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