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  * $Id: cm.c 4311 2005-12-05 18:42:01Z sean.hefty $
  36  */
  37
  38 #include <linux/completion.h>
  39 #include <linux/dma-mapping.h>
  40 #include <linux/device.h>
  41 #include <linux/err.h>
  42 #include <linux/idr.h>
  43 #include <linux/interrupt.h>
  44 #include <linux/random.h>
  45 #include <linux/rbtree.h>
  46 #include <linux/spinlock.h>
  47 #include <linux/sysfs.h>
  48 #include <linux/workqueue.h>
  49
  50 #include <rdma/ib_cache.h>
  51 #include <rdma/ib_cm.h>
  52 #include "cm_msgs.h"
  53
  54 MODULE_AUTHOR("Sean Hefty");
  55 MODULE_DESCRIPTION("InfiniBand CM");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 static void cm_add_one(struct ib_device *device);
  59 static void cm_remove_one(struct ib_device *device);
  60
  61 static struct ib_client cm_client = {
  62         .name   = "cm",
  63         .add    = cm_add_one,
  64         .remove = cm_remove_one
  65 };
  66
  67 static struct ib_cm {
  68         spinlock_t lock;
  69         struct list_head device_list;
  70         rwlock_t device_lock;
  71         struct rb_root listen_service_table;
  72         u64 listen_service_id;
  73         /* struct rb_root peer_service_table; todo: fix peer to peer */
  74         struct rb_root remote_qp_table;
  75         struct rb_root remote_id_table;
  76         struct rb_root remote_sidr_table;
  77         struct idr local_id_table;
  78         __be32 random_id_operand;
  79         struct list_head timewait_list;
  80         struct workqueue_struct *wq;
  81 } cm;
  82
  83 /* Counter indexes ordered by attribute ID */
  84 enum {
  85         CM_REQ_COUNTER,
  86         CM_MRA_COUNTER,
  87         CM_REJ_COUNTER,
  88         CM_REP_COUNTER,
  89         CM_RTU_COUNTER,
  90         CM_DREQ_COUNTER,
  91         CM_DREP_COUNTER,
  92         CM_SIDR_REQ_COUNTER,
  93         CM_SIDR_REP_COUNTER,
  94         CM_LAP_COUNTER,
  95         CM_APR_COUNTER,
  96         CM_ATTR_COUNT,
  97         CM_ATTR_ID_OFFSET = 0x0010,
  98 };
  99
 100 enum {
 101         CM_XMIT,
 102         CM_XMIT_RETRIES,
 103         CM_RECV,
 104         CM_RECV_DUPLICATES,
 105         CM_COUNTER_GROUPS
 106 };
 107
 108 static char const counter_group_names[CM_COUNTER_GROUPS]
 109                                      [sizeof("cm_rx_duplicates")] = {
 110         "cm_tx_msgs", "cm_tx_retries",
 111         "cm_rx_msgs", "cm_rx_duplicates"
 112 };
 113
 114 struct cm_counter_group {
 115         struct kobject obj;
 116         atomic_long_t counter[CM_ATTR_COUNT];
 117 };
 118
 119 struct cm_counter_attribute {
 120         struct attribute attr;
 121         int index;
 122 };
 123
 124 #define CM_COUNTER_ATTR(_name, _index) \
 125 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 126         .attr = { .name = __stringify(_name), .mode = 0444, .owner = THIS_MODULE }, \
 127         .index = _index \
 128 }
 129
 130 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 131 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 132 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 133 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 134 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 135 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 136 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 137 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 138 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 139 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 140 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 141
 142 static struct attribute *cm_counter_default_attrs[] = {
 143         &cm_req_counter_attr.attr,
 144         &cm_mra_counter_attr.attr,
 145         &cm_rej_counter_attr.attr,
 146         &cm_rep_counter_attr.attr,
 147         &cm_rtu_counter_attr.attr,
 148         &cm_dreq_counter_attr.attr,
 149         &cm_drep_counter_attr.attr,
 150         &cm_sidr_req_counter_attr.attr,
 151         &cm_sidr_rep_counter_attr.attr,
 152         &cm_lap_counter_attr.attr,
 153         &cm_apr_counter_attr.attr,
 154         NULL
 155 };
 156
 157 struct cm_port {
 158         struct cm_device *cm_dev;
 159         struct ib_mad_agent *mad_agent;
 160         struct kobject port_obj;
 161         u8 port_num;
 162         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 163 };
 164
 165 struct cm_device {
 166         struct list_head list;
 167         struct ib_device *device;
 168         struct kobject dev_obj;
 169         u8 ack_delay;
 170         struct cm_port *port[0];
 171 };
 172
 173 struct cm_av {
 174         struct cm_port *port;
 175         union ib_gid dgid;
 176         struct ib_ah_attr ah_attr;
 177         u16 pkey_index;
 178         u8 timeout;
 179 };
 180
 181 struct cm_work {
 182         struct delayed_work work;
 183         struct list_head list;
 184         struct cm_port *port;
 185         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 186         __be32 local_id;                        /* Established / timewait */
 187         __be32 remote_id;
 188         struct ib_cm_event cm_event;
 189         struct ib_sa_path_rec path[0];
 190 };
 191
 192 struct cm_timewait_info {
 193         struct cm_work work;                    /* Must be first. */
 194         struct list_head list;
 195         struct rb_node remote_qp_node;
 196         struct rb_node remote_id_node;
 197         __be64 remote_ca_guid;
 198         __be32 remote_qpn;
 199         u8 inserted_remote_qp;
 200         u8 inserted_remote_id;
 201 };
 202
 203 struct cm_id_private {
 204         struct ib_cm_id id;
 205
 206         struct rb_node service_node;
 207         struct rb_node sidr_id_node;
 208         spinlock_t lock;        /* Do not acquire inside cm.lock */
 209         struct completion comp;
 210         atomic_t refcount;
 211
 212         struct ib_mad_send_buf *msg;
 213         struct cm_timewait_info *timewait_info;
 214         /* todo: use alternate port on send failure */
 215         struct cm_av av;
 216         struct cm_av alt_av;
 217         struct ib_cm_compare_data *compare_data;
 218
 219         void *private_data;
 220         __be64 tid;
 221         __be32 local_qpn;
 222         __be32 remote_qpn;
 223         enum ib_qp_type qp_type;
 224         __be32 sq_psn;
 225         __be32 rq_psn;
 226         int timeout_ms;
 227         enum ib_mtu path_mtu;
 228         __be16 pkey;
 229         u8 private_data_len;
 230         u8 max_cm_retries;
 231         u8 peer_to_peer;
 232         u8 responder_resources;
 233         u8 initiator_depth;
 234         u8 retry_count;
 235         u8 rnr_retry_count;
 236         u8 service_timeout;
 237         u8 target_ack_delay;
 238
 239         struct list_head work_list;
 240         atomic_t work_count;
 241 };
 242
 243 static void cm_work_handler(struct work_struct *work);
 244
 245 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 246 {
 247         if (atomic_dec_and_test(&cm_id_priv->refcount))
 248                 complete(&cm_id_priv->comp);
 249 }
 250
 251 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 252                         struct ib_mad_send_buf **msg)
 253 {
 254         struct ib_mad_agent *mad_agent;
 255         struct ib_mad_send_buf *m;
 256         struct ib_ah *ah;
 257
 258         mad_agent = cm_id_priv->av.port->mad_agent;
 259         ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
 260         if (IS_ERR(ah))
 261                 return PTR_ERR(ah);
 262
 263         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 264                                cm_id_priv->av.pkey_index,
 265                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 266                                GFP_ATOMIC);
 267         if (IS_ERR(m)) {
 268                 ib_destroy_ah(ah);
 269                 return PTR_ERR(m);
 270         }
 271
 272         /* Timeout set by caller if response is expected. */
 273         m->ah = ah;
 274         m->retries = cm_id_priv->max_cm_retries;
 275
 276         atomic_inc(&cm_id_priv->refcount);
 277         m->context[0] = cm_id_priv;
 278         *msg = m;
 279         return 0;
 280 }
 281
 282 static int cm_alloc_response_msg(struct cm_port *port,
 283                                  struct ib_mad_recv_wc *mad_recv_wc,
 284                                  struct ib_mad_send_buf **msg)
 285 {
 286         struct ib_mad_send_buf *m;
 287         struct ib_ah *ah;
 288
 289         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 290                                   mad_recv_wc->recv_buf.grh, port->port_num);
 291         if (IS_ERR(ah))
 292                 return PTR_ERR(ah);
 293
 294         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 295                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 296                                GFP_ATOMIC);
 297         if (IS_ERR(m)) {
 298                 ib_destroy_ah(ah);
 299                 return PTR_ERR(m);
 300         }
 301         m->ah = ah;
 302         *msg = m;
 303         return 0;
 304 }
 305
 306 static void cm_free_msg(struct ib_mad_send_buf *msg)
 307 {
 308         ib_destroy_ah(msg->ah);
 309         if (msg->context[0])
 310                 cm_deref_id(msg->context[0]);
 311         ib_free_send_mad(msg);
 312 }
 313
 314 static void * cm_copy_private_data(const void *private_data,
 315                                    u8 private_data_len)
 316 {
 317         void *data;
 318
 319         if (!private_data || !private_data_len)
 320                 return NULL;
 321
 322         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 323         if (!data)
 324                 return ERR_PTR(-ENOMEM);
 325
 326         return data;
 327 }
 328
 329 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 330                                  void *private_data, u8 private_data_len)
 331 {
 332         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 333                 kfree(cm_id_priv->private_data);
 334
 335         cm_id_priv->private_data = private_data;
 336         cm_id_priv->private_data_len = private_data_len;
 337 }
 338
 339 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 340                                     struct ib_grh *grh, struct cm_av *av)
 341 {
 342         av->port = port;
 343         av->pkey_index = wc->pkey_index;
 344         ib_init_ah_from_wc(port->cm_dev->device, port->port_num, wc,
 345                            grh, &av->ah_attr);
 346 }
 347
 348 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
 349 {
 350         struct cm_device *cm_dev;
 351         struct cm_port *port = NULL;
 352         unsigned long flags;
 353         int ret;
 354         u8 p;
 355
 356         read_lock_irqsave(&cm.device_lock, flags);
 357         list_for_each_entry(cm_dev, &cm.device_list, list) {
 358                 if (!ib_find_cached_gid(cm_dev->device, &path->sgid,
 359                                         &p, NULL)) {
 360                         port = cm_dev->port[p-1];
 361                         break;
 362                 }
 363         }
 364         read_unlock_irqrestore(&cm.device_lock, flags);
 365
 366         if (!port)
 367                 return -EINVAL;
 368
 369         ret = ib_find_cached_pkey(cm_dev->device, port->port_num,
 370                                   be16_to_cpu(path->pkey), &av->pkey_index);
 371         if (ret)
 372                 return ret;
 373
 374         av->port = port;
 375         ib_init_ah_from_path(cm_dev->device, port->port_num, path,
 376                              &av->ah_attr);
 377         av->timeout = path->packet_life_time + 1;
 378         return 0;
 379 }
 380
 381 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 382 {
 383         unsigned long flags;
 384         int ret, id;
 385         static int next_id;
 386
 387         do {
 388                 spin_lock_irqsave(&cm.lock, flags);
 389                 ret = idr_get_new_above(&cm.local_id_table, cm_id_priv,
 390                                         next_id, &id);
 391                 if (!ret)
 392                         next_id = ((unsigned) id + 1) & MAX_ID_MASK;
 393                 spin_unlock_irqrestore(&cm.lock, flags);
 394         } while( (ret == -EAGAIN) && idr_pre_get(&cm.local_id_table, GFP_KERNEL) );
 395
 396         cm_id_priv->id.local_id = (__force __be32) (id ^ cm.random_id_operand);
 397         return ret;
 398 }
 399
 400 static void cm_free_id(__be32 local_id)
 401 {
 402         spin_lock_irq(&cm.lock);
 403         idr_remove(&cm.local_id_table,
 404                    (__force int) (local_id ^ cm.random_id_operand));
 405         spin_unlock_irq(&cm.lock);
 406 }
 407
 408 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 409 {
 410         struct cm_id_private *cm_id_priv;
 411
 412         cm_id_priv = idr_find(&cm.local_id_table,
 413                               (__force int) (local_id ^ cm.random_id_operand));
 414         if (cm_id_priv) {
 415                 if (cm_id_priv->id.remote_id == remote_id)
 416                         atomic_inc(&cm_id_priv->refcount);
 417                 else
 418                         cm_id_priv = NULL;
 419         }
 420
 421         return cm_id_priv;
 422 }
 423
 424 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 425 {
 426         struct cm_id_private *cm_id_priv;
 427
 428         spin_lock_irq(&cm.lock);
 429         cm_id_priv = cm_get_id(local_id, remote_id);
 430         spin_unlock_irq(&cm.lock);
 431
 432         return cm_id_priv;
 433 }
 434
 435 static void cm_mask_copy(u8 *dst, u8 *src, u8 *mask)
 436 {
 437         int i;
 438
 439         for (i = 0; i < IB_CM_COMPARE_SIZE / sizeof(unsigned long); i++)
 440                 ((unsigned long *) dst)[i] = ((unsigned long *) src)[i] &
 441                                              ((unsigned long *) mask)[i];
 442 }
 443
 444 static int cm_compare_data(struct ib_cm_compare_data *src_data,
 445                            struct ib_cm_compare_data *dst_data)
 446 {
 447         u8 src[IB_CM_COMPARE_SIZE];
 448         u8 dst[IB_CM_COMPARE_SIZE];
 449
 450         if (!src_data || !dst_data)
 451                 return 0;
 452
 453         cm_mask_copy(src, src_data->data, dst_data->mask);
 454         cm_mask_copy(dst, dst_data->data, src_data->mask);
 455         return memcmp(src, dst, IB_CM_COMPARE_SIZE);
 456 }
 457
 458 static int cm_compare_private_data(u8 *private_data,
 459                                    struct ib_cm_compare_data *dst_data)
 460 {
 461         u8 src[IB_CM_COMPARE_SIZE];
 462
 463         if (!dst_data)
 464                 return 0;
 465
 466         cm_mask_copy(src, private_data, dst_data->mask);
 467         return memcmp(src, dst_data->data, IB_CM_COMPARE_SIZE);
 468 }
 469
 470 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 471 {
 472         struct rb_node **link = &cm.listen_service_table.rb_node;
 473         struct rb_node *parent = NULL;
 474         struct cm_id_private *cur_cm_id_priv;
 475         __be64 service_id = cm_id_priv->id.service_id;
 476         __be64 service_mask = cm_id_priv->id.service_mask;
 477         int data_cmp;
 478
 479         while (*link) {
 480                 parent = *link;
 481                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 482                                           service_node);
 483                 data_cmp = cm_compare_data(cm_id_priv->compare_data,
 484                                            cur_cm_id_priv->compare_data);
 485                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 486                     (service_mask & cur_cm_id_priv->id.service_id) &&
 487                     (cm_id_priv->id.device == cur_cm_id_priv->id.device) &&
 488                     !data_cmp)
 489                         return cur_cm_id_priv;
 490
 491                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 492                         link = &(*link)->rb_left;
 493                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 494                         link = &(*link)->rb_right;
 495                 else if (service_id < cur_cm_id_priv->id.service_id)
 496                         link = &(*link)->rb_left;
 497                 else if (service_id > cur_cm_id_priv->id.service_id)
 498                         link = &(*link)->rb_right;
 499                 else if (data_cmp < 0)
 500                         link = &(*link)->rb_left;
 501                 else
 502                         link = &(*link)->rb_right;
 503         }
 504         rb_link_node(&cm_id_priv->service_node, parent, link);
 505         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 506         return NULL;
 507 }
 508
 509 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 510                                              __be64 service_id,
 511                                              u8 *private_data)
 512 {
 513         struct rb_node *node = cm.listen_service_table.rb_node;
 514         struct cm_id_private *cm_id_priv;
 515         int data_cmp;
 516
 517         while (node) {
 518                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 519                 data_cmp = cm_compare_private_data(private_data,
 520                                                    cm_id_priv->compare_data);
 521                 if ((cm_id_priv->id.service_mask & service_id) ==
 522                      cm_id_priv->id.service_id &&
 523                     (cm_id_priv->id.device == device) && !data_cmp)
 524                         return cm_id_priv;
 525
 526                 if (device < cm_id_priv->id.device)
 527                         node = node->rb_left;
 528                 else if (device > cm_id_priv->id.device)
 529                         node = node->rb_right;
 530                 else if (service_id < cm_id_priv->id.service_id)
 531                         node = node->rb_left;
 532                 else if (service_id > cm_id_priv->id.service_id)
 533                         node = node->rb_right;
 534                 else if (data_cmp < 0)
 535                         node = node->rb_left;
 536                 else
 537                         node = node->rb_right;
 538         }
 539         return NULL;
 540 }
 541
 542 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 543                                                      *timewait_info)
 544 {
 545         struct rb_node **link = &cm.remote_id_table.rb_node;
 546         struct rb_node *parent = NULL;
 547         struct cm_timewait_info *cur_timewait_info;
 548         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 549         __be32 remote_id = timewait_info->work.remote_id;
 550
 551         while (*link) {
 552                 parent = *link;
 553                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 554                                              remote_id_node);
 555                 if (remote_id < cur_timewait_info->work.remote_id)
 556                         link = &(*link)->rb_left;
 557                 else if (remote_id > cur_timewait_info->work.remote_id)
 558                         link = &(*link)->rb_right;
 559                 else if (remote_ca_guid < cur_timewait_info->remote_ca_guid)
 560                         link = &(*link)->rb_left;
 561                 else if (remote_ca_guid > cur_timewait_info->remote_ca_guid)
 562                         link = &(*link)->rb_right;
 563                 else
 564                         return cur_timewait_info;
 565         }
 566         timewait_info->inserted_remote_id = 1;
 567         rb_link_node(&timewait_info->remote_id_node, parent, link);
