git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/core/cm.c

   1 /*
   2  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  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/sysfs.h>
  46 #include <linux/workqueue.h>
  47 #include <linux/kdev_t.h>
  48
  49 #include <rdma/ib_cache.h>
  50 #include <rdma/ib_cm.h>
  51 #include "cm_msgs.h"
  52
  53 MODULE_AUTHOR("Sean Hefty");
  54 MODULE_DESCRIPTION("InfiniBand CM");
  55 MODULE_LICENSE("Dual BSD/GPL");
  56
  57 static void cm_add_one(struct ib_device *device);
  58 static void cm_remove_one(struct ib_device *device);
  59
  60 static struct ib_client cm_client = {
  61         .name   = "cm",
  62         .add    = cm_add_one,
  63         .remove = cm_remove_one
  64 };
  65
  66 static struct ib_cm {
  67         spinlock_t lock;
  68         struct list_head device_list;
  69         rwlock_t device_lock;
  70         struct rb_root listen_service_table;
  71         u64 listen_service_id;
  72         /* struct rb_root peer_service_table; todo: fix peer to peer */
  73         struct rb_root remote_qp_table;
  74         struct rb_root remote_id_table;
  75         struct rb_root remote_sidr_table;
  76         struct idr local_id_table;
  77         __be32 random_id_operand;
  78         struct list_head timewait_list;
  79         struct workqueue_struct *wq;
  80 } cm;
  81
  82 /* Counter indexes ordered by attribute ID */
  83 enum {
  84         CM_REQ_COUNTER,
  85         CM_MRA_COUNTER,
  86         CM_REJ_COUNTER,
  87         CM_REP_COUNTER,
  88         CM_RTU_COUNTER,
  89         CM_DREQ_COUNTER,
  90         CM_DREP_COUNTER,
  91         CM_SIDR_REQ_COUNTER,
  92         CM_SIDR_REP_COUNTER,
  93         CM_LAP_COUNTER,
  94         CM_APR_COUNTER,
  95         CM_ATTR_COUNT,
  96         CM_ATTR_ID_OFFSET = 0x0010,
  97 };
  98
  99 enum {
 100         CM_XMIT,
 101         CM_XMIT_RETRIES,
 102         CM_RECV,
 103         CM_RECV_DUPLICATES,
 104         CM_COUNTER_GROUPS
 105 };
 106
 107 static char const counter_group_names[CM_COUNTER_GROUPS]
 108                                      [sizeof("cm_rx_duplicates")] = {
 109         "cm_tx_msgs", "cm_tx_retries",
 110         "cm_rx_msgs", "cm_rx_duplicates"
 111 };
 112
 113 struct cm_counter_group {
 114         struct kobject obj;
 115         atomic_long_t counter[CM_ATTR_COUNT];
 116 };
 117
 118 struct cm_counter_attribute {
 119         struct attribute attr;
 120         int index;
 121 };
 122
 123 #define CM_COUNTER_ATTR(_name, _index) \
 124 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 125         .attr = { .name = __stringify(_name), .mode = 0444, .owner = THIS_MODULE }, \
 126         .index = _index \
 127 }
 128
 129 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 130 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 131 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 132 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 133 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 134 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 135 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 136 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 137 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 138 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 139 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 140
 141 static struct attribute *cm_counter_default_attrs[] = {
 142         &cm_req_counter_attr.attr,
 143         &cm_mra_counter_attr.attr,
 144         &cm_rej_counter_attr.attr,
 145         &cm_rep_counter_attr.attr,
 146         &cm_rtu_counter_attr.attr,
 147         &cm_dreq_counter_attr.attr,
 148         &cm_drep_counter_attr.attr,
 149         &cm_sidr_req_counter_attr.attr,
 150         &cm_sidr_rep_counter_attr.attr,
 151         &cm_lap_counter_attr.attr,
 152         &cm_apr_counter_attr.attr,
 153         NULL
 154 };
 155
 156 struct cm_port {
 157         struct cm_device *cm_dev;
 158         struct ib_mad_agent *mad_agent;
 159         struct kobject port_obj;
 160         u8 port_num;
 161         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 162 };
 163
 164 struct cm_device {
 165         struct list_head list;
 166         struct ib_device *ib_device;
 167         struct device *device;
 168         u8 ack_delay;
 169         struct cm_port *port[0];
 170 };
 171
 172 struct cm_av {
 173         struct cm_port *port;
 174         union ib_gid dgid;
 175         struct ib_ah_attr ah_attr;
 176         u16 pkey_index;
 177         u8 timeout;
 178 };
 179
 180 struct cm_work {
 181         struct delayed_work work;
 182         struct list_head list;
 183         struct cm_port *port;
 184         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 185         __be32 local_id;                        /* Established / timewait */
 186         __be32 remote_id;
 187         struct ib_cm_event cm_event;
 188         struct ib_sa_path_rec path[0];
 189 };
 190
 191 struct cm_timewait_info {
 192         struct cm_work work;                    /* Must be first. */
 193         struct list_head list;
 194         struct rb_node remote_qp_node;
 195         struct rb_node remote_id_node;
 196         __be64 remote_ca_guid;
 197         __be32 remote_qpn;
 198         u8 inserted_remote_qp;
 199         u8 inserted_remote_id;
 200 };
 201
 202 struct cm_id_private {
 203         struct ib_cm_id id;
 204
 205         struct rb_node service_node;
 206         struct rb_node sidr_id_node;
 207         spinlock_t lock;        /* Do not acquire inside cm.lock */
 208         struct completion comp;
 209         atomic_t refcount;
 210
 211         struct ib_mad_send_buf *msg;
 212         struct cm_timewait_info *timewait_info;
 213         /* todo: use alternate port on send failure */
 214         struct cm_av av;
 215         struct cm_av alt_av;
 216         struct ib_cm_compare_data *compare_data;
 217
 218         void *private_data;
 219         __be64 tid;
 220         __be32 local_qpn;
 221         __be32 remote_qpn;
 222         enum ib_qp_type qp_type;
 223         __be32 sq_psn;
 224         __be32 rq_psn;
 225         int timeout_ms;
 226         enum ib_mtu path_mtu;
 227         __be16 pkey;
 228         u8 private_data_len;
 229         u8 max_cm_retries;
 230         u8 peer_to_peer;
 231         u8 responder_resources;
 232         u8 initiator_depth;
 233         u8 retry_count;
 234         u8 rnr_retry_count;
 235         u8 service_timeout;
 236         u8 target_ack_delay;
 237
 238         struct list_head work_list;
 239         atomic_t work_count;
 240 };
 241
 242 static void cm_work_handler(struct work_struct *work);
 243
 244 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 245 {
 246         if (atomic_dec_and_test(&cm_id_priv->refcount))
 247                 complete(&cm_id_priv->comp);
 248 }
 249
 250 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 251                         struct ib_mad_send_buf **msg)
 252 {
 253         struct ib_mad_agent *mad_agent;
 254         struct ib_mad_send_buf *m;
 255         struct ib_ah *ah;
 256
 257         mad_agent = cm_id_priv->av.port->mad_agent;
 258         ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
 259         if (IS_ERR(ah))
 260                 return PTR_ERR(ah);
 261
 262         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 263                                cm_id_priv->av.pkey_index,
 264                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 265                                GFP_ATOMIC);
 266         if (IS_ERR(m)) {
 267                 ib_destroy_ah(ah);
 268                 return PTR_ERR(m);
 269         }
 270
 271         /* Timeout set by caller if response is expected. */
 272         m->ah = ah;
 273         m->retries = cm_id_priv->max_cm_retries;
 274
 275         atomic_inc(&cm_id_priv->refcount);
 276         m->context[0] = cm_id_priv;
 277         *msg = m;
 278         return 0;
 279 }
 280
 281 static int cm_alloc_response_msg(struct cm_port *port,
 282                                  struct ib_mad_recv_wc *mad_recv_wc,
 283                                  struct ib_mad_send_buf **msg)
 284 {
 285         struct ib_mad_send_buf *m;
 286         struct ib_ah *ah;
 287
 288         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 289                                   mad_recv_wc->recv_buf.grh, port->port_num);
 290         if (IS_ERR(ah))
 291                 return PTR_ERR(ah);
 292
 293         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 294                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 295                                GFP_ATOMIC);
 296         if (IS_ERR(m)) {
 297                 ib_destroy_ah(ah);
 298                 return PTR_ERR(m);
 299         }
 300         m->ah = ah;
 301         *msg = m;
 302         return 0;
 303 }
 304
 305 static void cm_free_msg(struct ib_mad_send_buf *msg)
 306 {
 307         ib_destroy_ah(msg->ah);
 308         if (msg->context[0])
 309                 cm_deref_id(msg->context[0]);
 310         ib_free_send_mad(msg);
 311 }
 312
 313 static void * cm_copy_private_data(const void *private_data,
 314                                    u8 private_data_len)
 315 {
 316         void *data;
 317
 318         if (!private_data || !private_data_len)
 319                 return NULL;
 320
 321         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 322         if (!data)
 323                 return ERR_PTR(-ENOMEM);
 324
 325         return data;
 326 }
 327
 328 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 329                                  void *private_data, u8 private_data_len)
 330 {
 331         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 332                 kfree(cm_id_priv->private_data);
 333
 334         cm_id_priv->private_data = private_data;
 335         cm_id_priv->private_data_len = private_data_len;
 336 }
 337
 338 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 339                                     struct ib_grh *grh, struct cm_av *av)
 340 {
 341         av->port = port;
 342         av->pkey_index = wc->pkey_index;
 343         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 344                            grh, &av->ah_attr);
 345 }
 346
 347 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
 348 {
 349         struct cm_device *cm_dev;
 350         struct cm_port *port = NULL;
 351         unsigned long flags;
 352         int ret;
 353         u8 p;
 354
 355         read_lock_irqsave(&cm.device_lock, flags);
 356         list_for_each_entry(cm_dev, &cm.device_list, list) {
 357                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 358                                         &p, NULL)) {
 359                         port = cm_dev->port[p-1];
 360                         break;
 361                 }
 362         }
 363         read_unlock_irqrestore(&cm.device_lock, flags);
 364
 365         if (!port)
 366                 return -EINVAL;
 367
 368         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 369                                   be16_to_cpu(path->pkey), &av->pkey_index);
 370         if (ret)
 371                 return ret;
 372
 373         av->port = port;
 374         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 375                              &av->ah_attr);
 376         av->timeout = path->packet_life_time + 1;
 377         return 0;
 378 }
 379
 380 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 381 {
 382         unsigned long flags;
 383         int ret, id;
 384         static int next_id;
 385
 386         do {
 387                 spin_lock_irqsave(&cm.lock, flags);
 388                 ret = idr_get_new_above(&cm.local_id_table, cm_id_priv,
 389                                         next_id, &id);
 390                 if (!ret)
 391                         next_id = ((unsigned) id + 1) & MAX_ID_MASK;
 392                 spin_unlock_irqrestore(&cm.lock, flags);
 393         } while( (ret == -EAGAIN) && idr_pre_get(&cm.local_id_table, GFP_KERNEL) );
 394
 395         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 396         return ret;
 397 }
 398
 399 static void cm_free_id(__be32 local_id)
 400 {
 401         spin_lock_irq(&cm.lock);
 402         idr_remove(&cm.local_id_table,
 403                    (__force int) (local_id ^ cm.random_id_operand));
 404         spin_unlock_irq(&cm.lock);
 405 }
 406
 407 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 408 {
 409         struct cm_id_private *cm_id_priv;
 410
 411         cm_id_priv = idr_find(&cm.local_id_table,
 412                               (__force int) (local_id ^ cm.random_id_operand));
 413         if (cm_id_priv) {
 414                 if (cm_id_priv->id.remote_id == remote_id)
 415                         atomic_inc(&cm_id_priv->refcount);
 416                 else
 417                         cm_id_priv = NULL;
 418         }
 419
 420         return cm_id_priv;
 421 }
 422
 423 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 424 {
 425         struct cm_id_private *cm_id_priv;
 426
 427         spin_lock_irq(&cm.lock);
 428         cm_id_priv = cm_get_id(local_id, remote_id);
 429         spin_unlock_irq(&cm.lock);
 430
 431         return cm_id_priv;
 432 }
 433
 434 static void cm_mask_copy(u8 *dst, u8 *src, u8 *mask)
 435 {
 436         int i;
 437
 438         for (i = 0; i < IB_CM_COMPARE_SIZE / sizeof(unsigned long); i++)
 439                 ((unsigned long *) dst)[i] = ((unsigned long *) src)[i] &
 440                                              ((unsigned long *) mask)[i];
 441 }
 442
 443 static int cm_compare_data(struct ib_cm_compare_data *src_data,
 444                            struct ib_cm_compare_data *dst_data)
 445 {
 446         u8 src[IB_CM_COMPARE_SIZE];
 447         u8 dst[IB_CM_COMPARE_SIZE];
 448
 449         if (!src_data || !dst_data)
 450                 return 0;
 451
 452         cm_mask_copy(src, src_data->data, dst_data->mask);
 453         cm_mask_copy(dst, dst_data->data, src_data->mask);
 454         return memcmp(src, dst, IB_CM_COMPARE_SIZE);
 455 }
 456
 457 static int cm_compare_private_data(u8 *private_data,
 458                                    struct ib_cm_compare_data *dst_data)
 459 {
 460         u8 src[IB_CM_COMPARE_SIZE];
 461
 462         if (!dst_data)
 463                 return 0;
 464
 465         cm_mask_copy(src, private_data, dst_data->mask);
 466         return memcmp(src, dst_data->data, IB_CM_COMPARE_SIZE);
 467 }
 468
 469 /*
 470  * Trivial helpers to strip endian annotation and compare; the
 471  * endianness doesn't actually matter since we just need a stable
 472  * order for the RB tree.
