drivers/infiniband/core/cm.c

   1 /*
   2  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/dma-mapping.h>
  38 #include <linux/device.h>
  39 #include <linux/module.h>
  40 #include <linux/err.h>
  41 #include <linux/idr.h>
  42 #include <linux/interrupt.h>
  43 #include <linux/random.h>
  44 #include <linux/rbtree.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/slab.h>
  47 #include <linux/sysfs.h>
  48 #include <linux/workqueue.h>
  49 #include <linux/kdev_t.h>
  50
  51 #include <rdma/ib_cache.h>
  52 #include <rdma/ib_cm.h>
  53 #include "cm_msgs.h"
  54
  55 MODULE_AUTHOR("Sean Hefty");
  56 MODULE_DESCRIPTION("InfiniBand CM");
  57 MODULE_LICENSE("Dual BSD/GPL");
  58
  59 static void cm_add_one(struct ib_device *device);
  60 static void cm_remove_one(struct ib_device *device);
  61
  62 static struct ib_client cm_client = {
  63         .name   = "cm",
  64         .add    = cm_add_one,
  65         .remove = cm_remove_one
  66 };
  67
  68 static struct ib_cm {
  69         spinlock_t lock;
  70         struct list_head device_list;
  71         rwlock_t device_lock;
  72         struct rb_root listen_service_table;
  73         u64 listen_service_id;
  74         /* struct rb_root peer_service_table; todo: fix peer to peer */
  75         struct rb_root remote_qp_table;
  76         struct rb_root remote_id_table;
  77         struct rb_root remote_sidr_table;
  78         struct idr local_id_table;
  79         __be32 random_id_operand;
  80         struct list_head timewait_list;
  81         struct workqueue_struct *wq;
  82 } cm;
  83
  84 /* Counter indexes ordered by attribute ID */
  85 enum {
  86         CM_REQ_COUNTER,
  87         CM_MRA_COUNTER,
  88         CM_REJ_COUNTER,
  89         CM_REP_COUNTER,
  90         CM_RTU_COUNTER,
  91         CM_DREQ_COUNTER,
  92         CM_DREP_COUNTER,
  93         CM_SIDR_REQ_COUNTER,
  94         CM_SIDR_REP_COUNTER,
  95         CM_LAP_COUNTER,
  96         CM_APR_COUNTER,
  97         CM_ATTR_COUNT,
  98         CM_ATTR_ID_OFFSET = 0x0010,
  99 };
 100
 101 enum {
 102         CM_XMIT,
 103         CM_XMIT_RETRIES,
 104         CM_RECV,
 105         CM_RECV_DUPLICATES,
 106         CM_COUNTER_GROUPS
 107 };
 108
 109 static char const counter_group_names[CM_COUNTER_GROUPS]
 110                                      [sizeof("cm_rx_duplicates")] = {
 111         "cm_tx_msgs", "cm_tx_retries",
 112         "cm_rx_msgs", "cm_rx_duplicates"
 113 };
 114
 115 struct cm_counter_group {
 116         struct kobject obj;
 117         atomic_long_t counter[CM_ATTR_COUNT];
 118 };
 119
 120 struct cm_counter_attribute {
 121         struct attribute attr;
 122         int index;
 123 };
 124
 125 #define CM_COUNTER_ATTR(_name, _index) \
 126 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 127         .attr = { .name = __stringify(_name), .mode = 0444 }, \
 128         .index = _index \
 129 }
 130
 131 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 132 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 133 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 134 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 135 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 136 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 137 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 138 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 139 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 140 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 141 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 142
 143 static struct attribute *cm_counter_default_attrs[] = {
 144         &cm_req_counter_attr.attr,
 145         &cm_mra_counter_attr.attr,
 146         &cm_rej_counter_attr.attr,
 147         &cm_rep_counter_attr.attr,
 148         &cm_rtu_counter_attr.attr,
 149         &cm_dreq_counter_attr.attr,
 150         &cm_drep_counter_attr.attr,
 151         &cm_sidr_req_counter_attr.attr,
 152         &cm_sidr_rep_counter_attr.attr,
 153         &cm_lap_counter_attr.attr,
 154         &cm_apr_counter_attr.attr,
 155         NULL
 156 };
 157
 158 struct cm_port {
 159         struct cm_device *cm_dev;
 160         struct ib_mad_agent *mad_agent;
 161         struct kobject port_obj;
 162         u8 port_num;
 163         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 164 };
 165
 166 struct cm_device {
 167         struct list_head list;
 168         struct ib_device *ib_device;
 169         struct device *device;
 170         u8 ack_delay;
 171         struct cm_port *port[0];
 172 };
 173
 174 struct cm_av {
 175         struct cm_port *port;
 176         union ib_gid dgid;
 177         struct ib_ah_attr ah_attr;
 178         u16 pkey_index;
 179         u8 timeout;
 180 };
 181
 182 struct cm_work {
 183         struct delayed_work work;
 184         struct list_head list;
 185         struct cm_port *port;
 186         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 187         __be32 local_id;                        /* Established / timewait */
 188         __be32 remote_id;
 189         struct ib_cm_event cm_event;
 190         struct ib_sa_path_rec path[0];
 191 };
 192
 193 struct cm_timewait_info {
 194         struct cm_work work;                    /* Must be first. */
 195         struct list_head list;
 196         struct rb_node remote_qp_node;
 197         struct rb_node remote_id_node;
 198         __be64 remote_ca_guid;
 199         __be32 remote_qpn;
 200         u8 inserted_remote_qp;
 201         u8 inserted_remote_id;
 202 };
 203
 204 struct cm_id_private {
 205         struct ib_cm_id id;
 206
 207         struct rb_node service_node;
 208         struct rb_node sidr_id_node;
 209         spinlock_t lock;        /* Do not acquire inside cm.lock */
 210         struct completion comp;
 211         atomic_t refcount;
 212
 213         struct ib_mad_send_buf *msg;
 214         struct cm_timewait_info *timewait_info;
 215         /* todo: use alternate port on send failure */
 216         struct cm_av av;
 217         struct cm_av alt_av;
 218         struct ib_cm_compare_data *compare_data;
 219
 220         void *private_data;
 221         __be64 tid;
 222         __be32 local_qpn;
 223         __be32 remote_qpn;
 224         enum ib_qp_type qp_type;
 225         __be32 sq_psn;
 226         __be32 rq_psn;
 227         int timeout_ms;
 228         enum ib_mtu path_mtu;
 229         __be16 pkey;
 230         u8 private_data_len;
 231         u8 max_cm_retries;
 232         u8 peer_to_peer;
 233         u8 responder_resources;
 234         u8 initiator_depth;
 235         u8 retry_count;
 236         u8 rnr_retry_count;
 237         u8 service_timeout;
 238         u8 target_ack_delay;
 239
 240         struct list_head work_list;
 241         atomic_t work_count;
 242 };
 243
 244 static void cm_work_handler(struct work_struct *work);
 245
 246 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 247 {
 248         if (atomic_dec_and_test(&cm_id_priv->refcount))
 249                 complete(&cm_id_priv->comp);
 250 }
 251
 252 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 253                         struct ib_mad_send_buf **msg)
 254 {
 255         struct ib_mad_agent *mad_agent;
 256         struct ib_mad_send_buf *m;
 257         struct ib_ah *ah;
 258
 259         mad_agent = cm_id_priv->av.port->mad_agent;
 260         ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
 261         if (IS_ERR(ah))
 262                 return PTR_ERR(ah);
 263
 264         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 265                                cm_id_priv->av.pkey_index,
 266                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 267                                GFP_ATOMIC);
 268         if (IS_ERR(m)) {
 269                 ib_destroy_ah(ah);
 270                 return PTR_ERR(m);
 271         }
 272
 273         /* Timeout set by caller if response is expected. */
 274         m->ah = ah;
 275         m->retries = cm_id_priv->max_cm_retries;
 276
 277         atomic_inc(&cm_id_priv->refcount);
 278         m->context[0] = cm_id_priv;
 279         *msg = m;
 280         return 0;
 281 }
 282
 283 static int cm_alloc_response_msg(struct cm_port *port,
 284                                  struct ib_mad_recv_wc *mad_recv_wc,
 285                                  struct ib_mad_send_buf **msg)
 286 {
 287         struct ib_mad_send_buf *m;
 288         struct ib_ah *ah;
 289
 290         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 291                                   mad_recv_wc->recv_buf.grh, port->port_num);
 292         if (IS_ERR(ah))
 293                 return PTR_ERR(ah);
 294
 295         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 296                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 297                                GFP_ATOMIC);
 298         if (IS_ERR(m)) {
 299                 ib_destroy_ah(ah);
 300                 return PTR_ERR(m);
 301         }
 302         m->ah = ah;
 303         *msg = m;
 304         return 0;
 305 }
 306
 307 static void cm_free_msg(struct ib_mad_send_buf *msg)
 308 {
 309         ib_destroy_ah(msg->ah);
 310         if (msg->context[0])
 311                 cm_deref_id(msg->context[0]);
 312         ib_free_send_mad(msg);
 313 }
 314
 315 static void * cm_copy_private_data(const void *private_data,
 316                                    u8 private_data_len)
 317 {
 318         void *data;
 319
 320         if (!private_data || !private_data_len)
 321                 return NULL;
 322
 323         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 324         if (!data)
 325                 return ERR_PTR(-ENOMEM);
 326
 327         return data;
 328 }
 329
 330 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 331                                  void *private_data, u8 private_data_len)
 332 {
 333         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 334                 kfree(cm_id_priv->private_data);
 335
 336         cm_id_priv->private_data = private_data;
 337         cm_id_priv->private_data_len = private_data_len;
 338 }
 339
 340 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 341                                     struct ib_grh *grh, struct cm_av *av)
 342 {
 343         av->port = port;
 344         av->pkey_index = wc->pkey_index;
 345         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 346                            grh, &av->ah_attr);
 347 }
 348
 349 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
 350 {
 351         struct cm_device *cm_dev;
 352         struct cm_port *port = NULL;
 353         unsigned long flags;
 354         int ret;
 355         u8 p;
 356
 357         read_lock_irqsave(&cm.device_lock, flags);
 358         list_for_each_entry(cm_dev, &cm.device_list, list) {
 359                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 360                                         &p, NULL)) {
 361                         port = cm_dev->port[p-1];
 362                         break;
 363                 }
 364         }
 365         read_unlock_irqrestore(&cm.device_lock, flags);
 366
 367         if (!port)
 368                 return -EINVAL;
 369
 370         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 371                                   be16_to_cpu(path->pkey), &av->pkey_index);
 372         if (ret)
 373                 return ret;
 374
 375         av->port = port;
 376         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 377                              &av->ah_attr);
 378         av->timeout = path->packet_life_time + 1;
 379         return 0;
 380 }
 381
 382 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 383 {
 384         unsigned long flags;
 385         int ret, id;
 386         static int next_id;
 387
 388         do {
 389                 spin_lock_irqsave(&cm.lock, flags);
 390                 ret = idr_get_new_above(&cm.local_id_table, cm_id_priv,
 391                                         next_id, &id);
 392                 if (!ret)
 393                         next_id = ((unsigned) id + 1) & MAX_ID_MASK;
 394                 spin_unlock_irqrestore(&cm.lock, flags);
 395         } while( (ret == -EAGAIN) && idr_pre_get(&cm.local_id_table, GFP_KERNEL) );
 396
 397         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 398         return ret;
 399 }
 400
 401 static void cm_free_id(__be32 local_id)
 402 {
 403         spin_lock_irq(&cm.lock);
 404         idr_remove(&cm.local_id_table,
 405                    (__force int) (local_id ^ cm.random_id_operand));
 406         spin_unlock_irq(&cm.lock);
 407 }
 408
 409 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 410 {
 411         struct cm_id_private *cm_id_priv;
 412
 413         cm_id_priv = idr_find(&cm.local_id_table,
 414                               (__force int) (local_id ^ cm.random_id_operand));
 415         if (cm_id_priv) {
 416                 if (cm_id_priv->id.remote_id == remote_id)
 417                         atomic_inc(&cm_id_priv->refcount);
 418                 else
 419                         cm_id_priv = NULL;
 420         }
 421
 422         return cm_id_priv;
 423 }
 424
 425 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 426 {
 427         struct cm_id_private *cm_id_priv;
 428
 429         spin_lock_irq(&cm.lock);
 430         cm_id_priv = cm_get_id(local_id, remote_id);
 431         spin_unlock_irq(&cm.lock);
 432
 433         return cm_id_priv;
 434 }
 435
 436 static void cm_mask_copy(u8 *dst, u8 *src, u8 *mask)
 437 {
 438         int i;
 439
 440         for (i = 0; i < IB_CM_COMPARE_SIZE / sizeof(unsigned long); i++)
 441                 ((unsigned long *) dst)[i] = ((unsigned long *) src)[i] &
 442                                              ((unsigned long *) mask)[i];
 443 }
 444
 445 static int cm_compare_data(struct ib_cm_compare_data *src_data,
 446                            struct ib_cm_compare_data *dst_data)
 447 {
 448         u8 src[IB_CM_COMPARE_SIZE];
 449         u8 dst[IB_CM_COMPARE_SIZE];
 450
 451         if (!src_data || !dst_data)
 452                 return 0;
 453
 454         cm_mask_copy(src, src_data->data, dst_data->mask);
 455         cm_mask_copy(dst, dst_data->data, src_data->mask);
 456         return memcmp(src, dst, IB_CM_COMPARE_SIZE);
 457 }
 458
 459 static int cm_compare_private_data(u8 *private_data,
 460                                    struct ib_cm_compare_data *dst_data)
 461 {
 462         u8 src[IB_CM_COMPARE_SIZE];
 463
 464         if (!dst_data)
 465                 return 0;
 466
 467         cm_mask_copy(src, private_data, dst_data->mask);
 468         return memcmp(src, dst_data->data, IB_CM_COMPARE_SIZE);
 469 }
 470
 471 /*
 472  * Trivial helpers to strip endian annotation and compare; the
 473  * endianness doesn't actually matter since we just need a stable
 474  * order for the RB tree.
