SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
 978                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
 979
 980         req_msg->local_comm_id = cm_id_priv->id.local_id;
 981         req_msg->service_id = param->service_id;
 982         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
 983         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
 984         cm_req_set_resp_res(req_msg, param->responder_resources);
 985         cm_req_set_init_depth(req_msg, param->initiator_depth);
 986         cm_req_set_remote_resp_timeout(req_msg,
 987                                        param->remote_cm_response_timeout);
 988         cm_req_set_qp_type(req_msg, param->qp_type);
 989         cm_req_set_flow_ctrl(req_msg, param->flow_control);
 990         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
 991         cm_req_set_local_resp_timeout(req_msg,
 992                                       param->local_cm_response_timeout);
 993         cm_req_set_retry_count(req_msg, param->retry_count);
 994         req_msg->pkey = param->primary_path->pkey;
 995         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
 996         cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
 997         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
 998         cm_req_set_srq(req_msg, param->srq);
 999
1000         req_msg->primary_local_lid = param->primary_path->slid;
1001         req_msg->primary_remote_lid = param->primary_path->dlid;
1002         req_msg->primary_local_gid = param->primary_path->sgid;
1003         req_msg->primary_remote_gid = param->primary_path->dgid;
1004         cm_req_set_primary_flow_label(req_msg, param->primary_path->flow_label);
1005         cm_req_set_primary_packet_rate(req_msg, param->primary_path->rate);
1006         req_msg->primary_traffic_class = param->primary_path->traffic_class;
1007         req_msg->primary_hop_limit = param->primary_path->hop_limit;
1008         cm_req_set_primary_sl(req_msg, param->primary_path->sl);
1009         cm_req_set_primary_subnet_local(req_msg, 1); /* local only... */
1010         cm_req_set_primary_local_ack_timeout(req_msg,
1011                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1012                                param->primary_path->packet_life_time));
1013
1014         if (param->alternate_path) {
1015                 req_msg->alt_local_lid = param->alternate_path->slid;
1016                 req_msg->alt_remote_lid = param->alternate_path->dlid;
1017                 req_msg->alt_local_gid = param->alternate_path->sgid;
1018                 req_msg->alt_remote_gid = param->alternate_path->dgid;
1019                 cm_req_set_alt_flow_label(req_msg,
1020                                           param->alternate_path->flow_label);
1021                 cm_req_set_alt_packet_rate(req_msg, param->alternate_path->rate);
1022                 req_msg->alt_traffic_class = param->alternate_path->traffic_class;
1023                 req_msg->alt_hop_limit = param->alternate_path->hop_limit;
1024                 cm_req_set_alt_sl(req_msg, param->alternate_path->sl);
1025                 cm_req_set_alt_subnet_local(req_msg, 1); /* local only... */
1026                 cm_req_set_alt_local_ack_timeout(req_msg,
1027                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1028                                        param->alternate_path->packet_life_time));
1029         }
1030
1031         if (param->private_data && param->private_data_len)
1032                 memcpy(req_msg->private_data, param->private_data,
1033                        param->private_data_len);
1034 }
1035
1036 static int cm_validate_req_param(struct ib_cm_req_param *param)
1037 {
1038         /* peer-to-peer not supported */
1039         if (param->peer_to_peer)
1040                 return -EINVAL;
1041
1042         if (!param->primary_path)
1043                 return -EINVAL;
1044
1045         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC)
1046                 return -EINVAL;
1047
1048         if (param->private_data &&
1049             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1050                 return -EINVAL;
1051
1052         if (param->alternate_path &&
1053             (param->alternate_path->pkey != param->primary_path->pkey ||
1054              param->alternate_path->mtu != param->primary_path->mtu))
1055                 return -EINVAL;
1056
1057         return 0;
1058 }
1059
1060 int ib_send_cm_req(struct ib_cm_id *cm_id,
1061                    struct ib_cm_req_param *param)
1062 {
1063         struct cm_id_private *cm_id_priv;
1064         struct cm_req_msg *req_msg;
1065         unsigned long flags;
1066         int ret;
1067
1068         ret = cm_validate_req_param(param);
1069         if (ret)
1070                 return ret;
1071
1072         /* Verify that we're not in timewait. */
1073         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1074         spin_lock_irqsave(&cm_id_priv->lock, flags);
1075         if (cm_id->state != IB_CM_IDLE) {
1076                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1077                 ret = -EINVAL;
1078                 goto out;
1079         }
1080         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1081
1082         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1083                                                             id.local_id);
1084         if (IS_ERR(cm_id_priv->timewait_info)) {
1085                 ret = PTR_ERR(cm_id_priv->timewait_info);
1086                 goto out;
1087         }
1088
1089         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av);
1090         if (ret)
1091                 goto error1;
1092         if (param->alternate_path) {
1093                 ret = cm_init_av_by_path(param->alternate_path,
1094                                          &cm_id_priv->alt_av);
1095                 if (ret)
1096                         goto error1;
1097         }
1098         cm_id->service_id = param->service_id;
1099         cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
1100         cm_id_priv->timeout_ms = cm_convert_to_ms(
1101                                     param->primary_path->packet_life_time) * 2 +
1102                                  cm_convert_to_ms(
1103                                     param->remote_cm_response_timeout);
1104         cm_id_priv->max_cm_retries = param->max_cm_retries;
1105         cm_id_priv->initiator_depth = param->initiator_depth;
1106         cm_id_priv->responder_resources = param->responder_resources;
1107         cm_id_priv->retry_count = param->retry_count;
1108         cm_id_priv->path_mtu = param->primary_path->mtu;
1109         cm_id_priv->pkey = param->primary_path->pkey;
1110         cm_id_priv->qp_type = param->qp_type;
1111
1112         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1113         if (ret)
1114                 goto error1;
1115
1116         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1117         cm_format_req(req_msg, cm_id_priv, param);
1118         cm_id_priv->tid = req_msg->hdr.tid;
1119         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1120         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1121
1122         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1123         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1124
1125         spin_lock_irqsave(&cm_id_priv->lock, flags);
1126         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1127         if (ret) {
1128                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1129                 goto error2;
1130         }
1131         BUG_ON(cm_id->state != IB_CM_IDLE);
1132         cm_id->state = IB_CM_REQ_SENT;
1133         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1134         return 0;
1135
1136 error2: cm_free_msg(cm_id_priv->msg);
1137 error1: kfree(cm_id_priv->timewait_info);
1138 out:    return ret;
1139 }
1140 EXPORT_SYMBOL(ib_send_cm_req);
1141
1142 static int cm_issue_rej(struct cm_port *port,
1143                         struct ib_mad_recv_wc *mad_recv_wc,
1144                         enum ib_cm_rej_reason reason,
1145                         enum cm_msg_response msg_rejected,
1146                         void *ari, u8 ari_length)
1147 {
1148         struct ib_mad_send_buf *msg = NULL;
1149         struct cm_rej_msg *rej_msg, *rcv_msg;
1150         int ret;
1151
1152         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1153         if (ret)
1154                 return ret;
1155
1156         /* We just need common CM header information.  Cast to any message. */
1157         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1158         rej_msg = (struct cm_rej_msg *) msg->mad;
1159
1160         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1161         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1162         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1163         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1164         rej_msg->reason = cpu_to_be16(reason);
1165
1166         if (ari && ari_length) {
1167                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1168                 memcpy(rej_msg->ari, ari, ari_length);
1169         }
1170
1171         ret = ib_post_send_mad(msg, NULL);
1172         if (ret)
1173                 cm_free_msg(msg);
1174
1175         return ret;
1176 }
1177
1178 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1179                                     __be32 local_qpn, __be32 remote_qpn)
1180 {
1181         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1182                 ((local_ca_guid == remote_ca_guid) &&
1183                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1184 }
1185
1186 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1187                                             struct ib_sa_path_rec *primary_path,
1188                                             struct ib_sa_path_rec *alt_path)
1189 {
1190         memset(primary_path, 0, sizeof *primary_path);
1191         primary_path->dgid = req_msg->primary_local_gid;
1192         primary_path->sgid = req_msg->primary_remote_gid;
1193         primary_path->dlid = req_msg->primary_local_lid;
1194         primary_path->slid = req_msg->primary_remote_lid;
1195         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1196         primary_path->hop_limit = req_msg->primary_hop_limit;
1197         primary_path->traffic_class = req_msg->primary_traffic_class;
1198         primary_path->reversible = 1;
1199         primary_path->pkey = req_msg->pkey;
1200         primary_path->sl = cm_req_get_primary_sl(req_msg);
1201         primary_path->mtu_selector = IB_SA_EQ;
1202         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1203         primary_path->rate_selector = IB_SA_EQ;
1204         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1205         primary_path->packet_life_time_selector = IB_SA_EQ;
1206         primary_path->packet_life_time =
1207                 cm_req_get_primary_local_ack_timeout(req_msg);
1208         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1209
1210         if (req_msg->alt_local_lid) {
1211                 memset(alt_path, 0, sizeof *alt_path);
1212                 alt_path->dgid = req_msg->alt_local_gid;
1213                 alt_path->sgid = req_msg->alt_remote_gid;
1214                 alt_path->dlid = req_msg->alt_local_lid;
1215                 alt_path->slid = req_msg->alt_remote_lid;
1216                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1217                 alt_path->hop_limit = req_msg->alt_hop_limit;
1218                 alt_path->traffic_class = req_msg->alt_traffic_class;
1219                 alt_path->reversible = 1;
1220                 alt_path->pkey = req_msg->pkey;
1221                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1222                 alt_path->mtu_selector = IB_SA_EQ;
1223                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1224                 alt_path->rate_selector = IB_SA_EQ;
1225                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1226                 alt_path->packet_life_time_selector = IB_SA_EQ;
1227                 alt_path->packet_life_time =
1228                         cm_req_get_alt_local_ack_timeout(req_msg);
1229                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1230         }
1231 }
1232
1233 static void cm_format_req_event(struct cm_work *work,
1234                                 struct cm_id_private *cm_id_priv,
1235                                 struct ib_cm_id *listen_id)
1236 {
1237         struct cm_req_msg *req_msg;
1238         struct ib_cm_req_event_param *param;
1239
1240         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1241         param = &work->cm_event.param.req_rcvd;
1242         param->listen_id = listen_id;
1243         param->port = cm_id_priv->av.port->port_num;
1244         param->primary_path = &work->path[0];
1245         if (req_msg->alt_local_lid)
1246                 param->alternate_path = &work->path[1];
1247         else
1248                 param->alternate_path = NULL;
1249         param->remote_ca_guid = req_msg->local_ca_guid;
1250         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1251         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1252         param->qp_type = cm_req_get_qp_type(req_msg);
1253         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1254         param->responder_resources = cm_req_get_init_depth(req_msg);
1255         param->initiator_depth = cm_req_get_resp_res(req_msg);
1256         param->local_cm_response_timeout =
1257                                         cm_req_get_remote_resp_timeout(req_msg);
1258         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1259         param->remote_cm_response_timeout =
1260                                         cm_req_get_local_resp_timeout(req_msg);
1261         param->retry_count = cm_req_get_retry_count(req_msg);
1262         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1263         param->srq = cm_req_get_srq(req_msg);
1264         work->cm_event.private_data = &req_msg->private_data;
1265 }
1266
1267 static void cm_process_work(struct cm_id_private *cm_id_priv,
1268                             struct cm_work *work)
1269 {
1270         int ret;
1271
1272         /* We will typically only have the current event to report. */
1273         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1274         cm_free_work(work);
1275
1276         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1277                 spin_lock_irq(&cm_id_priv->lock);
1278                 work = cm_dequeue_work(cm_id_priv);
1279                 spin_unlock_irq(&cm_id_priv->lock);
1280                 BUG_ON(!work);
1281                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1282                                                 &work->cm_event);
1283                 cm_free_work(work);
1284         }
1285         cm_deref_id(cm_id_priv);
1286         if (ret)
1287                 cm_destroy_id(&cm_id_priv->id, ret);
1288 }
1289
1290 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1291                           struct cm_id_private *cm_id_priv,
1292                           enum cm_msg_response msg_mraed, u8 service_timeout,
1293                           const void *private_data, u8 private_data_len)
1294 {