 473  */
 474 static int be32_lt(__be32 a, __be32 b)
 475 {
 476         return (__force u32) a < (__force u32) b;
 477 }
 478
 479 static int be32_gt(__be32 a, __be32 b)
 480 {
 481         return (__force u32) a > (__force u32) b;
 482 }
 483
 484 static int be64_lt(__be64 a, __be64 b)
 485 {
 486         return (__force u64) a < (__force u64) b;
 487 }
 488
 489 static int be64_gt(__be64 a, __be64 b)
 490 {
 491         return (__force u64) a > (__force u64) b;
 492 }
 493
 494 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 495 {
 496         struct rb_node **link = &cm.listen_service_table.rb_node;
 497         struct rb_node *parent = NULL;
 498         struct cm_id_private *cur_cm_id_priv;
 499         __be64 service_id = cm_id_priv->id.service_id;
 500         __be64 service_mask = cm_id_priv->id.service_mask;
 501         int data_cmp;
 502
 503         while (*link) {
 504                 parent = *link;
 505                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 506                                           service_node);
 507                 data_cmp = cm_compare_data(cm_id_priv->compare_data,
 508                                            cur_cm_id_priv->compare_data);
 509                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 510                     (service_mask & cur_cm_id_priv->id.service_id) &&
 511                     (cm_id_priv->id.device == cur_cm_id_priv->id.device) &&
 512                     !data_cmp)
 513                         return cur_cm_id_priv;
 514
 515                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 516                         link = &(*link)->rb_left;
 517                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 518                         link = &(*link)->rb_right;
 519                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 520                         link = &(*link)->rb_left;
 521                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 522                         link = &(*link)->rb_right;
 523                 else if (data_cmp < 0)
 524                         link = &(*link)->rb_left;
 525                 else
 526                         link = &(*link)->rb_right;
 527         }
 528         rb_link_node(&cm_id_priv->service_node, parent, link);
 529         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 530         return NULL;
 531 }
 532
 533 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 534                                              __be64 service_id,
 535                                              u8 *private_data)
 536 {
 537         struct rb_node *node = cm.listen_service_table.rb_node;
 538         struct cm_id_private *cm_id_priv;
 539         int data_cmp;
 540
 541         while (node) {
 542                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 543                 data_cmp = cm_compare_private_data(private_data,
 544                                                    cm_id_priv->compare_data);
 545                 if ((cm_id_priv->id.service_mask & service_id) ==
 546                      cm_id_priv->id.service_id &&
 547                     (cm_id_priv->id.device == device) && !data_cmp)
 548                         return cm_id_priv;
 549
 550                 if (device < cm_id_priv->id.device)
 551                         node = node->rb_left;
 552                 else if (device > cm_id_priv->id.device)
 553                         node = node->rb_right;
 554                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 555                         node = node->rb_left;
 556                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 557                         node = node->rb_right;
 558                 else if (data_cmp < 0)
 559                         node = node->rb_left;
 560                 else
 561                         node = node->rb_right;
 562         }
 563         return NULL;
 564 }
 565
 566 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 567                                                      *timewait_info)
 568 {
 569         struct rb_node **link = &cm.remote_id_table.rb_node;
 570         struct rb_node *parent = NULL;
 571         struct cm_timewait_info *cur_timewait_info;
 572         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 573         __be32 remote_id = timewait_info->work.remote_id;
 574
 575         while (*link) {
 576                 parent = *link;
 577                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 578                                              remote_id_node);
 579                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 580                         link = &(*link)->rb_left;
 581                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 582                         link = &(*link)->rb_right;
 583                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 584                         link = &(*link)->rb_left;
 585                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 586                         link = &(*link)->rb_right;
 587                 else
 588                         return cur_timewait_info;
 589         }
 590         timewait_info->inserted_remote_id = 1;
 591         rb_link_node(&timewait_info->remote_id_node, parent, link);
 592         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 593         return NULL;
 594 }
 595
 596 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 597                                                    __be32 remote_id)
 598 {
 599         struct rb_node *node = cm.remote_id_table.rb_node;
 600         struct cm_timewait_info *timewait_info;
 601
 602         while (node) {
 603                 timewait_info = rb_entry(node, struct cm_timewait_info,
 604                                          remote_id_node);
 605                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 606                         node = node->rb_left;
 607                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 608                         node = node->rb_right;
 609                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 610                         node = node->rb_left;
 611                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 612                         node = node->rb_right;
 613                 else
 614                         return timewait_info;
 615         }
 616         return NULL;
 617 }
 618
 619 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 620                                                       *timewait_info)
 621 {
 622         struct rb_node **link = &cm.remote_qp_table.rb_node;
 623         struct rb_node *parent = NULL;
 624         struct cm_timewait_info *cur_timewait_info;
 625         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 626         __be32 remote_qpn = timewait_info->remote_qpn;
 627
 628         while (*link) {
 629                 parent = *link;
 630                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 631                                              remote_qp_node);
 632                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 633                         link = &(*link)->rb_left;
 634                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 635                         link = &(*link)->rb_right;
 636                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 637                         link = &(*link)->rb_left;
 638                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 639                         link = &(*link)->rb_right;
 640                 else
 641                         return cur_timewait_info;
 642         }
 643         timewait_info->inserted_remote_qp = 1;
 644         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 645         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 646         return NULL;
 647 }
 648
 649 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 650                                                     *cm_id_priv)
 651 {
 652         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 653         struct rb_node *parent = NULL;
 654         struct cm_id_private *cur_cm_id_priv;
 655         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 656         __be32 remote_id = cm_id_priv->id.remote_id;
 657
 658         while (*link) {
 659                 parent = *link;
 660                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 661                                           sidr_id_node);
 662                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 663                         link = &(*link)->rb_left;
 664                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 665                         link = &(*link)->rb_right;
 666                 else {
 667                         int cmp;
 668                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 669                                      sizeof *port_gid);
 670                         if (cmp < 0)
 671                                 link = &(*link)->rb_left;
 672                         else if (cmp > 0)
 673                                 link = &(*link)->rb_right;
 674                         else
 675                                 return cur_cm_id_priv;
 676                 }
 677         }
 678         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 679         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 680         return NULL;
 681 }
 682
 683 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 684                                enum ib_cm_sidr_status status)
 685 {
 686         struct ib_cm_sidr_rep_param param;
 687
 688         memset(&param, 0, sizeof param);
 689         param.status = status;
 690         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 691 }
 692
 693 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 694                                  ib_cm_handler cm_handler,
 695                                  void *context)
 696 {
 697         struct cm_id_private *cm_id_priv;
 698         int ret;
 699
 700         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 701         if (!cm_id_priv)
 702                 return ERR_PTR(-ENOMEM);
 703
 704         cm_id_priv->id.state = IB_CM_IDLE;
 705         cm_id_priv->id.device = device;
 706         cm_id_priv->id.cm_handler = cm_handler;
 707         cm_id_priv->id.context = context;
 708         cm_id_priv->id.remote_cm_qpn = 1;
 709         ret = cm_alloc_id(cm_id_priv);
 710         if (ret)
 711                 goto error;
 712
 713         spin_lock_init(&cm_id_priv->lock);
 714         init_completion(&cm_id_priv->comp);
 715         INIT_LIST_HEAD(&cm_id_priv->work_list);
 716         atomic_set(&cm_id_priv->work_count, -1);
 717         atomic_set(&cm_id_priv->refcount, 1);
 718         return &cm_id_priv->id;
 719
 720 error:
 721         kfree(cm_id_priv);
 722         return ERR_PTR(-ENOMEM);
 723 }
 724 EXPORT_SYMBOL(ib_create_cm_id);
 725
 726 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 727 {
 728         struct cm_work *work;
 729
 730         if (list_empty(&cm_id_priv->work_list))
 731                 return NULL;
 732
 733         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 734         list_del(&work->list);
 735         return work;
 736 }
 737
 738 static void cm_free_work(struct cm_work *work)
 739 {
 740         if (work->mad_recv_wc)
 741                 ib_free_recv_mad(work->mad_recv_wc);
 742         kfree(work);
 743 }
 744
 745 static inline int cm_convert_to_ms(int iba_time)
 746 {
 747         /* approximate conversion to ms from 4.096us x 2^iba_time */
 748         return 1 << max(iba_time - 8, 0);
 749 }
 750
 751 /*
 752  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 753  * Because of how ack_timeout is stored, adding one doubles the timeout.
 754  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 755  * increment it (round up) only if the other is within 50%.
 756  */
 757 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 758 {
 759         int ack_timeout = packet_life_time + 1;
 760
 761         if (ack_timeout >= ca_ack_delay)
 762                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 763         else
 764                 ack_timeout = ca_ack_delay +
 765                               (ack_timeout >= (ca_ack_delay - 1));
 766
 767         return min(31, ack_timeout);
 768 }
 769
 770 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 771 {
 772         if (timewait_info->inserted_remote_id) {
 773                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 774                 timewait_info->inserted_remote_id = 0;
 775         }
 776
 777         if (timewait_info->inserted_remote_qp) {
 778                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 779                 timewait_info->inserted_remote_qp = 0;
 780         }
 781 }
 782
 783 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 784 {
 785         struct cm_timewait_info *timewait_info;
 786
 787         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 788         if (!timewait_info)
 789                 return ERR_PTR(-ENOMEM);
 790
 791         timewait_info->work.local_id = local_id;
 792         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 793         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 794         return timewait_info;
 795 }
 796
 797 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 798 {
 799         int wait_time;
 800         unsigned long flags;
 801
 802         spin_lock_irqsave(&cm.lock, flags);
 803         cm_cleanup_timewait(cm_id_priv->timewait_info);
 804         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 805         spin_unlock_irqrestore(&cm.lock, flags);
 806
 807         /*
 808          * The cm_id could be destroyed by the user before we exit timewait.
 809          * To protect against this, we search for the cm_id after exiting
 810          * timewait before notifying the user that we've exited timewait.
 811          */
 812         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 813         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 814         queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 815                            msecs_to_jiffies(wait_time));
 816         cm_id_priv->timewait_info = NULL;
 817 }
 818
 819 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 820 {
 821         unsigned long flags;
 822
 823         cm_id_priv->id.state = IB_CM_IDLE;
 824         if (cm_id_priv->timewait_info) {
 825                 spin_lock_irqsave(&cm.lock, flags);
 826                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 827                 spin_unlock_irqrestore(&cm.lock, flags);
 828                 kfree(cm_id_priv->timewait_info);
 829                 cm_id_priv->timewait_info = NULL;
 830         }
 831 }
 832
 833 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 834 {
 835         struct cm_id_private *cm_id_priv;
 836         struct cm_work *work;
 837
 838         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 839 retest:
 840         spin_lock_irq(&cm_id_priv->lock);
 841         switch (cm_id->state) {
 842         case IB_CM_LISTEN:
 843                 cm_id->state = IB_CM_IDLE;
 844                 spin_unlock_irq(&cm_id_priv->lock);
 845                 spin_lock_irq(&cm.lock);
 846                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 847                 spin_unlock_irq(&cm.lock);
 848                 break;
 849         case IB_CM_SIDR_REQ_SENT:
 850                 cm_id->state = IB_CM_IDLE;
 851                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 852                 spin_unlock_irq(&cm_id_priv->lock);
 853                 break;
 854         case IB_CM_SIDR_REQ_RCVD:
 855                 spin_unlock_irq(&cm_id_priv->lock);
 856                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 857                 break;
 858         case IB_CM_REQ_SENT:
 859                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 860                 spin_unlock_irq(&cm_id_priv->lock);
 861                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 862                                &cm_id_priv->id.device->node_guid,
 863                                sizeof cm_id_priv->id.device->node_guid,
 864                                NULL, 0);
 865                 break;
 866         case IB_CM_REQ_RCVD:
 867                 if (err == -ENOMEM) {
 868                         /* Do not reject to allow future retries. */
 869                         cm_reset_to_idle(cm_id_priv);
 870                         spin_unlock_irq(&cm_id_priv->lock);
 871                 } else {
 872                         spin_unlock_irq(&cm_id_priv->lock);
 873                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 874                                        NULL, 0, NULL, 0);
 875                 }
 876                 break;
 877         case IB_CM_MRA_REQ_RCVD:
 878         case IB_CM_REP_SENT:
 879         case IB_CM_MRA_REP_RCVD:
 880                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 881                 /* Fall through */
 882         case IB_CM_MRA_REQ_SENT:
 883         case IB_CM_REP_RCVD:
 884         case IB_CM_MRA_REP_SENT:
 885                 spin_unlock_irq(&cm_id_priv->lock);
 886                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 887                                NULL, 0, NULL, 0);
 888                 break;
 889         case IB_CM_ESTABLISHED:
 890                 spin_unlock_irq(&cm_id_priv->lock);
 891                 ib_send_cm_dreq(cm_id, NULL, 0);
 892                 goto retest;
 893         case IB_CM_DREQ_SENT:
 894                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 895                 cm_enter_timewait(cm_id_priv);
 896                 spin_unlock_irq(&cm_id_priv->lock);
 897                 break;
 898         case IB_CM_DREQ_RCVD:
 899                 spin_unlock_irq(&cm_id_priv->lock);
 900                 ib_send_cm_drep(cm_id, NULL, 0);
 901                 break;
 902         default:
 903                 spin_unlock_irq(&cm_id_priv->lock);
 904                 break;
 905         }
 906
 907         cm_free_id(cm_id->local_id);
 908         cm_deref_id(cm_id_priv);
 909         wait_for_completion(&cm_id_priv->comp);
 910         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
 911                 cm_free_work(work);
 912         kfree(cm_id_priv->compare_data);
 913         kfree(cm_id_priv->private_data);
 914         kfree(cm_id_priv);
 915 }
 916
 917 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
 918 {
 919         cm_destroy_id(cm_id, 0);
 920 }
 921 EXPORT_SYMBOL(ib_destroy_cm_id);
 922
 923 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask,
 924                  struct ib_cm_compare_data *compare_data)
 925 {
 926         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
 927         unsigned long flags;
 928         int ret = 0;
 929
 930         service_mask = service_mask ? service_mask :
 931                        __constant_cpu_to_be64(~0ULL);
 932         service_id &= service_mask;
 933         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
 934             (service_id != IB_CM_ASSIGN_SERVICE_ID))
 935                 return -EINVAL;
 936
 937         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 938         if (cm_id->state != IB_CM_IDLE)
 939                 return -EINVAL;
 940
 941         if (compare_data) {
 942                 cm_id_priv->compare_data = kzalloc(sizeof *compare_data,
 943                                                    GFP_KERNEL);
 944                 if (!cm_id_priv->compare_data)
 945                         return -ENOMEM;
 946                 cm_mask_copy(cm_id_priv->compare_data->data,
 947                              compare_data->data, compare_data->mask);
 948                 memcpy(cm_id_priv->compare_data->mask, compare_data->mask,
 949                        IB_CM_COMPARE_SIZE);
 950         }
 951
 952         cm_id->state = IB_CM_LISTEN;
 953
 954         spin_lock_irqsave(&cm.lock, flags);
 955         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
 956                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
 957                 cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
 958         } else {
 959                 cm_id->service_id = service_id;
 960                 cm_id->service_mask = service_mask;
 961         }
 962         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
 963         spin_unlock_irqrestore(&cm.lock, flags);
 964
 965         if (cur_cm_id_priv) {
 966                 cm_id->state = IB_CM_IDLE;
 967                 kfree(cm_id_priv->compare_data);
 968                 cm_id_priv->compare_data = NULL;
 969                 ret = -EBUSY;
 970         }
 971         return ret;
 972 }
 973 EXPORT_SYMBOL(ib_cm_listen);
 974
 975 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
 976                           enum cm_msg_sequence msg_seq)
 977 {
 978         u64 hi_tid, low_tid;
 979
 980         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
 981         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
 982                           (msg_seq << 30));
 983         return cpu_to_be64(hi_tid | low_tid);
 984 }
 985
 986 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
 987                               __be16 attr_id, __be64 tid)
 988 {
 989         hdr->base_version  = IB_MGMT_BASE_VERSION;
 990         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
 991         hdr->class_version = IB_CM_CLASS_VERSION;
 992         hdr->method        = IB_MGMT_METHOD_SEND;
 993         hdr->attr_id       = attr_id;
 994         hdr->tid           = tid;
 995 }
 996
 997 static void cm_format_req(struct cm_req_msg *req_msg,
 998                           struct cm_id_private *cm_id_priv,
 999                           struct ib_cm_req_param *param)
1000 {
1001         struct ib_sa_path_rec *pri_path = param->primary_path;
1002         struct ib_sa_path_rec *alt_path = param->alternate_path;
1003