 475  */
 476 static int be32_lt(__be32 a, __be32 b)
 477 {
 478         return (__force u32) a < (__force u32) b;
 479 }
 480
 481 static int be32_gt(__be32 a, __be32 b)
 482 {
 483         return (__force u32) a > (__force u32) b;
 484 }
 485
 486 static int be64_lt(__be64 a, __be64 b)
 487 {
 488         return (__force u64) a < (__force u64) b;
 489 }
 490
 491 static int be64_gt(__be64 a, __be64 b)
 492 {
 493         return (__force u64) a > (__force u64) b;
 494 }
 495
 496 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 497 {
 498         struct rb_node **link = &cm.listen_service_table.rb_node;
 499         struct rb_node *parent = NULL;
 500         struct cm_id_private *cur_cm_id_priv;
 501         __be64 service_id = cm_id_priv->id.service_id;
 502         __be64 service_mask = cm_id_priv->id.service_mask;
 503         int data_cmp;
 504
 505         while (*link) {
 506                 parent = *link;
 507                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 508                                           service_node);
 509                 data_cmp = cm_compare_data(cm_id_priv->compare_data,
 510                                            cur_cm_id_priv->compare_data);
 511                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 512                     (service_mask & cur_cm_id_priv->id.service_id) &&
 513                     (cm_id_priv->id.device == cur_cm_id_priv->id.device) &&
 514                     !data_cmp)
 515                         return cur_cm_id_priv;
 516
 517                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 518                         link = &(*link)->rb_left;
 519                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 520                         link = &(*link)->rb_right;
 521                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 522                         link = &(*link)->rb_left;
 523                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 524                         link = &(*link)->rb_right;
 525                 else if (data_cmp < 0)
 526                         link = &(*link)->rb_left;
 527                 else
 528                         link = &(*link)->rb_right;
 529         }
 530         rb_link_node(&cm_id_priv->service_node, parent, link);
 531         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 532         return NULL;
 533 }
 534
 535 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 536                                              __be64 service_id,
 537                                              u8 *private_data)
 538 {
 539         struct rb_node *node = cm.listen_service_table.rb_node;
 540         struct cm_id_private *cm_id_priv;
 541         int data_cmp;
 542
 543         while (node) {
 544                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 545                 data_cmp = cm_compare_private_data(private_data,
 546                                                    cm_id_priv->compare_data);
 547                 if ((cm_id_priv->id.service_mask & service_id) ==
 548                      cm_id_priv->id.service_id &&
 549                     (cm_id_priv->id.device == device) && !data_cmp)
 550                         return cm_id_priv;
 551
 552                 if (device < cm_id_priv->id.device)
 553                         node = node->rb_left;
 554                 else if (device > cm_id_priv->id.device)
 555                         node = node->rb_right;
 556                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 557                         node = node->rb_left;
 558                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 559                         node = node->rb_right;
 560                 else if (data_cmp < 0)
 561                         node = node->rb_left;
 562                 else
 563                         node = node->rb_right;
 564         }
 565         return NULL;
 566 }
 567
 568 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 569                                                      *timewait_info)
 570 {
 571         struct rb_node **link = &cm.remote_id_table.rb_node;
 572         struct rb_node *parent = NULL;
 573         struct cm_timewait_info *cur_timewait_info;
 574         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 575         __be32 remote_id = timewait_info->work.remote_id;
 576
 577         while (*link) {
 578                 parent = *link;
 579                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 580                                              remote_id_node);
 581                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 582                         link = &(*link)->rb_left;
 583                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 584                         link = &(*link)->rb_right;
 585                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 586                         link = &(*link)->rb_left;
 587                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 588                         link = &(*link)->rb_right;
 589                 else
 590                         return cur_timewait_info;
 591         }
 592         timewait_info->inserted_remote_id = 1;
 593         rb_link_node(&timewait_info->remote_id_node, parent, link);
 594         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 595         return NULL;
 596 }
 597
 598 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 599                                                    __be32 remote_id)
 600 {
 601         struct rb_node *node = cm.remote_id_table.rb_node;
 602         struct cm_timewait_info *timewait_info;
 603
 604         while (node) {
 605                 timewait_info = rb_entry(node, struct cm_timewait_info,
 606                                          remote_id_node);
 607                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 608                         node = node->rb_left;
 609                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 610                         node = node->rb_right;
 611                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 612                         node = node->rb_left;
 613                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 614                         node = node->rb_right;
 615                 else
 616                         return timewait_info;
 617         }
 618         return NULL;
 619 }
 620
 621 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 622                                                       *timewait_info)
 623 {
 624         struct rb_node **link = &cm.remote_qp_table.rb_node;
 625         struct rb_node *parent = NULL;
 626         struct cm_timewait_info *cur_timewait_info;
 627         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 628         __be32 remote_qpn = timewait_info->remote_qpn;
 629
 630         while (*link) {
 631                 parent = *link;
 632                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 633                                              remote_qp_node);
 634                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 635                         link = &(*link)->rb_left;
 636                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 637                         link = &(*link)->rb_right;
 638                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 639                         link = &(*link)->rb_left;
 640                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 641                         link = &(*link)->rb_right;
 642                 else
 643                         return cur_timewait_info;
 644         }
 645         timewait_info->inserted_remote_qp = 1;
 646         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 647         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 648         return NULL;
 649 }
 650
 651 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 652                                                     *cm_id_priv)
 653 {
 654         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 655         struct rb_node *parent = NULL;
 656         struct cm_id_private *cur_cm_id_priv;
 657         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 658         __be32 remote_id = cm_id_priv->id.remote_id;
 659
 660         while (*link) {
 661                 parent = *link;
 662                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 663                                           sidr_id_node);
 664                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 665                         link = &(*link)->rb_left;
 666                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 667                         link = &(*link)->rb_right;
 668                 else {
 669                         int cmp;
 670                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 671                                      sizeof *port_gid);
 672                         if (cmp < 0)
 673                                 link = &(*link)->rb_left;
 674                         else if (cmp > 0)
 675                                 link = &(*link)->rb_right;
 676                         else
 677                                 return cur_cm_id_priv;
 678                 }
 679         }
 680         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 681         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 682         return NULL;
 683 }
 684
 685 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 686                                enum ib_cm_sidr_status status)
 687 {
 688         struct ib_cm_sidr_rep_param param;
 689
 690         memset(&param, 0, sizeof param);
 691         param.status = status;
 692         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 693 }
 694
 695 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 696                                  ib_cm_handler cm_handler,
 697                                  void *context)
 698 {
 699         struct cm_id_private *cm_id_priv;
 700         int ret;
 701
 702         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 703         if (!cm_id_priv)
 704                 return ERR_PTR(-ENOMEM);
 705
 706         cm_id_priv->id.state = IB_CM_IDLE;
 707         cm_id_priv->id.device = device;
 708         cm_id_priv->id.cm_handler = cm_handler;
 709         cm_id_priv->id.context = context;
 710         cm_id_priv->id.remote_cm_qpn = 1;
 711         ret = cm_alloc_id(cm_id_priv);
 712         if (ret)
 713                 goto error;
 714
 715         spin_lock_init(&cm_id_priv->lock);
 716         init_completion(&cm_id_priv->comp);
 717         INIT_LIST_HEAD(&cm_id_priv->work_list);
 718         atomic_set(&cm_id_priv->work_count, -1);
 719         atomic_set(&cm_id_priv->refcount, 1);
 720         return &cm_id_priv->id;
 721
 722 error:
 723         kfree(cm_id_priv);
 724         return ERR_PTR(-ENOMEM);
 725 }
 726 EXPORT_SYMBOL(ib_create_cm_id);
 727
 728 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 729 {
 730         struct cm_work *work;
 731
 732         if (list_empty(&cm_id_priv->work_list))
 733                 return NULL;
 734
 735         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 736         list_del(&work->list);
 737         return work;
 738 }
 739
 740 static void cm_free_work(struct cm_work *work)
 741 {
 742         if (work->mad_recv_wc)
 743                 ib_free_recv_mad(work->mad_recv_wc);
 744         kfree(work);
 745 }
 746
 747 static inline int cm_convert_to_ms(int iba_time)
 748 {
 749         /* approximate conversion to ms from 4.096us x 2^iba_time */
 750         return 1 << max(iba_time - 8, 0);
 751 }
 752
 753 /*
 754  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 755  * Because of how ack_timeout is stored, adding one doubles the timeout.
 756  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 757  * increment it (round up) only if the other is within 50%.
 758  */
 759 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 760 {
 761         int ack_timeout = packet_life_time + 1;
 762
 763         if (ack_timeout >= ca_ack_delay)
 764                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 765         else
 766                 ack_timeout = ca_ack_delay +
 767                               (ack_timeout >= (ca_ack_delay - 1));
 768
 769         return min(31, ack_timeout);
 770 }
 771
 772 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 773 {
 774         if (timewait_info->inserted_remote_id) {
 775                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 776                 timewait_info->inserted_remote_id = 0;
 777         }
 778
 779         if (timewait_info->inserted_remote_qp) {
 780                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 781                 timewait_info->inserted_remote_qp = 0;
 782         }
 783 }
 784
 785 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 786 {
 787         struct cm_timewait_info *timewait_info;
 788
 789         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 790         if (!timewait_info)
 791                 return ERR_PTR(-ENOMEM);
 792
 793         timewait_info->work.local_id = local_id;
 794         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 795         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 796         return timewait_info;
 797 }
 798
 799 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 800 {
 801         int wait_time;
 802         unsigned long flags;
 803
 804         spin_lock_irqsave(&cm.lock, flags);
 805         cm_cleanup_timewait(cm_id_priv->timewait_info);
 806         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 807         spin_unlock_irqrestore(&cm.lock, flags);
 808
 809         /*
 810          * The cm_id could be destroyed by the user before we exit timewait.
 811          * To protect against this, we search for the cm_id after exiting
 812          * timewait before notifying the user that we've exited timewait.
 813          */
 814         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 815         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 816         queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 817                            msecs_to_jiffies(wait_time));
 818         cm_id_priv->timewait_info = NULL;
 819 }
 820
 821 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 822 {
 823         unsigned long flags;
 824
 825         cm_id_priv->id.state = IB_CM_IDLE;
 826         if (cm_id_priv->timewait_info) {
 827                 spin_lock_irqsave(&cm.lock, flags);
 828                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 829                 spin_unlock_irqrestore(&cm.lock, flags);
 830                 kfree(cm_id_priv->timewait_info);
 831                 cm_id_priv->timewait_info = NULL;
 832         }
 833 }
 834
 835 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 836 {
 837         struct cm_id_private *cm_id_priv;
 838         struct cm_work *work;
 839
 840         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 841 retest:
 842         spin_lock_irq(&cm_id_priv->lock);
 843         switch (cm_id->state) {
 844         case IB_CM_LISTEN:
 845                 cm_id->state = IB_CM_IDLE;
 846                 spin_unlock_irq(&cm_id_priv->lock);
 847                 spin_lock_irq(&cm.lock);
 848                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 849                 spin_unlock_irq(&cm.lock);
 850                 break;
 851         case IB_CM_SIDR_REQ_SENT:
 852                 cm_id->state = IB_CM_IDLE;
 853                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 854                 spin_unlock_irq(&cm_id_priv->lock);
 855                 break;
 856         case IB_CM_SIDR_REQ_RCVD:
 857                 spin_unlock_irq(&cm_id_priv->lock);
 858                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 859                 break;
 860         case IB_CM_REQ_SENT:
 861                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 862                 spin_unlock_irq(&cm_id_priv->lock);
 863                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 864                                &cm_id_priv->id.device->node_guid,
 865                                sizeof cm_id_priv->id.device->node_guid,
 866                                NULL, 0);
 867                 break;
 868         case IB_CM_REQ_RCVD:
 869                 if (err == -ENOMEM) {
 870                         /* Do not reject to allow future retries. */
 871                         cm_reset_to_idle(cm_id_priv);
 872                         spin_unlock_irq(&cm_id_priv->lock);
 873                 } else {
 874                         spin_unlock_irq(&cm_id_priv->lock);
 875                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 876                                        NULL, 0, NULL, 0);
 877                 }
 878                 break;
 879         case IB_CM_MRA_REQ_RCVD:
 880         case IB_CM_REP_SENT:
 881         case IB_CM_MRA_REP_RCVD:
 882                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 883                 /* Fall through */
 884         case IB_CM_MRA_REQ_SENT:
 885         case IB_CM_REP_RCVD:
 886         case IB_CM_MRA_REP_SENT:
 887                 spin_unlock_irq(&cm_id_priv->lock);
 888                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 889                                NULL, 0, NULL, 0);
 890                 break;
 891         case IB_CM_ESTABLISHED:
 892                 spin_unlock_irq(&cm_id_priv->lock);
 893                 ib_send_cm_dreq(cm_id, NULL, 0);
 894                 goto retest;
 895         case IB_CM_DREQ_SENT:
 896                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 897                 cm_enter_timewait(cm_id_priv);
 898                 spin_unlock_irq(&cm_id_priv->lock);
 899                 break;
 900         case IB_CM_DREQ_RCVD:
 901                 spin_unlock_irq(&cm_id_priv->lock);
 902                 ib_send_cm_drep(cm_id, NULL, 0);
 903                 break;
 904         default:
 905                 spin_unlock_irq(&cm_id_priv->lock);
 906                 break;
 907         }
 908
 909         cm_free_id(cm_id->local_id);
 910         cm_deref_id(cm_id_priv);
 911         wait_for_completion(&cm_id_priv->comp);
 912         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
 913                 cm_free_work(work);
 914         kfree(cm_id_priv->compare_data);
 915         kfree(cm_id_priv->private_data);
 916         kfree(cm_id_priv);
 917 }
 918
 919 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
 920 {
 921         cm_destroy_id(cm_id, 0);
 922 }
 923 EXPORT_SYMBOL(ib_destroy_cm_id);
 924
 925 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask,
 926                  struct ib_cm_compare_data *compare_data)
 927 {
 928         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
 929         unsigned long flags;
 930         int ret = 0;
 931
 932         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
 933         service_id &= service_mask;
 934         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
 935             (service_id != IB_CM_ASSIGN_SERVICE_ID))
 936                 return -EINVAL;
 937
 938         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 939         if (cm_id->state != IB_CM_IDLE)
 940                 return -EINVAL;
 941
 942         if (compare_data) {
 943                 cm_id_priv->compare_data = kzalloc(sizeof *compare_data,
 944                                                    GFP_KERNEL);
 945                 if (!cm_id_priv->compare_data)
 946                         return -ENOMEM;
 947                 cm_mask_copy(cm_id_priv->compare_data->data,
 948                              compare_data->data, compare_data->mask);
 949                 memcpy(cm_id_priv->compare_data->mask, compare_data->mask,
 950                        IB_CM_COMPARE_SIZE);
 951         }
 952
 953         cm_id->state = IB_CM_LISTEN;
 954
 955         spin_lock_irqsave(&cm.lock, flags);
 956         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
 957                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
 958                 cm_id->service_mask = ~cpu_to_be64(0);
 959         } else {
 960                 cm_id->service_id = service_id;
 961                 cm_id->service_mask = service_mask;
 962         }
 963         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
 964         spin_unlock_irqrestore(&cm.lock, flags);
 965
 966         if (cur_cm_id_priv) {
 967                 cm_id->state = IB_CM_IDLE;
 968                 kfree(cm_id_priv->compare_data);
 969                 cm_id_priv->compare_data = NULL;
 970                 ret = -EBUSY;
 971         }
 972         return ret;
 973 }
 974 EXPORT_SYMBOL(ib_cm_listen);
 975
 976 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
 977                           enum cm_msg_sequence msg_seq)
 978 {
 979         u64 hi_tid, low_tid;
 980
 981         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
 982         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
 983                           (msg_seq << 30));
 984         return cpu_to_be64(hi_tid | low_tid);
 985 }
 986
 987 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
 988                               __be16 attr_id, __be64 tid)
 989 {
 990         hdr->base_version  = IB_MGMT_BASE_VERSION;
 991         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
 992         hdr->class_version = IB_CM_CLASS_VERSION;
 993         hdr->method        = IB_MGMT_METHOD_SEND;
 994         hdr->attr_id       = attr_id;
 995         hdr->tid           = tid;
 996 }
 997
 998 static void cm_format_req(struct cm_req_msg *req_msg,
 999                           struct cm_id_private *cm_id_priv,
1000                           struct ib_cm_req_param *param)
1001 {
1002         struct ib_sa_path_rec *pri_path = param->primary_path;
1003         struct ib_sa_path_rec *alt_path = param->alternate_path;
1004