1295         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1296         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1297         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1298         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1299         cm_mra_set_service_timeout(mra_msg, service_timeout);
1300
1301         if (private_data && private_data_len)
1302                 memcpy(mra_msg->private_data, private_data, private_data_len);
1303 }
1304
1305 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1306                           struct cm_id_private *cm_id_priv,
1307                           enum ib_cm_rej_reason reason,
1308                           void *ari,
1309                           u8 ari_length,
1310                           const void *private_data,
1311                           u8 private_data_len)
1312 {
1313         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1314         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1315
1316         switch(cm_id_priv->id.state) {
1317         case IB_CM_REQ_RCVD:
1318                 rej_msg->local_comm_id = 0;
1319                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1320                 break;
1321         case IB_CM_MRA_REQ_SENT:
1322                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1323                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1324                 break;
1325         case IB_CM_REP_RCVD:
1326         case IB_CM_MRA_REP_SENT:
1327                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1328                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1329                 break;
1330         default:
1331                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1332                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1333                 break;
1334         }
1335
1336         rej_msg->reason = cpu_to_be16(reason);
1337         if (ari && ari_length) {
1338                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1339                 memcpy(rej_msg->ari, ari, ari_length);
1340         }
1341
1342         if (private_data && private_data_len)
1343                 memcpy(rej_msg->private_data, private_data, private_data_len);
1344 }
1345
1346 static void cm_dup_req_handler(struct cm_work *work,
1347                                struct cm_id_private *cm_id_priv)
1348 {
1349         struct ib_mad_send_buf *msg = NULL;
1350         int ret;
1351
1352         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1353                         counter[CM_REQ_COUNTER]);
1354
1355         /* Quick state check to discard duplicate REQs. */
1356         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1357                 return;
1358
1359         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1360         if (ret)
1361                 return;
1362
1363         spin_lock_irq(&cm_id_priv->lock);
1364         switch (cm_id_priv->id.state) {
1365         case IB_CM_MRA_REQ_SENT:
1366                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1367                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1368                               cm_id_priv->private_data,
1369                               cm_id_priv->private_data_len);
1370                 break;
1371         case IB_CM_TIMEWAIT:
1372                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1373                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1374                 break;
1375         default:
1376                 goto unlock;
1377         }
1378         spin_unlock_irq(&cm_id_priv->lock);
1379
1380         ret = ib_post_send_mad(msg, NULL);
1381         if (ret)
1382                 goto free;
1383         return;
1384
1385 unlock: spin_unlock_irq(&cm_id_priv->lock);
1386 free:   cm_free_msg(msg);
1387 }
1388
1389 static struct cm_id_private * cm_match_req(struct cm_work *work,
1390                                            struct cm_id_private *cm_id_priv)
1391 {
1392         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1393         struct cm_timewait_info *timewait_info;
1394         struct cm_req_msg *req_msg;
1395
1396         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1397
1398         /* Check for possible duplicate REQ. */
1399         spin_lock_irq(&cm.lock);
1400         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1401         if (timewait_info) {
1402                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1403                                            timewait_info->work.remote_id);
1404                 spin_unlock_irq(&cm.lock);
1405                 if (cur_cm_id_priv) {
1406                         cm_dup_req_handler(work, cur_cm_id_priv);
1407                         cm_deref_id(cur_cm_id_priv);
1408                 }
1409                 return NULL;
1410         }
1411
1412         /* Check for stale connections. */
1413         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1414         if (timewait_info) {
1415                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1416                 spin_unlock_irq(&cm.lock);
1417                 cm_issue_rej(work->port, work->mad_recv_wc,
1418                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1419                              NULL, 0);
1420                 return NULL;
1421         }
1422
1423         /* Find matching listen request. */
1424         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1425                                            req_msg->service_id,
1426                                            req_msg->private_data);
1427         if (!listen_cm_id_priv) {
1428                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1429                 spin_unlock_irq(&cm.lock);
1430                 cm_issue_rej(work->port, work->mad_recv_wc,
1431                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1432                              NULL, 0);
1433                 goto out;
1434         }
1435         atomic_inc(&listen_cm_id_priv->refcount);
1436         atomic_inc(&cm_id_priv->refcount);
1437         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1438         atomic_inc(&cm_id_priv->work_count);
1439         spin_unlock_irq(&cm.lock);
1440 out:
1441         return listen_cm_id_priv;
1442 }
1443
1444 static int cm_req_handler(struct cm_work *work)
1445 {
1446         struct ib_cm_id *cm_id;
1447         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1448         struct cm_req_msg *req_msg;
1449         int ret;
1450
1451         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1452
1453         cm_id = ib_create_cm_id(work->port->cm_dev->device, NULL, NULL);
1454         if (IS_ERR(cm_id))
1455                 return PTR_ERR(cm_id);
1456
1457         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1458         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1459         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1460                                 work->mad_recv_wc->recv_buf.grh,
1461                                 &cm_id_priv->av);
1462         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1463                                                             id.local_id);
1464         if (IS_ERR(cm_id_priv->timewait_info)) {
1465                 ret = PTR_ERR(cm_id_priv->timewait_info);
1466                 goto destroy;
1467         }
1468         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1469         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1470         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1471
1472         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1473         if (!listen_cm_id_priv) {
1474                 ret = -EINVAL;
1475                 kfree(cm_id_priv->timewait_info);
1476                 goto destroy;
1477         }
1478
1479         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1480         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1481         cm_id_priv->id.service_id = req_msg->service_id;
1482         cm_id_priv->id.service_mask = __constant_cpu_to_be64(~0ULL);
1483
1484         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1485         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
1486         if (ret) {
1487                 ib_get_cached_gid(work->port->cm_dev->device,
1488                                   work->port->port_num, 0, &work->path[0].sgid);
1489                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1490                                &work->path[0].sgid, sizeof work->path[0].sgid,
1491                                NULL, 0);
1492                 goto rejected;
1493         }
1494         if (req_msg->alt_local_lid) {
1495                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
1496                 if (ret) {
1497                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1498                                        &work->path[0].sgid,
1499                                        sizeof work->path[0].sgid, NULL, 0);
1500                         goto rejected;
1501                 }
1502         }
1503         cm_id_priv->tid = req_msg->hdr.tid;
1504         cm_id_priv->timeout_ms = cm_convert_to_ms(
1505                                         cm_req_get_local_resp_timeout(req_msg));
1506         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1507         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1508         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1509         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1510         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1511         cm_id_priv->pkey = req_msg->pkey;
1512         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1513         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1514         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1515         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1516
1517         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1518         cm_process_work(cm_id_priv, work);
1519         cm_deref_id(listen_cm_id_priv);
1520         return 0;
1521
1522 rejected:
1523         atomic_dec(&cm_id_priv->refcount);
1524         cm_deref_id(listen_cm_id_priv);
1525 destroy:
1526         ib_destroy_cm_id(cm_id);
1527         return ret;
1528 }
1529
1530 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1531                           struct cm_id_private *cm_id_priv,
1532                           struct ib_cm_rep_param *param)
1533 {
1534         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1535         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1536         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1537         cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1538         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1539         rep_msg->resp_resources = param->responder_resources;
1540         rep_msg->initiator_depth = param->initiator_depth;
1541         cm_rep_set_target_ack_delay(rep_msg,
1542                                     cm_id_priv->av.port->cm_dev->ack_delay);
1543         cm_rep_set_failover(rep_msg, param->failover_accepted);
1544         cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1545         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1546         cm_rep_set_srq(rep_msg, param->srq);
1547         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1548
1549         if (param->private_data && param->private_data_len)
1550                 memcpy(rep_msg->private_data, param->private_data,
1551                        param->private_data_len);
1552 }
1553
1554 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1555                    struct ib_cm_rep_param *param)
1556 {
1557         struct cm_id_private *cm_id_priv;
1558         struct ib_mad_send_buf *msg;
1559         struct cm_rep_msg *rep_msg;
1560         unsigned long flags;
1561         int ret;
1562
1563         if (param->private_data &&
1564             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1565                 return -EINVAL;
1566
1567         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1568         spin_lock_irqsave(&cm_id_priv->lock, flags);
1569         if (cm_id->state != IB_CM_REQ_RCVD &&
1570             cm_id->state != IB_CM_MRA_REQ_SENT) {
1571                 ret = -EINVAL;
1572                 goto out;
1573         }
1574
1575         ret = cm_alloc_msg(cm_id_priv, &msg);
1576         if (ret)
1577                 goto out;
1578
1579         rep_msg = (struct cm_rep_msg *) msg->mad;
1580         cm_format_rep(rep_msg, cm_id_priv, param);
1581         msg->timeout_ms = cm_id_priv->timeout_ms;
1582         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1583
1584         ret = ib_post_send_mad(msg, NULL);
1585         if (ret) {
1586                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1587                 cm_free_msg(msg);
1588                 return ret;
1589         }
1590
1591         cm_id->state = IB_CM_REP_SENT;
1592         cm_id_priv->msg = msg;
1593         cm_id_priv->initiator_depth = param->initiator_depth;
1594         cm_id_priv->responder_resources = param->responder_resources;
1595         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1596         cm_id_priv->local_qpn = cm_rep_get_local_qpn(rep_msg);
1597
1598 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1599         return ret;
1600 }
1601 EXPORT_SYMBOL(ib_send_cm_rep);
1602
1603 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1604                           struct cm_id_private *cm_id_priv,
1605                           const void *private_data,
1606                           u8 private_data_len)
1607 {
1608         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1609         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1610         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1611
1612         if (private_data && private_data_len)
1613                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1614 }
1615
1616 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1617                    const void *private_data,
1618                    u8 private_data_len)
1619 {
1620         struct cm_id_private *cm_id_priv;
1621         struct ib_mad_send_buf *msg;
1622         unsigned long flags;
1623         void *data;
1624         int ret;
1625
1626         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1627                 return -EINVAL;
1628
1629         data = cm_copy_private_data(private_data, private_data_len);
1630         if (IS_ERR(data))
1631                 return PTR_ERR(data);
1632
1633         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1634         spin_lock_irqsave(&cm_id_priv->lock, flags);
1635         if (cm_id->state != IB_CM_REP_RCVD &&
1636             cm_id->state != IB_CM_MRA_REP_SENT) {
1637                 ret = -EINVAL;
1638                 goto error;
1639         }
1640
1641         ret = cm_alloc_msg(cm_id_priv, &msg);
1642         if (ret)
1643                 goto error;
1644
1645         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1646                       private_data, private_data_len);
1647
1648         ret = ib_post_send_mad(msg, NULL);
1649         if (ret) {
1650                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1651                 cm_free_msg(msg);
1652                 kfree(data);
1653                 return ret;
1654         }
