drivers/infiniband/core/cm.c

   1 /*
   2  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/dma-mapping.h>
  38 #include <linux/device.h>
  39 #include <linux/err.h>
  40 #include <linux/idr.h>
  41 #include <linux/interrupt.h>
  42 #include <linux/random.h>
  43 #include <linux/rbtree.h>
  44 #include <linux/spinlock.h>
  45 #include <linux/slab.h>
  46 #include <linux/sysfs.h>
  47 #include <linux/workqueue.h>
  48 #include <linux/kdev_t.h>
  49
  50 #include <rdma/ib_cache.h>
  51 #include <rdma/ib_cm.h>
  52 #include "cm_msgs.h"
  53
  54 MODULE_AUTHOR("Sean Hefty");
  55 MODULE_DESCRIPTION("InfiniBand CM");
  56 MODULE_LICENSE("Dual BSD/GPL");
  57
  58 static void cm_add_one(struct ib_device *device);
  59 static void cm_remove_one(struct ib_device *device);
  60
  61 static struct ib_client cm_client = {
  62         .name   = "cm",
  63         .add    = cm_add_one,
  64         .remove = cm_remove_one
  65 };
  66
  67 static struct ib_cm {
  68         spinlock_t lock;
  69         struct list_head device_list;
  70         rwlock_t device_lock;
  71         struct rb_root listen_service_table;
  72         u64 listen_service_id;
  73         /* struct rb_root peer_service_table; todo: fix peer to peer */
  74         struct rb_root remote_qp_table;
  75         struct rb_root remote_id_table;
  76         struct rb_root remote_sidr_table;
  77         struct idr local_id_table;
  78         __be32 random_id_operand;
  79         struct list_head timewait_list;
  80         struct workqueue_struct *wq;
  81 } cm;
  82
  83 /* Counter indexes ordered by attribute ID */
  84 enum {
  85         CM_REQ_COUNTER,
  86         CM_MRA_COUNTER,
  87         CM_REJ_COUNTER,
  88         CM_REP_COUNTER,
  89         CM_RTU_COUNTER,
  90         CM_DREQ_COUNTER,
  91         CM_DREP_COUNTER,
  92         CM_SIDR_REQ_COUNTER,
  93         CM_SIDR_REP_COUNTER,
  94         CM_LAP_COUNTER,
  95         CM_APR_COUNTER,
  96         CM_ATTR_COUNT,
  97         CM_ATTR_ID_OFFSET = 0x0010,
  98 };
  99
 100 enum {
 101         CM_XMIT,
 102         CM_XMIT_RETRIES,
 103         CM_RECV,
 104         CM_RECV_DUPLICATES,
 105         CM_COUNTER_GROUPS
 106 };
 107
 108 static char const counter_group_names[CM_COUNTER_GROUPS]
 109                                      [sizeof("cm_rx_duplicates")] = {
 110         "cm_tx_msgs", "cm_tx_retries",
 111         "cm_rx_msgs", "cm_rx_duplicates"
 112 };
 113
 114 struct cm_counter_group {
 115         struct kobject obj;
 116         atomic_long_t counter[CM_ATTR_COUNT];
 117 };
 118
 119 struct cm_counter_attribute {
 120         struct attribute attr;
 121         int index;
 122 };
 123
 124 #define CM_COUNTER_ATTR(_name, _index) \
 125 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 126         .attr = { .name = __stringify(_name), .mode = 0444 }, \
 127         .index = _index \
 128 }
 129
 130 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 131 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 132 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 133 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 134 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 135 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 136 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 137 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 138 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 139 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 140 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 141
 142 static struct attribute *cm_counter_default_attrs[] = {
 143         &cm_req_counter_attr.attr,
 144         &cm_mra_counter_attr.attr,
 145         &cm_rej_counter_attr.attr,
 146         &cm_rep_counter_attr.attr,
 147         &cm_rtu_counter_attr.attr,
 148         &cm_dreq_counter_attr.attr,
 149         &cm_drep_counter_attr.attr,
 150         &cm_sidr_req_counter_attr.attr,
 151         &cm_sidr_rep_counter_attr.attr,
 152         &cm_lap_counter_attr.attr,
 153         &cm_apr_counter_attr.attr,
 154         NULL
 155 };
 156
 157 struct cm_port {
 158         struct cm_device *cm_dev;
 159         struct ib_mad_agent *mad_agent;
 160         struct kobject port_obj;
 161         u8 port_num;
 162         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 163 };
 164
 165 struct cm_device {
 166         struct list_head list;
 167         struct ib_device *ib_device;
 168         struct device *device;
 169         u8 ack_delay;
 170         struct cm_port *port[0];
 171 };
 172
 173 struct cm_av {
 174         struct cm_port *port;
 175         union ib_gid dgid;
 176         struct ib_ah_attr ah_attr;
 177         u16 pkey_index;
 178         u8 timeout;
 179 };
 180
 181 struct cm_work {
 182         struct delayed_work work;
 183         struct list_head list;
 184         struct cm_port *port;
 185         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 186         __be32 local_id;                        /* Established / timewait */
 187         __be32 remote_id;
 188         struct ib_cm_event cm_event;
 189         struct ib_sa_path_rec path[0];
 190 };
 191
 192 struct cm_timewait_info {
 193         struct cm_work work;                    /* Must be first. */
 194         struct list_head list;
 195         struct rb_node remote_qp_node;
 196         struct rb_node remote_id_node;
 197         __be64 remote_ca_guid;
 198         __be32 remote_qpn;
 199         u8 inserted_remote_qp;
 200         u8 inserted_remote_id;
 201 };
 202
 203 struct cm_id_private {
 204         struct ib_cm_id id;
 205
 206         struct rb_node service_node;
 207         struct rb_node sidr_id_node;
 208         spinlock_t lock;        /* Do not acquire inside cm.lock */
 209         struct completion comp;
 210         atomic_t refcount;
 211
 212         struct ib_mad_send_buf *msg;
 213         struct cm_timewait_info *timewait_info;
 214         /* todo: use alternate port on send failure */
 215         struct cm_av av;
 216         struct cm_av alt_av;
 217         struct ib_cm_compare_data *compare_data;
 218
 219         void *private_data;
 220         __be64 tid;
 221         __be32 local_qpn;
 222         __be32 remote_qpn;
 223         enum ib_qp_type qp_type;
 224         __be32 sq_psn;
 225         __be32 rq_psn;
 226         int timeout_ms;
 227         enum ib_mtu path_mtu;
 228         __be16 pkey;
 229         u8 private_data_len;
 230         u8 max_cm_retries;
 231         u8 peer_to_peer;
 232         u8 responder_resources;
 233         u8 initiator_depth;
 234         u8 retry_count;
 235         u8 rnr_retry_count;
 236         u8 service_timeout;
 237         u8 target_ack_delay;
 238
 239         struct list_head work_list;
 240         atomic_t work_count;
 241 };
 242
 243 static void cm_work_handler(struct work_struct *work);
 244
 245 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 246 {
 247         if (atomic_dec_and_test(&cm_id_priv->refcount))
 248                 complete(&cm_id_priv->comp);
 249 }
 250
 251 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 252                         struct ib_mad_send_buf **msg)
 253 {
 254         struct ib_mad_agent *mad_agent;
 255         struct ib_mad_send_buf *m;
 256         struct ib_ah *ah;
 257
 258         mad_agent = cm_id_priv->av.port->mad_agent;
 259         ah = ib_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr);
 260         if (IS_ERR(ah))
 261                 return PTR_ERR(ah);
 262
 263         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 264                                cm_id_priv->av.pkey_index,
 265                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 266                                GFP_ATOMIC);
 267         if (IS_ERR(m)) {
 268                 ib_destroy_ah(ah);
 269                 return PTR_ERR(m);
 270         }
 271
 272         /* Timeout set by caller if response is expected. */
 273         m->ah = ah;
 274         m->retries = cm_id_priv->max_cm_retries;
 275
 276         atomic_inc(&cm_id_priv->refcount);
 277         m->context[0] = cm_id_priv;
 278         *msg = m;
 279         return 0;
 280 }
 281
 282 static int cm_alloc_response_msg(struct cm_port *port,
 283                                  struct ib_mad_recv_wc *mad_recv_wc,
 284                                  struct ib_mad_send_buf **msg)
 285 {
 286         struct ib_mad_send_buf *m;
 287         struct ib_ah *ah;
 288
 289         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 290                                   mad_recv_wc->recv_buf.grh, port->port_num);
 291         if (IS_ERR(ah))
 292                 return PTR_ERR(ah);
 293
 294         m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 295                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 296                                GFP_ATOMIC);
 297         if (IS_ERR(m)) {
 298                 ib_destroy_ah(ah);
 299                 return PTR_ERR(m);
 300         }
 301         m->ah = ah;
 302         *msg = m;
 303         return 0;
 304 }
 305
 306 static void cm_free_msg(struct ib_mad_send_buf *msg)
 307 {
 308         ib_destroy_ah(msg->ah);
 309         if (msg->context[0])
 310                 cm_deref_id(msg->context[0]);
 311         ib_free_send_mad(msg);
 312 }
 313
 314 static void * cm_copy_private_data(const void *private_data,
 315                                    u8 private_data_len)
 316 {
 317         void *data;
 318
 319         if (!private_data || !private_data_len)
 320                 return NULL;
 321
 322         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 323         if (!data)
 324                 return ERR_PTR(-ENOMEM);
 325
 326         return data;
 327 }
 328
 329 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 330                                  void *private_data, u8 private_data_len)
 331 {
 332         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 333                 kfree(cm_id_priv->private_data);
 334
 335         cm_id_priv->private_data = private_data;
 336         cm_id_priv->private_data_len = private_data_len;
 337 }
 338
 339 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 340                                     struct ib_grh *grh, struct cm_av *av)
 341 {
 342         av->port = port;
 343         av->pkey_index = wc->pkey_index;
 344         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 345                            grh, &av->ah_attr);
 346 }
 347
 348 static int cm_init_av_by_path(struct ib_sa_path_rec *path, struct cm_av *av)
 349 {
 350         struct cm_device *cm_dev;
 351         struct cm_port *port = NULL;
 352         unsigned long flags;
 353         int ret;
 354         u8 p;
 355
 356         read_lock_irqsave(&cm.device_lock, flags);
 357         list_for_each_entry(cm_dev, &cm.device_list, list) {
 358                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 359                                         &p, NULL)) {
 360                         port = cm_dev->port[p-1];
 361                         break;
 362                 }
 363         }
 364         read_unlock_irqrestore(&cm.device_lock, flags);
 365
 366         if (!port)
 367                 return -EINVAL;
 368
 369         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 370                                   be16_to_cpu(path->pkey), &av->pkey_index);
 371         if (ret)
 372                 return ret;
 373
 374         av->port = port;
 375         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 376                              &av->ah_attr);
 377         av->timeout = path->packet_life_time + 1;
 378         return 0;
 379 }
 380
 381 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 382 {
 383         unsigned long flags;
 384         int ret, id;
 385         static int next_id;
 386
