_mad(mad_recv_wc);
3476                 return;
3477         }
3478
3479         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3480         work->cm_event.event = event;
3481         work->mad_recv_wc = mad_recv_wc;
3482         work->port = port;
3483
3484         /* Check if the device started its remove_one */
3485         spin_lock_irq(&cm.lock);
3486         if (!port->cm_dev->going_down)
3487                 queue_delayed_work(cm.wq, &work->work, 0);
3488         else
3489                 going_down = 1;
3490         spin_unlock_irq(&cm.lock);
3491
3492         if (going_down) {
3493                 kfree(work);
3494                 ib_free_recv_mad(mad_recv_wc);
3495         }
3496 }
3497
3498 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3499                                 struct ib_qp_attr *qp_attr,
3500                                 int *qp_attr_mask)
3501 {
3502         unsigned long flags;
3503         int ret;
3504
3505         spin_lock_irqsave(&cm_id_priv->lock, flags);
3506         switch (cm_id_priv->id.state) {
3507         case IB_CM_REQ_SENT:
3508         case IB_CM_MRA_REQ_RCVD:
3509         case IB_CM_REQ_RCVD:
3510         case IB_CM_MRA_REQ_SENT:
3511         case IB_CM_REP_RCVD:
3512         case IB_CM_MRA_REP_SENT:
3513         case IB_CM_REP_SENT:
3514         case IB_CM_MRA_REP_RCVD:
3515         case IB_CM_ESTABLISHED:
3516                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3517                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3518                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3519                 if (cm_id_priv->responder_resources)
3520                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3521                                                     IB_ACCESS_REMOTE_ATOMIC;
3522                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3523                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3524                 ret = 0;
3525                 break;
3526         default:
3527                 ret = -EINVAL;
3528                 break;
3529         }
3530         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3531         return ret;
3532 }
3533
3534 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3535                                struct ib_qp_attr *qp_attr,
3536                                int *qp_attr_mask)
3537 {
3538         unsigned long flags;
3539         int ret;
3540
3541         spin_lock_irqsave(&cm_id_priv->lock, flags);
3542         switch (cm_id_priv->id.state) {
3543         case IB_CM_REQ_RCVD:
3544         case IB_CM_MRA_REQ_SENT:
3545         case IB_CM_REP_RCVD:
3546         case IB_CM_MRA_REP_SENT:
3547         case IB_CM_REP_SENT:
3548         case IB_CM_MRA_REP_RCVD:
3549         case IB_CM_ESTABLISHED:
3550                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3551                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3552                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3553                 if (!cm_id_priv->av.valid) {
3554                         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3555                         return -EINVAL;
3556                 }
3557                 if (cm_id_priv->av.ah_attr.vlan_id != 0xffff) {
3558                         qp_attr->vlan_id = cm_id_priv->av.ah_attr.vlan_id;
3559                         *qp_attr_mask |= IB_QP_VID;
3560                 }
3561                 if (!is_zero_ether_addr(cm_id_priv->av.smac)) {
3562                         memcpy(qp_attr->smac, cm_id_priv->av.smac,
3563                                sizeof(qp_attr->smac));
3564                         *qp_attr_mask |= IB_QP_SMAC;
3565                 }
3566                 if (cm_id_priv->alt_av.valid) {
3567                         if (cm_id_priv->alt_av.ah_attr.vlan_id != 0xffff) {
3568                                 qp_attr->alt_vlan_id =
3569                                         cm_id_priv->alt_av.ah_attr.vlan_id;
3570                                 *qp_attr_mask |= IB_QP_ALT_VID;
3571                         }
3572                         if (!is_zero_ether_addr(cm_id_priv->alt_av.smac)) {
3573                                 memcpy(qp_attr->alt_smac,
3574                                        cm_id_priv->alt_av.smac,
3575                                        sizeof(qp_attr->alt_smac));
3576                                 *qp_attr_mask |= IB_QP_ALT_SMAC;
3577                         }
3578                 }
3579                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3580                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3581                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3582                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3583                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3584                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3585                                          IB_QP_MIN_RNR_TIMER;
3586                         qp_attr->max_dest_rd_atomic =