1004         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1005                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1006
1007         req_msg->local_comm_id = cm_id_priv->id.local_id;
1008         req_msg->service_id = param->service_id;
1009         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1010         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1011         cm_req_set_resp_res(req_msg, param->responder_resources);
1012         cm_req_set_init_depth(req_msg, param->initiator_depth);
1013         cm_req_set_remote_resp_timeout(req_msg,
1014                                        param->remote_cm_response_timeout);
1015         cm_req_set_qp_type(req_msg, param->qp_type);
1016         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1017         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1018         cm_req_set_local_resp_timeout(req_msg,
1019                                       param->local_cm_response_timeout);
1020         cm_req_set_retry_count(req_msg, param->retry_count);
1021         req_msg->pkey = param->primary_path->pkey;
1022         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1023         cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1024         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1025         cm_req_set_srq(req_msg, param->srq);
1026
1027         if (pri_path->hop_limit <= 1) {
1028                 req_msg->primary_local_lid = pri_path->slid;
1029                 req_msg->primary_remote_lid = pri_path->dlid;
1030         } else {
1031                 /* Work-around until there's a way to obtain remote LID info */
1032                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1033                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1034         }
1035         req_msg->primary_local_gid = pri_path->sgid;
1036         req_msg->primary_remote_gid = pri_path->dgid;
1037         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1038         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1039         req_msg->primary_traffic_class = pri_path->traffic_class;
1040         req_msg->primary_hop_limit = pri_path->hop_limit;
1041         cm_req_set_primary_sl(req_msg, pri_path->sl);
1042         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1043         cm_req_set_primary_local_ack_timeout(req_msg,
1044                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1045                                pri_path->packet_life_time));
1046
1047         if (alt_path) {
1048                 if (alt_path->hop_limit <= 1) {
1049                         req_msg->alt_local_lid = alt_path->slid;
1050                         req_msg->alt_remote_lid = alt_path->dlid;
1051                 } else {
1052                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1053                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1054                 }
1055                 req_msg->alt_local_gid = alt_path->sgid;
1056                 req_msg->alt_remote_gid = alt_path->dgid;
1057                 cm_req_set_alt_flow_label(req_msg,
1058                                           alt_path->flow_label);
1059                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1060                 req_msg->alt_traffic_class = alt_path->traffic_class;
1061                 req_msg->alt_hop_limit = alt_path->hop_limit;
1062                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1063                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1064                 cm_req_set_alt_local_ack_timeout(req_msg,
1065                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1066                                        alt_path->packet_life_time));
1067         }
1068
1069         if (param->private_data && param->private_data_len)
1070                 memcpy(req_msg->private_data, param->private_data,
1071                        param->private_data_len);
1072 }
1073
1074 static int cm_validate_req_param(struct ib_cm_req_param *param)
1075 {
1076         /* peer-to-peer not supported */
1077         if (param->peer_to_peer)
1078                 return -EINVAL;
1079
1080         if (!param->primary_path)
1081                 return -EINVAL;
1082
1083         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC)
1084                 return -EINVAL;
1085
1086         if (param->private_data &&
1087             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1088                 return -EINVAL;
1089
1090         if (param->alternate_path &&
1091             (param->alternate_path->pkey != param->primary_path->pkey ||
1092              param->alternate_path->mtu != param->primary_path->mtu))
1093                 return -EINVAL;
1094
1095         return 0;
1096 }
1097
1098 int ib_send_cm_req(struct ib_cm_id *cm_id,
1099                    struct ib_cm_req_param *param)
1100 {
1101         struct cm_id_private *cm_id_priv;
1102         struct cm_req_msg *req_msg;
1103         unsigned long flags;
1104         int ret;
1105
1106         ret = cm_validate_req_param(param);
1107         if (ret)
1108                 return ret;
1109
1110         /* Verify that we're not in timewait. */
1111         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1112         spin_lock_irqsave(&cm_id_priv->lock, flags);
1113         if (cm_id->state != IB_CM_IDLE) {
1114                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1115                 ret = -EINVAL;
1116                 goto out;
1117         }
1118         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1119
1120         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1121                                                             id.local_id);
1122         if (IS_ERR(cm_id_priv->timewait_info)) {
1123                 ret = PTR_ERR(cm_id_priv->timewait_info);
1124                 goto out;
1125         }
1126
1127         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av);
1128         if (ret)
1129                 goto error1;
1130         if (param->alternate_path) {
1131                 ret = cm_init_av_by_path(param->alternate_path,
1132                                          &cm_id_priv->alt_av);
1133                 if (ret)
1134                         goto error1;
1135         }
1136         cm_id->service_id = param->service_id;
1137         cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
1138         cm_id_priv->timeout_ms = cm_convert_to_ms(
1139                                     param->primary_path->packet_life_time) * 2 +
1140                                  cm_convert_to_ms(
1141                                     param->remote_cm_response_timeout);
1142         cm_id_priv->max_cm_retries = param->max_cm_retries;
1143         cm_id_priv->initiator_depth = param->initiator_depth;
1144         cm_id_priv->responder_resources = param->responder_resources;
1145         cm_id_priv->retry_count = param->retry_count;
1146         cm_id_priv->path_mtu = param->primary_path->mtu;
1147         cm_id_priv->pkey = param->primary_path->pkey;
1148         cm_id_priv->qp_type = param->qp_type;
1149
1150         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1151         if (ret)
1152                 goto error1;
1153
1154         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1155         cm_format_req(req_msg, cm_id_priv, param);
1156         cm_id_priv->tid = req_msg->hdr.tid;
1157         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1158         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1159
1160         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1161         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1162
1163         spin_lock_irqsave(&cm_id_priv->lock, flags);
1164         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1165         if (ret) {
1166                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1167                 goto error2;
1168         }
1169         BUG_ON(cm_id->state != IB_CM_IDLE);
1170         cm_id->state = IB_CM_REQ_SENT;
1171         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1172         return 0;
1173
1174 error2: cm_free_msg(cm_id_priv->msg);
1175 error1: kfree(cm_id_priv->timewait_info);
1176 out:    return ret;
1177 }
1178 EXPORT_SYMBOL(ib_send_cm_req);
1179
1180 static int cm_issue_rej(struct cm_port *port,
1181                         struct ib_mad_recv_wc *mad_recv_wc,
1182                         enum ib_cm_rej_reason reason,
1183                         enum cm_msg_response msg_rejected,
1184                         void *ari, u8 ari_length)
1185 {
1186         struct ib_mad_send_buf *msg = NULL;
1187         struct cm_rej_msg *rej_msg, *rcv_msg;
1188         int ret;
1189
1190         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1191         if (ret)
1192                 return ret;
1193
1194         /* We just need common CM header information.  Cast to any message. */
1195         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1196         rej_msg = (struct cm_rej_msg *) msg->mad;
1197
1198         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1199         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1200         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1201         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1202         rej_msg->reason = cpu_to_be16(reason);
1203
1204         if (ari && ari_length) {
1205                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1206                 memcpy(rej_msg->ari, ari, ari_length);
1207         }
1208
1209         ret = ib_post_send_mad(msg, NULL);
1210         if (ret)
1211                 cm_free_msg(msg);
1212
1213         return ret;
1214 }
1215
1216 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1217                                     __be32 local_qpn, __be32 remote_qpn)
1218 {
1219         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1220                 ((local_ca_guid == remote_ca_guid) &&
1221                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1222 }
1223
1224 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1225                                             struct ib_sa_path_rec *primary_path,
1226                                             struct ib_sa_path_rec *alt_path)
1227 {
1228         memset(primary_path, 0, sizeof *primary_path);
1229         primary_path->dgid = req_msg->primary_local_gid;
1230         primary_path->sgid = req_msg->primary_remote_gid;
1231         primary_path->dlid = req_msg->primary_local_lid;
1232         primary_path->slid = req_msg->primary_remote_lid;
1233         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1234         primary_path->hop_limit = req_msg->primary_hop_limit;
1235         primary_path->traffic_class = req_msg->primary_traffic_class;
1236         primary_path->reversible = 1;
1237         primary_path->pkey = req_msg->pkey;
1238         primary_path->sl = cm_req_get_primary_sl(req_msg);
1239         primary_path->mtu_selector = IB_SA_EQ;
1240         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1241         primary_path->rate_selector = IB_SA_EQ;
1242         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1243         primary_path->packet_life_time_selector = IB_SA_EQ;
1244         primary_path->packet_life_time =
1245                 cm_req_get_primary_local_ack_timeout(req_msg);
1246         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1247
1248         if (req_msg->alt_local_lid) {
1249                 memset(alt_path, 0, sizeof *alt_path);
1250                 alt_path->dgid = req_msg->alt_local_gid;
1251                 alt_path->sgid = req_msg->alt_remote_gid;
1252                 alt_path->dlid = req_msg->alt_local_lid;
1253                 alt_path->slid = req_msg->alt_remote_lid;
1254                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1255                 alt_path->hop_limit = req_msg->alt_hop_limit;
1256                 alt_path->traffic_class = req_msg->alt_traffic_class;
1257                 alt_path->reversible = 1;
1258                 alt_path->pkey = req_msg->pkey;
1259                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1260                 alt_path->mtu_selector = IB_SA_EQ;
1261                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1262                 alt_path->rate_selector = IB_SA_EQ;
1263                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1264                 alt_path->packet_life_time_selector = IB_SA_EQ;
1265                 alt_path->packet_life_time =
1266                         cm_req_get_alt_local_ack_timeout(req_msg);
1267                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1268         }
1269 }
1270
1271 static void cm_format_req_event(struct cm_work *work,
1272                                 struct cm_id_private *cm_id_priv,
1273                                 struct ib_cm_id *listen_id)
1274 {
1275         struct cm_req_msg *req_msg;
1276         struct ib_cm_req_event_param *param;
1277
1278         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1279         param = &work->cm_event.param.req_rcvd;
1280         param->listen_id = listen_id;
1281         param->port = cm_id_priv->av.port->port_num;
1282         param->primary_path = &work->path[0];
1283         if (req_msg->alt_local_lid)
1284                 param->alternate_path = &work->path[1];
1285         else
1286                 param->alternate_path = NULL;
1287         param->remote_ca_guid = req_msg->local_ca_guid;
1288         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1289         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1290         param->qp_type = cm_req_get_qp_type(req_msg);
1291         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1292         param->responder_resources = cm_req_get_init_depth(req_msg);
1293         param->initiator_depth = cm_req_get_resp_res(req_msg);
1294         param->local_cm_response_timeout =
1295                                         cm_req_get_remote_resp_timeout(req_msg);
1296         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1297         param->remote_cm_response_timeout =
1298                                         cm_req_get_local_resp_timeout(req_msg);
1299         param->retry_count = cm_req_get_retry_count(req_msg);
1300         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1301         param->srq = cm_req_get_srq(req_msg);
1302         work->cm_event.private_data = &req_msg->private_data;
1303 }
1304
1305 static void cm_process_work(struct cm_id_private *cm_id_priv,
1306                             struct cm_work *work)
1307 {
1308         int ret;
1309
1310         /* We will typically only have the current event to report. */
1311         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1312         cm_free_work(work);
1313
1314         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1315                 spin_lock_irq(&cm_id_priv->lock);
1316                 work = cm_dequeue_work(cm_id_priv);
1317                 spin_unlock_irq(&cm_id_priv->lock);
1318                 BUG_ON(!work);
1319                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1320                                                 &work->cm_event);
1321                 cm_free_work(work);
1322         }
1323         cm_deref_id(cm_id_priv);
1324         if (ret)
1325                 cm_destroy_id(&cm_id_priv->id, ret);
1326 }
1327
1328 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1329                           struct cm_id_private *cm_id_priv,
1330                           enum cm_msg_response msg_mraed, u8 service_timeout,
1331                           const void *private_data, u8 private_data_len)
1332 {
1333         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1334         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1335         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1336         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1337         cm_mra_set_service_timeout(mra_msg, service_timeout);
1338
1339         if (private_data && private_data_len)
1340                 memcpy(mra_msg->private_data, private_data, private_data_len);
1341 }
1342
1343 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1344                           struct cm_id_private *cm_id_priv,
1345                           enum ib_cm_rej_reason reason,
1346                           void *ari,
1347                           u8 ari_length,
1348                           const void *private_data,
1349                           u8 private_data_len)
1350 {
1351         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1352         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1353
1354         switch(cm_id_priv->id.state) {
1355         case IB_CM_REQ_RCVD:
1356                 rej_msg->local_comm_id = 0;
1357                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1358                 break;
1359         case IB_CM_MRA_REQ_SENT:
1360                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1361                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1362                 break;
1363         case IB_CM_REP_RCVD:
1364         case IB_CM_MRA_REP_SENT:
1365                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1366                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1367                 break;
1368         default:
1369                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1370                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1371                 break;
1372         }
1373
1374         rej_msg->reason = cpu_to_be16(reason);
1375         if (ari && ari_length) {
1376                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1377                 memcpy(rej_msg->ari, ari, ari_length);
1378         }
1379
1380         if (private_data && private_data_len)
1381                 memcpy(rej_msg->private_data, private_data, private_data_len);
1382 }
1383
1384 static void cm_dup_req_handler(struct cm_work *work,
1385                                struct cm_id_private *cm_id_priv)
1386 {
1387         struct ib_mad_send_buf *msg = NULL;
1388         int ret;
1389
1390         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1391                         counter[CM_REQ_COUNTER]);
1392
1393         /* Quick state check to discard duplicate REQs. */
1394         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1395                 return;
1396
1397         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1398         if (ret)
1399                 return;
1400
1401         spin_lock_irq(&cm_id_priv->lock);
1402         switch (cm_id_priv->id.state) {
1403         case IB_CM_MRA_REQ_SENT:
1404                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1405                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1406                               cm_id_priv->private_data,
1407                               cm_id_priv->private_data_len);
1408                 break;
1409         case IB_CM_TIMEWAIT:
1410                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1411                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1412                 break;
1413         default:
1414                 goto unlock;
1415         }
1416         spin_unlock_irq(&cm_id_priv->lock);
1417
1418         ret = ib_post_send_mad(msg, NULL);
1419         if (ret)
1420                 goto free;
1421         return;
1422
1423 unlock: spin_unlock_irq(&cm_id_priv->lock);
1424 free:   cm_free_msg(msg);
1425 }
1426
1427 static struct cm_id_private * cm_match_req(struct cm_work *work,
1428                                            struct cm_id_private *cm_id_priv)
1429 {
1430         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1431         struct cm_timewait_info *timewait_info;
1432         struct cm_req_msg *req_msg;
1433
1434         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1435
1436         /* Check for possible duplicate REQ. */
1437         spin_lock_irq(&cm.lock);
1438         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1439         if (timewait_info) {
1440                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1441                                            timewait_info->work.remote_id);
1442                 spin_unlock_irq(&cm.lock);
1443                 if (cur_cm_id_priv) {
1444                         cm_dup_req_handler(work, cur_cm_id_priv);
1445                         cm_deref_id(cur_cm_id_priv);
1446                 }
1447                 return NULL;
1448         }
1449
1450         /* Check for stale connections. */
1451         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1452         if (timewait_info) {
1453                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1454                 spin_unlock_irq(&cm.lock);
1455                 cm_issue_rej(work->port, work->mad_recv_wc,
1456                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1457                              NULL, 0);
1458                 return NULL;
1459         }
1460
1461         /* Find matching listen request. */
1462         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1463                                            req_msg->service_id,
1464                                            req_msg->private_data);
1465         if (!listen_cm_id_priv) {
1466                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1467                 spin_unlock_irq(&cm.lock);
1468                 cm_issue_rej(work->port, work->mad_recv_wc,
1469                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1470                              NULL, 0);
1471                 goto out;
1472         }
1473         atomic_inc(&listen_cm_id_priv->refcount);
1474         atomic_inc(&cm_id_priv->refcount);
1475         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1476         atomic_inc(&cm_id_priv->work_count);
1477         spin_unlock_irq(&cm.lock);
1478 out:
1479         return listen_cm_id_priv;
1480 }
1481
1482 /*
1483  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1484  * we need to override the LID/SL data in the REQ with the LID information
1485  * in the work completion.