 387         do {
 388                 spin_lock_irqsave(&cm.lock, flags);
 389                 ret = idr_get_new_above(&cm.local_id_table, cm_id_priv,
 390                                         next_id, &id);
 391                 if (!ret)
 392                         next_id = ((unsigned) id + 1) & MAX_ID_MASK;
 393                 spin_unlock_irqrestore(&cm.lock, flags);
 394         } while( (ret == -EAGAIN) && idr_pre_get(&cm.local_id_table, GFP_KERNEL) );
 395
 396         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 397         return ret;
 398 }
 399
 400 static void cm_free_id(__be32 local_id)
 401 {
 402         spin_lock_irq(&cm.lock);
 403         idr_remove(&cm.local_id_table,
 404                    (__force int) (local_id ^ cm.random_id_operand));
 405         spin_unlock_irq(&cm.lock);
 406 }
 407
 408 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 409 {
 410         struct cm_id_private *cm_id_priv;
 411
 412         cm_id_priv = idr_find(&cm.local_id_table,
 413                               (__force int) (local_id ^ cm.random_id_operand));
 414         if (cm_id_priv) {
 415                 if (cm_id_priv->id.remote_id == remote_id)
 416                         atomic_inc(&cm_id_priv->refcount);
 417                 else
 418                         cm_id_priv = NULL;
 419         }
 420
 421         return cm_id_priv;
 422 }
 423
 424 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 425 {
 426         struct cm_id_private *cm_id_priv;
 427
 428         spin_lock_irq(&cm.lock);
 429         cm_id_priv = cm_get_id(local_id, remote_id);
 430         spin_unlock_irq(&cm.lock);
 431
 432         return cm_id_priv;
 433 }
 434
 435 static void cm_mask_copy(u8 *dst, u8 *src, u8 *mask)
 436 {
 437         int i;
 438
 439         for (i = 0; i < IB_CM_COMPARE_SIZE / sizeof(unsigned long); i++)
 440                 ((unsigned long *) dst)[i] = ((unsigned long *) src)[i] &
 441                                              ((unsigned long *) mask)[i];
 442 }
 443
 444 static int cm_compare_data(struct ib_cm_compare_data *src_data,
 445                            struct ib_cm_compare_data *dst_data)
 446 {
 447         u8 src[IB_CM_COMPARE_SIZE];
 448         u8 dst[IB_CM_COMPARE_SIZE];
 449
 450         if (!src_data || !dst_data)
 451                 return 0;
 452
 453         cm_mask_copy(src, src_data->data, dst_data->mask);
 454         cm_mask_copy(dst, dst_data->data, src_data->mask);
 455         return memcmp(src, dst, IB_CM_COMPARE_SIZE);
 456 }
 457
 458 static int cm_compare_private_data(u8 *private_data,
 459                                    struct ib_cm_compare_data *dst_data)
 460 {
 461         u8 src[IB_CM_COMPARE_SIZE];
 462
 463         if (!dst_data)
 464                 return 0;
 465
 466         cm_mask_copy(src, private_data, dst_data->mask);
 467         return memcmp(src, dst_data->data, IB_CM_COMPARE_SIZE);
 468 }
 469
 470 /*
 471  * Trivial helpers to strip endian annotation and compare; the
 472  * endianness doesn't actually matter since we just need a stable
 473  * order for the RB tree.
 474  */
 475 static int be32_lt(__be32 a, __be32 b)
 476 {
 477         return (__force u32) a < (__force u32) b;
 478 }
 479
 480 static int be32_gt(__be32 a, __be32 b)
 481 {
 482         return (__force u32) a > (__force u32) b;
 483 }
 484
 485 static int be64_lt(__be64 a, __be64 b)
 486 {
 487         return (__force u64) a < (__force u64) b;
 488 }
 489
 490 static int be64_gt(__be64 a, __be64 b)
 491 {
 492         return (__force u64) a > (__force u64) b;
 493 }
 494
 495 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 496 {
 497         struct rb_node **link = &cm.listen_service_table.rb_node;
 498         struct rb_node *parent = NULL;
 499         struct cm_id_private *cur_cm_id_priv;
 500         __be64 service_id = cm_id_priv->id.service_id;
 501         __be64 service_mask = cm_id_priv->id.service_mask;
 502         int data_cmp;
 503
 504         while (*link) {
 505                 parent = *link;
 506                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 507                                           service_node);
 508                 data_cmp = cm_compare_data(cm_id_priv->compare_data,
 509                                            cur_cm_id_priv->compare_data);
 510                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 511                     (service_mask & cur_cm_id_priv->id.service_id) &&
 512                     (cm_id_priv->id.device == cur_cm_id_priv->id.device) &&
 513                     !data_cmp)
 514                         return cur_cm_id_priv;
 515
 516                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 517                         link = &(*link)->rb_left;
 518                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 519                         link = &(*link)->rb_right;
 520                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 521                         link = &(*link)->rb_left;
 522                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 523                         link = &(*link)->rb_right;
 524                 else if (data_cmp < 0)
 525                         link = &(*link)->rb_left;
 526                 else
 527                         link = &(*link)->rb_right;
 528         }
 529         rb_link_node(&cm_id_priv->service_node, parent, link);
 530         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 531         return NULL;
 532 }
 533
 534 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 535                                              __be64 service_id,
 536                                              u8 *private_data)
 537 {
 538         struct rb_node *node = cm.listen_service_table.rb_node;
 539         struct cm_id_private *cm_id_priv;
 540         int data_cmp;
 541
 542         while (node) {
 543                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 544                 data_cmp = cm_compare_private_data(private_data,
 545                                                    cm_id_priv->compare_data);
 546                 if ((cm_id_priv->id.service_mask & service_id) ==
 547                      cm_id_priv->id.service_id &&
 548                     (cm_id_priv->id.device == device) && !data_cmp)
 549                         return cm_id_priv;
 550
 551                 if (device < cm_id_priv->id.device)
 552                         node = node->rb_left;
 553                 else if (device > cm_id_priv->id.device)
 554                         node = node->rb_right;
 555                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 556                         node = node->rb_left;
 557                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 558                         node = node->rb_right;
 559                 else if (data_cmp < 0)
 560                         node = node->rb_left;
 561                 else
 562                         node = node->rb_right;
 563         }
 564         return NULL;
 565 }
 566
 567 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 568                                                      *timewait_info)
 569 {
 570         struct rb_node **link = &cm.remote_id_table.rb_node;
 571         struct rb_node *parent = NULL;
 572         struct cm_timewait_info *cur_timewait_info;
 573         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 574         __be32 remote_id = timewait_info->work.remote_id;
 575
 576         while (*link) {
 577                 parent = *link;
 578                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 579                                              remote_id_node);
 580                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 581                         link = &(*link)->rb_left;
 582                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 583                         link = &(*link)->rb_right;
 584                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 585                         link = &(*link)->rb_left;
 586                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 587                         link = &(*link)->rb_right;
 588                 else
 589                         return cur_timewait_info;
 590         }
 591         timewait_info->inserted_remote_id = 1;
 592         rb_link_node(&timewait_info->remote_id_node, parent, link);
 593         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 594         return NULL;
 595 }
 596
 597 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 598                                                    __be32 remote_id)
 599 {
 600         struct rb_node *node = cm.remote_id_table.rb_node;
 601         struct cm_timewait_info *timewait_info;
 602
 603         while (node) {
 604                 timewait_info = rb_entry(node, struct cm_timewait_info,
 605                                          remote_id_node);
 606                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 607                         node = node->rb_left;
 608                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 609                         node = node->rb_right;
 610                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 611                         node = node->rb_left;
 612                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 613                         node = node->rb_right;
 614                 else
 615                         return timewait_info;
 616         }
 617         return NULL;
 618 }
 619
 620 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 621                                                       *timewait_info)
 622 {
 623         struct rb_node **link = &cm.remote_qp_table.rb_node;
 624         struct rb_node *parent = NULL;
 625         struct cm_timewait_info *cur_timewait_info;
 626         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 627         __be32 remote_qpn = timewait_info->remote_qpn;
 628
 629         while (*link) {
 630                 parent = *link;
 631                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 632                                              remote_qp_node);
 633                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 634                         link = &(*link)->rb_left;
 635                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 636                         link = &(*link)->rb_right;
 637                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 638                         link = &(*link)->rb_left;
 639                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 640                         link = &(*link)->rb_right;
 641                 else
 642                         return cur_timewait_info;
 643         }
 644         timewait_info->inserted_remote_qp = 1;
 645         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 646         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 647         return NULL;
 648 }
 649
 650 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 651                                                     *cm_id_priv)
 652 {
 653         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 654         struct rb_node *parent = NULL;
 655         struct cm_id_private *cur_cm_id_priv;
 656         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 657         __be32 remote_id = cm_id_priv->id.remote_id;
 658
 659         while (*link) {
 660                 parent = *link;
 661                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 662                                           sidr_id_node);
 663                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 664                         link = &(*link)->rb_left;
