drivers/infiniband/core/mad.c

   1 /*
   2  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
   3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id: mad.c 5596 2006-03-03 01:00:07Z sean.hefty $
  35  */
  36 #include <linux/dma-mapping.h>
  37
  38 #include "mad_priv.h"
  39 #include "mad_rmpp.h"
  40 #include "smi.h"
  41 #include "agent.h"
  42
  43 MODULE_LICENSE("Dual BSD/GPL");
  44 MODULE_DESCRIPTION("kernel IB MAD API");
  45 MODULE_AUTHOR("Hal Rosenstock");
  46 MODULE_AUTHOR("Sean Hefty");
  47
  48
  49 kmem_cache_t *ib_mad_cache;
  50
  51 static struct list_head ib_mad_port_list;
  52 static u32 ib_mad_client_id = 0;
  53
  54 /* Port list lock */
  55 static spinlock_t ib_mad_port_list_lock;
  56
  57
  58 /* Forward declarations */
  59 static int method_in_use(struct ib_mad_mgmt_method_table **method,
  60                          struct ib_mad_reg_req *mad_reg_req);
  61 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
  62 static struct ib_mad_agent_private *find_mad_agent(
  63                                         struct ib_mad_port_private *port_priv,
  64                                         struct ib_mad *mad);
  65 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
  66                                     struct ib_mad_private *mad);
  67 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
  68 static void timeout_sends(void *data);
  69 static void local_completions(void *data);
  70 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  71                               struct ib_mad_agent_private *agent_priv,
  72                               u8 mgmt_class);
  73 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  74                            struct ib_mad_agent_private *agent_priv);
  75
  76 /*
  77  * Returns a ib_mad_port_private structure or NULL for a device/port
  78  * Assumes ib_mad_port_list_lock is being held
  79  */
  80 static inline struct ib_mad_port_private *
  81 __ib_get_mad_port(struct ib_device *device, int port_num)
  82 {
  83         struct ib_mad_port_private *entry;
  84
  85         list_for_each_entry(entry, &ib_mad_port_list, port_list) {
  86                 if (entry->device == device && entry->port_num == port_num)
  87                         return entry;
  88         }
  89         return NULL;
  90 }
  91
  92 /*
  93  * Wrapper function to return a ib_mad_port_private structure or NULL
  94  * for a device/port
  95  */
  96 static inline struct ib_mad_port_private *
  97 ib_get_mad_port(struct ib_device *device, int port_num)
  98 {
  99         struct ib_mad_port_private *entry;
 100         unsigned long flags;
 101
 102         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
 103         entry = __ib_get_mad_port(device, port_num);
 104         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
 105
 106         return entry;
 107 }
 108
 109 static inline u8 convert_mgmt_class(u8 mgmt_class)
 110 {
 111         /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
 112         return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
 113                 0 : mgmt_class;
 114 }
 115
 116 static int get_spl_qp_index(enum ib_qp_type qp_type)
 117 {
 118         switch (qp_type)
 119         {
 120         case IB_QPT_SMI:
 121                 return 0;
 122         case IB_QPT_GSI:
 123                 return 1;
 124         default:
 125                 return -1;
 126         }
 127 }
 128
 129 static int vendor_class_index(u8 mgmt_class)
 130 {
 131         return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
 132 }
 133
 134 static int is_vendor_class(u8 mgmt_class)
 135 {
 136         if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
 137             (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
 138                 return 0;
 139         return 1;
 140 }
 141
 142 static int is_vendor_oui(char *oui)
 143 {
 144         if (oui[0] || oui[1] || oui[2])
 145                 return 1;
 146         return 0;
 147 }
 148
 149 static int is_vendor_method_in_use(
 150                 struct ib_mad_mgmt_vendor_class *vendor_class,
 151                 struct ib_mad_reg_req *mad_reg_req)
 152 {
 153         struct ib_mad_mgmt_method_table *method;
 154         int i;
 155
 156         for (i = 0; i < MAX_MGMT_OUI; i++) {
 157                 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
 158                         method = vendor_class->method_table[i];
 159                         if (method) {
 160                                 if (method_in_use(&method, mad_reg_req))
 161                                         return 1;
 162                                 else
 163                                         break;
 164                         }
 165                 }
 166         }
 167         return 0;
 168 }
 169
 170 /*
 171  * ib_register_mad_agent - Register to send/receive MADs
 172  */
 173 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 174                                            u8 port_num,
 175                                            enum ib_qp_type qp_type,
 176                                            struct ib_mad_reg_req *mad_reg_req,
 177                                            u8 rmpp_version,
 178                                            ib_mad_send_handler send_handler,
 179                                            ib_mad_recv_handler recv_handler,
 180                                            void *context)
 181 {
 182         struct ib_mad_port_private *port_priv;
 183         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
 184         struct ib_mad_agent_private *mad_agent_priv;
 185         struct ib_mad_reg_req *reg_req = NULL;
 186         struct ib_mad_mgmt_class_table *class;
 187         struct ib_mad_mgmt_vendor_class_table *vendor;
 188         struct ib_mad_mgmt_vendor_class *vendor_class;
 189         struct ib_mad_mgmt_method_table *method;
 190         int ret2, qpn;
 191         unsigned long flags;
 192         u8 mgmt_class, vclass;
 193
 194         /* Validate parameters */
 195         qpn = get_spl_qp_index(qp_type);
 196         if (qpn == -1)