1005         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1006                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1007
1008         req_msg->local_comm_id = cm_id_priv->id.local_id;
1009         req_msg->service_id = param->service_id;
1010         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1011         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1012         cm_req_set_resp_res(req_msg, param->responder_resources);
1013         cm_req_set_init_depth(req_msg, param->initiator_depth);
1014         cm_req_set_remote_resp_timeout(req_msg,
1015                                        param->remote_cm_response_timeout);
1016         cm_req_set_qp_type(req_msg, param->qp_type);
1017         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1018         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1019         cm_req_set_local_resp_timeout(req_msg,
1020                                       param->local_cm_response_timeout);
1021         cm_req_set_retry_count(req_msg, param->retry_count);
1022         req_msg->pkey = param->primary_path->pkey;
1023         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1024         cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1025         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1026         cm_req_set_srq(req_msg, param->srq);
1027
1028         if (pri_path->hop_limit <= 1) {
1029                 req_msg->primary_local_lid = pri_path->slid;
1030                 req_msg->primary_remote_lid = pri_path->dlid;
1031         } else {
1032                 /* Work-around until there's a way to obtain remote LID info */
1033                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1034                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1035         }
1036         req_msg->primary_local_gid = pri_path->sgid;
1037         req_msg->primary_remote_gid = pri_path->dgid;
1038         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1039         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1040         req_msg->primary_traffic_class = pri_path->traffic_class;
1041         req_msg->primary_hop_limit = pri_path->hop_limit;
1042         cm_req_set_primary_sl(req_msg, pri_path->sl);
1043         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1044         cm_req_set_primary_local_ack_timeout(req_msg,
1045                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1046                                pri_path->packet_life_time));
1047
1048         if (alt_path) {
1049                 if (alt_path->hop_limit <= 1) {
1050                         req_msg->alt_local_lid = alt_path->slid;
1051                         req_msg->alt_remote_lid = alt_path->dlid;
1052                 } else {
1053                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1054                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1055                 }
1056                 req_msg->alt_local_gid = alt_path->sgid;
1057                 req_msg->alt_remote_gid = alt_path->dgid;
1058                 cm_req_set_alt_flow_label(req_msg,
1059                                           alt_path->flow_label);
1060                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1061                 req_msg->alt_traffic_class = alt_path->traffic_class;
1062                 req_msg->alt_hop_limit = alt_path->hop_limit;
1063                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1064                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1065                 cm_req_set_alt_local_ack_timeout(req_msg,
1066                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1067                                        alt_path->packet_life_time));
1068         }
1069
1070         if (param->private_data && param->private_data_len)
1071                 memcpy(req_msg->private_data, param->private_data,
1072                        param->private_data_len);
1073 }
1074
1075 static int cm_validate_req_param(struct ib_cm_req_param *param)
1076 {
1077         /* peer-to-peer not supported */
1078         if (param->peer_to_peer)
1079                 return -EINVAL;
1080
1081         if (!param->primary_path)
1082                 return -EINVAL;
1083
1084         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC)
1085                 return -EINVAL;
1086
1087         if (param->private_data &&
1088             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1089                 return -EINVAL;
1090
1091         if (param->alternate_path &&
1092             (param->alternate_path->pkey != param->primary_path->pkey ||
1093              param->alternate_path->mtu != param->primary_path->mtu))
1094                 return -EINVAL;
1095
1096         return 0;
1097 }
1098
1099 int ib_send_cm_req(struct ib_cm_id *cm_id,
1100                    struct ib_cm_req_param *param)
1101 {
1102         struct cm_id_private *cm_id_priv;
1103         struct cm_req_msg *req_msg;
1104         unsigned long flags;
1105         int ret;
1106
1107         ret = cm_validate_req_param(param);
1108         if (ret)
1109                 return ret;
1110
1111         /* Verify that we're not in timewait. */
1112         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1113         spin_lock_irqsave(&cm_id_priv->lock, flags);
1114         if (cm_id->state != IB_CM_IDLE) {
1115                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1116                 ret = -EINVAL;
1117                 goto out;
1118         }
1119         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1120
1121         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1122                                                             id.local_id);
1123         if (IS_ERR(cm_id_priv->timewait_info)) {
1124                 ret = PTR_ERR(cm_id_priv->timewait_info);
1125                 goto out;
1126         }
1127
1128         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av);
1129         if (ret)
1130                 goto error1;
1131         if (param->alternate_path) {
1132                 ret = cm_init_av_by_path(param->alternate_path,
1133                                          &cm_id_priv->alt_av);
1134                 if (ret)
1135                         goto error1;
1136         }
1137         cm_id->service_id = param->service_id;
1138         cm_id->service_mask = ~cpu_to_be64(0);
1139         cm_id_priv->timeout_ms = cm_convert_to_ms(
1140                                     param->primary_path->packet_life_time) * 2 +
1141                                  cm_convert_to_ms(
1142                                     param->remote_cm_response_timeout);
1143         cm_id_priv->max_cm_retries = param->max_cm_retries;
1144         cm_id_priv->initiator_depth = param->initiator_depth;
1145         cm_id_priv->responder_resources = param->responder_resources;
1146         cm_id_priv->retry_count = param->retry_count;
1147         cm_id_priv->path_mtu = param->primary_path->mtu;
1148         cm_id_priv->pkey = param->primary_path->pkey;
1149         cm_id_priv->qp_type = param->qp_type;
1150
1151         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1152         if (ret)
1153                 goto error1;
1154
1155         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1156         cm_format_req(req_msg, cm_id_priv, param);
1157         cm_id_priv->tid = req_msg->hdr.tid;
1158         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1159         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1160
1161         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1162         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1163
1164         spin_lock_irqsave(&cm_id_priv->lock, flags);
1165         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1166         if (ret) {
1167                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1168                 goto error2;
1169         }
1170         BUG_ON(cm_id->state != IB_CM_IDLE);
1171         cm_id->state = IB_CM_REQ_SENT;
1172         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1173         return 0;
1174
1175 error2: cm_free_msg(cm_id_priv->msg);
1176 error1: kfree(cm_id_priv->timewait_info);
1177 out:    return ret;
1178 }
1179 EXPORT_SYMBOL(ib_send_cm_req);
1180
1181 static int cm_issue_rej(struct cm_port *port,
1182                         struct ib_mad_recv_wc *mad_recv_wc,
1183                         enum ib_cm_rej_reason reason,
1184                         enum cm_msg_response msg_rejected,
1185                         void *ari, u8 ari_length)
1186 {
1187         struct ib_mad_send_buf *msg = NULL;
1188         struct cm_rej_msg *rej_msg, *rcv_msg;
1189         int ret;
1190
1191         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1192         if (ret)
1193                 return ret;
1194
1195         /* We just need common CM header information.  Cast to any message. */
1196         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1197         rej_msg = (struct cm_rej_msg *) msg->mad;
1198
1199         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1200         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1201         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1202         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1203         rej_msg->reason = cpu_to_be16(reason);
1204
1205         if (ari && ari_length) {
1206                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1207                 memcpy(rej_msg->ari, ari, ari_length);
1208         }
1209
1210         ret = ib_post_send_mad(msg, NULL);
1211         if (ret)
1212                 cm_free_msg(msg);
1213
1214         return ret;
1215 }
1216
1217 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1218                                     __be32 local_qpn, __be32 remote_qpn)
1219 {
1220         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1221                 ((local_ca_guid == remote_ca_guid) &&
1222                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1223 }
1224
1225 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1226                                             struct ib_sa_path_rec *primary_path,
1227                                             struct ib_sa_path_rec *alt_path)
1228 {
1229         memset(primary_path, 0, sizeof *primary_path);
1230         primary_path->dgid = req_msg->primary_local_gid;
1231         primary_path->sgid = req_msg->primary_remote_gid;
1232         primary_path->dlid = req_msg->primary_local_lid;
1233         primary_path->slid = req_msg->primary_remote_lid;
1234         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1235         primary_path->hop_limit = req_msg->primary_hop_limit;
1236         primary_path->traffic_class = req_msg->primary_traffic_class;
1237         primary_path->reversible = 1;
1238         primary_path->pkey = req_msg->pkey;
1239         primary_path->sl = cm_req_get_primary_sl(req_msg);
1240         primary_path->mtu_selector = IB_SA_EQ;
1241         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1242         primary_path->rate_selector = IB_SA_EQ;
1243         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1244         primary_path->packet_life_time_selector = IB_SA_EQ;
1245         primary_path->packet_life_time =
1246                 cm_req_get_primary_local_ack_timeout(req_msg);
1247         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1248
1249         if (req_msg->alt_local_lid) {
1250                 memset(alt_path, 0, sizeof *alt_path);
1251                 alt_path->dgid = req_msg->alt_local_gid;
1252                 alt_path->sgid = req_msg->alt_remote_gid;
1253                 alt_path->dlid = req_msg->alt_local_lid;
1254                 alt_path->slid = req_msg->alt_remote_lid;
1255                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1256                 alt_path->hop_limit = req_msg->alt_hop_limit;
1257                 alt_path->traffic_class = req_msg->alt_traffic_class;
1258                 alt_path->reversible = 1;
1259                 alt_path->pkey = req_msg->pkey;
1260                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1261                 alt_path->mtu_selector = IB_SA_EQ;
1262                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1263                 alt_path->rate_selector = IB_SA_EQ;
1264                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1265                 alt_path->packet_life_time_selector = IB_SA_EQ;
1266                 alt_path->packet_life_time =
1267                         cm_req_get_alt_local_ack_timeout(req_msg);
1268                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1269         }
1270 }
1271
1272 static void cm_format_req_event(struct cm_work *work,
1273                                 struct cm_id_private *cm_id_priv,
1274                                 struct ib_cm_id *listen_id)
1275 {
1276         struct cm_req_msg *req_msg;
1277         struct ib_cm_req_event_param *param;
1278
1279         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1280         param = &work->cm_event.param.req_rcvd;
1281         param->listen_id = listen_id;
1282         param->port = cm_id_priv->av.port->port_num;
1283         param->primary_path = &work->path[0];
1284         if (req_msg->alt_local_lid)
1285                 param->alternate_path = &work->path[1];
1286         else
1287                 param->alternate_path = NULL;
1288         param->remote_ca_guid = req_msg->local_ca_guid;
1289         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1290         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1291         param->qp_type = cm_req_get_qp_type(req_msg);
1292         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1293         param->responder_resources = cm_req_get_init_depth(req_msg);
1294         param->initiator_depth = cm_req_get_resp_res(req_msg);
1295         param->local_cm_response_timeout =
1296                                         cm_req_get_remote_resp_timeout(req_msg);
1297         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1298         param->remote_cm_response_timeout =
1299                                         cm_req_get_local_resp_timeout(req_msg);
1300         param->retry_count = cm_req_get_retry_count(req_msg);
1301         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1302         param->srq = cm_req_get_srq(req_msg);
1303         work->cm_event.private_data = &req_msg->private_data;
1304 }
1305
1306 static void cm_process_work(struct cm_id_private *cm_id_priv,
1307                             struct cm_work *work)
1308 {
1309         int ret;
1310
1311         /* We will typically only have the current event to report. */
1312         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1313         cm_free_work(work);
1314
1315         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1316                 spin_lock_irq(&cm_id_priv->lock);
1317                 work = cm_dequeue_work(cm_id_priv);
1318                 spin_unlock_irq(&cm_id_priv->lock);
1319                 BUG_ON(!work);
1320                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1321                                                 &work->cm_event);
1322                 cm_free_work(work);
1323         }
1324         cm_deref_id(cm_id_priv);
1325         if (ret)
1326                 cm_destroy_id(&cm_id_priv->id, ret);
1327 }
1328
1329 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1330                           struct cm_id_private *cm_id_priv,
1331                           enum cm_msg_response msg_mraed, u8 service_timeout,
1332                           const void *private_data, u8 private_data_len)
1333 {
1334         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1335         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1336         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1337         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1338         cm_mra_set_service_timeout(mra_msg, service_timeout);
1339
1340         if (private_data && private_data_len)
1341                 memcpy(mra_msg->private_data, private_data, private_data_len);
1342 }
1343
1344 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1345                           struct cm_id_private *cm_id_priv,
1346                           enum ib_cm_rej_reason reason,
1347                           void *ari,
1348                           u8 ari_length,
1349                           const void *private_data,
1350                           u8 private_data_len)
1351 {
1352         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1353         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1354
1355         switch(cm_id_priv->id.state) {
1356         case IB_CM_REQ_RCVD:
1357                 rej_msg->local_comm_id = 0;
1358                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1359                 break;
1360         case IB_CM_MRA_REQ_SENT:
1361                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1362                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1363                 break;
1364         case IB_CM_REP_RCVD:
1365         case IB_CM_MRA_REP_SENT:
1366                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1367                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1368                 break;
1369         default:
1370                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1371                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1372                 break;
1373         }
1374
1375         rej_msg->reason = cpu_to_be16(reason);
1376         if (ari && ari_length) {
1377                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1378                 memcpy(rej_msg->ari, ari, ari_length);
1379         }
1380
1381         if (private_data && private_data_len)
1382                 memcpy(rej_msg->private_data, private_data, private_data_len);
1383 }
1384
1385 static void cm_dup_req_handler(struct cm_work *work,
1386                                struct cm_id_private *cm_id_priv)
1387 {
1388         struct ib_mad_send_buf *msg = NULL;
1389         int ret;
1390
1391         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1392                         counter[CM_REQ_COUNTER]);
1393
1394         /* Quick state check to discard duplicate REQs. */
1395         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1396                 return;
1397
1398         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1399         if (ret)
1400                 return;
1401
1402         spin_lock_irq(&cm_id_priv->lock);
1403         switch (cm_id_priv->id.state) {
1404         case IB_CM_MRA_REQ_SENT:
1405                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1406                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1407                               cm_id_priv->private_data,
1408                               cm_id_priv->private_data_len);
1409                 break;
1410         case IB_CM_TIMEWAIT:
1411                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1412                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1413                 break;
1414         default:
1415                 goto unlock;
1416         }
1417         spin_unlock_irq(&cm_id_priv->lock);
1418
1419         ret = ib_post_send_mad(msg, NULL);
1420         if (ret)
1421                 goto free;
1422         return;
1423
1424 unlock: spin_unlock_irq(&cm_id_priv->lock);
1425 free:   cm_free_msg(msg);
1426 }
1427
1428 static struct cm_id_private * cm_match_req(struct cm_work *work,
1429                                            struct cm_id_private *cm_id_priv)
1430 {
1431         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1432         struct cm_timewait_info *timewait_info;
1433         struct cm_req_msg *req_msg;
1434
1435         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1436
1437         /* Check for possible duplicate REQ. */
1438         spin_lock_irq(&cm.lock);
1439         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1440         if (timewait_info) {
1441                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1442                                            timewait_info->work.remote_id);
1443                 spin_unlock_irq(&cm.lock);
1444                 if (cur_cm_id_priv) {
1445                         cm_dup_req_handler(work, cur_cm_id_priv);
1446                         cm_deref_id(cur_cm_id_priv);
1447                 }
1448                 return NULL;
1449         }
1450
1451         /* Check for stale connections. */
1452         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1453         if (timewait_info) {
1454                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1455                 spin_unlock_irq(&cm.lock);
1456                 cm_issue_rej(work->port, work->mad_recv_wc,
1457                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1458                              NULL, 0);
1459                 return NULL;
1460         }
1461
1462         /* Find matching listen request. */
1463         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1464                                            req_msg->service_id,
1465                                            req_msg->private_data);
1466         if (!listen_cm_id_priv) {
1467                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1468                 spin_unlock_irq(&cm.lock);
1469                 cm_issue_rej(work->port, work->mad_recv_wc,
1470                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1471                              NULL, 0);
1472                 goto out;
1473         }
1474         atomic_inc(&listen_cm_id_priv->refcount);
1475         atomic_inc(&cm_id_priv->refcount);
1476         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1477         atomic_inc(&cm_id_priv->work_count);
1478         spin_unlock_irq(&cm.lock);
1479 out:
1480         return listen_cm_id_priv;
1481 }
1482
1483 /*
1484  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1485  * we need to override the LID/SL data in the REQ with the LID information
1486  * in the work completion.