1655
1656         cm_id->state = IB_CM_ESTABLISHED;
1657         cm_set_private_data(cm_id_priv, data, private_data_len);
1658         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1659         return 0;
1660
1661 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1662         kfree(data);
1663         return ret;
1664 }
1665 EXPORT_SYMBOL(ib_send_cm_rtu);
1666
1667 static void cm_format_rep_event(struct cm_work *work)
1668 {
1669         struct cm_rep_msg *rep_msg;
1670         struct ib_cm_rep_event_param *param;
1671
1672         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1673         param = &work->cm_event.param.rep_rcvd;
1674         param->remote_ca_guid = rep_msg->local_ca_guid;
1675         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1676         param->remote_qpn = be32_to_cpu(cm_rep_get_local_qpn(rep_msg));
1677         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1678         param->responder_resources = rep_msg->initiator_depth;
1679         param->initiator_depth = rep_msg->resp_resources;
1680         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1681         param->failover_accepted = cm_rep_get_failover(rep_msg);
1682         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1683         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1684         param->srq = cm_rep_get_srq(rep_msg);
1685         work->cm_event.private_data = &rep_msg->private_data;
1686 }
1687
1688 static void cm_dup_rep_handler(struct cm_work *work)
1689 {
1690         struct cm_id_private *cm_id_priv;
1691         struct cm_rep_msg *rep_msg;
1692         struct ib_mad_send_buf *msg = NULL;
1693         int ret;
1694
1695         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1696         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1697                                    rep_msg->local_comm_id);
1698         if (!cm_id_priv)
1699                 return;
1700
1701         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1702                         counter[CM_REP_COUNTER]);
1703         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1704         if (ret)
1705                 goto deref;
1706
1707         spin_lock_irq(&cm_id_priv->lock);
1708         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1709                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1710                               cm_id_priv->private_data,
1711                               cm_id_priv->private_data_len);
1712         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1713                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1714                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1715                               cm_id_priv->private_data,
1716                               cm_id_priv->private_data_len);
1717         else
1718                 goto unlock;
1719         spin_unlock_irq(&cm_id_priv->lock);
1720
1721         ret = ib_post_send_mad(msg, NULL);
1722         if (ret)
1723                 goto free;
1724         goto deref;
1725
1726 unlock: spin_unlock_irq(&cm_id_priv->lock);
1727 free:   cm_free_msg(msg);
1728 deref:  cm_deref_id(cm_id_priv);
1729 }
1730
1731 static int cm_rep_handler(struct cm_work *work)
1732 {
1733         struct cm_id_private *cm_id_priv;
1734         struct cm_rep_msg *rep_msg;
1735         int ret;
1736
1737         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1738         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1739         if (!cm_id_priv) {
1740                 cm_dup_rep_handler(work);
1741                 return -EINVAL;
1742         }
1743
1744         cm_format_rep_event(work);
1745
1746         spin_lock_irq(&cm_id_priv->lock);
1747         switch (cm_id_priv->id.state) {
1748         case IB_CM_REQ_SENT:
1749         case IB_CM_MRA_REQ_RCVD:
1750                 break;
1751         default:
1752                 spin_unlock_irq(&cm_id_priv->lock);
1753                 ret = -EINVAL;
1754                 goto error;
1755         }
1756
1757         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1758         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1759         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1760
1761         spin_lock(&cm.lock);
1762         /* Check for duplicate REP. */
1763         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1764                 spin_unlock(&cm.lock);
1765                 spin_unlock_irq(&cm_id_priv->lock);
1766                 ret = -EINVAL;
1767                 goto error;
1768         }
1769         /* Check for a stale connection. */
1770         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
1771                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
1772                          &cm.remote_id_table);
1773                 cm_id_priv->timewait_info->inserted_remote_id = 0;
1774                 spin_unlock(&cm.lock);
1775                 spin_unlock_irq(&cm_id_priv->lock);
1776                 cm_issue_rej(work->port, work->mad_recv_wc,
1777                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
1778                              NULL, 0);
1779                 ret = -EINVAL;
1780                 goto error;
1781         }
1782         spin_unlock(&cm.lock);
1783
1784         cm_id_priv->id.state = IB_CM_REP_RCVD;
1785         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
1786         cm_id_priv->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1787         cm_id_priv->initiator_depth = rep_msg->resp_resources;
1788         cm_id_priv->responder_resources = rep_msg->initiator_depth;
1789         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
1790         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1791         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1792         cm_id_priv->av.timeout =
1793                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1794                                        cm_id_priv->av.timeout - 1);
1795         cm_id_priv->alt_av.timeout =
1796                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1797                                        cm_id_priv->alt_av.timeout - 1);
1798
1799         /* todo: handle peer_to_peer */
1800
1801         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1802         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1803         if (!ret)
1804                 list_add_tail(&work->list, &cm_id_priv->work_list);
1805         spin_unlock_irq(&cm_id_priv->lock);
1806
1807         if (ret)
1808                 cm_process_work(cm_id_priv, work);
1809         else
1810                 cm_deref_id(cm_id_priv);
1811         return 0;
1812
1813 error:
1814         cm_deref_id(cm_id_priv);
1815         return ret;
1816 }
1817
1818 static int cm_establish_handler(struct cm_work *work)
1819 {
1820         struct cm_id_private *cm_id_priv;
1821         int ret;
1822
1823         /* See comment in cm_establish about lookup. */
1824         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
1825         if (!cm_id_priv)
1826                 return -EINVAL;
1827
1828         spin_lock_irq(&cm_id_priv->lock);
1829         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
1830                 spin_unlock_irq(&cm_id_priv->lock);
1831                 goto out;
1832         }
1833
1834         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1835         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1836         if (!ret)
1837                 list_add_tail(&work->list, &cm_id_priv->work_list);
1838         spin_unlock_irq(&cm_id_priv->lock);
1839
1840         if (ret)
1841                 cm_process_work(cm_id_priv, work);
1842         else
1843                 cm_deref_id(cm_id_priv);
1844         return 0;
1845 out:
1846         cm_deref_id(cm_id_priv);
1847         return -EINVAL;
1848 }
1849
1850 static int cm_rtu_handler(struct cm_work *work)
1851 {
1852         struct cm_id_private *cm_id_priv;
1853         struct cm_rtu_msg *rtu_msg;
1854         int ret;
1855
1856         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
1857         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
1858                                    rtu_msg->local_comm_id);
1859         if (!cm_id_priv)
1860                 return -EINVAL;
1861
1862         work->cm_event.private_data = &rtu_msg->private_data;
1863
1864         spin_lock_irq(&cm_id_priv->lock);
1865         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
1866             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
1867                 spin_unlock_irq(&cm_id_priv->lock);
1868                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1869                                 counter[CM_RTU_COUNTER]);
1870                 goto out;
1871         }
1872         cm_id_priv->id.state = IB_CM_ESTABLISHED;
1873
1874         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1875         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1876         if (!ret)
1877                 list_add_tail(&work->list, &cm_id_priv->work_list);
1878         spin_unlock_irq(&cm_id_priv->lock);
1879
1880         if (ret)
1881                 cm_process_work(cm_id_priv, work);
1882         else
1883                 cm_deref_id(cm_id_priv);
1884         return 0;
1885 out:
1886         cm_deref_id(cm_id_priv);
1887         return -EINVAL;
1888 }
1889
1890 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
1891                           struct cm_id_private *cm_id_priv,
1892                           const void *private_data,
1893                           u8 private_data_len)
1894 {
1895         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
1896                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
1897         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
1898         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
1899         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
1900
1901         if (private_data && private_data_len)
1902                 memcpy(dreq_msg->private_data, private_data, private_data_len);
1903 }
1904
1905 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
1906                     const void *private_data,
1907                     u8 private_data_len)
1908 {
1909         struct cm_id_private *cm_id_priv;
1910         struct ib_mad_send_buf *msg;
1911         unsigned long flags;
1912         int ret;
1913
1914         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
1915                 return -EINVAL;
1916
1917         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1918         spin_lock_irqsave(&cm_id_priv->lock, flags);
1919         if (cm_id->state != IB_CM_ESTABLISHED) {
1920                 ret = -EINVAL;
1921                 goto out;
1922         }
1923
1924         ret = cm_alloc_msg(cm_id_priv, &msg);
1925         if (ret) {
1926                 cm_enter_timewait(cm_id_priv);
1927                 goto out;
1928         }
1929
1930         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
1931                        private_data, private_data_len);
1932         msg->timeout_ms = cm_id_priv->timeout_ms;
1933         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
1934
1935         ret = ib_post_send_mad(msg, NULL);
1936         if (ret) {
1937                 cm_enter_timewait(cm_id_priv);
1938                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1939                 cm_free_msg(msg);
1940                 return ret;
1941         }
1942
1943         cm_id->state = IB_CM_DREQ_SENT;
1944         cm_id_priv->msg = msg;
1945 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1946         return ret;
1947 }
1948 EXPORT_SYMBOL(ib_send_cm_dreq);
1949
1950 static void cm_format_drep(struct cm_drep_msg *drep_msg,
1951                           struct cm_id_private *cm_id_priv,
1952                           const void *private_data,
1953                           u8 private_data_len)
1954 {
1955         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
1956         drep_msg->local_comm_id = cm_id_priv->id.local_id;
1957         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1958
1959         if (private_data && private_data_len)
1960                 memcpy(drep_msg->private_data, private_data, private_data_len);
1961 }
1962
1963 int ib_send_cm_drep(struct ib_cm_id *cm_id,
1964                     const void *private_data,
1965                     u8 private_data_len)
1966 {
1967         struct cm_id_private *cm_id_priv;
1968         struct ib_mad_send_buf *msg;
1969         unsigned long flags;
1970         void *data;
1971         int ret;
1972
1973         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
1974                 return -EINVAL;
1975
1976         data = cm_copy_private_data(private_data, private_data_len);
1977         if (IS_ERR(data))
1978                 return PTR_ERR(data);
1979
1980         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1981         spin_lock_irqsave(&cm_id_priv->lock, flags);
1982         if (cm_id->state != IB_CM_DREQ_RCVD) {
1983                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1984                 kfree(data);
1985                 return -EINVAL;
1986         }
1987
1988         cm_set_private_data(cm_id_priv, data, private_data_len);
1989         cm_enter_timewait(cm_id_priv);
1990
1991         ret = cm_alloc_msg(cm_id_priv, &msg);
1992         if (ret)
1993                 goto out;
1994
1995         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
1996                        private_data, private_data_len);
1997
1998         ret = ib_post_send_mad(msg, NULL);
1999         if (ret) {
2000                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2001                 cm_free_msg(msg);
2002                 return ret;
2003         }
2004
2005 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2006         return ret;
2007 }
2008 EXPORT_SYMBOL(ib_send_cm_drep);
2009
2010 static int cm_issue_drep(struct cm_port *port,
2011                          struct ib_mad_recv_wc *mad_recv_wc)
2012 {
2013         struct ib_mad_send_buf *msg = NULL;
2014         struct cm_dreq_msg *dreq_msg;
2015         struct cm_drep_msg *drep_msg;
2016         int ret;
2017
2018         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2019         if (ret)
2020                 return ret;
2021
2022         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2023         drep_msg = (struct cm_drep_msg *) msg->mad;
2024
2025         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2026         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2027         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2028
2029         ret = ib_post_send_mad(msg, NULL);
2030         if (ret)
2031                 cm_free_msg(msg);
2032
2033         return ret;
2034 }
2035
2036 static int cm_dreq_handler(struct cm_work *work)
2037 {
2038         struct cm_id_private *cm_id_priv;
2039         struct cm_dreq_msg *dreq_msg;
2040         struct ib_mad_send_buf *msg = NULL;
2041         int ret;
2042
2043         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2044         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2045                                    dreq_msg->local_comm_id);
2046         if (!cm_id_priv) {
2047                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2048                                 counter[CM_DREQ_COUNTER]);