3587                                         cm_id_priv->responder_resources;
3588                         qp_attr->min_rnr_timer = 0;
3589                 }
3590                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3591                         *qp_attr_mask |= IB_QP_ALT_PATH;
3592                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3593                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3594                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3595                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3596                 }
3597                 ret = 0;
3598                 break;
3599         default:
3600                 ret = -EINVAL;
3601                 break;
3602         }
3603         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3604         return ret;
3605 }
3606
3607 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3608                                struct ib_qp_attr *qp_attr,
3609                                int *qp_attr_mask)
3610 {
3611         unsigned long flags;
3612         int ret;
3613
3614         spin_lock_irqsave(&cm_id_priv->lock, flags);
3615         switch (cm_id_priv->id.state) {
3616         /* Allow transition to RTS before sending REP */
3617         case IB_CM_REQ_RCVD:
3618         case IB_CM_MRA_REQ_SENT:
3619
3620         case IB_CM_REP_RCVD:
3621         case IB_CM_MRA_REP_SENT:
3622         case IB_CM_REP_SENT:
3623         case IB_CM_MRA_REP_RCVD:
3624         case IB_CM_ESTABLISHED:
3625                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3626                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3627                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3628                         switch (cm_id_priv->qp_type) {
3629                         case IB_QPT_RC:
3630                         case IB_QPT_XRC_INI:
3631                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3632                                                  IB_QP_MAX_QP_RD_ATOMIC;
3633                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3634                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3635                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3636                                 /* fall through */
3637                         case IB_QPT_XRC_TGT:
3638                                 *qp_attr_mask |= IB_QP_TIMEOUT;
3639                                 qp_attr->timeout = cm_id_priv->av.timeout;
3640                                 break;
3641                         default:
3642                                 break;
3643                         }
3644                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3645                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3646                                 qp_attr->path_mig_state = IB_MIG_REARM;
3647                         }
3648                 } else {
3649                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3650                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3651                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3652                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3653                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3654                         qp_attr->path_mig_state = IB_MIG_REARM;
3655                 }
3656                 ret = 0;
3657                 break;
3658         default:
3659                 ret = -EINVAL;
3660                 break;
3661         }
3662         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3663         return ret;
3664 }
3665
3666 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3667                        struct ib_qp_attr *qp_attr,
3668                        int *qp_attr_mask)
3669 {
3670         struct cm_id_private *cm_id_priv;
3671         int ret;
3672
3673         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3674         switch (qp_attr->qp_state) {
3675         case IB_QPS_INIT:
3676                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3677                 break;
3678         case IB_QPS_RTR:
3679                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3680                 break;
3681         case IB_QPS_RTS:
3682                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3683                 break;
3684         default:
3685                 ret = -EINVAL;
3686                 break;
3687         }
3688         return ret;
3689 }
3690 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3691
3692 static void cm_get_ack_delay(struct cm_device *cm_dev)
3693 {
3694         struct ib_device_attr attr;
3695
3696         if (ib_query_device(cm_dev->ib_device, &attr))
3697                 cm_dev->ack_delay = 0; /* acks will rely on packet life time */
3698         else
3699                 cm_dev->ack_delay = attr.local_ca_ack_delay;
3700 }
3701
3702 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3703                                char *buf)