 197                 goto error1;
 198
 199         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
 200                 goto error1;
 201
 202         /* Validate MAD registration request if supplied */
 203         if (mad_reg_req) {
 204                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
 205                         goto error1;
 206                 if (!recv_handler)
 207                         goto error1;
 208                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
 209                         /*
 210                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
 211                          * one in this range currently allowed
 212                          */
 213                         if (mad_reg_req->mgmt_class !=
 214                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 215                                 goto error1;
 216                 } else if (mad_reg_req->mgmt_class == 0) {
 217                         /*
 218                          * Class 0 is reserved in IBA and is used for
 219                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
 220                          */
 221                         goto error1;
 222                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
 223                         /*
 224                          * If class is in "new" vendor range,
 225                          * ensure supplied OUI is not zero
 226                          */
 227                         if (!is_vendor_oui(mad_reg_req->oui))
 228                                 goto error1;
 229                 }
 230                 /* Make sure class supplied is consistent with QP type */
 231                 if (qp_type == IB_QPT_SMI) {
 232                         if ((mad_reg_req->mgmt_class !=
 233                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
 234                             (mad_reg_req->mgmt_class !=
 235                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 236                                 goto error1;
 237                 } else {
 238                         if ((mad_reg_req->mgmt_class ==
 239                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
 240                             (mad_reg_req->mgmt_class ==
 241                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 242                                 goto error1;
 243                 }
 244         } else {
 245                 /* No registration request supplied */
 246                 if (!send_handler)
 247                         goto error1;
 248         }
 249
 250         /* Validate device and port */
 251         port_priv = ib_get_mad_port(device, port_num);
 252         if (!port_priv) {
 253                 ret = ERR_PTR(-ENODEV);
 254                 goto error1;
 255         }
 256
 257         /* Allocate structures */
 258         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
 259         if (!mad_agent_priv) {
 260                 ret = ERR_PTR(-ENOMEM);
 261                 goto error1;
 262         }
 263
 264         mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
 265                                                  IB_ACCESS_LOCAL_WRITE);
 266         if (IS_ERR(mad_agent_priv->agent.mr)) {
 267                 ret = ERR_PTR(-ENOMEM);
 268                 goto error2;
 269         }
 270
 271         if (mad_reg_req) {
 272                 reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
 273                 if (!reg_req) {
 274                         ret = ERR_PTR(-ENOMEM);
 275                         goto error3;
 276                 }
 277                 /* Make a copy of the MAD registration request */
 278                 memcpy(reg_req, mad_reg_req, sizeof *reg_req);
 279         }
 280
 281         /* Now, fill in the various structures */
 282         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 283         mad_agent_priv->reg_req = reg_req;
 284         mad_agent_priv->agent.rmpp_version = rmpp_version;
 285         mad_agent_priv->agent.device = device;
 286         mad_agent_priv->agent.recv_handler = recv_handler;
 287         mad_agent_priv->agent.send_handler = send_handler;
 288         mad_agent_priv->agent.context = context;
 289         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 290         mad_agent_priv->agent.port_num = port_num;
 291
 292         spin_lock_irqsave(&port_priv->reg_lock, flags);
 293         mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
 294
 295         /*
 296          * Make sure MAD registration (if supplied)
 297          * is non overlapping with any existing ones
 298          */
 299         if (mad_reg_req) {
 300                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 301                 if (!is_vendor_class(mgmt_class)) {
 302                         class = port_priv->version[mad_reg_req->
 303                                                    mgmt_class_version].class;
 304                         if (class) {
 305                                 method = class->method_table[mgmt_class];
 306                                 if (method) {
 307                                         if (method_in_use(&method,
 308                                                            mad_reg_req))
 309                                                 goto error4;
 310                                 }
 311                         }
 312                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 313                                                   mgmt_class);
 314                 } else {
 315                         /* "New" vendor class range */
 316                         vendor = port_priv->version[mad_reg_req->
 317                                                     mgmt_class_version].vendor;
 318                         if (vendor) {
 319                                 vclass = vendor_class_index(mgmt_class);
 320                                 vendor_class = vendor->vendor_class[vclass];
 321                                 if (vendor_class) {
 322                                         if (is_vendor_method_in_use(