 665                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 666                         link = &(*link)->rb_right;
 667                 else {
 668                         int cmp;
 669                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 670                                      sizeof *port_gid);
 671                         if (cmp < 0)
 672                                 link = &(*link)->rb_left;
 673                         else if (cmp > 0)
 674                                 link = &(*link)->rb_right;
 675                         else
 676                                 return cur_cm_id_priv;
 677                 }
 678         }
 679         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 680         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 681         return NULL;
 682 }
 683
 684 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 685                                enum ib_cm_sidr_status status)
 686 {
 687         struct ib_cm_sidr_rep_param param;
 688
 689         memset(&param, 0, sizeof param);
 690         param.status = status;
 691         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 692 }
 693
 694 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 695                                  ib_cm_handler cm_handler,
 696                                  void *context)
 697 {
 698         struct cm_id_private *cm_id_priv;
 699         int ret;
 700
 701         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 702         if (!cm_id_priv)
 703                 return ERR_PTR(-ENOMEM);
 704
 705         cm_id_priv->id.state = IB_CM_IDLE;
 706         cm_id_priv->id.device = device;
 707         cm_id_priv->id.cm_handler = cm_handler;
 708         cm_id_priv->id.context = context;
 709         cm_id_priv->id.remote_cm_qpn = 1;
 710         ret = cm_alloc_id(cm_id_priv);
 711         if (ret)
 712                 goto error;
 713
 714         spin_lock_init(&cm_id_priv->lock);
 715         init_completion(&cm_id_priv->comp);
 716         INIT_LIST_HEAD(&cm_id_priv->work_list);
 717         atomic_set(&cm_id_priv->work_count, -1);
 718         atomic_set(&cm_id_priv->refcount, 1);
 719         return &cm_id_priv->id;
 720
 721 error:
 722         kfree(cm_id_priv);
 723         return ERR_PTR(-ENOMEM);
 724 }
 725 EXPORT_SYMBOL(ib_create_cm_id);
 726
 727 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 728 {
 729         struct cm_work *work;
 730
 731         if (list_empty(&cm_id_priv->work_list))
 732                 return NULL;
 733
 734         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 735         list_del(&work->list);
 736         return work;
 737 }
 738
 739 static void cm_free_work(struct cm_work *work)
 740 {
 741         if (work->mad_recv_wc)
 742                 ib_free_recv_mad(work->mad_recv_wc);
 743         kfree(work);
 744 }
 745
 746 static inline int cm_convert_to_ms(int iba_time)
 747 {
 748         /* approximate conversion to ms from 4.096us x 2^iba_time */
 749         return 1 << max(iba_time - 8, 0);
 750 }
 751
 752 /*
 753  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 754  * Because of how ack_timeout is stored, adding one doubles the timeout.
 755  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 756  * increment it (round up) only if the other is within 50%.
 757  */
 758 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 759 {
 760         int ack_timeout = packet_life_time + 1;
 761
 762         if (ack_timeout >= ca_ack_delay)
 763                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 764         else
 765                 ack_timeout = ca_ack_delay +
 766                               (ack_timeout >= (ca_ack_delay - 1));
 767
 768         return min(31, ack_timeout);
 769 }
 770
 771 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 772 {
 773         if (timewait_info->inserted_remote_id) {
 774                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 775                 timewait_info->inserted_remote_id = 0;
 776         }
 777
 778         if (timewait_info->inserted_remote_qp) {
 779                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 780                 timewait_info->inserted_remote_qp = 0;
 781         }
 782 }
 783
 784 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 785 {
 786         struct cm_timewait_info *timewait_info;
 787
 788         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 789         if (!timewait_info)
 790                 return ERR_PTR(-ENOMEM);
 791
 792         timewait_info->work.local_id = local_id;
 793         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 794         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 795         return timewait_info;
 796 }
 797
 798 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 799 {
 800         int wait_time;
 801         unsigned long flags;
 802
 803         spin_lock_irqsave(&cm.lock, flags);
 804         cm_cleanup_timewait(cm_id_priv->timewait_info);
 805         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 806         spin_unlock_irqrestore(&cm.lock, flags);
 807
 808         /*
 809          * The cm_id could be destroyed by the user before we exit timewait.
 810          * To protect against this, we search for the cm_id after exiting
 811          * timewait before notifying the user that we've exited timewait.
 812          */
 813         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 814         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 815         queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 816                            msecs_to_jiffies(wait_time));
 817         cm_id_priv->timewait_info = NULL;
 818 }
 819
 820 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 821 {
 822         unsigned long flags;
 823
 824         cm_id_priv->id.state = IB_CM_IDLE;
 825         if (cm_id_priv->timewait_info) {
 826                 spin_lock_irqsave(&cm.lock, flags);
 827                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 828                 spin_unlock_irqrestore(&cm.lock, flags);
 829                 kfree(cm_id_priv->timewait_info);
 830                 cm_id_priv->timewait_info = NULL;
 831         }
 832 }
 833
 834 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 835 {
 836         struct cm_id_private *cm_id_priv;
 837         struct cm_work *work;
 838
 839         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 840 retest:
 841         spin_lock_irq(&cm_id_priv->lock);
 842         switch (cm_id->state) {
 843         case IB_CM_LISTEN:
 844                 cm_id->state = IB_CM_IDLE;
 845                 spin_unlock_irq(&cm_id_priv->lock);
 846                 spin_lock_irq(&cm.lock);
 847                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 848                 spin_unlock_irq(&cm.lock);
 849                 break;
 850         case IB_CM_SIDR_REQ_SENT:
 851                 cm_id->state = IB_CM_IDLE;
 852                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 853                 spin_unlock_irq(&cm_id_priv->lock);
 854                 break;
 855         case IB_CM_SIDR_REQ_RCVD:
 856                 spin_unlock_irq(&cm_id_priv->lock);
 857                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 858                 break;
 859         case IB_CM_REQ_SENT:
 860                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 861                 spin_unlock_irq(&cm_id_priv->lock);
 862                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 863                                &cm_id_priv->id.device->node_guid,
 864                                sizeof cm_id_priv->id.device->node_guid,
 865                                NULL, 0);
 866                 break;
 867         case IB_CM_REQ_RCVD:
 868                 if (err == -ENOMEM) {
 869                         /* Do not reject to allow future retries. */
 870                         cm_reset_to_idle(cm_id_priv);
 871                         spin_unlock_irq(&cm_id_priv->lock);
 872                 } else {
 873                         spin_unlock_irq(&cm_id_priv->lock);
 874                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 875                                        NULL, 0, NULL, 0);
 876                 }
 877                 break;
 878         case IB_CM_MRA_REQ_RCVD:
 879         case IB_CM_REP_SENT:
 880         case IB_CM_MRA_REP_RCVD:
 881                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 882                 /* Fall through */
 883         case IB_CM_MRA_REQ_SENT:
 884         case IB_CM_REP_RCVD:
 885         case IB_CM_MRA_REP_SENT:
 886                 spin_unlock_irq(&cm_id_priv->lock);
 887                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 888                                NULL, 0, NULL, 0);
 889                 break;
 890         case IB_CM_ESTABLISHED:
 891                 spin_unlock_irq(&cm_id_priv->lock);
 892                 ib_send_cm_dreq(cm_id, NULL, 0);
 893                 goto retest;
 894         case IB_CM_DREQ_SENT:
 895                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 896                 cm_enter_timewait(cm_id_priv);
 897                 spin_unlock_irq(&cm_id_priv->lock);
 898                 break;
 899         case IB_CM_DREQ_RCVD:
 900                 spin_unlock_irq(&cm_id_priv->lock);
 901                 ib_send_cm_drep(cm_id, NULL, 0);
 902                 break;
 903         default:
 904                 spin_unlock_irq(&cm_id_priv->lock);
 905                 break;
 906         }
 907
 908         cm_free_id(cm_id->local_id);
 909         cm_deref_id(cm_id_priv);
 910         wait_for_completion(&cm_id_priv->comp);
 911         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
 912                 cm_free_work(work);
 913         kfree(cm_id_priv->compare_data);
 914         kfree(cm_id_priv->private_data);
 915         kfree(cm_id_priv);
 916 }
 917
 918 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
 919 {
 920         cm_destroy_id(cm_id, 0);
 921 }
 922 EXPORT_SYMBOL(ib_destroy_cm_id);
 923
 924 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask,
 925                  struct ib_cm_compare_data *compare_data)
 926 {
 927         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
 928         unsigned long flags;
 929         int ret = 0;
 930
 931         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
 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 = ~cpu_to_be64(0);
 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 = ~cpu_to_be64(0);
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 = ~cpu_to_be64(0);
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                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2413                         cm_state = cm_id->state;
2414                         lap_state = IB_CM_MRA_LAP_SENT;
2415                         msg_response = CM_MSG_RESPONSE_OTHER;
2416                         break;
2417                 }
2418         default:
2419                 ret = -EINVAL;
2420                 goto error1;
2421         }
2422
2423         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2424                 ret = cm_alloc_msg(cm_id_priv, &msg);
2425                 if (ret)
2426                         goto error1;
2427
2428                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2429                               msg_response, service_timeout,
2430                               private_data, private_data_len);
2431                 ret = ib_post_send_mad(msg, NULL);
2432                 if (ret)
2433                         goto error2;
2434         }
2435
2436         cm_id->state = cm_state;
2437         cm_id->lap_state = lap_state;
2438         cm_id_priv->service_timeout = service_timeout;