1486  */
1487 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1488 {
1489         if (!cm_req_get_primary_subnet_local(req_msg)) {
1490                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1491                         req_msg->primary_local_lid = cpu_to_be16(wc->slid);
1492                         cm_req_set_primary_sl(req_msg, wc->sl);
1493                 }
1494
1495                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1496                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1497         }
1498
1499         if (!cm_req_get_alt_subnet_local(req_msg)) {
1500                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1501                         req_msg->alt_local_lid = cpu_to_be16(wc->slid);
1502                         cm_req_set_alt_sl(req_msg, wc->sl);
1503                 }
1504
1505                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1506                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1507         }
1508 }
1509
1510 static int cm_req_handler(struct cm_work *work)
1511 {
1512         struct ib_cm_id *cm_id;
1513         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1514         struct cm_req_msg *req_msg;
1515         int ret;
1516
1517         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1518
1519         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1520         if (IS_ERR(cm_id))
1521                 return PTR_ERR(cm_id);
1522
1523         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1524         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1525         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1526                                 work->mad_recv_wc->recv_buf.grh,
1527                                 &cm_id_priv->av);
1528         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1529                                                             id.local_id);
1530         if (IS_ERR(cm_id_priv->timewait_info)) {
1531                 ret = PTR_ERR(cm_id_priv->timewait_info);
1532                 goto destroy;
1533         }
1534         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1535         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1536         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1537
1538         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1539         if (!listen_cm_id_priv) {
1540                 ret = -EINVAL;
1541                 kfree(cm_id_priv->timewait_info);
1542                 goto destroy;
1543         }
1544
1545         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1546         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1547         cm_id_priv->id.service_id = req_msg->service_id;
1548         cm_id_priv->id.service_mask = __constant_cpu_to_be64(~0ULL);
1549
1550         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1551         cm_format_paths_from_req(req_msg, &work->path[0], &work->path[1]);
1552         ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av);
1553         if (ret) {
1554                 ib_get_cached_gid(work->port->cm_dev->ib_device,
1555                                   work->port->port_num, 0, &work->path[0].sgid);
1556                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1557                                &work->path[0].sgid, sizeof work->path[0].sgid,
1558                                NULL, 0);
1559                 goto rejected;
1560         }
1561         if (req_msg->alt_local_lid) {
1562                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av);
1563                 if (ret) {
1564                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1565                                        &work->path[0].sgid,
1566                                        sizeof work->path[0].sgid, NULL, 0);
1567                         goto rejected;
1568                 }
1569         }
1570         cm_id_priv->tid = req_msg->hdr.tid;
1571         cm_id_priv->timeout_ms = cm_convert_to_ms(
1572                                         cm_req_get_local_resp_timeout(req_msg));
1573         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1574         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1575         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1576         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1577         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1578         cm_id_priv->pkey = req_msg->pkey;
1579         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1580         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1581         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1582         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1583
1584         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1585         cm_process_work(cm_id_priv, work);
1586         cm_deref_id(listen_cm_id_priv);
1587         return 0;
1588
1589 rejected:
1590         atomic_dec(&cm_id_priv->refcount);
1591         cm_deref_id(listen_cm_id_priv);
1592 destroy:
1593         ib_destroy_cm_id(cm_id);
1594         return ret;
1595 }
1596
1597 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1598                           struct cm_id_private *cm_id_priv,
1599                           struct ib_cm_rep_param *param)
1600 {
1601         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1602         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1603         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1604         cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1605         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1606         rep_msg->resp_resources = param->responder_resources;
1607         rep_msg->initiator_depth = param->initiator_depth;
1608         cm_rep_set_target_ack_delay(rep_msg,
1609                                     cm_id_priv->av.port->cm_dev->ack_delay);
1610         cm_rep_set_failover(rep_msg, param->failover_accepted);
1611         cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1612         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1613         cm_rep_set_srq(rep_msg, param->srq);
1614         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1615
1616         if (param->private_data && param->private_data_len)
1617                 memcpy(rep_msg->private_data, param->private_data,
1618                        param->private_data_len);
1619 }
1620
1621 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1622                    struct ib_cm_rep_param *param)
1623 {
1624         struct cm_id_private *cm_id_priv;
1625         struct ib_mad_send_buf *msg;
1626         struct cm_rep_msg *rep_msg;
1627         unsigned long flags;
1628         int ret;
1629
1630         if (param->private_data &&
1631             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1632                 return -EINVAL;
1633
1634         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1635         spin_lock_irqsave(&cm_id_priv->lock, flags);
1636         if (cm_id->state != IB_CM_REQ_RCVD &&
1637             cm_id->state != IB_CM_MRA_REQ_SENT) {
1638                 ret = -EINVAL;
1639                 goto out;
1640         }
1641
1642         ret = cm_alloc_msg(cm_id_priv, &msg);
1643         if (ret)
1644                 goto out;
1645
1646         rep_msg = (struct cm_rep_msg *) msg->mad;
1647         cm_format_rep(rep_msg, cm_id_priv, param);
1648         msg->timeout_ms = cm_id_priv->timeout_ms;
1649         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
1650
1651         ret = ib_post_send_mad(msg, NULL);
1652         if (ret) {
1653                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1654                 cm_free_msg(msg);
1655                 return ret;
1656         }
1657
1658         cm_id->state = IB_CM_REP_SENT;
1659         cm_id_priv->msg = msg;
1660         cm_id_priv->initiator_depth = param->initiator_depth;
1661         cm_id_priv->responder_resources = param->responder_resources;
1662         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
1663         cm_id_priv->local_qpn = cm_rep_get_local_qpn(rep_msg);
1664
1665 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1666         return ret;
1667 }
1668 EXPORT_SYMBOL(ib_send_cm_rep);
1669
1670 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
1671                           struct cm_id_private *cm_id_priv,
1672                           const void *private_data,
1673                           u8 private_data_len)
1674 {
1675         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
1676         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
1677         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
1678
1679         if (private_data && private_data_len)
1680                 memcpy(rtu_msg->private_data, private_data, private_data_len);
1681 }
1682
1683 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
1684                    const void *private_data,
1685                    u8 private_data_len)
1686 {
1687         struct cm_id_private *cm_id_priv;
1688         struct ib_mad_send_buf *msg;
1689         unsigned long flags;
1690         void *data;
1691         int ret;
1692
1693         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
1694                 return -EINVAL;
1695
1696         data = cm_copy_private_data(private_data, private_data_len);
1697         if (IS_ERR(data))
1698                 return PTR_ERR(data);
1699
1700         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1701         spin_lock_irqsave(&cm_id_priv->lock, flags);
1702         if (cm_id->state != IB_CM_REP_RCVD &&
1703             cm_id->state != IB_CM_MRA_REP_SENT) {
1704                 ret = -EINVAL;
1705                 goto error;
1706         }
1707
1708         ret = cm_alloc_msg(cm_id_priv, &msg);
1709         if (ret)
1710                 goto error;
1711
1712         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1713                       private_data, private_data_len);
1714
1715         ret = ib_post_send_mad(msg, NULL);
1716         if (ret) {
1717                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1718                 cm_free_msg(msg);
1719                 kfree(data);
1720                 return ret;
1721         }
1722
1723         cm_id->state = IB_CM_ESTABLISHED;
1724         cm_set_private_data(cm_id_priv, data, private_data_len);
1725         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1726         return 0;
1727
1728 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1729         kfree(data);
1730         return ret;
1731 }
1732 EXPORT_SYMBOL(ib_send_cm_rtu);
1733
1734 static void cm_format_rep_event(struct cm_work *work)
1735 {
1736         struct cm_rep_msg *rep_msg;
1737         struct ib_cm_rep_event_param *param;
1738
1739         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1740         param = &work->cm_event.param.rep_rcvd;
1741         param->remote_ca_guid = rep_msg->local_ca_guid;
1742         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
1743         param->remote_qpn = be32_to_cpu(cm_rep_get_local_qpn(rep_msg));
1744         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
1745         param->responder_resources = rep_msg->initiator_depth;
1746         param->initiator_depth = rep_msg->resp_resources;
1747         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1748         param->failover_accepted = cm_rep_get_failover(rep_msg);
1749         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
1750         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1751         param->srq = cm_rep_get_srq(rep_msg);
1752         work->cm_event.private_data = &rep_msg->private_data;
1753 }
1754
1755 static void cm_dup_rep_handler(struct cm_work *work)
1756 {
1757         struct cm_id_private *cm_id_priv;
1758         struct cm_rep_msg *rep_msg;
1759         struct ib_mad_send_buf *msg = NULL;
1760         int ret;
1761
1762         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
1763         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
1764                                    rep_msg->local_comm_id);
1765         if (!cm_id_priv)
1766                 return;
1767
1768         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1769                         counter[CM_REP_COUNTER]);
1770         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1771         if (ret)
1772                 goto deref;
1773
1774         spin_lock_irq(&cm_id_priv->lock);
1775         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
1776                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
1777                               cm_id_priv->private_data,
1778                               cm_id_priv->private_data_len);
1779         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
1780                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1781                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
1782                               cm_id_priv->private_data,
1783                               cm_id_priv->private_data_len);
1784         else
1785                 goto unlock;
1786         spin_unlock_irq(&cm_id_priv->lock);
1787
1788         ret = ib_post_send_mad(msg, NULL);
1789         if (ret)
1790                 goto free;
1791         goto deref;
1792
1793 unlock: spin_unlock_irq(&cm_id_priv->lock);
1794 free:   cm_free_msg(msg);
1795 deref:  cm_deref_id(cm_id_priv);
1796 }
1797
1798 static int cm_rep_handler(struct cm_work *work)
1799 {
1800         struct cm_id_private *cm_id_priv;
1801         struct cm_rep_msg *rep_msg;
1802         int ret;
1803
1804         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
1805         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
1806         if (!cm_id_priv) {
1807                 cm_dup_rep_handler(work);
1808                 return -EINVAL;
1809         }
1810
1811         cm_format_rep_event(work);
1812
1813         spin_lock_irq(&cm_id_priv->lock);
1814         switch (cm_id_priv->id.state) {
1815         case IB_CM_REQ_SENT:
1816         case IB_CM_MRA_REQ_RCVD:
1817                 break;
1818         default:
1819                 spin_unlock_irq(&cm_id_priv->lock);
1820                 ret = -EINVAL;
1821                 goto error;
1822         }
1823
1824         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
1825         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
1826         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1827
1828         spin_lock(&cm.lock);
1829         /* Check for duplicate REP. */
1830         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
1831                 spin_unlock(&cm.lock);
1832                 spin_unlock_irq(&cm_id_priv->lock);
1833                 ret = -EINVAL;
1834                 goto error;
1835         }
1836         /* Check for a stale connection. */
1837         if (cm_insert_remote_qpn(cm_id_priv->timewait_info)) {
1838                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
1839                          &cm.remote_id_table);
1840                 cm_id_priv->timewait_info->inserted_remote_id = 0;
1841                 spin_unlock(&cm.lock);
1842                 spin_unlock_irq(&cm_id_priv->lock);
1843                 cm_issue_rej(work->port, work->mad_recv_wc,
1844                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
1845                              NULL, 0);
1846                 ret = -EINVAL;
1847                 goto error;
1848         }
1849         spin_unlock(&cm.lock);
1850
1851         cm_id_priv->id.state = IB_CM_REP_RCVD;
1852         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
1853         cm_id_priv->remote_qpn = cm_rep_get_local_qpn(rep_msg);
1854         cm_id_priv->initiator_depth = rep_msg->resp_resources;
1855         cm_id_priv->responder_resources = rep_msg->initiator_depth;
1856         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
1857         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
1858         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
1859         cm_id_priv->av.timeout =
1860                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1861                                        cm_id_priv->av.timeout - 1);
1862         cm_id_priv->alt_av.timeout =
1863                         cm_ack_timeout(cm_id_priv->target_ack_delay,
1864                                        cm_id_priv->alt_av.timeout - 1);
1865
1866         /* todo: handle peer_to_peer */
1867
1868         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1869         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1870         if (!ret)
1871                 list_add_tail(&work->list, &cm_id_priv->work_list);
1872         spin_unlock_irq(&cm_id_priv->lock);
1873
1874         if (ret)
1875                 cm_process_work(cm_id_priv, work);
1876         else
1877                 cm_deref_id(cm_id_priv);
1878         return 0;
1879
1880 error:
1881         cm_deref_id(cm_id_priv);
1882         return ret;
1883 }
1884
1885 static int cm_establish_handler(struct cm_work *work)
1886 {
1887         struct cm_id_private *cm_id_priv;
1888         int ret;
1889
1890         /* See comment in cm_establish about lookup. */
1891         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
1892         if (!cm_id_priv)
1893                 return -EINVAL;
1894
1895         spin_lock_irq(&cm_id_priv->lock);
1896         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
1897                 spin_unlock_irq(&cm_id_priv->lock);
1898                 goto out;
1899         }
1900
1901         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1902         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1903         if (!ret)
1904                 list_add_tail(&work->list, &cm_id_priv->work_list);
1905         spin_unlock_irq(&cm_id_priv->lock);
1906
1907         if (ret)
1908                 cm_process_work(cm_id_priv, work);
1909         else
1910                 cm_deref_id(cm_id_priv);
1911         return 0;
1912 out:
1913         cm_deref_id(cm_id_priv);
1914         return -EINVAL;
1915 }
1916
1917 static int cm_rtu_handler(struct cm_work *work)
1918 {
1919         struct cm_id_private *cm_id_priv;
1920         struct cm_rtu_msg *rtu_msg;
1921         int ret;
1922
1923         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
1924         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
1925                                    rtu_msg->local_comm_id);
1926         if (!cm_id_priv)
1927                 return -EINVAL;
1928
1929         work->cm_event.private_data = &rtu_msg->private_data;
1930
1931         spin_lock_irq(&cm_id_priv->lock);
1932         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
1933             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
1934                 spin_unlock_irq(&cm_id_priv->lock);
1935                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1936                                 counter[CM_RTU_COUNTER]);
1937                 goto out;
1938         }
1939         cm_id_priv->id.state = IB_CM_ESTABLISHED;
1940
1941         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1942         ret = atomic_inc_and_test(&cm_id_priv->work_count);
1943         if (!ret)
1944                 list_add_tail(&work->list, &cm_id_priv->work_list);
1945         spin_unlock_irq(&cm_id_priv->lock);
1946
1947         if (ret)
1948                 cm_process_work(cm_id_priv, work);
1949         else
1950                 cm_deref_id(cm_id_priv);
1951         return 0;
1952 out:
1953         cm_deref_id(cm_id_priv);
1954         return -EINVAL;
1955 }
1956
1957 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
1958                           struct cm_id_private *cm_id_priv,
1959                           const void *private_data,
1960                           u8 private_data_len)
1961 {
1962         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
1963                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
1964         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
1965         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
1966         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
1967
1968         if (private_data && private_data_len)
1969                 memcpy(dreq_msg->private_data, private_data, private_data_len);
1970 }
1971
1972 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
1973                     const void *private_data,
1974                     u8 private_data_len)
1975 {
1976         struct cm_id_private *cm_id_priv;