 323                                                         vendor_class,
 324                                                         mad_reg_req))
 325                                                 goto error4;
 326                                 }
 327                         }
 328                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 329                 }
 330                 if (ret2) {
 331                         ret = ERR_PTR(ret2);
 332                         goto error4;
 333                 }
 334         }
 335
 336         /* Add mad agent into port's agent list */
 337         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
 338         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 339
 340         spin_lock_init(&mad_agent_priv->lock);
 341         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 342         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 343         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 344         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 345         INIT_WORK(&mad_agent_priv->timed_work, timeout_sends, mad_agent_priv);
 346         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 347         INIT_WORK(&mad_agent_priv->local_work, local_completions,
 348                    mad_agent_priv);
 349         atomic_set(&mad_agent_priv->refcount, 1);
 350         init_waitqueue_head(&mad_agent_priv->wait);
 351
 352         return &mad_agent_priv->agent;
 353
 354 error4:
 355         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 356         kfree(reg_req);
 357 error3:
 358         ib_dereg_mr(mad_agent_priv->agent.mr);
 359 error2:
 360         kfree(mad_agent_priv);
 361 error1:
 362         return ret;
 363 }
 364 EXPORT_SYMBOL(ib_register_mad_agent);
 365
 366 static inline int is_snooping_sends(int mad_snoop_flags)
 367 {
 368         return (mad_snoop_flags &
 369                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 370                  IB_MAD_SNOOP_RMPP_SENDS |*/
 371                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 372                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 373 }
 374
 375 static inline int is_snooping_recvs(int mad_snoop_flags)
 376 {
 377         return (mad_snoop_flags &
 378                 (IB_MAD_SNOOP_RECVS /*|
 379                  IB_MAD_SNOOP_RMPP_RECVS*/));
 380 }
 381
 382 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 383                                 struct ib_mad_snoop_private *mad_snoop_priv)
 384 {
 385         struct ib_mad_snoop_private **new_snoop_table;
 386         unsigned long flags;
 387         int i;
 388
 389         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 390         /* Check for empty slot in array. */
 391         for (i = 0; i < qp_info->snoop_table_size; i++)
 392                 if (!qp_info->snoop_table[i])
 393                         break;
 394
 395         if (i == qp_info->snoop_table_size) {
 396                 /* Grow table. */
 397                 new_snoop_table = kmalloc(sizeof mad_snoop_priv *
 398                                           qp_info->snoop_table_size + 1,
 399                                           GFP_ATOMIC);
 400                 if (!new_snoop_table) {
 401                         i = -ENOMEM;
 402                         goto out;
 403                 }
 404                 if (qp_info->snoop_table) {
 405                         memcpy(new_snoop_table, qp_info->snoop_table,
 406                                sizeof mad_snoop_priv *
 407                                qp_info->snoop_table_size);
 408                         kfree(qp_info->snoop_table);
 409                 }
 410                 qp_info->snoop_table = new_snoop_table;
 411                 qp_info->snoop_table_size++;
 412         }
 413         qp_info->snoop_table[i] = mad_snoop_priv;
 414         atomic_inc(&qp_info->snoop_count);
 415 out:
 416         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 417         return i;
 418 }
 419
 420 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 421                                            u8 port_num,
 422                                            enum ib_qp_type qp_type,
 423                                            int mad_snoop_flags,
 424                                            ib_mad_snoop_handler snoop_handler,
 425                                            ib_mad_recv_handler recv_handler,
 426                                            void *context)
 427 {
 428         struct ib_mad_port_private *port_priv;
 429         struct ib_mad_agent *ret;
 430         struct ib_mad_snoop_private *mad_snoop_priv;
 431         int qpn;
 432
 433         /* Validate parameters */
 434         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 435             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 436                 ret = ERR_PTR(-EINVAL);
 437                 goto error1;
 438         }
 439         qpn = get_spl_qp_index(qp_type);
 440         if (qpn == -1) {
 441                 ret = ERR_PTR(-EINVAL);
 442                 goto error1;
 443         }
 444         port_priv = ib_get_mad_port(device, port_num);
 445         if (!port_priv) {
 446                 ret = ERR_PTR(-ENODEV);
 447                 goto error1;
 448         }
 449         /* Allocate structures */
 450         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 451         if (!mad_snoop_priv) {
 452                 ret = ERR_PTR(-ENOMEM);
 453                 goto error1;
 454         }
 455
 456         /* Now, fill in the various structures */
 457         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 458         mad_snoop_priv->agent.device = device;
 459         mad_snoop_priv->agent.recv_handler = recv_handler;
 460         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 461         mad_snoop_priv->agent.context = context;
 462         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 463         mad_snoop_priv->agent.port_num = port_num;
 464         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 465         init_waitqueue_head(&mad_snoop_priv->wait);