 568         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 569         return NULL;
 570 }
 571
 572 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 573                                                    __be32 remote_id)
 574 {
 575         struct rb_node *node = cm.remote_id_table.rb_node;
 576         struct cm_timewait_info *timewait_info;
 577
 578         while (node) {
 579                 timewait_info = rb_entry(node, struct cm_timewait_info,
 580                                          remote_id_node);
 581                 if (remote_id < timewait_info->work.remote_id)
 582                         node = node->rb_left;
 583                 else if (remote_id > timewait_info->work.remote_id)
 584                         node = node->rb_right;
 585                 else if (remote_ca_guid < timewait_info->remote_ca_guid)
 586                         node = node->rb_left;
 587                 else if (remote_ca_guid > timewait_info->remote_ca_guid)
 588                         node = node->rb_right;
 589                 else
 590                         return timewait_info;
 591         }
 592         return NULL;
 593 }
 594
 595 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 596                                                       *timewait_info)
 597 {
 598         struct rb_node **link = &cm.remote_qp_table.rb_node;
 599         struct rb_node *parent = NULL;
 600         struct cm_timewait_info *cur_timewait_info;
 601         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 602         __be32 remote_qpn = timewait_info->remote_qpn;
 603
 604         while (*link) {
 605                 parent = *link;
 606                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 607                                              remote_qp_node);
 608                 if (remote_qpn < cur_timewait_info->remote_qpn)
 609                         link = &(*link)->rb_left;
 610                 else if (remote_qpn > cur_timewait_info->remote_qpn)
 611                         link = &(*link)->rb_right;
 612                 else if (remote_ca_guid < cur_timewait_info->remote_ca_guid)
 613                         link = &(*link)->rb_left;
 614                 else if (remote_ca_guid > cur_timewait_info->remote_ca_guid)
 615                         link = &(*link)->rb_right;
 616                 else
 617                         return cur_timewait_info;
 618         }
 619         timewait_info->inserted_remote_qp = 1;
 620         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 621         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 622         return NULL;
 623 }
 624
 625 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 626                                                     *cm_id_priv)
 627 {
 628         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 629         struct rb_node *parent = NULL;
 630         struct cm_id_private *cur_cm_id_priv;
 631         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 632         __be32 remote_id = cm_id_priv->id.remote_id;
 633
 634         while (*link) {
 635                 parent = *link;
 636                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 637                                           sidr_id_node);
 638                 if (remote_id < cur_cm_id_priv->id.remote_id)
 639                         link = &(*link)->rb_left;
 640                 else if (remote_id > cur_cm_id_priv->id.remote_id)
 641                         link = &(*link)->rb_right;
 642                 else {
 643                         int cmp;
 644                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 645                                      sizeof *port_gid);
 646                         if (cmp < 0)
 647                                 link = &(*link)->rb_left;
 648                         else if (cmp > 0)
 649                                 link = &(*link)->rb_right;
 650                         else
 651                                 return cur_cm_id_priv;
 652                 }
 653         }
 654         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 655         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 656         return NULL;
 657 }
 658
 659 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 660                                enum ib_cm_sidr_status status)
 661 {
 662         struct ib_cm_sidr_rep_param param;
 663
 664         memset(&param, 0, sizeof param);
 665         param.status = status;
 666         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 667 }
 668
 669 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 670                                  ib_cm_handler cm_handler,
 671                                  void *context)
 672 {
 673         struct cm_id_private *cm_id_priv;
 674         int ret;
 675
 676         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 677         if (!cm_id_priv)
 678                 return ERR_PTR(-ENOMEM);
 679
 680         cm_id_priv->id.state = IB_CM_IDLE;
 681         cm_id_priv->id.device = device;
 682         cm_id_priv->id.cm_handler = cm_handler;
 683         cm_id_priv->id.context = context;
 684         cm_id_priv->id.remote_cm_qpn = 1;
 685         ret = cm_alloc_id(cm_id_priv);
 686         if (ret)
 687                 goto error;
 688
 689         spin_lock_init(&cm_id_priv->lock);
 690         init_completion(&cm_id_priv->comp);
 691         INIT_LIST_HEAD(&cm_id_priv->work_list);
 692         atomic_set(&cm_id_priv->work_count, -1);
 693         atomic_set(&cm_id_priv->refcount, 1);
 694         return &cm_id_priv->id;
 695
 696 error:
 697         kfree(cm_id_priv);
 698         return ERR_PTR(-ENOMEM);
 699 }
 700 EXPORT_SYMBOL(ib_create_cm_id);
 701
 702 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 703 {
 704         struct cm_work *work;
 705
 706         if (list_empty(&cm_id_priv->work_list))
 707                 return NULL;
 708
 709         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 710         list_del(&work->list);
 711         return work;
 712 }
 713
 714 static void cm_free_work(struct cm_work *work)
 715 {
 716         if (work->mad_recv_wc)
 717                 ib_free_recv_mad(work->mad_recv_wc);
 718         kfree(work);
 719 }
 720
 721 static inline int cm_convert_to_ms(int iba_time)
 722 {
 723         /* approximate conversion to ms from 4.096us x 2^iba_time */
 724         return 1 << max(iba_time - 8, 0);
 725 }
 726
 727 /*
 728  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 729  * Because of how ack_timeout is stored, adding one doubles the timeout.
 730  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 731  * increment it (round up) only if the other is within 50%.
 732  */
 733 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 734 {
 735         int ack_timeout = packet_life_time + 1;
 736
 737         if (ack_timeout >= ca_ack_delay)
 738                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 739         else
 740                 ack_timeout = ca_ack_delay +
 741                               (ack_timeout >= (ca_ack_delay - 1));
 742
 743         return min(31, ack_timeout);
 744 }
 745
 746 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 747 {
 748         if (timewait_info->inserted_remote_id) {
 749                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 750                 timewait_info->inserted_remote_id = 0;
 751         }
 752
 753         if (timewait_info->inserted_remote_qp) {
 754                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 755                 timewait_info->inserted_remote_qp = 0;
 756         }
 757 }
 758
 759 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 760 {
 761         struct cm_timewait_info *timewait_info;
 762
 763         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 764         if (!timewait_info)
 765                 return ERR_PTR(-ENOMEM);
 766
 767         timewait_info->work.local_id = local_id;
 768         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 769         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 770         return timewait_info;
 771 }
 772
 773 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 774 {
 775         int wait_time;
 776         unsigned long flags;
 777
 778         spin_lock_irqsave(&cm.lock, flags);
 779         cm_cleanup_timewait(cm_id_priv->timewait_info);
 780         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 781         spin_unlock_irqrestore(&cm.lock, flags);
 782
 783         /*
 784          * The cm_id could be destroyed by the user before we exit timewait.
 785          * To protect against this, we search for the cm_id after exiting
 786          * timewait before notifying the user that we've exited timewait.
 787          */
 788         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 789         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 790         queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 791                            msecs_to_jiffies(wait_time));
 792         cm_id_priv->timewait_info = NULL;
 793 }
 794
 795 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 796 {
 797         unsigned long flags;
 798
 799         cm_id_priv->id.state = IB_CM_IDLE;
 800         if (cm_id_priv->timewait_info) {
 801                 spin_lock_irqsave(&cm.lock, flags);
 802                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 803                 spin_unlock_irqrestore(&cm.lock, flags);
 804                 kfree(cm_id_priv->timewait_info);
 805                 cm_id_priv->timewait_info = NULL;
 806         }
 807 }
 808
 809 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 810 {
 811         struct cm_id_private *cm_id_priv;
 812         struct cm_work *work;
 813
 814         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 815 retest:
 816         spin_lock_irq(&cm_id_priv->lock);
 817         switch (cm_id->state) {
 818         case IB_CM_LISTEN:
 819                 cm_id->state = IB_CM_IDLE;
 820                 spin_unlock_irq(&cm_id_priv->lock);
 821                 spin_lock_irq(&cm.lock);
 822                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 823                 spin_unlock_irq(&cm.lock);
 824                 break;
 825         case IB_CM_SIDR_REQ_SENT:
 826                 cm_id->state = IB_CM_IDLE;
 827                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 828                 spin_unlock_irq(&cm_id_priv->lock);
 829                 break;
 830         case IB_CM_SIDR_REQ_RCVD:
 831                 spin_unlock_irq(&cm_id_priv->lock);
 832                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 833                 break;
 834         case IB_CM_REQ_SENT:
 835                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 836                 spin_unlock_irq(&cm_id_priv->lock);
 837                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 838                                &cm_id_priv->id.device->node_guid,
 839                                sizeof cm_id_priv->id.device->node_guid,
 840                                NULL, 0);
 841                 break;
 842         case IB_CM_REQ_RCVD:
 843                 if (err == -ENOMEM) {
 844                         /* Do not reject to allow future retries. */
 845                         cm_reset_to_idle(cm_id_priv);
 846                         spin_unlock_irq(&cm_id_priv->lock);
 847                 } else {
 848                         spin_unlock_irq(&cm_id_priv->lock);
 849                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 850                                        NULL, 0, NULL, 0);
 851                 }
 852                 break;
 853         case IB_CM_MRA_REQ_RCVD:
 854         case IB_CM_REP_SENT:
 855         case IB_CM_MRA_REP_RCVD:
 856                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 857                 /* Fall through */
 858         case IB_CM_MRA_REQ_SENT:
 859         case IB_CM_REP_RCVD:
 860         case IB_CM_MRA_REP_SENT:
 861                 spin_unlock_irq(&cm_id_priv->lock);
 862                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 863                                NULL, 0, NULL, 0);
 864                 break;
 865         case IB_CM_ESTABLISHED:
 866                 spin_unlock_irq(&cm_id_priv->lock);
 867                 ib_send_cm_dreq(cm_id, NULL, 0);
 868                 goto retest;
 869         case IB_CM_DREQ_SENT:
 870                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 871                 cm_enter_timewait(cm_id_priv);
 872                 spin_unlock_irq(&cm_id_priv->lock);
 873                 break;
 874         case IB_CM_DREQ_RCVD:
 875                 spin_unlock_irq(&cm_id_priv->lock);
 876                 ib_send_cm_drep(cm_id, NULL, 0);
 877                 break;
 878         default:
 879                 spin_unlock_irq(&cm_id_priv->lock);
 880                 break;
 881         }
 882
 883         cm_free_id(cm_id->local_id);
 884         cm_deref_id(cm_id_priv);
 885         wait_for_completion(&cm_id_priv->comp);
 886         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
 887                 cm_free_work(work);
 888         kfree(cm_id_priv->compare_data);
 889         kfree(cm_id_priv->private_data);
 890         kfree(cm_id_priv);
 891 }
 892
 893 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
 894 {
 895         cm_destroy_id(cm_id, 0);
 896 }
 897 EXPORT_SYMBOL(ib_destroy_cm_id);
 898
 899 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask,
 900                  struct ib_cm_compare_data *compare_data)
 901 {
 902         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
 903         unsigned long flags;
 904         int ret = 0;
 905
 906         service_mask = service_mask ? service_mask :
 907                        __constant_cpu_to_be64(~0ULL);
 908         service_id &= service_mask;
 909         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
 910             (service_id != IB_CM_ASSIGN_SERVICE_ID))
 911                 return -EINVAL;
 912
 913         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 914         if (cm_id->state != IB_CM_IDLE)
 915                 return -EINVAL;
 916
 917         if (compare_data) {
 918                 cm_id_priv->compare_data = kzalloc(sizeof *compare_data,
 919                                                    GFP_KERNEL);
 920                 if (!cm_id_priv->compare_data)
 921                         return -ENOMEM;
 922                 cm_mask_copy(cm_id_priv->compare_data->data,
 923                              compare_data->data, compare_data->mask);
 924                 memcpy(cm_id_priv->compare_data->mask, compare_data->mask,
 925                        IB_CM_COMPARE_SIZE);
 926         }
 927
 928         cm_id->state = IB_CM_LISTEN;
 929
 930         spin_lock_irqsave(&cm.lock, flags);
 931         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
 932                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
 933                 cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
 934         } else {
 935                 cm_id->service_id = service_id;
 936                 cm_id->service_mask = service_mask;
 937         }
 938         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
 939         spin_unlock_irqrestore(&cm.lock, flags);
 940
 941         if (cur_cm_id_priv) {
 942                 cm_id->state = IB_CM_IDLE;
 943                 kfree(cm_id_priv->compare_data);
 944                 cm_id_priv->compare_data = NULL;
 945                 ret = -EBUSY;
 946         }
 947         return ret;
 948 }
 949 EXPORT_SYMBOL(ib_cm_listen);
 950
 951 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
 952                           enum cm_msg_sequence msg_seq)
 953 {
 954         u64 hi_tid, low_tid;
 955
 956         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
 957         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
 958                           (msg_seq << 30));
 959         return cpu_to_be64(hi_tid | low_tid);
 960 }
 961
 962 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
 963                               __be16 attr_id, __be64 tid)
 964 {
 965         hdr->base_version  = IB_MGMT_BASE_VERSION;
 966         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
 967         hdr->class_version = IB_CM_CLASS_VERSION;
 968         hdr->method        = IB_MGMT_METHOD_SEND;
 969         hdr->attr_id       = attr_id;
 970         hdr->tid           = tid;
 971 }
 972
 973 static void cm_format_req(struct cm_req_msg *req_msg,
 974                           struct cm_id_private *cm_id_priv,
 975                           struct ib_cm_req_param *param)
 976 {
 977         struct ib_sa_path_rec *pri_path = param->primary_path;
 978         struct ib_sa_path_rec *alt_path = param->alternate_path;
 979
 980         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
 981                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
 982
 983         req_msg->local_comm_id = cm_id_priv->id.local_id;
 984         req_msg->service_id = param->service_id;
 985         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
 986         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
 987         cm_req_set_resp_res(req_msg, param->responder_resources);
 988         cm_req_set_init_depth(req_msg, param->initiator_depth);
 989         cm_req_set_remote_resp_timeout(req_msg,
 990                                        param->remote_cm_response_timeout);
 991         cm_req_set_qp_type(req_msg, param->qp_type);
 992         cm_req_set_flow_ctrl(req_msg, param->flow_control);
 993         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
 994         cm_req_set_local_resp_timeout(req_msg,
 995                                       param->local_cm_response_timeout);
 996         cm_req_set_retry_count(req_msg, param->retry_count);
 997         req_msg->pkey = param->primary_path->pkey;
 998         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
 999         cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1000         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1001         cm_req_set_srq(req_msg, param->srq);
1002
1003         if (pri_path->hop_limit <= 1) {
1004                 req_msg->primary_local_lid = pri_path->slid;
1005                 req_msg->primary_remote_lid = pri_path->dlid;
1006         } else {
1007                 /* Work-around until there's a way to obtain remote LID info */
1008                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1009                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1010         }
1011         req_msg->primary_local_gid = pri_path->sgid;
1012         req_msg->primary_remote_gid = pri_path->dgid;
1013         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1014         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1015         req_msg->primary_traffic_class = pri_path->traffic_class;
1016         req_msg->primary_hop_limit = pri_path->hop_limit;
1017         cm_req_set_primary_sl(req_msg, pri_path->sl);
1018         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1019         cm_req_set_primary_local_ack_timeout(req_msg,
1020                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1021                                pri_path->packet_life_time));
1022
1023         if (alt_path) {
1024                 if (alt_path->hop_limit <= 1) {
1025                         req_msg->alt_local_lid = alt_path->slid;
