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/err.h>
  40 #include <linux/idr.h>
  41 #include <linux/interrupt.h>
  42 #include <linux/random.h>
  43 #include <linux/rbtree.h>
  44 #include <linux/spinlock.h>
  45 #include <linux/slab.h>
  46 #include <linux/sysfs.h>
  47 #include <linux/workqueue.h>
  48 #include <linux/kdev_t.h>
  49
  50 #include <rdma/ib_cache.h>
  51 #include <rdma/ib_cm.h>
  52 #include "cm_msgs.h"
  53
  54 MODULE_AUTHOR("Sean Hefty");
  55 MODULE_DESCRIPTION("InfiniBand CM");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 static void cm_add_one(struct ib_device *device);
  59 static void cm_remove_one(struct ib_device *device);
  60
  61 static struct ib_client cm_client = {
  62         .name   = "cm",
  63         .add    = cm_add_one,
  64         .remove = cm_remove_one
  65 };
  66
  67 static struct ib_cm {
  68         spinlock_t lock;
  69         struct list_head device_list;
  70         rwlock_t device_lock;
  71         struct rb_root listen_service_table;
  72         u64 listen_service_id;
  73         /* struct rb_root peer_service_table; todo: fix peer to peer */
  74         struct rb_root remote_qp_table;
  75         struct rb_root remote_id_table;
  76         struct rb_root remote_sidr_table;
  77         struct idr local_id_table;
  78         __be32 random_id_operand;
  79         struct list_head timewait_list;
  80         struct workqueue_struct *wq;
  81 } cm;
  82
  83 /* Counter indexes ordered by attribute ID */
  84 enum {
  85         CM_REQ_COUNTER,
  86         CM_MRA_COUNTER,
  87         CM_REJ_COUNTER,
  88         CM_REP_COUNTER,
  89         CM_RTU_COUNTER,
  90         CM_DREQ_COUNTER,
  91         CM_DREP_COUNTER,
  92         CM_SIDR_REQ_COUNTER,
  93         CM_SIDR_REP_COUNTER,
  94         CM_LAP_COUNTER,
  95         CM_APR_COUNTER,
  96         CM_ATTR_COUNT,
  97         CM_ATTR_ID_OFFSET = 0x0010,
  98 };
  99
 100 enum {
 101         CM_XMIT,
 102         CM_XMIT_RETRIES,
 103         CM_RECV,
 104         CM_RECV_DUPLICATES,
 105         CM_COUNTER_GROUPS
 106 };
 107
 108 static char const counter_group_names[CM_COUNTER_GROUPS]
 109                                      [sizeof("cm_rx_duplicates")] = {
 110         "cm_tx_msgs", "cm_tx_retries",
 111         "cm_rx_msgs", "cm_rx_duplicates"
 112 };
 113
 114 struct cm_counter_group {
 115         struct kobject obj;
 116         atomic_long_t counter[CM_ATTR_COUNT];
 117 };
 118
 119 struct cm_counter_attribute {
 120         struct attribute attr;
 121         int index;
 122 };
 123
 124 #define CM_COUNTER_ATTR(_name, _index) \
 125 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 126         .attr = { .name = __stringify(_name), .mode = 0444 }, \
 127         .index = _index \
 128 }
 129
 130 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 131 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 132 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 133 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 134 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 135 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 136 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 137 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 138 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 139 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 140 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 141
 142 static struct attribute *cm_counter_default_attrs[] = {
 143         &cm_req_counter_attr.attr,
 144         &cm_mra_counter_attr.attr,
 145         &cm_rej_counter_attr.attr,
 146         &cm_rep_counter_attr.attr,
 147         &cm_rtu_counter_attr.attr,
 148         &cm_dreq_counter_attr.attr,
 149         &cm_drep_counter_attr.attr,
 150         &cm_sidr_req_counter_attr.attr,
 151         &cm_sidr_rep_counter_attr.attr,
 152         &cm_lap_counter_attr.attr,
 153         &cm_apr_counter_attr.attr,
 154         NULL
 155 };
 156
 157 struct cm_port {
 158         struct cm_device *cm_dev;
 159         struct ib_mad_agent *mad_agent;
 160         struct kobject port_obj;
 161         u8 port_num;
 162         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 163 };
 164
 165 struct cm_device {
 166         struct list_head list;
 167         struct ib_device *ib_device;
 168         struct device *device;
 169         u8 ack_delay;
 170         struct cm_port *port[0];
 171 };
 172
 173 struct cm_av {
 174         struct cm_port *port;
 175         union ib_gid dgid;
 176         struct ib_ah_attr ah_attr;
 177         u16 pkey_index;
 178         u8 timeout;
 179 };
 180
 181 struct cm_work {
 182         struct delayed_work work;
 183         struct list_head list;
 184         struct cm_port *port;
 185         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 186         __be32 local_id;                        /* Established / timewait */
 187         __be32 remote_id;
 188         struct ib_cm_event cm_event;
 189         struct ib_sa_path_rec path[0];
 190 };
 191
 192 struct cm_timewait_info {
 193         struct cm_work work;                    /* Must be first. */
 194         struct list_head list;
 195         struct rb_node remote_qp_node;
 196         struct rb_node remote_id_node;
 197         __be64 remote_ca_guid;
 198         __be32 remote_qpn;
 199         u8 inserted_remote_qp;
 200         u8 inserted_remote_id;
 201 };
 202
 203 struct cm_id_private {
 204         struct ib_cm_id id;
 205
 206         struct rb_node service_node;
 207         struct rb_node sidr_id_node;
 208         spinlock_t lock;        /* Do not acquire inside cm.lock */
 209         struct completion comp;
 210         atomic_t refcount;
 211
 212         struct ib_mad_send_buf *msg;
 213         struct cm_timewait_info *timewait_info;
 214         /* todo: use alternate port on send failure */
 215         struct cm_av av;
 216         struct cm_av alt_av;
 217         struct ib_cm_compare_data *compare_data;
 218
 219         void *private_data;
 220         __be64 tid;
 221         __be32 local_qpn;
 222         __be32 remote_qpn;
 223         enum ib_qp_type qp_type;
 224         __be32 sq_psn;
 225         __be32 rq_psn;
 226         int timeout_ms;
 227         enum ib_mtu path_mtu;
 228         __be16 pkey;
 229         u8 private_data_len;
 230         u8 max_cm_retries;
 231         u8 peer_to_peer;
 232         u8 responder_resources;
 233         u8 initiator_depth;
 234         u8 retry_count;
 235         u8 rnr_retry_count;
 236         u8 service_timeout;
 237         u8 target_ack_delay;
 238
 239         struct list_head work_list;
 240         atomic_t work_count;
 241 };
 242
 243 static void cm_work_handler(struct work_struct *work);
 244
 245 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 246 {
 247         if (atomic_dec_and_test(&cm_id_priv->refcount))
 248                 complete(&cm_id_priv->comp);
 249 }
 250
 251 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 252                         struct ib_mad_send_buf **msg)
 253 {
 254         struct ib_mad_agent *mad_agent;
 255         struct ib_mad_send_buf *m;
 256         struct ib_ah *ah;
 257
 258         mad_agent = cm_id_priv->av.port->mad_agent;
 259         ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
 260         if (IS_ERR(ah))
 261                 return PTR_ERR(ah);
 262
 263         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 264                                cm_id_priv->av.pkey_index,
 265                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 266                                GFP_ATOMIC);
 267         if (IS_ERR(m)) {
 268                 ib_destroy_ah(ah);
 269                 return PTR_ERR(m);
 270         }
 271
 272         /* Timeout set by caller if response is expected. */
 273         m->ah = ah;
 274         m->retries = cm_id_priv->max_cm_retries;
 275
 276         atomic_inc(&cm_id_priv->refcount);
 277         m->context[0] = cm_id_priv;
 278         *msg = m;
 279         return 0;
 280 }
 281
 282 static int cm_alloc_response_msg(struct cm_port *port,
 283                                  struct ib_mad_recv_wc *mad_recv_wc,
 284                                  struct ib_mad_send_buf **msg)
 285 {
 286         struct ib_mad_send_buf *m;
 287         struct ib_ah *ah;
 288
 289         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 290                                   mad_recv_wc->recv_buf.grh, port->port_num);
 291         if (IS_ERR(ah))
 292                 return PTR_ERR(ah);
 293
 294         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 295                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 296                                GFP_ATOMIC);
 297         if (IS_ERR(m)) {
 298                 ib_destroy_ah(ah);
 299                 return PTR_ERR(m);
 300         }
 301         m->ah = ah;
 302         *msg = m;
 303         return 0;
 304 }
 305
 306 static void cm_free_msg(struct ib_mad_send_buf *msg)
 307 {
 308         ib_destroy_ah(msg->ah);
 309         if (msg->context[0])
 310                 cm_deref_id(msg->context[0]);
 311         ib_free_send_mad(msg);
 312 }
 313
 314 static void * cm_copy_private_data(const void *private_data,
 315                                    u8 private_data_len)
 316 {
 317         void *data;
 318
 319         if (!private_data || !private_data_len)
 320                 return NULL;
 321
 322         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 323         if (!data)
 324                 return ERR_PTR(-ENOMEM);
 325
 326         return data;
 327 }
 328
 329 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 330                                  void *private_data, u8 private_data_len)
 331 {
 332         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 333                 kfree(cm_id_priv->private_data);
 334
 335         cm_id_priv->private_data = private_data;
 336         cm_id_priv->private_data_len = private_data_len;
 337 }
 338
 339 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 340                                     struct ib_grh *grh, struct cm_av *av)
 341 {
 342         av->port = port;
 343         av->pkey_index = wc->pkey_index;
 344         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 345                            grh, &av->ah_attr);
 346 }
 347
 348 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
 349 {
 350         struct cm_device *cm_dev;
 351         struct cm_port *port = NULL;
 352         unsigned long flags;
 353         int ret;
 354         u8 p;
 355
 356         read_lock_irqsave(&cm.device_lock, flags);
 357         list_for_each_entry(cm_dev, &cm.device_list, list) {
 358                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 359                                         &p, NULL)) {
 360                         port = cm_dev->port[p-1];
 361                         break;
 362                 }
 363         }
 364         read_unlock_irqrestore(&cm.device_lock, flags);
 365
 366         if (!port)
 367                 return -EINVAL;
 368
 369         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 370                                   be16_to_cpu(path->pkey), &av->pkey_index);
 371         if (ret)
 372                 return ret;
 373
 374         av->port = port;
 375         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 376                              &av->ah_attr);
 377         av->timeout = path->packet_life_time + 1;
 378         return 0;
 379 }
 380
 381 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 382 {
 383         unsigned long flags;
 384         int ret, id;
 385         static int next_id;
 386
 387         do {
 388                 spin_lock_irqsave(&cm.lock, flags);
 389                 ret = idr_get_new_above(&cm.local_id_table, cm_id_priv,
 390                                         next_id, &id);
 391                 if (!ret)
 392                         next_id = ((unsigned) id + 1) & MAX_ID_MASK;
 393                 spin_unlock_irqrestore(&cm.lock, flags);
 394         } while( (ret == -EAGAIN) && idr_pre_get(&cm.local_id_table, GFP_KERNEL) );
 395
 396         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 397         return ret;
 398 }
 399
 400 static void cm_free_id(__be32 local_id)
 401 {
 402         spin_lock_irq(&cm.lock);
 403         idr_remove(&cm.local_id_table,
 404                    (__force int) (local_id ^ cm.random_id_operand));
 405         spin_unlock_irq(&cm.lock);
 406 }
 407
 408 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 409 {
 410         struct cm_id_private *cm_id_priv;
 411
 412         cm_id_priv = idr_find(&cm.local_id_table,
 413                               (__force int) (local_id ^ cm.random_id_operand));
 414         if (cm_id_priv) {
 415                 if (cm_id_priv->id.remote_id == remote_id)
 416                         atomic_inc(&cm_id_priv->refcount);
 417                 else
 418                         cm_id_priv = NULL;
 419         }
 420
 421         return cm_id_priv;
 422 }
 423
 424 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 425 {
 426         struct cm_id_private *cm_id_priv;
 427
 428         spin_lock_irq(&cm.lock);
 429         cm_id_priv = cm_get_id(local_id, remote_id);
 430         spin_unlock_irq(&cm.lock);
 431
 432         return cm_id_priv;
 433 }
 434
 435 static void cm_mask_copy(u8 *dst, u8 *src, u8 *mask)
 436 {
 437         int i;
 438
 439         for (i = 0; i < IB_CM_COMPARE_SIZE / sizeof(unsigned long); i++)
 440                 ((unsigned long *) dst)[i] = ((unsigned long *) src)[i] &
 441                                              ((unsigned long *) mask)[i];
 442 }
 443
 444 static int cm_compare_data(struct ib_cm_compare_data *src_data,
 445                            struct ib_cm_compare_data *dst_data)
 446 {
 447         u8 src[IB_CM_COMPARE_SIZE];
 448         u8 dst[IB_CM_COMPARE_SIZE];
 449
 450         if (!src_data || !dst_data)
 451                 return 0;
 452
 453         cm_mask_copy(src, src_data->data, dst_data->mask);
 454         cm_mask_copy(dst, dst_data->data, src_data->mask);
 455         return memcmp(src, dst, IB_CM_COMPARE_SIZE);
 456 }
 457
 458 static int cm_compare_private_data(u8 *private_data,
 459                                    struct ib_cm_compare_data *dst_data)
 460 {
 461         u8 src[IB_CM_COMPARE_SIZE];
 462
 463         if (!dst_data)
 464                 return 0;
 465
 466         cm_mask_copy(src, private_data, dst_data->mask);
 467         return memcmp(src, dst_data->data, IB_CM_COMPARE_SIZE);
 468 }
 469
 470 /*
 471  * Trivial helpers to strip endian annotation and compare; the
 472  * endianness doesn't actually matter since we just need a stable
 473  * order for the RB tree.