 466         mad_snoop_priv->snoop_index = register_snoop_agent(
 467                                                 &port_priv->qp_info[qpn],
 468                                                 mad_snoop_priv);
 469         if (mad_snoop_priv->snoop_index < 0) {
 470                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 471                 goto error2;
 472         }
 473
 474         atomic_set(&mad_snoop_priv->refcount, 1);
 475         return &mad_snoop_priv->agent;
 476
 477 error2:
 478         kfree(mad_snoop_priv);
 479 error1:
 480         return ret;
 481 }
 482 EXPORT_SYMBOL(ib_register_mad_snoop);
 483
 484 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 485 {
 486         struct ib_mad_port_private *port_priv;
 487         unsigned long flags;
 488
 489         /* Note that we could still be handling received MADs */
 490
 491         /*
 492          * Canceling all sends results in dropping received response
 493          * MADs, preventing us from queuing additional work
 494          */
 495         cancel_mads(mad_agent_priv);
 496         port_priv = mad_agent_priv->qp_info->port_priv;
 497         cancel_delayed_work(&mad_agent_priv->timed_work);
 498
 499         spin_lock_irqsave(&port_priv->reg_lock, flags);
 500         remove_mad_reg_req(mad_agent_priv);
 501         list_del(&mad_agent_priv->agent_list);
 502         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 503
 504         flush_workqueue(port_priv->wq);
 505         ib_cancel_rmpp_recvs(mad_agent_priv);
 506
 507         atomic_dec(&mad_agent_priv->refcount);
 508         wait_event(mad_agent_priv->wait,
 509                    !atomic_read(&mad_agent_priv->refcount));
 510
 511         kfree(mad_agent_priv->reg_req);
 512         ib_dereg_mr(mad_agent_priv->agent.mr);
 513         kfree(mad_agent_priv);
 514 }
 515
 516 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 517 {
 518         struct ib_mad_qp_info *qp_info;
 519         unsigned long flags;
 520
 521         qp_info = mad_snoop_priv->qp_info;
 522         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 523         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 524         atomic_dec(&qp_info->snoop_count);
 525         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 526
 527         atomic_dec(&mad_snoop_priv->refcount);
 528         wait_event(mad_snoop_priv->wait,
 529                    !atomic_read(&mad_snoop_priv->refcount));
 530
 531         kfree(mad_snoop_priv);
 532 }
 533
 534 /*
 535  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 536  */
 537 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 538 {
 539         struct ib_mad_agent_private *mad_agent_priv;
 540         struct ib_mad_snoop_private *mad_snoop_priv;
 541
 542         /* If the TID is zero, the agent can only snoop. */
 543         if (mad_agent->hi_tid) {
 544                 mad_agent_priv = container_of(mad_agent,
 545                                               struct ib_mad_agent_private,
 546                                               agent);
 547                 unregister_mad_agent(mad_agent_priv);
 548         } else {
 549                 mad_snoop_priv = container_of(mad_agent,
 550                                               struct ib_mad_snoop_private,
 551                                               agent);
 552                 unregister_mad_snoop(mad_snoop_priv);
 553         }
 554         return 0;
 555 }
 556 EXPORT_SYMBOL(ib_unregister_mad_agent);
 557
 558 static inline int response_mad(struct ib_mad *mad)
 559 {
 560         /* Trap represses are responses although response bit is reset */
 561         return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
 562                 (mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
 563 }
 564
 565 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 566 {
 567         struct ib_mad_queue *mad_queue;
 568         unsigned long flags;
 569
 570         BUG_ON(!mad_list->mad_queue);
 571         mad_queue = mad_list->mad_queue;
 572         spin_lock_irqsave(&mad_queue->lock, flags);
 573         list_del(&mad_list->list);
 574         mad_queue->count--;
 575         spin_unlock_irqrestore(&mad_queue->lock, flags);
 576 }
 577
 578 static void snoop_send(struct ib_mad_qp_info *qp_info,
 579                        struct ib_mad_send_buf *send_buf,
 580                        struct ib_mad_send_wc *mad_send_wc,
 581                        int mad_snoop_flags)
 582 {
 583         struct ib_mad_snoop_private *mad_snoop_priv;
 584         unsigned long flags;
 585         int i;
 586
 587         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 588         for (i = 0; i < qp_info->snoop_table_size; i++) {
 589                 mad_snoop_priv = qp_info->snoop_table[i];
 590                 if (!mad_snoop_priv ||
 591                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 592                         continue;
 593
 594                 atomic_inc(&mad_snoop_priv->refcount);
 595                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 596                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 597                                                     send_buf, mad_send_wc);
 598                 if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 599                         wake_up(&mad_snoop_priv->wait);
 600                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 601         }
 602         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 603 }
 604
 605 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 606                        struct ib_mad_recv_wc *mad_recv_wc,
 607                        int mad_snoop_flags)
 608 {
 609         struct ib_mad_snoop_private *mad_snoop_priv;
 610         unsigned long flags;
 611         int i;
 612
 613         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 614         for (i = 0; i < qp_info->snoop_table_size; i++) {
 615                 mad_snoop_priv = qp_info->snoop_table[i];
 616                 if (!mad_snoop_priv ||