1487  */
1488 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1489 {
1490         if (!cm_req_get_primary_subnet_local(req_msg)) {
1491                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1492                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1493                         cm_req_set_primary_sl(req_msg, wc->sl);
1494                 }
1495
1496                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1497                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1498         }
1499
1500         if (!cm_req_get_alt_subnet_local(req_msg)) {
1501                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1502                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1503                         cm_req_set_alt_sl(req_msg, wc->sl);
1504                 }
1505
1506                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1507                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1508         }
1509 }
1510
1511 static int cm_req_handler(struct cm_work *work)
1512 {
1513         struct ib_cm_id *cm_id;
1514         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1515         struct cm_req_msg *req_msg;
1516         int ret;
1517
1518         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1519
1520         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1521         if (IS_ERR(cm_id))
1522                 return PTR_ERR(cm_id);
1523
1524         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1525         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1526         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1527                                 work->mad_recv_wc->recv_buf.grh,
1528                                 &cm_id_priv->av);
1529         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1530                                                             id.local_id);
1531         if (IS_ERR(cm_id_priv->timewait_info)) {
1532                 ret = PTR_ERR(cm_id_priv->timewait_info);
1533                 goto destroy;
1534         }
1535         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1536         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1537         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1538
1539         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1540         if (!listen_cm_id_priv) {
1541                 ret = -EINVAL;
1542                 kfree(cm_id_priv->timewait_info);
1543                 goto destroy;
1544         }
1545
1546         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1547         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1548         cm_id_priv->id.service_id = req_msg->service_id;
1549         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1550
1551         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1552         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1553         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
1554         if (ret) {
1555                 ib_get_cached_gid(work->port->cm_dev->ib_device,
1556                                   work->port->port_num, 0, &work->path[0].sgid);
1557                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1558                                &work->path[0].sgid, sizeof work->path[0].sgid,
1559                                NULL, 0);
1560                 goto rejected;
1561         }
1562         if (req_msg->alt_local_lid) {
1563                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
1564                 if (ret) {
1565                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1566                                        &work->path[0].sgid,
1567                                        sizeof work->path[0].sgid, NULL, 0);
1568                         goto rejected;
1569                 }
1570         }
1571         cm_id_priv->tid = req_msg->hdr.tid;
1572         cm_id_priv->timeout_ms = cm_convert_to_ms(
1573                                         cm_req_get_local_resp_timeout(req_msg));
1574         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1575         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1576         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1577         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1578         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1579         cm_id_priv->pkey = req_msg->pkey;
1580         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1581         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1582         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1583         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1584
1585         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1586         cm_process_work(cm_id_priv, work);
1587         cm_deref_id(listen_cm_id_priv);
1588         return 0;
1589
1590 rejected:
1591         atomic_dec(&cm_id_priv->refcount);
1592         cm_deref_id(listen_cm_id_priv);
1593 destroy:
1594         ib_destroy_cm_id(cm_id);
1595         return ret;
1596 }
1597
1598 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1599                           struct cm_id_private *cm_id_priv,
1600                           struct ib_cm_rep_param *param)
1601 {
1602         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1603         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1604         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1605         cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1606         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1607         rep_msg->resp_resources = param->responder_resources;
1608         rep_msg->initiator_depth = param->initiator_depth;
1609         cm_rep_set_target_ack_delay(rep_msg,
1610                                     cm_id_priv->av.port->cm_dev->ack_delay);
1611         cm_rep_set_failover(rep_msg, param->failover_accepted);
1612         cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1613         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1614         cm_rep_set_srq(rep_msg, param->srq);
1615         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1616
1617         if (param->private_data && param->private_data_len)
1618                 memcpy(rep_msg->private_data, param->private_data,
1619                        param->private_data_len);
1620 }
1621
1622 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1623                    struct ib_cm_rep_param *param)
1624 {
1625         struct cm_id_private *cm_id_priv;
1626         struct ib_mad_send_buf *msg;
1627         struct cm_rep_msg *rep_msg;
1628         unsigned long flags;
1629         int ret;
1630
1631         if (param->private_data &&
1632             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1633                 return -EINVAL;
1634
1635         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1636         spin_lock_irqsave(&cm_id_priv->lock, flags);
1637         if (cm_id->state != IB_CM_REQ_RCVD &&
1638             cm_id->state != IB_CM_MRA_REQ_SENT) {
1639                 ret = -EINVAL;
1640                 goto out;
1641         }
1642
1643         ret = cm_alloc_msg(cm_id_priv, &msg);
1644         if (ret)
1645                 goto out;
1646
1647         rep_msg = (struct cm_rep_msg *) msg->mad;
1648         cm_format_rep(rep_msg, cm_id_priv, param);
1649         msg->timeout_ms = cm_id_priv->timeout_ms;
1650         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1651
1652         ret = ib_post_send_mad(msg, NULL);
1653         if (ret) {
1654                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1655                 cm_free_msg(msg);
1656                 return ret;
1657         }
1658
1659         cm_id->state = IB_CM_REP_SENT;
1660         cm_id_priv->msg = msg;
1661         cm_id_priv->initiator_depth = param->initiator_depth;
1662         cm_id_priv->responder_resources = param->responder_resources;
1663         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1664         cm_id_priv->local_qpn = cm_rep_get_local_qpn(rep_msg);
1665
1666 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1667         return ret;
1668 }
1669 EXPORT_SYMBOL(ib_send_cm_rep);
1670
1671 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1672                           struct cm_id_private *cm_id_priv,
1673                           const void *private_data,
1674                           u8 private_data_len)
1675 {
1676         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1677         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1678         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1679
1680         if (private_data && private_data_len)
1681                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1682 }
1683
1684 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1685                    const void *private_data,
1686                    u8 private_data_len)
1687 {
1688         struct cm_id_private *cm_id_priv;
1689         struct ib_mad_send_buf *msg;
1690         unsigned long flags;
1691         void *data;
1692         int ret;
1693
1694         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1695                 return -EINVAL;
1696
1697         data = cm_copy_private_data(private_data, private_data_len);
1698         if (IS_ERR(data))
1699                 return PTR_ERR(data);
1700
1701         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1702         spin_lock_irqsave(&cm_id_priv->lock, flags);
1703         if (cm_id->state != IB_CM_REP_RCVD &&
1704             cm_id->state != IB_CM_MRA_REP_SENT) {
1705                 ret = -EINVAL;
1706                 goto error;
1707         }
1708
1709         ret = cm_alloc_msg(cm_id_priv, &msg);
1710         if (ret)
1711                 goto error;
1712
1713         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1714                       private_data, private_data_len);
1715
1716         ret = ib_post_send_mad(msg, NULL);
1717         if (ret) {
1718                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1719                 cm_free_msg(msg);
1720                 kfree(data);
1721                 return ret;
1722         }
1723
1724         cm_id->state = IB_CM_ESTABLISHED;
1725         cm_set_private_data(cm_id_priv, data, private_data_len);
1726         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1727         return 0;
1728
1729 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1730         kfree(data);
1731         return ret;
1732 }
1733 EXPORT_SYMBOL(ib_send_cm_rtu);
1734
1735 static void cm_format_rep_event(struct cm_work *work)
1736 {
1737         struct cm_rep_msg *rep_msg;
1738         struct ib_cm_rep_event_param *param;
1739
1740         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1741         param = &work->cm_event.param.rep_rcvd;
1742         param->remote_ca_guid = rep_msg->local_ca_guid;
1743         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1744         param->remote_qpn = be32_to_cpu(cm_rep_get_local_qpn(rep_msg));
1745         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1746         param->responder_resources = rep_msg->initiator_depth;
1747         param->initiator_depth = rep_msg->resp_resources;
1748         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1749         param->failover_accepted = cm_rep_get_failover(rep_msg);
1750         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1751         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1752         param->srq = cm_rep_get_srq(rep_msg);
1753         work->cm_event.private_data = &rep_msg->private_data;
1754 }
1755
1756 static void cm_dup_rep_handler(struct cm_work *work)
1757 {
1758         struct cm_id_private *cm_id_priv;
1759         struct cm_rep_msg *rep_msg;
1760         struct ib_mad_send_buf *msg = NULL;
1761         int ret;
1762
1763         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1764         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1765                                    rep_msg->local_comm_id);
1766         if (!cm_id_priv)
1767                 return;
1768
1769         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1770                         counter[CM_REP_COUNTER]);
1771         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1772         if (ret)
1773                 goto deref;
1774
1775         spin_lock_irq(&cm_id_priv->lock);
1776         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1777                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1778                               cm_id_priv->private_data,
1779                               cm_id_priv->private_data_len);
1780         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1781                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1782                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1783                               cm_id_priv->private_data,
1784                               cm_id_priv->private_data_len);
1785         else
1786                 goto unlock;
1787         spin_unlock_irq(&cm_id_priv->lock);
1788
1789         ret = ib_post_send_mad(msg, NULL);
1790         if (ret)
1791                 goto free;
1792         goto deref;
1793
1794 unlock: spin_unlock_irq(&cm_id_priv->lock);
1795 free:   cm_free_msg(msg);
1796 deref:  cm_deref_id(cm_id_priv);
1797 }
1798
1799 static int cm_rep_handler(struct cm_work *work)
1800 {
1801         struct cm_id_private *cm_id_priv;
1802         struct cm_rep_msg *rep_msg;
1803         int ret;
1804
1805         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1806         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1807         if (!cm_id_priv) {
1808                 cm_dup_rep_handler(work);
1809                 return -EINVAL;
1810         }
1811
1812         cm_format_rep_event(work);
1813
1814         spin_lock_irq(&cm_id_priv->lock);
1815         switch (cm_id_priv->id.state) {
1816         case IB_CM_REQ_SENT:
1817         case IB_CM_MRA_REQ_RCVD:
1818                 break;
1819         default:
1820                 spin_unlock_irq(&cm_id_priv->lock);
1821                 ret = -EINVAL;
1822                 goto error;
1823         }
1824
1825         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1826         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1827         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1828
1829         spin_lock(&cm.lock);
1830         /* Check for duplicate REP. */
1831         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1832                 spin_unlock(&cm.lock);
1833                 spin_unlock_irq(&cm_id_priv->lock);
1834                 ret = -EINVAL;
1835                 goto error;
1836         }
1837         /* Check for a stale connection. */
1838         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
1839                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
1840                          &cm.remote_id_table);
1841                 cm_id_priv->timewait_info->inserted_remote_id = 0;
1842                 spin_unlock(&cm.lock);
1843                 spin_unlock_irq(&cm_id_priv->lock);
1844                 cm_issue_rej(work->port, work->mad_recv_wc,
1845                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
1846                              NULL, 0);
1847                 ret = -EINVAL;
1848                 goto error;
1849         }
1850         spin_unlock(&cm.lock);
1851
1852         cm_id_priv->id.state = IB_CM_REP_RCVD;
1853         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
1854         cm_id_priv->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1855         cm_id_priv->initiator_depth = rep_msg->resp_resources;
1856         cm_id_priv->responder_resources = rep_msg->initiator_depth;
1857         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
1858         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1859         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1860         cm_id_priv->av.timeout =
1861                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1862                                        cm_id_priv->av.timeout - 1);
1863         cm_id_priv->alt_av.timeout =
1864                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1865                                        cm_id_priv->alt_av.timeout - 1);
1866
1867         /* todo: handle peer_to_peer */
1868
1869         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1870         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1871         if (!ret)
1872                 list_add_tail(&work->list, &cm_id_priv->work_list);
1873         spin_unlock_irq(&cm_id_priv->lock);
1874
1875         if (ret)
1876                 cm_process_work(cm_id_priv, work);
1877         else
1878                 cm_deref_id(cm_id_priv);
1879         return 0;
1880
1881 error:
1882         cm_deref_id(cm_id_priv);
1883         return ret;
1884 }
1885
1886 static int cm_establish_handler(struct cm_work *work)
1887 {
1888         struct cm_id_private *cm_id_priv;
1889         int ret;
1890
1891         /* See comment in cm_establish about lookup. */
1892         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
1893         if (!cm_id_priv)
1894                 return -EINVAL;
1895
1896         spin_lock_irq(&cm_id_priv->lock);
1897         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
1898                 spin_unlock_irq(&cm_id_priv->lock);
1899                 goto out;
1900         }
1901
1902         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1903         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1904         if (!ret)
1905                 list_add_tail(&work->list, &cm_id_priv->work_list);
1906         spin_unlock_irq(&cm_id_priv->lock);
1907
1908         if (ret)
1909                 cm_process_work(cm_id_priv, work);
1910         else
1911                 cm_deref_id(cm_id_priv);
1912         return 0;
1913 out:
1914         cm_deref_id(cm_id_priv);
1915         return -EINVAL;
1916 }
1917
1918 static int cm_rtu_handler(struct cm_work *work)
1919 {
1920         struct cm_id_private *cm_id_priv;
1921         struct cm_rtu_msg *rtu_msg;
1922         int ret;
1923
1924         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
1925         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
1926                                    rtu_msg->local_comm_id);
1927         if (!cm_id_priv)
1928                 return -EINVAL;
1929
1930         work->cm_event.private_data = &rtu_msg->private_data;
1931
1932         spin_lock_irq(&cm_id_priv->lock);
1933         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
1934             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
1935                 spin_unlock_irq(&cm_id_priv->lock);
1936                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1937                                 counter[CM_RTU_COUNTER]);
1938                 goto out;
1939         }
1940         cm_id_priv->id.state = IB_CM_ESTABLISHED;
1941
1942         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1943         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1944         if (!ret)
1945                 list_add_tail(&work->list, &cm_id_priv->work_list);
1946         spin_unlock_irq(&cm_id_priv->lock);
1947
1948         if (ret)
1949                 cm_process_work(cm_id_priv, work);
1950         else
1951                 cm_deref_id(cm_id_priv);
1952         return 0;
1953 out:
1954         cm_deref_id(cm_id_priv);
1955         return -EINVAL;
1956 }
1957
1958 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
1959                           struct cm_id_private *cm_id_priv,
1960                           const void *private_data,
1961                           u8 private_data_len)
1962 {
1963         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
1964                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
1965         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
1966         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
1967         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
1968
1969         if (private_data && private_data_len)
1970                 memcpy(dreq_msg->private_data, private_data, private_data_len);
1971 }
1972
1973 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
1974                     const void *private_data,
1975                     u8 private_data_len)
1976 {
1977         struct cm_id_private *cm_id_priv;
1978         struct ib_mad_send_buf *msg;
1979         unsigned long flags;
1980         int ret;
1981
1982         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
1983                 return -EINVAL;
1984
1985         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1986         spin_lock_irqsave(&cm_id_priv->lock, flags);
1987         if (cm_id->state != IB_CM_ESTABLISHED) {
1988                 ret = -EINVAL;
1989                 goto out;
1990         }
1991
1992         if (cm_id->lap_state == IB_CM_LAP_SENT ||
1993             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