1026                         req_msg->alt_remote_lid = alt_path->dlid;
1027                 } else {
1028                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1029                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1030                 }
1031                 req_msg->alt_local_gid = alt_path->sgid;
1032                 req_msg->alt_remote_gid = alt_path->dgid;
1033                 cm_req_set_alt_flow_label(req_msg,
1034                                           alt_path->flow_label);
1035                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1036                 req_msg->alt_traffic_class = alt_path->traffic_class;
1037                 req_msg->alt_hop_limit = alt_path->hop_limit;
1038                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1039                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1040                 cm_req_set_alt_local_ack_timeout(req_msg,
1041                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1042                                        alt_path->packet_life_time));
1043         }
1044
1045         if (param->private_data && param->private_data_len)
1046                 memcpy(req_msg->private_data, param->private_data,
1047                        param->private_data_len);
1048 }
1049
1050 static int cm_validate_req_param(struct ib_cm_req_param *param)
1051 {
1052         /* peer-to-peer not supported */
1053         if (param->peer_to_peer)
1054                 return -EINVAL;
1055
1056         if (!param->primary_path)
1057                 return -EINVAL;
1058
1059         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC)
1060                 return -EINVAL;
1061
1062         if (param->private_data &&
1063             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1064                 return -EINVAL;
1065
1066         if (param->alternate_path &&
1067             (param->alternate_path->pkey != param->primary_path->pkey ||
1068              param->alternate_path->mtu != param->primary_path->mtu))
1069                 return -EINVAL;
1070
1071         return 0;
1072 }
1073
1074 int ib_send_cm_req(struct ib_cm_id *cm_id,
1075                    struct ib_cm_req_param *param)
1076 {
1077         struct cm_id_private *cm_id_priv;
1078         struct cm_req_msg *req_msg;
1079         unsigned long flags;
1080         int ret;
1081
1082         ret = cm_validate_req_param(param);
1083         if (ret)
1084                 return ret;
1085
1086         /* Verify that we're not in timewait. */
1087         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1088         spin_lock_irqsave(&cm_id_priv->lock, flags);
1089         if (cm_id->state != IB_CM_IDLE) {
1090                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1091                 ret = -EINVAL;
1092                 goto out;
1093         }
1094         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1095
1096         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1097                                                             id.local_id);
1098         if (IS_ERR(cm_id_priv->timewait_info)) {
1099                 ret = PTR_ERR(cm_id_priv->timewait_info);
1100                 goto out;
1101         }
1102
1103         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av);
1104         if (ret)
1105                 goto error1;
1106         if (param->alternate_path) {
1107                 ret = cm_init_av_by_path(param->alternate_path,
1108                                          &cm_id_priv->alt_av);
1109                 if (ret)
1110                         goto error1;
1111         }
1112         cm_id->service_id = param->service_id;
1113         cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
1114         cm_id_priv->timeout_ms = cm_convert_to_ms(
1115                                     param->primary_path->packet_life_time) * 2 +
1116                                  cm_convert_to_ms(
1117                                     param->remote_cm_response_timeout);
1118         cm_id_priv->max_cm_retries = param->max_cm_retries;
1119         cm_id_priv->initiator_depth = param->initiator_depth;
1120         cm_id_priv->responder_resources = param->responder_resources;
1121         cm_id_priv->retry_count = param->retry_count;
1122         cm_id_priv->path_mtu = param->primary_path->mtu;
1123         cm_id_priv->pkey = param->primary_path->pkey;
1124         cm_id_priv->qp_type = param->qp_type;
1125
1126         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1127         if (ret)
1128                 goto error1;
1129
1130         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1131         cm_format_req(req_msg, cm_id_priv, param);
1132         cm_id_priv->tid = req_msg->hdr.tid;
1133         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1134         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1135
1136         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1137         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1138
1139         spin_lock_irqsave(&cm_id_priv->lock, flags);
1140         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1141         if (ret) {
1142                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1143                 goto error2;
1144         }
1145         BUG_ON(cm_id->state != IB_CM_IDLE);
1146         cm_id->state = IB_CM_REQ_SENT;
1147         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1148         return 0;
1149
1150 error2: cm_free_msg(cm_id_priv->msg);
1151 error1: kfree(cm_id_priv->timewait_info);
1152 out:    return ret;
1153 }
1154 EXPORT_SYMBOL(ib_send_cm_req);
1155
1156 static int cm_issue_rej(struct cm_port *port,
1157                         struct ib_mad_recv_wc *mad_recv_wc,
1158                         enum ib_cm_rej_reason reason,
1159                         enum cm_msg_response msg_rejected,
1160                         void *ari, u8 ari_length)
1161 {
1162         struct ib_mad_send_buf *msg = NULL;
1163         struct cm_rej_msg *rej_msg, *rcv_msg;
1164         int ret;
1165
1166         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1167         if (ret)
1168                 return ret;
1169
1170         /* We just need common CM header information.  Cast to any message. */
1171         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1172         rej_msg = (struct cm_rej_msg *) msg->mad;
1173
1174         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1175         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1176         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1177         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1178         rej_msg->reason = cpu_to_be16(reason);
1179
1180         if (ari && ari_length) {
1181                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1182                 memcpy(rej_msg->ari, ari, ari_length);
1183         }
1184
1185         ret = ib_post_send_mad(msg, NULL);
1186         if (ret)
1187                 cm_free_msg(msg);
1188
1189         return ret;
1190 }
1191
1192 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1193                                     __be32 local_qpn, __be32 remote_qpn)
1194 {
1195         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1196                 ((local_ca_guid == remote_ca_guid) &&
1197                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1198 }
1199
1200 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1201                                             struct ib_sa_path_rec *primary_path,
1202                                             struct ib_sa_path_rec *alt_path)
1203 {
1204         memset(primary_path, 0, sizeof *primary_path);
1205         primary_path->dgid = req_msg->primary_local_gid;
1206         primary_path->sgid = req_msg->primary_remote_gid;
1207         primary_path->dlid = req_msg->primary_local_lid;
1208         primary_path->slid = req_msg->primary_remote_lid;
1209         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1210         primary_path->hop_limit = req_msg->primary_hop_limit;
1211         primary_path->traffic_class = req_msg->primary_traffic_class;
1212         primary_path->reversible = 1;
1213         primary_path->pkey = req_msg->pkey;
1214         primary_path->sl = cm_req_get_primary_sl(req_msg);
1215         primary_path->mtu_selector = IB_SA_EQ;
1216         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1217         primary_path->rate_selector = IB_SA_EQ;
1218         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1219         primary_path->packet_life_time_selector = IB_SA_EQ;
1220         primary_path->packet_life_time =
1221                 cm_req_get_primary_local_ack_timeout(req_msg);
1222         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1223
1224         if (req_msg->alt_local_lid) {
1225                 memset(alt_path, 0, sizeof *alt_path);
1226                 alt_path->dgid = req_msg->alt_local_gid;
1227                 alt_path->sgid = req_msg->alt_remote_gid;
1228                 alt_path->dlid = req_msg->alt_local_lid;
1229                 alt_path->slid = req_msg->alt_remote_lid;
1230                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1231                 alt_path->hop_limit = req_msg->alt_hop_limit;
1232                 alt_path->traffic_class = req_msg->alt_traffic_class;
1233                 alt_path->reversible = 1;
1234                 alt_path->pkey = req_msg->pkey;
1235                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1236                 alt_path->mtu_selector = IB_SA_EQ;
1237                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1238                 alt_path->rate_selector = IB_SA_EQ;
1239                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1240                 alt_path->packet_life_time_selector = IB_SA_EQ;
1241                 alt_path->packet_life_time =
1242                         cm_req_get_alt_local_ack_timeout(req_msg);
1243                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1244         }
1245 }
1246
1247 static void cm_format_req_event(struct cm_work *work,
1248                                 struct cm_id_private *cm_id_priv,
1249                                 struct ib_cm_id *listen_id)
1250 {
1251         struct cm_req_msg *req_msg;
1252         struct ib_cm_req_event_param *param;
1253
1254         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1255         param = &work->cm_event.param.req_rcvd;
1256         param->listen_id = listen_id;
1257         param->port = cm_id_priv->av.port->port_num;
1258         param->primary_path = &work->path[0];
1259         if (req_msg->alt_local_lid)
1260                 param->alternate_path = &work->path[1];
1261         else
1262                 param->alternate_path = NULL;
1263         param->remote_ca_guid = req_msg->local_ca_guid;
1264         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1265         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1266         param->qp_type = cm_req_get_qp_type(req_msg);
1267         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1268         param->responder_resources = cm_req_get_init_depth(req_msg);
1269         param->initiator_depth = cm_req_get_resp_res(req_msg);
1270         param->local_cm_response_timeout =
1271                                         cm_req_get_remote_resp_timeout(req_msg);
1272         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1273         param->remote_cm_response_timeout =
1274                                         cm_req_get_local_resp_timeout(req_msg);
1275         param->retry_count = cm_req_get_retry_count(req_msg);
1276         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1277         param->srq = cm_req_get_srq(req_msg);
1278         work->cm_event.private_data = &req_msg->private_data;
1279 }
1280
1281 static void cm_process_work(struct cm_id_private *cm_id_priv,
1282                             struct cm_work *work)
1283 {
1284         int ret;
1285
1286         /* We will typically only have the current event to report. */
1287         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1288         cm_free_work(work);
1289
1290         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1291                 spin_lock_irq(&cm_id_priv->lock);
1292                 work = cm_dequeue_work(cm_id_priv);
1293                 spin_unlock_irq(&cm_id_priv->lock);
1294                 BUG_ON(!work);
1295                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1296                                                 &work->cm_event);
1297                 cm_free_work(work);
1298         }
1299         cm_deref_id(cm_id_priv);
1300         if (ret)
1301                 cm_destroy_id(&cm_id_priv->id, ret);
1302 }
1303
1304 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1305                           struct cm_id_private *cm_id_priv,
1306                           enum cm_msg_response msg_mraed, u8 service_timeout,
1307                           const void *private_data, u8 private_data_len)
1308 {
1309         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1310         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1311         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1312         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1313         cm_mra_set_service_timeout(mra_msg, service_timeout);
1314
1315         if (private_data && private_data_len)
1316                 memcpy(mra_msg->private_data, private_data, private_data_len);
1317 }
1318
1319 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1320                           struct cm_id_private *cm_id_priv,
1321                           enum ib_cm_rej_reason reason,
1322                           void *ari,
1323                           u8 ari_length,
1324                           const void *private_data,
1325                           u8 private_data_len)
1326 {
1327         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1328         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1329
1330         switch(cm_id_priv->id.state) {
1331         case IB_CM_REQ_RCVD:
1332                 rej_msg->local_comm_id = 0;
1333                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1334                 break;
1335         case IB_CM_MRA_REQ_SENT:
1336                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1337                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1338                 break;
1339         case IB_CM_REP_RCVD:
1340         case IB_CM_MRA_REP_SENT:
1341                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1342                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1343                 break;
1344         default:
1345                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1346                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1347                 break;
1348         }
1349
1350         rej_msg->reason = cpu_to_be16(reason);
1351         if (ari && ari_length) {
1352                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1353                 memcpy(rej_msg->ari, ari, ari_length);
1354         }
1355
1356         if (private_data && private_data_len)
1357                 memcpy(rej_msg->private_data, private_data, private_data_len);
1358 }
1359
1360 static void cm_dup_req_handler(struct cm_work *work,
1361                                struct cm_id_private *cm_id_priv)
1362 {
1363         struct ib_mad_send_buf *msg = NULL;
1364         int ret;
1365
1366         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1367                         counter[CM_REQ_COUNTER]);
1368
1369         /* Quick state check to discard duplicate REQs. */
1370         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1371                 return;
1372
1373         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1374         if (ret)
1375                 return;
1376
1377         spin_lock_irq(&cm_id_priv->lock);
1378         switch (cm_id_priv->id.state) {
1379         case IB_CM_MRA_REQ_SENT:
1380                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1381                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1382                               cm_id_priv->private_data,
1383                               cm_id_priv->private_data_len);
1384                 break;
1385         case IB_CM_TIMEWAIT:
1386                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1387                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1388                 break;
1389         default:
1390                 goto unlock;
1391         }
1392         spin_unlock_irq(&cm_id_priv->lock);
1393
1394         ret = ib_post_send_mad(msg, NULL);
1395         if (ret)
1396                 goto free;
1397         return;
1398
1399 unlock: spin_unlock_irq(&cm_id_priv->lock);
1400 free:   cm_free_msg(msg);
1401 }
1402
1403 static struct cm_id_private * cm_match_req(struct cm_work *work,
1404                                            struct cm_id_private *cm_id_priv)
1405 {
1406         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1407         struct cm_timewait_info *timewait_info;
1408         struct cm_req_msg *req_msg;
1409
1410         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1411
1412         /* Check for possible duplicate REQ. */
1413         spin_lock_irq(&cm.lock);
1414         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1415         if (timewait_info) {
1416                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1417                                            timewait_info->work.remote_id);
1418                 spin_unlock_irq(&cm.lock);
1419                 if (cur_cm_id_priv) {
1420                         cm_dup_req_handler(work, cur_cm_id_priv);
1421                         cm_deref_id(cur_cm_id_priv);
1422                 }
1423                 return NULL;
1424         }
1425
1426         /* Check for stale connections. */
1427         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1428         if (timewait_info) {
1429                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1430                 spin_unlock_irq(&cm.lock);
1431                 cm_issue_rej(work->port, work->mad_recv_wc,
1432                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1433                              NULL, 0);
1434                 return NULL;
1435         }
1436
1437         /* Find matching listen request. */
1438         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1439                                            req_msg->service_id,
1440                                            req_msg->private_data);
1441         if (!listen_cm_id_priv) {
1442                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1443                 spin_unlock_irq(&cm.lock);
1444                 cm_issue_rej(work->port, work->mad_recv_wc,
1445                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1446                              NULL, 0);
1447                 goto out;
1448         }
1449         atomic_inc(&listen_cm_id_priv->refcount);
1450         atomic_inc(&cm_id_priv->refcount);
1451         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1452         atomic_inc(&cm_id_priv->work_count);
1453         spin_unlock_irq(&cm.lock);
1454 out:
1455         return listen_cm_id_priv;
1456 }
1457
1458 /*
1459  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1460  * we need to override the LID/SL data in the REQ with the LID information
1461  * in the work completion.