 617                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 618                         continue;
 619
 620                 atomic_inc(&mad_snoop_priv->refcount);
 621                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 622                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
 623                                                    mad_recv_wc);
 624                 if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 625                         wake_up(&mad_snoop_priv->wait);
 626                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 627         }
 628         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 629 }
 630
 631 static void build_smp_wc(u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
 632                          struct ib_wc *wc)
 633 {
 634         memset(wc, 0, sizeof *wc);
 635         wc->wr_id = wr_id;
 636         wc->status = IB_WC_SUCCESS;
 637         wc->opcode = IB_WC_RECV;
 638         wc->pkey_index = pkey_index;
 639         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
 640         wc->src_qp = IB_QP0;
 641         wc->qp_num = IB_QP0;
 642         wc->slid = slid;
 643         wc->sl = 0;
 644         wc->dlid_path_bits = 0;
 645         wc->port_num = port_num;
 646 }
 647
 648 /*
 649  * Return 0 if SMP is to be sent
 650  * Return 1 if SMP was consumed locally (whether or not solicited)
 651  * Return < 0 if error
 652  */
 653 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 654                                   struct ib_mad_send_wr_private *mad_send_wr)
 655 {
 656         int ret;
 657         struct ib_smp *smp = mad_send_wr->send_buf.mad;
 658         unsigned long flags;
 659         struct ib_mad_local_private *local;
 660         struct ib_mad_private *mad_priv;
 661         struct ib_mad_port_private *port_priv;
 662         struct ib_mad_agent_private *recv_mad_agent = NULL;
 663         struct ib_device *device = mad_agent_priv->agent.device;
 664         u8 port_num = mad_agent_priv->agent.port_num;
 665         struct ib_wc mad_wc;
 666         struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
 667
 668         /*
 669          * Directed route handling starts if the initial LID routed part of
 670          * a request or the ending LID routed part of a response is empty.
 671          * If we are at the start of the LID routed part, don't update the
 672          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
 673          */
 674         if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
 675              IB_LID_PERMISSIVE &&
 676             !smi_handle_dr_smp_send(smp, device->node_type, port_num)) {
 677                 ret = -EINVAL;
 678                 printk(KERN_ERR PFX "Invalid directed route\n");
 679                 goto out;
 680         }
 681         /* Check to post send on QP or process locally */
 682         ret = smi_check_local_smp(smp, device);
 683         if (!ret)
 684                 goto out;
 685
 686         local = kmalloc(sizeof *local, GFP_ATOMIC);
 687         if (!local) {
 688                 ret = -ENOMEM;
 689                 printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
 690                 goto out;
 691         }
 692         local->mad_priv = NULL;
 693         local->recv_mad_agent = NULL;
 694         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
 695         if (!mad_priv) {
 696                 ret = -ENOMEM;
 697                 printk(KERN_ERR PFX "No memory for local response MAD\n");
 698                 kfree(local);
 699                 goto out;
 700         }
 701
 702         build_smp_wc(send_wr->wr_id, be16_to_cpu(smp->dr_slid),
 703                      send_wr->wr.ud.pkey_index,
 704                      send_wr->wr.ud.port_num, &mad_wc);
 705
 706         /* No GRH for DR SMP */
 707         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 708                                   (struct ib_mad *)smp,
 709                                   (struct ib_mad *)&mad_priv->mad);
 710         switch (ret)
 711         {
 712         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 713                 if (response_mad(&mad_priv->mad.mad) &&
 714                     mad_agent_priv->agent.recv_handler) {
 715                         local->mad_priv = mad_priv;
 716                         local->recv_mad_agent = mad_agent_priv;
 717                         /*
 718                          * Reference MAD agent until receive
 719                          * side of local completion handled
 720                          */
 721                         atomic_inc(&mad_agent_priv->refcount);
 722                 } else
 723                         kmem_cache_free(ib_mad_cache, mad_priv);
 724                 break;
 725         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 726                 kmem_cache_free(ib_mad_cache, mad_priv);
 727                 break;
 728         case IB_MAD_RESULT_SUCCESS:
 729                 /* Treat like an incoming receive MAD */
 730                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 731                                             mad_agent_priv->agent.port_num);
 732                 if (port_priv) {
 733                         mad_priv->mad.mad.mad_hdr.tid =
 734                                 ((struct ib_mad *)smp)->mad_hdr.tid;
 735                         recv_mad_agent = find_mad_agent(port_priv,
 736                                                         &mad_priv->mad.mad);
 737                 }
 738                 if (!port_priv || !recv_mad_agent) {
 739                         kmem_cache_free(ib_mad_cache, mad_priv);
 740                         kfree(local);
 741                         ret = 0;
 742                         goto out;
 743                 }
 744                 local->mad_priv = mad_priv;
 745                 local->recv_mad_agent = recv_mad_agent;
 746                 break;
 747         default:
 748                 kmem_cache_free(ib_mad_cache, mad_priv);
 749                 kfree(local);
 750                 ret = -EINVAL;
 751                 goto out;
 752         }
 753
 754         local->mad_send_wr = mad_send_wr;