2439         cm_set_private_data(cm_id_priv, data, private_data_len);
2440         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2441         return 0;
2442
2443 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2444         kfree(data);
2445         return ret;
2446
2447 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2448         kfree(data);
2449         cm_free_msg(msg);
2450         return ret;
2451 }
2452 EXPORT_SYMBOL(ib_send_cm_mra);
2453
2454 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2455 {
2456         switch (cm_mra_get_msg_mraed(mra_msg)) {
2457         case CM_MSG_RESPONSE_REQ:
2458                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2459         case CM_MSG_RESPONSE_REP:
2460         case CM_MSG_RESPONSE_OTHER:
2461                 return cm_acquire_id(mra_msg->remote_comm_id,
2462                                      mra_msg->local_comm_id);
2463         default:
2464                 return NULL;
2465         }
2466 }
2467
2468 static int cm_mra_handler(struct cm_work *work)
2469 {
2470         struct cm_id_private *cm_id_priv;
2471         struct cm_mra_msg *mra_msg;
2472         int timeout, ret;
2473
2474         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2475         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2476         if (!cm_id_priv)
2477                 return -EINVAL;
2478
2479         work->cm_event.private_data = &mra_msg->private_data;
2480         work->cm_event.param.mra_rcvd.service_timeout =
2481                                         cm_mra_get_service_timeout(mra_msg);
2482         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2483                   cm_convert_to_ms(cm_id_priv->av.timeout);
2484
2485         spin_lock_irq(&cm_id_priv->lock);
2486         switch (cm_id_priv->id.state) {
2487         case IB_CM_REQ_SENT:
2488                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2489                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2490                                   cm_id_priv->msg, timeout))
2491                         goto out;
2492                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2493                 break;
2494         case IB_CM_REP_SENT:
2495                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2496                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2497                                   cm_id_priv->msg, timeout))
2498                         goto out;
2499                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2500                 break;
2501         case IB_CM_ESTABLISHED:
2502                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2503                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2504                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2505                                   cm_id_priv->msg, timeout)) {
2506                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2507                                 atomic_long_inc(&work->port->
2508                                                 counter_group[CM_RECV_DUPLICATES].
2509                                                 counter[CM_MRA_COUNTER]);
2510                         goto out;
2511                 }
2512                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2513                 break;
2514         case IB_CM_MRA_REQ_RCVD:
2515         case IB_CM_MRA_REP_RCVD:
2516                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2517                                 counter[CM_MRA_COUNTER]);
2518                 /* fall through */
2519         default:
2520                 goto out;
2521         }
2522
2523         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2524                                       cm_id_priv->id.state;
2525         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2526         if (!ret)
2527                 list_add_tail(&work->list, &cm_id_priv->work_list);
2528         spin_unlock_irq(&cm_id_priv->lock);
2529
2530         if (ret)
2531                 cm_process_work(cm_id_priv, work);
2532         else
2533                 cm_deref_id(cm_id_priv);
2534         return 0;
2535 out:
2536         spin_unlock_irq(&cm_id_priv->lock);
2537         cm_deref_id(cm_id_priv);
2538         return -EINVAL;
2539 }
2540
2541 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2542                           struct cm_id_private *cm_id_priv,
2543                           struct ib_sa_path_rec *alternate_path,
2544                           const void *private_data,
2545                           u8 private_data_len)
2546 {
2547         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2548                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2549         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2550         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2551         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2552         /* todo: need remote CM response timeout */
2553         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2554         lap_msg->alt_local_lid = alternate_path->slid;
2555         lap_msg->alt_remote_lid = alternate_path->dlid;
2556         lap_msg->alt_local_gid = alternate_path->sgid;
2557         lap_msg->alt_remote_gid = alternate_path->dgid;
2558         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2559         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2560         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2561         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2562         cm_lap_set_sl(lap_msg, alternate_path->sl);
2563         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2564         cm_lap_set_local_ack_timeout(lap_msg,
2565                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2566                                alternate_path->packet_life_time));
2567
2568         if (private_data && private_data_len)
2569                 memcpy(lap_msg->private_data, private_data, private_data_len);
2570 }
2571
2572 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2573                    struct ib_sa_path_rec *alternate_path,
2574                    const void *private_data,
2575                    u8 private_data_len)
2576 {
2577         struct cm_id_private *cm_id_priv;
2578         struct ib_mad_send_buf *msg;
2579         unsigned long flags;
2580         int ret;
2581
2582         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2583                 return -EINVAL;
2584
2585         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2586         spin_lock_irqsave(&cm_id_priv->lock, flags);
2587         if (cm_id->state != IB_CM_ESTABLISHED ||
2588             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2589              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2590                 ret = -EINVAL;
2591                 goto out;
2592         }
2593
2594         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av);
2595         if (ret)
2596                 goto out;
2597         cm_id_priv->alt_av.timeout =
2598                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2599                                        cm_id_priv->alt_av.timeout - 1);
2600
2601         ret = cm_alloc_msg(cm_id_priv, &msg);
2602         if (ret)
2603                 goto out;
2604
2605         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
2606                       alternate_path, private_data, private_data_len);
2607         msg->timeout_ms = cm_id_priv->timeout_ms;
2608         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
2609
2610         ret = ib_post_send_mad(msg, NULL);
2611         if (ret) {
2612                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2613                 cm_free_msg(msg);
2614                 return ret;
2615         }
2616
2617         cm_id->lap_state = IB_CM_LAP_SENT;
2618         cm_id_priv->msg = msg;
2619
2620 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2621         return ret;
2622 }
2623 EXPORT_SYMBOL(ib_send_cm_lap);
2624
2625 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
2626                                     struct ib_sa_path_rec *path,
2627                                     struct cm_lap_msg *lap_msg)
2628 {
2629         memset(path, 0, sizeof *path);
2630         path->dgid = lap_msg->alt_local_gid;
2631         path->sgid = lap_msg->alt_remote_gid;
2632         path->dlid = lap_msg->alt_local_lid;
2633         path->slid = lap_msg->alt_remote_lid;
2634         path->flow_label = cm_lap_get_flow_label(lap_msg);
2635         path->hop_limit = lap_msg->alt_hop_limit;
2636         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
2637         path->reversible = 1;
2638         path->pkey = cm_id_priv->pkey;
2639         path->sl = cm_lap_get_sl(lap_msg);
2640         path->mtu_selector = IB_SA_EQ;
2641         path->mtu = cm_id_priv->path_mtu;
2642         path->rate_selector = IB_SA_EQ;
2643         path->rate = cm_lap_get_packet_rate(lap_msg);
2644         path->packet_life_time_selector = IB_SA_EQ;
2645         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
2646         path->packet_life_time -= (path->packet_life_time > 0);
2647 }
2648
2649 static int cm_lap_handler(struct cm_work *work)
2650 {
2651         struct cm_id_private *cm_id_priv;
2652         struct cm_lap_msg *lap_msg;
2653         struct ib_cm_lap_event_param *param;
2654         struct ib_mad_send_buf *msg = NULL;
2655         int ret;
2656
2657         /* todo: verify LAP request and send reject APR if invalid. */
2658         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
2659         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
2660                                    lap_msg->local_comm_id);
2661         if (!cm_id_priv)
2662                 return -EINVAL;
2663
2664         param = &work->cm_event.param.lap_rcvd;
2665         param->alternate_path = &work->path[0];
2666         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
2667         work->cm_event.private_data = &lap_msg->private_data;
2668
2669         spin_lock_irq(&cm_id_priv->lock);
2670         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
2671                 goto unlock;
2672
2673         switch (cm_id_priv->id.lap_state) {
2674         case IB_CM_LAP_UNINIT:
2675         case IB_CM_LAP_IDLE:
2676                 break;
2677         case IB_CM_MRA_LAP_SENT:
2678                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2679                                 counter[CM_LAP_COUNTER]);
2680                 if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
2681                         goto unlock;
2682
2683                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2684                               CM_MSG_RESPONSE_OTHER,
2685                               cm_id_priv->service_timeout,
2686                               cm_id_priv->private_data,
2687                               cm_id_priv->private_data_len);
2688                 spin_unlock_irq(&cm_id_priv->lock);
2689
2690                 if (ib_post_send_mad(msg, NULL))
2691                         cm_free_msg(msg);
2692                 goto deref;
2693         case IB_CM_LAP_RCVD:
2694                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2695                                 counter[CM_LAP_COUNTER]);
2696                 goto unlock;
2697         default:
2698                 goto unlock;
2699         }
2700
2701         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
2702         cm_id_priv->tid = lap_msg->hdr.tid;
2703         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2704                                 work->mad_recv_wc->recv_buf.grh,
2705                                 &cm_id_priv->av);