1462  */
1463 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1464 {
1465         if (!cm_req_get_primary_subnet_local(req_msg)) {
1466                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1467                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1468                         cm_req_set_primary_sl(req_msg, wc->sl);
1469                 }
1470
1471                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1472                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1473         }
1474
1475         if (!cm_req_get_alt_subnet_local(req_msg)) {
1476                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1477                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1478                         cm_req_set_alt_sl(req_msg, wc->sl);
1479                 }
1480
1481                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1482                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1483         }
1484 }
1485
1486 static int cm_req_handler(struct cm_work *work)
1487 {
1488         struct ib_cm_id *cm_id;
1489         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1490         struct cm_req_msg *req_msg;
1491         int ret;
1492
1493         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1494
1495         cm_id = ib_create_cm_id(work->port->cm_dev->device, NULL, NULL);
1496         if (IS_ERR(cm_id))
1497                 return PTR_ERR(cm_id);
1498
1499         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1500         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1501         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1502                                 work->mad_recv_wc->recv_buf.grh,
1503                                 &cm_id_priv->av);
1504         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1505                                                             id.local_id);
1506         if (IS_ERR(cm_id_priv->timewait_info)) {
1507                 ret = PTR_ERR(cm_id_priv->timewait_info);
1508                 goto destroy;
1509         }
1510         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1511         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1512         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1513
1514         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1515         if (!listen_cm_id_priv) {
1516                 ret = -EINVAL;
1517                 kfree(cm_id_priv->timewait_info);
1518                 goto destroy;
1519         }
1520
1521         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1522         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1523         cm_id_priv->id.service_id = req_msg->service_id;
1524         cm_id_priv->id.service_mask = __constant_cpu_to_be64(~0ULL);
1525
1526         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1527         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1528         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
1529         if (ret) {
1530                 ib_get_cached_gid(work->port->cm_dev->device,
1531                                   work->port->port_num, 0, &work->path[0].sgid);
1532                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1533                                &work->path[0].sgid, sizeof work->path[0].sgid,
1534                                NULL, 0);
1535                 goto rejected;
1536         }
1537         if (req_msg->alt_local_lid) {
1538                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
1539                 if (ret) {
1540                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1541                                        &work->path[0].sgid,
1542                                        sizeof work->path[0].sgid, NULL, 0);
1543                         goto rejected;
1544                 }
1545         }
1546         cm_id_priv->tid = req_msg->hdr.tid;
1547         cm_id_priv->timeout_ms = cm_convert_to_ms(
1548                                         cm_req_get_local_resp_timeout(req_msg));
1549         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1550         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1551         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1552         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1553         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1554         cm_id_priv->pkey = req_msg->pkey;
1555         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1556         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1557         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1558         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1559
1560         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1561         cm_process_work(cm_id_priv, work);
1562         cm_deref_id(listen_cm_id_priv);
1563         return 0;
1564
1565 rejected:
1566         atomic_dec(&cm_id_priv->refcount);
1567         cm_deref_id(listen_cm_id_priv);
1568 destroy:
1569         ib_destroy_cm_id(cm_id);
1570         return ret;
1571 }
1572
1573 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1574                           struct cm_id_private *cm_id_priv,
1575                           struct ib_cm_rep_param *param)
1576 {
1577         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1578         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1579         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1580         cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1581         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1582         rep_msg->resp_resources = param->responder_resources;
1583         rep_msg->initiator_depth = param->initiator_depth;
1584         cm_rep_set_target_ack_delay(rep_msg,
1585                                     cm_id_priv->av.port->cm_dev->ack_delay);
1586         cm_rep_set_failover(rep_msg, param->failover_accepted);
1587         cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1588         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1589         cm_rep_set_srq(rep_msg, param->srq);
1590         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1591
1592         if (param->private_data && param->private_data_len)
1593                 memcpy(rep_msg->private_data, param->private_data,
1594                        param->private_data_len);
1595 }
1596
1597 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1598                    struct ib_cm_rep_param *param)
1599 {
1600         struct cm_id_private *cm_id_priv;
1601         struct ib_mad_send_buf *msg;
1602         struct cm_rep_msg *rep_msg;
1603         unsigned long flags;
1604         int ret;
1605
1606         if (param->private_data &&
1607             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1608                 return -EINVAL;
1609
1610         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1611         spin_lock_irqsave(&cm_id_priv->lock, flags);
1612         if (cm_id->state != IB_CM_REQ_RCVD &&
1613             cm_id->state != IB_CM_MRA_REQ_SENT) {
1614                 ret = -EINVAL;
1615                 goto out;
1616         }
1617
1618         ret = cm_alloc_msg(cm_id_priv, &msg);
1619         if (ret)
1620                 goto out;
1621
1622         rep_msg = (struct cm_rep_msg *) msg->mad;
1623         cm_format_rep(rep_msg, cm_id_priv, param);
1624         msg->timeout_ms = cm_id_priv->timeout_ms;
1625         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1626
1627         ret = ib_post_send_mad(msg, NULL);
1628         if (ret) {
1629                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1630                 cm_free_msg(msg);
1631                 return ret;
1632         }
1633
1634         cm_id->state = IB_CM_REP_SENT;
1635         cm_id_priv->msg = msg;
1636         cm_id_priv->initiator_depth = param->initiator_depth;
1637         cm_id_priv->responder_resources = param->responder_resources;
1638         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1639         cm_id_priv->local_qpn = cm_rep_get_local_qpn(rep_msg);
1640
1641 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1642         return ret;
1643 }
1644 EXPORT_SYMBOL(ib_send_cm_rep);
1645
1646 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1647                           struct cm_id_private *cm_id_priv,
1648                           const void *private_data,
1649                           u8 private_data_len)
1650 {
1651         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1652         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1653         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1654
1655         if (private_data && private_data_len)
1656                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1657 }
1658
1659 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1660                    const void *private_data,
1661                    u8 private_data_len)
1662 {
1663         struct cm_id_private *cm_id_priv;
1664         struct ib_mad_send_buf *msg;
1665         unsigned long flags;
1666         void *data;
1667         int ret;
1668
1669         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1670                 return -EINVAL;
1671
1672         data = cm_copy_private_data(private_data, private_data_len);
1673         if (IS_ERR(data))
1674                 return PTR_ERR(data);
1675
1676         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1677         spin_lock_irqsave(&cm_id_priv->lock, flags);
1678         if (cm_id->state != IB_CM_REP_RCVD &&
1679             cm_id->state != IB_CM_MRA_REP_SENT) {
1680                 ret = -EINVAL;
1681                 goto error;
1682         }
1683
1684         ret = cm_alloc_msg(cm_id_priv, &msg);
1685         if (ret)
1686                 goto error;
1687
1688         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1689                       private_data, private_data_len);
1690
1691         ret = ib_post_send_mad(msg, NULL);
1692         if (ret) {
1693                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1694                 cm_free_msg(msg);
1695                 kfree(data);
1696                 return ret;
1697         }
1698
1699         cm_id->state = IB_CM_ESTABLISHED;
1700         cm_set_private_data(cm_id_priv, data, private_data_len);
1701         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1702         return 0;
1703
1704 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1705         kfree(data);
1706         return ret;
1707 }
1708 EXPORT_SYMBOL(ib_send_cm_rtu);
1709
1710 static void cm_format_rep_event(struct cm_work *work)
1711 {
1712         struct cm_rep_msg *rep_msg;
1713         struct ib_cm_rep_event_param *param;
1714
1715         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1716         param = &work->cm_event.param.rep_rcvd;
1717         param->remote_ca_guid = rep_msg->local_ca_guid;
1718         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1719         param->remote_qpn = be32_to_cpu(cm_rep_get_local_qpn(rep_msg));
1720         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1721         param->responder_resources = rep_msg->initiator_depth;
1722         param->initiator_depth = rep_msg->resp_resources;
1723         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1724         param->failover_accepted = cm_rep_get_failover(rep_msg);
1725         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1726         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1727         param->srq = cm_rep_get_srq(rep_msg);
1728         work->cm_event.private_data = &rep_msg->private_data;
1729 }
1730
1731 static void cm_dup_rep_handler(struct cm_work *work)
1732 {
1733         struct cm_id_private *cm_id_priv;
1734         struct cm_rep_msg *rep_msg;
1735         struct ib_mad_send_buf *msg = NULL;
1736         int ret;
1737
1738         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1739         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1740                                    rep_msg->local_comm_id);
1741         if (!cm_id_priv)
1742                 return;
1743
1744         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1745                         counter[CM_REP_COUNTER]);
1746         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1747         if (ret)
1748                 goto deref;
1749
1750         spin_lock_irq(&cm_id_priv->lock);
1751         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1752                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1753                               cm_id_priv->private_data,
1754                               cm_id_priv->private_data_len);
1755         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1756                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1757                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1758                               cm_id_priv->private_data,
1759                               cm_id_priv->private_data_len);
1760         else
1761                 goto unlock;
1762         spin_unlock_irq(&cm_id_priv->lock);
1763
1764         ret = ib_post_send_mad(msg, NULL);
1765         if (ret)
1766                 goto free;
1767         goto deref;
1768
1769 unlock: spin_unlock_irq(&cm_id_priv->lock);
1770 free:   cm_free_msg(msg);
1771 deref:  cm_deref_id(cm_id_priv);
1772 }
1773
1774 static int cm_rep_handler(struct cm_work *work)
1775 {
1776         struct cm_id_private *cm_id_priv;
1777         struct cm_rep_msg *rep_msg;
1778         int ret;
1779
1780         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1781         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1782         if (!cm_id_priv) {
1783                 cm_dup_rep_handler(work);
1784                 return -EINVAL;
1785         }
1786
1787         cm_format_rep_event(work);
1788
1789         spin_lock_irq(&cm_id_priv->lock);
1790         switch (cm_id_priv->id.state) {
1791         case IB_CM_REQ_SENT:
1792         case IB_CM_MRA_REQ_RCVD:
1793                 break;
1794         default:
1795                 spin_unlock_irq(&cm_id_priv->lock);
1796                 ret = -EINVAL;
1797                 goto error;
1798         }
1799
1800         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1801         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1802         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1803
1804         spin_lock(&cm.lock);
1805         /* Check for duplicate REP. */
1806         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1807                 spin_unlock(&cm.lock);
1808                 spin_unlock_irq(&cm_id_priv->lock);
1809                 ret = -EINVAL;
1810                 goto error;
1811         }
1812         /* Check for a stale connection. */
1813         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
1814                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
1815                          &cm.remote_id_table);
1816                 cm_id_priv->timewait_info->inserted_remote_id = 0;
1817                 spin_unlock(&cm.lock);
1818                 spin_unlock_irq(&cm_id_priv->lock);
1819                 cm_issue_rej(work->port, work->mad_recv_wc,
1820                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
1821                              NULL, 0);
1822                 ret = -EINVAL;
1823                 goto error;
1824         }
1825         spin_unlock(&cm.lock);
1826
1827         cm_id_priv->id.state = IB_CM_REP_RCVD;
1828         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
1829         cm_id_priv->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1830         cm_id_priv->initiator_depth = rep_msg->resp_resources;
1831         cm_id_priv->responder_resources = rep_msg->initiator_depth;
1832         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
1833         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1834         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1835         cm_id_priv->av.timeout =
1836                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1837                                        cm_id_priv->av.timeout - 1);
1838         cm_id_priv->alt_av.timeout =
1839                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1840                                        cm_id_priv->alt_av.timeout - 1);
1841
1842         /* todo: handle peer_to_peer */
1843
1844         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1845         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1846         if (!ret)
1847                 list_add_tail(&work->list, &cm_id_priv->work_list);
1848         spin_unlock_irq(&cm_id_priv->lock);
1849
1850         if (ret)
1851                 cm_process_work(cm_id_priv, work);
1852         else
1853                 cm_deref_id(cm_id_priv);
1854         return 0;
1855
1856 error:
1857         cm_deref_id(cm_id_priv);
1858         return ret;
1859 }
1860
1861 static int cm_establish_handler(struct cm_work *work)
1862 {
1863         struct cm_id_private *cm_id_priv;
1864         int ret;
1865
1866         /* See comment in cm_establish about lookup. */
1867         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
1868         if (!cm_id_priv)
1869                 return -EINVAL;
1870
1871         spin_lock_irq(&cm_id_priv->lock);
1872         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
1873                 spin_unlock_irq(&cm_id_priv->lock);
1874                 goto out;
1875         }
1876
1877         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1878         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1879         if (!ret)
1880                 list_add_tail(&work->list, &cm_id_priv->work_list);
1881         spin_unlock_irq(&cm_id_priv->lock);
1882
1883         if (ret)
1884                 cm_process_work(cm_id_priv, work);
1885         else
1886                 cm_deref_id(cm_id_priv);
1887         return 0;
1888 out:
1889         cm_deref_id(cm_id_priv);
1890         return -EINVAL;
1891 }
1892
1893 static int cm_rtu_handler(struct cm_work *work)
1894 {
1895         struct cm_id_private *cm_id_priv;
1896         struct cm_rtu_msg *rtu_msg;
1897         int ret;
1898
1899         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
1900         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
1901                                    rtu_msg->local_comm_id);
1902         if (!cm_id_priv)
1903                 return -EINVAL;
1904
1905         work->cm_event.private_data = &rtu_msg->private_data;
1906
1907         spin_lock_irq(&cm_id_priv->lock);
1908         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
1909             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
1910                 spin_unlock_irq(&cm_id_priv->lock);
1911                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1912                                 counter[CM_RTU_COUNTER]);
1913                 goto out;
1914         }
1915         cm_id_priv->id.state = IB_CM_ESTABLISHED;
1916
1917         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1918         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1919         if (!ret)
1920                 list_add_tail(&work->list, &cm_id_priv->work_list);
1921         spin_unlock_irq(&cm_id_priv->lock);
1922
1923         if (ret)
1924                 cm_process_work(cm_id_priv, work);
1925         else
1926                 cm_deref_id(cm_id_priv);
1927         return 0;
1928 out:
1929         cm_deref_id(cm_id_priv);
1930         return -EINVAL;
1931 }
1932
1933 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
1934                           struct cm_id_private *cm_id_priv,
1935                           const void *private_data,
1936                           u8 private_data_len)
1937 {
1938         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
1939                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
1940         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
1941         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
1942         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
1943
1944         if (private_data && private_data_len)
1945                 memcpy(dreq_msg->private_data, private_data, private_data_len);
1946 }
1947
1948 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
1949                     const void *private_data,
1950                     u8 private_data_len)
1951 {
1952         struct cm_id_private *cm_id_priv;
1953         struct ib_mad_send_buf *msg;