 755         /* Reference MAD agent until send side of local completion handled */
 756         atomic_inc(&mad_agent_priv->refcount);
 757         /* Queue local completion to local list */
 758         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 759         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 760         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 761         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 762                    &mad_agent_priv->local_work);
 763         ret = 1;
 764 out:
 765         return ret;
 766 }
 767
 768 static int get_pad_size(int hdr_len, int data_len)
 769 {
 770         int seg_size, pad;
 771
 772         seg_size = sizeof(struct ib_mad) - hdr_len;
 773         if (data_len && seg_size) {
 774                 pad = seg_size - data_len % seg_size;
 775                 return pad == seg_size ? 0 : pad;
 776         } else
 777                 return seg_size;
 778 }
 779
 780 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 781 {
 782         struct ib_rmpp_segment *s, *t;
 783
 784         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 785                 list_del(&s->list);
 786                 kfree(s);
 787         }
 788 }
 789
 790 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 791                                 gfp_t gfp_mask)
 792 {
 793         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 794         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 795         struct ib_rmpp_segment *seg = NULL;
 796         int left, seg_size, pad;
 797
 798         send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
 799         seg_size = send_buf->seg_size;
 800         pad = send_wr->pad;
 801
 802         /* Allocate data segments. */
 803         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 804                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 805                 if (!seg) {
 806                         printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
 807                                "alloc failed for len %zd, gfp %#x\n",
 808                                sizeof (*seg) + seg_size, gfp_mask);
 809                         free_send_rmpp_list(send_wr);
 810                         return -ENOMEM;
 811                 }
 812                 seg->num = ++send_buf->seg_count;
 813                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 814         }
 815
 816         /* Zero any padding */
 817         if (pad)
 818                 memset(seg->data + seg_size - pad, 0, pad);
 819
 820         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
 821                                           agent.rmpp_version;
 822         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 823         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
 824
 825         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
 826                                         struct ib_rmpp_segment, list);
 827         send_wr->last_ack_seg = send_wr->cur_seg;
 828         return 0;
 829 }
 830
 831 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 832                                             u32 remote_qpn, u16 pkey_index,
 833                                             int rmpp_active,
 834                                             int hdr_len, int data_len,
 835                                             gfp_t gfp_mask)
 836 {
 837         struct ib_mad_agent_private *mad_agent_priv;
 838         struct ib_mad_send_wr_private *mad_send_wr;
 839         int pad, message_size, ret, size;
 840         void *buf;
 841
 842         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 843                                       agent);
 844         pad = get_pad_size(hdr_len, data_len);
 845         message_size = hdr_len + data_len + pad;
 846
 847         if ((!mad_agent->rmpp_version &&
 848              (rmpp_active || message_size > sizeof(struct ib_mad))) ||
 849             (!rmpp_active && message_size > sizeof(struct ib_mad)))
 850                 return ERR_PTR(-EINVAL);
 851
 852         size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
 853         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
 854         if (!buf)
 855                 return ERR_PTR(-ENOMEM);
 856
 857         mad_send_wr = buf + size;
 858         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
 859         mad_send_wr->send_buf.mad = buf;
 860         mad_send_wr->send_buf.hdr_len = hdr_len;
 861         mad_send_wr->send_buf.data_len = data_len;
 862         mad_send_wr->pad = pad;
 863
 864         mad_send_wr->mad_agent_priv = mad_agent_priv;
 865         mad_send_wr->sg_list[0].length = hdr_len;
 866         mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
 867         mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
 868         mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
 869
 870         mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
 871         mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 872         mad_send_wr->send_wr.num_sge = 2;
 873         mad_send_wr->send_wr.opcode = IB_WR_SEND;
 874         mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
 875         mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
 876         mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 877         mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
 878
 879         if (rmpp_active) {
 880                 ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
 881                 if (ret) {
 882                         kfree(buf);
 883                         return ERR_PTR(ret);
 884                 }
 885         }
 886
 887         mad_send_wr->send_buf.mad_agent = mad_agent;
 888         atomic_inc(&mad_agent_priv->refcount);
 889         return &mad_send_wr->send_buf;
 890 }
 891 EXPORT_SYMBOL(ib_create_send_mad);
 892
 893 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
 894 {
 895         struct ib_mad_send_wr_private *mad_send_wr;
 896         struct list_head *list;
 897