2049                 cm_issue_drep(work->port, work->mad_recv_wc);
2050                 return -EINVAL;
2051         }
2052
2053         work->cm_event.private_data = &dreq_msg->private_data;
2054
2055         spin_lock_irq(&cm_id_priv->lock);
2056         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2057                 goto unlock;
2058
2059         switch (cm_id_priv->id.state) {
2060         case IB_CM_REP_SENT:
2061         case IB_CM_DREQ_SENT:
2062                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2063                 break;
2064         case IB_CM_ESTABLISHED:
2065         case IB_CM_MRA_REP_RCVD:
2066                 break;
2067         case IB_CM_TIMEWAIT:
2068                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2069                                 counter[CM_DREQ_COUNTER]);
2070                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2071                         goto unlock;
2072
2073                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2074                                cm_id_priv->private_data,
2075                                cm_id_priv->private_data_len);
2076                 spin_unlock_irq(&cm_id_priv->lock);
2077
2078                 if (ib_post_send_mad(msg, NULL))
2079                         cm_free_msg(msg);
2080                 goto deref;
2081         case IB_CM_DREQ_RCVD:
2082                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2083                                 counter[CM_DREQ_COUNTER]);
2084                 goto unlock;
2085         default:
2086                 goto unlock;
2087         }
2088         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2089         cm_id_priv->tid = dreq_msg->hdr.tid;
2090         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2091         if (!ret)
2092                 list_add_tail(&work->list, &cm_id_priv->work_list);
2093         spin_unlock_irq(&cm_id_priv->lock);
2094
2095         if (ret)
2096                 cm_process_work(cm_id_priv, work);
2097         else
2098                 cm_deref_id(cm_id_priv);
2099         return 0;
2100
2101 unlock: spin_unlock_irq(&cm_id_priv->lock);
2102 deref:  cm_deref_id(cm_id_priv);
2103         return -EINVAL;
2104 }
2105
2106 static int cm_drep_handler(struct cm_work *work)
2107 {
2108         struct cm_id_private *cm_id_priv;
2109         struct cm_drep_msg *drep_msg;
2110         int ret;
2111
2112         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2113         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2114                                    drep_msg->local_comm_id);
2115         if (!cm_id_priv)
2116                 return -EINVAL;
2117
2118         work->cm_event.private_data = &drep_msg->private_data;
2119
2120         spin_lock_irq(&cm_id_priv->lock);
2121         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2122             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2123                 spin_unlock_irq(&cm_id_priv->lock);
2124                 goto out;
2125         }
2126         cm_enter_timewait(cm_id_priv);
2127
2128         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2129         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2130         if (!ret)
2131                 list_add_tail(&work->list, &cm_id_priv->work_list);
2132         spin_unlock_irq(&cm_id_priv->lock);
2133
2134         if (ret)
2135                 cm_process_work(cm_id_priv, work);
2136         else
2137                 cm_deref_id(cm_id_priv);
2138         return 0;
2139 out:
2140         cm_deref_id(cm_id_priv);
2141         return -EINVAL;
2142 }
2143
2144 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2145                    enum ib_cm_rej_reason reason,
2146                    void *ari,
2147                    u8 ari_length,
2148                    const void *private_data,
2149                    u8 private_data_len)
2150 {
2151         struct cm_id_private *cm_id_priv;
2152         struct ib_mad_send_buf *msg;
2153         unsigned long flags;
2154         int ret;
2155
2156         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2157             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2158                 return -EINVAL;
2159
2160         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2161
2162         spin_lock_irqsave(&cm_id_priv->lock, flags);
2163         switch (cm_id->state) {
2164         case IB_CM_REQ_SENT:
2165         case IB_CM_MRA_REQ_RCVD:
2166         case IB_CM_REQ_RCVD:
2167         case IB_CM_MRA_REQ_SENT:
2168         case IB_CM_REP_RCVD:
2169         case IB_CM_MRA_REP_SENT:
2170                 ret = cm_alloc_msg(cm_id_priv, &msg);
2171                 if (!ret)
2172                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2173                                       cm_id_priv, reason, ari, ari_length,
2174                                       private_data, private_data_len);
2175
2176                 cm_reset_to_idle(cm_id_priv);
2177                 break;
2178         case IB_CM_REP_SENT:
2179         case IB_CM_MRA_REP_RCVD:
2180                 ret = cm_alloc_msg(cm_id_priv, &msg);
2181                 if (!ret)
2182                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2183                                       cm_id_priv, reason, ari, ari_length,
2184                                       private_data, private_data_len);
2185
2186                 cm_enter_timewait(cm_id_priv);
2187                 break;
2188         default:
2189                 ret = -EINVAL;
2190                 goto out;
2191         }
2192
2193         if (ret)
2194                 goto out;
2195
2196         ret = ib_post_send_mad(msg, NULL);
2197         if (ret)
2198                 cm_free_msg(msg);
2199
2200 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2201         return ret;
2202 }
2203 EXPORT_SYMBOL(ib_send_cm_rej);
2204
2205 static void cm_format_rej_event(struct cm_work *work)
2206 {
2207         struct cm_rej_msg *rej_msg;
2208         struct ib_cm_rej_event_param *param;
2209
2210         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2211         param = &work->cm_event.param.rej_rcvd;
2212         param->ari = rej_msg->ari;
2213         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2214         param->reason = __be16_to_cpu(rej_msg->reason);
2215         work->cm_event.private_data = &rej_msg->private_data;
2216 }
2217
2218 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2219 {
2220         struct cm_timewait_info *timewait_info;
2221         struct cm_id_private *cm_id_priv;
2222         __be32 remote_id;
2223
2224         remote_id = rej_msg->local_comm_id;
2225
2226         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2227                 spin_lock_irq(&cm.lock);
2228                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2229                                                   remote_id);
2230                 if (!timewait_info) {
2231                         spin_unlock_irq(&cm.lock);
2232                         return NULL;
2233                 }
2234                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2235                                       (timewait_info->work.local_id ^
2236                                        cm.random_id_operand));
2237                 if (cm_id_priv) {
2238                         if (cm_id_priv->id.remote_id == remote_id)
2239                                 atomic_inc(&cm_id_priv->refcount);
2240                         else
2241                                 cm_id_priv = NULL;
2242                 }
2243                 spin_unlock_irq(&cm.lock);
2244         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2245                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2246         else
2247                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2248
2249         return cm_id_priv;
2250 }
2251
2252 static int cm_rej_handler(struct cm_work *work)
2253 {
2254         struct cm_id_private *cm_id_priv;
2255         struct cm_rej_msg *rej_msg;
2256         int ret;
2257
2258         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2259         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2260         if (!cm_id_priv)
2261                 return -EINVAL;
2262
2263         cm_format_rej_event(work);
2264
2265         spin_lock_irq(&cm_id_priv->lock);
2266         switch (cm_id_priv->id.state) {
2267         case IB_CM_REQ_SENT:
2268         case IB_CM_MRA_REQ_RCVD:
2269         case IB_CM_REP_SENT:
2270         case IB_CM_MRA_REP_RCVD:
2271                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2272                 /* fall through */
2273         case IB_CM_REQ_RCVD:
2274         case IB_CM_MRA_REQ_SENT:
2275                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2276                         cm_enter_timewait(cm_id_priv);
2277                 else
2278                         cm_reset_to_idle(cm_id_priv);
2279                 break;
2280         case IB_CM_DREQ_SENT:
2281                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2282                 /* fall through */
2283         case IB_CM_REP_RCVD:
2284         case IB_CM_MRA_REP_SENT:
2285         case IB_CM_ESTABLISHED:
2286                 cm_enter_timewait(cm_id_priv);
2287                 break;
2288         default:
2289                 spin_unlock_irq(&cm_id_priv->lock);
2290                 ret = -EINVAL;
2291                 goto out;
2292         }
2293
2294         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2295         if (!ret)
2296                 list_add_tail(&work->list, &cm_id_priv->work_list);
2297         spin_unlock_irq(&cm_id_priv->lock);
2298
2299         if (ret)
2300                 cm_process_work(cm_id_priv, work);
2301         else
2302                 cm_deref_id(cm_id_priv);
2303         return 0;
2304 out:
2305         cm_deref_id(cm_id_priv);
2306         return -EINVAL;
2307 }
2308
2309 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2310                    u8 service_timeout,
2311                    const void *private_data,
2312                    u8 private_data_len)
2313 {
2314         struct cm_id_private *cm_id_priv;
2315         struct ib_mad_send_buf *msg;
2316         enum ib_cm_state cm_state;
2317         enum ib_cm_lap_state lap_state;
2318         enum cm_msg_response msg_response;
2319         void *data;
2320         unsigned long flags;
2321         int ret;
2322
2323         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2324                 return -EINVAL;
2325
2326         data = cm_copy_private_data(private_data, private_data_len);
2327         if (IS_ERR(data))
2328                 return PTR_ERR(data);
2329
2330         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2331
2332         spin_lock_irqsave(&cm_id_priv->lock, flags);
2333         switch(cm_id_priv->id.state) {
2334         case IB_CM_REQ_RCVD:
2335                 cm_state = IB_CM_MRA_REQ_SENT;
2336                 lap_state = cm_id->lap_state;
2337                 msg_response = CM_MSG_RESPONSE_REQ;
2338                 break;
2339         case IB_CM_REP_RCVD:
2340                 cm_state = IB_CM_MRA_REP_SENT;
2341                 lap_state = cm_id->lap_state;
2342                 msg_response = CM_MSG_RESPONSE_REP;
2343                 break;
2344         case IB_CM_ESTABLISHED:
2345                 cm_state = cm_id->state;
2346                 lap_state = IB_CM_MRA_LAP_SENT;
2347                 msg_response = CM_MSG_RESPONSE_OTHER;
2348                 break;
2349         default:
2350                 ret = -EINVAL;
2351                 goto error1;
2352         }
2353
2354         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2355                 ret = cm_alloc_msg(cm_id_priv, &msg);
2356                 if (ret)
2357                         goto error1;
2358
2359                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2360                               msg_response, service_timeout,
2361                               private_data, private_data_len);
2362                 ret = ib_post_send_mad(msg, NULL);
2363                 if (ret)
2364                         goto error2;
2365         }
2366
2367         cm_id->state = cm_state;
2368         cm_id->lap_state = lap_state;
2369         cm_id_priv->service_timeout = service_timeout;
2370         cm_set_private_data(cm_id_priv, data, private_data_len);
2371         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2372         return 0;
2373
2374 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2375         kfree(data);
2376         return ret;
2377
2378 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2379         kfree(data);
2380         cm_free_msg(msg);
2381         return ret;
2382 }
2383 EXPORT_SYMBOL(ib_send_cm_mra);
2384
2385 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2386 {
2387         switch (cm_mra_get_msg_mraed(mra_msg)) {
2388         case CM_MSG_RESPONSE_REQ:
2389                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2390         case CM_MSG_RESPONSE_REP:
2391         case CM_MSG_RESPONSE_OTHER:
2392                 return cm_acquire_id(mra_msg->remote_comm_id,
2393                                      mra_msg->local_comm_id);
2394         default:
2395                 return NULL;
2396         }
2397 }
2398
2399 static int cm_mra_handler(struct cm_work *work)
2400 {
2401         struct cm_id_private *cm_id_priv;
2402         struct cm_mra_msg *mra_msg;
2403         int timeout, ret;
2404
2405         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2406         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2407         if (!cm_id_priv)
2408                 return -EINVAL;
2409
2410         work->cm_event.private_data = &mra_msg->private_data;
2411         work->cm_event.param.mra_rcvd.service_timeout =
2412                                         cm_mra_get_service_timeout(mra_msg);
2413         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2414                   cm_convert_to_ms(cm_id_priv->av.timeout);
2415
2416         spin_lock_irq(&cm_id_priv->lock);
2417         switch (cm_id_priv->id.state) {
2418         case IB_CM_REQ_SENT:
2419                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2420                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2421                                   cm_id_priv->msg, timeout))
2422                         goto out;
2423                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2424                 break;
2425         case IB_CM_REP_SENT:
2426                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2427                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2428                                   cm_id_priv->msg, timeout))
2429                         goto out;
2430                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2431                 break;
2432         case IB_CM_ESTABLISHED:
2433                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2434                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2435                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2436                                   cm_id_priv->msg, timeout)) {
2437                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2438                                 atomic_long_inc(&work->port->
2439                                                 counter_group[CM_RECV_DUPLICATES].