1954         unsigned long flags;
1955         int ret;
1956
1957         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
1958                 return -EINVAL;
1959
1960         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1961         spin_lock_irqsave(&cm_id_priv->lock, flags);
1962         if (cm_id->state != IB_CM_ESTABLISHED) {
1963                 ret = -EINVAL;
1964                 goto out;
1965         }
1966
1967         ret = cm_alloc_msg(cm_id_priv, &msg);
1968         if (ret) {
1969                 cm_enter_timewait(cm_id_priv);
1970                 goto out;
1971         }
1972
1973         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
1974                        private_data, private_data_len);
1975         msg->timeout_ms = cm_id_priv->timeout_ms;
1976         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
1977
1978         ret = ib_post_send_mad(msg, NULL);
1979         if (ret) {
1980                 cm_enter_timewait(cm_id_priv);
1981                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1982                 cm_free_msg(msg);
1983                 return ret;
1984         }
1985
1986         cm_id->state = IB_CM_DREQ_SENT;
1987         cm_id_priv->msg = msg;
1988 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1989         return ret;
1990 }
1991 EXPORT_SYMBOL(ib_send_cm_dreq);
1992
1993 static void cm_format_drep(struct cm_drep_msg *drep_msg,
1994                           struct cm_id_private *cm_id_priv,
1995                           const void *private_data,
1996                           u8 private_data_len)
1997 {
1998         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
1999         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2000         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2001
2002         if (private_data && private_data_len)
2003                 memcpy(drep_msg->private_data, private_data, private_data_len);
2004 }
2005
2006 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2007                     const void *private_data,
2008                     u8 private_data_len)
2009 {
2010         struct cm_id_private *cm_id_priv;
2011         struct ib_mad_send_buf *msg;
2012         unsigned long flags;
2013         void *data;
2014         int ret;
2015
2016         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2017                 return -EINVAL;
2018
2019         data = cm_copy_private_data(private_data, private_data_len);
2020         if (IS_ERR(data))
2021                 return PTR_ERR(data);
2022
2023         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2024         spin_lock_irqsave(&cm_id_priv->lock, flags);
2025         if (cm_id->state != IB_CM_DREQ_RCVD) {
2026                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2027                 kfree(data);
2028                 return -EINVAL;
2029         }
2030
2031         cm_set_private_data(cm_id_priv, data, private_data_len);
2032         cm_enter_timewait(cm_id_priv);
2033
2034         ret = cm_alloc_msg(cm_id_priv, &msg);
2035         if (ret)
2036                 goto out;
2037
2038         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2039                        private_data, private_data_len);
2040
2041         ret = ib_post_send_mad(msg, NULL);
2042         if (ret) {
2043                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2044                 cm_free_msg(msg);
2045                 return ret;
2046         }
2047
2048 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2049         return ret;
2050 }
2051 EXPORT_SYMBOL(ib_send_cm_drep);
2052
2053 static int cm_issue_drep(struct cm_port *port,
2054                          struct ib_mad_recv_wc *mad_recv_wc)
2055 {
2056         struct ib_mad_send_buf *msg = NULL;
2057         struct cm_dreq_msg *dreq_msg;
2058         struct cm_drep_msg *drep_msg;
2059         int ret;
2060
2061         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2062         if (ret)
2063                 return ret;
2064
2065         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2066         drep_msg = (struct cm_drep_msg *) msg->mad;
2067
2068         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2069         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2070         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2071
2072         ret = ib_post_send_mad(msg, NULL);
2073         if (ret)
2074                 cm_free_msg(msg);
2075
2076         return ret;
2077 }
2078
2079 static int cm_dreq_handler(struct cm_work *work)
2080 {
2081         struct cm_id_private *cm_id_priv;
2082         struct cm_dreq_msg *dreq_msg;
2083         struct ib_mad_send_buf *msg = NULL;
2084         int ret;
2085
2086         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2087         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2088                                    dreq_msg->local_comm_id);
2089         if (!cm_id_priv) {
2090                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2091                                 counter[CM_DREQ_COUNTER]);
2092                 cm_issue_drep(work->port, work->mad_recv_wc);
2093                 return -EINVAL;
2094         }
2095
2096         work->cm_event.private_data = &dreq_msg->private_data;
2097
2098         spin_lock_irq(&cm_id_priv->lock);
2099         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2100                 goto unlock;
2101
2102         switch (cm_id_priv->id.state) {
2103         case IB_CM_REP_SENT:
2104         case IB_CM_DREQ_SENT:
2105                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2106                 break;
2107         case IB_CM_ESTABLISHED:
2108         case IB_CM_MRA_REP_RCVD:
2109                 break;
2110         case IB_CM_TIMEWAIT:
2111                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2112                                 counter[CM_DREQ_COUNTER]);
2113                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2114                         goto unlock;
2115
2116                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2117                                cm_id_priv->private_data,
2118                                cm_id_priv->private_data_len);
2119                 spin_unlock_irq(&cm_id_priv->lock);
2120
2121                 if (ib_post_send_mad(msg, NULL))
2122                         cm_free_msg(msg);
2123                 goto deref;
2124         case IB_CM_DREQ_RCVD:
2125                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2126                                 counter[CM_DREQ_COUNTER]);
2127                 goto unlock;
2128         default:
2129                 goto unlock;
2130         }
2131         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2132         cm_id_priv->tid = dreq_msg->hdr.tid;
2133         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2134         if (!ret)
2135                 list_add_tail(&work->list, &cm_id_priv->work_list);
2136         spin_unlock_irq(&cm_id_priv->lock);
2137
2138         if (ret)
2139                 cm_process_work(cm_id_priv, work);
2140         else
2141                 cm_deref_id(cm_id_priv);
2142         return 0;
2143
2144 unlock: spin_unlock_irq(&cm_id_priv->lock);
2145 deref:  cm_deref_id(cm_id_priv);
2146         return -EINVAL;
2147 }
2148
2149 static int cm_drep_handler(struct cm_work *work)
2150 {
2151         struct cm_id_private *cm_id_priv;
2152         struct cm_drep_msg *drep_msg;
2153         int ret;
2154
2155         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2156         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2157                                    drep_msg->local_comm_id);
2158         if (!cm_id_priv)
2159                 return -EINVAL;
2160
2161         work->cm_event.private_data = &drep_msg->private_data;
2162
2163         spin_lock_irq(&cm_id_priv->lock);
2164         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2165             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2166                 spin_unlock_irq(&cm_id_priv->lock);
2167                 goto out;
2168         }
2169         cm_enter_timewait(cm_id_priv);
2170
2171         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2172         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2173         if (!ret)
2174                 list_add_tail(&work->list, &cm_id_priv->work_list);
2175         spin_unlock_irq(&cm_id_priv->lock);
2176
2177         if (ret)
2178                 cm_process_work(cm_id_priv, work);
2179         else
2180                 cm_deref_id(cm_id_priv);
2181         return 0;
2182 out:
2183         cm_deref_id(cm_id_priv);
2184         return -EINVAL;
2185 }
2186
2187 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2188                    enum ib_cm_rej_reason reason,
2189                    void *ari,
2190                    u8 ari_length,
2191                    const void *private_data,
2192                    u8 private_data_len)
2193 {
2194         struct cm_id_private *cm_id_priv;
2195         struct ib_mad_send_buf *msg;
2196         unsigned long flags;
2197         int ret;
2198
2199         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2200             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2201                 return -EINVAL;
2202
2203         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2204
2205         spin_lock_irqsave(&cm_id_priv->lock, flags);
2206         switch (cm_id->state) {
2207         case IB_CM_REQ_SENT:
2208         case IB_CM_MRA_REQ_RCVD:
2209         case IB_CM_REQ_RCVD:
2210         case IB_CM_MRA_REQ_SENT:
2211         case IB_CM_REP_RCVD:
2212         case IB_CM_MRA_REP_SENT:
2213                 ret = cm_alloc_msg(cm_id_priv, &msg);
2214                 if (!ret)
2215                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2216                                       cm_id_priv, reason, ari, ari_length,
2217                                       private_data, private_data_len);
2218
2219                 cm_reset_to_idle(cm_id_priv);
2220                 break;
2221         case IB_CM_REP_SENT:
2222         case IB_CM_MRA_REP_RCVD:
2223                 ret = cm_alloc_msg(cm_id_priv, &msg);
2224                 if (!ret)
2225                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2226                                       cm_id_priv, reason, ari, ari_length,
2227                                       private_data, private_data_len);
2228
2229                 cm_enter_timewait(cm_id_priv);
2230                 break;
2231         default:
2232                 ret = -EINVAL;
2233                 goto out;
2234         }
2235
2236         if (ret)
2237                 goto out;
2238
2239         ret = ib_post_send_mad(msg, NULL);
2240         if (ret)
2241                 cm_free_msg(msg);
2242
2243 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2244         return ret;
2245 }
2246 EXPORT_SYMBOL(ib_send_cm_rej);
2247
2248 static void cm_format_rej_event(struct cm_work *work)
2249 {
2250         struct cm_rej_msg *rej_msg;
2251         struct ib_cm_rej_event_param *param;
2252
2253         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2254         param = &work->cm_event.param.rej_rcvd;
2255         param->ari = rej_msg->ari;
2256         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2257         param->reason = __be16_to_cpu(rej_msg->reason);
2258         work->cm_event.private_data = &rej_msg->private_data;
2259 }
2260
2261 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2262 {
2263         struct cm_timewait_info *timewait_info;
2264         struct cm_id_private *cm_id_priv;
2265         __be32 remote_id;
2266
2267         remote_id = rej_msg->local_comm_id;
2268
2269         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2270                 spin_lock_irq(&cm.lock);
2271                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2272                                                   remote_id);
2273                 if (!timewait_info) {
2274                         spin_unlock_irq(&cm.lock);
2275                         return NULL;
2276                 }
2277                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2278                                       (timewait_info->work.local_id ^
2279                                        cm.random_id_operand));
2280                 if (cm_id_priv) {
2281                         if (cm_id_priv->id.remote_id == remote_id)
2282                                 atomic_inc(&cm_id_priv->refcount);
2283                         else
2284                                 cm_id_priv = NULL;
2285                 }
2286                 spin_unlock_irq(&cm.lock);
2287         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2288                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2289         else
2290                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2291
2292         return cm_id_priv;
2293 }
2294
2295 static int cm_rej_handler(struct cm_work *work)
2296 {
2297         struct cm_id_private *cm_id_priv;
2298         struct cm_rej_msg *rej_msg;
2299         int ret;
2300
2301         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2302         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2303         if (!cm_id_priv)
2304                 return -EINVAL;
2305
2306         cm_format_rej_event(work);
2307
2308         spin_lock_irq(&cm_id_priv->lock);
2309         switch (cm_id_priv->id.state) {
2310         case IB_CM_REQ_SENT:
2311         case IB_CM_MRA_REQ_RCVD:
2312         case IB_CM_REP_SENT:
2313         case IB_CM_MRA_REP_RCVD:
2314                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2315                 /* fall through */
2316         case IB_CM_REQ_RCVD:
2317         case IB_CM_MRA_REQ_SENT:
2318                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2319                         cm_enter_timewait(cm_id_priv);
2320                 else
2321                         cm_reset_to_idle(cm_id_priv);
2322                 break;
2323         case IB_CM_DREQ_SENT:
2324                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2325                 /* fall through */
2326         case IB_CM_REP_RCVD:
2327         case IB_CM_MRA_REP_SENT:
2328         case IB_CM_ESTABLISHED:
2329                 cm_enter_timewait(cm_id_priv);
2330                 break;
2331         default:
2332                 spin_unlock_irq(&cm_id_priv->lock);
2333                 ret = -EINVAL;
2334                 goto out;
2335         }
2336
2337         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2338         if (!ret)
2339                 list_add_tail(&work->list, &cm_id_priv->work_list);
2340         spin_unlock_irq(&cm_id_priv->lock);
2341
2342         if (ret)
2343                 cm_process_work(cm_id_priv, work);
2344         else
2345                 cm_deref_id(cm_id_priv);
2346         return 0;
2347 out:
2348         cm_deref_id(cm_id_priv);
2349         return -EINVAL;
2350 }
2351
2352 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2353                    u8 service_timeout,
2354                    const void *private_data,
2355                    u8 private_data_len)
2356 {
2357         struct cm_id_private *cm_id_priv;
2358         struct ib_mad_send_buf *msg;
2359         enum ib_cm_state cm_state;
2360         enum ib_cm_lap_state lap_state;
2361         enum cm_msg_response msg_response;
2362         void *data;
2363         unsigned long flags;
2364         int ret;
2365
2366         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2367                 return -EINVAL;
2368
2369         data = cm_copy_private_data(private_data, private_data_len);
2370         if (IS_ERR(data))
2371                 return PTR_ERR(data);
2372
2373         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2374
2375         spin_lock_irqsave(&cm_id_priv->lock, flags);
2376         switch(cm_id_priv->id.state) {
2377         case IB_CM_REQ_RCVD:
2378                 cm_state = IB_CM_MRA_REQ_SENT;
2379                 lap_state = cm_id->lap_state;
2380                 msg_response = CM_MSG_RESPONSE_REQ;
2381                 break;
2382         case IB_CM_REP_RCVD:
2383                 cm_state = IB_CM_MRA_REP_SENT;
2384                 lap_state = cm_id->lap_state;
2385                 msg_response = CM_MSG_RESPONSE_REP;
2386                 break;
2387         case IB_CM_ESTABLISHED:
2388                 cm_state = cm_id->state;
2389                 lap_state = IB_CM_MRA_LAP_SENT;
2390                 msg_response = CM_MSG_RESPONSE_OTHER;
2391                 break;
2392         default:
2393                 ret = -EINVAL;
2394                 goto error1;
2395         }
2396
2397         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2398                 ret = cm_alloc_msg(cm_id_priv, &msg);
2399                 if (ret)
2400                         goto error1;
2401
2402                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2403                               msg_response, service_timeout,
2404                               private_data, private_data_len);
2405                 ret = ib_post_send_mad(msg, NULL);
2406                 if (ret)
2407                         goto error2;
2408         }
2409
2410         cm_id->state = cm_state;
2411         cm_id->lap_state = lap_state;
2412         cm_id_priv->service_timeout = service_timeout;
2413         cm_set_private_data(cm_id_priv, data, private_data_len);
2414         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2415         return 0;
2416
2417 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2418         kfree(data);
2419         return ret;
2420
2421 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2422         kfree(data);
2423         cm_free_msg(msg);
2424         return ret;
2425 }
2426 EXPORT_SYMBOL(ib_send_cm_mra);
2427
2428 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2429 {
2430         switch (cm_mra_get_msg_mraed(mra_msg)) {
2431         case CM_MSG_RESPONSE_REQ:
2432                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2433         case CM_MSG_RESPONSE_REP:
2434         case CM_MSG_RESPONSE_OTHER:
2435                 return cm_acquire_id(mra_msg->remote_comm_id,
2436                                      mra_msg->local_comm_id);
2437         default:
2438                 return NULL;
2439         }
2440 }
2441
2442 static int cm_mra_handler(struct cm_work *work)
2443 {
2444         struct cm_id_private *cm_id_priv;
2445         struct cm_mra_msg *mra_msg;
2446         int timeout, ret;
2447
2448         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2449         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2450         if (!cm_id_priv)
2451                 return -EINVAL;
2452
2453         work->cm_event.private_data = &mra_msg->private_data;
2454         work->cm_event.param.mra_rcvd.service_timeout =
2455                                         cm_mra_get_service_timeout(mra_msg);
2456         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2457                   cm_convert_to_ms(cm_id_priv->av.timeout);
2458
2459         spin_lock_irq(&cm_id_priv->lock);
2460         switch (cm_id_priv->id.state) {
2461         case IB_CM_REQ_SENT:
2462                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2463                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2464                                   cm_id_priv->msg, timeout))
2465                         goto out;
2466                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2467                 break;
2468         case IB_CM_REP_SENT:
2469                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2470                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2471                                   cm_id_priv->msg, timeout))
2472                         goto out;
2473                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2474                 break;
2475         case IB_CM_ESTABLISHED:
2476                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2477                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2478                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2479                                   cm_id_priv->msg, timeout)) {
2480                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2481                                 atomic_long_inc(&work->port->
2482                                                 counter_group[CM_RECV_DUPLICATES].