 898         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 899                                    send_buf);
 900         list = &mad_send_wr->cur_seg->list;
 901
 902         if (mad_send_wr->cur_seg->num < seg_num) {
 903                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
 904                         if (mad_send_wr->cur_seg->num == seg_num)
 905                                 break;
 906         } else if (mad_send_wr->cur_seg->num > seg_num) {
 907                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
 908                         if (mad_send_wr->cur_seg->num == seg_num)
 909                                 break;
 910         }
 911         return mad_send_wr->cur_seg->data;
 912 }
 913 EXPORT_SYMBOL(ib_get_rmpp_segment);
 914
 915 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
 916 {
 917         if (mad_send_wr->send_buf.seg_count)
 918                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
 919                                            mad_send_wr->seg_num);
 920         else
 921                 return mad_send_wr->send_buf.mad +
 922                        mad_send_wr->send_buf.hdr_len;
 923 }
 924
 925 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
 926 {
 927         struct ib_mad_agent_private *mad_agent_priv;
 928         struct ib_mad_send_wr_private *mad_send_wr;
 929
 930         mad_agent_priv = container_of(send_buf->mad_agent,
 931                                       struct ib_mad_agent_private, agent);
 932         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 933                                    send_buf);
 934
 935         free_send_rmpp_list(mad_send_wr);
 936         kfree(send_buf->mad);
 937         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 938                 wake_up(&mad_agent_priv->wait);
 939 }
 940 EXPORT_SYMBOL(ib_free_send_mad);
 941
 942 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 943 {
 944         struct ib_mad_qp_info *qp_info;
 945         struct list_head *list;
 946         struct ib_send_wr *bad_send_wr;
 947         struct ib_mad_agent *mad_agent;
 948         struct ib_sge *sge;
 949         unsigned long flags;
 950         int ret;
 951
 952         /* Set WR ID to find mad_send_wr upon completion */
 953         qp_info = mad_send_wr->mad_agent_priv->qp_info;
 954         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
 955         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
 956
 957         mad_agent = mad_send_wr->send_buf.mad_agent;
 958         sge = mad_send_wr->sg_list;
 959         sge[0].addr = dma_map_single(mad_agent->device->dma_device,
 960                                      mad_send_wr->send_buf.mad,
 961                                      sge[0].length,
 962                                      DMA_TO_DEVICE);
 963         pci_unmap_addr_set(mad_send_wr, header_mapping, sge[0].addr);
 964
 965         sge[1].addr = dma_map_single(mad_agent->device->dma_device,
 966                                      ib_get_payload(mad_send_wr),
 967                                      sge[1].length,
 968                                      DMA_TO_DEVICE);
 969         pci_unmap_addr_set(mad_send_wr, payload_mapping, sge[1].addr);
 970
 971         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
 972         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
 973                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
 974                                    &bad_send_wr);
 975                 list = &qp_info->send_queue.list;
 976         } else {
 977                 ret = 0;
 978                 list = &qp_info->overflow_list;
 979         }
 980
 981         if (!ret) {
 982                 qp_info->send_queue.count++;
 983                 list_add_tail(&mad_send_wr->mad_list.list, list);
 984         }
 985         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
 986         if (ret) {
 987                 dma_unmap_single(mad_agent->device->dma_device,
 988                                  pci_unmap_addr(mad_send_wr, header_mapping),
 989                                  sge[0].length, DMA_TO_DEVICE);
 990                 dma_unmap_single(mad_agent->device->dma_device,
 991                                  pci_unmap_addr(mad_send_wr, payload_mapping),
 992                                  sge[1].length, DMA_TO_DEVICE);
 993         }
 994         return ret;
 995 }
 996
 997 /*
 998  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
 999  *  with the registered client
1000  */
1001 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1002                      struct ib_mad_send_buf **bad_send_buf)
1003 {
1004         struct ib_mad_agent_private *mad_agent_priv;
1005         struct ib_mad_send_buf *next_send_buf;
1006         struct ib_mad_send_wr_private *mad_send_wr;
1007         unsigned long flags;
1008         int ret = -EINVAL;
1009
1010         /* Walk list of send WRs and post each on send list */
1011         for (; send_buf; send_buf = next_send_buf) {
1012
1013                 mad_send_wr = container_of(send_buf,
1014                                            struct ib_mad_send_wr_private,
1015                                            send_buf);
1016                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1017
1018                 if (!send_buf->mad_agent->send_handler ||
1019                     (send_buf->timeout_ms &&
1020                      !send_buf->mad_agent->recv_handler)) {
1021                         ret = -EINVAL;
1022                         goto error;
1023                 }
1024
1025                 /*
1026                  * Save pointer to next work request to post in case the
1027                  * current one completes, and the user modifies the work
1028                  * request associated with the completion
1029                  */
1030                 next_send_buf = send_buf->next;
1031                 mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1032
1033                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1034                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1035                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1036                                                      mad_send_wr);