2440                                                 counter[CM_MRA_COUNTER]);
2441                         goto out;
2442                 }
2443                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2444                 break;
2445         case IB_CM_MRA_REQ_RCVD:
2446         case IB_CM_MRA_REP_RCVD:
2447                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2448                                 counter[CM_MRA_COUNTER]);
2449                 /* fall through */
2450         default:
2451                 goto out;
2452         }
2453
2454         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2455                                       cm_id_priv->id.state;
2456         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2457         if (!ret)
2458                 list_add_tail(&work->list, &cm_id_priv->work_list);
2459         spin_unlock_irq(&cm_id_priv->lock);
2460
2461         if (ret)
2462                 cm_process_work(cm_id_priv, work);
2463         else
2464                 cm_deref_id(cm_id_priv);
2465         return 0;
2466 out:
2467         spin_unlock_irq(&cm_id_priv->lock);
2468         cm_deref_id(cm_id_priv);
2469         return -EINVAL;
2470 }
2471
2472 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2473                           struct cm_id_private *cm_id_priv,
2474                           struct ib_sa_path_rec *alternate_path,
2475                           const void *private_data,
2476                           u8 private_data_len)
2477 {
2478         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2479                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2480         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2481         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2482         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2483         /* todo: need remote CM response timeout */
2484         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2485         lap_msg->alt_local_lid = alternate_path->slid;
2486         lap_msg->alt_remote_lid = alternate_path->dlid;
2487         lap_msg->alt_local_gid = alternate_path->sgid;
2488         lap_msg->alt_remote_gid = alternate_path->dgid;
2489         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2490         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2491         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2492         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2493         cm_lap_set_sl(lap_msg, alternate_path->sl);
2494         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2495         cm_lap_set_local_ack_timeout(lap_msg,
2496                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2497                                alternate_path->packet_life_time));
2498
2499         if (private_data && private_data_len)
2500                 memcpy(lap_msg->private_data, private_data, private_data_len);
2501 }
2502
2503 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2504                    struct ib_sa_path_rec *alternate_path,
2505                    const void *private_data,
2506                    u8 private_data_len)
2507 {
2508         struct cm_id_private *cm_id_priv;
2509         struct ib_mad_send_buf *msg;
2510         unsigned long flags;
2511         int ret;
2512
2513         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2514                 return -EINVAL;
2515
2516         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2517         spin_lock_irqsave(&cm_id_priv->lock, flags);
2518         if (cm_id->state != IB_CM_ESTABLISHED ||
2519             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2520              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2521                 ret = -EINVAL;
2522                 goto out;
2523         }
2524
2525         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
2526         if (ret)
2527                 goto out;
2528         cm_id_priv->alt_av.timeout =
2529                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2530                                        cm_id_priv->alt_av.timeout - 1);
2531
2532         ret = cm_alloc_msg(cm_id_priv, &msg);
2533         if (ret)
2534                 goto out;
2535
2536         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2537                       alternate_path, private_data, private_data_len);
2538         msg->timeout_ms = cm_id_priv->timeout_ms;
2539         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2540
2541         ret = ib_post_send_mad(msg, NULL);
2542         if (ret) {
2543                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2544                 cm_free_msg(msg);
2545                 return ret;
2546         }
2547
2548         cm_id->lap_state = IB_CM_LAP_SENT;
2549         cm_id_priv->msg = msg;
2550
2551 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2552         return ret;
2553 }
2554 EXPORT_SYMBOL(ib_send_cm_lap);
2555
2556 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2557                                     struct ib_sa_path_rec *path,
2558                                     struct cm_lap_msg *lap_msg)
2559 {
2560         memset(path, 0, sizeof *path);
2561         path->dgid = lap_msg->alt_local_gid;
2562         path->sgid = lap_msg->alt_remote_gid;
2563         path->dlid = lap_msg->alt_local_lid;
2564         path->slid = lap_msg->alt_remote_lid;
2565         path->flow_label = cm_lap_get_flow_label(lap_msg);
2566         path->hop_limit = lap_msg->alt_hop_limit;
2567         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2568         path->reversible = 1;
2569         path->pkey = cm_id_priv->pkey;
2570         path->sl = cm_lap_get_sl(lap_msg);
2571         path->mtu_selector = IB_SA_EQ;
2572         path->mtu = cm_id_priv->path_mtu;
2573         path->rate_selector = IB_SA_EQ;
2574         path->rate = cm_lap_get_packet_rate(lap_msg);
2575         path->packet_life_time_selector = IB_SA_EQ;
2576         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2577         path->packet_life_time -= (path->packet_life_time > 0);
2578 }
2579
2580 static int cm_lap_handler(struct cm_work *work)
2581 {
2582         struct cm_id_private *cm_id_priv;
2583         struct cm_lap_msg *lap_msg;
2584         struct ib_cm_lap_event_param *param;
2585         struct ib_mad_send_buf *msg = NULL;
2586         int ret;
2587
2588         /* todo: verify LAP request and send reject APR if invalid. */
2589         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2590         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2591                                    lap_msg->local_comm_id);
2592         if (!cm_id_priv)
2593                 return -EINVAL;
2594
2595         param = &work->cm_event.param.lap_rcvd;
2596         param->alternate_path = &work->path[0];
2597         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2598         work->cm_event.private_data = &lap_msg->private_data;
2599
2600         spin_lock_irq(&cm_id_priv->lock);
2601         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2602                 goto unlock;
2603
2604         switch (cm_id_priv->id.lap_state) {
2605         case IB_CM_LAP_UNINIT:
2606         case IB_CM_LAP_IDLE:
2607                 break;
2608         case IB_CM_MRA_LAP_SENT:
2609                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2610                                 counter[CM_LAP_COUNTER]);
2611                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2612                         goto unlock;
2613
2614                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2615                               CM_MSG_RESPONSE_OTHER,
2616                               cm_id_priv->service_timeout,
2617                               cm_id_priv->private_data,
2618                               cm_id_priv->private_data_len);
2619                 spin_unlock_irq(&cm_id_priv->lock);
2620
2621                 if (ib_post_send_mad(msg, NULL))
2622                         cm_free_msg(msg);
2623                 goto deref;
2624         case IB_CM_LAP_RCVD:
2625                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2626                                 counter[CM_LAP_COUNTER]);
2627                 goto unlock;
2628         default:
2629                 goto unlock;
2630         }
2631
2632         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2633         cm_id_priv->tid = lap_msg->hdr.tid;
2634         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2635                                 work->mad_recv_wc->recv_buf.grh,
2636                                 &cm_id_priv->av);
2637         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av);
2638         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2639         if (!ret)
2640                 list_add_tail(&work->list, &cm_id_priv->work_list);
2641         spin_unlock_irq(&cm_id_priv->lock);
2642
2643         if (ret)
2644                 cm_process_work(cm_id_priv, work);
2645         else
2646                 cm_deref_id(cm_id_priv);
2647         return 0;
2648
2649 unlock: spin_unlock_irq(&cm_id_priv->lock);
2650 deref:  cm_deref_id(cm_id_priv);
2651         return -EINVAL;
2652 }
2653
2654 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2655                           struct cm_id_private *cm_id_priv,
2656                           enum ib_cm_apr_status status,
2657                           void *info,
2658                           u8 info_length,
2659                           const void *private_data,
2660                           u8 private_data_len)
2661 {
2662         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2663         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2664         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2665         apr_msg->ap_status = (u8) status;
2666
2667         if (info && info_length) {
2668                 apr_msg->info_length = info_length;
2669                 memcpy(apr_msg->info, info, info_length);
2670         }
2671
2672         if (private_data && private_data_len)
2673                 memcpy(apr_msg->private_data, private_data, private_data_len);
2674 }
2675
2676 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2677                    enum ib_cm_apr_status status,
2678                    void *info,
2679                    u8 info_length,
2680                    const void *private_data,
2681                    u8 private_data_len)
2682 {
2683         struct cm_id_private *cm_id_priv;
2684         struct ib_mad_send_buf *msg;
2685         unsigned long flags;
2686         int ret;
2687
2688         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2689             (info && info_length > IB_CM_APR_INFO_LENGTH))
2690                 return -EINVAL;
2691
2692         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2693         spin_lock_irqsave(&cm_id_priv->lock, flags);
2694         if (cm_id->state != IB_CM_ESTABLISHED ||
2695             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2696              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2697                 ret = -EINVAL;
2698                 goto out;
2699         }
2700
2701         ret = cm_alloc_msg(cm_id_priv, &msg);
2702         if (ret)
2703                 goto out;
2704
2705         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2706                       info, info_length, private_data, private_data_len);
2707         ret = ib_post_send_mad(msg, NULL);
2708         if (ret) {
2709                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2710                 cm_free_msg(msg);
2711                 return ret;
2712         }
2713
2714         cm_id->lap_state = IB_CM_LAP_IDLE;
2715 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2716         return ret;
2717 }
2718 EXPORT_SYMBOL(ib_send_cm_apr);
2719
2720 static int cm_apr_handler(struct cm_work *work)
2721 {
2722         struct cm_id_private *cm_id_priv;
2723         struct cm_apr_msg *apr_msg;
2724         int ret;
2725
2726         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2727         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2728                                    apr_msg->local_comm_id);
2729         if (!cm_id_priv)
2730                 return -EINVAL; /* Unmatched reply. */
2731
2732         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2733         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2734         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2735         work->cm_event.private_data = &apr_msg->private_data;
2736
2737         spin_lock_irq(&cm_id_priv->lock);
2738         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2739             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2740              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2741                 spin_unlock_irq(&cm_id_priv->lock);
2742                 goto out;
2743         }
2744         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2745         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2746         cm_id_priv->msg = NULL;
2747
2748         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2749         if (!ret)
2750                 list_add_tail(&work->list, &cm_id_priv->work_list);
2751         spin_unlock_irq(&cm_id_priv->lock);
2752
2753         if (ret)
2754                 cm_process_work(cm_id_priv, work);
2755         else
2756                 cm_deref_id(cm_id_priv);
2757         return 0;
2758 out:
2759         cm_deref_id(cm_id_priv);
2760         return -EINVAL;
2761 }
2762
2763 static int cm_timewait_handler(struct cm_work *work)
2764 {
2765         struct cm_timewait_info *timewait_info;
2766         struct cm_id_private *cm_id_priv;
2767         int ret;
2768
2769         timewait_info = (struct cm_timewait_info *)work;
2770         spin_lock_irq(&cm.lock);
2771         list_del(&timewait_info->list);
2772         spin_unlock_irq(&cm.lock);
2773
2774         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2775                                    timewait_info->work.remote_id);