2706         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av);
2707         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2708         if (!ret)
2709                 list_add_tail(&work->list, &cm_id_priv->work_list);
2710         spin_unlock_irq(&cm_id_priv->lock);
2711
2712         if (ret)
2713                 cm_process_work(cm_id_priv, work);
2714         else
2715                 cm_deref_id(cm_id_priv);
2716         return 0;
2717
2718 unlock: spin_unlock_irq(&cm_id_priv->lock);
2719 deref:  cm_deref_id(cm_id_priv);
2720         return -EINVAL;
2721 }
2722
2723 static void cm_format_apr(struct cm_apr_msg *apr_msg,
2724                           struct cm_id_private *cm_id_priv,
2725                           enum ib_cm_apr_status status,
2726                           void *info,
2727                           u8 info_length,
2728                           const void *private_data,
2729                           u8 private_data_len)
2730 {
2731         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
2732         apr_msg->local_comm_id = cm_id_priv->id.local_id;
2733         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
2734         apr_msg->ap_status = (u8) status;
2735
2736         if (info && info_length) {
2737                 apr_msg->info_length = info_length;
2738                 memcpy(apr_msg->info, info, info_length);
2739         }
2740
2741         if (private_data && private_data_len)
2742                 memcpy(apr_msg->private_data, private_data, private_data_len);
2743 }
2744
2745 int ib_send_cm_apr(struct ib_cm_id *cm_id,
2746                    enum ib_cm_apr_status status,
2747                    void *info,
2748                    u8 info_length,
2749                    const void *private_data,
2750                    u8 private_data_len)
2751 {
2752         struct cm_id_private *cm_id_priv;
2753         struct ib_mad_send_buf *msg;
2754         unsigned long flags;
2755         int ret;
2756
2757         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
2758             (info && info_length > IB_CM_APR_INFO_LENGTH))
2759                 return -EINVAL;
2760
2761         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2762         spin_lock_irqsave(&cm_id_priv->lock, flags);
2763         if (cm_id->state != IB_CM_ESTABLISHED ||
2764             (cm_id->lap_state != IB_CM_LAP_RCVD &&
2765              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
2766                 ret = -EINVAL;
2767                 goto out;
2768         }
2769
2770         ret = cm_alloc_msg(cm_id_priv, &msg);
2771         if (ret)
2772                 goto out;
2773
2774         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
2775                       info, info_length, private_data, private_data_len);
2776         ret = ib_post_send_mad(msg, NULL);
2777         if (ret) {
2778                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2779                 cm_free_msg(msg);
2780                 return ret;
2781         }
2782
2783         cm_id->lap_state = IB_CM_LAP_IDLE;
2784 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2785         return ret;
2786 }
2787 EXPORT_SYMBOL(ib_send_cm_apr);
2788
2789 static int cm_apr_handler(struct cm_work *work)
2790 {
2791         struct cm_id_private *cm_id_priv;
2792         struct cm_apr_msg *apr_msg;
2793         int ret;
2794
2795         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
2796         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
2797                                    apr_msg->local_comm_id);
2798         if (!cm_id_priv)
2799                 return -EINVAL; /* Unmatched reply. */
2800
2801         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
2802         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
2803         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
2804         work->cm_event.private_data = &apr_msg->private_data;
2805
2806         spin_lock_irq(&cm_id_priv->lock);
2807         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
2808             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
2809              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
2810                 spin_unlock_irq(&cm_id_priv->lock);
2811                 goto out;
2812         }
2813         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
2814         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2815         cm_id_priv->msg = NULL;
2816
2817         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2818         if (!ret)
2819                 list_add_tail(&work->list, &cm_id_priv->work_list);
2820         spin_unlock_irq(&cm_id_priv->lock);
2821
2822         if (ret)
2823                 cm_process_work(cm_id_priv, work);
2824         else
2825                 cm_deref_id(cm_id_priv);
2826         return 0;
2827 out:
2828         cm_deref_id(cm_id_priv);
2829         return -EINVAL;
2830 }
2831
2832 static int cm_timewait_handler(struct cm_work *work)
2833 {
2834         struct cm_timewait_info *timewait_info;
2835         struct cm_id_private *cm_id_priv;
2836         int ret;
2837
2838         timewait_info = (struct cm_timewait_info *)work;
2839         spin_lock_irq(&cm.lock);
2840         list_del(&timewait_info->list);
2841         spin_unlock_irq(&cm.lock);
2842
2843         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2844                                    timewait_info->work.remote_id);
2845         if (!cm_id_priv)
2846                 return -EINVAL;
2847
2848         spin_lock_irq(&cm_id_priv->lock);
2849         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
2850             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
2851                 spin_unlock_irq(&cm_id_priv->lock);
2852                 goto out;
2853         }
2854         cm_id_priv->id.state = IB_CM_IDLE;
2855         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2856         if (!ret)
2857                 list_add_tail(&work->list, &cm_id_priv->work_list);
2858         spin_unlock_irq(&cm_id_priv->lock);
2859
2860         if (ret)
2861                 cm_process_work(cm_id_priv, work);
2862         else
2863                 cm_deref_id(cm_id_priv);
2864         return 0;
2865 out:
2866         cm_deref_id(cm_id_priv);
2867         return -EINVAL;
2868 }
2869
2870 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
2871                                struct cm_id_private *cm_id_priv,
2872                                struct ib_cm_sidr_req_param *param)
2873 {
2874         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
2875                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
2876         sidr_req_msg->request_id = cm_id_priv->id.local_id;
2877         sidr_req_msg->pkey = param->path->pkey;
2878         sidr_req_msg->service_id = param->service_id;
2879
2880         if (param->private_data && param->private_data_len)
2881                 memcpy(sidr_req_msg->private_data, param->private_data,
2882                        param->private_data_len);
2883 }
2884
2885 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
2886                         struct ib_cm_sidr_req_param *param)
2887 {
2888         struct cm_id_private *cm_id_priv;
2889         struct ib_mad_send_buf *msg;
2890         unsigned long flags;
2891         int ret;
2892
2893         if (!param->path || (param->private_data &&
2894              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
2895                 return -EINVAL;
2896
2897         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2898         ret = cm_init_av_by_path(param->path, &cm_id_priv->av);
2899         if (ret)
2900                 goto out;
2901
2902         cm_id->service_id = param->service_id;
2903         cm_id->service_mask = ~cpu_to_be64(0);
2904         cm_id_priv->timeout_ms = param->timeout_ms;
2905         cm_id_priv->max_cm_retries = param->max_cm_retries;
2906         ret = cm_alloc_msg(cm_id_priv, &msg);
2907         if (ret)
2908                 goto out;
2909
2910         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
2911                            param);
2912         msg->timeout_ms = cm_id_priv->timeout_ms;
2913         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
2914
2915         spin_lock_irqsave(&cm_id_priv->lock, flags);
2916         if (cm_id->state == IB_CM_IDLE)
2917                 ret = ib_post_send_mad(msg, NULL);
2918         else
2919                 ret = -EINVAL;
2920
2921         if (ret) {
2922                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2923                 cm_free_msg(msg);
2924                 goto out;
2925         }
2926         cm_id->state = IB_CM_SIDR_REQ_SENT;
2927         cm_id_priv->msg = msg;
2928         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2929 out:
2930         return ret;
2931 }
2932 EXPORT_SYMBOL(ib_send_cm_sidr_req);
2933
2934 static void cm_format_sidr_req_event(struct cm_work *work,
2935                                      struct ib_cm_id *listen_id)
2936 {
2937         struct cm_sidr_req_msg *sidr_req_msg;
2938         struct ib_cm_sidr_req_event_param *param;
2939
2940         sidr_req_msg = (struct cm_sidr_req_msg *)
2941                                 work->mad_recv_wc->recv_buf.mad;
2942         param = &work->cm_event.param.sidr_req_rcvd;
2943         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
2944         param->listen_id = listen_id;
2945         param->port = work->port->port_num;
2946         work->cm_event.private_data = &sidr_req_msg->private_data;
2947 }
2948
2949 static int cm_sidr_req_handler(struct cm_work *work)
2950 {
2951         struct ib_cm_id *cm_id;
2952         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
2953         struct cm_sidr_req_msg *sidr_req_msg;
2954         struct ib_wc *wc;
2955
2956         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
2957         if (IS_ERR(cm_id))
2958                 return PTR_ERR(cm_id);
2959         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2960
2961         /* Record SGID/SLID and request ID for lookup. */
2962         sidr_req_msg = (struct cm_sidr_req_msg *)
2963                                 work->mad_recv_wc->recv_buf.mad;
2964         wc = work->mad_recv_wc->wc;
2965         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
2966         cm_id_priv->av.dgid.global.interface_id = 0;
2967         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2968                                 work->mad_recv_wc->recv_buf.grh,
2969                                 &cm_id_priv->av);
2970         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
2971         cm_id_priv->tid = sidr_req_msg->hdr.tid;
2972         atomic_inc(&cm_id_priv->work_count);
2973
2974         spin_lock_irq(&cm.lock);
2975         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
2976         if (cur_cm_id_priv) {
2977                 spin_unlock_irq(&cm.lock);
2978                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2979                                 counter[CM_SIDR_REQ_COUNTER]);
2980                 goto out; /* Duplicate message. */
2981         }
2982         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
2983         cur_cm_id_priv = cm_find_listen(cm_id->device,
2984                                         sidr_req_msg->service_id,
2985                                         sidr_req_msg->private_data);
2986         if (!cur_cm_id_priv) {
2987                 spin_unlock_irq(&cm.lock);
2988                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
2989                 goto out; /* No match. */
2990         }
2991         atomic_inc(&cur_cm_id_priv->refcount);
2992         spin_unlock_irq(&cm.lock);
2993
2994         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