1977         struct ib_mad_send_buf *msg;
1978         unsigned long flags;
1979         int ret;
1980
1981         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
1982                 return -EINVAL;
1983
1984         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1985         spin_lock_irqsave(&cm_id_priv->lock, flags);
1986         if (cm_id->state != IB_CM_ESTABLISHED) {
1987                 ret = -EINVAL;
1988                 goto out;
1989         }
1990
1991         ret = cm_alloc_msg(cm_id_priv, &msg);
1992         if (ret) {
1993                 cm_enter_timewait(cm_id_priv);
1994                 goto out;
1995         }
1996
1997         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
1998                        private_data, private_data_len);
1999         msg->timeout_ms = cm_id_priv->timeout_ms;
2000         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2001
2002         ret = ib_post_send_mad(msg, NULL);
2003         if (ret) {
2004                 cm_enter_timewait(cm_id_priv);
2005                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2006                 cm_free_msg(msg);
2007                 return ret;
2008         }
2009
2010         cm_id->state = IB_CM_DREQ_SENT;
2011         cm_id_priv->msg = msg;
2012 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2013         return ret;
2014 }
2015 EXPORT_SYMBOL(ib_send_cm_dreq);
2016
2017 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2018                           struct cm_id_private *cm_id_priv,
2019                           const void *private_data,
2020                           u8 private_data_len)
2021 {
2022         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2023         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2024         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2025
2026         if (private_data && private_data_len)
2027                 memcpy(drep_msg->private_data, private_data, private_data_len);
2028 }
2029
2030 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2031                     const void *private_data,
2032                     u8 private_data_len)
2033 {
2034         struct cm_id_private *cm_id_priv;
2035         struct ib_mad_send_buf *msg;
2036         unsigned long flags;
2037         void *data;
2038         int ret;
2039
2040         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2041                 return -EINVAL;
2042
2043         data = cm_copy_private_data(private_data, private_data_len);
2044         if (IS_ERR(data))
2045                 return PTR_ERR(data);
2046
2047         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2048         spin_lock_irqsave(&cm_id_priv->lock, flags);
2049         if (cm_id->state != IB_CM_DREQ_RCVD) {
2050                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2051                 kfree(data);
2052                 return -EINVAL;
2053         }
2054
2055         cm_set_private_data(cm_id_priv, data, private_data_len);
2056         cm_enter_timewait(cm_id_priv);
2057
2058         ret = cm_alloc_msg(cm_id_priv, &msg);
2059         if (ret)
2060                 goto out;
2061
2062         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2063                        private_data, private_data_len);
2064
2065         ret = ib_post_send_mad(msg, NULL);
2066         if (ret) {
2067                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2068                 cm_free_msg(msg);
2069                 return ret;
2070         }
2071
2072 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2073         return ret;
2074 }
2075 EXPORT_SYMBOL(ib_send_cm_drep);
2076
2077 static int cm_issue_drep(struct cm_port *port,
2078                          struct ib_mad_recv_wc *mad_recv_wc)
2079 {
2080         struct ib_mad_send_buf *msg = NULL;
2081         struct cm_dreq_msg *dreq_msg;
2082         struct cm_drep_msg *drep_msg;
2083         int ret;
2084
2085         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2086         if (ret)
2087                 return ret;
2088
2089         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2090         drep_msg = (struct cm_drep_msg *) msg->mad;
2091
2092         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2093         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2094         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2095
2096         ret = ib_post_send_mad(msg, NULL);
2097         if (ret)
2098                 cm_free_msg(msg);
2099
2100         return ret;
2101 }
2102
2103 static int cm_dreq_handler(struct cm_work *work)
2104 {
2105         struct cm_id_private *cm_id_priv;
2106         struct cm_dreq_msg *dreq_msg;
2107         struct ib_mad_send_buf *msg = NULL;
2108         int ret;
2109
2110         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2111         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2112                                    dreq_msg->local_comm_id);
2113         if (!cm_id_priv) {
2114                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2115                                 counter[CM_DREQ_COUNTER]);
2116                 cm_issue_drep(work->port, work->mad_recv_wc);
2117                 return -EINVAL;
2118         }
2119
2120         work->cm_event.private_data = &dreq_msg->private_data;
2121
2122         spin_lock_irq(&cm_id_priv->lock);
2123         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2124                 goto unlock;
2125
2126         switch (cm_id_priv->id.state) {
2127         case IB_CM_REP_SENT:
2128         case IB_CM_DREQ_SENT:
2129                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2130                 break;
2131         case IB_CM_ESTABLISHED:
2132         case IB_CM_MRA_REP_RCVD:
2133                 break;
2134         case IB_CM_TIMEWAIT:
2135                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2136                                 counter[CM_DREQ_COUNTER]);
2137                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2138                         goto unlock;
2139
2140                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2141                                cm_id_priv->private_data,
2142                                cm_id_priv->private_data_len);
2143                 spin_unlock_irq(&cm_id_priv->lock);
2144
2145                 if (ib_post_send_mad(msg, NULL))
2146                         cm_free_msg(msg);
2147                 goto deref;
2148         case IB_CM_DREQ_RCVD:
2149                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2150                                 counter[CM_DREQ_COUNTER]);
2151                 goto unlock;
2152         default:
2153                 goto unlock;
2154         }
2155         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2156         cm_id_priv->tid = dreq_msg->hdr.tid;
2157         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2158         if (!ret)
2159                 list_add_tail(&work->list, &cm_id_priv->work_list);
2160         spin_unlock_irq(&cm_id_priv->lock);
2161
2162         if (ret)
2163                 cm_process_work(cm_id_priv, work);
2164         else
2165                 cm_deref_id(cm_id_priv);
2166         return 0;
2167
2168 unlock: spin_unlock_irq(&cm_id_priv->lock);
2169 deref:  cm_deref_id(cm_id_priv);
2170         return -EINVAL;
2171 }
2172
2173 static int cm_drep_handler(struct cm_work *work)
2174 {
2175         struct cm_id_private *cm_id_priv;
2176         struct cm_drep_msg *drep_msg;
2177         int ret;
2178
2179         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2180         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2181                                    drep_msg->local_comm_id);
2182         if (!cm_id_priv)
2183                 return -EINVAL;
2184
2185         work->cm_event.private_data = &drep_msg->private_data;
2186
2187         spin_lock_irq(&cm_id_priv->lock);
2188         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2189             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2190                 spin_unlock_irq(&cm_id_priv->lock);
2191                 goto out;
2192         }
2193         cm_enter_timewait(cm_id_priv);
2194
2195         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2196         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2197         if (!ret)
2198                 list_add_tail(&work->list, &cm_id_priv->work_list);
2199         spin_unlock_irq(&cm_id_priv->lock);
2200
2201         if (ret)
2202                 cm_process_work(cm_id_priv, work);
2203         else
2204                 cm_deref_id(cm_id_priv);
2205         return 0;
2206 out:
2207         cm_deref_id(cm_id_priv);
2208         return -EINVAL;
2209 }
2210
2211 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2212                    enum ib_cm_rej_reason reason,
2213                    void *ari,
2214                    u8 ari_length,
2215                    const void *private_data,
2216                    u8 private_data_len)
2217 {
2218         struct cm_id_private *cm_id_priv;
2219         struct ib_mad_send_buf *msg;
2220         unsigned long flags;
2221         int ret;
2222
2223         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2224             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2225                 return -EINVAL;
2226
2227         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2228
2229         spin_lock_irqsave(&cm_id_priv->lock, flags);
2230         switch (cm_id->state) {
2231         case IB_CM_REQ_SENT:
2232         case IB_CM_MRA_REQ_RCVD:
2233         case IB_CM_REQ_RCVD:
2234         case IB_CM_MRA_REQ_SENT:
2235         case IB_CM_REP_RCVD:
2236         case IB_CM_MRA_REP_SENT:
2237                 ret = cm_alloc_msg(cm_id_priv, &msg);
2238                 if (!ret)
2239                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2240                                       cm_id_priv, reason, ari, ari_length,
2241                                       private_data, private_data_len);
2242
2243                 cm_reset_to_idle(cm_id_priv);
2244                 break;
2245         case IB_CM_REP_SENT:
2246         case IB_CM_MRA_REP_RCVD:
2247                 ret = cm_alloc_msg(cm_id_priv, &msg);
2248                 if (!ret)
2249                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2250                                       cm_id_priv, reason, ari, ari_length,
2251                                       private_data, private_data_len);
2252
2253                 cm_enter_timewait(cm_id_priv);
2254                 break;
2255         default:
2256                 ret = -EINVAL;
2257                 goto out;
2258         }
2259
2260         if (ret)
2261                 goto out;
2262
2263         ret = ib_post_send_mad(msg, NULL);
2264         if (ret)
2265                 cm_free_msg(msg);
2266
2267 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2268         return ret;
2269 }
2270 EXPORT_SYMBOL(ib_send_cm_rej);
2271
2272 static void cm_format_rej_event(struct cm_work *work)
2273 {
2274         struct cm_rej_msg *rej_msg;
2275         struct ib_cm_rej_event_param *param;
2276
2277         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2278         param = &work->cm_event.param.rej_rcvd;
2279         param->ari = rej_msg->ari;
2280         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2281         param->reason = __be16_to_cpu(rej_msg->reason);
2282         work->cm_event.private_data = &rej_msg->private_data;
2283 }
2284
2285 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2286 {
2287         struct cm_timewait_info *timewait_info;
2288         struct cm_id_private *cm_id_priv;
2289         __be32 remote_id;
2290
2291         remote_id = rej_msg->local_comm_id;
2292
2293         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2294                 spin_lock_irq(&cm.lock);
2295                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2296                                                   remote_id);
2297                 if (!timewait_info) {
2298                         spin_unlock_irq(&cm.lock);
2299                         return NULL;
2300                 }
2301                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2302                                       (timewait_info->work.local_id ^
2303                                        cm.random_id_operand));
2304                 if (cm_id_priv) {
2305                         if (cm_id_priv->id.remote_id == remote_id)
2306                                 atomic_inc(&cm_id_priv->refcount);
2307                         else
2308                                 cm_id_priv = NULL;
2309                 }
2310                 spin_unlock_irq(&cm.lock);
2311         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2312                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2313         else
2314                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2315
2316         return cm_id_priv;
2317 }
2318
2319 static int cm_rej_handler(struct cm_work *work)
2320 {
2321         struct cm_id_private *cm_id_priv;
2322         struct cm_rej_msg *rej_msg;
2323         int ret;
2324
2325         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2326         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2327         if (!cm_id_priv)
2328                 return -EINVAL;
2329
2330         cm_format_rej_event(work);
2331
2332         spin_lock_irq(&cm_id_priv->lock);
2333         switch (cm_id_priv->id.state) {
2334         case IB_CM_REQ_SENT:
2335         case IB_CM_MRA_REQ_RCVD:
2336         case IB_CM_REP_SENT:
2337         case IB_CM_MRA_REP_RCVD:
2338                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2339                 /* fall through */
2340         case IB_CM_REQ_RCVD:
2341         case IB_CM_MRA_REQ_SENT:
2342                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2343                         cm_enter_timewait(cm_id_priv);
2344                 else
2345                         cm_reset_to_idle(cm_id_priv);
2346                 break;
2347         case IB_CM_DREQ_SENT:
2348                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2349                 /* fall through */
2350         case IB_CM_REP_RCVD:
2351         case IB_CM_MRA_REP_SENT:
2352         case IB_CM_ESTABLISHED:
2353                 cm_enter_timewait(cm_id_priv);
2354                 break;
2355         default:
2356                 spin_unlock_irq(&cm_id_priv->lock);
2357                 ret = -EINVAL;
2358                 goto out;
2359         }
2360
2361         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2362         if (!ret)
2363                 list_add_tail(&work->list, &cm_id_priv->work_list);
2364         spin_unlock_irq(&cm_id_priv->lock);
2365
2366         if (ret)
2367                 cm_process_work(cm_id_priv, work);
2368         else
2369                 cm_deref_id(cm_id_priv);
2370         return 0;
2371 out:
2372         cm_deref_id(cm_id_priv);
2373         return -EINVAL;
2374 }
2375
2376 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2377                    u8 service_timeout,
2378                    const void *private_data,
2379                    u8 private_data_len)
2380 {
2381         struct cm_id_private *cm_id_priv;
2382         struct ib_mad_send_buf *msg;
2383         enum ib_cm_state cm_state;
2384         enum ib_cm_lap_state lap_state;
2385         enum cm_msg_response msg_response;
2386         void *data;
2387         unsigned long flags;
2388         int ret;
2389
2390         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2391                 return -EINVAL;
2392
2393         data = cm_copy_private_data(private_data, private_data_len);
2394         if (IS_ERR(data))
2395                 return PTR_ERR(data);
2396
2397         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2398
2399         spin_lock_irqsave(&cm_id_priv->lock, flags);
2400         switch(cm_id_priv->id.state) {
2401         case IB_CM_REQ_RCVD:
2402                 cm_state = IB_CM_MRA_REQ_SENT;
2403                 lap_state = cm_id->lap_state;
2404                 msg_response = CM_MSG_RESPONSE_REQ;
2405                 break;
2406         case IB_CM_REP_RCVD:
2407                 cm_state = IB_CM_MRA_REP_SENT;
2408                 lap_state = cm_id->lap_state;
2409                 msg_response = CM_MSG_RESPONSE_REP;
2410                 break;
2411         case IB_CM_ESTABLISHED:
2412                 cm_state = cm_id->state;
2413                 lap_state = IB_CM_MRA_LAP_SENT;
2414                 msg_response = CM_MSG_RESPONSE_OTHER;
2415                 break;
2416         default:
2417                 ret = -EINVAL;
2418                 goto error1;
2419         }
2420
2421         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2422                 ret = cm_alloc_msg(cm_id_priv, &msg);
2423                 if (ret)
2424                         goto error1;
2425
2426                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2427                               msg_response, service_timeout,
2428                               private_data, private_data_len);
2429                 ret = ib_post_send_mad(msg, NULL);
2430                 if (ret)
2431                         goto error2;
2432         }
2433
2434         cm_id->state = cm_state;
2435         cm_id->lap_state = lap_state;
2436         cm_id_priv->service_timeout = service_timeout;
2437         cm_set_private_data(cm_id_priv, data, private_data_len);
2438         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2439         return 0;
2440
2441 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2442         kfree(data);
2443         return ret;
2444
2445 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2446         kfree(data);
2447         cm_free_msg(msg);
2448         return ret;
2449 }
2450 EXPORT_SYMBOL(ib_send_cm_mra);
2451
2452 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2453 {
2454         switch (cm_mra_get_msg_mraed(mra_msg)) {
2455         case CM_MSG_RESPONSE_REQ:
2456                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2457         case CM_MSG_RESPONSE_REP:
2458         case CM_MSG_RESPONSE_OTHER:
2459                 return cm_acquire_id(mra_msg->remote_comm_id,
2460                                      mra_msg->local_comm_id);
2461         default:
2462                 return NULL;
2463         }
2464 }
2465
2466 static int cm_mra_handler(struct cm_work *work)
2467 {
2468         struct cm_id_private *cm_id_priv;
2469         struct cm_mra_msg *mra_msg;
2470         int timeout, ret;
2471
2472         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2473         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2474         if (!cm_id_priv)
2475                 return -EINVAL;
2476
2477         work->cm_event.private_data = &mra_msg->private_data;
2478         work->cm_event.param.mra_rcvd.service_timeout =
2479                                         cm_mra_get_service_timeout(mra_msg);
2480         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2481                   cm_convert_to_ms(cm_id_priv->av.timeout);
2482
2483         spin_lock_irq(&cm_id_priv->lock);
2484         switch (cm_id_priv->id.state) {
2485         case IB_CM_REQ_SENT:
2486                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2487                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2488                                   cm_id_priv->msg, timeout))
2489                         goto out;
2490                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2491                 break;
2492         case IB_CM_REP_SENT:
2493                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2494                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2495                                   cm_id_priv->msg, timeout))
2496                         goto out;
2497                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2498                 break;
2499         case IB_CM_ESTABLISHED:
2500                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2501                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2502                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2503                                   cm_id_priv->msg, timeout)) {
2504                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2505                                 atomic_long_inc(&work->port->
2506                                                 counter_group[CM_RECV_DUPLICATES].