2483                                                 counter[CM_MRA_COUNTER]);
2484                         goto out;
2485                 }
2486                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2487                 break;
2488         case IB_CM_MRA_REQ_RCVD:
2489         case IB_CM_MRA_REP_RCVD:
2490                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2491                                 counter[CM_MRA_COUNTER]);
2492                 /* fall through */
2493         default:
2494                 goto out;
2495         }
2496
2497         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2498                                       cm_id_priv->id.state;
2499         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2500         if (!ret)
2501                 list_add_tail(&work->list, &cm_id_priv->work_list);
2502         spin_unlock_irq(&cm_id_priv->lock);
2503
2504         if (ret)
2505                 cm_process_work(cm_id_priv, work);
2506         else
2507                 cm_deref_id(cm_id_priv);
2508         return 0;
2509 out:
2510         spin_unlock_irq(&cm_id_priv->lock);
2511         cm_deref_id(cm_id_priv);
2512         return -EINVAL;
2513 }
2514
2515 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2516                           struct cm_id_private *cm_id_priv,
2517                           struct ib_sa_path_rec *alternate_path,
2518                           const void *private_data,
2519                           u8 private_data_len)
2520 {
2521         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2522                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2523         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2524         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2525         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2526         /* todo: need remote CM response timeout */
2527         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2528         lap_msg->alt_local_lid = alternate_path->slid;
2529         lap_msg->alt_remote_lid = alternate_path->dlid;
2530         lap_msg->alt_local_gid = alternate_path->sgid;
2531         lap_msg->alt_remote_gid = alternate_path->dgid;
2532         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2533         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2534         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2535         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2536         cm_lap_set_sl(lap_msg, alternate_path->sl);
2537         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2538         cm_lap_set_local_ack_timeout(lap_msg,
2539                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2540                                alternate_path->packet_life_time));
2541
2542         if (private_data && private_data_len)
2543                 memcpy(lap_msg->private_data, private_data, private_data_len);
2544 }
2545
2546 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2547                    struct ib_sa_path_rec *alternate_path,
2548                    const void *private_data,
2549                    u8 private_data_len)
2550 {
2551         struct cm_id_private *cm_id_priv;
2552         struct ib_mad_send_buf *msg;
2553         unsigned long flags;
2554         int ret;
2555
2556         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2557                 return -EINVAL;
2558
2559         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2560         spin_lock_irqsave(&cm_id_priv->lock, flags);
2561         if (cm_id->state != IB_CM_ESTABLISHED ||
2562             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2563              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2564                 ret = -EINVAL;
2565                 goto out;
2566         }
2567
2568         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
2569         if (ret)
2570                 goto out;
2571         cm_id_priv->alt_av.timeout =
2572                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2573                                        cm_id_priv->alt_av.timeout - 1);
2574
2575         ret = cm_alloc_msg(cm_id_priv, &msg);
2576         if (ret)
2577                 goto out;
2578
2579         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2580                       alternate_path, private_data, private_data_len);
2581         msg->timeout_ms = cm_id_priv->timeout_ms;
2582         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2583
2584         ret = ib_post_send_mad(msg, NULL);
2585         if (ret) {
2586                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2587                 cm_free_msg(msg);
2588                 return ret;
2589         }
2590
2591         cm_id->lap_state = IB_CM_LAP_SENT;
2592         cm_id_priv->msg = msg;
2593
2594 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2595         return ret;
2596 }
2597 EXPORT_SYMBOL(ib_send_cm_lap);
2598
2599 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2600                                     struct ib_sa_path_rec *path,
2601                                     struct cm_lap_msg *lap_msg)
2602 {
2603         memset(path, 0, sizeof *path);
2604         path->dgid = lap_msg->alt_local_gid;
2605         path->sgid = lap_msg->alt_remote_gid;
2606         path->dlid = lap_msg->alt_local_lid;
2607         path->slid = lap_msg->alt_remote_lid;
2608         path->flow_label = cm_lap_get_flow_label(lap_msg);
2609         path->hop_limit = lap_msg->alt_hop_limit;
2610         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2611         path->reversible = 1;
2612         path->pkey = cm_id_priv->pkey;
2613         path->sl = cm_lap_get_sl(lap_msg);
2614         path->mtu_selector = IB_SA_EQ;
2615         path->mtu = cm_id_priv->path_mtu;
2616         path->rate_selector = IB_SA_EQ;
2617         path->rate = cm_lap_get_packet_rate(lap_msg);
2618         path->packet_life_time_selector = IB_SA_EQ;
2619         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2620         path->packet_life_time -= (path->packet_life_time > 0);
2621 }
2622
2623 static int cm_lap_handler(struct cm_work *work)
2624 {
2625         struct cm_id_private *cm_id_priv;
2626         struct cm_lap_msg *lap_msg;
2627         struct ib_cm_lap_event_param *param;
2628         struct ib_mad_send_buf *msg = NULL;
2629         int ret;
2630
2631         /* todo: verify LAP request and send reject APR if invalid. */
2632         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2633         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2634                                    lap_msg->local_comm_id);
2635         if (!cm_id_priv)
2636                 return -EINVAL;
2637
2638         param = &work->cm_event.param.lap_rcvd;
2639         param->alternate_path = &work->path[0];
2640         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2641         work->cm_event.private_data = &lap_msg->private_data;
2642
2643         spin_lock_irq(&cm_id_priv->lock);
2644         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2645                 goto unlock;
2646
2647         switch (cm_id_priv->id.lap_state) {
2648         case IB_CM_LAP_UNINIT:
2649         case IB_CM_LAP_IDLE:
2650                 break;
2651         case IB_CM_MRA_LAP_SENT:
2652                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2653                                 counter[CM_LAP_COUNTER]);
2654                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2655                         goto unlock;
2656
2657                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2658                               CM_MSG_RESPONSE_OTHER,
2659                               cm_id_priv->service_timeout,
2660                               cm_id_priv->private_data,
2661                               cm_id_priv->private_data_len);
2662                 spin_unlock_irq(&cm_id_priv->lock);
2663
2664                 if (ib_post_send_mad(msg, NULL))
2665                         cm_free_msg(msg);
2666                 goto deref;
2667         case IB_CM_LAP_RCVD:
2668                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2669                                 counter[CM_LAP_COUNTER]);
2670                 goto unlock;
2671         default:
2672                 goto unlock;
2673         }
2674
2675         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2676         cm_id_priv->tid = lap_msg->hdr.tid;
2677         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2678                                 work->mad_recv_wc->recv_buf.grh,
2679                                 &cm_id_priv->av);
2680         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av);
2681         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2682         if (!ret)
2683                 list_add_tail(&work->list, &cm_id_priv->work_list);
2684         spin_unlock_irq(&cm_id_priv->lock);
2685
2686         if (ret)
2687                 cm_process_work(cm_id_priv, work);
2688         else
2689                 cm_deref_id(cm_id_priv);
2690         return 0;
2691
2692 unlock: spin_unlock_irq(&cm_id_priv->lock);
2693 deref:  cm_deref_id(cm_id_priv);
2694         return -EINVAL;
2695 }
2696
2697 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2698                           struct cm_id_private *cm_id_priv,
2699                           enum ib_cm_apr_status status,
2700                           void *info,
2701                           u8 info_length,
2702                           const void *private_data,
2703                           u8 private_data_len)
2704 {
2705         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2706         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2707         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2708         apr_msg->ap_status = (u8) status;
2709
2710         if (info && info_length) {
2711                 apr_msg->info_length = info_length;
2712                 memcpy(apr_msg->info, info, info_length);
2713         }
2714
2715         if (private_data && private_data_len)
2716                 memcpy(apr_msg->private_data, private_data, private_data_len);
2717 }
2718
2719 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2720                    enum ib_cm_apr_status status,
2721                    void *info,
2722                    u8 info_length,
2723                    const void *private_data,
2724                    u8 private_data_len)
2725 {
2726         struct cm_id_private *cm_id_priv;
2727         struct ib_mad_send_buf *msg;
2728         unsigned long flags;
2729         int ret;
2730
2731         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2732             (info && info_length > IB_CM_APR_INFO_LENGTH))
2733                 return -EINVAL;
2734
2735         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2736         spin_lock_irqsave(&cm_id_priv->lock, flags);
2737         if (cm_id->state != IB_CM_ESTABLISHED ||
2738             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2739              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2740                 ret = -EINVAL;
2741                 goto out;
2742         }
2743
2744         ret = cm_alloc_msg(cm_id_priv, &msg);
2745         if (ret)
2746                 goto out;
2747
2748         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2749                       info, info_length, private_data, private_data_len);
2750         ret = ib_post_send_mad(msg, NULL);
2751         if (ret) {
2752                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2753                 cm_free_msg(msg);
2754                 return ret;
2755         }
2756
2757         cm_id->lap_state = IB_CM_LAP_IDLE;
2758 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2759         return ret;
2760 }
2761 EXPORT_SYMBOL(ib_send_cm_apr);
2762
2763 static int cm_apr_handler(struct cm_work *work)
2764 {
2765         struct cm_id_private *cm_id_priv;
2766         struct cm_apr_msg *apr_msg;
2767         int ret;
2768
2769         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2770         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2771                                    apr_msg->local_comm_id);
2772         if (!cm_id_priv)
2773                 return -EINVAL; /* Unmatched reply. */
2774
2775         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2776         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2777         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2778         work->cm_event.private_data = &apr_msg->private_data;
2779
2780         spin_lock_irq(&cm_id_priv->lock);
2781         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2782             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2783              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2784                 spin_unlock_irq(&cm_id_priv->lock);
2785                 goto out;
2786         }
2787         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2788         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2789         cm_id_priv->msg = NULL;
2790
2791         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2792         if (!ret)
2793                 list_add_tail(&work->list, &cm_id_priv->work_list);
2794         spin_unlock_irq(&cm_id_priv->lock);
2795
2796         if (ret)
2797                 cm_process_work(cm_id_priv, work);
2798         else
2799                 cm_deref_id(cm_id_priv);
2800         return 0;
2801 out:
2802         cm_deref_id(cm_id_priv);
2803         return -EINVAL;
2804 }
2805
2806 static int cm_timewait_handler(struct cm_work *work)
2807 {
2808         struct cm_timewait_info *timewait_info;
2809         struct cm_id_private *cm_id_priv;
2810         int ret;
2811
2812         timewait_info = (struct cm_timewait_info *)work;
2813         spin_lock_irq(&cm.lock);
2814         list_del(&timewait_info->list);
2815         spin_unlock_irq(&cm.lock);
2816
2817         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2818                                    timewait_info->work.remote_id);
2819         if (!cm_id_priv)
2820                 return -EINVAL;
2821
2822         spin_lock_irq(&cm_id_priv->lock);
2823         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2824             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2825                 spin_unlock_irq(&cm_id_priv->lock);
2826                 goto out;
2827         }
2828         cm_id_priv->id.state = IB_CM_IDLE;
2829         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2830         if (!ret)
2831                 list_add_tail(&work->list, &cm_id_priv->work_list);
2832         spin_unlock_irq(&cm_id_priv->lock);
2833
2834         if (ret)
2835                 cm_process_work(cm_id_priv, work);
2836         else
2837                 cm_deref_id(cm_id_priv);
2838         return 0;
2839 out:
2840         cm_deref_id(cm_id_priv);
2841         return -EINVAL;
2842 }
2843
2844 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
2845                                struct cm_id_private *cm_id_priv,
2846                                struct ib_cm_sidr_req_param *param)
2847 {
2848         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
2849                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
2850         sidr_req_msg->request_id = cm_id_priv->id.local_id;
2851         sidr_req_msg->pkey = cpu_to_be16(param->path->pkey);
2852         sidr_req_msg->service_id = param->service_id;
2853
2854         if (param->private_data && param->private_data_len)
2855                 memcpy(sidr_req_msg->private_data, param->private_data,
2856                        param->private_data_len);
2857 }
2858
2859 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
2860                         struct ib_cm_sidr_req_param *param)
2861 {
2862         struct cm_id_private *cm_id_priv;
2863         struct ib_mad_send_buf *msg;
2864         unsigned long flags;
2865         int ret;
2866
2867         if (!param->path || (param->private_data &&
2868              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
2869                 return -EINVAL;
2870
2871         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2872         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
2873         if (ret)
2874                 goto out;
2875
2876         cm_id->service_id = param->service_id;
2877         cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
2878         cm_id_priv->timeout_ms = param->timeout_ms;
2879         cm_id_priv->max_cm_retries = param->max_cm_retries;
2880         ret = cm_alloc_msg(cm_id_priv, &msg);
2881         if (ret)
2882                 goto out;
2883
2884         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
2885                            param);
2886         msg->timeout_ms = cm_id_priv->timeout_ms;
2887         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
2888
2889         spin_lock_irqsave(&cm_id_priv->lock, flags);
2890         if (cm_id->state == IB_CM_IDLE)
2891                 ret = ib_post_send_mad(msg, NULL);
2892         else
2893                 ret = -EINVAL;
2894
2895         if (ret) {
2896                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2897                 cm_free_msg(msg);
2898                 goto out;
2899         }
2900         cm_id->state = IB_CM_SIDR_REQ_SENT;
2901         cm_id_priv->msg = msg;
2902         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2903 out:
2904         return ret;
2905 }
2906 EXPORT_SYMBOL(ib_send_cm_sidr_req);
2907
2908 static void cm_format_sidr_req_event(struct cm_work *work,
2909                                      struct ib_cm_id *listen_id)
2910 {
2911         struct cm_sidr_req_msg *sidr_req_msg;
2912         struct ib_cm_sidr_req_event_param *param;
2913
2914         sidr_req_msg = (struct cm_sidr_req_msg *)
2915                                 work->mad_recv_wc->recv_buf.mad;
2916         param = &work->cm_event.param.sidr_req_rcvd;