 474  */
 475 static int be32_lt(__be32 a, __be32 b)
 476 {
 477         return (__force u32) a < (__force u32) b;
 478 }
 479
 480 static int be32_gt(__be32 a, __be32 b)
 481 {
 482         return (__force u32) a > (__force u32) b;
 483 }
 484
 485 static int be64_lt(__be64 a, __be64 b)
 486 {
 487         return (__force u64) a < (__force u64) b;
 488 }
 489
 490 static int be64_gt(__be64 a, __be64 b)
 491 {
 492         return (__force u64) a > (__force u64) b;
 493 }
 494
 495 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 496 {
 497         struct rb_node **link = &cm.listen_service_table.rb_node;
 498         struct rb_node *parent = NULL;
 499         struct cm_id_private *cur_cm_id_priv;
 500         __be64 service_id = cm_id_priv->id.service_id;
 501         __be64 service_mask = cm_id_priv->id.service_mask;
 502         int data_cmp;
 503
 504         while (*link) {
 505                 parent = *link;
 506                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 507                                           service_node);
 508                 data_cmp = cm_compare_data(cm_id_priv->compare_data,
 509                                            cur_cm_id_priv->compare_data);
 510                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 511                     (service_mask & cur_cm_id_priv->id.service_id) &&
 512                     (cm_id_priv->id.device == cur_cm_id_priv->id.device) &&
 513                     !data_cmp)
 514                         return cur_cm_id_priv;
 515
 516                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 517                         link = &(*link)->rb_left;
 518                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 519                         link = &(*link)->rb_right;
 520                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 521                         link = &(*link)->rb_left;
 522                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 523                         link = &(*link)->rb_right;
 524                 else if (data_cmp < 0)
 525                         link = &(*link)->rb_left;
 526                 else
 527                         link = &(*link)->rb_right;
 528         }
 529         rb_link_node(&cm_id_priv->service_node, parent, link);
 530         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 531         return NULL;
 532 }
 533
 534 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 535                                              __be64 service_id,
 536                                              u8 *private_data)
 537 {
 538         struct rb_node *node = cm.listen_service_table.rb_node;
 539         struct cm_id_private *cm_id_priv;
 540         int data_cmp;
 541
 542         while (node) {
 543                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 544                 data_cmp = cm_compare_private_data(private_data,
 545                                                    cm_id_priv->compare_data);
 546                 if ((cm_id_priv->id.service_mask & service_id) ==
 547                      cm_id_priv->id.service_id &&
 548                     (cm_id_priv->id.device == device) && !data_cmp)
 549                         return cm_id_priv;
 550
 551                 if (device < cm_id_priv->id.device)
 552                         node = node->rb_left;
 553                 else if (device > cm_id_priv->id.device)
 554                         node = node->rb_right;
 555                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 556                         node = node->rb_left;
 557                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 558                         node = node->rb_right;
 559                 else if (data_cmp < 0)
 560                         node = node->rb_left;
 561                 else
 562                         node = node->rb_right;
 563         }
 564         return NULL;
 565 }
 566
 567 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 568                                                      *timewait_info)
 569 {
 570         struct rb_node **link = &cm.remote_id_table.rb_node;
 571         struct rb_node *parent = NULL;
 572         struct cm_timewait_info *cur_timewait_info;
 573         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 574         __be32 remote_id = timewait_info->work.remote_id;
 575
 576         while (*link) {
 577                 parent = *link;
 578                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 579                                              remote_id_node);
 580                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 581                         link = &(*link)->rb_left;
 582                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 583                         link = &(*link)->rb_right;
 584                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 585                         link = &(*link)->rb_left;
 586                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 587                         link = &(*link)->rb_right;
 588                 else
 589                         return cur_timewait_info;
 590         }
 591         timewait_info->inserted_remote_id = 1;
 592         rb_link_node(&timewait_info->remote_id_node, parent, link);
 593         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 594         return NULL;
 595 }
 596
 597 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 598                                                    __be32 remote_id)
 599 {
 600         struct rb_node *node = cm.remote_id_table.rb_node;
 601         struct cm_timewait_info *timewait_info;
 602
 603         while (node) {
 604                 timewait_info = rb_entry(node, struct cm_timewait_info,
 605                                          remote_id_node);
 606                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 607                         node = node->rb_left;
 608                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 609                         node = node->rb_right;
 610                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 611                         node = node->rb_left;
 612                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 613                         node = node->rb_right;
 614                 else
 615                         return timewait_info;
 616         }
 617         return NULL;
 618 }
 619
 620 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 621                                                       *timewait_info)
 622 {
 623         struct rb_node **link = &cm.remote_qp_table.rb_node;
 624         struct rb_node *parent = NULL;
 625         struct cm_timewait_info *cur_timewait_info;
 626         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 627         __be32 remote_qpn = timewait_info->remote_qpn;
 628
 629         while (*link) {
 630                 parent = *link;
 631                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 632                                              remote_qp_node);
 633                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 634                         link = &(*link)->rb_left;
 635                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 636                         link = &(*link)->rb_right;
 637                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 638                         link = &(*link)->rb_left;
 639                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 640                         link = &(*link)->rb_right;
 641                 else
 642                         return cur_timewait_info;
 643         }
 644         timewait_info->inserted_remote_qp = 1;
 645         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 646         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 647         return NULL;
 648 }
 649
 650 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 651                                                     *cm_id_priv)
 652 {
 653         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 654         struct rb_node *parent = NULL;
 655         struct cm_id_private *cur_cm_id_priv;
 656         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 657         __be32 remote_id = cm_id_priv->id.remote_id;
 658
 659         while (*link) {
 660                 parent = *link;
 661                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 662                                           sidr_id_node);
 663                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 664                         link = &(*link)->rb_left;
 665                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 666                         link = &(*link)->rb_right;
 667                 else {
 668                         int cmp;
 669                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 670                                      sizeof *port_gid);
 671                         if (cmp < 0)
 672                                 link = &(*link)->rb_left;
 673                         else if (cmp > 0)
 674                                 link = &(*link)->rb_right;
 675                         else
 676                                 return cur_cm_id_priv;
 677                 }
 678         }
 679         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 680         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 681         return NULL;
 682 }
 683
 684 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 685                                enum ib_cm_sidr_status status)
 686 {
 687         struct ib_cm_sidr_rep_param param;
 688
 689         memset(&param, 0, sizeof param);
 690         param.status = status;
 691         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 692 }
 693
 694 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 695                                  ib_cm_handler cm_handler,
 696                                  void *context)
 697 {
 698         struct cm_id_private *cm_id_priv;
 699         int ret;
 700
 701         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 702         if (!cm_id_priv)
 703                 return ERR_PTR(-ENOMEM);
 704
 705         cm_id_priv->id.state = IB_CM_IDLE;
 706         cm_id_priv->id.device = device;
 707         cm_id_priv->id.cm_handler = cm_handler;
 708         cm_id_priv->id.context = context;
 709         cm_id_priv->id.remote_cm_qpn = 1;
 710         ret = cm_alloc_id(cm_id_priv);
 711         if (ret)
 712                 goto error;
 713
 714         spin_lock_init(&cm_id_priv->lock);
 715         init_completion(&cm_id_priv->comp);
 716         INIT_LIST_HEAD(&cm_id_priv->work_list);
 717         atomic_set(&cm_id_priv->work_count, -1);
 718         atomic_set(&cm_id_priv->refcount, 1);
 719         return &cm_id_priv->id;
 720
 721 error:
 722         kfree(cm_id_priv);
 723         return ERR_PTR(-ENOMEM);
 724 }
 725 EXPORT_SYMBOL(ib_create_cm_id);
 726
 727 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 728 {
 729         struct cm_work *work;
 730
 731         if (list_empty(&cm_id_priv->work_list))
 732                 return NULL;
 733
 734         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 735         list_del(&work->list);
 736         return work;
 737 }
 738
 739 static void cm_free_work(struct cm_work *work)
 740 {
 741         if (work->mad_recv_wc)
 742                 ib_free_recv_mad(work->mad_recv_wc);
 743         kfree(work);
 744 }
 745
 746 static inline int cm_convert_to_ms(int iba_time)
 747 {
 748         /* approximate conversion to ms from 4.096us x 2^iba_time */
 749         return 1 << max(iba_time - 8, 0);
 750 }
 751
 752 /*
 753  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 754  * Because of how ack_timeout is stored, adding one doubles the timeout.
 755  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 756  * increment it (round up) only if the other is within 50%.
 757  */
 758 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 759 {
 760         int ack_timeout = packet_life_time + 1;
 761
 762         if (ack_timeout >= ca_ack_delay)
 763                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 764         else
 765                 ack_timeout = ca_ack_delay +
 766                               (ack_timeout >= (ca_ack_delay - 1));
 767
 768         return min(31, ack_timeout);
 769 }
 770
 771 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 772 {
 773         if (timewait_info->inserted_remote_id) {
 774                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 775                 timewait_info->inserted_remote_id = 0;
 776         }
 777
 778         if (timewait_info->inserted_remote_qp) {
 779                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 780                 timewait_info->inserted_remote_qp = 0;
 781         }
 782 }
 783
 784 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 785 {
 786         struct cm_timewait_info *timewait_info;
 787
 788         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 789         if (!timewait_info)
 790                 return ERR_PTR(-ENOMEM);
 791
 792         timewait_info->work.local_id = local_id;
 793         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 794         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 795         return timewait_info;
 796 }
 797
 798 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 799 {
 800         int wait_time;
 801         unsigned long flags;
 802
 803         spin_lock_irqsave(&cm.lock, flags);
 804         cm_cleanup_timewait(cm_id_priv->timewait_info);
 805         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 806         spin_unlock_irqrestore(&cm.lock, flags);
 807
 808         /*
 809          * The cm_id could be destroyed by the user before we exit timewait.
 810          * To protect against this, we search for the cm_id after exiting
 811          * timewait before notifying the user that we've exited timewait.
 812          */
 813         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 814         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 815         queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 816                            msecs_to_jiffies(wait_time));
 817         cm_id_priv->timewait_info = NULL;
 818 }
 819
 820 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 821 {
 822         unsigned long flags;
 823
 824         cm_id_priv->id.state = IB_CM_IDLE;
 825         if (cm_id_priv->timewait_info) {
 826                 spin_lock_irqsave(&cm.lock, flags);
 827                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 828                 spin_unlock_irqrestore(&cm.lock, flags);
 829                 kfree(cm_id_priv->timewait_info);
 830                 cm_id_priv->timewait_info = NULL;
 831         }
 832 }
 833
 834 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 835 {
 836         struct cm_id_private *cm_id_priv;
 837         struct cm_work *work;
 838
 839         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 840 retest:
 841         spin_lock_irq(&cm_id_priv->lock);
 842         switch (cm_id->state) {
 843         case IB_CM_LISTEN:
 844                 cm_id->state = IB_CM_IDLE;
 845                 spin_unlock_irq(&cm_id_priv->lock);
 846                 spin_lock_irq(&cm.lock);
 847                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 848                 spin_unlock_irq(&cm.lock);
 849                 break;
 850         case IB_CM_SIDR_REQ_SENT:
 851                 cm_id->state = IB_CM_IDLE;
 852                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 853                 spin_unlock_irq(&cm_id_priv->lock);
 854                 break;
 855         case IB_CM_SIDR_REQ_RCVD:
 856                 spin_unlock_irq(&cm_id_priv->lock);
 857                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 858                 break;
 859         case IB_CM_REQ_SENT:
 860                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 861                 spin_unlock_irq(&cm_id_priv->lock);
 862                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 863                                &cm_id_priv->id.device->node_guid,
 864                                sizeof cm_id_priv->id.device->node_guid,
 865                                NULL, 0);
 866                 break;
 867         case IB_CM_REQ_RCVD:
 868                 if (err == -ENOMEM) {
 869                         /* Do not reject to allow future retries. */
 870                         cm_reset_to_idle(cm_id_priv);
 871                         spin_unlock_irq(&cm_id_priv->lock);
 872                 } else {
 873                         spin_unlock_irq(&cm_id_priv->lock);
 874                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 875                                        NULL, 0, NULL, 0);
 876                 }
 877                 break;
 878         case IB_CM_MRA_REQ_RCVD:
 879         case IB_CM_REP_SENT:
 880         case IB_CM_MRA_REP_RCVD:
 881                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 882                 /* Fall through */
 883         case IB_CM_MRA_REQ_SENT:
 884         case IB_CM_REP_RCVD:
 885         case IB_CM_MRA_REP_SENT:
 886                 spin_unlock_irq(&cm_id_priv->lock);
 887                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 888                                NULL, 0, NULL, 0);
 889                 break;
 890         case IB_CM_ESTABLISHED:
 891                 spin_unlock_irq(&cm_id_priv->lock);
 892                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT)
 893                         break;
 894                 ib_send_cm_dreq(cm_id, NULL, 0);
 895                 goto retest;
 896         case IB_CM_DREQ_SENT:
 897                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 898                 cm_enter_timewait(cm_id_priv);
 899                 spin_unlock_irq(&cm_id_priv->lock);
 900                 break;
 901         case IB_CM_DREQ_RCVD:
 902                 spin_unlock_irq(&cm_id_priv->lock);
 903                 ib_send_cm_drep(cm_id, NULL, 0);
 904                 break;
 905         default:
 906                 spin_unlock_irq(&cm_id_priv->lock);
 907                 break;
 908         }
 909
 910         cm_free_id(cm_id->local_id);
 911         cm_deref_id(cm_id_priv);
 912         wait_for_completion(&cm_id_priv->comp);
 913         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
 914                 cm_free_work(work);
 915         kfree(cm_id_priv->compare_data);
 916         kfree(cm_id_priv->private_data);
 917         kfree(cm_id_priv);
 918 }
 919
 920 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
 921 {
 922         cm_destroy_id(cm_id, 0);
 923 }
 924 EXPORT_SYMBOL(ib_destroy_cm_id);
 925
 926 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask,
 927                  struct ib_cm_compare_data *compare_data)
 928 {
 929         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
 930         unsigned long flags;
 931         int ret = 0;
 932
 933         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
 934         service_id &= service_mask;
 935         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
 936             (service_id != IB_CM_ASSIGN_SERVICE_ID))
 937                 return -EINVAL;
 938
 939         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 940         if (cm_id->state != IB_CM_IDLE)
 941                 return -EINVAL;
 942
 943         if (compare_data) {
 944                 cm_id_priv->compare_data = kzalloc(sizeof *compare_data,
 945                                                    GFP_KERNEL);
 946                 if (!cm_id_priv->compare_data)
 947                         return -ENOMEM;
 948                 cm_mask_copy(cm_id_priv->compare_data->data,
 949                              compare_data->data, compare_data->mask);
 950                 memcpy(cm_id_priv->compare_data->mask, compare_data->mask,
 951                        IB_CM_COMPARE_SIZE);
 952         }
 953
 954         cm_id->state = IB_CM_LISTEN;
 955
 956         spin_lock_irqsave(&cm.lock, flags);
 957         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
 958                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
 959                 cm_id->service_mask = ~cpu_to_be64(0);
 960         } else {
 961                 cm_id->service_id = service_id;
 962                 cm_id->service_mask = service_mask;
 963         }
 964         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
 965         spin_unlock_irqrestore(&cm.lock, flags);
 966
 967         if (cur_cm_id_priv) {
 968                 cm_id->state = IB_CM_IDLE;
 969                 kfree(cm_id_priv->compare_data);
 970                 cm_id_priv->compare_data = NULL;
 971                 ret = -EBUSY;
 972         }
 973         return ret;
 974 }
 975 EXPORT_SYMBOL(ib_cm_listen);
 976
 977 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
 978                           enum cm_msg_sequence msg_seq)
 979 {
 980         u64 hi_tid, low_tid;
 981
 982         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
 983         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
 984                           (msg_seq << 30));
 985         return cpu_to_be64(hi_tid | low_tid);
 986 }
 987
 988 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
 989                               __be16 attr_id, __be64 tid)
 990 {
 991         hdr->base_version  = IB_MGMT_BASE_VERSION;
 992         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
 993         hdr->class_version = IB_CM_CLASS_VERSION;
 994         hdr->method        = IB_MGMT_METHOD_SEND;
 995         hdr->attr_id       = attr_id;
 996         hdr->tid           = tid;
 997 }
 998
 999 static void cm_format_req(struct cm_req_msg *req_msg,
1000                           struct cm_id_private *cm_id_priv,
1001                           struct ib_cm_req_param *param)
1002 {
1003         struct ib_sa_path_rec *pri_path = param->primary_path;
1004         struct ib_sa_path_rec *alt_path = param->alternate_path;
1005
1006         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1007                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1008