2995         cm_id_priv->id.context = cur_cm_id_priv->id.context;
2996         cm_id_priv->id.service_id = sidr_req_msg->service_id;
2997         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
2998
2999         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3000         cm_process_work(cm_id_priv, work);
3001         cm_deref_id(cur_cm_id_priv);
3002         return 0;
3003 out:
3004         ib_destroy_cm_id(&cm_id_priv->id);
3005         return -EINVAL;
3006 }
3007
3008 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3009                                struct cm_id_private *cm_id_priv,
3010                                struct ib_cm_sidr_rep_param *param)
3011 {
3012         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3013                           cm_id_priv->tid);
3014         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3015         sidr_rep_msg->status = param->status;
3016         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3017         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3018         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3019
3020         if (param->info && param->info_length)
3021                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3022
3023         if (param->private_data && param->private_data_len)
3024                 memcpy(sidr_rep_msg->private_data, param->private_data,
3025                        param->private_data_len);
3026 }
3027
3028 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3029                         struct ib_cm_sidr_rep_param *param)
3030 {
3031         struct cm_id_private *cm_id_priv;
3032         struct ib_mad_send_buf *msg;
3033         unsigned long flags;
3034         int ret;
3035
3036         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3037             (param->private_data &&
3038              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3039                 return -EINVAL;
3040
3041         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3042         spin_lock_irqsave(&cm_id_priv->lock, flags);
3043         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3044                 ret = -EINVAL;
3045                 goto error;
3046         }
3047
3048         ret = cm_alloc_msg(cm_id_priv, &msg);
3049         if (ret)
3050                 goto error;
3051
3052         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3053                            param);
3054         ret = ib_post_send_mad(msg, NULL);
3055         if (ret) {
3056                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3057                 cm_free_msg(msg);
3058                 return ret;
3059         }
3060         cm_id->state = IB_CM_IDLE;
3061         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3062
3063         spin_lock_irqsave(&cm.lock, flags);
3064         rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3065         spin_unlock_irqrestore(&cm.lock, flags);
3066         return 0;
3067
3068 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3069         return ret;
3070 }
3071 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3072
3073 static void cm_format_sidr_rep_event(struct cm_work *work)
3074 {
3075         struct cm_sidr_rep_msg *sidr_rep_msg;
3076         struct ib_cm_sidr_rep_event_param *param;
3077
3078         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3079                                 work->mad_recv_wc->recv_buf.mad;
3080         param = &work->cm_event.param.sidr_rep_rcvd;
3081         param->status = sidr_rep_msg->status;
3082         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3083         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3084         param->info = &sidr_rep_msg->info;
3085         param->info_len = sidr_rep_msg->info_length;
3086         work->cm_event.private_data = &sidr_rep_msg->private_data;
3087 }
3088
3089 static int cm_sidr_rep_handler(struct cm_work *work)
3090 {
3091         struct cm_sidr_rep_msg *sidr_rep_msg;
3092         struct cm_id_private *cm_id_priv;
3093
3094         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3095                                 work->mad_recv_wc->recv_buf.mad;
3096         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3097         if (!cm_id_priv)
3098                 return -EINVAL; /* Unmatched reply. */
3099
3100         spin_lock_irq(&cm_id_priv->lock);
3101         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3102                 spin_unlock_irq(&cm_id_priv->lock);
3103                 goto out;
3104         }
3105         cm_id_priv->id.state = IB_CM_IDLE;
3106         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3107         spin_unlock_irq(&cm_id_priv->lock);
3108
3109         cm_format_sidr_rep_event(work);
3110         cm_process_work(cm_id_priv, work);
3111         return 0;
3112 out:
3113         cm_deref_id(cm_id_priv);
3114         return -EINVAL;
3115 }
3116
3117 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3118                                   enum ib_wc_status wc_status)
3119 {
3120         struct cm_id_private *cm_id_priv;
3121         struct ib_cm_event cm_event;
3122         enum ib_cm_state state;
3123         int ret;
3124
3125         memset(&cm_event, 0, sizeof cm_event);
3126         cm_id_priv = msg->context[0];
3127
3128         /* Discard old sends or ones without a response. */
3129         spin_lock_irq(&cm_id_priv->lock);
3130         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3131         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3132                 goto discard;
3133
3134         switch (state) {
3135         case IB_CM_REQ_SENT:
3136         case IB_CM_MRA_REQ_RCVD:
3137                 cm_reset_to_idle(cm_id_priv);
3138                 cm_event.event = IB_CM_REQ_ERROR;
3139                 break;
3140         case IB_CM_REP_SENT:
3141         case IB_CM_MRA_REP_RCVD:
3142                 cm_reset_to_idle(cm_id_priv);
3143                 cm_event.event = IB_CM_REP_ERROR;
3144                 break;
3145         case IB_CM_DREQ_SENT:
3146                 cm_enter_timewait(cm_id_priv);
3147                 cm_event.event = IB_CM_DREQ_ERROR;
3148                 break;
3149         case IB_CM_SIDR_REQ_SENT:
3150                 cm_id_priv->id.state = IB_CM_IDLE;
3151                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3152                 break;
3153         default:
3154                 goto discard;
3155         }
3156         spin_unlock_irq(&cm_id_priv->lock);
3157         cm_event.param.send_status = wc_status;
3158
3159         /* No other events can occur on the cm_id at this point. */
3160         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3161         cm_free_msg(msg);
3162         if (ret)
3163                 ib_destroy_cm_id(&cm_id_priv->id);
3164         return;
3165 discard:
3166         spin_unlock_irq(&cm_id_priv->lock);
3167         cm_free_msg(msg);
3168 }
3169
3170 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3171                             struct ib_mad_send_wc *mad_send_wc)
3172 {
3173         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3174         struct cm_port *port;
3175         u16 attr_index;
3176
3177         port = mad_agent->context;
3178         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3179                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3180
3181         /*
3182          * If the send was in response to a received message (context[0] is not
3183          * set to a cm_id), and is not a REJ, then it is a send that was
3184          * manually retried.
3185          */
3186         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3187                 msg->retries = 1;
3188
3189         atomic_long_add(1 + msg->retries,
3190                         &port->counter_group[CM_XMIT].counter[attr_index]);
3191         if (msg->retries)
3192                 atomic_long_add(msg->retries,
3193                                 &port->counter_group[CM_XMIT_RETRIES].
3194                                 counter[attr_index]);
3195
3196         switch (mad_send_wc->status) {
3197         case IB_WC_SUCCESS:
3198         case IB_WC_WR_FLUSH_ERR:
3199                 cm_free_msg(msg);
3200                 break;
3201         default:
3202                 if (msg->context[0] && msg->context[1])
3203                         cm_process_send_error(msg, mad_send_wc->status);
3204                 else
3205                         cm_free_msg(msg);
3206                 break;
3207         }
3208 }
3209
3210 static void cm_work_handler(struct work_struct *_work)
3211 {
3212         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3213         int ret;
3214
3215         switch (work->cm_event.event) {
3216         case IB_CM_REQ_RECEIVED:
3217                 ret = cm_req_handler(work);
3218                 break;
3219         case IB_CM_MRA_RECEIVED:
3220                 ret = cm_mra_handler(work);
3221                 break;
3222         case IB_CM_REJ_RECEIVED:
3223                 ret = cm_rej_handler(work);
3224                 break;
3225         case IB_CM_REP_RECEIVED:
3226                 ret = cm_rep_handler(work);
3227                 break;
3228         case IB_CM_RTU_RECEIVED:
3229                 ret = cm_rtu_handler(work);
3230                 break;
3231         case IB_CM_USER_ESTABLISHED:
3232                 ret = cm_establish_handler(work);
3233                 break;
3234         case IB_CM_DREQ_RECEIVED:
3235                 ret = cm_dreq_handler(work);
3236                 break;
3237         case IB_CM_DREP_RECEIVED:
3238                 ret = cm_drep_handler(work);
3239                 break;
3240         case IB_CM_SIDR_REQ_RECEIVED:
3241                 ret = cm_sidr_req_handler(work);
3242                 break;
3243         case IB_CM_SIDR_REP_RECEIVED:
3244                 ret = cm_sidr_rep_handler(work);
3245                 break;
3246         case IB_CM_LAP_RECEIVED:
3247                 ret = cm_lap_handler(work);
3248                 break;
3249         case IB_CM_APR_RECEIVED:
3250                 ret = cm_apr_handler(work);
3251                 break;
3252         case IB_CM_TIMEWAIT_EXIT:
3253                 ret = cm_timewait_handler(work);
3254                 break;
3255         default:
3256                 ret = -EINVAL;
3257                 break;
3258         }
3259         if (ret)
3260                 cm_free_work(work);
3261 }
3262
3263 static int cm_establish(struct ib_cm_id *cm_id)
3264 {
3265         struct cm_id_private *cm_id_priv;
3266         struct cm_work *work;
3267         unsigned long flags;
3268         int ret = 0;
3269
3270         work = kmalloc(sizeof *work, GFP_ATOMIC);
3271         if (!work)
3272                 return -ENOMEM;
3273
3274         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3275         spin_lock_irqsave(&cm_id_priv->lock, flags);
3276         switch (cm_id->state)
3277         {
3278         case IB_CM_REP_SENT:
3279         case IB_CM_MRA_REP_RCVD:
3280                 cm_id->state = IB_CM_ESTABLISHED;
3281                 break;
3282         case IB_CM_ESTABLISHED:
3283                 ret = -EISCONN;
3284                 break;
3285         default:
3286                 ret = -EINVAL;
3287                 break;
3288         }
3289         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3290
3291         if (ret) {
3292                 kfree(work);
3293                 goto out;
3294         }
3295
3296         /*
3297          * The CM worker thread may try to destroy the cm_id before it
3298          * can execute this work item.  To prevent potential deadlock,
3299          * we need to find the cm_id once we're in the context of the
3300          * worker thread, rather than holding a reference on it.