2776         if (!cm_id_priv)
2777                 return -EINVAL;
2778
2779         spin_lock_irq(&cm_id_priv->lock);
2780         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2781             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2782                 spin_unlock_irq(&cm_id_priv->lock);
2783                 goto out;
2784         }
2785         cm_id_priv->id.state = IB_CM_IDLE;
2786         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2787         if (!ret)
2788                 list_add_tail(&work->list, &cm_id_priv->work_list);
2789         spin_unlock_irq(&cm_id_priv->lock);
2790
2791         if (ret)
2792                 cm_process_work(cm_id_priv, work);
2793         else
2794                 cm_deref_id(cm_id_priv);
2795         return 0;
2796 out:
2797         cm_deref_id(cm_id_priv);
2798         return -EINVAL;
2799 }
2800
2801 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
2802                                struct cm_id_private *cm_id_priv,
2803                                struct ib_cm_sidr_req_param *param)
2804 {
2805         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
2806                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
2807         sidr_req_msg->request_id = cm_id_priv->id.local_id;
2808         sidr_req_msg->pkey = cpu_to_be16(param->path->pkey);
2809         sidr_req_msg->service_id = param->service_id;
2810
2811         if (param->private_data && param->private_data_len)
2812                 memcpy(sidr_req_msg->private_data, param->private_data,
2813                        param->private_data_len);
2814 }
2815
2816 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
2817                         struct ib_cm_sidr_req_param *param)
2818 {
2819         struct cm_id_private *cm_id_priv;
2820         struct ib_mad_send_buf *msg;
2821         unsigned long flags;
2822         int ret;
2823
2824         if (!param->path || (param->private_data &&
2825              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
2826                 return -EINVAL;
2827
2828         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2829         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
2830         if (ret)
2831                 goto out;
2832
2833         cm_id->service_id = param->service_id;
2834         cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
2835         cm_id_priv->timeout_ms = param->timeout_ms;
2836         cm_id_priv->max_cm_retries = param->max_cm_retries;
2837         ret = cm_alloc_msg(cm_id_priv, &msg);
2838         if (ret)
2839                 goto out;
2840
2841         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
2842                            param);
2843         msg->timeout_ms = cm_id_priv->timeout_ms;
2844         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
2845
2846         spin_lock_irqsave(&cm_id_priv->lock, flags);
2847         if (cm_id->state == IB_CM_IDLE)
2848                 ret = ib_post_send_mad(msg, NULL);
2849         else
2850                 ret = -EINVAL;
2851
2852         if (ret) {
2853                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2854                 cm_free_msg(msg);
2855                 goto out;
2856         }
2857         cm_id->state = IB_CM_SIDR_REQ_SENT;
2858         cm_id_priv->msg = msg;
2859         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2860 out:
2861         return ret;
2862 }
2863 EXPORT_SYMBOL(ib_send_cm_sidr_req);
2864
2865 static void cm_format_sidr_req_event(struct cm_work *work,
2866                                      struct ib_cm_id *listen_id)
2867 {
2868         struct cm_sidr_req_msg *sidr_req_msg;
2869         struct ib_cm_sidr_req_event_param *param;
2870
2871         sidr_req_msg = (struct cm_sidr_req_msg *)
2872                                 work->mad_recv_wc->recv_buf.mad;
2873         param = &work->cm_event.param.sidr_req_rcvd;
2874         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
2875         param->listen_id = listen_id;
2876         param->port = work->port->port_num;
2877         work->cm_event.private_data = &sidr_req_msg->private_data;
2878 }
2879
2880 static int cm_sidr_req_handler(struct cm_work *work)
2881 {
2882         struct ib_cm_id *cm_id;
2883         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
2884         struct cm_sidr_req_msg *sidr_req_msg;
2885         struct ib_wc *wc;
2886
2887         cm_id = ib_create_cm_id(work->port->cm_dev->device, NULL, NULL);
2888         if (IS_ERR(cm_id))
2889                 return PTR_ERR(cm_id);
2890         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2891
2892         /* Record SGID/SLID and request ID for lookup. */
2893         sidr_req_msg = (struct cm_sidr_req_msg *)
2894                                 work->mad_recv_wc->recv_buf.mad;
2895         wc = work->mad_recv_wc->wc;
2896         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
2897         cm_id_priv->av.dgid.global.interface_id = 0;
2898         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2899                                 work->mad_recv_wc->recv_buf.grh,
2900                                 &cm_id_priv->av);
2901         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
2902         cm_id_priv->tid = sidr_req_msg->hdr.tid;
2903         atomic_inc(&cm_id_priv->work_count);
2904
2905         spin_lock_irq(&cm.lock);
2906         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
2907         if (cur_cm_id_priv) {
2908                 spin_unlock_irq(&cm.lock);
2909                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2910                                 counter[CM_SIDR_REQ_COUNTER]);
2911                 goto out; /* Duplicate message. */
2912         }
2913         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
2914         cur_cm_id_priv = cm_find_listen(cm_id->device,
2915                                         sidr_req_msg->service_id,
2916                                         sidr_req_msg->private_data);
2917         if (!cur_cm_id_priv) {
2918                 spin_unlock_irq(&cm.lock);
2919                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
2920                 goto out; /* No match. */
2921         }
2922         atomic_inc(&cur_cm_id_priv->refcount);
2923         spin_unlock_irq(&cm.lock);
2924
2925         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
2926         cm_id_priv->id.context = cur_cm_id_priv->id.context;
2927         cm_id_priv->id.service_id = sidr_req_msg->service_id;
2928         cm_id_priv->id.service_mask = __constant_cpu_to_be64(~0ULL);
2929
2930         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
2931         cm_process_work(cm_id_priv, work);
2932         cm_deref_id(cur_cm_id_priv);
2933         return 0;
2934 out:
2935         ib_destroy_cm_id(&cm_id_priv->id);
2936         return -EINVAL;
2937 }
2938
2939 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
2940                                struct cm_id_private *cm_id_priv,
2941                                struct ib_cm_sidr_rep_param *param)
2942 {
2943         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
2944                           cm_id_priv->tid);
2945         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
2946         sidr_rep_msg->status = param->status;
2947         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
2948         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
2949         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
2950
2951         if (param->info && param->info_length)
2952                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
2953
2954         if (param->private_data && param->private_data_len)
2955                 memcpy(sidr_rep_msg->private_data, param->private_data,
2956                        param->private_data_len);
2957 }
2958
2959 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
2960                         struct ib_cm_sidr_rep_param *param)
2961 {
2962         struct cm_id_private *cm_id_priv;
2963         struct ib_mad_send_buf *msg;
2964         unsigned long flags;
2965         int ret;
2966
2967         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
2968             (param->private_data &&
2969              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
2970                 return -EINVAL;
2971
2972         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2973         spin_lock_irqsave(&cm_id_priv->lock, flags);
2974         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
2975                 ret = -EINVAL;
2976                 goto error;
2977         }
2978
2979         ret = cm_alloc_msg(cm_id_priv, &msg);
2980         if (ret)
2981                 goto error;
2982
2983         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
2984                            param);
2985         ret = ib_post_send_mad(msg, NULL);
2986         if (ret) {
2987                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2988                 cm_free_msg(msg);
2989                 return ret;
2990         }
2991         cm_id->state = IB_CM_IDLE;
2992         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2993
2994         spin_lock_irqsave(&cm.lock, flags);
2995         rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
2996         spin_unlock_irqrestore(&cm.lock, flags);
2997         return 0;
2998
2999 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3000         return ret;
3001 }
3002 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3003
3004 static void cm_format_sidr_rep_event(struct cm_work *work)
3005 {
3006         struct cm_sidr_rep_msg *sidr_rep_msg;
3007         struct ib_cm_sidr_rep_event_param *param;
3008
3009         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3010                                 work->mad_recv_wc->recv_buf.mad;
3011         param = &work->cm_event.param.sidr_rep_rcvd;
3012         param->status = sidr_rep_msg->status;
3013         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3014         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3015         param->info = &sidr_rep_msg->info;
3016         param->info_len = sidr_rep_msg->info_length;
3017         work->cm_event.private_data = &sidr_rep_msg->private_data;
3018 }
3019
3020 static int cm_sidr_rep_handler(struct cm_work *work)
3021 {
3022         struct cm_sidr_rep_msg *sidr_rep_msg;
3023         struct cm_id_private *cm_id_priv;
3024
3025         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3026                                 work->mad_recv_wc->recv_buf.mad;
3027         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3028         if (!cm_id_priv)
3029                 return -EINVAL; /* Unmatched reply. */
3030
3031         spin_lock_irq(&cm_id_priv->lock);
3032         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3033                 spin_unlock_irq(&cm_id_priv->lock);
3034                 goto out;
3035         }
3036         cm_id_priv->id.state = IB_CM_IDLE;
3037         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3038         spin_unlock_irq(&cm_id_priv->lock);
3039
3040         cm_format_sidr_rep_event(work);
3041         cm_process_work(cm_id_priv, work);
3042         return 0;
3043 out:
3044         cm_deref_id(cm_id_priv);
3045         return -EINVAL;
3046 }
3047
3048 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3049                                   enum ib_wc_status wc_status)
3050 {
3051         struct cm_id_private *cm_id_priv;
3052         struct ib_cm_event cm_event;
3053         enum ib_cm_state state;
3054         int ret;
3055
3056         memset(&cm_event, 0, sizeof cm_event);
3057         cm_id_priv = msg->context[0];
3058
3059         /* Discard old sends or ones without a response. */
3060         spin_lock_irq(&cm_id_priv->lock);
3061         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3062         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3063                 goto discard;
3064
3065         switch (state) {
3066         case IB_CM_REQ_SENT:
3067         case IB_CM_MRA_REQ_RCVD:
3068                 cm_reset_to_idle(cm_id_priv);
3069                 cm_event.event = IB_CM_REQ_ERROR;
3070                 break;
3071         case IB_CM_REP_SENT:
3072         case IB_CM_MRA_REP_RCVD:
3073                 cm_reset_to_idle(cm_id_priv);
3074                 cm_event.event = IB_CM_REP_ERROR;
3075                 break;
3076         case IB_CM_DREQ_SENT:
3077                 cm_enter_timewait(cm_id_priv);
3078                 cm_event.event = IB_CM_DREQ_ERROR;
3079                 break;
3080         case IB_CM_SIDR_REQ_SENT:
3081                 cm_id_priv->id.state = IB_CM_IDLE;
3082                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3083                 break;
3084         default:
3085                 goto discard;
3086         }
3087         spin_unlock_irq(&cm_id_priv->lock);
3088         cm_event.param.send_status = wc_status;
3089
3090         /* No other events can occur on the cm_id at this point. */
3091         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3092         cm_free_msg(msg);
3093         if (ret)
3094                 ib_destroy_cm_id(&cm_id_priv->id);
3095         return;
3096 discard:
3097         spin_unlock_irq(&cm_id_priv->lock);
3098         cm_free_msg(msg);
3099 }
3100
3101 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3102                             struct ib_mad_send_wc *mad_send_wc)
3103 {
3104         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3105         struct cm_port *port;
3106         u16 attr_index;
3107
3108         port = mad_agent->context;
3109         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3110                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3111
3112         /*
3113          * If the send was in response to a received message (context[0] is not
3114          * set to a cm_id), and is not a REJ, then it is a send that was
3115          * manually retried.