2917         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
2918         param->listen_id = listen_id;
2919         param->port = work->port->port_num;
2920         work->cm_event.private_data = &sidr_req_msg->private_data;
2921 }
2922
2923 static int cm_sidr_req_handler(struct cm_work *work)
2924 {
2925         struct ib_cm_id *cm_id;
2926         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
2927         struct cm_sidr_req_msg *sidr_req_msg;
2928         struct ib_wc *wc;
2929
2930         cm_id = ib_create_cm_id(work->port->cm_dev->device, NULL, NULL);
2931         if (IS_ERR(cm_id))
2932                 return PTR_ERR(cm_id);
2933         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2934
2935         /* Record SGID/SLID and request ID for lookup. */
2936         sidr_req_msg = (struct cm_sidr_req_msg *)
2937                                 work->mad_recv_wc->recv_buf.mad;
2938         wc = work->mad_recv_wc->wc;
2939         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
2940         cm_id_priv->av.dgid.global.interface_id = 0;
2941         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2942                                 work->mad_recv_wc->recv_buf.grh,
2943                                 &cm_id_priv->av);
2944         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
2945         cm_id_priv->tid = sidr_req_msg->hdr.tid;
2946         atomic_inc(&cm_id_priv->work_count);
2947
2948         spin_lock_irq(&cm.lock);
2949         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
2950         if (cur_cm_id_priv) {
2951                 spin_unlock_irq(&cm.lock);
2952                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2953                                 counter[CM_SIDR_REQ_COUNTER]);
2954                 goto out; /* Duplicate message. */
2955         }
2956         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
2957         cur_cm_id_priv = cm_find_listen(cm_id->device,
2958                                         sidr_req_msg->service_id,
2959                                         sidr_req_msg->private_data);
2960         if (!cur_cm_id_priv) {
2961                 spin_unlock_irq(&cm.lock);
2962                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
2963                 goto out; /* No match. */
2964         }
2965         atomic_inc(&cur_cm_id_priv->refcount);
2966         spin_unlock_irq(&cm.lock);
2967
2968         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
2969         cm_id_priv->id.context = cur_cm_id_priv->id.context;
2970         cm_id_priv->id.service_id = sidr_req_msg->service_id;
2971         cm_id_priv->id.service_mask = __constant_cpu_to_be64(~0ULL);
2972
2973         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
2974         cm_process_work(cm_id_priv, work);
2975         cm_deref_id(cur_cm_id_priv);
2976         return 0;
2977 out:
2978         ib_destroy_cm_id(&cm_id_priv->id);
2979         return -EINVAL;
2980 }
2981
2982 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
2983                                struct cm_id_private *cm_id_priv,
2984                                struct ib_cm_sidr_rep_param *param)
2985 {
2986         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
2987                           cm_id_priv->tid);
2988         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
2989         sidr_rep_msg->status = param->status;
2990         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
2991         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
2992         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
2993
2994         if (param->info && param->info_length)
2995                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
2996
2997         if (param->private_data && param->private_data_len)
2998                 memcpy(sidr_rep_msg->private_data, param->private_data,
2999                        param->private_data_len);
3000 }
3001
3002 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3003                         struct ib_cm_sidr_rep_param *param)
3004 {
3005         struct cm_id_private *cm_id_priv;
3006         struct ib_mad_send_buf *msg;
3007         unsigned long flags;
3008         int ret;
3009
3010         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3011             (param->private_data &&
3012              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3013                 return -EINVAL;
3014
3015         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3016         spin_lock_irqsave(&cm_id_priv->lock, flags);
3017         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3018                 ret = -EINVAL;
3019                 goto error;
3020         }
3021
3022         ret = cm_alloc_msg(cm_id_priv, &msg);
3023         if (ret)
3024                 goto error;
3025
3026         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3027                            param);
3028         ret = ib_post_send_mad(msg, NULL);
3029         if (ret) {
3030                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3031                 cm_free_msg(msg);
3032                 return ret;
3033         }
3034         cm_id->state = IB_CM_IDLE;
3035         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3036
3037         spin_lock_irqsave(&cm.lock, flags);
3038         rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3039         spin_unlock_irqrestore(&cm.lock, flags);
3040         return 0;
3041
3042 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3043         return ret;
3044 }
3045 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3046
3047 static void cm_format_sidr_rep_event(struct cm_work *work)
3048 {
3049         struct cm_sidr_rep_msg *sidr_rep_msg;
3050         struct ib_cm_sidr_rep_event_param *param;
3051
3052         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3053                                 work->mad_recv_wc->recv_buf.mad;
3054         param = &work->cm_event.param.sidr_rep_rcvd;
3055         param->status = sidr_rep_msg->status;
3056         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3057         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3058         param->info = &sidr_rep_msg->info;
3059         param->info_len = sidr_rep_msg->info_length;
3060         work->cm_event.private_data = &sidr_rep_msg->private_data;
3061 }
3062
3063 static int cm_sidr_rep_handler(struct cm_work *work)
3064 {
3065         struct cm_sidr_rep_msg *sidr_rep_msg;
3066         struct cm_id_private *cm_id_priv;
3067
3068         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3069                                 work->mad_recv_wc->recv_buf.mad;
3070         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3071         if (!cm_id_priv)
3072                 return -EINVAL; /* Unmatched reply. */
3073
3074         spin_lock_irq(&cm_id_priv->lock);
3075         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3076                 spin_unlock_irq(&cm_id_priv->lock);
3077                 goto out;
3078         }
3079         cm_id_priv->id.state = IB_CM_IDLE;
3080         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3081         spin_unlock_irq(&cm_id_priv->lock);
3082
3083         cm_format_sidr_rep_event(work);
3084         cm_process_work(cm_id_priv, work);
3085         return 0;
3086 out:
3087         cm_deref_id(cm_id_priv);
3088         return -EINVAL;
3089 }
3090
3091 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3092                                   enum ib_wc_status wc_status)
3093 {
3094         struct cm_id_private *cm_id_priv;
3095         struct ib_cm_event cm_event;
3096         enum ib_cm_state state;
3097         int ret;
3098
3099         memset(&cm_event, 0, sizeof cm_event);
3100         cm_id_priv = msg->context[0];
3101
3102         /* Discard old sends or ones without a response. */
3103         spin_lock_irq(&cm_id_priv->lock);
3104         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3105         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3106                 goto discard;
3107
3108         switch (state) {
3109         case IB_CM_REQ_SENT:
3110         case IB_CM_MRA_REQ_RCVD:
3111                 cm_reset_to_idle(cm_id_priv);
3112                 cm_event.event = IB_CM_REQ_ERROR;
3113                 break;
3114         case IB_CM_REP_SENT:
3115         case IB_CM_MRA_REP_RCVD:
3116                 cm_reset_to_idle(cm_id_priv);
3117                 cm_event.event = IB_CM_REP_ERROR;
3118                 break;
3119         case IB_CM_DREQ_SENT:
3120                 cm_enter_timewait(cm_id_priv);
3121                 cm_event.event = IB_CM_DREQ_ERROR;
3122                 break;
3123         case IB_CM_SIDR_REQ_SENT:
3124                 cm_id_priv->id.state = IB_CM_IDLE;
3125                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3126                 break;
3127         default:
3128                 goto discard;
3129         }
3130         spin_unlock_irq(&cm_id_priv->lock);
3131         cm_event.param.send_status = wc_status;
3132
3133         /* No other events can occur on the cm_id at this point. */
3134         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3135         cm_free_msg(msg);
3136         if (ret)
3137                 ib_destroy_cm_id(&cm_id_priv->id);
3138         return;
3139 discard:
3140         spin_unlock_irq(&cm_id_priv->lock);
3141         cm_free_msg(msg);
3142 }
3143
3144 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3145                             struct ib_mad_send_wc *mad_send_wc)
3146 {
3147         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3148         struct cm_port *port;
3149         u16 attr_index;
3150
3151         port = mad_agent->context;
3152         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3153                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3154
3155         /*
3156          * If the send was in response to a received message (context[0] is not
3157          * set to a cm_id), and is not a REJ, then it is a send that was
3158          * manually retried.
3159          */
3160         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3161                 msg->retries = 1;
3162
3163         atomic_long_add(1 + msg->retries,
3164                         &port->counter_group[CM_XMIT].counter[attr_index]);
3165         if (msg->retries)
3166                 atomic_long_add(msg->retries,
3167                                 &port->counter_group[CM_XMIT_RETRIES].
3168                                 counter[attr_index]);
3169
3170         switch (mad_send_wc->status) {
3171         case IB_WC_SUCCESS:
3172         case IB_WC_WR_FLUSH_ERR:
3173                 cm_free_msg(msg);
3174                 break;
3175         default:
3176                 if (msg->context[0] && msg->context[1])
3177                         cm_process_send_error(msg, mad_send_wc->status);
3178                 else
3179                         cm_free_msg(msg);
3180                 break;
3181         }
3182 }
3183
3184 static void cm_work_handler(struct work_struct *_work)
3185 {
3186         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3187         int ret;
3188
3189         switch (work->cm_event.event) {
3190         case IB_CM_REQ_RECEIVED:
3191                 ret = cm_req_handler(work);
3192                 break;
3193         case IB_CM_MRA_RECEIVED:
3194                 ret = cm_mra_handler(work);
3195                 break;
3196         case IB_CM_REJ_RECEIVED:
3197                 ret = cm_rej_handler(work);
3198                 break;
3199         case IB_CM_REP_RECEIVED:
3200                 ret = cm_rep_handler(work);
3201                 break;
3202         case IB_CM_RTU_RECEIVED:
3203                 ret = cm_rtu_handler(work);
3204                 break;
3205         case IB_CM_USER_ESTABLISHED:
3206                 ret = cm_establish_handler(work);
3207                 break;
3208         case IB_CM_DREQ_RECEIVED:
3209                 ret = cm_dreq_handler(work);
3210                 break;
3211         case IB_CM_DREP_RECEIVED:
3212                 ret = cm_drep_handler(work);
3213                 break;
3214         case IB_CM_SIDR_REQ_RECEIVED:
3215                 ret = cm_sidr_req_handler(work);
3216                 break;
3217         case IB_CM_SIDR_REP_RECEIVED:
3218                 ret = cm_sidr_rep_handler(work);
3219                 break;
3220         case IB_CM_LAP_RECEIVED:
3221                 ret = cm_lap_handler(work);
3222                 break;
3223         case IB_CM_APR_RECEIVED:
3224                 ret = cm_apr_handler(work);
3225                 break;
3226         case IB_CM_TIMEWAIT_EXIT:
3227                 ret = cm_timewait_handler(work);
3228                 break;
3229         default:
3230                 ret = -EINVAL;
3231                 break;
3232         }
3233         if (ret)
3234                 cm_free_work(work);
3235 }
3236
3237 static int cm_establish(struct ib_cm_id *cm_id)
3238 {
3239         struct cm_id_private *cm_id_priv;
3240         struct cm_work *work;
3241         unsigned long flags;
3242         int ret = 0;
3243
3244         work = kmalloc(sizeof *work, GFP_ATOMIC);
3245         if (!work)
3246                 return -ENOMEM;
3247
3248         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3249         spin_lock_irqsave(&cm_id_priv->lock, flags);
3250         switch (cm_id->state)
3251         {
3252         case IB_CM_REP_SENT:
3253         case IB_CM_MRA_REP_RCVD:
3254                 cm_id->state = IB_CM_ESTABLISHED;
3255                 break;
3256         case IB_CM_ESTABLISHED:
3257                 ret = -EISCONN;
3258                 break;
3259         default:
3260                 ret = -EINVAL;
3261                 break;
3262         }
3263         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3264
3265         if (ret) {
3266                 kfree(work);
3267                 goto out;
3268         }
3269
3270         /*
3271          * The CM worker thread may try to destroy the cm_id before it
3272          * can execute this work item.  To prevent potential deadlock,
3273          * we need to find the cm_id once we're in the context of the
3274          * worker thread, rather than holding a reference on it.
3275          */
3276         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3277         work->local_id = cm_id->local_id;
3278         work->remote_id = cm_id->remote_id;
3279         work->mad_recv_wc = NULL;
3280         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3281         queue_delayed_work(cm.wq, &work->work, 0);
3282 out:
3283         return ret;
3284 }
3285
3286 static int cm_migrate(struct ib_cm_id *cm_id)
3287 {
3288         struct cm_id_private *cm_id_priv;
3289         unsigned long flags;
3290         int ret = 0;
3291
3292         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3293         spin_lock_irqsave(&cm_id_priv->lock, flags);
3294         if (cm_id->state == IB_CM_ESTABLISHED &&
3295             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3296              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3297                 cm_id->lap_state = IB_CM_LAP_IDLE;
3298                 cm_id_priv->av = cm_id_priv->alt_av;
3299         } else
3300                 ret = -EINVAL;
3301         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3302
3303         return ret;
3304 }
3305
3306 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3307 {
3308         int ret;
3309
3310         switch (event) {
3311         case IB_EVENT_COMM_EST:
3312                 ret = cm_establish(cm_id);
3313                 break;
3314         case IB_EVENT_PATH_MIG:
3315                 ret = cm_migrate(cm_id);
3316                 break;
3317         default:
3318                 ret = -EINVAL;
3319         }
3320         return ret;
3321 }
3322 EXPORT_SYMBOL(ib_cm_notify);
3323
3324 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3325                             struct ib_mad_recv_wc *mad_recv_wc)
3326 {
3327         struct cm_port *port = mad_agent->context;
3328         struct cm_work *work;
3329         enum ib_cm_event_type event;
3330         u16 attr_id;
3331         int paths = 0;
3332
3333         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3334         case CM_REQ_ATTR_ID:
3335                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3336                                                     alt_local_lid != 0);
3337                 event = IB_CM_REQ_RECEIVED;
3338                 break;
3339         case CM_MRA_ATTR_ID:
3340                 event = IB_CM_MRA_RECEIVED;
3341                 break;
3342         case CM_REJ_ATTR_ID:
3343                 event = IB_CM_REJ_RECEIVED;
3344                 break;
3345         case CM_REP_ATTR_ID:
3346                 event = IB_CM_REP_RECEIVED;
3347                 break;
3348         case CM_RTU_ATTR_ID:
3349                 event = IB_CM_RTU_RECEIVED;
3350                 break;
3351         case CM_DREQ_ATTR_ID:
3352                 event = IB_CM_DREQ_RECEIVED;
3353                 break;
3354         case CM_DREP_ATTR_ID:
3355                 event = IB_CM_DREP_RECEIVED;
3356                 break;
3357         case CM_SIDR_REQ_ATTR_ID:
3358                 event = IB_CM_SIDR_REQ_RECEIVED;
3359                 break;
3360         case CM_SIDR_REP_ATTR_ID:
3361                 event = IB_CM_SIDR_REP_RECEIVED;
3362                 break;
3363         case CM_LAP_ATTR_ID:
3364                 paths = 1;
3365                 event = IB_CM_LAP_RECEIVED;
3366                 break;
3367         case CM_APR_ATTR_ID:
3368                 event = IB_CM_APR_RECEIVED;
3369                 break;
3370         default:
3371                 ib_free_recv_mad(mad_recv_wc);
3372                 return;
3373         }
3374
3375         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3376         atomic_long_inc(&port->counter_group[CM_RECV].