2507                                                 counter[CM_MRA_COUNTER]);
2508                         goto out;
2509                 }
2510                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2511                 break;
2512         case IB_CM_MRA_REQ_RCVD:
2513         case IB_CM_MRA_REP_RCVD:
2514                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2515                                 counter[CM_MRA_COUNTER]);
2516                 /* fall through */
2517         default:
2518                 goto out;
2519         }
2520
2521         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2522                                       cm_id_priv->id.state;
2523         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2524         if (!ret)
2525                 list_add_tail(&work->list, &cm_id_priv->work_list);
2526         spin_unlock_irq(&cm_id_priv->lock);
2527
2528         if (ret)
2529                 cm_process_work(cm_id_priv, work);
2530         else
2531                 cm_deref_id(cm_id_priv);
2532         return 0;
2533 out:
2534         spin_unlock_irq(&cm_id_priv->lock);
2535         cm_deref_id(cm_id_priv);
2536         return -EINVAL;
2537 }
2538
2539 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2540                           struct cm_id_private *cm_id_priv,
2541                           struct ib_sa_path_rec *alternate_path,
2542                           const void *private_data,
2543                           u8 private_data_len)
2544 {
2545         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2546                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2547         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2548         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2549         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2550         /* todo: need remote CM response timeout */
2551         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2552         lap_msg->alt_local_lid = alternate_path->slid;
2553         lap_msg->alt_remote_lid = alternate_path->dlid;
2554         lap_msg->alt_local_gid = alternate_path->sgid;
2555         lap_msg->alt_remote_gid = alternate_path->dgid;
2556         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2557         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2558         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2559         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2560         cm_lap_set_sl(lap_msg, alternate_path->sl);
2561         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2562         cm_lap_set_local_ack_timeout(lap_msg,
2563                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2564                                alternate_path->packet_life_time));
2565
2566         if (private_data && private_data_len)
2567                 memcpy(lap_msg->private_data, private_data, private_data_len);
2568 }
2569
2570 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2571                    struct ib_sa_path_rec *alternate_path,
2572                    const void *private_data,
2573                    u8 private_data_len)
2574 {
2575         struct cm_id_private *cm_id_priv;
2576         struct ib_mad_send_buf *msg;
2577         unsigned long flags;
2578         int ret;
2579
2580         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2581                 return -EINVAL;
2582
2583         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2584         spin_lock_irqsave(&cm_id_priv->lock, flags);
2585         if (cm_id->state != IB_CM_ESTABLISHED ||
2586             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2587              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2588                 ret = -EINVAL;
2589                 goto out;
2590         }
2591
2592         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
2593         if (ret)
2594                 goto out;
2595         cm_id_priv->alt_av.timeout =
2596                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2597                                        cm_id_priv->alt_av.timeout - 1);
2598
2599         ret = cm_alloc_msg(cm_id_priv, &msg);
2600         if (ret)
2601                 goto out;
2602
2603         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2604                       alternate_path, private_data, private_data_len);
2605         msg->timeout_ms = cm_id_priv->timeout_ms;
2606         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2607
2608         ret = ib_post_send_mad(msg, NULL);
2609         if (ret) {
2610                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2611                 cm_free_msg(msg);
2612                 return ret;
2613         }
2614
2615         cm_id->lap_state = IB_CM_LAP_SENT;
2616         cm_id_priv->msg = msg;
2617
2618 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2619         return ret;
2620 }
2621 EXPORT_SYMBOL(ib_send_cm_lap);
2622
2623 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2624                                     struct ib_sa_path_rec *path,
2625                                     struct cm_lap_msg *lap_msg)
2626 {
2627         memset(path, 0, sizeof *path);
2628         path->dgid = lap_msg->alt_local_gid;
2629         path->sgid = lap_msg->alt_remote_gid;
2630         path->dlid = lap_msg->alt_local_lid;
2631         path->slid = lap_msg->alt_remote_lid;
2632         path->flow_label = cm_lap_get_flow_label(lap_msg);
2633         path->hop_limit = lap_msg->alt_hop_limit;
2634         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2635         path->reversible = 1;
2636         path->pkey = cm_id_priv->pkey;
2637         path->sl = cm_lap_get_sl(lap_msg);
2638         path->mtu_selector = IB_SA_EQ;
2639         path->mtu = cm_id_priv->path_mtu;
2640         path->rate_selector = IB_SA_EQ;
2641         path->rate = cm_lap_get_packet_rate(lap_msg);
2642         path->packet_life_time_selector = IB_SA_EQ;
2643         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2644         path->packet_life_time -= (path->packet_life_time > 0);
2645 }
2646
2647 static int cm_lap_handler(struct cm_work *work)
2648 {
2649         struct cm_id_private *cm_id_priv;
2650         struct cm_lap_msg *lap_msg;
2651         struct ib_cm_lap_event_param *param;
2652         struct ib_mad_send_buf *msg = NULL;
2653         int ret;
2654
2655         /* todo: verify LAP request and send reject APR if invalid. */
2656         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2657         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2658                                    lap_msg->local_comm_id);
2659         if (!cm_id_priv)
2660                 return -EINVAL;
2661
2662         param = &work->cm_event.param.lap_rcvd;
2663         param->alternate_path = &work->path[0];
2664         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2665         work->cm_event.private_data = &lap_msg->private_data;
2666
2667         spin_lock_irq(&cm_id_priv->lock);
2668         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2669                 goto unlock;
2670
2671         switch (cm_id_priv->id.lap_state) {
2672         case IB_CM_LAP_UNINIT:
2673         case IB_CM_LAP_IDLE:
2674                 break;
2675         case IB_CM_MRA_LAP_SENT:
2676                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2677                                 counter[CM_LAP_COUNTER]);
2678                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2679                         goto unlock;
2680
2681                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2682                               CM_MSG_RESPONSE_OTHER,
2683                               cm_id_priv->service_timeout,
2684                               cm_id_priv->private_data,
2685                               cm_id_priv->private_data_len);
2686                 spin_unlock_irq(&cm_id_priv->lock);
2687
2688                 if (ib_post_send_mad(msg, NULL))
2689                         cm_free_msg(msg);
2690                 goto deref;
2691         case IB_CM_LAP_RCVD:
2692                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2693                                 counter[CM_LAP_COUNTER]);
2694                 goto unlock;
2695         default:
2696                 goto unlock;
2697         }
2698
2699         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2700         cm_id_priv->tid = lap_msg->hdr.tid;
2701         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2702                                 work->mad_recv_wc->recv_buf.grh,
2703                                 &cm_id_priv->av);
2704         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av);
2705         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2706         if (!ret)
2707                 list_add_tail(&work->list, &cm_id_priv->work_list);
2708         spin_unlock_irq(&cm_id_priv->lock);
2709
2710         if (ret)
2711                 cm_process_work(cm_id_priv, work);
2712         else
2713                 cm_deref_id(cm_id_priv);
2714         return 0;
2715
2716 unlock: spin_unlock_irq(&cm_id_priv->lock);
2717 deref:  cm_deref_id(cm_id_priv);
2718         return -EINVAL;
2719 }
2720
2721 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2722                           struct cm_id_private *cm_id_priv,
2723                           enum ib_cm_apr_status status,
2724                           void *info,
2725                           u8 info_length,
2726                           const void *private_data,
2727                           u8 private_data_len)
2728 {
2729         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2730         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2731         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2732         apr_msg->ap_status = (u8) status;
2733
2734         if (info && info_length) {
2735                 apr_msg->info_length = info_length;
2736                 memcpy(apr_msg->info, info, info_length);
2737         }
2738
2739         if (private_data && private_data_len)
2740                 memcpy(apr_msg->private_data, private_data, private_data_len);
2741 }
2742
2743 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2744                    enum ib_cm_apr_status status,
2745                    void *info,
2746                    u8 info_length,
2747                    const void *private_data,
2748                    u8 private_data_len)
2749 {
2750         struct cm_id_private *cm_id_priv;
2751         struct ib_mad_send_buf *msg;
2752         unsigned long flags;
2753         int ret;
2754
2755         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2756             (info && info_length > IB_CM_APR_INFO_LENGTH))
2757                 return -EINVAL;
2758
2759         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2760         spin_lock_irqsave(&cm_id_priv->lock, flags);
2761         if (cm_id->state != IB_CM_ESTABLISHED ||
2762             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2763              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2764                 ret = -EINVAL;
2765                 goto out;
2766         }
2767
2768         ret = cm_alloc_msg(cm_id_priv, &msg);
2769         if (ret)
2770                 goto out;
2771
2772         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2773                       info, info_length, private_data, private_data_len);
2774         ret = ib_post_send_mad(msg, NULL);
2775         if (ret) {
2776                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2777                 cm_free_msg(msg);
2778                 return ret;
2779         }
2780
2781         cm_id->lap_state = IB_CM_LAP_IDLE;
2782 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2783         return ret;
2784 }
2785 EXPORT_SYMBOL(ib_send_cm_apr);
2786
2787 static int cm_apr_handler(struct cm_work *work)
2788 {
2789         struct cm_id_private *cm_id_priv;
2790         struct cm_apr_msg *apr_msg;
2791         int ret;
2792
2793         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2794         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2795                                    apr_msg->local_comm_id);
2796         if (!cm_id_priv)
2797                 return -EINVAL; /* Unmatched reply. */
2798
2799         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2800         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2801         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2802         work->cm_event.private_data = &apr_msg->private_data;
2803
2804         spin_lock_irq(&cm_id_priv->lock);
2805         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2806             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2807              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2808                 spin_unlock_irq(&cm_id_priv->lock);
2809                 goto out;
2810         }
2811         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2812         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2813         cm_id_priv->msg = NULL;
2814
2815         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2816         if (!ret)
2817                 list_add_tail(&work->list, &cm_id_priv->work_list);
2818         spin_unlock_irq(&cm_id_priv->lock);
2819
2820         if (ret)
2821                 cm_process_work(cm_id_priv, work);
2822         else
2823                 cm_deref_id(cm_id_priv);
2824         return 0;
2825 out:
2826         cm_deref_id(cm_id_priv);
2827         return -EINVAL;
2828 }
2829
2830 static int cm_timewait_handler(struct cm_work *work)
2831 {
2832         struct cm_timewait_info *timewait_info;
2833         struct cm_id_private *cm_id_priv;
2834         int ret;
2835
2836         timewait_info = (struct cm_timewait_info *)work;
2837         spin_lock_irq(&cm.lock);
2838         list_del(&timewait_info->list);
2839         spin_unlock_irq(&cm.lock);
2840
2841         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2842                                    timewait_info->work.remote_id);
2843         if (!cm_id_priv)
2844                 return -EINVAL;
2845
2846         spin_lock_irq(&cm_id_priv->lock);
2847         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2848             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2849                 spin_unlock_irq(&cm_id_priv->lock);
2850                 goto out;
2851         }
2852         cm_id_priv->id.state = IB_CM_IDLE;
2853         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2854         if (!ret)
2855                 list_add_tail(&work->list, &cm_id_priv->work_list);
2856         spin_unlock_irq(&cm_id_priv->lock);
2857
2858         if (ret)
2859                 cm_process_work(cm_id_priv, work);
2860         else
2861                 cm_deref_id(cm_id_priv);
2862         return 0;
2863 out:
2864         cm_deref_id(cm_id_priv);
2865         return -EINVAL;
2866 }
2867
2868 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
2869                                struct cm_id_private *cm_id_priv,
2870                                struct ib_cm_sidr_req_param *param)
2871 {
2872         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
2873                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
2874         sidr_req_msg->request_id = cm_id_priv->id.local_id;
2875         sidr_req_msg->pkey = param->path->pkey;
2876         sidr_req_msg->service_id = param->service_id;
2877
2878         if (param->private_data && param->private_data_len)
2879                 memcpy(sidr_req_msg->private_data, param->private_data,
2880                        param->private_data_len);
2881 }
2882
2883 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
2884                         struct ib_cm_sidr_req_param *param)
2885 {
2886         struct cm_id_private *cm_id_priv;
2887         struct ib_mad_send_buf *msg;
2888         unsigned long flags;
2889         int ret;
2890
2891         if (!param->path || (param->private_data &&
2892              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
2893                 return -EINVAL;
2894
2895         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2896         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
2897         if (ret)
2898                 goto out;
2899
2900         cm_id->service_id = param->service_id;
2901         cm_id->service_mask = __constant_cpu_to_be64(~0ULL);
2902         cm_id_priv->timeout_ms = param->timeout_ms;
2903         cm_id_priv->max_cm_retries = param->max_cm_retries;
2904         ret = cm_alloc_msg(cm_id_priv, &msg);
2905         if (ret)
2906                 goto out;
2907
2908         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
2909                            param);
2910         msg->timeout_ms = cm_id_priv->timeout_ms;
2911         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
2912
2913         spin_lock_irqsave(&cm_id_priv->lock, flags);
2914         if (cm_id->state == IB_CM_IDLE)
2915                 ret = ib_post_send_mad(msg, NULL);
2916         else
2917                 ret = -EINVAL;
2918
2919         if (ret) {
2920                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2921                 cm_free_msg(msg);
2922                 goto out;
2923         }
2924         cm_id->state = IB_CM_SIDR_REQ_SENT;
2925         cm_id_priv->msg = msg;
2926         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2927 out:
2928         return ret;
2929 }
2930 EXPORT_SYMBOL(ib_send_cm_sidr_req);
2931
2932 static void cm_format_sidr_req_event(struct cm_work *work,
2933                                      struct ib_cm_id *listen_id)
2934 {
2935         struct cm_sidr_req_msg *sidr_req_msg;
2936         struct ib_cm_sidr_req_event_param *param;
2937
2938         sidr_req_msg = (struct cm_sidr_req_msg *)
2939                                 work->mad_recv_wc->recv_buf.mad;
2940         param = &work->cm_event.param.sidr_req_rcvd;