1994                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1995
1996         ret = cm_alloc_msg(cm_id_priv, &msg);
1997         if (ret) {
1998                 cm_enter_timewait(cm_id_priv);
1999                 goto out;
2000         }
2001
2002         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2003                        private_data, private_data_len);
2004         msg->timeout_ms = cm_id_priv->timeout_ms;
2005         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2006
2007         ret = ib_post_send_mad(msg, NULL);
2008         if (ret) {
2009                 cm_enter_timewait(cm_id_priv);
2010                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2011                 cm_free_msg(msg);
2012                 return ret;
2013         }
2014
2015         cm_id->state = IB_CM_DREQ_SENT;
2016         cm_id_priv->msg = msg;
2017 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2018         return ret;
2019 }
2020 EXPORT_SYMBOL(ib_send_cm_dreq);
2021
2022 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2023                           struct cm_id_private *cm_id_priv,
2024                           const void *private_data,
2025                           u8 private_data_len)
2026 {
2027         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2028         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2029         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2030
2031         if (private_data && private_data_len)
2032                 memcpy(drep_msg->private_data, private_data, private_data_len);
2033 }
2034
2035 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2036                     const void *private_data,
2037                     u8 private_data_len)
2038 {
2039         struct cm_id_private *cm_id_priv;
2040         struct ib_mad_send_buf *msg;
2041         unsigned long flags;
2042         void *data;
2043         int ret;
2044
2045         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2046                 return -EINVAL;
2047
2048         data = cm_copy_private_data(private_data, private_data_len);
2049         if (IS_ERR(data))
2050                 return PTR_ERR(data);
2051
2052         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2053         spin_lock_irqsave(&cm_id_priv->lock, flags);
2054         if (cm_id->state != IB_CM_DREQ_RCVD) {
2055                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2056                 kfree(data);
2057                 return -EINVAL;
2058         }
2059
2060         cm_set_private_data(cm_id_priv, data, private_data_len);
2061         cm_enter_timewait(cm_id_priv);
2062
2063         ret = cm_alloc_msg(cm_id_priv, &msg);
2064         if (ret)
2065                 goto out;
2066
2067         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2068                        private_data, private_data_len);
2069
2070         ret = ib_post_send_mad(msg, NULL);
2071         if (ret) {
2072                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2073                 cm_free_msg(msg);
2074                 return ret;
2075         }
2076
2077 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2078         return ret;
2079 }
2080 EXPORT_SYMBOL(ib_send_cm_drep);
2081
2082 static int cm_issue_drep(struct cm_port *port,
2083                          struct ib_mad_recv_wc *mad_recv_wc)
2084 {
2085         struct ib_mad_send_buf *msg = NULL;
2086         struct cm_dreq_msg *dreq_msg;
2087         struct cm_drep_msg *drep_msg;
2088         int ret;
2089
2090         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2091         if (ret)
2092                 return ret;
2093
2094         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2095         drep_msg = (struct cm_drep_msg *) msg->mad;
2096
2097         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2098         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2099         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2100
2101         ret = ib_post_send_mad(msg, NULL);
2102         if (ret)
2103                 cm_free_msg(msg);
2104
2105         return ret;
2106 }
2107
2108 static int cm_dreq_handler(struct cm_work *work)
2109 {
2110         struct cm_id_private *cm_id_priv;
2111         struct cm_dreq_msg *dreq_msg;
2112         struct ib_mad_send_buf *msg = NULL;
2113         int ret;
2114
2115         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2116         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2117                                    dreq_msg->local_comm_id);
2118         if (!cm_id_priv) {
2119                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2120                                 counter[CM_DREQ_COUNTER]);
2121                 cm_issue_drep(work->port, work->mad_recv_wc);
2122                 return -EINVAL;
2123         }
2124
2125         work->cm_event.private_data = &dreq_msg->private_data;
2126
2127         spin_lock_irq(&cm_id_priv->lock);
2128         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2129                 goto unlock;
2130
2131         switch (cm_id_priv->id.state) {
2132         case IB_CM_REP_SENT:
2133         case IB_CM_DREQ_SENT:
2134                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2135                 break;
2136         case IB_CM_ESTABLISHED:
2137                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2138                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2139                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2140                 break;
2141         case IB_CM_MRA_REP_RCVD:
2142                 break;
2143         case IB_CM_TIMEWAIT:
2144                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2145                                 counter[CM_DREQ_COUNTER]);
2146                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2147                         goto unlock;
2148
2149                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2150                                cm_id_priv->private_data,
2151                                cm_id_priv->private_data_len);
2152                 spin_unlock_irq(&cm_id_priv->lock);
2153
2154                 if (ib_post_send_mad(msg, NULL))
2155                         cm_free_msg(msg);
2156                 goto deref;
2157         case IB_CM_DREQ_RCVD:
2158                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2159                                 counter[CM_DREQ_COUNTER]);
2160                 goto unlock;
2161         default:
2162                 goto unlock;
2163         }
2164         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2165         cm_id_priv->tid = dreq_msg->hdr.tid;
2166         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2167         if (!ret)
2168                 list_add_tail(&work->list, &cm_id_priv->work_list);
2169         spin_unlock_irq(&cm_id_priv->lock);
2170
2171         if (ret)
2172                 cm_process_work(cm_id_priv, work);
2173         else
2174                 cm_deref_id(cm_id_priv);
2175         return 0;
2176
2177 unlock: spin_unlock_irq(&cm_id_priv->lock);
2178 deref:  cm_deref_id(cm_id_priv);
2179         return -EINVAL;
2180 }
2181
2182 static int cm_drep_handler(struct cm_work *work)
2183 {
2184         struct cm_id_private *cm_id_priv;
2185         struct cm_drep_msg *drep_msg;
2186         int ret;
2187
2188         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2189         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2190                                    drep_msg->local_comm_id);
2191         if (!cm_id_priv)
2192                 return -EINVAL;
2193
2194         work->cm_event.private_data = &drep_msg->private_data;
2195
2196         spin_lock_irq(&cm_id_priv->lock);
2197         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2198             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2199                 spin_unlock_irq(&cm_id_priv->lock);
2200                 goto out;
2201         }
2202         cm_enter_timewait(cm_id_priv);
2203
2204         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2205         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2206         if (!ret)
2207                 list_add_tail(&work->list, &cm_id_priv->work_list);
2208         spin_unlock_irq(&cm_id_priv->lock);
2209
2210         if (ret)
2211                 cm_process_work(cm_id_priv, work);
2212         else
2213                 cm_deref_id(cm_id_priv);
2214         return 0;
2215 out:
2216         cm_deref_id(cm_id_priv);
2217         return -EINVAL;
2218 }
2219
2220 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2221                    enum ib_cm_rej_reason reason,
2222                    void *ari,
2223                    u8 ari_length,
2224                    const void *private_data,
2225                    u8 private_data_len)
2226 {
2227         struct cm_id_private *cm_id_priv;
2228         struct ib_mad_send_buf *msg;
2229         unsigned long flags;
2230         int ret;
2231
2232         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2233             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2234                 return -EINVAL;
2235
2236         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2237
2238         spin_lock_irqsave(&cm_id_priv->lock, flags);
2239         switch (cm_id->state) {
2240         case IB_CM_REQ_SENT:
2241         case IB_CM_MRA_REQ_RCVD:
2242         case IB_CM_REQ_RCVD:
2243         case IB_CM_MRA_REQ_SENT:
2244         case IB_CM_REP_RCVD:
2245         case IB_CM_MRA_REP_SENT:
2246                 ret = cm_alloc_msg(cm_id_priv, &msg);
2247                 if (!ret)
2248                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2249                                       cm_id_priv, reason, ari, ari_length,
2250                                       private_data, private_data_len);
2251
2252                 cm_reset_to_idle(cm_id_priv);
2253                 break;
2254         case IB_CM_REP_SENT:
2255         case IB_CM_MRA_REP_RCVD:
2256                 ret = cm_alloc_msg(cm_id_priv, &msg);
2257                 if (!ret)
2258                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2259                                       cm_id_priv, reason, ari, ari_length,
2260                                       private_data, private_data_len);
2261
2262                 cm_enter_timewait(cm_id_priv);
2263                 break;
2264         default:
2265                 ret = -EINVAL;
2266                 goto out;
2267         }
2268
2269         if (ret)
2270                 goto out;
2271
2272         ret = ib_post_send_mad(msg, NULL);
2273         if (ret)
2274                 cm_free_msg(msg);
2275
2276 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2277         return ret;
2278 }
2279 EXPORT_SYMBOL(ib_send_cm_rej);
2280
2281 static void cm_format_rej_event(struct cm_work *work)
2282 {
2283         struct cm_rej_msg *rej_msg;
2284         struct ib_cm_rej_event_param *param;
2285
2286         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2287         param = &work->cm_event.param.rej_rcvd;
2288         param->ari = rej_msg->ari;
2289         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2290         param->reason = __be16_to_cpu(rej_msg->reason);
2291         work->cm_event.private_data = &rej_msg->private_data;
2292 }
2293
2294 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2295 {
2296         struct cm_timewait_info *timewait_info;
2297         struct cm_id_private *cm_id_priv;
2298         __be32 remote_id;
2299
2300         remote_id = rej_msg->local_comm_id;
2301
2302         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2303                 spin_lock_irq(&cm.lock);
2304                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2305                                                   remote_id);
2306                 if (!timewait_info) {
2307                         spin_unlock_irq(&cm.lock);
2308                         return NULL;
2309                 }
2310                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2311                                       (timewait_info->work.local_id ^
2312                                        cm.random_id_operand));
2313                 if (cm_id_priv) {
2314                         if (cm_id_priv->id.remote_id == remote_id)
2315                                 atomic_inc(&cm_id_priv->refcount);
2316                         else
2317                                 cm_id_priv = NULL;
2318                 }
2319                 spin_unlock_irq(&cm.lock);
2320         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2321                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2322         else
2323                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2324
2325         return cm_id_priv;
2326 }
2327
2328 static int cm_rej_handler(struct cm_work *work)
2329 {
2330         struct cm_id_private *cm_id_priv;
2331         struct cm_rej_msg *rej_msg;
2332         int ret;
2333
2334         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2335         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2336         if (!cm_id_priv)
2337                 return -EINVAL;
2338
2339         cm_format_rej_event(work);
2340
2341         spin_lock_irq(&cm_id_priv->lock);
2342         switch (cm_id_priv->id.state) {
2343         case IB_CM_REQ_SENT:
2344         case IB_CM_MRA_REQ_RCVD:
2345         case IB_CM_REP_SENT:
2346         case IB_CM_MRA_REP_RCVD:
2347                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2348                 /* fall through */
2349         case IB_CM_REQ_RCVD:
2350         case IB_CM_MRA_REQ_SENT:
2351                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2352                         cm_enter_timewait(cm_id_priv);
2353                 else
2354                         cm_reset_to_idle(cm_id_priv);
2355                 break;
2356         case IB_CM_DREQ_SENT:
2357                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2358                 /* fall through */
2359         case IB_CM_REP_RCVD:
2360         case IB_CM_MRA_REP_SENT:
2361                 cm_enter_timewait(cm_id_priv);
2362                 break;
2363         case IB_CM_ESTABLISHED:
2364                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2365                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2366                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2367                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2368                                               cm_id_priv->msg);
2369                         cm_enter_timewait(cm_id_priv);
2370                         break;
2371                 }
2372                 /* fall through */
2373         default:
2374                 spin_unlock_irq(&cm_id_priv->lock);
2375                 ret = -EINVAL;
2376                 goto out;
2377         }
2378
2379         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2380         if (!ret)
2381                 list_add_tail(&work->list, &cm_id_priv->work_list);
2382         spin_unlock_irq(&cm_id_priv->lock);
2383
2384         if (ret)
2385                 cm_process_work(cm_id_priv, work);
2386         else
2387                 cm_deref_id(cm_id_priv);
2388         return 0;
2389 out:
2390         cm_deref_id(cm_id_priv);
2391         return -EINVAL;
2392 }
2393
2394 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2395                    u8 service_timeout,
2396                    const void *private_data,
2397                    u8 private_data_len)
2398 {
2399         struct cm_id_private *cm_id_priv;
2400         struct ib_mad_send_buf *msg;
2401         enum ib_cm_state cm_state;
2402         enum ib_cm_lap_state lap_state;
2403         enum cm_msg_response msg_response;
2404         void *data;
2405         unsigned long flags;
2406         int ret;
2407
2408         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2409                 return -EINVAL;
2410
2411         data = cm_copy_private_data(private_data, private_data_len);
2412         if (IS_ERR(data))
2413                 return PTR_ERR(data);
2414
2415         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2416
2417         spin_lock_irqsave(&cm_id_priv->lock, flags);
2418         switch(cm_id_priv->id.state) {
2419         case IB_CM_REQ_RCVD:
2420                 cm_state = IB_CM_MRA_REQ_SENT;
2421                 lap_state = cm_id->lap_state;
2422                 msg_response = CM_MSG_RESPONSE_REQ;
2423                 break;
2424         case IB_CM_REP_RCVD:
2425                 cm_state = IB_CM_MRA_REP_SENT;
2426                 lap_state = cm_id->lap_state;
2427                 msg_response = CM_MSG_RESPONSE_REP;
2428                 break;
2429         case IB_CM_ESTABLISHED:
2430                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2431                         cm_state = cm_id->state;
2432                         lap_state = IB_CM_MRA_LAP_SENT;
2433                         msg_response = CM_MSG_RESPONSE_OTHER;
2434                         break;
2435                 }
2436         default:
2437                 ret = -EINVAL;
2438                 goto error1;
2439         }
2440
2441         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2442                 ret = cm_alloc_msg(cm_id_priv, &msg);
2443                 if (ret)
2444                         goto error1;
2445
2446                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2447                               msg_response, service_timeout,
2448                               private_data, private_data_len);
2449                 ret = ib_post_send_mad(msg, NULL);
2450                 if (ret)
2451                         goto error2;
2452         }
2453
2454         cm_id->state = cm_state;
2455         cm_id->lap_state = lap_state;
2456         cm_id_priv->service_timeout = service_timeout;
2457         cm_set_private_data(cm_id_priv, data, private_data_len);
2458         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2459         return 0;
2460
2461 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2462         kfree(data);
2463         return ret;
2464
2465 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2466         kfree(data);
2467         cm_free_msg(msg);
2468         return ret;
2469 }
2470 EXPORT_SYMBOL(ib_send_cm_mra);
2471
2472 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2473 {
2474         switch (cm_mra_get_msg_mraed(mra_msg)) {
2475         case CM_MSG_RESPONSE_REQ:
2476                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2477         case CM_MSG_RESPONSE_REP:
2478         case CM_MSG_RESPONSE_OTHER:
2479                 return cm_acquire_id(mra_msg->remote_comm_id,
2480                                      mra_msg->local_comm_id);
2481         default:
2482                 return NULL;
2483         }
2484 }
2485
2486 static int cm_mra_handler(struct cm_work *work)
2487 {
2488         struct cm_id_private *cm_id_priv;
2489         struct cm_mra_msg *mra_msg;
2490         int timeout, ret;
2491
2492         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2493         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2494         if (!cm_id_priv)
2495                 return -EINVAL;
2496
2497         work->cm_event.private_data = &mra_msg->private_data;
2498         work->cm_event.param.mra_rcvd.service_timeout =
2499                                         cm_mra_get_service_timeout(mra_msg);
2500         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2501                   cm_convert_to_ms(cm_id_priv->av.timeout);
2502
2503         spin_lock_irq(&cm_id_priv->lock);
2504         switch (cm_id_priv->id.state) {
2505         case IB_CM_REQ_SENT:
2506                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2507                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2508                                   cm_id_priv->msg, timeout))
2509                         goto out;
2510                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2511                 break;
2512         case IB_CM_REP_SENT:
2513                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2514                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2515                                   cm_id_priv->msg, timeout))
2516                         goto out;
2517                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2518                 break;
2519         case IB_CM_ESTABLISHED:
2520                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2521                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2522                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2523                                   cm_id_priv->msg, timeout)) {
2524                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2525                                 atomic_long_inc(&work->port->
2526                                                 counter_group[CM_RECV_DUPLICATES].