3377                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3378
3379         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3380                        GFP_KERNEL);
3381         if (!work) {
3382                 ib_free_recv_mad(mad_recv_wc);
3383                 return;
3384         }
3385
3386         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3387         work->cm_event.event = event;
3388         work->mad_recv_wc = mad_recv_wc;
3389         work->port = port;
3390         queue_delayed_work(cm.wq, &work->work, 0);
3391 }
3392
3393 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3394                                 struct ib_qp_attr *qp_attr,
3395                                 int *qp_attr_mask)
3396 {
3397         unsigned long flags;
3398         int ret;
3399
3400         spin_lock_irqsave(&cm_id_priv->lock, flags);
3401         switch (cm_id_priv->id.state) {
3402         case IB_CM_REQ_SENT:
3403         case IB_CM_MRA_REQ_RCVD:
3404         case IB_CM_REQ_RCVD:
3405         case IB_CM_MRA_REQ_SENT:
3406         case IB_CM_REP_RCVD:
3407         case IB_CM_MRA_REP_SENT:
3408         case IB_CM_REP_SENT:
3409         case IB_CM_MRA_REP_RCVD:
3410         case IB_CM_ESTABLISHED:
3411                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3412                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3413                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3414                 if (cm_id_priv->responder_resources)
3415                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3416                                                     IB_ACCESS_REMOTE_ATOMIC;
3417                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3418                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3419                 ret = 0;
3420                 break;
3421         default:
3422                 ret = -EINVAL;
3423                 break;
3424         }
3425         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3426         return ret;
3427 }
3428
3429 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3430                                struct ib_qp_attr *qp_attr,
3431                                int *qp_attr_mask)
3432 {
3433         unsigned long flags;
3434         int ret;
3435
3436         spin_lock_irqsave(&cm_id_priv->lock, flags);
3437         switch (cm_id_priv->id.state) {
3438         case IB_CM_REQ_RCVD:
3439         case IB_CM_MRA_REQ_SENT:
3440         case IB_CM_REP_RCVD:
3441         case IB_CM_MRA_REP_SENT:
3442         case IB_CM_REP_SENT:
3443         case IB_CM_MRA_REP_RCVD:
3444         case IB_CM_ESTABLISHED:
3445                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3446                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3447                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3448                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3449                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3450                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3451                 if (cm_id_priv->qp_type == IB_QPT_RC) {
3452                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3453                                          IB_QP_MIN_RNR_TIMER;
3454                         qp_attr->max_dest_rd_atomic =
3455                                         cm_id_priv->responder_resources;
3456                         qp_attr->min_rnr_timer = 0;
3457                 }
3458                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3459                         *qp_attr_mask |= IB_QP_ALT_PATH;
3460                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3461                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3462                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3463                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3464                 }
3465                 ret = 0;
3466                 break;
3467         default:
3468                 ret = -EINVAL;
3469                 break;
3470         }
3471         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3472         return ret;
3473 }
3474
3475 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3476                                struct ib_qp_attr *qp_attr,
3477                                int *qp_attr_mask)
3478 {
3479         unsigned long flags;
3480         int ret;
3481
3482         spin_lock_irqsave(&cm_id_priv->lock, flags);
3483         switch (cm_id_priv->id.state) {
3484         /* Allow transition to RTS before sending REP */
3485         case IB_CM_REQ_RCVD:
3486         case IB_CM_MRA_REQ_SENT:
3487
3488         case IB_CM_REP_RCVD:
3489         case IB_CM_MRA_REP_SENT:
3490         case IB_CM_REP_SENT:
3491         case IB_CM_MRA_REP_RCVD:
3492         case IB_CM_ESTABLISHED:
3493                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3494                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3495                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3496                         if (cm_id_priv->qp_type == IB_QPT_RC) {
3497                                 *qp_attr_mask |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
3498                                                  IB_QP_RNR_RETRY |
3499                                                  IB_QP_MAX_QP_RD_ATOMIC;
3500                                 qp_attr->timeout = cm_id_priv->av.timeout;
3501                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3502                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3503                                 qp_attr->max_rd_atomic =
3504                                         cm_id_priv->initiator_depth;
3505                         }
3506                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3507                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3508                                 qp_attr->path_mig_state = IB_MIG_REARM;
3509                         }
3510                 } else {
3511                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3512                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3513                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3514                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3515                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3516                         qp_attr->path_mig_state = IB_MIG_REARM;
3517                 }
3518                 ret = 0;
3519                 break;
3520         default:
3521                 ret = -EINVAL;
3522                 break;
3523         }
3524         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3525         return ret;
3526 }
3527
3528 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3529                        struct ib_qp_attr *qp_attr,
3530                        int *qp_attr_mask)
3531 {
3532         struct cm_id_private *cm_id_priv;
3533         int ret;
3534
3535         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3536         switch (qp_attr->qp_state) {
3537         case IB_QPS_INIT:
3538                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3539                 break;
3540         case IB_QPS_RTR:
3541                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3542                 break;
3543         case IB_QPS_RTS:
3544                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3545                 break;
3546         default:
3547                 ret = -EINVAL;
3548                 break;
3549         }
3550         return ret;
3551 }
3552 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3553
3554 static void cm_get_ack_delay(struct cm_device *cm_dev)
3555 {
3556         struct ib_device_attr attr;
3557
3558         if (ib_query_device(cm_dev->device, &attr))
3559                 cm_dev->ack_delay = 0; /* acks will rely on packet life time */
3560         else
3561                 cm_dev->ack_delay = attr.local_ca_ack_delay;
3562 }
3563
3564 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3565                                char *buf)
3566 {
3567         struct cm_counter_group *group;
3568         struct cm_counter_attribute *cm_attr;
3569
3570         group = container_of(obj, struct cm_counter_group, obj);
3571         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3572
3573         return sprintf(buf, "%ld\n",
3574                        atomic_long_read(&group->counter[cm_attr->index]));
3575 }
3576
3577 static struct sysfs_ops cm_counter_ops = {
3578         .show = cm_show_counter
3579 };
3580
3581 static struct kobj_type cm_counter_obj_type = {
3582         .sysfs_ops = &cm_counter_ops,
3583         .default_attrs = cm_counter_default_attrs
3584 };
3585
3586 static void cm_release_port_obj(struct kobject *obj)
3587 {
3588         struct cm_port *cm_port;
3589
3590         cm_port = container_of(obj, struct cm_port, port_obj);
3591         kfree(cm_port);
3592 }
3593
3594 static struct kobj_type cm_port_obj_type = {
3595         .release = cm_release_port_obj
3596 };
3597
3598 static void cm_release_dev_obj(struct kobject *obj)
3599 {
3600         struct cm_device *cm_dev;
3601
3602         cm_dev = container_of(obj, struct cm_device, dev_obj);
3603         kfree(cm_dev);
3604 }
3605
3606 static struct kobj_type cm_dev_obj_type = {
3607         .release = cm_release_dev_obj
3608 };
3609
3610 struct class cm_class = {
3611         .name    = "infiniband_cm",
3612 };
3613 EXPORT_SYMBOL(cm_class);
3614
3615 static int cm_create_port_fs(struct cm_port *port)
3616 {
3617         int i, ret;
3618
3619         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3620                                    &port->cm_dev->dev_obj,
3621                                    "%d", port->port_num);
3622         if (ret) {
3623                 kfree(port);
3624                 return ret;
3625         }
3626
3627         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3628                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3629                                            &cm_counter_obj_type,
3630                                            &port->port_obj,
3631                                            "%s", counter_group_names[i]);
3632                 if (ret)
3633                         goto error;
3634         }
3635
3636         return 0;
3637
3638 error:
3639         while (i--)
3640                 kobject_put(&port->counter_group[i].obj);
3641         kobject_put(&port->port_obj);
3642         return ret;
3643
3644 }
3645
3646 static void cm_remove_port_fs(struct cm_port *port)
3647 {
3648         int i;
3649
3650         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3651                 kobject_put(&port->counter_group[i].obj);
3652
3653         kobject_put(&port->port_obj);
3654 }
3655
3656 static void cm_add_one(struct ib_device *device)
3657 {
3658         struct cm_device *cm_dev;
3659         struct cm_port *port;
3660         struct ib_mad_reg_req reg_req = {
3661                 .mgmt_class = IB_MGMT_CLASS_CM,
3662                 .mgmt_class_version = IB_CM_CLASS_VERSION
3663         };
3664         struct ib_port_modify port_modify = {
3665                 .set_port_cap_mask = IB_PORT_CM_SUP
3666         };
3667         unsigned long flags;
3668         int ret;
3669         u8 i;
3670
3671         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
3672                 return;
3673
3674         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3675                          device->phys_port_cnt, GFP_KERNEL);
3676         if (!cm_dev)
3677                 return;
3678
3679         cm_dev->device = device;
3680         cm_get_ack_delay(cm_dev);
3681
3682         ret = kobject_init_and_add(&cm_dev->dev_obj, &cm_dev_obj_type,
3683                                    &cm_class.subsys.kobj, "%s", device->name);
3684         if (ret) {
3685                 kfree(cm_dev);
3686                 return;
3687         }
3688
3689         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3690         for (i = 1; i <= device->phys_port_cnt; i++) {
3691                 port = kzalloc(sizeof *port, GFP_KERNEL);
3692                 if (!port)
3693                         goto error1;
3694
3695                 cm_dev->port[i-1] = port;
3696                 port->cm_dev = cm_dev;
3697                 port->port_num = i;
3698
3699                 ret = cm_create_port_fs(port);
3700                 if (ret)
3701                         goto error1;
3702
3703                 port->mad_agent = ib_register_mad_agent(device, i,
3704                                                         IB_QPT_GSI,
3705                                                         &reg_req,
3706                                                         0,
3707                                                         cm_send_handler,
3708                                                         cm_recv_handler,
3709                                                         port);
3710                 if (IS_ERR(port->mad_agent))
3711                         goto error2;
3712
3713                 ret = ib_modify_port(device, i, 0, &port_modify);
3714                 if (ret)
3715                         goto error3;
3716         }
3717         ib_set_client_data(device, &cm_client, cm_dev);
3718
3719         write_lock_irqsave(&cm.device_lock, flags);
3720         list_add_tail(&cm_dev->list, &cm.device_list);
3721         write_unlock_irqrestore(&cm.device_lock, flags);
3722         return;
3723
3724 error3:
3725         ib_unregister_mad_agent(port->mad_agent);
3726 error2:
3727         cm_remove_port_fs(port);
3728 error1:
3729         port_modify.set_port_cap_mask = 0;
3730         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3731         while (--i) {
3732                 port = cm_dev->port[i-1];
3733                 ib_modify_port(device, port->port_num, 0, &port_modify);
3734                 ib_unregister_mad_agent(port->mad_agent);
3735                 cm_remove_port_fs(port);
3736         }
3737         kobject_put(&cm_dev->dev_obj);
3738 }
3739
3740 static void cm_remove_one(struct ib_device *device)
3741 {
3742         struct cm_device *cm_dev;
3743         struct cm_port *port;
3744         struct ib_port_modify port_modify = {
3745                 .clr_port_cap_mask = IB_PORT_CM_SUP
3746         };
3747         unsigned long flags;
3748         int i;
3749
3750         cm_dev = ib_get_client_data(device, &cm_client);
3751         if (!cm_dev)
3752                 return;
3753
3754         write_lock_irqsave(&cm.device_lock, flags);
3755         list_del(&cm_dev->list);
3756         write_unlock_irqrestore(&cm.device_lock, flags);
3757
3758         for (i = 1; i <= device->phys_port_cnt; i++) {
3759                 port = cm_dev->port[i-1];
3760                 ib_modify_port(device, port->port_num, 0, &port_modify);
3761                 ib_unregister_mad_agent(port->mad_agent);
3762                 flush_workqueue(cm.wq);
3763                 cm_remove_port_fs(port);
3764         }
3765         kobject_put(&cm_dev->dev_obj);
3766 }
3767
3768 static int __init ib_cm_init(void)
3769 {
3770         int ret;
3771
3772         memset(&cm, 0, sizeof cm);
3773         INIT_LIST_HEAD(&cm.device_list);
3774         rwlock_init(&cm.device_lock);
3775         spin_lock_init(&cm.lock);
3776         cm.listen_service_table = RB_ROOT;
3777         cm.listen_service_id = __constant_be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3778         cm.remote_id_table = RB_ROOT;
3779         cm.remote_qp_table = RB_ROOT;
3780         cm.remote_sidr_table = RB_ROOT;
3781         idr_init(&cm.local_id_table);
3782         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
3783         idr_pre_get(&cm.local_id_table, GFP_KERNEL);
3784         INIT_LIST_HEAD(&cm.timewait_list);
3785
3786         ret = class_register(&cm_class);
3787         if (ret)
3788                 return -ENOMEM;
3789
3790         cm.wq = create_workqueue("ib_cm");
3791         if (!cm.wq) {
3792                 ret = -ENOMEM;
3793                 goto error1;
3794         }
3795
3796         ret = ib_register_client(&cm_client);
3797         if (ret)
3798                 goto error2;
3799
3800         return 0;
3801 error2:
3802         destroy_workqueue(cm.wq);
3803 error1:
3804         class_unregister(&cm_class);
3805         return ret;
3806 }
3807
3808 static void __exit ib_cm_cleanup(void)
3809 {
3810         struct cm_timewait_info *timewait_info, *tmp;
3811
3812         spin_lock_irq(&cm.lock);
3813         list_for_each_entry(timewait_info, &cm.timewait_list, list)
3814                 cancel_delayed_work(&timewait_info->work.work);
3815         spin_unlock_irq(&cm.lock);
3816
3817         ib_unregister_client(&cm_client);
3818         destroy_workqueue(cm.wq);
3819
3820         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
3821                 list_del(&timewait_info->list);
3822                 kfree(timewait_info);
3823         }
3824
3825         class_unregister(&cm_class);
3826         idr_destroy(&cm.local_id_table);
3827 }
3828
3829 module_init(ib_cm_init);
3830 module_exit(ib_cm_cleanup);
3831