2941         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
2942         param->listen_id = listen_id;
2943         param->port = work->port->port_num;
2944         work->cm_event.private_data = &sidr_req_msg->private_data;
2945 }
2946
2947 static int cm_sidr_req_handler(struct cm_work *work)
2948 {
2949         struct ib_cm_id *cm_id;
2950         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
2951         struct cm_sidr_req_msg *sidr_req_msg;
2952         struct ib_wc *wc;
2953
2954         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
2955         if (IS_ERR(cm_id))
2956                 return PTR_ERR(cm_id);
2957         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2958
2959         /* Record SGID/SLID and request ID for lookup. */
2960         sidr_req_msg = (struct cm_sidr_req_msg *)
2961                                 work->mad_recv_wc->recv_buf.mad;
2962         wc = work->mad_recv_wc->wc;
2963         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
2964         cm_id_priv->av.dgid.global.interface_id = 0;
2965         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2966                                 work->mad_recv_wc->recv_buf.grh,
2967                                 &cm_id_priv->av);
2968         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
2969         cm_id_priv->tid = sidr_req_msg->hdr.tid;
2970         atomic_inc(&cm_id_priv->work_count);
2971
2972         spin_lock_irq(&cm.lock);
2973         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
2974         if (cur_cm_id_priv) {
2975                 spin_unlock_irq(&cm.lock);
2976                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2977                                 counter[CM_SIDR_REQ_COUNTER]);
2978                 goto out; /* Duplicate message. */
2979         }
2980         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
2981         cur_cm_id_priv = cm_find_listen(cm_id->device,
2982                                         sidr_req_msg->service_id,
2983                                         sidr_req_msg->private_data);
2984         if (!cur_cm_id_priv) {
2985                 spin_unlock_irq(&cm.lock);
2986                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
2987                 goto out; /* No match. */
2988         }
2989         atomic_inc(&cur_cm_id_priv->refcount);
2990         spin_unlock_irq(&cm.lock);
2991
2992         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
2993         cm_id_priv->id.context = cur_cm_id_priv->id.context;
2994         cm_id_priv->id.service_id = sidr_req_msg->service_id;
2995         cm_id_priv->id.service_mask = __constant_cpu_to_be64(~0ULL);
2996
2997         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
2998         cm_process_work(cm_id_priv, work);
2999         cm_deref_id(cur_cm_id_priv);
3000         return 0;
3001 out:
3002         ib_destroy_cm_id(&cm_id_priv->id);
3003         return -EINVAL;
3004 }
3005
3006 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3007                                struct cm_id_private *cm_id_priv,
3008                                struct ib_cm_sidr_rep_param *param)
3009 {
3010         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3011                           cm_id_priv->tid);
3012         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3013         sidr_rep_msg->status = param->status;
3014         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3015         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3016         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3017
3018         if (param->info && param->info_length)
3019                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3020
3021         if (param->private_data && param->private_data_len)
3022                 memcpy(sidr_rep_msg->private_data, param->private_data,
3023                        param->private_data_len);
3024 }
3025
3026 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3027                         struct ib_cm_sidr_rep_param *param)
3028 {
3029         struct cm_id_private *cm_id_priv;
3030         struct ib_mad_send_buf *msg;
3031         unsigned long flags;
3032         int ret;
3033
3034         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3035             (param->private_data &&
3036              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3037                 return -EINVAL;
3038
3039         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3040         spin_lock_irqsave(&cm_id_priv->lock, flags);
3041         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3042                 ret = -EINVAL;
3043                 goto error;
3044         }
3045
3046         ret = cm_alloc_msg(cm_id_priv, &msg);
3047         if (ret)
3048                 goto error;
3049
3050         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3051                            param);
3052         ret = ib_post_send_mad(msg, NULL);
3053         if (ret) {
3054                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3055                 cm_free_msg(msg);
3056                 return ret;
3057         }
3058         cm_id->state = IB_CM_IDLE;
3059         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3060
3061         spin_lock_irqsave(&cm.lock, flags);
3062         rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3063         spin_unlock_irqrestore(&cm.lock, flags);
3064         return 0;
3065
3066 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3067         return ret;
3068 }
3069 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3070
3071 static void cm_format_sidr_rep_event(struct cm_work *work)
3072 {
3073         struct cm_sidr_rep_msg *sidr_rep_msg;
3074         struct ib_cm_sidr_rep_event_param *param;
3075
3076         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3077                                 work->mad_recv_wc->recv_buf.mad;
3078         param = &work->cm_event.param.sidr_rep_rcvd;
3079         param->status = sidr_rep_msg->status;
3080         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3081         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3082         param->info = &sidr_rep_msg->info;
3083         param->info_len = sidr_rep_msg->info_length;
3084         work->cm_event.private_data = &sidr_rep_msg->private_data;
3085 }
3086
3087 static int cm_sidr_rep_handler(struct cm_work *work)
3088 {
3089         struct cm_sidr_rep_msg *sidr_rep_msg;
3090         struct cm_id_private *cm_id_priv;
3091
3092         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3093                                 work->mad_recv_wc->recv_buf.mad;
3094         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3095         if (!cm_id_priv)
3096                 return -EINVAL; /* Unmatched reply. */
3097
3098         spin_lock_irq(&cm_id_priv->lock);
3099         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3100                 spin_unlock_irq(&cm_id_priv->lock);
3101                 goto out;
3102         }
3103         cm_id_priv->id.state = IB_CM_IDLE;
3104         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3105         spin_unlock_irq(&cm_id_priv->lock);
3106
3107         cm_format_sidr_rep_event(work);
3108         cm_process_work(cm_id_priv, work);
3109         return 0;
3110 out:
3111         cm_deref_id(cm_id_priv);
3112         return -EINVAL;
3113 }
3114
3115 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3116                                   enum ib_wc_status wc_status)
3117 {
3118         struct cm_id_private *cm_id_priv;
3119         struct ib_cm_event cm_event;
3120         enum ib_cm_state state;
3121         int ret;
3122
3123         memset(&cm_event, 0, sizeof cm_event);
3124         cm_id_priv = msg->context[0];
3125
3126         /* Discard old sends or ones without a response. */
3127         spin_lock_irq(&cm_id_priv->lock);
3128         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3129         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3130                 goto discard;
3131
3132         switch (state) {
3133         case IB_CM_REQ_SENT:
3134         case IB_CM_MRA_REQ_RCVD:
3135                 cm_reset_to_idle(cm_id_priv);
3136                 cm_event.event = IB_CM_REQ_ERROR;
3137                 break;
3138         case IB_CM_REP_SENT:
3139         case IB_CM_MRA_REP_RCVD:
3140                 cm_reset_to_idle(cm_id_priv);
3141                 cm_event.event = IB_CM_REP_ERROR;
3142                 break;
3143         case IB_CM_DREQ_SENT:
3144                 cm_enter_timewait(cm_id_priv);
3145                 cm_event.event = IB_CM_DREQ_ERROR;
3146                 break;
3147         case IB_CM_SIDR_REQ_SENT:
3148                 cm_id_priv->id.state = IB_CM_IDLE;
3149                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3150                 break;
3151         default:
3152                 goto discard;
3153         }
3154         spin_unlock_irq(&cm_id_priv->lock);
3155         cm_event.param.send_status = wc_status;
3156
3157         /* No other events can occur on the cm_id at this point. */
3158         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3159         cm_free_msg(msg);
3160         if (ret)
3161                 ib_destroy_cm_id(&cm_id_priv->id);
3162         return;
3163 discard:
3164         spin_unlock_irq(&cm_id_priv->lock);
3165         cm_free_msg(msg);
3166 }
3167
3168 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3169                             struct ib_mad_send_wc *mad_send_wc)
3170 {
3171         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3172         struct cm_port *port;
3173         u16 attr_index;
3174
3175         port = mad_agent->context;
3176         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3177                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3178
3179         /*
3180          * If the send was in response to a received message (context[0] is not
3181          * set to a cm_id), and is not a REJ, then it is a send that was
3182          * manually retried.
3183          */
3184         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3185                 msg->retries = 1;
3186
3187         atomic_long_add(1 + msg->retries,
3188                         &port->counter_group[CM_XMIT].counter[attr_index]);
3189         if (msg->retries)
3190                 atomic_long_add(msg->retries,
3191                                 &port->counter_group[CM_XMIT_RETRIES].
3192                                 counter[attr_index]);
3193
3194         switch (mad_send_wc->status) {
3195         case IB_WC_SUCCESS:
3196         case IB_WC_WR_FLUSH_ERR:
3197                 cm_free_msg(msg);
3198                 break;
3199         default:
3200                 if (msg->context[0] && msg->context[1])
3201                         cm_process_send_error(msg, mad_send_wc->status);
3202                 else
3203                         cm_free_msg(msg);
3204                 break;
3205         }
3206 }
3207
3208 static void cm_work_handler(struct work_struct *_work)
3209 {
3210         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3211         int ret;
3212
3213         switch (work->cm_event.event) {
3214         case IB_CM_REQ_RECEIVED:
3215                 ret = cm_req_handler(work);
3216                 break;
3217         case IB_CM_MRA_RECEIVED:
3218                 ret = cm_mra_handler(work);
3219                 break;
3220         case IB_CM_REJ_RECEIVED:
3221                 ret = cm_rej_handler(work);
3222                 break;
3223         case IB_CM_REP_RECEIVED:
3224                 ret = cm_rep_handler(work);
3225                 break;
3226         case IB_CM_RTU_RECEIVED:
3227                 ret = cm_rtu_handler(work);
3228                 break;
3229         case IB_CM_USER_ESTABLISHED:
3230                 ret = cm_establish_handler(work);
3231                 break;
3232         case IB_CM_DREQ_RECEIVED:
3233                 ret = cm_dreq_handler(work);
3234                 break;
3235         case IB_CM_DREP_RECEIVED:
3236                 ret = cm_drep_handler(work);
3237                 break;
3238         case IB_CM_SIDR_REQ_RECEIVED:
3239                 ret = cm_sidr_req_handler(work);
3240                 break;
3241         case IB_CM_SIDR_REP_RECEIVED:
3242                 ret = cm_sidr_rep_handler(work);
3243                 break;
3244         case IB_CM_LAP_RECEIVED:
3245                 ret = cm_lap_handler(work);
3246                 break;
3247         case IB_CM_APR_RECEIVED:
3248                 ret = cm_apr_handler(work);
3249                 break;
3250         case IB_CM_TIMEWAIT_EXIT:
3251                 ret = cm_timewait_handler(work);
3252                 break;
3253         default:
3254                 ret = -EINVAL;
3255                 break;
3256         }
3257         if (ret)
3258                 cm_free_work(work);
3259 }
3260
3261 static int cm_establish(struct ib_cm_id *cm_id)
3262 {
3263         struct cm_id_private *cm_id_priv;
3264         struct cm_work *work;
3265         unsigned long flags;
3266         int ret = 0;
3267
3268         work = kmalloc(sizeof *work, GFP_ATOMIC);
3269         if (!work)
3270                 return -ENOMEM;
3271
3272         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3273         spin_lock_irqsave(&cm_id_priv->lock, flags);
3274         switch (cm_id->state)
3275         {
3276         case IB_CM_REP_SENT:
3277         case IB_CM_MRA_REP_RCVD:
3278                 cm_id->state = IB_CM_ESTABLISHED;
3279                 break;
3280         case IB_CM_ESTABLISHED:
3281                 ret = -EISCONN;
3282                 break;
3283         default:
3284                 ret = -EINVAL;
3285                 break;
3286         }
3287         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3288
3289         if (ret) {
3290                 kfree(work);
3291                 goto out;
3292         }
3293
3294         /*
3295          * The CM worker thread may try to destroy the cm_id before it
3296          * can execute this work item.  To prevent potential deadlock,
3297          * we need to find the cm_id once we're in the context of the
3298          * worker thread, rather than holding a reference on it.
3299          */
3300         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3301         work->local_id = cm_id->local_id;
3302         work->remote_id = cm_id->remote_id;
3303         work->mad_recv_wc = NULL;
3304         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3305         queue_delayed_work(cm.wq, &work->work, 0);
3306 out:
3307         return ret;
3308 }
3309
3310 static int cm_migrate(struct ib_cm_id *cm_id)
3311 {
3312         struct cm_id_private *cm_id_priv;
3313         unsigned long flags;
3314         int ret = 0;
3315
3316         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3317         spin_lock_irqsave(&cm_id_priv->lock, flags);
3318         if (cm_id->state == IB_CM_ESTABLISHED &&
3319             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3320              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3321                 cm_id->lap_state = IB_CM_LAP_IDLE;
3322                 cm_id_priv->av = cm_id_priv->alt_av;
3323         } else
3324                 ret = -EINVAL;
3325         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3326
3327         return ret;
3328 }
3329
3330 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3331 {
3332         int ret;
3333
3334         switch (event) {
3335         case IB_EVENT_COMM_EST:
3336                 ret = cm_establish(cm_id);
3337                 break;
3338         case IB_EVENT_PATH_MIG:
3339                 ret = cm_migrate(cm_id);
3340                 break;
3341         default:
3342                 ret = -EINVAL;
3343         }
3344         return ret;
3345 }
3346 EXPORT_SYMBOL(ib_cm_notify);
3347
3348 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3349                             struct ib_mad_recv_wc *mad_recv_wc)
3350 {
3351         struct cm_port *port = mad_agent->context;
3352         struct cm_work *work;
3353         enum ib_cm_event_type event;
3354         u16 attr_id;
3355         int paths = 0;
3356
3357         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3358         case CM_REQ_ATTR_ID:
3359                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3360                                                     alt_local_lid != 0);
3361                 event = IB_CM_REQ_RECEIVED;
3362                 break;
3363         case CM_MRA_ATTR_ID:
3364                 event = IB_CM_MRA_RECEIVED;
3365                 break;
3366         case CM_REJ_ATTR_ID:
3367                 event = IB_CM_REJ_RECEIVED;
3368                 break;
3369         case CM_REP_ATTR_ID:
3370                 event = IB_CM_REP_RECEIVED;
3371                 break;
3372         case CM_RTU_ATTR_ID:
3373                 event = IB_CM_RTU_RECEIVED;
3374                 break;
3375         case CM_DREQ_ATTR_ID:
3376                 event = IB_CM_DREQ_RECEIVED;
3377                 break;
3378         case CM_DREP_ATTR_ID:
3379                 event = IB_CM_DREP_RECEIVED;
3380                 break;
3381         case CM_SIDR_REQ_ATTR_ID:
3382                 event = IB_CM_SIDR_REQ_RECEIVED;
3383                 break;
3384         case CM_SIDR_REP_ATTR_ID:
3385                 event = IB_CM_SIDR_REP_RECEIVED;
3386                 break;
3387         case CM_LAP_ATTR_ID:
3388                 paths = 1;
3389                 event = IB_CM_LAP_RECEIVED;
3390                 break;
3391         case CM_APR_ATTR_ID:
3392                 event = IB_CM_APR_RECEIVED;
3393                 break;
3394         default:
3395                 ib_free_recv_mad(mad_recv_wc);
3396                 return;
3397         }
3398
3399         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3400         atomic_long_inc(&port->counter_group[CM_RECV].