2527                                                 counter[CM_MRA_COUNTER]);
2528                         goto out;
2529                 }
2530                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2531                 break;
2532         case IB_CM_MRA_REQ_RCVD:
2533         case IB_CM_MRA_REP_RCVD:
2534                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2535                                 counter[CM_MRA_COUNTER]);
2536                 /* fall through */
2537         default:
2538                 goto out;
2539         }
2540
2541         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2542                                       cm_id_priv->id.state;
2543         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2544         if (!ret)
2545                 list_add_tail(&work->list, &cm_id_priv->work_list);
2546         spin_unlock_irq(&cm_id_priv->lock);
2547
2548         if (ret)
2549                 cm_process_work(cm_id_priv, work);
2550         else
2551                 cm_deref_id(cm_id_priv);
2552         return 0;
2553 out:
2554         spin_unlock_irq(&cm_id_priv->lock);
2555         cm_deref_id(cm_id_priv);
2556         return -EINVAL;
2557 }
2558
2559 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2560                           struct cm_id_private *cm_id_priv,
2561                           struct ib_sa_path_rec *alternate_path,
2562                           const void *private_data,
2563                           u8 private_data_len)
2564 {
2565         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2566                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2567         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2568         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2569         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2570         /* todo: need remote CM response timeout */
2571         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2572         lap_msg->alt_local_lid = alternate_path->slid;
2573         lap_msg->alt_remote_lid = alternate_path->dlid;
2574         lap_msg->alt_local_gid = alternate_path->sgid;
2575         lap_msg->alt_remote_gid = alternate_path->dgid;
2576         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2577         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2578         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2579         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2580         cm_lap_set_sl(lap_msg, alternate_path->sl);
2581         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2582         cm_lap_set_local_ack_timeout(lap_msg,
2583                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2584                                alternate_path->packet_life_time));
2585
2586         if (private_data && private_data_len)
2587                 memcpy(lap_msg->private_data, private_data, private_data_len);
2588 }
2589
2590 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2591                    struct ib_sa_path_rec *alternate_path,
2592                    const void *private_data,
2593                    u8 private_data_len)
2594 {
2595         struct cm_id_private *cm_id_priv;
2596         struct ib_mad_send_buf *msg;
2597         unsigned long flags;
2598         int ret;
2599
2600         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2601                 return -EINVAL;
2602
2603         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2604         spin_lock_irqsave(&cm_id_priv->lock, flags);
2605         if (cm_id->state != IB_CM_ESTABLISHED ||
2606             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2607              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2608                 ret = -EINVAL;
2609                 goto out;
2610         }
2611
2612         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
2613         if (ret)
2614                 goto out;
2615         cm_id_priv->alt_av.timeout =
2616                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2617                                        cm_id_priv->alt_av.timeout - 1);
2618
2619         ret = cm_alloc_msg(cm_id_priv, &msg);
2620         if (ret)
2621                 goto out;
2622
2623         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2624                       alternate_path, private_data, private_data_len);
2625         msg->timeout_ms = cm_id_priv->timeout_ms;
2626         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2627
2628         ret = ib_post_send_mad(msg, NULL);
2629         if (ret) {
2630                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2631                 cm_free_msg(msg);
2632                 return ret;
2633         }
2634
2635         cm_id->lap_state = IB_CM_LAP_SENT;
2636         cm_id_priv->msg = msg;
2637
2638 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2639         return ret;
2640 }
2641 EXPORT_SYMBOL(ib_send_cm_lap);
2642
2643 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2644                                     struct ib_sa_path_rec *path,
2645                                     struct cm_lap_msg *lap_msg)
2646 {
2647         memset(path, 0, sizeof *path);
2648         path->dgid = lap_msg->alt_local_gid;
2649         path->sgid = lap_msg->alt_remote_gid;
2650         path->dlid = lap_msg->alt_local_lid;
2651         path->slid = lap_msg->alt_remote_lid;
2652         path->flow_label = cm_lap_get_flow_label(lap_msg);
2653         path->hop_limit = lap_msg->alt_hop_limit;
2654         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2655         path->reversible = 1;
2656         path->pkey = cm_id_priv->pkey;
2657         path->sl = cm_lap_get_sl(lap_msg);
2658         path->mtu_selector = IB_SA_EQ;
2659         path->mtu = cm_id_priv->path_mtu;
2660         path->rate_selector = IB_SA_EQ;
2661         path->rate = cm_lap_get_packet_rate(lap_msg);
2662         path->packet_life_time_selector = IB_SA_EQ;
2663         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2664         path->packet_life_time -= (path->packet_life_time > 0);
2665 }
2666
2667 static int cm_lap_handler(struct cm_work *work)
2668 {
2669         struct cm_id_private *cm_id_priv;
2670         struct cm_lap_msg *lap_msg;
2671         struct ib_cm_lap_event_param *param;
2672         struct ib_mad_send_buf *msg = NULL;
2673         int ret;
2674
2675         /* todo: verify LAP request and send reject APR if invalid. */
2676         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2677         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2678                                    lap_msg->local_comm_id);
2679         if (!cm_id_priv)
2680                 return -EINVAL;
2681
2682         param = &work->cm_event.param.lap_rcvd;
2683         param->alternate_path = &work->path[0];
2684         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2685         work->cm_event.private_data = &lap_msg->private_data;
2686
2687         spin_lock_irq(&cm_id_priv->lock);
2688         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2689                 goto unlock;
2690
2691         switch (cm_id_priv->id.lap_state) {
2692         case IB_CM_LAP_UNINIT:
2693         case IB_CM_LAP_IDLE:
2694                 break;
2695         case IB_CM_MRA_LAP_SENT:
2696                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2697                                 counter[CM_LAP_COUNTER]);
2698                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2699                         goto unlock;
2700
2701                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2702                               CM_MSG_RESPONSE_OTHER,
2703                               cm_id_priv->service_timeout,
2704                               cm_id_priv->private_data,
2705                               cm_id_priv->private_data_len);
2706                 spin_unlock_irq(&cm_id_priv->lock);
2707
2708                 if (ib_post_send_mad(msg, NULL))
2709                         cm_free_msg(msg);
2710                 goto deref;
2711         case IB_CM_LAP_RCVD:
2712                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2713                                 counter[CM_LAP_COUNTER]);
2714                 goto unlock;
2715         default:
2716                 goto unlock;
2717         }
2718
2719         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2720         cm_id_priv->tid = lap_msg->hdr.tid;
2721         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2722                                 work->mad_recv_wc->recv_buf.grh,
2723                                 &cm_id_priv->av);
2724         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av);
2725         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2726         if (!ret)
2727                 list_add_tail(&work->list, &cm_id_priv->work_list);
2728         spin_unlock_irq(&cm_id_priv->lock);
2729
2730         if (ret)
2731                 cm_process_work(cm_id_priv, work);
2732         else
2733                 cm_deref_id(cm_id_priv);
2734         return 0;
2735
2736 unlock: spin_unlock_irq(&cm_id_priv->lock);
2737 deref:  cm_deref_id(cm_id_priv);
2738         return -EINVAL;
2739 }
2740
2741 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2742                           struct cm_id_private *cm_id_priv,
2743                           enum ib_cm_apr_status status,
2744                           void *info,
2745                           u8 info_length,
2746                           const void *private_data,
2747                           u8 private_data_len)
2748 {
2749         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2750         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2751         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2752         apr_msg->ap_status = (u8) status;
2753
2754         if (info && info_length) {
2755                 apr_msg->info_length = info_length;
2756                 memcpy(apr_msg->info, info, info_length);
2757         }
2758
2759         if (private_data && private_data_len)
2760                 memcpy(apr_msg->private_data, private_data, private_data_len);
2761 }
2762
2763 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2764                    enum ib_cm_apr_status status,
2765                    void *info,
2766                    u8 info_length,
2767                    const void *private_data,
2768                    u8 private_data_len)
2769 {
2770         struct cm_id_private *cm_id_priv;
2771         struct ib_mad_send_buf *msg;
2772         unsigned long flags;
2773         int ret;
2774
2775         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2776             (info && info_length > IB_CM_APR_INFO_LENGTH))
2777                 return -EINVAL;
2778
2779         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2780         spin_lock_irqsave(&cm_id_priv->lock, flags);
2781         if (cm_id->state != IB_CM_ESTABLISHED ||
2782             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2783              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2784                 ret = -EINVAL;
2785                 goto out;
2786         }
2787
2788         ret = cm_alloc_msg(cm_id_priv, &msg);
2789         if (ret)
2790                 goto out;
2791
2792         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2793                       info, info_length, private_data, private_data_len);
2794         ret = ib_post_send_mad(msg, NULL);
2795         if (ret) {
2796                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2797                 cm_free_msg(msg);
2798                 return ret;
2799         }
2800
2801         cm_id->lap_state = IB_CM_LAP_IDLE;
2802 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2803         return ret;
2804 }
2805 EXPORT_SYMBOL(ib_send_cm_apr);
2806
2807 static int cm_apr_handler(struct cm_work *work)
2808 {
2809         struct cm_id_private *cm_id_priv;
2810         struct cm_apr_msg *apr_msg;
2811         int ret;
2812
2813         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2814         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2815                                    apr_msg->local_comm_id);
2816         if (!cm_id_priv)
2817                 return -EINVAL; /* Unmatched reply. */
2818
2819         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2820         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2821         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2822         work->cm_event.private_data = &apr_msg->private_data;
2823
2824         spin_lock_irq(&cm_id_priv->lock);
2825         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2826             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2827              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2828                 spin_unlock_irq(&cm_id_priv->lock);
2829                 goto out;
2830         }
2831         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2832         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2833         cm_id_priv->msg = NULL;
2834
2835         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2836         if (!ret)
2837                 list_add_tail(&work->list, &cm_id_priv->work_list);
2838         spin_unlock_irq(&cm_id_priv->lock);
2839
2840         if (ret)
2841                 cm_process_work(cm_id_priv, work);
2842         else
2843                 cm_deref_id(cm_id_priv);
2844         return 0;
2845 out:
2846         cm_deref_id(cm_id_priv);
2847         return -EINVAL;
2848 }
2849
2850 static int cm_timewait_handler(struct cm_work *work)
2851 {
2852         struct cm_timewait_info *timewait_info;
2853         struct cm_id_private *cm_id_priv;
2854         int ret;
2855
2856         timewait_info = (struct cm_timewait_info *)work;
2857         spin_lock_irq(&cm.lock);
2858         list_del(&timewait_info->list);
2859         spin_unlock_irq(&cm.lock);
2860
2861         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2862                                    timewait_info->work.remote_id);
2863         if (!cm_id_priv)
2864                 return -EINVAL;
2865
2866         spin_lock_irq(&cm_id_priv->lock);
2867         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2868             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2869                 spin_unlock_irq(&cm_id_priv->lock);
2870                 goto out;
2871         }
2872         cm_id_priv->id.state = IB_CM_IDLE;
2873         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2874         if (!ret)
2875                 list_add_tail(&work->list, &cm_id_priv->work_list);
2876         spin_unlock_irq(&cm_id_priv->lock);
2877
2878         if (ret)
2879                 cm_process_work(cm_id_priv, work);
2880         else
2881                 cm_deref_id(cm_id_priv);
2882         return 0;
2883 out:
2884         cm_deref_id(cm_id_priv);
2885         return -EINVAL;
2886 }
2887
2888 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
2889                                struct cm_id_private *cm_id_priv,
2890                                struct ib_cm_sidr_req_param *param)
2891 {
2892         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
2893                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
2894         sidr_req_msg->request_id = cm_id_priv->id.local_id;
2895         sidr_req_msg->pkey = param->path->pkey;
2896         sidr_req_msg->service_id = param->service_id;
2897
2898         if (param->private_data && param->private_data_len)
2899                 memcpy(sidr_req_msg->private_data, param->private_data,
2900                        param->private_data_len);
2901 }
2902
2903 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
2904                         struct ib_cm_sidr_req_param *param)
2905 {
2906         struct cm_id_private *cm_id_priv;
2907         struct ib_mad_send_buf *msg;
2908         unsigned long flags;
2909         int ret;
2910
2911         if (!param->path || (param->private_data &&
2912              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
2913                 return -EINVAL;
2914
2915         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2916         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
2917         if (ret)
2918                 goto out;
2919
2920         cm_id->service_id = param->service_id;
2921         cm_id->service_mask = ~cpu_to_be64(0);
2922         cm_id_priv->timeout_ms = param->timeout_ms;
2923         cm_id_priv->max_cm_retries = param->max_cm_retries;
2924         ret = cm_alloc_msg(cm_id_priv, &msg);
2925         if (ret)
2926                 goto out;
2927
2928         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
2929                            param);
2930         msg->timeout_ms = cm_id_priv->timeout_ms;
2931         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
2932
2933         spin_lock_irqsave(&cm_id_priv->lock, flags);
2934         if (cm_id->state == IB_CM_IDLE)
2935                 ret = ib_post_send_mad(msg, NULL);
2936         else
2937                 ret = -EINVAL;
2938
2939         if (ret) {
2940                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2941                 cm_free_msg(msg);
2942                 goto out;
2943         }
2944         cm_id->state = IB_CM_SIDR_REQ_SENT;
2945         cm_id_priv->msg = msg;
2946         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2947 out:
2948         return ret;
2949 }
2950 EXPORT_SYMBOL(ib_send_cm_sidr_req);
2951
2952 static void cm_format_sidr_req_event(struct cm_work *work,
2953                                      struct ib_cm_id *listen_id)
2954 {
2955         struct cm_sidr_req_msg *sidr_req_msg;
2956         struct ib_cm_sidr_req_event_param *param;
2957
2958         sidr_req_msg = (struct cm_sidr_req_msg *)
2959                                 work->mad_recv_wc->recv_buf.mad;
2960         param = &work->cm_event.param.sidr_req_rcvd;