1009         req_msg->local_comm_id = cm_id_priv->id.local_id;
1010         req_msg->service_id = param->service_id;
1011         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1012         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1013         cm_req_set_init_depth(req_msg, param->initiator_depth);
1014         cm_req_set_remote_resp_timeout(req_msg,
1015                                        param->remote_cm_response_timeout);
1016         cm_req_set_qp_type(req_msg, param->qp_type);
1017         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1018         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1019         cm_req_set_local_resp_timeout(req_msg,
1020                                       param->local_cm_response_timeout);
1021         req_msg->pkey = param->primary_path->pkey;
1022         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1023         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1024
1025         if (param->qp_type != IB_QPT_XRC_INI) {
1026                 cm_req_set_resp_res(req_msg, param->responder_resources);
1027                 cm_req_set_retry_count(req_msg, param->retry_count);
1028                 cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1029                 cm_req_set_srq(req_msg, param->srq);
1030         }
1031
1032         if (pri_path->hop_limit <= 1) {
1033                 req_msg->primary_local_lid = pri_path->slid;
1034                 req_msg->primary_remote_lid = pri_path->dlid;
1035         } else {
1036                 /* Work-around until there's a way to obtain remote LID info */
1037                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1038                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1039         }
1040         req_msg->primary_local_gid = pri_path->sgid;
1041         req_msg->primary_remote_gid = pri_path->dgid;
1042         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1043         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1044         req_msg->primary_traffic_class = pri_path->traffic_class;
1045         req_msg->primary_hop_limit = pri_path->hop_limit;
1046         cm_req_set_primary_sl(req_msg, pri_path->sl);
1047         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1048         cm_req_set_primary_local_ack_timeout(req_msg,
1049                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1050                                pri_path->packet_life_time));
1051
1052         if (alt_path) {
1053                 if (alt_path->hop_limit <= 1) {
1054                         req_msg->alt_local_lid = alt_path->slid;
1055                         req_msg->alt_remote_lid = alt_path->dlid;
1056                 } else {
1057                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1058                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1059                 }
1060                 req_msg->alt_local_gid = alt_path->sgid;
1061                 req_msg->alt_remote_gid = alt_path->dgid;
1062                 cm_req_set_alt_flow_label(req_msg,
1063                                           alt_path->flow_label);
1064                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1065                 req_msg->alt_traffic_class = alt_path->traffic_class;
1066                 req_msg->alt_hop_limit = alt_path->hop_limit;
1067                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1068                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1069                 cm_req_set_alt_local_ack_timeout(req_msg,
1070                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1071                                        alt_path->packet_life_time));
1072         }
1073
1074         if (param->private_data && param->private_data_len)
1075                 memcpy(req_msg->private_data, param->private_data,
1076                        param->private_data_len);
1077 }
1078
1079 static int cm_validate_req_param(struct ib_cm_req_param *param)
1080 {
1081         /* peer-to-peer not supported */
1082         if (param->peer_to_peer)
1083                 return -EINVAL;
1084
1085         if (!param->primary_path)
1086                 return -EINVAL;
1087
1088         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1089             param->qp_type != IB_QPT_XRC_INI)
1090                 return -EINVAL;
1091
1092         if (param->private_data &&
1093             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1094                 return -EINVAL;
1095
1096         if (param->alternate_path &&
1097             (param->alternate_path->pkey != param->primary_path->pkey ||
1098              param->alternate_path->mtu != param->primary_path->mtu))
1099                 return -EINVAL;
1100
1101         return 0;
1102 }
1103
1104 int ib_send_cm_req(struct ib_cm_id *cm_id,
1105                    struct ib_cm_req_param *param)
1106 {
1107         struct cm_id_private *cm_id_priv;
1108         struct cm_req_msg *req_msg;
1109         unsigned long flags;
1110         int ret;
1111
1112         ret = cm_validate_req_param(param);
1113         if (ret)
1114                 return ret;
1115
1116         /* Verify that we're not in timewait. */
1117         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1118         spin_lock_irqsave(&cm_id_priv->lock, flags);
1119         if (cm_id->state != IB_CM_IDLE) {
1120                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1121                 ret = -EINVAL;
1122                 goto out;
1123         }
1124         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1125
1126         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1127                                                             id.local_id);
1128         if (IS_ERR(cm_id_priv->timewait_info)) {
1129                 ret = PTR_ERR(cm_id_priv->timewait_info);
1130                 goto out;
1131         }
1132
1133         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av);
1134         if (ret)
1135                 goto error1;
1136         if (param->alternate_path) {
1137                 ret = cm_init_av_by_path(param->alternate_path,
1138                                          &cm_id_priv->alt_av);
1139                 if (ret)
1140                         goto error1;
1141         }
1142         cm_id->service_id = param->service_id;
1143         cm_id->service_mask = ~cpu_to_be64(0);
1144         cm_id_priv->timeout_ms = cm_convert_to_ms(
1145                                     param->primary_path->packet_life_time) * 2 +
1146                                  cm_convert_to_ms(
1147                                     param->remote_cm_response_timeout);
1148         cm_id_priv->max_cm_retries = param->max_cm_retries;
1149         cm_id_priv->initiator_depth = param->initiator_depth;
1150         cm_id_priv->responder_resources = param->responder_resources;
1151         cm_id_priv->retry_count = param->retry_count;
1152         cm_id_priv->path_mtu = param->primary_path->mtu;
1153         cm_id_priv->pkey = param->primary_path->pkey;
1154         cm_id_priv->qp_type = param->qp_type;
1155
1156         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1157         if (ret)
1158                 goto error1;
1159
1160         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1161         cm_format_req(req_msg, cm_id_priv, param);
1162         cm_id_priv->tid = req_msg->hdr.tid;
1163         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1164         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1165
1166         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1167         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1168
1169         spin_lock_irqsave(&cm_id_priv->lock, flags);
1170         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1171         if (ret) {
1172                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1173                 goto error2;
1174         }
1175         BUG_ON(cm_id->state != IB_CM_IDLE);
1176         cm_id->state = IB_CM_REQ_SENT;
1177         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1178         return 0;
1179
1180 error2: cm_free_msg(cm_id_priv->msg);
1181 error1: kfree(cm_id_priv->timewait_info);
1182 out:    return ret;
1183 }
1184 EXPORT_SYMBOL(ib_send_cm_req);
1185
1186 static int cm_issue_rej(struct cm_port *port,
1187                         struct ib_mad_recv_wc *mad_recv_wc,
1188                         enum ib_cm_rej_reason reason,
1189                         enum cm_msg_response msg_rejected,
1190                         void *ari, u8 ari_length)
1191 {
1192         struct ib_mad_send_buf *msg = NULL;
1193         struct cm_rej_msg *rej_msg, *rcv_msg;
1194         int ret;
1195
1196         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1197         if (ret)
1198                 return ret;
1199
1200         /* We just need common CM header information.  Cast to any message. */
1201         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1202         rej_msg = (struct cm_rej_msg *) msg->mad;
1203
1204         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1205         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1206         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1207         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1208         rej_msg->reason = cpu_to_be16(reason);
1209
1210         if (ari && ari_length) {
1211                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1212                 memcpy(rej_msg->ari, ari, ari_length);
1213         }
1214
1215         ret = ib_post_send_mad(msg, NULL);
1216         if (ret)
1217                 cm_free_msg(msg);
1218
1219         return ret;
1220 }
1221
1222 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1223                                     __be32 local_qpn, __be32 remote_qpn)
1224 {
1225         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1226                 ((local_ca_guid == remote_ca_guid) &&
1227                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1228 }
1229
1230 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1231                                             struct ib_sa_path_rec *primary_path,
1232                                             struct ib_sa_path_rec *alt_path)
1233 {
1234         memset(primary_path, 0, sizeof *primary_path);
1235         primary_path->dgid = req_msg->primary_local_gid;
1236         primary_path->sgid = req_msg->primary_remote_gid;
1237         primary_path->dlid = req_msg->primary_local_lid;
1238         primary_path->slid = req_msg->primary_remote_lid;
1239         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1240         primary_path->hop_limit = req_msg->primary_hop_limit;
1241         primary_path->traffic_class = req_msg->primary_traffic_class;
1242         primary_path->reversible = 1;
1243         primary_path->pkey = req_msg->pkey;
1244         primary_path->sl = cm_req_get_primary_sl(req_msg);
1245         primary_path->mtu_selector = IB_SA_EQ;
1246         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1247         primary_path->rate_selector = IB_SA_EQ;
1248         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1249         primary_path->packet_life_time_selector = IB_SA_EQ;
1250         primary_path->packet_life_time =
1251                 cm_req_get_primary_local_ack_timeout(req_msg);
1252         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1253
1254         if (req_msg->alt_local_lid) {
1255                 memset(alt_path, 0, sizeof *alt_path);
1256                 alt_path->dgid = req_msg->alt_local_gid;
1257                 alt_path->sgid = req_msg->alt_remote_gid;
1258                 alt_path->dlid = req_msg->alt_local_lid;
1259                 alt_path->slid = req_msg->alt_remote_lid;
1260                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1261                 alt_path->hop_limit = req_msg->alt_hop_limit;
1262                 alt_path->traffic_class = req_msg->alt_traffic_class;
1263                 alt_path->reversible = 1;
1264                 alt_path->pkey = req_msg->pkey;
1265                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1266                 alt_path->mtu_selector = IB_SA_EQ;
1267                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1268                 alt_path->rate_selector = IB_SA_EQ;
1269                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1270                 alt_path->packet_life_time_selector = IB_SA_EQ;
1271                 alt_path->packet_life_time =
1272                         cm_req_get_alt_local_ack_timeout(req_msg);
1273                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1274         }
1275 }
1276
1277 static void cm_format_req_event(struct cm_work *work,
1278                                 struct cm_id_private *cm_id_priv,
1279                                 struct ib_cm_id *listen_id)
1280 {
1281         struct cm_req_msg *req_msg;
1282         struct ib_cm_req_event_param *param;
1283
1284         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1285         param = &work->cm_event.param.req_rcvd;
1286         param->listen_id = listen_id;
1287         param->port = cm_id_priv->av.port->port_num;
1288         param->primary_path = &work->path[0];
1289         if (req_msg->alt_local_lid)
1290                 param->alternate_path = &work->path[1];
1291         else
1292                 param->alternate_path = NULL;
1293         param->remote_ca_guid = req_msg->local_ca_guid;
1294         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1295         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1296         param->qp_type = cm_req_get_qp_type(req_msg);
1297         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1298         param->responder_resources = cm_req_get_init_depth(req_msg);
1299         param->initiator_depth = cm_req_get_resp_res(req_msg);
1300         param->local_cm_response_timeout =
1301                                         cm_req_get_remote_resp_timeout(req_msg);
1302         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1303         param->remote_cm_response_timeout =
1304                                         cm_req_get_local_resp_timeout(req_msg);
1305         param->retry_count = cm_req_get_retry_count(req_msg);
1306         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1307         param->srq = cm_req_get_srq(req_msg);
1308         work->cm_event.private_data = &req_msg->private_data;
1309 }
1310
1311 static void cm_process_work(struct cm_id_private *cm_id_priv,
1312                             struct cm_work *work)
1313 {
1314         int ret;
1315
1316         /* We will typically only have the current event to report. */
1317         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1318         cm_free_work(work);
1319
1320         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1321                 spin_lock_irq(&cm_id_priv->lock);
1322                 work = cm_dequeue_work(cm_id_priv);
1323                 spin_unlock_irq(&cm_id_priv->lock);
1324                 BUG_ON(!work);
1325                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1326                                                 &work->cm_event);
1327                 cm_free_work(work);
1328         }
1329         cm_deref_id(cm_id_priv);
1330         if (ret)
1331                 cm_destroy_id(&cm_id_priv->id, ret);
1332 }
1333
1334 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1335                           struct cm_id_private *cm_id_priv,
1336                           enum cm_msg_response msg_mraed, u8 service_timeout,
1337                           const void *private_data, u8 private_data_len)
1338 {
1339         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1340         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1341         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1342         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1343         cm_mra_set_service_timeout(mra_msg, service_timeout);
1344
1345         if (private_data && private_data_len)
1346                 memcpy(mra_msg->private_data, private_data, private_data_len);
1347 }
1348
1349 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1350                           struct cm_id_private *cm_id_priv,
1351                           enum ib_cm_rej_reason reason,
1352                           void *ari,
1353                           u8 ari_length,
1354                           const void *private_data,
1355                           u8 private_data_len)
1356 {
1357         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1358         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1359
1360         switch(cm_id_priv->id.state) {
1361         case IB_CM_REQ_RCVD:
1362                 rej_msg->local_comm_id = 0;
1363                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1364                 break;
1365         case IB_CM_MRA_REQ_SENT:
1366                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1367                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1368                 break;
1369         case IB_CM_REP_RCVD:
1370         case IB_CM_MRA_REP_SENT:
1371                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1372                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1373                 break;
1374         default:
1375                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1376                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1377                 break;
1378         }
1379
1380         rej_msg->reason = cpu_to_be16(reason);
1381         if (ari && ari_length) {
1382                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1383                 memcpy(rej_msg->ari, ari, ari_length);
1384         }
1385
1386         if (private_data && private_data_len)
1387                 memcpy(rej_msg->private_data, private_data, private_data_len);
1388 }
1389
1390 static void cm_dup_req_handler(struct cm_work *work,
1391                                struct cm_id_private *cm_id_priv)
1392 {
1393         struct ib_mad_send_buf *msg = NULL;
1394         int ret;
1395
1396         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1397                         counter[CM_REQ_COUNTER]);
1398
1399         /* Quick state check to discard duplicate REQs. */
1400         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1401                 return;
1402
1403         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1404         if (ret)
1405                 return;
1406
1407         spin_lock_irq(&cm_id_priv->lock);
1408         switch (cm_id_priv->id.state) {
1409         case IB_CM_MRA_REQ_SENT:
1410                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1411                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1412                               cm_id_priv->private_data,
1413                               cm_id_priv->private_data_len);
1414                 break;
1415         case IB_CM_TIMEWAIT:
1416                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1417                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1418                 break;
1419         default:
1420                 goto unlock;
1421         }
1422         spin_unlock_irq(&cm_id_priv->lock);
1423
1424         ret = ib_post_send_mad(msg, NULL);
1425         if (ret)
1426                 goto free;
1427         return;
1428
1429 unlock: spin_unlock_irq(&cm_id_priv->lock);
1430 free:   cm_free_msg(msg);
1431 }
1432
1433 static struct cm_id_private * cm_match_req(struct cm_work *work,
1434                                            struct cm_id_private *cm_id_priv)
1435 {
1436         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1437         struct cm_timewait_info *timewait_info;
1438         struct cm_req_msg *req_msg;
1439
1440         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1441
1442         /* Check for possible duplicate REQ. */
1443         spin_lock_irq(&cm.lock);
1444         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1445         if (timewait_info) {
1446                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1447                                            timewait_info->work.remote_id);
1448                 spin_unlock_irq(&cm.lock);
1449                 if (cur_cm_id_priv) {
1450                         cm_dup_req_handler(work, cur_cm_id_priv);
1451                         cm_deref_id(cur_cm_id_priv);
1452                 }
1453                 return NULL;
1454         }
1455
1456         /* Check for stale connections. */
1457         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1458         if (timewait_info) {
1459                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1460                 spin_unlock_irq(&cm.lock);
1461                 cm_issue_rej(work->port, work->mad_recv_wc,
1462                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1463                              NULL, 0);
1464                 return NULL;
1465         }
1466
1467         /* Find matching listen request. */
1468         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1469                                            req_msg->service_id,
1470                                            req_msg->private_data);
1471         if (!listen_cm_id_priv) {
1472                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1473                 spin_unlock_irq(&cm.lock);
1474                 cm_issue_rej(work->port, work->mad_recv_wc,
1475                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1476                              NULL, 0);
1477                 goto out;
1478         }
1479         atomic_inc(&listen_cm_id_priv->refcount);
1480         atomic_inc(&cm_id_priv->refcount);
1481         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1482         atomic_inc(&cm_id_priv->work_count);
1483         spin_unlock_irq(&cm.lock);
1484 out:
1485         return listen_cm_id_priv;
1486 }
1487
1488 /*
1489  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1490  * we need to override the LID/SL data in the REQ with the LID information
1491  * in the work completion.