3401                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3402
3403         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3404                        GFP_KERNEL);
3405         if (!work) {
3406                 ib_free_recv_mad(mad_recv_wc);
3407                 return;
3408         }
3409
3410         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3411         work->cm_event.event = event;
3412         work->mad_recv_wc = mad_recv_wc;
3413         work->port = port;
3414         queue_delayed_work(cm.wq, &work->work, 0);
3415 }
3416
3417 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3418                                 struct ib_qp_attr *qp_attr,
3419                                 int *qp_attr_mask)
3420 {
3421         unsigned long flags;
3422         int ret;
3423
3424         spin_lock_irqsave(&cm_id_priv->lock, flags);
3425         switch (cm_id_priv->id.state) {
3426         case IB_CM_REQ_SENT:
3427         case IB_CM_MRA_REQ_RCVD:
3428         case IB_CM_REQ_RCVD:
3429         case IB_CM_MRA_REQ_SENT:
3430         case IB_CM_REP_RCVD:
3431         case IB_CM_MRA_REP_SENT:
3432         case IB_CM_REP_SENT:
3433         case IB_CM_MRA_REP_RCVD:
3434         case IB_CM_ESTABLISHED:
3435                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3436                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3437                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3438                 if (cm_id_priv->responder_resources)
3439                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3440                                                     IB_ACCESS_REMOTE_ATOMIC;
3441                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3442                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3443                 ret = 0;
3444                 break;
3445         default:
3446                 ret = -EINVAL;
3447                 break;
3448         }
3449         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3450         return ret;
3451 }
3452
3453 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3454                                struct ib_qp_attr *qp_attr,
3455                                int *qp_attr_mask)
3456 {
3457         unsigned long flags;
3458         int ret;
3459
3460         spin_lock_irqsave(&cm_id_priv->lock, flags);
3461         switch (cm_id_priv->id.state) {
3462         case IB_CM_REQ_RCVD:
3463         case IB_CM_MRA_REQ_SENT:
3464         case IB_CM_REP_RCVD:
3465         case IB_CM_MRA_REP_SENT:
3466         case IB_CM_REP_SENT:
3467         case IB_CM_MRA_REP_RCVD:
3468         case IB_CM_ESTABLISHED:
3469                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3470                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3471                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3472                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3473                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3474                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3475                 if (cm_id_priv->qp_type == IB_QPT_RC) {
3476                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3477                                          IB_QP_MIN_RNR_TIMER;
3478                         qp_attr->max_dest_rd_atomic =
3479                                         cm_id_priv->responder_resources;
3480                         qp_attr->min_rnr_timer = 0;
3481                 }
3482                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3483                         *qp_attr_mask |= IB_QP_ALT_PATH;
3484                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3485                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3486                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3487                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3488                 }
3489                 ret = 0;
3490                 break;
3491         default:
3492                 ret = -EINVAL;
3493                 break;
3494         }
3495         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3496         return ret;
3497 }
3498
3499 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3500                                struct ib_qp_attr *qp_attr,
3501                                int *qp_attr_mask)
3502 {
3503         unsigned long flags;
3504         int ret;
3505
3506         spin_lock_irqsave(&cm_id_priv->lock, flags);
3507         switch (cm_id_priv->id.state) {
3508         /* Allow transition to RTS before sending REP */
3509         case IB_CM_REQ_RCVD:
3510         case IB_CM_MRA_REQ_SENT:
3511
3512         case IB_CM_REP_RCVD:
3513         case IB_CM_MRA_REP_SENT:
3514         case IB_CM_REP_SENT:
3515         case IB_CM_MRA_REP_RCVD:
3516         case IB_CM_ESTABLISHED:
3517                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3518                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3519                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3520                         if (cm_id_priv->qp_type == IB_QPT_RC) {
3521                                 *qp_attr_mask |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
3522                                                  IB_QP_RNR_RETRY |
3523                                                  IB_QP_MAX_QP_RD_ATOMIC;
3524                                 qp_attr->timeout = cm_id_priv->av.timeout;
3525                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3526                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3527                                 qp_attr->max_rd_atomic =
3528                                         cm_id_priv->initiator_depth;
3529                         }
3530                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3531                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3532                                 qp_attr->path_mig_state = IB_MIG_REARM;
3533                         }
3534                 } else {
3535                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3536                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3537                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3538                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3539                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3540                         qp_attr->path_mig_state = IB_MIG_REARM;
3541                 }
3542                 ret = 0;
3543                 break;
3544         default:
3545                 ret = -EINVAL;
3546                 break;
3547         }
3548         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3549         return ret;
3550 }
3551
3552 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3553                        struct ib_qp_attr *qp_attr,
3554                        int *qp_attr_mask)
3555 {
3556         struct cm_id_private *cm_id_priv;
3557         int ret;
3558
3559         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3560         switch (qp_attr->qp_state) {
3561         case IB_QPS_INIT:
3562                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3563                 break;
3564         case IB_QPS_RTR:
3565                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3566                 break;
3567         case IB_QPS_RTS:
3568                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3569                 break;
3570         default:
3571                 ret = -EINVAL;
3572                 break;
3573         }
3574         return ret;
3575 }
3576 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3577
3578 static void cm_get_ack_delay(struct cm_device *cm_dev)
3579 {
3580         struct ib_device_attr attr;
3581
3582         if (ib_query_device(cm_dev->ib_device, &attr))
3583                 cm_dev->ack_delay = 0; /* acks will rely on packet life time */
3584         else
3585                 cm_dev->ack_delay = attr.local_ca_ack_delay;
3586 }
3587
3588 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3589                                char *buf)
3590 {
3591         struct cm_counter_group *group;
3592         struct cm_counter_attribute *cm_attr;
3593
3594         group = container_of(obj, struct cm_counter_group, obj);
3595         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3596
3597         return sprintf(buf, "%ld\n",
3598                        atomic_long_read(&group->counter[cm_attr->index]));
3599 }
3600
3601 static struct sysfs_ops cm_counter_ops = {
3602         .show = cm_show_counter
3603 };
3604
3605 static struct kobj_type cm_counter_obj_type = {
3606         .sysfs_ops = &cm_counter_ops,
3607         .default_attrs = cm_counter_default_attrs
3608 };
3609
3610 static void cm_release_port_obj(struct kobject *obj)
3611 {
3612         struct cm_port *cm_port;
3613
3614         cm_port = container_of(obj, struct cm_port, port_obj);
3615         kfree(cm_port);
3616 }
3617
3618 static struct kobj_type cm_port_obj_type = {
3619         .release = cm_release_port_obj
3620 };
3621
3622 struct class cm_class = {
3623         .name    = "infiniband_cm",
3624 };
3625 EXPORT_SYMBOL(cm_class);
3626
3627 static int cm_create_port_fs(struct cm_port *port)
3628 {
3629         int i, ret;
3630
3631         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3632                                    &port->cm_dev->device->kobj,
3633                                    "%d", port->port_num);
3634         if (ret) {
3635                 kfree(port);
3636                 return ret;
3637         }
3638
3639         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3640                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3641                                            &cm_counter_obj_type,
3642                                            &port->port_obj,
3643                                            "%s", counter_group_names[i]);
3644                 if (ret)
3645                         goto error;
3646         }
3647
3648         return 0;
3649
3650 error:
3651         while (i--)
3652                 kobject_put(&port->counter_group[i].obj);
3653         kobject_put(&port->port_obj);
3654         return ret;
3655
3656 }
3657
3658 static void cm_remove_port_fs(struct cm_port *port)
3659 {
3660         int i;
3661
3662         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3663                 kobject_put(&port->counter_group[i].obj);
3664
3665         kobject_put(&port->port_obj);
3666 }
3667
3668 static void cm_add_one(struct ib_device *ib_device)
3669 {
3670         struct cm_device *cm_dev;
3671         struct cm_port *port;
3672         struct ib_mad_reg_req reg_req = {
3673                 .mgmt_class = IB_MGMT_CLASS_CM,
3674                 .mgmt_class_version = IB_CM_CLASS_VERSION
3675         };
3676         struct ib_port_modify port_modify = {
3677                 .set_port_cap_mask = IB_PORT_CM_SUP
3678         };
3679         unsigned long flags;
3680         int ret;
3681         u8 i;
3682
3683         if (rdma_node_get_transport(ib_device->node_type) != RDMA_TRANSPORT_IB)
3684                 return;
3685
3686         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3687                          ib_device->phys_port_cnt, GFP_KERNEL);
3688         if (!cm_dev)
3689                 return;
3690
3691         cm_dev->ib_device = ib_device;
3692         cm_get_ack_delay(cm_dev);
3693
3694         cm_dev->device = device_create_drvdata(&cm_class, &ib_device->dev,
3695                                                MKDEV(0, 0), NULL,
3696                                                "%s", ib_device->name);
3697         if (!cm_dev->device) {
3698                 kfree(cm_dev);
3699                 return;
3700         }
3701
3702         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3703         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3704                 port = kzalloc(sizeof *port, GFP_KERNEL);
3705                 if (!port)
3706                         goto error1;
3707
3708                 cm_dev->port[i-1] = port;
3709                 port->cm_dev = cm_dev;
3710                 port->port_num = i;
3711
3712                 ret = cm_create_port_fs(port);
3713                 if (ret)
3714                         goto error1;
3715
3716                 port->mad_agent = ib_register_mad_agent(ib_device, i,
3717                                                         IB_QPT_GSI,
3718                                                         &reg_req,
3719                                                         0,
3720                                                         cm_send_handler,
3721                                                         cm_recv_handler,
3722                                                         port);
3723                 if (IS_ERR(port->mad_agent))
3724                         goto error2;
3725
3726                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
3727                 if (ret)
3728                         goto error3;
3729         }
3730         ib_set_client_data(ib_device, &cm_client, cm_dev);
3731
3732         write_lock_irqsave(&cm.device_lock, flags);
3733         list_add_tail(&cm_dev->list, &cm.device_list);
3734         write_unlock_irqrestore(&cm.device_lock, flags);
3735         return;
3736
3737 error3:
3738         ib_unregister_mad_agent(port->mad_agent);
3739 error2:
3740         cm_remove_port_fs(port);
3741 error1:
3742         port_modify.set_port_cap_mask = 0;
3743         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3744         while (--i) {
3745                 port = cm_dev->port[i-1];
3746                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3747                 ib_unregister_mad_agent(port->mad_agent);
3748                 cm_remove_port_fs(port);
3749         }
3750         device_unregister(cm_dev->device);
3751 }
3752
3753 static void cm_remove_one(struct ib_device *ib_device)
3754 {
3755         struct cm_device *cm_dev;
3756         struct cm_port *port;
3757         struct ib_port_modify port_modify = {
3758                 .clr_port_cap_mask = IB_PORT_CM_SUP
3759         };
3760         unsigned long flags;
3761         int i;
3762
3763         cm_dev = ib_get_client_data(ib_device, &cm_client);
3764         if (!cm_dev)
3765                 return;
3766
3767         write_lock_irqsave(&cm.device_lock, flags);
3768         list_del(&cm_dev->list);
3769         write_unlock_irqrestore(&cm.device_lock, flags);
3770
3771         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3772                 port = cm_dev->port[i-1];
3773                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3774                 ib_unregister_mad_agent(port->mad_agent);
3775                 flush_workqueue(cm.wq);
3776                 cm_remove_port_fs(port);
3777         }
3778         device_unregister(cm_dev->device);
3779 }
3780
3781 static int __init ib_cm_init(void)
3782 {
3783         int ret;
3784
3785         memset(&cm, 0, sizeof cm);
3786         INIT_LIST_HEAD(&cm.device_list);
3787         rwlock_init(&cm.device_lock);
3788         spin_lock_init(&cm.lock);
3789         cm.listen_service_table = RB_ROOT;
3790         cm.listen_service_id = __constant_be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3791         cm.remote_id_table = RB_ROOT;
3792         cm.remote_qp_table = RB_ROOT;
3793         cm.remote_sidr_table = RB_ROOT;
3794         idr_init(&cm.local_id_table);
3795         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
3796         idr_pre_get(&cm.local_id_table, GFP_KERNEL);
3797         INIT_LIST_HEAD(&cm.timewait_list);
3798
3799         ret = class_register(&cm_class);
3800         if (ret)
3801                 return -ENOMEM;
3802
3803         cm.wq = create_workqueue("ib_cm");
3804         if (!cm.wq) {
3805                 ret = -ENOMEM;
3806                 goto error1;
3807         }
3808
3809         ret = ib_register_client(&cm_client);
3810         if (ret)
3811                 goto error2;
3812
3813         return 0;
3814 error2:
3815         destroy_workqueue(cm.wq);
3816 error1:
3817         class_unregister(&cm_class);
3818         return ret;
3819 }
3820
3821 static void __exit ib_cm_cleanup(void)
3822 {
3823         struct cm_timewait_info *timewait_info, *tmp;
3824
3825         spin_lock_irq(&cm.lock);
3826         list_for_each_entry(timewait_info, &cm.timewait_list, list)
3827                 cancel_delayed_work(&timewait_info->work.work);
3828         spin_unlock_irq(&cm.lock);
3829
3830         ib_unregister_client(&cm_client);
3831         destroy_workqueue(cm.wq);
3832
3833         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
3834                 list_del(&timewait_info->list);
3835                 kfree(timewait_info);
3836         }
3837
3838         class_unregister(&cm_class);
3839         idr_destroy(&cm.local_id_table);
3840 }
3841
3842 module_init(ib_cm_init);
3843 module_exit(ib_cm_cleanup);
3844