2961         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
2962         param->listen_id = listen_id;
2963         param->port = work->port->port_num;
2964         work->cm_event.private_data = &sidr_req_msg->private_data;
2965 }
2966
2967 static int cm_sidr_req_handler(struct cm_work *work)
2968 {
2969         struct ib_cm_id *cm_id;
2970         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
2971         struct cm_sidr_req_msg *sidr_req_msg;
2972         struct ib_wc *wc;
2973
2974         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
2975         if (IS_ERR(cm_id))
2976                 return PTR_ERR(cm_id);
2977         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2978
2979         /* Record SGID/SLID and request ID for lookup. */
2980         sidr_req_msg = (struct cm_sidr_req_msg *)
2981                                 work->mad_recv_wc->recv_buf.mad;
2982         wc = work->mad_recv_wc->wc;
2983         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
2984         cm_id_priv->av.dgid.global.interface_id = 0;
2985         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2986                                 work->mad_recv_wc->recv_buf.grh,
2987                                 &cm_id_priv->av);
2988         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
2989         cm_id_priv->tid = sidr_req_msg->hdr.tid;
2990         atomic_inc(&cm_id_priv->work_count);
2991
2992         spin_lock_irq(&cm.lock);
2993         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
2994         if (cur_cm_id_priv) {
2995                 spin_unlock_irq(&cm.lock);
2996                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2997                                 counter[CM_SIDR_REQ_COUNTER]);
2998                 goto out; /* Duplicate message. */
2999         }
3000         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3001         cur_cm_id_priv = cm_find_listen(cm_id->device,
3002                                         sidr_req_msg->service_id,
3003                                         sidr_req_msg->private_data);
3004         if (!cur_cm_id_priv) {
3005                 spin_unlock_irq(&cm.lock);
3006                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3007                 goto out; /* No match. */
3008         }
3009         atomic_inc(&cur_cm_id_priv->refcount);
3010         atomic_inc(&cm_id_priv->refcount);
3011         spin_unlock_irq(&cm.lock);
3012
3013         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3014         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3015         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3016         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3017
3018         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3019         cm_process_work(cm_id_priv, work);
3020         cm_deref_id(cur_cm_id_priv);
3021         return 0;
3022 out:
3023         ib_destroy_cm_id(&cm_id_priv->id);
3024         return -EINVAL;
3025 }
3026
3027 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3028                                struct cm_id_private *cm_id_priv,
3029                                struct ib_cm_sidr_rep_param *param)
3030 {
3031         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3032                           cm_id_priv->tid);
3033         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3034         sidr_rep_msg->status = param->status;
3035         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3036         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3037         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3038
3039         if (param->info && param->info_length)
3040                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3041
3042         if (param->private_data && param->private_data_len)
3043                 memcpy(sidr_rep_msg->private_data, param->private_data,
3044                        param->private_data_len);
3045 }
3046
3047 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3048                         struct ib_cm_sidr_rep_param *param)
3049 {
3050         struct cm_id_private *cm_id_priv;
3051         struct ib_mad_send_buf *msg;
3052         unsigned long flags;
3053         int ret;
3054
3055         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3056             (param->private_data &&
3057              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3058                 return -EINVAL;
3059
3060         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3061         spin_lock_irqsave(&cm_id_priv->lock, flags);
3062         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3063                 ret = -EINVAL;
3064                 goto error;
3065         }
3066
3067         ret = cm_alloc_msg(cm_id_priv, &msg);
3068         if (ret)
3069                 goto error;
3070
3071         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3072                            param);
3073         ret = ib_post_send_mad(msg, NULL);
3074         if (ret) {
3075                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3076                 cm_free_msg(msg);
3077                 return ret;
3078         }
3079         cm_id->state = IB_CM_IDLE;
3080         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3081
3082         spin_lock_irqsave(&cm.lock, flags);
3083         rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3084         spin_unlock_irqrestore(&cm.lock, flags);
3085         return 0;
3086
3087 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3088         return ret;
3089 }
3090 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3091
3092 static void cm_format_sidr_rep_event(struct cm_work *work)
3093 {
3094         struct cm_sidr_rep_msg *sidr_rep_msg;
3095         struct ib_cm_sidr_rep_event_param *param;
3096
3097         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3098                                 work->mad_recv_wc->recv_buf.mad;
3099         param = &work->cm_event.param.sidr_rep_rcvd;
3100         param->status = sidr_rep_msg->status;
3101         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3102         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3103         param->info = &sidr_rep_msg->info;
3104         param->info_len = sidr_rep_msg->info_length;
3105         work->cm_event.private_data = &sidr_rep_msg->private_data;
3106 }
3107
3108 static int cm_sidr_rep_handler(struct cm_work *work)
3109 {
3110         struct cm_sidr_rep_msg *sidr_rep_msg;
3111         struct cm_id_private *cm_id_priv;
3112
3113         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3114                                 work->mad_recv_wc->recv_buf.mad;
3115         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3116         if (!cm_id_priv)
3117                 return -EINVAL; /* Unmatched reply. */
3118
3119         spin_lock_irq(&cm_id_priv->lock);
3120         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3121                 spin_unlock_irq(&cm_id_priv->lock);
3122                 goto out;
3123         }
3124         cm_id_priv->id.state = IB_CM_IDLE;
3125         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3126         spin_unlock_irq(&cm_id_priv->lock);
3127
3128         cm_format_sidr_rep_event(work);
3129         cm_process_work(cm_id_priv, work);
3130         return 0;
3131 out:
3132         cm_deref_id(cm_id_priv);
3133         return -EINVAL;
3134 }
3135
3136 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3137                                   enum ib_wc_status wc_status)
3138 {
3139         struct cm_id_private *cm_id_priv;
3140         struct ib_cm_event cm_event;
3141         enum ib_cm_state state;
3142         int ret;
3143
3144         memset(&cm_event, 0, sizeof cm_event);
3145         cm_id_priv = msg->context[0];
3146
3147         /* Discard old sends or ones without a response. */
3148         spin_lock_irq(&cm_id_priv->lock);
3149         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3150         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3151                 goto discard;
3152
3153         switch (state) {
3154         case IB_CM_REQ_SENT:
3155         case IB_CM_MRA_REQ_RCVD:
3156                 cm_reset_to_idle(cm_id_priv);
3157                 cm_event.event = IB_CM_REQ_ERROR;
3158                 break;
3159         case IB_CM_REP_SENT:
3160         case IB_CM_MRA_REP_RCVD:
3161                 cm_reset_to_idle(cm_id_priv);
3162                 cm_event.event = IB_CM_REP_ERROR;
3163                 break;
3164         case IB_CM_DREQ_SENT:
3165                 cm_enter_timewait(cm_id_priv);
3166                 cm_event.event = IB_CM_DREQ_ERROR;
3167                 break;
3168         case IB_CM_SIDR_REQ_SENT:
3169                 cm_id_priv->id.state = IB_CM_IDLE;
3170                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3171                 break;
3172         default:
3173                 goto discard;
3174         }
3175         spin_unlock_irq(&cm_id_priv->lock);
3176         cm_event.param.send_status = wc_status;
3177
3178         /* No other events can occur on the cm_id at this point. */
3179         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3180         cm_free_msg(msg);
3181         if (ret)
3182                 ib_destroy_cm_id(&cm_id_priv->id);
3183         return;
3184 discard:
3185         spin_unlock_irq(&cm_id_priv->lock);
3186         cm_free_msg(msg);
3187 }
3188
3189 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3190                             struct ib_mad_send_wc *mad_send_wc)
3191 {
3192         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3193         struct cm_port *port;
3194         u16 attr_index;
3195
3196         port = mad_agent->context;
3197         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3198                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3199
3200         /*
3201          * If the send was in response to a received message (context[0] is not
3202          * set to a cm_id), and is not a REJ, then it is a send that was
3203          * manually retried.
3204          */
3205         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3206                 msg->retries = 1;
3207
3208         atomic_long_add(1 + msg->retries,
3209                         &port->counter_group[CM_XMIT].counter[attr_index]);
3210         if (msg->retries)
3211                 atomic_long_add(msg->retries,
3212                                 &port->counter_group[CM_XMIT_RETRIES].
3213                                 counter[attr_index]);
3214
3215         switch (mad_send_wc->status) {
3216         case IB_WC_SUCCESS:
3217         case IB_WC_WR_FLUSH_ERR:
3218                 cm_free_msg(msg);
3219                 break;
3220         default:
3221                 if (msg->context[0] && msg->context[1])
3222                         cm_process_send_error(msg, mad_send_wc->status);
3223                 else
3224                         cm_free_msg(msg);
3225                 break;
3226         }
3227 }
3228
3229 static void cm_work_handler(struct work_struct *_work)
3230 {
3231         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3232         int ret;
3233
3234         switch (work->cm_event.event) {
3235         case IB_CM_REQ_RECEIVED:
3236                 ret = cm_req_handler(work);
3237                 break;
3238         case IB_CM_MRA_RECEIVED:
3239                 ret = cm_mra_handler(work);
3240                 break;
3241         case IB_CM_REJ_RECEIVED:
3242                 ret = cm_rej_handler(work);
3243                 break;
3244         case IB_CM_REP_RECEIVED:
3245                 ret = cm_rep_handler(work);
3246                 break;
3247         case IB_CM_RTU_RECEIVED:
3248                 ret = cm_rtu_handler(work);
3249                 break;
3250         case IB_CM_USER_ESTABLISHED:
3251                 ret = cm_establish_handler(work);
3252                 break;
3253         case IB_CM_DREQ_RECEIVED:
3254                 ret = cm_dreq_handler(work);
3255                 break;
3256         case IB_CM_DREP_RECEIVED:
3257                 ret = cm_drep_handler(work);
3258                 break;
3259         case IB_CM_SIDR_REQ_RECEIVED:
3260                 ret = cm_sidr_req_handler(work);
3261                 break;
3262         case IB_CM_SIDR_REP_RECEIVED:
3263                 ret = cm_sidr_rep_handler(work);
3264                 break;
3265         case IB_CM_LAP_RECEIVED:
3266                 ret = cm_lap_handler(work);
3267                 break;
3268         case IB_CM_APR_RECEIVED:
3269                 ret = cm_apr_handler(work);
3270                 break;
3271         case IB_CM_TIMEWAIT_EXIT:
3272                 ret = cm_timewait_handler(work);
3273                 break;
3274         default:
3275                 ret = -EINVAL;
3276                 break;
3277         }
3278         if (ret)
3279                 cm_free_work(work);
3280 }
3281
3282 static int cm_establish(struct ib_cm_id *cm_id)
3283 {
3284         struct cm_id_private *cm_id_priv;
3285         struct cm_work *work;
3286         unsigned long flags;
3287         int ret = 0;
3288
3289         work = kmalloc(sizeof *work, GFP_ATOMIC);
3290         if (!work)
3291                 return -ENOMEM;
3292
3293         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3294         spin_lock_irqsave(&cm_id_priv->lock, flags);
3295         switch (cm_id->state)
3296         {
3297         case IB_CM_REP_SENT:
3298         case IB_CM_MRA_REP_RCVD:
3299                 cm_id->state = IB_CM_ESTABLISHED;
3300                 break;
3301         case IB_CM_ESTABLISHED:
3302                 ret = -EISCONN;
3303                 break;
3304         default:
3305                 ret = -EINVAL;
3306                 break;
3307         }
3308         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3309
3310         if (ret) {
3311                 kfree(work);
3312                 goto out;
3313         }
3314
3315         /*
3316          * The CM worker thread may try to destroy the cm_id before it
3317          * can execute this work item.  To prevent potential deadlock,
3318          * we need to find the cm_id once we're in the context of the
3319          * worker thread, rather than holding a reference on it.
3320          */
3321         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3322         work->local_id = cm_id->local_id;
3323         work->remote_id = cm_id->remote_id;
3324         work->mad_recv_wc = NULL;
3325         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3326         queue_delayed_work(cm.wq, &work->work, 0);
3327 out:
3328         return ret;
3329 }
3330
3331 static int cm_migrate(struct ib_cm_id *cm_id)
3332 {
3333         struct cm_id_private *cm_id_priv;
3334         unsigned long flags;
3335         int ret = 0;
3336
3337         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3338         spin_lock_irqsave(&cm_id_priv->lock, flags);
3339         if (cm_id->state == IB_CM_ESTABLISHED &&
3340             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3341              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3342                 cm_id->lap_state = IB_CM_LAP_IDLE;
3343                 cm_id_priv->av = cm_id_priv->alt_av;
3344         } else
3345                 ret = -EINVAL;
3346         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3347
3348         return ret;
3349 }
3350
3351 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3352 {
3353         int ret;
3354
3355         switch (event) {
3356         case IB_EVENT_COMM_EST:
3357                 ret = cm_establish(cm_id);
3358                 break;
3359         case IB_EVENT_PATH_MIG:
3360                 ret = cm_migrate(cm_id);
3361                 break;
3362         default:
3363                 ret = -EINVAL;
3364         }
3365         return ret;
3366 }
3367 EXPORT_SYMBOL(ib_cm_notify);
3368
3369 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3370                             struct ib_mad_recv_wc *mad_recv_wc)
3371 {
3372         struct cm_port *port = mad_agent->context;
3373         struct cm_work *work;
3374         enum ib_cm_event_type event;
3375         u16 attr_id;
3376         int paths = 0;
3377
3378         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3379         case CM_REQ_ATTR_ID:
3380                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3381                                                     alt_local_lid != 0);
3382                 event = IB_CM_REQ_RECEIVED;
3383                 break;
3384         case CM_MRA_ATTR_ID:
3385                 event = IB_CM_MRA_RECEIVED;
3386                 break;
3387         case CM_REJ_ATTR_ID:
3388                 event = IB_CM_REJ_RECEIVED;
3389                 break;
3390         case CM_REP_ATTR_ID:
3391                 event = IB_CM_REP_RECEIVED;
3392                 break;
3393         case CM_RTU_ATTR_ID:
3394                 event = IB_CM_RTU_RECEIVED;
3395                 break;
3396         case CM_DREQ_ATTR_ID:
3397                 event = IB_CM_DREQ_RECEIVED;
3398                 break;
3399         case CM_DREP_ATTR_ID:
3400                 event = IB_CM_DREP_RECEIVED;
3401                 break;
3402         case CM_SIDR_REQ_ATTR_ID:
3403                 event = IB_CM_SIDR_REQ_RECEIVED;
3404                 break;
3405         case CM_SIDR_REP_ATTR_ID:
3406                 event = IB_CM_SIDR_REP_RECEIVED;
3407                 break;
3408         case CM_LAP_ATTR_ID:
3409                 paths = 1;
3410                 event = IB_CM_LAP_RECEIVED;
3411                 break;
3412         case CM_APR_ATTR_ID:
3413                 event = IB_CM_APR_RECEIVED;
3414                 break;
3415         default:
3416                 ib_free_recv_mad(mad_recv_wc);
3417                 return;
3418         }
3419
3420         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3421         atomic_long_inc(&port->counter_group[CM_RECV].