3704 {
3705         struct cm_counter_group *group;
3706         struct cm_counter_attribute *cm_attr;
3707
3708         group = container_of(obj, struct cm_counter_group, obj);
3709         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3710
3711         return sprintf(buf, "%ld\n",
3712                        atomic_long_read(&group->counter[cm_attr->index]));
3713 }
3714
3715 static const struct sysfs_ops cm_counter_ops = {
3716         .show = cm_show_counter
3717 };
3718
3719 static struct kobj_type cm_counter_obj_type = {
3720         .sysfs_ops = &cm_counter_ops,
3721         .default_attrs = cm_counter_default_attrs
3722 };
3723
3724 static void cm_release_port_obj(struct kobject *obj)
3725 {
3726         struct cm_port *cm_port;
3727
3728         cm_port = container_of(obj, struct cm_port, port_obj);
3729         kfree(cm_port);
3730 }
3731
3732 static struct kobj_type cm_port_obj_type = {
3733         .release = cm_release_port_obj
3734 };
3735
3736 static char *cm_devnode(struct device *dev, umode_t *mode)
3737 {
3738         if (mode)
3739                 *mode = 0666;
3740         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3741 }
3742
3743 struct class cm_class = {
3744         .owner   = THIS_MODULE,
3745         .name    = "infiniband_cm",
3746         .devnode = cm_devnode,
3747 };
3748 EXPORT_SYMBOL(cm_class);
3749
3750 static int cm_create_port_fs(struct cm_port *port)
3751 {
3752         int i, ret;
3753
3754         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3755                                    &port->cm_dev->device->kobj,
3756                                    "%d", port->port_num);
3757         if (ret) {
3758                 kfree(port);
3759                 return ret;
3760         }
3761
3762         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3763                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3764                                            &cm_counter_obj_type,
3765                                            &port->port_obj,
3766                                            "%s", counter_group_names[i]);
3767                 if (ret)
3768                         goto error;
3769         }
3770
3771         return 0;
3772
3773 error:
3774         while (i--)
3775                 kobject_put(&port->counter_group[i].obj);
3776         kobject_put(&port->port_obj);
3777         return ret;
3778
3779 }
3780
3781 static void cm_remove_port_fs(struct cm_port *port)
3782 {
3783         int i;
3784
3785         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3786                 kobject_put(&port->counter_group[i].obj);
3787
3788         kobject_put(&port->port_obj);
3789 }
3790
3791 static void cm_add_one(struct ib_device *ib_device)
3792 {
3793         struct cm_device *cm_dev;
3794         struct cm_port *port;
3795         struct ib_mad_reg_req reg_req = {
3796                 .mgmt_class = IB_MGMT_CLASS_CM,
3797                 .mgmt_class_version = IB_CM_CLASS_VERSION,
3798         };
3799         struct ib_port_modify port_modify = {
3800                 .set_port_cap_mask = IB_PORT_CM_SUP
3801         };
3802         unsigned long flags;
3803         int ret;
3804         int count = 0;
3805         u8 i;
3806
3807         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3808                          ib_device->phys_port_cnt, GFP_KERNEL);
3809         if (!cm_dev)
3810                 return;
3811
3812         cm_dev->ib_device = ib_device;
3813         cm_get_ack_delay(cm_dev);
3814         cm_dev->going_down = 0;
3815         cm_dev->device = device_create(&cm_class, &ib_device->dev,
3816                                        MKDEV(0, 0), NULL,
3817                                        "%s", ib_device->name);
3818         if (IS_ERR(cm_dev->device)) {
3819                 kfree(cm_dev);
3820                 return;
3821         }
3822
3823         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3824         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3825                 if (!rdma_cap_ib_cm(ib_device, i))
3826                         continue;
3827
3828                 port = kzalloc(sizeof *port, GFP_KERNEL);
3829                 if (!port)
3830                         goto error1;
3831
3832                 cm_dev->port[i-1] = port;
3833                 port->cm_dev = cm_dev;
3834                 port->port_num = i;
3835
3836                 ret = cm_create_port_fs(port);
3837                 if (ret)
3838                         goto error1;
3839
3840                 port->mad_agent = ib_register_mad_agent(ib_device, i,
3841                                                         IB_QPT_GSI,
3842                                                         &reg_req,
3843                                                         0,
3844                                                         cm_send_handler,
3845                                                         cm_recv_handler,
3846                                                         port,