1492  */
1493 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1494 {
1495         if (!cm_req_get_primary_subnet_local(req_msg)) {
1496                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1497                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1498                         cm_req_set_primary_sl(req_msg, wc->sl);
1499                 }
1500
1501                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1502                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1503         }
1504
1505         if (!cm_req_get_alt_subnet_local(req_msg)) {
1506                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1507                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1508                         cm_req_set_alt_sl(req_msg, wc->sl);
1509                 }
1510
1511                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1512                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1513         }
1514 }
1515
1516 static int cm_req_handler(struct cm_work *work)
1517 {
1518         struct ib_cm_id *cm_id;
1519         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1520         struct cm_req_msg *req_msg;
1521         int ret;
1522
1523         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1524
1525         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1526         if (IS_ERR(cm_id))
1527                 return PTR_ERR(cm_id);
1528
1529         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1530         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1531         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1532                                 work->mad_recv_wc->recv_buf.grh,
1533                                 &cm_id_priv->av);
1534         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1535                                                             id.local_id);
1536         if (IS_ERR(cm_id_priv->timewait_info)) {
1537                 ret = PTR_ERR(cm_id_priv->timewait_info);
1538                 goto destroy;
1539         }
1540         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1541         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1542         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1543
1544         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1545         if (!listen_cm_id_priv) {
1546                 ret = -EINVAL;
1547                 kfree(cm_id_priv->timewait_info);
1548                 goto destroy;
1549         }
1550
1551         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1552         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1553         cm_id_priv->id.service_id = req_msg->service_id;
1554         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1555
1556         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1557         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1558         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
1559         if (ret) {
1560                 ib_get_cached_gid(work->port->cm_dev->ib_device,
1561                                   work->port->port_num, 0, &work->path[0].sgid);
1562                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1563                                &work->path[0].sgid, sizeof work->path[0].sgid,
1564                                NULL, 0);
1565                 goto rejected;
1566         }
1567         if (req_msg->alt_local_lid) {
1568                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
1569                 if (ret) {
1570                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1571                                        &work->path[0].sgid,
1572                                        sizeof work->path[0].sgid, NULL, 0);
1573                         goto rejected;
1574                 }
1575         }
1576         cm_id_priv->tid = req_msg->hdr.tid;
1577         cm_id_priv->timeout_ms = cm_convert_to_ms(
1578                                         cm_req_get_local_resp_timeout(req_msg));
1579         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1580         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1581         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1582         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1583         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1584         cm_id_priv->pkey = req_msg->pkey;
1585         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1586         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1587         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1588         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1589
1590         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1591         cm_process_work(cm_id_priv, work);
1592         cm_deref_id(listen_cm_id_priv);
1593         return 0;
1594
1595 rejected:
1596         atomic_dec(&cm_id_priv->refcount);
1597         cm_deref_id(listen_cm_id_priv);
1598 destroy:
1599         ib_destroy_cm_id(cm_id);
1600         return ret;
1601 }
1602
1603 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1604                           struct cm_id_private *cm_id_priv,
1605                           struct ib_cm_rep_param *param)
1606 {
1607         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1608         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1609         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1610         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1611         rep_msg->resp_resources = param->responder_resources;
1612         cm_rep_set_target_ack_delay(rep_msg,
1613                                     cm_id_priv->av.port->cm_dev->ack_delay);
1614         cm_rep_set_failover(rep_msg, param->failover_accepted);
1615         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1616         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1617
1618         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1619                 rep_msg->initiator_depth = param->initiator_depth;
1620                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1621                 cm_rep_set_srq(rep_msg, param->srq);
1622                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1623         } else {
1624                 cm_rep_set_srq(rep_msg, 1);
1625                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1626         }
1627
1628         if (param->private_data && param->private_data_len)
1629                 memcpy(rep_msg->private_data, param->private_data,
1630                        param->private_data_len);
1631 }
1632
1633 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1634                    struct ib_cm_rep_param *param)
1635 {
1636         struct cm_id_private *cm_id_priv;
1637         struct ib_mad_send_buf *msg;
1638         struct cm_rep_msg *rep_msg;
1639         unsigned long flags;
1640         int ret;
1641
1642         if (param->private_data &&
1643             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1644                 return -EINVAL;
1645
1646         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1647         spin_lock_irqsave(&cm_id_priv->lock, flags);
1648         if (cm_id->state != IB_CM_REQ_RCVD &&
1649             cm_id->state != IB_CM_MRA_REQ_SENT) {
1650                 ret = -EINVAL;
1651                 goto out;
1652         }
1653
1654         ret = cm_alloc_msg(cm_id_priv, &msg);
1655         if (ret)
1656                 goto out;
1657
1658         rep_msg = (struct cm_rep_msg *) msg->mad;
1659         cm_format_rep(rep_msg, cm_id_priv, param);
1660         msg->timeout_ms = cm_id_priv->timeout_ms;
1661         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1662
1663         ret = ib_post_send_mad(msg, NULL);
1664         if (ret) {
1665                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1666                 cm_free_msg(msg);
1667                 return ret;
1668         }
1669
1670         cm_id->state = IB_CM_REP_SENT;
1671         cm_id_priv->msg = msg;
1672         cm_id_priv->initiator_depth = param->initiator_depth;
1673         cm_id_priv->responder_resources = param->responder_resources;
1674         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1675         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
1676
1677 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1678         return ret;
1679 }
1680 EXPORT_SYMBOL(ib_send_cm_rep);
1681
1682 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1683                           struct cm_id_private *cm_id_priv,
1684                           const void *private_data,
1685                           u8 private_data_len)
1686 {
1687         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1688         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1689         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1690
1691         if (private_data && private_data_len)
1692                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1693 }
1694
1695 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1696                    const void *private_data,
1697                    u8 private_data_len)
1698 {
1699         struct cm_id_private *cm_id_priv;
1700         struct ib_mad_send_buf *msg;
1701         unsigned long flags;
1702         void *data;
1703         int ret;
1704
1705         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1706                 return -EINVAL;
1707
1708         data = cm_copy_private_data(private_data, private_data_len);
1709         if (IS_ERR(data))
1710                 return PTR_ERR(data);
1711
1712         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1713         spin_lock_irqsave(&cm_id_priv->lock, flags);
1714         if (cm_id->state != IB_CM_REP_RCVD &&
1715             cm_id->state != IB_CM_MRA_REP_SENT) {
1716                 ret = -EINVAL;
1717                 goto error;
1718         }
1719
1720         ret = cm_alloc_msg(cm_id_priv, &msg);
1721         if (ret)
1722                 goto error;
1723
1724         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1725                       private_data, private_data_len);
1726
1727         ret = ib_post_send_mad(msg, NULL);
1728         if (ret) {
1729                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1730                 cm_free_msg(msg);
1731                 kfree(data);
1732                 return ret;
1733         }
1734
1735         cm_id->state = IB_CM_ESTABLISHED;
1736         cm_set_private_data(cm_id_priv, data, private_data_len);
1737         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1738         return 0;
1739
1740 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1741         kfree(data);
1742         return ret;
1743 }
1744 EXPORT_SYMBOL(ib_send_cm_rtu);
1745
1746 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
1747 {
1748         struct cm_rep_msg *rep_msg;
1749         struct ib_cm_rep_event_param *param;
1750
1751         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1752         param = &work->cm_event.param.rep_rcvd;
1753         param->remote_ca_guid = rep_msg->local_ca_guid;
1754         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1755         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
1756         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1757         param->responder_resources = rep_msg->initiator_depth;
1758         param->initiator_depth = rep_msg->resp_resources;
1759         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1760         param->failover_accepted = cm_rep_get_failover(rep_msg);
1761         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1762         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1763         param->srq = cm_rep_get_srq(rep_msg);
1764         work->cm_event.private_data = &rep_msg->private_data;
1765 }
1766
1767 static void cm_dup_rep_handler(struct cm_work *work)
1768 {
1769         struct cm_id_private *cm_id_priv;
1770         struct cm_rep_msg *rep_msg;
1771         struct ib_mad_send_buf *msg = NULL;
1772         int ret;
1773
1774         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1775         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1776                                    rep_msg->local_comm_id);
1777         if (!cm_id_priv)
1778                 return;
1779
1780         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1781                         counter[CM_REP_COUNTER]);
1782         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1783         if (ret)
1784                 goto deref;
1785
1786         spin_lock_irq(&cm_id_priv->lock);
1787         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1788                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1789                               cm_id_priv->private_data,
1790                               cm_id_priv->private_data_len);
1791         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1792                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1793                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1794                               cm_id_priv->private_data,
1795                               cm_id_priv->private_data_len);
1796         else
1797                 goto unlock;
1798         spin_unlock_irq(&cm_id_priv->lock);
1799
1800         ret = ib_post_send_mad(msg, NULL);
1801         if (ret)
1802                 goto free;
1803         goto deref;
1804
1805 unlock: spin_unlock_irq(&cm_id_priv->lock);
1806 free:   cm_free_msg(msg);
1807 deref:  cm_deref_id(cm_id_priv);
1808 }
1809
1810 static int cm_rep_handler(struct cm_work *work)
1811 {
1812         struct cm_id_private *cm_id_priv;
1813         struct cm_rep_msg *rep_msg;
1814         int ret;
1815
1816         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1817         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1818         if (!cm_id_priv) {
1819                 cm_dup_rep_handler(work);
1820                 return -EINVAL;
1821         }
1822
1823         cm_format_rep_event(work, cm_id_priv->qp_type);
1824
1825         spin_lock_irq(&cm_id_priv->lock);
1826         switch (cm_id_priv->id.state) {
1827         case IB_CM_REQ_SENT:
1828         case IB_CM_MRA_REQ_RCVD:
1829                 break;
1830         default:
1831                 spin_unlock_irq(&cm_id_priv->lock);
1832                 ret = -EINVAL;
1833                 goto error;
1834         }
1835
1836         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1837         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1838         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
1839
1840         spin_lock(&cm.lock);
1841         /* Check for duplicate REP. */
1842         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1843                 spin_unlock(&cm.lock);
1844                 spin_unlock_irq(&cm_id_priv->lock);
1845                 ret = -EINVAL;
1846                 goto error;
1847         }
1848         /* Check for a stale connection. */
1849         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
1850                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
1851                          &cm.remote_id_table);
1852                 cm_id_priv->timewait_info->inserted_remote_id = 0;
1853                 spin_unlock(&cm.lock);
1854                 spin_unlock_irq(&cm_id_priv->lock);
1855                 cm_issue_rej(work->port, work->mad_recv_wc,
1856                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
1857                              NULL, 0);
1858                 ret = -EINVAL;
1859                 goto error;
1860         }
1861         spin_unlock(&cm.lock);
1862
1863         cm_id_priv->id.state = IB_CM_REP_RCVD;
1864         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
1865         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
1866         cm_id_priv->initiator_depth = rep_msg->resp_resources;
1867         cm_id_priv->responder_resources = rep_msg->initiator_depth;
1868         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
1869         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1870         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1871         cm_id_priv->av.timeout =
1872                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1873                                        cm_id_priv->av.timeout - 1);
1874         cm_id_priv->alt_av.timeout =
1875                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1876                                        cm_id_priv->alt_av.timeout - 1);
1877
1878         /* todo: handle peer_to_peer */
1879
1880         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1881         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1882         if (!ret)
1883                 list_add_tail(&work->list, &cm_id_priv->work_list);
1884         spin_unlock_irq(&cm_id_priv->lock);
1885
1886         if (ret)
1887                 cm_process_work(cm_id_priv, work);
1888         else
1889                 cm_deref_id(cm_id_priv);
1890         return 0;
1891
1892 error:
1893         cm_deref_id(cm_id_priv);
1894         return ret;
1895 }
1896
1897 static int cm_establish_handler(struct cm_work *work)
1898 {
1899         struct cm_id_private *cm_id_priv;
1900         int ret;
1901
1902         /* See comment in cm_establish about lookup. */
1903         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
1904         if (!cm_id_priv)
1905                 return -EINVAL;
1906
1907         spin_lock_irq(&cm_id_priv->lock);
1908         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
1909                 spin_unlock_irq(&cm_id_priv->lock);
1910                 goto out;
1911         }
1912
1913         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1914         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1915         if (!ret)
1916                 list_add_tail(&work->list, &cm_id_priv->work_list);
1917         spin_unlock_irq(&cm_id_priv->lock);
1918
1919         if (ret)
1920                 cm_process_work(cm_id_priv, work);
1921         else
1922                 cm_deref_id(cm_id_priv);
1923         return 0;
1924 out:
1925         cm_deref_id(cm_id_priv);
1926         return -EINVAL;
1927 }
1928
1929 static int cm_rtu_handler(struct cm_work *work)
1930 {
1931         struct cm_id_private *cm_id_priv;
1932         struct cm_rtu_msg *rtu_msg;
1933         int ret;
1934
1935         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
1936         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
1937                                    rtu_msg->local_comm_id);
1938         if (!cm_id_priv)
1939                 return -EINVAL;
1940
1941         work->cm_event.private_data = &rtu_msg->private_data;
1942
1943         spin_lock_irq(&cm_id_priv->lock);
1944         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
1945             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
1946                 spin_unlock_irq(&cm_id_priv->lock);
1947                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1948                                 counter[CM_RTU_COUNTER]);
1949                 goto out;
1950         }
1951         cm_id_priv->id.state = IB_CM_ESTABLISHED;
1952
1953         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1954         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1955         if (!ret)
1956                 list_add_tail(&work->list, &cm_id_priv->work_list);
1957         spin_unlock_irq(&cm_id_priv->lock);
1958
1959         if (ret)
1960                 cm_process_work(cm_id_priv, work);
1961         else
1962                 cm_deref_id(cm_id_priv);
1963         return 0;
1964 out:
1965         cm_deref_id(cm_id_priv);
1966         return -EINVAL;
1967 }
1968
1969 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
1970                           struct cm_id_private *cm_id_priv,
1971                           const void *private_data,
1972                           u8 private_data_len)
1973 {
1974         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
1975                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
1976         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
1977         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
1978         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
1979
1980         if (private_data && private_data_len)
1981                 memcpy(dreq_msg->private_data, private_data, private_data_len);
1982 }
1983
1984 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
1985                     const void *private_data,
1986                     u8 private_data_len)
1987 {
1988         struct cm_id_private *cm_id_priv;
1989         struct ib_mad_send_buf *msg;
1990         unsigned long flags;
1991         int ret;
1992
1993         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
1994                 return -EINVAL;
1995
1996         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1997         spin_lock_irqsave(&cm_id_priv->lock, flags);
1998         if (cm_id->state != IB_CM_ESTABLISHED) {
1999                 ret = -EINVAL;
2000                 goto out;