3116          */
3117         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3118                 msg->retries = 1;
3119
3120         atomic_long_add(1 + msg->retries,
3121                         &port->counter_group[CM_XMIT].counter[attr_index]);
3122         if (msg->retries)
3123                 atomic_long_add(msg->retries,
3124                                 &port->counter_group[CM_XMIT_RETRIES].
3125                                 counter[attr_index]);
3126
3127         switch (mad_send_wc->status) {
3128         case IB_WC_SUCCESS:
3129         case IB_WC_WR_FLUSH_ERR:
3130                 cm_free_msg(msg);
3131                 break;
3132         default:
3133                 if (msg->context[0] && msg->context[1])
3134                         cm_process_send_error(msg, mad_send_wc->status);
3135                 else
3136                         cm_free_msg(msg);
3137                 break;
3138         }
3139 }
3140
3141 static void cm_work_handler(struct work_struct *_work)
3142 {
3143         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3144         int ret;
3145
3146         switch (work->cm_event.event) {
3147         case IB_CM_REQ_RECEIVED:
3148                 ret = cm_req_handler(work);
3149                 break;
3150         case IB_CM_MRA_RECEIVED:
3151                 ret = cm_mra_handler(work);
3152                 break;
3153         case IB_CM_REJ_RECEIVED:
3154                 ret = cm_rej_handler(work);
3155                 break;
3156         case IB_CM_REP_RECEIVED:
3157                 ret = cm_rep_handler(work);
3158                 break;
3159         case IB_CM_RTU_RECEIVED:
3160                 ret = cm_rtu_handler(work);
3161                 break;
3162         case IB_CM_USER_ESTABLISHED:
3163                 ret = cm_establish_handler(work);
3164                 break;
3165         case IB_CM_DREQ_RECEIVED:
3166                 ret = cm_dreq_handler(work);
3167                 break;
3168         case IB_CM_DREP_RECEIVED:
3169                 ret = cm_drep_handler(work);
3170                 break;
3171         case IB_CM_SIDR_REQ_RECEIVED:
3172                 ret = cm_sidr_req_handler(work);
3173                 break;
3174         case IB_CM_SIDR_REP_RECEIVED:
3175                 ret = cm_sidr_rep_handler(work);
3176                 break;
3177         case IB_CM_LAP_RECEIVED:
3178                 ret = cm_lap_handler(work);
3179                 break;
3180         case IB_CM_APR_RECEIVED:
3181                 ret = cm_apr_handler(work);
3182                 break;
3183         case IB_CM_TIMEWAIT_EXIT:
3184                 ret = cm_timewait_handler(work);
3185                 break;
3186         default:
3187                 ret = -EINVAL;
3188                 break;
3189         }
3190         if (ret)
3191                 cm_free_work(work);
3192 }
3193
3194 static int cm_establish(struct ib_cm_id *cm_id)
3195 {
3196         struct cm_id_private *cm_id_priv;
3197         struct cm_work *work;
3198         unsigned long flags;
3199         int ret = 0;
3200
3201         work = kmalloc(sizeof *work, GFP_ATOMIC);
3202         if (!work)
3203                 return -ENOMEM;
3204
3205         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3206         spin_lock_irqsave(&cm_id_priv->lock, flags);
3207         switch (cm_id->state)
3208         {
3209         case IB_CM_REP_SENT:
3210         case IB_CM_MRA_REP_RCVD:
3211                 cm_id->state = IB_CM_ESTABLISHED;
3212                 break;
3213         case IB_CM_ESTABLISHED:
3214                 ret = -EISCONN;
3215                 break;
3216         default:
3217                 ret = -EINVAL;
3218                 break;
3219         }
3220         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3221
3222         if (ret) {
3223                 kfree(work);
3224                 goto out;
3225         }
3226
3227         /*
3228          * The CM worker thread may try to destroy the cm_id before it
3229          * can execute this work item.  To prevent potential deadlock,
3230          * we need to find the cm_id once we're in the context of the
3231          * worker thread, rather than holding a reference on it.
3232          */
3233         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3234         work->local_id = cm_id->local_id;
3235         work->remote_id = cm_id->remote_id;
3236         work->mad_recv_wc = NULL;
3237         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3238         queue_delayed_work(cm.wq, &work->work, 0);
3239 out:
3240         return ret;
3241 }
3242
3243 static int cm_migrate(struct ib_cm_id *cm_id)
3244 {
3245         struct cm_id_private *cm_id_priv;
3246         unsigned long flags;
3247         int ret = 0;
3248
3249         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3250         spin_lock_irqsave(&cm_id_priv->lock, flags);
3251         if (cm_id->state == IB_CM_ESTABLISHED &&
3252             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3253              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3254                 cm_id->lap_state = IB_CM_LAP_IDLE;
3255                 cm_id_priv->av = cm_id_priv->alt_av;
3256         } else
3257                 ret = -EINVAL;
3258         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3259
3260         return ret;
3261 }
3262
3263 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3264 {
3265         int ret;
3266
3267         switch (event) {
3268         case IB_EVENT_COMM_EST:
3269                 ret = cm_establish(cm_id);
3270                 break;
3271         case IB_EVENT_PATH_MIG:
3272                 ret = cm_migrate(cm_id);
3273                 break;
3274         default:
3275                 ret = -EINVAL;
3276         }
3277         return ret;
3278 }
3279 EXPORT_SYMBOL(ib_cm_notify);
3280
3281 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3282                             struct ib_mad_recv_wc *mad_recv_wc)
3283 {
3284         struct cm_port *port = mad_agent->context;
3285         struct cm_work *work;
3286         enum ib_cm_event_type event;
3287         u16 attr_id;
3288         int paths = 0;
3289
3290         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3291         case CM_REQ_ATTR_ID:
3292                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3293                                                     alt_local_lid != 0);
3294                 event = IB_CM_REQ_RECEIVED;
3295                 break;
3296         case CM_MRA_ATTR_ID:
3297                 event = IB_CM_MRA_RECEIVED;
3298                 break;
3299         case CM_REJ_ATTR_ID:
3300                 event = IB_CM_REJ_RECEIVED;
3301                 break;
3302         case CM_REP_ATTR_ID:
3303                 event = IB_CM_REP_RECEIVED;
3304                 break;
3305         case CM_RTU_ATTR_ID:
3306                 event = IB_CM_RTU_RECEIVED;
3307                 break;
3308         case CM_DREQ_ATTR_ID:
3309                 event = IB_CM_DREQ_RECEIVED;
3310                 break;
3311         case CM_DREP_ATTR_ID:
3312                 event = IB_CM_DREP_RECEIVED;
3313                 break;
3314         case CM_SIDR_REQ_ATTR_ID:
3315                 event = IB_CM_SIDR_REQ_RECEIVED;
3316                 break;
3317         case CM_SIDR_REP_ATTR_ID:
3318                 event = IB_CM_SIDR_REP_RECEIVED;
3319                 break;
3320         case CM_LAP_ATTR_ID:
3321                 paths = 1;
3322                 event = IB_CM_LAP_RECEIVED;
3323                 break;
3324         case CM_APR_ATTR_ID:
3325                 event = IB_CM_APR_RECEIVED;
3326                 break;
3327         default:
3328                 ib_free_recv_mad(mad_recv_wc);
3329                 return;
3330         }
3331
3332         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3333         atomic_long_inc(&port->counter_group[CM_RECV].