3301          */
3302         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3303         work->local_id = cm_id->local_id;
3304         work->remote_id = cm_id->remote_id;
3305         work->mad_recv_wc = NULL;
3306         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3307         queue_delayed_work(cm.wq, &work->work, 0);
3308 out:
3309         return ret;
3310 }
3311
3312 static int cm_migrate(struct ib_cm_id *cm_id)
3313 {
3314         struct cm_id_private *cm_id_priv;
3315         unsigned long flags;
3316         int ret = 0;
3317
3318         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3319         spin_lock_irqsave(&cm_id_priv->lock, flags);
3320         if (cm_id->state == IB_CM_ESTABLISHED &&
3321             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3322              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3323                 cm_id->lap_state = IB_CM_LAP_IDLE;
3324                 cm_id_priv->av = cm_id_priv->alt_av;
3325         } else
3326                 ret = -EINVAL;
3327         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3328
3329         return ret;
3330 }
3331
3332 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3333 {
3334         int ret;
3335
3336         switch (event) {
3337         case IB_EVENT_COMM_EST:
3338                 ret = cm_establish(cm_id);
3339                 break;
3340         case IB_EVENT_PATH_MIG:
3341                 ret = cm_migrate(cm_id);
3342                 break;
3343         default:
3344                 ret = -EINVAL;
3345         }
3346         return ret;
3347 }
3348 EXPORT_SYMBOL(ib_cm_notify);
3349
3350 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3351                             struct ib_mad_recv_wc *mad_recv_wc)
3352 {
3353         struct cm_port *port = mad_agent->context;
3354         struct cm_work *work;
3355         enum ib_cm_event_type event;
3356         u16 attr_id;
3357         int paths = 0;
3358
3359         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3360         case CM_REQ_ATTR_ID:
3361                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3362                                                     alt_local_lid != 0);
3363                 event = IB_CM_REQ_RECEIVED;
3364                 break;
3365         case CM_MRA_ATTR_ID:
3366                 event = IB_CM_MRA_RECEIVED;
3367                 break;
3368         case CM_REJ_ATTR_ID:
3369                 event = IB_CM_REJ_RECEIVED;
3370                 break;
3371         case CM_REP_ATTR_ID:
3372                 event = IB_CM_REP_RECEIVED;
3373                 break;
3374         case CM_RTU_ATTR_ID:
3375                 event = IB_CM_RTU_RECEIVED;
3376                 break;
3377         case CM_DREQ_ATTR_ID:
3378                 event = IB_CM_DREQ_RECEIVED;
3379                 break;
3380         case CM_DREP_ATTR_ID:
3381                 event = IB_CM_DREP_RECEIVED;
3382                 break;
3383         case CM_SIDR_REQ_ATTR_ID:
3384                 event = IB_CM_SIDR_REQ_RECEIVED;
3385                 break;
3386         case CM_SIDR_REP_ATTR_ID:
3387                 event = IB_CM_SIDR_REP_RECEIVED;
3388                 break;
3389         case CM_LAP_ATTR_ID:
3390                 paths = 1;
3391                 event = IB_CM_LAP_RECEIVED;
3392                 break;
3393         case CM_APR_ATTR_ID:
3394                 event = IB_CM_APR_RECEIVED;
3395                 break;
3396         default:
3397                 ib_free_recv_mad(mad_recv_wc);
3398                 return;
3399         }
3400
3401         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3402         atomic_long_inc(&port->counter_group[CM_RECV].
3403                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3404
3405         work = kmalloc(sizeof *work + sizeof(struct ib_sa_path_rec) * paths,
3406                        GFP_KERNEL);
3407         if (!work) {
3408                 ib_free_recv_mad(mad_recv_wc);
3409                 return;
3410         }
3411
3412         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3413         work->cm_event.event = event;
3414         work->mad_recv_wc = mad_recv_wc;
3415         work->port = port;
3416         queue_delayed_work(cm.wq, &work->work, 0);
3417 }
3418
3419 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3420                                 struct ib_qp_attr *qp_attr,
3421                                 int *qp_attr_mask)
3422 {
3423         unsigned long flags;
3424         int ret;
3425
3426         spin_lock_irqsave(&cm_id_priv->lock, flags);
3427         switch (cm_id_priv->id.state) {
3428         case IB_CM_REQ_SENT:
3429         case IB_CM_MRA_REQ_RCVD:
3430         case IB_CM_REQ_RCVD:
3431         case IB_CM_MRA_REQ_SENT:
3432         case IB_CM_REP_RCVD:
3433         case IB_CM_MRA_REP_SENT:
3434         case IB_CM_REP_SENT:
3435         case IB_CM_MRA_REP_RCVD:
3436         case IB_CM_ESTABLISHED:
3437                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3438                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3439                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3440                 if (cm_id_priv->responder_resources)
3441                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3442                                                     IB_ACCESS_REMOTE_ATOMIC;
3443                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3444                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3445                 ret = 0;
3446                 break;
3447         default:
3448                 ret = -EINVAL;
3449                 break;
3450         }
3451         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3452         return ret;
3453 }
3454
3455 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3456                                struct ib_qp_attr *qp_attr,
3457                                int *qp_attr_mask)
3458 {
3459         unsigned long flags;
3460         int ret;
3461
3462         spin_lock_irqsave(&cm_id_priv->lock, flags);
3463         switch (cm_id_priv->id.state) {
3464         case IB_CM_REQ_RCVD:
3465         case IB_CM_MRA_REQ_SENT:
3466         case IB_CM_REP_RCVD:
3467         case IB_CM_MRA_REP_SENT:
3468         case IB_CM_REP_SENT:
3469         case IB_CM_MRA_REP_RCVD:
3470         case IB_CM_ESTABLISHED:
3471                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3472                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3473                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3474                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3475                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3476                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3477                 if (cm_id_priv->qp_type == IB_QPT_RC) {
3478                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3479                                          IB_QP_MIN_RNR_TIMER;
3480                         qp_attr->max_dest_rd_atomic =
3481                                         cm_id_priv->responder_resources;
3482                         qp_attr->min_rnr_timer = 0;
3483                 }
3484                 if (cm_id_priv->alt_av.ah_attr.dlid) {
3485                         *qp_attr_mask |= IB_QP_ALT_PATH;
3486                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3487                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3488                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3489                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3490                 }
3491                 ret = 0;
3492                 break;
3493         default:
3494                 ret = -EINVAL;
3495                 break;
3496         }
3497         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3498         return ret;
3499 }
3500
3501 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3502                                struct ib_qp_attr *qp_attr,
3503                                int *qp_attr_mask)
3504 {
3505         unsigned long flags;
3506         int ret;
3507
3508         spin_lock_irqsave(&cm_id_priv->lock, flags);
3509         switch (cm_id_priv->id.state) {
3510         /* Allow transition to RTS before sending REP */
3511         case IB_CM_REQ_RCVD:
3512         case IB_CM_MRA_REQ_SENT:
3513
3514         case IB_CM_REP_RCVD:
3515         case IB_CM_MRA_REP_SENT:
3516         case IB_CM_REP_SENT:
3517         case IB_CM_MRA_REP_RCVD:
3518         case IB_CM_ESTABLISHED:
3519                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3520                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3521                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
3522                         if (cm_id_priv->qp_type == IB_QPT_RC) {
3523                                 *qp_attr_mask |= IB_QP_TIMEOUT | IB_QP_RETRY_CNT |
3524                                                  IB_QP_RNR_RETRY |
3525                                                  IB_QP_MAX_QP_RD_ATOMIC;
3526                                 qp_attr->timeout = cm_id_priv->av.timeout;
3527                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
3528                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
3529                                 qp_attr->max_rd_atomic =
3530                                         cm_id_priv->initiator_depth;
3531                         }
3532                         if (cm_id_priv->alt_av.ah_attr.dlid) {
3533                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
3534                                 qp_attr->path_mig_state = IB_MIG_REARM;
3535                         }
3536                 } else {
3537                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
3538                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3539                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3540                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3541                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3542                         qp_attr->path_mig_state = IB_MIG_REARM;
3543                 }
3544                 ret = 0;
3545                 break;
3546         default:
3547                 ret = -EINVAL;
3548                 break;
3549         }