2001         }
2002
2003         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2004             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2005                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2006
2007         ret = cm_alloc_msg(cm_id_priv, &msg);
2008         if (ret) {
2009                 cm_enter_timewait(cm_id_priv);
2010                 goto out;
2011         }
2012
2013         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2014                        private_data, private_data_len);
2015         msg->timeout_ms = cm_id_priv->timeout_ms;
2016         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2017
2018         ret = ib_post_send_mad(msg, NULL);
2019         if (ret) {
2020                 cm_enter_timewait(cm_id_priv);
2021                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2022                 cm_free_msg(msg);
2023                 return ret;
2024         }
2025
2026         cm_id->state = IB_CM_DREQ_SENT;
2027         cm_id_priv->msg = msg;
2028 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2029         return ret;
2030 }
2031 EXPORT_SYMBOL(ib_send_cm_dreq);
2032
2033 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2034                           struct cm_id_private *cm_id_priv,
2035                           const void *private_data,
2036                           u8 private_data_len)
2037 {
2038         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2039         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2040         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2041
2042         if (private_data && private_data_len)
2043                 memcpy(drep_msg->private_data, private_data, private_data_len);
2044 }
2045
2046 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2047                     const void *private_data,
2048                     u8 private_data_len)
2049 {
2050         struct cm_id_private *cm_id_priv;
2051         struct ib_mad_send_buf *msg;
2052         unsigned long flags;
2053         void *data;
2054         int ret;
2055
2056         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2057                 return -EINVAL;
2058
2059         data = cm_copy_private_data(private_data, private_data_len);
2060         if (IS_ERR(data))
2061                 return PTR_ERR(data);
2062
2063         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2064         spin_lock_irqsave(&cm_id_priv->lock, flags);
2065         if (cm_id->state != IB_CM_DREQ_RCVD) {
2066                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2067                 kfree(data);
2068                 return -EINVAL;
2069         }
2070
2071         cm_set_private_data(cm_id_priv, data, private_data_len);
2072         cm_enter_timewait(cm_id_priv);
2073
2074         ret = cm_alloc_msg(cm_id_priv, &msg);
2075         if (ret)
2076                 goto out;
2077
2078         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2079                        private_data, private_data_len);
2080
2081         ret = ib_post_send_mad(msg, NULL);
2082         if (ret) {
2083                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2084                 cm_free_msg(msg);
2085                 return ret;
2086         }
2087
2088 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2089         return ret;
2090 }
2091 EXPORT_SYMBOL(ib_send_cm_drep);
2092
2093 static int cm_issue_drep(struct cm_port *port,
2094                          struct ib_mad_recv_wc *mad_recv_wc)
2095 {
2096         struct ib_mad_send_buf *msg = NULL;
2097         struct cm_dreq_msg *dreq_msg;
2098         struct cm_drep_msg *drep_msg;
2099         int ret;
2100
2101         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2102         if (ret)
2103                 return ret;
2104
2105         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2106         drep_msg = (struct cm_drep_msg *) msg->mad;
2107
2108         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2109         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2110         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2111
2112         ret = ib_post_send_mad(msg, NULL);
2113         if (ret)
2114                 cm_free_msg(msg);
2115
2116         return ret;
2117 }
2118
2119 static int cm_dreq_handler(struct cm_work *work)
2120 {
2121         struct cm_id_private *cm_id_priv;
2122         struct cm_dreq_msg *dreq_msg;
2123         struct ib_mad_send_buf *msg = NULL;
2124         int ret;
2125
2126         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2127         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2128                                    dreq_msg->local_comm_id);
2129         if (!cm_id_priv) {
2130                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2131                                 counter[CM_DREQ_COUNTER]);
2132                 cm_issue_drep(work->port, work->mad_recv_wc);
2133                 return -EINVAL;
2134         }
2135
2136         work->cm_event.private_data = &dreq_msg->private_data;
2137
2138         spin_lock_irq(&cm_id_priv->lock);
2139         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2140                 goto unlock;
2141
2142         switch (cm_id_priv->id.state) {
2143         case IB_CM_REP_SENT:
2144         case IB_CM_DREQ_SENT:
2145                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2146                 break;
2147         case IB_CM_ESTABLISHED:
2148                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2149                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2150                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2151                 break;
2152         case IB_CM_MRA_REP_RCVD:
2153                 break;
2154         case IB_CM_TIMEWAIT:
2155                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2156                                 counter[CM_DREQ_COUNTER]);
2157                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2158                         goto unlock;
2159
2160                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2161                                cm_id_priv->private_data,
2162                                cm_id_priv->private_data_len);
2163                 spin_unlock_irq(&cm_id_priv->lock);
2164
2165                 if (ib_post_send_mad(msg, NULL))
2166                         cm_free_msg(msg);
2167                 goto deref;
2168         case IB_CM_DREQ_RCVD:
2169                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2170                                 counter[CM_DREQ_COUNTER]);
2171                 goto unlock;
2172         default:
2173                 goto unlock;
2174         }
2175         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2176         cm_id_priv->tid = dreq_msg->hdr.tid;
2177         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2178         if (!ret)
2179                 list_add_tail(&work->list, &cm_id_priv->work_list);
2180         spin_unlock_irq(&cm_id_priv->lock);
2181
2182         if (ret)
2183                 cm_process_work(cm_id_priv, work);
2184         else
2185                 cm_deref_id(cm_id_priv);
2186         return 0;
2187
2188 unlock: spin_unlock_irq(&cm_id_priv->lock);
2189 deref:  cm_deref_id(cm_id_priv);
2190         return -EINVAL;
2191 }
2192
2193 static int cm_drep_handler(struct cm_work *work)
2194 {
2195         struct cm_id_private *cm_id_priv;
2196         struct cm_drep_msg *drep_msg;
2197         int ret;
2198
2199         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2200         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2201                                    drep_msg->local_comm_id);
2202         if (!cm_id_priv)
2203                 return -EINVAL;
2204
2205         work->cm_event.private_data = &drep_msg->private_data;
2206
2207         spin_lock_irq(&cm_id_priv->lock);
2208         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2209             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2210                 spin_unlock_irq(&cm_id_priv->lock);
2211                 goto out;
2212         }
2213         cm_enter_timewait(cm_id_priv);
2214
2215         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2216         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2217         if (!ret)
2218                 list_add_tail(&work->list, &cm_id_priv->work_list);
2219         spin_unlock_irq(&cm_id_priv->lock);
2220
2221         if (ret)
2222                 cm_process_work(cm_id_priv, work);
2223         else
2224                 cm_deref_id(cm_id_priv);
2225         return 0;
2226 out:
2227         cm_deref_id(cm_id_priv);
2228         return -EINVAL;
2229 }
2230
2231 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2232                    enum ib_cm_rej_reason reason,
2233                    void *ari,
2234                    u8 ari_length,
2235                    const void *private_data,
2236                    u8 private_data_len)
2237 {
2238         struct cm_id_private *cm_id_priv;
2239         struct ib_mad_send_buf *msg;
2240         unsigned long flags;
2241         int ret;
2242
2243         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2244             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2245                 return -EINVAL;
2246
2247         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2248
2249         spin_lock_irqsave(&cm_id_priv->lock, flags);
2250         switch (cm_id->state) {
2251         case IB_CM_REQ_SENT:
2252         case IB_CM_MRA_REQ_RCVD:
2253         case IB_CM_REQ_RCVD:
2254         case IB_CM_MRA_REQ_SENT:
2255         case IB_CM_REP_RCVD:
2256         case IB_CM_MRA_REP_SENT:
2257                 ret = cm_alloc_msg(cm_id_priv, &msg);
2258                 if (!ret)
2259                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2260                                       cm_id_priv, reason, ari, ari_length,
2261                                       private_data, private_data_len);
2262
2263                 cm_reset_to_idle(cm_id_priv);
2264                 break;
2265         case IB_CM_REP_SENT:
2266         case IB_CM_MRA_REP_RCVD:
2267                 ret = cm_alloc_msg(cm_id_priv, &msg);
2268                 if (!ret)
2269                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2270                                       cm_id_priv, reason, ari, ari_length,
2271                                       private_data, private_data_len);
2272
2273                 cm_enter_timewait(cm_id_priv);
2274                 break;
2275         default:
2276                 ret = -EINVAL;
2277                 goto out;
2278         }
2279
2280         if (ret)
2281                 goto out;
2282
2283         ret = ib_post_send_mad(msg, NULL);
2284         if (ret)
2285                 cm_free_msg(msg);
2286
2287 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2288         return ret;
2289 }
2290 EXPORT_SYMBOL(ib_send_cm_rej);
2291
2292 static void cm_format_rej_event(struct cm_work *work)
2293 {
2294         struct cm_rej_msg *rej_msg;
2295         struct ib_cm_rej_event_param *param;
2296
2297         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2298         param = &work->cm_event.param.rej_rcvd;
2299         param->ari = rej_msg->ari;
2300         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2301         param->reason = __be16_to_cpu(rej_msg->reason);
2302         work->cm_event.private_data = &rej_msg->private_data;
2303 }
2304
2305 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2306 {
2307         struct cm_timewait_info *timewait_info;
2308         struct cm_id_private *cm_id_priv;
2309         __be32 remote_id;
2310
2311         remote_id = rej_msg->local_comm_id;
2312
2313         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2314                 spin_lock_irq(&cm.lock);
2315                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2316                                                   remote_id);
2317                 if (!timewait_info) {
2318                         spin_unlock_irq(&cm.lock);
2319                         return NULL;
2320                 }
2321                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2322                                       (timewait_info->work.local_id ^
2323                                        cm.random_id_operand));
2324                 if (cm_id_priv) {
2325                         if (cm_id_priv->id.remote_id == remote_id)
2326                                 atomic_inc(&cm_id_priv->refcount);
2327                         else
2328                                 cm_id_priv = NULL;
2329                 }
2330                 spin_unlock_irq(&cm.lock);
2331         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2332                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2333         else
2334                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2335
2336         return cm_id_priv;
2337 }
2338
2339 static int cm_rej_handler(struct cm_work *work)
2340 {
2341         struct cm_id_private *cm_id_priv;
2342         struct cm_rej_msg *rej_msg;
2343         int ret;
2344
2345         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2346         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2347         if (!cm_id_priv)
2348                 return -EINVAL;
2349
2350         cm_format_rej_event(work);
2351
2352         spin_lock_irq(&cm_id_priv->lock);
2353         switch (cm_id_priv->id.state) {
2354         case IB_CM_REQ_SENT:
2355         case IB_CM_MRA_REQ_RCVD:
2356         case IB_CM_REP_SENT:
2357         case IB_CM_MRA_REP_RCVD:
2358                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2359                 /* fall through */
2360         case IB_CM_REQ_RCVD:
2361         case IB_CM_MRA_REQ_SENT:
2362                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2363                         cm_enter_timewait(cm_id_priv);
2364                 else
2365                         cm_reset_to_idle(cm_id_priv);
2366                 break;
2367         case IB_CM_DREQ_SENT:
2368                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2369                 /* fall through */
2370         case IB_CM_REP_RCVD:
2371         case IB_CM_MRA_REP_SENT:
2372                 cm_enter_timewait(cm_id_priv);
2373                 break;
2374         case IB_CM_ESTABLISHED:
2375                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2376                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2377                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2378                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2379                                               cm_id_priv->msg);
2380                         cm_enter_timewait(cm_id_priv);
2381                         break;
2382                 }
2383                 /* fall through */
2384         default:
2385                 spin_unlock_irq(&cm_id_priv->lock);
2386                 ret = -EINVAL;
2387                 goto out;
2388         }
2389
2390         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2391         if (!ret)
2392                 list_add_tail(&work->list, &cm_id_priv->work_list);
2393         spin_unlock_irq(&cm_id_priv->lock);
2394
2395         if (ret)
2396                 cm_process_work(cm_id_priv, work);
2397         else
2398                 cm_deref_id(cm_id_priv);
2399         return 0;
2400 out:
2401         cm_deref_id(cm_id_priv);
2402         return -EINVAL;
2403 }
2404
2405 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2406                    u8 service_timeout,
2407                    const void *private_data,
2408                    u8 private_data_len)
2409 {
2410         struct cm_id_private *cm_id_priv;
2411         struct ib_mad_send_buf *msg;
2412         enum ib_cm_state cm_state;
2413         enum ib_cm_lap_state lap_state;
2414         enum cm_msg_response msg_response;
2415         void *data;
2416         unsigned long flags;
2417         int ret;
2418
2419         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2420                 return -EINVAL;
2421
2422         data = cm_copy_private_data(private_data, private_data_len);
2423         if (IS_ERR(data))
2424                 return PTR_ERR(data);
2425
2426         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2427
2428         spin_lock_irqsave(&cm_id_priv->lock, flags);
2429         switch(cm_id_priv->id.state) {
2430         case IB_CM_REQ_RCVD:
2431                 cm_state = IB_CM_MRA_REQ_SENT;
2432                 lap_state = cm_id->lap_state;
2433                 msg_response = CM_MSG_RESPONSE_REQ;
2434                 break;
2435         case IB_CM_REP_RCVD:
2436                 cm_state = IB_CM_MRA_REP_SENT;
2437                 lap_state = cm_id->lap_state;
2438                 msg_response = CM_MSG_RESPONSE_REP;
2439                 break;
2440         case IB_CM_ESTABLISHED:
2441                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2442                         cm_state = cm_id->state;
2443                         lap_state = IB_CM_MRA_LAP_SENT;
2444                         msg_response = CM_MSG_RESPONSE_OTHER;
2445                         break;
2446                 }
2447         default:
2448                 ret = -EINVAL;
2449                 goto error1;
2450         }
2451
2452         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2453                 ret = cm_alloc_msg(cm_id_priv, &msg);
2454                 if (ret)
2455                         goto error1;
2456
2457                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2458                               msg_response, service_timeout,
2459                               private_data, private_data_len);
2460                 ret = ib_post_send_mad(msg, NULL);
2461                 if (ret)
2462                         goto error2;
2463         }
2464
2465         cm_id->state = cm_state;
2466         cm_id->lap_state = lap_state;
2467         cm_id_priv->service_timeout = service_timeout;
2468         cm_set_private_data(cm_id_priv, data, private_data_len);
2469         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2470         return 0;
2471
2472 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2473         kfree(data);
2474         return ret;
2475
2476 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2477         kfree(data);
2478         cm_free_msg(msg);
2479         return ret;
2480 }
2481 EXPORT_SYMBOL(ib_send_cm_mra);
2482
2483 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2484 {
2485         switch (cm_mra_get_msg_mraed(mra_msg)) {
2486         case CM_MSG_RESPONSE_REQ:
2487                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2488         case CM_MSG_RESPONSE_REP:
2489         case CM_MSG_RESPONSE_OTHER:
2490                 return cm_acquire_id(mra_msg->remote_comm_id,
2491                                      mra_msg->local_comm_id);
2492         default:
2493                 return NULL;
2494         }
2495 }
2496
2497 static int cm_mra_handler(struct cm_work *work)
2498 {
2499         struct cm_id_private *cm_id_priv;
2500         struct cm_mra_msg *mra_msg;
2501         int timeout, ret;
2502
2503         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2504         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2505         if (!cm_id_priv)
2506                 return -EINVAL;
2507
2508         work->cm_event.private_data = &mra_msg->private_data;
2509         work->cm_event.param.mra_rcvd.service_timeout =
2510                                         cm_mra_get_service_timeout(mra_msg);
2511         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2512                   cm_convert_to_ms(cm_id_priv->av.timeout);
2513
2514         spin_lock_irq(&cm_id_priv->lock);
2515         switch (cm_id_priv->id.state) {
2516         case IB_CM_REQ_SENT:
2517                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2518                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2519                                   cm_id_priv->msg, timeout))
2520                         goto out;
2521                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2522                 break;
2523         case IB_CM_REP_SENT:
2524                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2525                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2526                                   cm_id_priv->msg, timeout))
2527                         goto out;
2528                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2529                 break;
2530         case IB_CM_ESTABLISHED:
2531                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2532                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2533                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2534                                   cm_id_priv->msg, timeout)) {
2535                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2536                                 atomic_long_inc(&work->port->
2537                                                 counter_group[CM_RECV_DUPLICATES].