3422                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3423
3424         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3425                        GFP_KERNEL);
3426         if (!work) {
3427                 ib_free_recv_mad(mad_recv_wc);
3428                 return;
3429         }
3430
3431         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3432         work->cm_event.event = event;
3433         work->mad_recv_wc = mad_recv_wc;
3434         work->port = port;
3435         queue_delayed_work(cm.wq, &work->work, 0);
3436 }
3437
3438 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3439                                 struct ib_qp_attr *qp_attr,
3440                                 int *qp_attr_mask)
3441 {
3442         unsigned long flags;
3443         int ret;
3444
3445         spin_lock_irqsave(&cm_id_priv->lock, flags);
3446         switch (cm_id_priv->id.state) {
3447         case IB_CM_REQ_SENT:
3448         case IB_CM_MRA_REQ_RCVD:
3449         case IB_CM_REQ_RCVD:
3450         case IB_CM_MRA_REQ_SENT:
3451         case IB_CM_REP_RCVD:
3452         case IB_CM_MRA_REP_SENT:
3453         case IB_CM_REP_SENT:
3454         case IB_CM_MRA_REP_RCVD:
3455         case IB_CM_ESTABLISHED:
3456                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3457                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3458                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3459                 if (cm_id_priv->responder_resources)
3460                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3461                                                     IB_ACCESS_REMOTE_ATOMIC;
3462                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3463                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3464                 ret = 0;
3465                 break;
3466         default:
3467                 ret = -EINVAL;
3468                 break;
3469         }
3470         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3471         return ret;
3472 }
3473
3474 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3475                                struct ib_qp_attr *qp_attr,
3476                                int *qp_attr_mask)
3477 {
3478         unsigned long flags;
3479         int ret;
3480
3481         spin_lock_irqsave(&cm_id_priv->lock, flags);
3482         switch (cm_id_priv->id.state) {
3483         case IB_CM_REQ_RCVD:
3484         case IB_CM_MRA_REQ_SENT:
3485         case IB_CM_REP_RCVD:
3486         case IB_CM_MRA_REP_SENT:
3487         case IB_CM_REP_SENT:
3488         case IB_CM_MRA_REP_RCVD:
3489         case IB_CM_ESTABLISHED:
3490                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3491                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3492                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3493                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3494                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3495                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3496                 if (cm_id_priv->qp_type == IB_QPT_RC) {
3497                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3498                                          IB_QP_MIN_RNR_TIMER;
3499                         qp_attr->max_dest_rd_atomic =
3500                                         cm_id_priv->responder_resources;
3501                         qp_attr->min_rnr_timer = 0;
3502                 }
3503                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3504                         *qp_attr_mask |= IB_QP_ALT_PATH;
3505                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3506                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3507                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3508                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3509                 }
3510                 ret = 0;
3511                 break;
3512         default:
3513                 ret = -EINVAL;
3514                 break;
3515         }
3516         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3517         return ret;
3518 }
3519
3520 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3521                                struct ib_qp_attr *qp_attr,
3522                                int *qp_attr_mask)
3523 {
3524         unsigned long flags;
3525         int ret;
3526
3527         spin_lock_irqsave(&cm_id_priv->lock, flags);
3528         switch (cm_id_priv->id.state) {
3529         /* Allow transition to RTS before sending REP */
3530         case IB_CM_REQ_RCVD:
3531         case IB_CM_MRA_REQ_SENT:
3532
3533         case IB_CM_REP_RCVD:
3534         case IB_CM_MRA_REP_SENT:
3535         case IB_CM_REP_SENT:
3536         case IB_CM_MRA_REP_RCVD:
3537         case IB_CM_ESTABLISHED:
3538                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3539                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3540                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3541                         if (cm_id_priv->qp_type == IB_QPT_RC) {
3542                                 *qp_attr_mask |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
3543                                                  IB_QP_RNR_RETRY |
3544                                                  IB_QP_MAX_QP_RD_ATOMIC;
3545                                 qp_attr->timeout = cm_id_priv->av.timeout;
3546                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3547                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3548                                 qp_attr->max_rd_atomic =
3549                                         cm_id_priv->initiator_depth;
3550                         }
3551                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3552                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3553                                 qp_attr->path_mig_state = IB_MIG_REARM;
3554                         }
3555                 } else {
3556                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3557                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3558                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3559                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3560                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3561                         qp_attr->path_mig_state = IB_MIG_REARM;
3562                 }
3563                 ret = 0;
3564                 break;
3565         default:
3566                 ret = -EINVAL;
3567                 break;
3568         }
3569         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3570         return ret;
3571 }
3572
3573 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3574                        struct ib_qp_attr *qp_attr,
3575                        int *qp_attr_mask)
3576 {
3577         struct cm_id_private *cm_id_priv;
3578         int ret;
3579
3580         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3581         switch (qp_attr->qp_state) {
3582         case IB_QPS_INIT:
3583                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3584                 break;
3585         case IB_QPS_RTR:
3586                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3587                 break;
3588         case IB_QPS_RTS:
3589                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3590                 break;
3591         default:
3592                 ret = -EINVAL;
3593                 break;
3594         }
3595         return ret;
3596 }
3597 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3598
3599 static void cm_get_ack_delay(struct cm_device *cm_dev)
3600 {
3601         struct ib_device_attr attr;
3602
3603         if (ib_query_device(cm_dev->ib_device, &attr))
3604                 cm_dev->ack_delay = 0; /* acks will rely on packet life time */
3605         else
3606                 cm_dev->ack_delay = attr.local_ca_ack_delay;
3607 }
3608
3609 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3610                                char *buf)
3611 {
3612         struct cm_counter_group *group;
3613         struct cm_counter_attribute *cm_attr;
3614
3615         group = container_of(obj, struct cm_counter_group, obj);
3616         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3617
3618         return sprintf(buf, "%ld\n",
3619                        atomic_long_read(&group->counter[cm_attr->index]));
3620 }
3621
3622 static const struct sysfs_ops cm_counter_ops = {
3623         .show = cm_show_counter
3624 };
3625
3626 static struct kobj_type cm_counter_obj_type = {
3627         .sysfs_ops = &cm_counter_ops,
3628         .default_attrs = cm_counter_default_attrs
3629 };
3630
3631 static void cm_release_port_obj(struct kobject *obj)
3632 {
3633         struct cm_port *cm_port;
3634
3635         cm_port = container_of(obj, struct cm_port, port_obj);
3636         kfree(cm_port);
3637 }
3638
3639 static struct kobj_type cm_port_obj_type = {
3640         .release = cm_release_port_obj
3641 };
3642
3643 static char *cm_devnode(struct device *dev, mode_t *mode)
3644 {
3645         if (mode)
3646                 *mode = 0666;
3647         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3648 }
3649
3650 struct class cm_class = {
3651         .owner   = THIS_MODULE,
3652         .name    = "infiniband_cm",
3653         .devnode = cm_devnode,
3654 };
3655 EXPORT_SYMBOL(cm_class);
3656
3657 static int cm_create_port_fs(struct cm_port *port)
3658 {
3659         int i, ret;
3660
3661         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3662                                    &port->cm_dev->device->kobj,
3663                                    "%d", port->port_num);
3664         if (ret) {
3665                 kfree(port);
3666                 return ret;
3667         }
3668
3669         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3670                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3671                                            &cm_counter_obj_type,
3672                                            &port->port_obj,
3673                                            "%s", counter_group_names[i]);
3674                 if (ret)
3675                         goto error;
3676         }
3677
3678         return 0;
3679
3680 error:
3681         while (i--)
3682                 kobject_put(&port->counter_group[i].obj);
3683         kobject_put(&port->port_obj);
3684         return ret;
3685
3686 }
3687
3688 static void cm_remove_port_fs(struct cm_port *port)
3689 {
3690         int i;
3691
3692         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3693                 kobject_put(&port->counter_group[i].obj);
3694
3695         kobject_put(&port->port_obj);
3696 }
3697
3698 static void cm_add_one(struct ib_device *ib_device)
3699 {
3700         struct cm_device *cm_dev;
3701         struct cm_port *port;
3702         struct ib_mad_reg_req reg_req = {
3703                 .mgmt_class = IB_MGMT_CLASS_CM,
3704                 .mgmt_class_version = IB_CM_CLASS_VERSION
3705         };
3706         struct ib_port_modify port_modify = {
3707                 .set_port_cap_mask = IB_PORT_CM_SUP
3708         };
3709         unsigned long flags;
3710         int ret;
3711         u8 i;
3712
3713         if (rdma_node_get_transport(ib_device->node_type) != RDMA_TRANSPORT_IB)
3714                 return;
3715
3716         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3717                          ib_device->phys_port_cnt, GFP_KERNEL);
3718         if (!cm_dev)
3719                 return;
3720
3721         cm_dev->ib_device = ib_device;
3722         cm_get_ack_delay(cm_dev);
3723
3724         cm_dev->device = device_create(&cm_class, &ib_device->dev,
3725                                        MKDEV(0, 0), NULL,
3726                                        "%s", ib_device->name);
3727         if (IS_ERR(cm_dev->device)) {
3728                 kfree(cm_dev);
3729                 return;
3730         }
3731
3732         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3733         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3734                 port = kzalloc(sizeof *port, GFP_KERNEL);
3735                 if (!port)
3736                         goto error1;
3737
3738                 cm_dev->port[i-1] = port;
3739                 port->cm_dev = cm_dev;
3740                 port->port_num = i;
3741
3742                 ret = cm_create_port_fs(port);
3743                 if (ret)
3744                         goto error1;
3745
3746                 port->mad_agent = ib_register_mad_agent(ib_device, i,
3747                                                         IB_QPT_GSI,
3748                                                         &reg_req,
3749                                                         0,
3750                                                         cm_send_handler,
3751                                                         cm_recv_handler,
3752                                                         port);
3753                 if (IS_ERR(port->mad_agent))
3754                         goto error2;
3755
3756                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
3757                 if (ret)
3758                         goto error3;
3759         }
3760         ib_set_client_data(ib_device, &cm_client, cm_dev);
3761
3762         write_lock_irqsave(&cm.device_lock, flags);
3763         list_add_tail(&cm_dev->list, &cm.device_list);
3764         write_unlock_irqrestore(&cm.device_lock, flags);
3765         return;
3766
3767 error3:
3768         ib_unregister_mad_agent(port->mad_agent);
3769 error2:
3770         cm_remove_port_fs(port);
3771 error1:
3772         port_modify.set_port_cap_mask = 0;
3773         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3774         while (--i) {
3775                 port = cm_dev->port[i-1];
3776                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3777                 ib_unregister_mad_agent(port->mad_agent);
3778                 cm_remove_port_fs(port);
3779         }
3780         device_unregister(cm_dev->device);
3781         kfree(cm_dev);
3782 }
3783
3784 static void cm_remove_one(struct ib_device *ib_device)
3785 {
3786         struct cm_device *cm_dev;
3787         struct cm_port *port;
3788         struct ib_port_modify port_modify = {
3789                 .clr_port_cap_mask = IB_PORT_CM_SUP
3790         };
3791         unsigned long flags;
3792         int i;
3793
3794         cm_dev = ib_get_client_data(ib_device, &cm_client);
3795         if (!cm_dev)
3796                 return;
3797
3798         write_lock_irqsave(&cm.device_lock, flags);
3799         list_del(&cm_dev->list);
3800         write_unlock_irqrestore(&cm.device_lock, flags);
3801
3802         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3803                 port = cm_dev->port[i-1];
3804                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3805                 ib_unregister_mad_agent(port->mad_agent);
3806                 flush_workqueue(cm.wq);
3807                 cm_remove_port_fs(port);
3808         }
3809         device_unregister(cm_dev->device);
3810         kfree(cm_dev);
3811 }
3812
3813 static int __init ib_cm_init(void)
3814 {
3815         int ret;
3816
3817         memset(&cm, 0, sizeof cm);
3818         INIT_LIST_HEAD(&cm.device_list);
3819         rwlock_init(&cm.device_lock);
3820         spin_lock_init(&cm.lock);
3821         cm.listen_service_table = RB_ROOT;
3822         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3823         cm.remote_id_table = RB_ROOT;
3824         cm.remote_qp_table = RB_ROOT;
3825         cm.remote_sidr_table = RB_ROOT;
3826         idr_init(&cm.local_id_table);
3827         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
3828         idr_pre_get(&cm.local_id_table, GFP_KERNEL);
3829         INIT_LIST_HEAD(&cm.timewait_list);
3830
3831         ret = class_register(&cm_class);
3832         if (ret)
3833                 return -ENOMEM;
3834
3835         cm.wq = create_workqueue("ib_cm");
3836         if (!cm.wq) {
3837                 ret = -ENOMEM;
3838                 goto error1;
3839         }
3840
3841         ret = ib_register_client(&cm_client);
3842         if (ret)
3843                 goto error2;
3844
3845         return 0;
3846 error2:
3847         destroy_workqueue(cm.wq);
3848 error1:
3849         class_unregister(&cm_class);
3850         return ret;
3851 }
3852
3853 static void __exit ib_cm_cleanup(void)
3854 {
3855         struct cm_timewait_info *timewait_info, *tmp;
3856
3857         spin_lock_irq(&cm.lock);
3858         list_for_each_entry(timewait_info, &cm.timewait_list, list)
3859                 cancel_delayed_work(&timewait_info->work.work);
3860         spin_unlock_irq(&cm.lock);
3861
3862         ib_unregister_client(&cm_client);
3863         destroy_workqueue(cm.wq);
3864
3865         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
3866                 list_del(&timewait_info->list);
3867                 kfree(timewait_info);
3868         }
3869
3870         class_unregister(&cm_class);
3871         idr_destroy(&cm.local_id_table);
3872 }
3873
3874 module_init(ib_cm_init);
3875 module_exit(ib_cm_cleanup);
3876