3847                                                         0);
3848                 if (IS_ERR(port->mad_agent))
3849                         goto error2;
3850
3851                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
3852                 if (ret)
3853                         goto error3;
3854
3855                 count++;
3856         }
3857
3858         if (!count)
3859                 goto free;
3860
3861         ib_set_client_data(ib_device, &cm_client, cm_dev);
3862
3863         write_lock_irqsave(&cm.device_lock, flags);
3864         list_add_tail(&cm_dev->list, &cm.device_list);
3865         write_unlock_irqrestore(&cm.device_lock, flags);
3866         return;
3867
3868 error3:
3869         ib_unregister_mad_agent(port->mad_agent);
3870 error2:
3871         cm_remove_port_fs(port);
3872 error1:
3873         port_modify.set_port_cap_mask = 0;
3874         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3875         while (--i) {
3876                 if (!rdma_cap_ib_cm(ib_device, i))
3877                         continue;
3878
3879                 port = cm_dev->port[i-1];
3880                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3881                 ib_unregister_mad_agent(port->mad_agent);
3882                 cm_remove_port_fs(port);
3883         }
3884 free:
3885         device_unregister(cm_dev->device);
3886         kfree(cm_dev);
3887 }
3888
3889 static void cm_remove_one(struct ib_device *ib_device)
3890 {
3891         struct cm_device *cm_dev;
3892         struct cm_port *port;
3893         struct ib_port_modify port_modify = {
3894                 .clr_port_cap_mask = IB_PORT_CM_SUP
3895         };
3896         unsigned long flags;
3897         int i;
3898
3899         cm_dev = ib_get_client_data(ib_device, &cm_client);
3900         if (!cm_dev)
3901                 return;
3902
3903         write_lock_irqsave(&cm.device_lock, flags);
3904         list_del(&cm_dev->list);
3905         write_unlock_irqrestore(&cm.device_lock, flags);
3906
3907         spin_lock_irq(&cm.lock);
3908         cm_dev->going_down = 1;
3909         spin_unlock_irq(&cm.lock);
3910
3911         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3912                 if (!rdma_cap_ib_cm(ib_device, i))
3913                         continue;
3914
3915                 port = cm_dev->port[i-1];
3916                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3917                 /*
3918                  * We flush the queue here after the going_down set, this
3919                  * verify that no new works will be queued in the recv handler,
3920                  * after that we can call the unregister_mad_agent
3921                  */
3922                 flush_workqueue(cm.wq);
3923                 ib_unregister_mad_agent(port->mad_agent);
3924                 cm_remove_port_fs(port);
3925         }
3926         device_unregister(cm_dev->device);
3927         kfree(cm_dev);
3928 }
3929
3930 static int __init ib_cm_init(void)
3931 {
3932         int ret;
3933
3934         memset(&cm, 0, sizeof cm);
3935         INIT_LIST_HEAD(&cm.device_list);
3936         rwlock_init(&cm.device_lock);
3937         spin_lock_init(&cm.lock);
3938         cm.listen_service_table = RB_ROOT;
3939         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3940         cm.remote_id_table = RB_ROOT;
3941         cm.remote_qp_table = RB_ROOT;
3942         cm.remote_sidr_table = RB_ROOT;
3943         idr_init(&cm.local_id_table);
3944         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
3945         INIT_LIST_HEAD(&cm.timewait_list);
3946
3947         ret = class_register(&cm_class);
3948         if (ret) {
3949                 ret = -ENOMEM;
3950                 goto error1;
3951         }
3952
3953         cm.wq = create_workqueue("ib_cm");
3954         if (!cm.wq) {
3955                 ret = -ENOMEM;
3956                 goto error2;
3957         }
3958
3959         ret = ib_register_client(&cm_client);
3960         if (ret)
3961                 goto error3;
3962
3963         return 0;
3964 error3:
3965         destroy_workqueue(cm.wq);
3966 error2:
3967         class_unregister(&cm_class);
3968 error1:
3969         idr_destroy(&cm.local_id_table);
3970         return ret;
3971 }
3972
3973 static void __exit ib_cm_cleanup(void)
3974 {
3975         struct cm_timewait_info *timewait_info, *tmp;
3976
3977         spin_lock_irq(&cm.lock);
3978         list_for_each_entry(timewait_info, &cm.timewait_list, list)
3979                 cancel_delayed_work(&timewait_info->work.work);
3980         spin_unlock_irq(&cm.lock);
3981
3982         ib_unregister_client(&cm_client);
3983         destroy_workqueue(cm.wq);
3984
3985         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
3986                 list_del(&timewait_info->list);
3987                 kfree(timewait_info);
3988         }
3989
3990         class_unregister(&cm_class);
3991         idr_destroy(&cm.local_id_table);
3992 }
3993
3994 module_init(ib_cm_init);
3995 module_exit(ib_cm_cleanup);
3996