2538                                                 counter[CM_MRA_COUNTER]);
2539                         goto out;
2540                 }
2541                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2542                 break;
2543         case IB_CM_MRA_REQ_RCVD:
2544         case IB_CM_MRA_REP_RCVD:
2545                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2546                                 counter[CM_MRA_COUNTER]);
2547                 /* fall through */
2548         default:
2549                 goto out;
2550         }
2551
2552         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2553                                       cm_id_priv->id.state;
2554         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2555         if (!ret)
2556                 list_add_tail(&work->list, &cm_id_priv->work_list);
2557         spin_unlock_irq(&cm_id_priv->lock);
2558
2559         if (ret)
2560                 cm_process_work(cm_id_priv, work);
2561         else
2562                 cm_deref_id(cm_id_priv);
2563         return 0;
2564 out:
2565         spin_unlock_irq(&cm_id_priv->lock);
2566         cm_deref_id(cm_id_priv);
2567         return -EINVAL;
2568 }
2569
2570 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2571                           struct cm_id_private *cm_id_priv,
2572                           struct ib_sa_path_rec *alternate_path,
2573                           const void *private_data,
2574                           u8 private_data_len)
2575 {
2576         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2577                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2578         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2579         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2580         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2581         /* todo: need remote CM response timeout */
2582         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2583         lap_msg->alt_local_lid = alternate_path->slid;
2584         lap_msg->alt_remote_lid = alternate_path->dlid;
2585         lap_msg->alt_local_gid = alternate_path->sgid;
2586         lap_msg->alt_remote_gid = alternate_path->dgid;
2587         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2588         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2589         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2590         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2591         cm_lap_set_sl(lap_msg, alternate_path->sl);
2592         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2593         cm_lap_set_local_ack_timeout(lap_msg,
2594                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2595                                alternate_path->packet_life_time));
2596
2597         if (private_data && private_data_len)
2598                 memcpy(lap_msg->private_data, private_data, private_data_len);
2599 }
2600
2601 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2602                    struct ib_sa_path_rec *alternate_path,
2603                    const void *private_data,
2604                    u8 private_data_len)
2605 {
2606         struct cm_id_private *cm_id_priv;
2607         struct ib_mad_send_buf *msg;
2608         unsigned long flags;
2609         int ret;
2610
2611         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2612                 return -EINVAL;
2613
2614         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2615         spin_lock_irqsave(&cm_id_priv->lock, flags);
2616         if (cm_id->state != IB_CM_ESTABLISHED ||
2617             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2618              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2619                 ret = -EINVAL;
2620                 goto out;
2621         }
2622
2623         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
2624         if (ret)
2625                 goto out;
2626         cm_id_priv->alt_av.timeout =
2627                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2628                                        cm_id_priv->alt_av.timeout - 1);
2629
2630         ret = cm_alloc_msg(cm_id_priv, &msg);
2631         if (ret)
2632                 goto out;
2633
2634         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2635                       alternate_path, private_data, private_data_len);
2636         msg->timeout_ms = cm_id_priv->timeout_ms;
2637         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2638
2639         ret = ib_post_send_mad(msg, NULL);
2640         if (ret) {
2641                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2642                 cm_free_msg(msg);
2643                 return ret;
2644         }
2645
2646         cm_id->lap_state = IB_CM_LAP_SENT;
2647         cm_id_priv->msg = msg;
2648
2649 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2650         return ret;
2651 }
2652 EXPORT_SYMBOL(ib_send_cm_lap);
2653
2654 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2655                                     struct ib_sa_path_rec *path,
2656                                     struct cm_lap_msg *lap_msg)
2657 {
2658         memset(path, 0, sizeof *path);
2659         path->dgid = lap_msg->alt_local_gid;
2660         path->sgid = lap_msg->alt_remote_gid;
2661         path->dlid = lap_msg->alt_local_lid;
2662         path->slid = lap_msg->alt_remote_lid;
2663         path->flow_label = cm_lap_get_flow_label(lap_msg);
2664         path->hop_limit = lap_msg->alt_hop_limit;
2665         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2666         path->reversible = 1;
2667         path->pkey = cm_id_priv->pkey;
2668         path->sl = cm_lap_get_sl(lap_msg);
2669         path->mtu_selector = IB_SA_EQ;
2670         path->mtu = cm_id_priv->path_mtu;
2671         path->rate_selector = IB_SA_EQ;
2672         path->rate = cm_lap_get_packet_rate(lap_msg);
2673         path->packet_life_time_selector = IB_SA_EQ;
2674         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2675         path->packet_life_time -= (path->packet_life_time > 0);
2676 }
2677
2678 static int cm_lap_handler(struct cm_work *work)
2679 {
2680         struct cm_id_private *cm_id_priv;
2681         struct cm_lap_msg *lap_msg;
2682         struct ib_cm_lap_event_param *param;
2683         struct ib_mad_send_buf *msg = NULL;
2684         int ret;
2685
2686         /* todo: verify LAP request and send reject APR if invalid. */
2687         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2688         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2689                                    lap_msg->local_comm_id);
2690         if (!cm_id_priv)
2691                 return -EINVAL;
2692
2693         param = &work->cm_event.param.lap_rcvd;
2694         param->alternate_path = &work->path[0];
2695         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2696         work->cm_event.private_data = &lap_msg->private_data;
2697
2698         spin_lock_irq(&cm_id_priv->lock);
2699         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2700                 goto unlock;
2701
2702         switch (cm_id_priv->id.lap_state) {
2703         case IB_CM_LAP_UNINIT:
2704         case IB_CM_LAP_IDLE:
2705                 break;
2706         case IB_CM_MRA_LAP_SENT:
2707                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2708                                 counter[CM_LAP_COUNTER]);
2709                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2710                         goto unlock;
2711
2712                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2713                               CM_MSG_RESPONSE_OTHER,
2714                               cm_id_priv->service_timeout,
2715                               cm_id_priv->private_data,
2716                               cm_id_priv->private_data_len);
2717                 spin_unlock_irq(&cm_id_priv->lock);
2718
2719                 if (ib_post_send_mad(msg, NULL))
2720                         cm_free_msg(msg);
2721                 goto deref;
2722         case IB_CM_LAP_RCVD:
2723                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2724                                 counter[CM_LAP_COUNTER]);
2725                 goto unlock;
2726         default:
2727                 goto unlock;
2728         }
2729
2730         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2731         cm_id_priv->tid = lap_msg->hdr.tid;
2732         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2733                                 work->mad_recv_wc->recv_buf.grh,
2734                                 &cm_id_priv->av);
2735         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av);
2736         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2737         if (!ret)
2738                 list_add_tail(&work->list, &cm_id_priv->work_list);
2739         spin_unlock_irq(&cm_id_priv->lock);
2740
2741         if (ret)
2742                 cm_process_work(cm_id_priv, work);
2743         else
2744                 cm_deref_id(cm_id_priv);
2745         return 0;
2746
2747 unlock: spin_unlock_irq(&cm_id_priv->lock);
2748 deref:  cm_deref_id(cm_id_priv);
2749         return -EINVAL;
2750 }
2751
2752 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2753                           struct cm_id_private *cm_id_priv,
2754                           enum ib_cm_apr_status status,
2755                           void *info,
2756                           u8 info_length,
2757                           const void *private_data,
2758                           u8 private_data_len)
2759 {
2760         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2761         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2762         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2763         apr_msg->ap_status = (u8) status;
2764
2765         if (info && info_length) {
2766                 apr_msg->info_length = info_length;
2767                 memcpy(apr_msg->info, info, info_length);
2768         }
2769
2770         if (private_data && private_data_len)
2771                 memcpy(apr_msg->private_data, private_data, private_data_len);
2772 }
2773
2774 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2775                    enum ib_cm_apr_status status,
2776                    void *info,
2777                    u8 info_length,
2778                    const void *private_data,
2779                    u8 private_data_len)
2780 {
2781         struct cm_id_private *cm_id_priv;
2782         struct ib_mad_send_buf *msg;
2783         unsigned long flags;
2784         int ret;
2785
2786         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2787             (info && info_length > IB_CM_APR_INFO_LENGTH))
2788                 return -EINVAL;
2789
2790         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2791         spin_lock_irqsave(&cm_id_priv->lock, flags);
2792         if (cm_id->state != IB_CM_ESTABLISHED ||
2793             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2794              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2795                 ret = -EINVAL;
2796                 goto out;
2797         }
2798
2799         ret = cm_alloc_msg(cm_id_priv, &msg);
2800         if (ret)
2801                 goto out;
2802
2803         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2804                       info, info_length, private_data, private_data_len);
2805         ret = ib_post_send_mad(msg, NULL);
2806         if (ret) {
2807                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2808                 cm_free_msg(msg);
2809                 return ret;
2810         }
2811
2812         cm_id->lap_state = IB_CM_LAP_IDLE;
2813 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2814         return ret;
2815 }
2816 EXPORT_SYMBOL(ib_send_cm_apr);
2817
2818 static int cm_apr_handler(struct cm_work *work)
2819 {
2820         struct cm_id_private *cm_id_priv;
2821         struct cm_apr_msg *apr_msg;
2822         int ret;
2823
2824         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2825         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2826                                    apr_msg->local_comm_id);
2827         if (!cm_id_priv)
2828                 return -EINVAL; /* Unmatched reply. */
2829
2830         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2831         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2832         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2833         work->cm_event.private_data = &apr_msg->private_data;
2834
2835         spin_lock_irq(&cm_id_priv->lock);
2836         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2837             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2838              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2839                 spin_unlock_irq(&cm_id_priv->lock);
2840                 goto out;
2841         }
2842         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2843         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2844         cm_id_priv->msg = NULL;
2845
2846         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2847         if (!ret)
2848                 list_add_tail(&work->list, &cm_id_priv->work_list);
2849         spin_unlock_irq(&cm_id_priv->lock);
2850
2851         if (ret)
2852                 cm_process_work(cm_id_priv, work);
2853         else
2854                 cm_deref_id(cm_id_priv);
2855         return 0;
2856 out:
2857         cm_deref_id(cm_id_priv);
2858         return -EINVAL;
2859 }
2860
2861 static int cm_timewait_handler(struct cm_work *work)
2862 {
2863         struct cm_timewait_info *timewait_info;
2864         struct cm_id_private *cm_id_priv;
2865         int ret;
2866
2867         timewait_info = (struct cm_timewait_info *)work;
2868         spin_lock_irq(&cm.lock);
2869         list_del(&timewait_info->list);
2870         spin_unlock_irq(&cm.lock);
2871
2872         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2873                                    timewait_info->work.remote_id);
2874         if (!cm_id_priv)
2875                 return -EINVAL;
2876
2877         spin_lock_irq(&cm_id_priv->lock);
2878         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2879             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2880                 spin_unlock_irq(&cm_id_priv->lock);
2881                 goto out;
2882         }
2883         cm_id_priv->id.state = IB_CM_IDLE;
2884         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2885         if (!ret)
2886                 list_add_tail(&work->list, &cm_id_priv->work_list);
2887         spin_unlock_irq(&cm_id_priv->lock);
2888
2889         if (ret)
2890                 cm_process_work(cm_id_priv, work);
2891         else
2892                 cm_deref_id(cm_id_priv);
2893         return 0;
2894 out:
2895         cm_deref_id(cm_id_priv);
2896         return -EINVAL;
2897 }
2898
2899 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
2900                                struct cm_id_private *cm_id_priv,
2901                                struct ib_cm_sidr_req_param *param)
2902 {
2903         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
2904                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
2905         sidr_req_msg->request_id = cm_id_priv->id.local_id;
2906         sidr_req_msg->pkey = param->path->pkey;
2907         sidr_req_msg->service_id = param->service_id;
2908
2909         if (param->private_data && param->private_data_len)
2910                 memcpy(sidr_req_msg->private_data, param->private_data,
2911                        param->private_data_len);
2912 }
2913
2914 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
2915                         struct ib_cm_sidr_req_param *param)
2916 {
2917         struct cm_id_private *cm_id_priv;
2918         struct ib_mad_send_buf *msg;
2919         unsigned long flags;
2920         int ret;
2921
2922         if (!param->path || (param->private_data &&
2923              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
2924                 return -EINVAL;
2925
2926         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2927         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
2928         if (ret)
2929                 goto out;
2930
2931         cm_id->service_id = param->service_id;
2932         cm_id->service_mask = ~cpu_to_be64(0);
2933         cm_id_priv->timeout_ms = param->timeout_ms;
2934         cm_id_priv->max_cm_retries = param->max_cm_retries;
2935         ret = cm_alloc_msg(cm_id_priv, &msg);
2936         if (ret)
2937                 goto out;
2938
2939         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
2940                            param);
2941         msg->timeout_ms = cm_id_priv->timeout_ms;
2942         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
2943
2944         spin_lock_irqsave(&cm_id_priv->lock, flags);
2945         if (cm_id->state == IB_CM_IDLE)
2946                 ret = ib_post_send_mad(msg, NULL);
2947         else
2948                 ret = -EINVAL;
2949
2950         if (ret) {
2951                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2952                 cm_free_msg(msg);
2953                 goto out;
2954         }
2955         cm_id->state = IB_CM_SIDR_REQ_SENT;
2956         cm_id_priv->msg = msg;
2957         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2958 out:
2959         return ret;
2960 }
2961 EXPORT_SYMBOL(ib_send_cm_sidr_req);
2962
2963 static void cm_format_sidr_req_event(struct cm_work *work,
2964                                      struct ib_cm_id *listen_id)
2965 {
2966         struct cm_sidr_req_msg *sidr_req_msg;
2967         struct ib_cm_sidr_req_event_param *param;
2968
2969         sidr_req_msg = (struct cm_sidr_req_msg *)
2970                                 work->mad_recv_wc->recv_buf.mad;
2971         param = &work->cm_event.param.sidr_req_rcvd;