3334                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3335
3336         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3337                        GFP_KERNEL);
3338         if (!work) {
3339                 ib_free_recv_mad(mad_recv_wc);
3340                 return;
3341         }
3342
3343         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3344         work->cm_event.event = event;
3345         work->mad_recv_wc = mad_recv_wc;
3346         work->port = port;
3347         queue_delayed_work(cm.wq, &work->work, 0);
3348 }
3349
3350 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3351                                 struct ib_qp_attr *qp_attr,
3352                                 int *qp_attr_mask)
3353 {
3354         unsigned long flags;
3355         int ret;
3356
3357         spin_lock_irqsave(&cm_id_priv->lock, flags);
3358         switch (cm_id_priv->id.state) {
3359         case IB_CM_REQ_SENT:
3360         case IB_CM_MRA_REQ_RCVD:
3361         case IB_CM_REQ_RCVD:
3362         case IB_CM_MRA_REQ_SENT:
3363         case IB_CM_REP_RCVD:
3364         case IB_CM_MRA_REP_SENT:
3365         case IB_CM_REP_SENT:
3366         case IB_CM_MRA_REP_RCVD:
3367         case IB_CM_ESTABLISHED:
3368                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3369                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3370                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3371                 if (cm_id_priv->responder_resources)
3372                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3373                                                     IB_ACCESS_REMOTE_ATOMIC;
3374                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3375                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3376                 ret = 0;
3377                 break;
3378         default:
3379                 ret = -EINVAL;
3380                 break;
3381         }
3382         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3383         return ret;
3384 }
3385
3386 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3387                                struct ib_qp_attr *qp_attr,
3388                                int *qp_attr_mask)
3389 {
3390         unsigned long flags;
3391         int ret;
3392
3393         spin_lock_irqsave(&cm_id_priv->lock, flags);
3394         switch (cm_id_priv->id.state) {
3395         case IB_CM_REQ_RCVD:
3396         case IB_CM_MRA_REQ_SENT:
3397         case IB_CM_REP_RCVD:
3398         case IB_CM_MRA_REP_SENT:
3399         case IB_CM_REP_SENT:
3400         case IB_CM_MRA_REP_RCVD:
3401         case IB_CM_ESTABLISHED:
3402                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3403                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3404                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3405                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3406                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3407                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3408                 if (cm_id_priv->qp_type == IB_QPT_RC) {
3409                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3410                                          IB_QP_MIN_RNR_TIMER;
3411                         qp_attr->max_dest_rd_atomic =
3412                                         cm_id_priv->responder_resources;
3413                         qp_attr->min_rnr_timer = 0;
3414                 }
3415                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3416                         *qp_attr_mask |= IB_QP_ALT_PATH;
3417                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3418                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3419                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3420                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3421                 }
3422                 ret = 0;
3423                 break;
3424         default:
3425                 ret = -EINVAL;
3426                 break;
3427         }
3428         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3429         return ret;
3430 }
3431
3432 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3433                                struct ib_qp_attr *qp_attr,
3434                                int *qp_attr_mask)
3435 {
3436         unsigned long flags;
3437         int ret;
3438
3439         spin_lock_irqsave(&cm_id_priv->lock, flags);
3440         switch (cm_id_priv->id.state) {
3441         /* Allow transition to RTS before sending REP */
3442         case IB_CM_REQ_RCVD:
3443         case IB_CM_MRA_REQ_SENT:
3444
3445         case IB_CM_REP_RCVD:
3446         case IB_CM_MRA_REP_SENT:
3447         case IB_CM_REP_SENT:
3448         case IB_CM_MRA_REP_RCVD:
3449         case IB_CM_ESTABLISHED:
3450                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3451                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3452                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3453                         if (cm_id_priv->qp_type == IB_QPT_RC) {
3454                                 *qp_attr_mask |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
3455                                                  IB_QP_RNR_RETRY |
3456                                                  IB_QP_MAX_QP_RD_ATOMIC;
3457                                 qp_attr->timeout = cm_id_priv->av.timeout;
3458                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3459                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3460                                 qp_attr->max_rd_atomic =
3461                                         cm_id_priv->initiator_depth;
3462                         }
3463                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3464                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3465                                 qp_attr->path_mig_state = IB_MIG_REARM;
3466                         }
3467                 } else {
3468                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3469                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3470                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3471                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3472                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3473                         qp_attr->path_mig_state = IB_MIG_REARM;
3474                 }
3475                 ret = 0;
3476                 break;
3477         default:
3478                 ret = -EINVAL;
3479                 break;
3480         }
3481         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3482         return ret;
3483 }
3484
3485 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3486                        struct ib_qp_attr *qp_attr,
3487                        int *qp_attr_mask)
3488 {
3489         struct cm_id_private *cm_id_priv;
3490         int ret;
3491
3492         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3493         switch (qp_attr->qp_state) {
3494         case IB_QPS_INIT:
3495                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3496                 break;
3497         case IB_QPS_RTR:
3498                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3499                 break;
3500         case IB_QPS_RTS:
3501                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3502                 break;
3503         default:
3504                 ret = -EINVAL;
3505                 break;
3506         }
3507         return ret;
3508 }
3509 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3510
3511 static void cm_get_ack_delay(struct cm_device *cm_dev)
3512 {
3513         struct ib_device_attr attr;
3514
3515         if (ib_query_device(cm_dev->device, &attr))
3516                 cm_dev->ack_delay = 0; /* acks will rely on packet life time */
3517         else
3518                 cm_dev->ack_delay = attr.local_ca_ack_delay;
3519 }
3520
3521 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3522                                char *buf)
3523 {
3524         struct cm_counter_group *group;
3525         struct cm_counter_attribute *cm_attr;
3526
3527         group = container_of(obj, struct cm_counter_group, obj);
3528         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3529
3530         return sprintf(buf, "%ld\n",
3531                        atomic_long_read(&group->counter[cm_attr->index]));
3532 }
3533
3534 static struct sysfs_ops cm_counter_ops = {
3535         .show = cm_show_counter
3536 };
3537
3538 static struct kobj_type cm_counter_obj_type = {
3539         .sysfs_ops = &cm_counter_ops,
3540         .default_attrs = cm_counter_default_attrs
3541 };
3542
3543 static void cm_release_port_obj(struct kobject *obj)
3544 {
3545         struct cm_port *cm_port;
3546
3547         printk(KERN_ERR "free cm port\n");
3548
3549         cm_port = container_of(obj, struct cm_port, port_obj);
3550         kfree(cm_port);
3551 }
3552
3553 static struct kobj_type cm_port_obj_type = {
3554         .release = cm_release_port_obj
3555 };
3556
3557 static void cm_release_dev_obj(struct kobject *obj)
3558 {
3559         struct cm_device *cm_dev;
3560
3561         printk(KERN_ERR "free cm dev\n");
3562
3563         cm_dev = container_of(obj, struct cm_device, dev_obj);
3564         kfree(cm_dev);
3565 }
3566
3567 static struct kobj_type cm_dev_obj_type = {
3568         .release = cm_release_dev_obj
3569 };
3570
3571 struct class cm_class = {
3572         .name    = "infiniband_cm",
3573 };
3574 EXPORT_SYMBOL(cm_class);
3575
3576 static void cm_remove_fs_obj(struct kobject *obj)
3577 {
3578         kobject_put(obj->parent);
3579         kobject_put(obj);
3580 }
3581
3582 static int cm_create_port_fs(struct cm_port *port)
3583 {
3584         int i, ret;
3585
3586         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3587                                    kobject_get(&port->cm_dev->dev_obj),
3588                                    "%d", port->port_num);
3589         if (ret) {
3590                 kfree(port);
3591                 return ret;
3592         }
3593
3594         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3595                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3596                                            &cm_counter_obj_type,
3597                                            kobject_get(&port->port_obj),
3598                                            "%s", counter_group_names[i]);
3599                 if (ret)
3600                         goto error;
3601         }
3602
3603         return 0;
3604
3605 error:
3606         while (i--)
3607                 cm_remove_fs_obj(&port->counter_group[i].obj);
3608         cm_remove_fs_obj(&port->port_obj);
3609         return ret;
3610
3611 }
3612
3613 static void cm_remove_port_fs(struct cm_port *port)
3614 {
3615         int i;
3616
3617         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3618                 cm_remove_fs_obj(&port->counter_group[i].obj);
3619
3620         cm_remove_fs_obj(&port->port_obj);
3621 }
3622
3623 static void cm_add_one(struct ib_device *device)
3624 {
3625         struct cm_device *cm_dev;
3626         struct cm_port *port;
3627         struct ib_mad_reg_req reg_req = {
3628                 .mgmt_class = IB_MGMT_CLASS_CM,
3629                 .mgmt_class_version = IB_CM_CLASS_VERSION
3630         };
3631         struct ib_port_modify port_modify = {
3632                 .set_port_cap_mask = IB_PORT_CM_SUP
3633         };
3634         unsigned long flags;
3635         int ret;
3636         u8 i;
3637
3638         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
3639                 return;
3640
3641         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3642                          device->phys_port_cnt, GFP_KERNEL);
3643         if (!cm_dev)
3644                 return;
3645
3646         cm_dev->device = device;
3647         cm_get_ack_delay(cm_dev);
3648
3649         ret = kobject_init_and_add(&cm_dev->dev_obj, &cm_dev_obj_type,
3650                                    &cm_class.subsys.kobj, "%s", device->name);
3651         if (ret) {
3652                 kfree(cm_dev);
3653                 return;
3654         }
3655
3656         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3657         for (i = 1; i <= device->phys_port_cnt; i++) {
3658                 port = kzalloc(sizeof *port, GFP_KERNEL);
3659                 if (!port)
3660                         goto error1;
3661
3662                 cm_dev->port[i-1] = port;
3663                 port->cm_dev = cm_dev;
3664                 port->port_num = i;
3665
3666                 ret = cm_create_port_fs(port);
3667                 if (ret)
3668                         goto error1;
3669
3670                 port->mad_agent = ib_register_mad_agent(device, i,
3671                                                         IB_QPT_GSI,
3672                                                         &reg_req,
3673                                                         0,
3674                                                         cm_send_handler,
3675                                                         cm_recv_handler,
3676                                                         port);
3677                 if (IS_ERR(port->mad_agent))
3678                         goto error2;
3679
3680                 ret = ib_modify_port(device, i, 0, &port_modify);
3681                 if (ret)
3682                         goto error3;
3683         }
3684         ib_set_client_data(device, &cm_client, cm_dev);
3685
3686         write_lock_irqsave(&cm.device_lock, flags);
3687         list_add_tail(&cm_dev->list, &cm.device_list);
3688         write_unlock_irqrestore(&cm.device_lock, flags);
3689         return;
3690
3691 error3:
3692         ib_unregister_mad_agent(port->mad_agent);
3693 error2:
3694         cm_remove_port_fs(port);
3695 error1:
3696         port_modify.set_port_cap_mask = 0;
3697         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3698         while (--i) {
3699                 port = cm_dev->port[i-1];
3700                 ib_modify_port(device, port->port_num, 0, &port_modify);
3701                 ib_unregister_mad_agent(port->mad_agent);
3702                 cm_remove_port_fs(port);
3703         }
3704         cm_remove_fs_obj(&cm_dev->dev_obj);
3705 }
3706
3707 static void cm_remove_one(struct ib_device *device)
3708 {
3709         struct cm_device *cm_dev;
3710         struct cm_port *port;
3711         struct ib_port_modify port_modify = {
3712                 .clr_port_cap_mask = IB_PORT_CM_SUP
3713         };
3714         unsigned long flags;
3715         int i;
3716
3717         cm_dev = ib_get_client_data(device, &cm_client);
3718         if (!cm_dev)
3719                 return;
3720
3721         write_lock_irqsave(&cm.device_lock, flags);
3722         list_del(&cm_dev->list);
3723         write_unlock_irqrestore(&cm.device_lock, flags);
3724
3725         for (i = 1; i <= device->phys_port_cnt; i++) {
3726                 port = cm_dev->port[i-1];
3727                 ib_modify_port(device, port->port_num, 0, &port_modify);
3728                 ib_unregister_mad_agent(port->mad_agent);
3729                 cm_remove_port_fs(port);
3730         }
3731         cm_remove_fs_obj(&cm_dev->dev_obj);
3732 }
3733
3734 static int __init ib_cm_init(void)
3735 {
3736         int ret;
3737
3738         memset(&cm, 0, sizeof cm);
3739         INIT_LIST_HEAD(&cm.device_list);
3740         rwlock_init(&cm.device_lock);
3741         spin_lock_init(&cm.lock);
3742         cm.listen_service_table = RB_ROOT;
3743         cm.listen_service_id = __constant_be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3744         cm.remote_id_table = RB_ROOT;
3745         cm.remote_qp_table = RB_ROOT;
3746         cm.remote_sidr_table = RB_ROOT;
3747         idr_init(&cm.local_id_table);
3748         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
3749         idr_pre_get(&cm.local_id_table, GFP_KERNEL);
3750         INIT_LIST_HEAD(&cm.timewait_list);
3751
3752         ret = class_register(&cm_class);
3753         if (ret)
3754                 return -ENOMEM;
3755
3756         cm.wq = create_workqueue("ib_cm");
3757         if (!cm.wq) {
3758                 ret = -ENOMEM;
3759                 goto error1;
3760         }
3761
3762         ret = ib_register_client(&cm_client);
3763         if (ret)
3764                 goto error2;
3765
3766         return 0;
3767 error2:
3768         destroy_workqueue(cm.wq);
3769 error1:
3770         class_unregister(&cm_class);
3771         return ret;
3772 }
3773
3774 static void __exit ib_cm_cleanup(void)
3775 {
3776         struct cm_timewait_info *timewait_info, *tmp;
3777
3778         spin_lock_irq(&cm.lock);
3779         list_for_each_entry(timewait_info, &cm.timewait_list, list)
3780                 cancel_delayed_work(&timewait_info->work.work);
3781         spin_unlock_irq(&cm.lock);
3782
3783         destroy_workqueue(cm.wq);
3784
3785         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
3786                 list_del(&timewait_info->list);
3787                 kfree(timewait_info);
3788         }
3789
3790         ib_unregister_client(&cm_client);
3791         class_unregister(&cm_class);
3792         idr_destroy(&cm.local_id_table);
3793 }
3794
3795 module_init(ib_cm_init);
3796 module_exit(ib_cm_cleanup);
3797