3550         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3551         return ret;
3552 }
3553
3554 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
3555                        struct ib_qp_attr *qp_attr,
3556                        int *qp_attr_mask)
3557 {
3558         struct cm_id_private *cm_id_priv;
3559         int ret;
3560
3561         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3562         switch (qp_attr->qp_state) {
3563         case IB_QPS_INIT:
3564                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
3565                 break;
3566         case IB_QPS_RTR:
3567                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
3568                 break;
3569         case IB_QPS_RTS:
3570                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
3571                 break;
3572         default:
3573                 ret = -EINVAL;
3574                 break;
3575         }
3576         return ret;
3577 }
3578 EXPORT_SYMBOL(ib_cm_init_qp_attr);
3579
3580 static void cm_get_ack_delay(struct cm_device *cm_dev)
3581 {
3582         struct ib_device_attr attr;
3583
3584         if (ib_query_device(cm_dev->ib_device, &attr))
3585                 cm_dev->ack_delay = 0; /* acks will rely on packet life time */
3586         else
3587                 cm_dev->ack_delay = attr.local_ca_ack_delay;
3588 }
3589
3590 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
3591                                char *buf)
3592 {
3593         struct cm_counter_group *group;
3594         struct cm_counter_attribute *cm_attr;
3595
3596         group = container_of(obj, struct cm_counter_group, obj);
3597         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
3598
3599         return sprintf(buf, "%ld\n",
3600                        atomic_long_read(&group->counter[cm_attr->index]));
3601 }
3602
3603 static const struct sysfs_ops cm_counter_ops = {
3604         .show = cm_show_counter
3605 };
3606
3607 static struct kobj_type cm_counter_obj_type = {
3608         .sysfs_ops = &cm_counter_ops,
3609         .default_attrs = cm_counter_default_attrs
3610 };
3611
3612 static void cm_release_port_obj(struct kobject *obj)
3613 {
3614         struct cm_port *cm_port;
3615
3616         cm_port = container_of(obj, struct cm_port, port_obj);
3617         kfree(cm_port);
3618 }
3619
3620 static struct kobj_type cm_port_obj_type = {
3621         .release = cm_release_port_obj
3622 };
3623
3624 struct class cm_class = {
3625         .name    = "infiniband_cm",
3626 };
3627 EXPORT_SYMBOL(cm_class);
3628
3629 static int cm_create_port_fs(struct cm_port *port)
3630 {
3631         int i, ret;
3632
3633         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
3634                                    &port->cm_dev->device->kobj,
3635                                    "%d", port->port_num);
3636         if (ret) {
3637                 kfree(port);
3638                 return ret;
3639         }
3640
3641         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
3642                 ret = kobject_init_and_add(&port->counter_group[i].obj,
3643                                            &cm_counter_obj_type,
3644                                            &port->port_obj,
3645                                            "%s", counter_group_names[i]);
3646                 if (ret)
3647                         goto error;
3648         }
3649
3650         return 0;
3651
3652 error:
3653         while (i--)
3654                 kobject_put(&port->counter_group[i].obj);
3655         kobject_put(&port->port_obj);
3656         return ret;
3657
3658 }
3659
3660 static void cm_remove_port_fs(struct cm_port *port)
3661 {
3662         int i;
3663
3664         for (i = 0; i < CM_COUNTER_GROUPS; i++)
3665                 kobject_put(&port->counter_group[i].obj);
3666
3667         kobject_put(&port->port_obj);
3668 }
3669
3670 static void cm_add_one(struct ib_device *ib_device)
3671 {
3672         struct cm_device *cm_dev;
3673         struct cm_port *port;
3674         struct ib_mad_reg_req reg_req = {
3675                 .mgmt_class = IB_MGMT_CLASS_CM,
3676                 .mgmt_class_version = IB_CM_CLASS_VERSION
3677         };
3678         struct ib_port_modify port_modify = {
3679                 .set_port_cap_mask = IB_PORT_CM_SUP
3680         };
3681         unsigned long flags;
3682         int ret;
3683         u8 i;
3684
3685         if (rdma_node_get_transport(ib_device->node_type) != RDMA_TRANSPORT_IB)
3686                 return;
3687
3688         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
3689                          ib_device->phys_port_cnt, GFP_KERNEL);
3690         if (!cm_dev)
3691                 return;
3692
3693         cm_dev->ib_device = ib_device;
3694         cm_get_ack_delay(cm_dev);
3695
3696         cm_dev->device = device_create(&cm_class, &ib_device->dev,
3697                                        MKDEV(0, 0), NULL,
3698                                        "%s", ib_device->name);
3699         if (IS_ERR(cm_dev->device)) {
3700                 kfree(cm_dev);
3701                 return;
3702         }
3703
3704         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
3705         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3706                 port = kzalloc(sizeof *port, GFP_KERNEL);
3707                 if (!port)
3708                         goto error1;
3709
3710                 cm_dev->port[i-1] = port;
3711                 port->cm_dev = cm_dev;
3712                 port->port_num = i;
3713
3714                 ret = cm_create_port_fs(port);
3715                 if (ret)
3716                         goto error1;
3717
3718                 port->mad_agent = ib_register_mad_agent(ib_device, i,
3719                                                         IB_QPT_GSI,
3720                                                         &reg_req,
3721                                                         0,
3722                                                         cm_send_handler,
3723                                                         cm_recv_handler,
3724                                                         port);
3725                 if (IS_ERR(port->mad_agent))
3726                         goto error2;
3727
3728                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
3729                 if (ret)
3730                         goto error3;
3731         }
3732         ib_set_client_data(ib_device, &cm_client, cm_dev);
3733
3734         write_lock_irqsave(&cm.device_lock, flags);
3735         list_add_tail(&cm_dev->list, &cm.device_list);
3736         write_unlock_irqrestore(&cm.device_lock, flags);
3737         return;
3738
3739 error3:
3740         ib_unregister_mad_agent(port->mad_agent);
3741 error2:
3742         cm_remove_port_fs(port);
3743 error1:
3744         port_modify.set_port_cap_mask = 0;
3745         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
3746         while (--i) {
3747                 port = cm_dev->port[i-1];
3748                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3749                 ib_unregister_mad_agent(port->mad_agent);
3750                 cm_remove_port_fs(port);
3751         }
3752         device_unregister(cm_dev->device);
3753         kfree(cm_dev);
3754 }
3755
3756 static void cm_remove_one(struct ib_device *ib_device)
3757 {
3758         struct cm_device *cm_dev;
3759         struct cm_port *port;
3760         struct ib_port_modify port_modify = {
3761                 .clr_port_cap_mask = IB_PORT_CM_SUP
3762         };
3763         unsigned long flags;
3764         int i;
3765
3766         cm_dev = ib_get_client_data(ib_device, &cm_client);
3767         if (!cm_dev)
3768                 return;
3769
3770         write_lock_irqsave(&cm.device_lock, flags);
3771         list_del(&cm_dev->list);
3772         write_unlock_irqrestore(&cm.device_lock, flags);
3773
3774         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
3775                 port = cm_dev->port[i-1];
3776                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
3777                 ib_unregister_mad_agent(port->mad_agent);
3778                 flush_workqueue(cm.wq);
3779                 cm_remove_port_fs(port);
3780         }
3781         device_unregister(cm_dev->device);
3782         kfree(cm_dev);
3783 }
3784
3785 static int __init ib_cm_init(void)
3786 {
3787         int ret;
3788
3789         memset(&cm, 0, sizeof cm);
3790         INIT_LIST_HEAD(&cm.device_list);
3791         rwlock_init(&cm.device_lock);
3792         spin_lock_init(&cm.lock);
3793         cm.listen_service_table = RB_ROOT;
3794         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
3795         cm.remote_id_table = RB_ROOT;
3796         cm.remote_qp_table = RB_ROOT;
3797         cm.remote_sidr_table = RB_ROOT;
3798         idr_init(&cm.local_id_table);
3799         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
3800         idr_pre_get(&cm.local_id_table, GFP_KERNEL);
3801         INIT_LIST_HEAD(&cm.timewait_list);
3802
3803         ret = class_register(&cm_class);
3804         if (ret)
3805                 return -ENOMEM;
3806
3807         cm.wq = create_workqueue("ib_cm");
3808         if (!cm.wq) {
3809                 ret = -ENOMEM;
3810                 goto error1;
3811         }
3812
3813         ret = ib_register_client(&cm_client);
3814         if (ret)
3815                 goto error2;
3816
3817         return 0;
3818 error2:
3819         destroy_workqueue(cm.wq);
3820 error1:
3821         class_unregister(&cm_class);
3822         return ret;
3823 }
3824
3825 static void __exit ib_cm_cleanup(void)
3826 {
3827         struct cm_timewait_info *timewait_info, *tmp;
3828
3829         spin_lock_irq(&cm.lock);
3830         list_for_each_entry(timewait_info, &cm.timewait_list, list)
3831                 cancel_delayed_work(&timewait_info->work.work);
3832         spin_unlock_irq(&cm.lock);
3833
3834         ib_unregister_client(&cm_client);
3835         destroy_workqueue(cm.wq);
3836
3837         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
3838                 list_del(&timewait_info->list);
3839                 kfree(timewait_info);
3840         }
3841
3842         class_unregister(&cm_class);
3843         idr_destroy(&cm.local_id_table);
3844 }
3845
3846 module_init(ib_cm_init);
3847 module_exit(ib_cm_cleanup);
3848