2972         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
2973         param->listen_id = listen_id;
2974         param->port = work->port->port_num;
2975         work->cm_event.private_data = &sidr_req_msg->private_data;
2976 }
2977
2978 static int cm_sidr_req_handler(struct cm_work *work)
2979 {
2980         struct ib_cm_id *cm_id;
2981         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
2982         struct cm_sidr_req_msg *sidr_req_msg;
2983         struct ib_wc *wc;
2984
2985         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
2986         if (IS_ERR(cm_id))
2987                 return PTR_ERR(cm_id);
2988         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2989
2990         /* Record SGID/SLID and request ID for lookup. */
2991         sidr_req_msg = (struct cm_sidr_req_msg *)
2992                                 work->mad_recv_wc->recv_buf.mad;
2993         wc = work->mad_recv_wc->wc;
2994         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
2995         cm_id_priv->av.dgid.global.interface_id = 0;
2996         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2997                                 work->mad_recv_wc->recv_buf.grh,
2998                                 &cm_id_priv->av);
2999         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3000         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3001         atomic_inc(&cm_id_priv->work_count);
3002
3003         spin_lock_irq(&cm.lock);
3004         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3005         if (cur_cm_id_priv) {
3006                 spin_unlock_irq(&cm.lock);
3007                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3008                                 counter[CM_SIDR_REQ_COUNTER]);
3009                 goto out; /* Duplicate message. */
3010         }
3011         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3012         cur_cm_id_priv = cm_find_listen(cm_id->device,
3013                                         sidr_req_msg->service_id,
3014                                         sidr_req_msg->private_data);
3015         if (!cur_cm_id_priv) {
3016                 spin_unlock_irq(&cm.lock);
3017                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3018                 goto out; /* No match. */
3019         }
3020         atomic_inc(&cur_cm_id_priv->refcount);
3021         atomic_inc(&cm_id_priv->refcount);
3022         spin_unlock_irq(&cm.lock);
3023
3024         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3025         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3026         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3027         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3028
3029         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3030         cm_process_work(cm_id_priv, work);
3031         cm_deref_id(cur_cm_id_priv);
3032         return 0;
3033 out:
3034         ib_destroy_cm_id(&cm_id_priv->id);
3035         return -EINVAL;
3036 }
3037
3038 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3039                                struct cm_id_private *cm_id_priv,
3040                                struct ib_cm_sidr_rep_param *param)
3041 {
3042         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3043                           cm_id_priv->tid);
3044         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3045         sidr_rep_msg->status = param->status;
3046         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3047         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3048         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3049
3050         if (param->info && param->info_length)
3051                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3052
3053         if (param->private_data && param->private_data_len)
3054                 memcpy(sidr_rep_msg->private_data, param->private_data,
3055                        param->private_data_len);
3056 }
3057
3058 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3059                         struct ib_cm_sidr_rep_param *param)
3060 {
3061         struct cm_id_private *cm_id_priv;
3062         struct ib_mad_send_buf *msg;
3063         unsigned long flags;
3064         int ret;
3065
3066         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3067             (param->private_data &&
3068              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3069                 return -EINVAL;
3070
3071         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3072         spin_lock_irqsave(&cm_id_priv->lock, flags);
3073         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3074                 ret = -EINVAL;
3075                 goto error;
3076         }
3077
3078         ret = cm_alloc_msg(cm_id_priv, &msg);
3079         if (ret)
3080                 goto error;
3081
3082         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3083                            param);
3084         ret = ib_post_send_mad(msg, NULL);
3085         if (ret) {
3086                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3087                 cm_free_msg(msg);
3088                 return ret;
3089         }
3090         cm_id->state = IB_CM_IDLE;
3091         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3092
3093         spin_lock_irqsave(&cm.lock, flags);
3094         rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3095         spin_unlock_irqrestore(&cm.lock, flags);
3096         return 0;
3097
3098 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3099         return ret;
3100 }
3101 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3102
3103 static void cm_format_sidr_rep_event(struct cm_work *work)
3104 {
3105         struct cm_sidr_rep_msg *sidr_rep_msg;
3106         struct ib_cm_sidr_rep_event_param *param;
3107
3108         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3109                                 work->mad_recv_wc->recv_buf.mad;
3110         param = &work->cm_event.param.sidr_rep_rcvd;
3111         param->status = sidr_rep_msg->status;
3112         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3113         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3114         param->info = &sidr_rep_msg->info;
3115         param->info_len = sidr_rep_msg->info_length;
3116         work->cm_event.private_data = &sidr_rep_msg->private_data;
3117 }
3118
3119 static int cm_sidr_rep_handler(struct cm_work *work)
3120 {
3121         struct cm_sidr_rep_msg *sidr_rep_msg;
3122         struct cm_id_private *cm_id_priv;
3123
3124         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3125                                 work->mad_recv_wc->recv_buf.mad;
3126         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3127         if (!cm_id_priv)
3128                 return -EINVAL; /* Unmatched reply. */
3129
3130         spin_lock_irq(&cm_id_priv->lock);
3131         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3132                 spin_unlock_irq(&cm_id_priv->lock);
3133                 goto out;
3134         }
3135         cm_id_priv->id.state = IB_CM_IDLE;
3136         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3137         spin_unlock_irq(&cm_id_priv->lock);
3138
3139         cm_format_sidr_rep_event(work);
3140         cm_process_work(cm_id_priv, work);
3141         return 0;
3142 out:
3143         cm_deref_id(cm_id_priv);
3144         return -EINVAL;
3145 }
3146
3147 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3148                                   enum ib_wc_status wc_status)
3149 {
3150         struct cm_id_private *cm_id_priv;
3151         struct ib_cm_event cm_event;
3152         enum ib_cm_state state;
3153         int ret;
3154
3155         memset(&cm_event, 0, sizeof cm_event);
3156         cm_id_priv = msg->context[0];
3157
3158         /* Discard old sends or ones without a response. */
3159         spin_lock_irq(&cm_id_priv->lock);
3160         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3161         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3162                 goto discard;
3163
3164         switch (state) {
3165         case IB_CM_REQ_SENT:
3166         case IB_CM_MRA_REQ_RCVD:
3167                 cm_reset_to_idle(cm_id_priv);
3168                 cm_event.event = IB_CM_REQ_ERROR;
3169                 break;
3170         case IB_CM_REP_SENT:
3171         case IB_CM_MRA_REP_RCVD:
3172                 cm_reset_to_idle(cm_id_priv);
3173                 cm_event.event = IB_CM_REP_ERROR;
3174                 break;
3175         case IB_CM_DREQ_SENT:
3176                 cm_enter_timewait(cm_id_priv);
3177                 cm_event.event = IB_CM_DREQ_ERROR;
3178                 break;
3179         case IB_CM_SIDR_REQ_SENT:
3180                 cm_id_priv->id.state = IB_CM_IDLE;
3181                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3182                 break;
3183         default:
3184                 goto discard;
3185         }
3186         spin_unlock_irq(&cm_id_priv->lock);
3187         cm_event.param.send_status = wc_status;
3188
3189         /* No other events can occur on the cm_id at this point. */
3190         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3191         cm_free_msg(msg);
3192         if (ret)
3193                 ib_destroy_cm_id(&cm_id_priv->id);
3194         return;
3195 discard:
3196         spin_unlock_irq(&cm_id_priv->lock);
3197         cm_free_msg(msg);
3198 }
3199
3200 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3201                             struct ib_mad_send_wc *mad_send_wc)
3202 {
3203         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3204         struct cm_port *port;
3205         u16 attr_index;
3206
3207         port = mad_agent->context;
3208         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3209                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3210
3211         /*
3212          * If the send was in response to a received message (context[0] is not
3213          * set to a cm_id), and is not a REJ, then it is a send that was
3214          * manually retried.
3215          */
3216         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3217                 msg->retries = 1;
3218
3219         atomic_long_add(1 + msg->retries,
3220                         &port->counter_group[CM_XMIT].counter[attr_index]);
3221         if (msg->retries)
3222                 atomic_long_add(msg->retries,
3223                                 &port->counter_group[CM_XMIT_RETRIES].
3224                                 counter[attr_index]);
3225
3226         switch (mad_send_wc->status) {
3227         case IB_WC_SUCCESS:
3228         case IB_WC_WR_FLUSH_ERR:
3229                 cm_free_msg(msg);
3230                 break;
3231         default:
3232                 if (msg->context[0] && msg->context[1])
3233                         cm_process_send_error(msg, mad_send_wc->status);
3234                 else
3235                         cm_free_msg(msg);
3236                 break;
3237         }
3238 }
3239
3240 static void cm_work_handler(struct work_struct *_work)
3241 {
3242         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3243         int ret;
3244
3245         switch (work->cm_event.event) {
3246         case IB_CM_REQ_RECEIVED:
3247                 ret = cm_req_handler(work);
3248                 break;
3249         case IB_CM_MRA_RECEIVED:
3250                 ret = cm_mra_handler(work);
3251                 break;
3252         case IB_CM_REJ_RECEIVED:
3253                 ret = cm_rej_handler(work);
3254                 break;
3255         case IB_CM_REP_RECEIVED:
3256                 ret = cm_rep_handler(work);
3257                 break;
3258         case IB_CM_RTU_RECEIVED:
3259                 ret = cm_rtu_handler(work);
3260                 break;
3261         case IB_CM_USER_ESTABLISHED:
3262                 ret = cm_establish_handler(work);
3263                 break;
3264         case IB_CM_DREQ_RECEIVED:
3265                 ret = cm_dreq_handler(work);
3266                 break;
3267         case IB_CM_DREP_RECEIVED:
3268                 ret = cm_drep_handler(work);
3269                 break;
3270         case IB_CM_SIDR_REQ_RECEIVED:
3271                 ret = cm_sidr_req_handler(work);
3272                 break;
3273         case IB_CM_SIDR_REP_RECEIVED:
3274                 ret = cm_sidr_rep_handler(work);
3275                 break;
3276         case IB_CM_LAP_RECEIVED:
3277                 ret = cm_lap_handler(work);
3278                 break;
3279         case IB_CM_APR_RECEIVED:
3280                 ret = cm_apr_handler(work);
3281                 break;
3282         case IB_CM_TIMEWAIT_EXIT:
3283                 ret = cm_timewait_handler(work);
3284                 break;
3285         default:
3286                 ret = -EINVAL;
3287                 break;
3288         }
3289         if (ret)
3290                 cm_free_work(work);
3291 }
3292
3293 static int cm_establish(struct ib_cm_id *cm_id)
3294 {
3295         struct cm_id_private *cm_id_priv;
3296         struct cm_work *work;
3297         unsigned long flags;
3298         int ret = 0;
3299
3300         work = kmalloc(sizeof *work, GFP_ATOMIC);
3301         if (!work)
3302                 return -ENOMEM;
3303
3304         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3305         spin_lock_irqsave(&cm_id_priv->lock, flags);
3306         switch (cm_id->state)
3307         {
3308         case IB_CM_REP_SENT:
3309         case IB_CM_MRA_REP_RCVD:
3310                 cm_id->state = IB_CM_ESTABLISHED;
3311                 break;
3312         case IB_CM_ESTABLISHED:
3313                 ret = -EISCONN;
3314                 break;
3315         default:
3316                 ret = -EINVAL;
3317                 break;
3318         }
3319         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3320
3321         if (ret) {
3322                 kfree(work);
3323                 goto out;
3324         }
3325
3326         /*
3327          * The CM worker thread may try to destroy the cm_id before it
3328          * can execute this work item.  To prevent potential deadlock,
3329          * we need to find the cm_id once we're in the context of the
3330          * worker thread, rather than holding a reference on it.
3331          */
3332         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3333         work->local_id = cm_id->local_id;
3334         work->remote_id = cm_id->remote_id;
3335         work->mad_recv_wc = NULL;
3336         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3337         queue_delayed_work(cm.wq, &work->work, 0);
3338 out:
3339         return ret;
3340 }
3341
3342 static int cm_migrate(struct ib_cm_id *cm_id)
3343 {
3344         struct cm_id_private *cm_id_priv;
3345         unsigned long flags;
3346         int ret = 0;
3347
3348         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3349         spin_lock_irqsave(&cm_id_priv->lock, flags);
3350         if (cm_id->state == IB_CM_ESTABLISHED &&
3351             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3352              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3353                 cm_id->lap_state = IB_CM_LAP_IDLE;
3354                 cm_id_priv->av = cm_id_priv->alt_av;
3355         } else
3356                 ret = -EINVAL;
3357         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3358
3359         return ret;
3360 }
3361
3362 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3363 {
3364         int ret;
3365
3366         switch (event) {
3367         case IB_EVENT_COMM_EST:
3368                 ret = cm_establish(cm_id);
3369                 break;
3370         case IB_EVENT_PATH_MIG:
3371                 ret = cm_migrate(cm_id);
3372                 break;
3373         default:
3374                 ret = -EINVAL;
3375         }
3376         return ret;
3377 }
3378 EXPORT_SYMBOL(ib_cm_notify);
3379
3380 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3381                             struct ib_mad_recv_wc *mad_recv_wc)
3382 {
3383         struct cm_port *port = mad_agent->context;
3384         struct cm_work *work;
3385         enum ib_cm_event_type event;
3386         u16 attr_id;
3387         int paths = 0;
3388
3389         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3390         case CM_REQ_ATTR_ID:
3391                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3392                                                     alt_local_lid != 0);
3393                 event = IB_CM_REQ_RECEIVED;
3394                 break;
3395         case CM_MRA_ATTR_ID:
3396                 event = IB_CM_MRA_RECEIVED;
3397                 break;
3398         case CM_REJ_ATTR_ID:
3399                 event = IB_CM_REJ_RECEIVED;
3400                 break;
3401         case CM_REP_ATTR_ID:
3402                 event = IB_CM_REP_RECEIVED;
3403                 break;
3404         case CM_RTU_ATTR_ID:
3405                 event = IB_CM_RTU_RECEIVED;
3406                 break;
3407         case CM_DREQ_ATTR_ID:
3408                 event = IB_CM_DREQ_RECEIVED;
3409                 break;
3410         case CM_DREP_ATTR_ID:
3411                 event = IB_CM_DREP_RECEIVED;
3412                 break;
3413         case CM_SIDR_REQ_ATTR_ID:
3414                 event = IB_CM_SIDR_REQ_RECEIVED;
3415                 break;
3416         case CM_SIDR_REP_ATTR_ID:
3417                 event = IB_CM_SIDR_REP_RECEIVED;
3418                 break;
3419         case CM_LAP_ATTR_ID:
3420                 paths = 1;
3421                 event = IB_CM_LAP_RECEIVED;
3422                 break;
3423         case CM_APR_ATTR_ID:
3424                 event = IB_CM_APR_RECEIVED;
3425                 break;
3426         default:
3427                 ib_free_recv_mad(mad_recv_wc);
3428                 return;
3429         }
3430
3431         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3432         atomic_long_inc(&port->counter_group[CM_RECV].
3433                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3434
3435         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3436                        GFP_KERNEL);
3437         if (!work) {
3438                 ib_free_recv_mad(mad_recv_wc);
3439                 return;
3440         }
3441
3442         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3443         work->cm_event.event = event;
3444         work->mad_recv_wc = mad_recv_wc;
3445         work->port = port;
3446         queue_delayed_work(cm.wq, &work->work, 0);
3447 }
3448
3449 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3450                                 struct ib_qp_attr *qp_attr,
3451                                 int *qp_attr_mask)
3452 {
3453         unsigned long flags;
3454         int ret;
3455
3456         spin_lock_irqsave(&cm_id_priv->lock, flags);
3457         switch (cm_id_priv->id.state) {
3458         case IB_CM_REQ_SENT:
3459         case IB_CM_MRA_REQ_RCVD:
3460         case IB_CM_REQ_RCVD:
3461         case IB_CM_MRA_REQ_SENT:
3462         case IB_CM_REP_RCVD:
3463         case IB_CM_MRA_REP_SENT:
3464         case IB_CM_REP_SENT:
3465         case IB_CM_MRA_REP_RCVD:
3466         case IB_CM_ESTABLISHED:
3467                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3468                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3469                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3470                 if (cm_id_priv->responder_resources)
3471                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3472                                                     IB_ACCESS_REMOTE_ATOMIC;
3473                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3474                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3475                 ret = 0;
3476                 break;
3477         default:
3478                 ret = -EINVAL;
3479                 break;
3480         }
3481         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3482         return ret;
3483 }
3484
3485 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3486                                struct ib_qp_attr *qp_attr,
3487                                int *qp_attr_mask)
3488 {
3489         unsigned long flags;
3490         int ret;
3491
3492         spin_lock_irqsave(&cm_id_priv->lock, flags);
3493         switch (cm_id_priv->id.state) {
3494         case IB_CM_REQ_RCVD:
3495         case IB_CM_MRA_REQ_SENT:
3496         case IB_CM_REP_RCVD:
3497         case IB_CM_MRA_REP_SENT:
3498         case IB_CM_REP_SENT:
3499         case IB_CM_MRA_REP_RCVD:
3500         case IB_CM_ESTABLISHED:
3501                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3502                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3503                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3504                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3505                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3506                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3507                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3508                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3509                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3510                                          IB_QP_MIN_RNR_TIMER;
3511                         qp_attr->max_dest_rd_atomic =
3512                                         cm_id_priv->responder_resources;
3513                         qp_attr->min_rnr_timer = 0;
3514                 }
3515                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3516                         *qp_attr_mask |= IB_QP_ALT_PATH;
3517                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3518                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3519                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3520                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3521                 }
3522                 ret = 0;
3523                 break;
3524         default:
3525                 ret = -EINVAL;
3526                 break;
3527         }
3528         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3529         return ret;
3530 }
3531
3532 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3533                                struct ib_qp_attr *qp_attr,
3534                                int *qp_attr_mask)
3535 {
3536         unsigned long flags;
3537         int ret;
3538
3539         spin_lock_irqsave(&cm_id_priv->lock, flags);
3540         switch (cm_id_priv->id.state) {
3541         /* Allow transition to RTS before sending REP */
3542         case IB_CM_REQ_RCVD:
3543         case IB_CM_MRA_REQ_SENT:
3544
3545         case IB_CM_REP_RCVD:
3546         case IB_CM_MRA_REP_SENT:
3547         case IB_CM_REP_SENT:
3548         case IB_CM_MRA_REP_RCVD:
3549         case IB_CM_ESTABLISHED:
3550                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3551                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3552                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3553                         switch (cm_id_priv->qp_type) {
3554                         case IB_QPT_RC:
3555                         case IB_QPT_XRC_INI:
3556                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
3557                                                  IB_QP_MAX_QP_RD_ATOMIC;
3558                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3559                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3560                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
3561                                 /* fall through */
3562                         case IB_QPT_XRC_TGT:
3563                                 *qp_attr_mask |= IB_QP_TIMEOUT;
3564                                 qp_attr->timeout = cm_id_priv->av.timeout;
3565                                 break;
3566                         default:
3567                                 break;
3568                         }
3569                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3570                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3571                                 qp_attr->path_mig_state = IB_MIG_REARM;
3572                         }
3573                 } else {
3574                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3575                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3576                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3577                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3578                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3579                         qp_attr->path_mig_state = IB_MIG_REARM;
3580                 }
3581                 ret = 0;
3582                 break;
3583         default:
3584                 ret = -EINVAL;
3585                 break;
3586         }
3587         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3588         return ret;
3589 }
3590
3591 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3592                        struct ib_qp_attr *qp_attr,
3593                        int *qp_attr_mask)
3594 {
3595         struct cm_id_private *cm_id_priv;
3596         int ret;
3597
3598         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3599         switch (qp_attr->qp_state) {
3600         case IB_QPS_INIT:
3601                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3602                 break;
3603         case IB_QPS_RTR:
3604                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3605                 break;
3606         case IB_QPS_RTS:
3607                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3608                 break;
3609         default:
3610                 ret = -EINVAL;
3611                 break;
3612         }
3613         return ret;
3614 }
3615 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3616
3617 static void cm_get_ack_delay(struct cm_device *cm_dev)
3618 {
3619         struct ib_device_attr attr;
3620
3621         if (ib_query_device(cm_dev->ib_device, &attr))
3622                 cm_dev->ack_delay = 0; /* acks will rely on packet life time */
3623         else
3624                 cm_dev->ack_delay = attr.local_ca_ack_delay;
3625 }
3626
3627 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3628                                char *buf)
3629 {
3630         struct cm_counter_group *group;
3631         struct cm_counter_attribute *cm_attr;
3632
3633         group = container_of(obj, struct cm_counter_group, obj);
3634         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3635
3636         return sprintf(buf, "%ld\n",
3637                        atomic_long_read(&group->counter[cm_attr->index]));
3638 }
3639
3640 static const struct sysfs_ops cm_counter_ops = {
3641         .show = cm_show_counter
3642 };
3643
3644 static struct kobj_type cm_counter_obj_type = {
3645         .sysfs_ops = &cm_counter_ops,
3646         .default_attrs = cm_counter_default_attrs
3647 };
3648
3649 static void cm_release_port_obj(struct kobject *obj)
3650 {
3651         struct cm_port *cm_port;
3652
3653         cm_port = container_of(obj, struct cm_port, port_obj);
3654         kfree(cm_port);
3655 }
3656
3657 static struct kobj_type cm_port_obj_type = {
3658         .release = cm_release_port_obj
3659 };
3660
3661 static char *cm_devnode(struct device *dev, mode_t *mode)
3662 {
3663         if (mode)
3664                 *mode = 0666;
3665         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
3666 }
3667
3668 struct class cm_class = {
3669         .owner   = THIS_MODULE,
3670         .name    = "infiniband_cm",
3671         .devnode = cm_devnode,
3672 };
3673 EXPORT_SYMBOL(cm_class);
3674
3675 static int cm_create_port_fs(struct cm_port *port)
3676 {
3677         int i, ret;
3678
3679         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3680                                    &port->cm_dev->device->kobj,
3681                                    "%d", port->port_num);
3682         if (ret) {
3683                 kfree(port);
3684                 return ret;
3685         }
3686
3687         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3688                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3689                                            &cm_counter_obj_type,
3690                                            &port->port_obj,
3691                                            "%s", counter_group_names[i]);
3692                 if (ret)
3693                         goto error;
3694         }
3695
3696         return 0;
3697
3698 error:
3699         while (i--)
3700                 kobject_put(&port->counter_group[i].obj);
3701         kobject_put(&port->port_obj);
3702         return ret;
3703
3704 }
3705
3706 static void cm_remove_port_fs(struct cm_port *port)
3707 {
3708         int i;
3709
3710         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3711                 kobject_put(&port->counter_group[i].obj);
3712
3713         kobject_put(&port->port_obj);
3714 }
3715
3716 static void cm_add_one(struct ib_device *ib_device)
3717 {
3718         struct cm_device *cm_dev;
3719         struct cm_port *port;
3720         struct ib_mad_reg_req reg_req = {
3721                 .mgmt_class = IB_MGMT_CLASS_CM,
3722                 .mgmt_class_version = IB_CM_CLASS_VERSION
3723         };
3724         struct ib_port_modify port_modify = {
3725                 .set_port_cap_mask = IB_PORT_CM_SUP
3726         };
3727         unsigned long flags;
3728         int ret;
3729         u8 i;
3730
3731         if (rdma_node_get_transport(ib_device->node_type) != RDMA_TRANSPORT_IB)
3732                 return;
3733
3734         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3735                          ib_device->phys_port_cnt, GFP_KERNEL);
3736         if (!cm_dev)
3737                 return;
3738
3739         cm_dev->ib_device = ib_device;
3740         cm_get_ack_delay(cm_dev);
3741
3742         cm_dev->device = device_create(&cm_class, &ib_device->dev,
3743                                        MKDEV(0, 0), NULL,
3744                                        "%s", ib_device->name);
3745         if (IS_ERR(cm_dev->device)) {
3746                 kfree(cm_dev);
3747                 return;
3748         }
3749
3750         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3751         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3752                 port = kzalloc(sizeof *port, GFP_KERNEL);
3753                 if (!port)
3754                         goto error1;
3755
3756                 cm_dev->port[i-1] = port;
3757                 port->cm_dev = cm_dev;
3758                 port->port_num = i;
3759
3760                 ret = cm_create_port_fs(port);
3761                 if (ret)
3762                         goto error1;
3763
3764                 port->mad_agent = ib_register_mad_agent(ib_device, i,
3765                                                         IB_QPT_GSI,
3766                                                         &reg_req,
3767                                                         0,
3768                                                         cm_send_handler,
3769                                                         cm_recv_handler,
3770                                                         port);
3771                 if (IS_ERR(port->mad_agent))
3772                         goto error2;
3773
3774                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
3775                 if (ret)
3776                         goto error3;
3777         }
3778         ib_set_client_data(ib_device, &cm_client, cm_dev);
3779
3780         write_lock_irqsave(&cm.device_lock, flags);
3781         list_add_tail(&cm_dev->list, &cm.device_list);
3782         write_unlock_irqrestore(&cm.device_lock, flags);
3783         return;
3784
3785 error3:
3786         ib_unregister_mad_agent(port->mad_agent);
3787 error2:
3788         cm_remove_port_fs(port);
3789 error1:
3790         port_modify.set_port_cap_mask = 0;
3791         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3792         while (--i) {
3793                 port = cm_dev->port[i-1];
3794                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3795                 ib_unregister_mad_agent(port->mad_agent);
3796                 cm_remove_port_fs(port);
3797         }
3798         device_unregister(cm_dev->device);
3799         kfree(cm_dev);
3800 }
3801
3802 static void cm_remove_one(struct ib_device *ib_device)
3803 {
3804         struct cm_device *cm_dev;
3805         struct cm_port *port;
3806         struct ib_port_modify port_modify = {
3807                 .clr_port_cap_mask = IB_PORT_CM_SUP
3808         };
3809         unsigned long flags;
3810         int i;
3811
3812         cm_dev = ib_get_client_data(ib_device, &cm_client);
3813         if (!cm_dev)
3814                 return;
3815
3816         write_lock_irqsave(&cm.device_lock, flags);
3817         list_del(&cm_dev->list);
3818         write_unlock_irqrestore(&cm.device_lock, flags);
3819
3820         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3821                 port = cm_dev->port[i-1];
3822                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3823                 ib_unregister_mad_agent(port->mad_agent);
3824                 flush_workqueue(cm.wq);
3825                 cm_remove_port_fs(port);
3826         }
3827         device_unregister(cm_dev->device);
3828         kfree(cm_dev);
3829 }
3830
3831 static int __init ib_cm_init(void)
3832 {
3833         int ret;
3834
3835         memset(&cm, 0, sizeof cm);
3836         INIT_LIST_HEAD(&cm.device_list);
3837         rwlock_init(&cm.device_lock);
3838         spin_lock_init(&cm.lock);
3839         cm.listen_service_table = RB_ROOT;
3840         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3841         cm.remote_id_table = RB_ROOT;
3842         cm.remote_qp_table = RB_ROOT;
3843         cm.remote_sidr_table = RB_ROOT;
3844         idr_init(&cm.local_id_table);
3845         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
3846         idr_pre_get(&cm.local_id_table, GFP_KERNEL);
3847         INIT_LIST_HEAD(&cm.timewait_list);
3848
3849         ret = class_register(&cm_class);
3850         if (ret)
3851                 return -ENOMEM;
3852
3853         cm.wq = create_workqueue("ib_cm");
3854         if (!cm.wq) {
3855                 ret = -ENOMEM;
3856                 goto error1;
3857         }
3858
3859         ret = ib_register_client(&cm_client);
3860         if (ret)
3861                 goto error2;
3862
3863         return 0;
3864 error2:
3865         destroy_workqueue(cm.wq);
3866 error1:
3867         class_unregister(&cm_class);
3868         return ret;
3869 }
3870
3871 static void __exit ib_cm_cleanup(void)
3872 {
3873         struct cm_timewait_info *timewait_info, *tmp;
3874
3875         spin_lock_irq(&cm.lock);
3876         list_for_each_entry(timewait_info, &cm.timewait_list, list)
3877                 cancel_delayed_work(&timewait_info->work.work);
3878         spin_unlock_irq(&cm.lock);
3879
3880         ib_unregister_client(&cm_client);
3881         destroy_workqueue(cm.wq);
3882
3883         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
3884                 list_del(&timewait_info->list);
3885                 kfree(timewait_info);
3886         }
3887
3888         class_unregister(&cm_class);
3889         idr_destroy(&cm.local_id_table);
3890 }
3891
3892 module_init(ib_cm_init);
3893 module_exit(ib_cm_cleanup);
3894