1037                         if (ret < 0)            /* error */
1038                                 goto error;
1039                         else if (ret == 1)      /* locally consumed */
1040                                 continue;
1041                 }
1042
1043                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1044                 /* Timeout will be updated after send completes */
1045                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1046                 mad_send_wr->retries = send_buf->retries;
1047                 /* Reference for work request to QP + response */
1048                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1049                 mad_send_wr->status = IB_WC_SUCCESS;
1050
1051                 /* Reference MAD agent until send completes */
1052                 atomic_inc(&mad_agent_priv->refcount);
1053                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1054                 list_add_tail(&mad_send_wr->agent_list,
1055                               &mad_agent_priv->send_list);
1056                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1057
1058                 if (mad_agent_priv->agent.rmpp_version) {
1059                         ret = ib_send_rmpp_mad(mad_send_wr);
1060                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1061                                 ret = ib_send_mad(mad_send_wr);
1062                 } else
1063                         ret = ib_send_mad(mad_send_wr);
1064                 if (ret < 0) {
1065                         /* Fail send request */
1066                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1067                         list_del(&mad_send_wr->agent_list);
1068                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1069                         atomic_dec(&mad_agent_priv->refcount);
1070                         goto error;
1071                 }
1072         }
1073         return 0;
1074 error:
1075         if (bad_send_buf)
1076                 *bad_send_buf = send_buf;
1077         return ret;
1078 }
1079 EXPORT_SYMBOL(ib_post_send_mad);
1080
1081 /*
1082  * ib_free_recv_mad - Returns data buffers used to receive
1083  *  a MAD to the access layer
1084  */
1085 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1086 {
1087         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1088         struct ib_mad_private_header *mad_priv_hdr;
1089         struct ib_mad_private *priv;
1090         struct list_head free_list;
1091
1092         INIT_LIST_HEAD(&free_list);
1093         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1094
1095         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1096                                         &free_list, list) {
1097                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1098                                            recv_buf);
1099                 mad_priv_hdr = container_of(mad_recv_wc,
1100                                             struct ib_mad_private_header,
1101                                             recv_wc);
1102                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1103                                     header);
1104                 kmem_cache_free(ib_mad_cache, priv);
1105         }
1106 }
1107 EXPORT_SYMBOL(ib_free_recv_mad);
1108
1109 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1110                                         u8 rmpp_version,
1111                                         ib_mad_send_handler send_handler,
1112                                         ib_mad_recv_handler recv_handler,
1113                                         void *context)
1114 {
1115         return ERR_PTR(-EINVAL);        /* XXX: for now */
1116 }
1117 EXPORT_SYMBOL(ib_redirect_mad_qp);
1118
1119 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1120                       struct ib_wc *wc)
1121 {
1122         printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1123         return 0;
1124 }
1125 EXPORT_SYMBOL(ib_process_mad_wc);
1126
1127 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1128                          struct ib_mad_reg_req *mad_reg_req)
1129 {
1130         int i;
1131
1132         for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1133              i < IB_MGMT_MAX_METHODS;
1134              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1135                                1+i)) {
1136                 if ((*method)->agent[i]) {
1137                         printk(KERN_ERR PFX "Method %d already in use\n", i);
1138                         return -EINVAL;
1139                 }
1140         }
1141         return 0;
1142 }
1143
1144 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1145 {
1146         /* Allocate management method table */
1147         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1148         if (!*method) {
1149                 printk(KERN_ERR PFX "No memory for "
1150                        "ib_mad_mgmt_method_table\n");
1151                 return -ENOMEM;
1152         }
1153
1154         return 0;
1155 }
1156
1157 /*
1158  * Check to see if there are any methods still in use
1159  */
1160 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1161 {
1162         int i;
1163
1164         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1165                 if (method->agent[i])
1166                         return 1;
1167         return 0;
1168 }
1169
1170 /*
1171  * Check to see if there are any method tables for this class still in use
1172  */
1173 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1174 {
1175         int i;
1176
1177         for (i = 0; i < MAX_MGMT_CLASS; i++)
1178                 if (class->method_table[i])
1179                         return 1;
1180         return 0;
1181 }
1182
1183 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1184 {
1185         int i;
1186
1187         for (i = 0; i < MAX_MGMT_OUI; i++)
1188                 if (vendor_class->method_table[i])
1189                         return 1;
1190         return 0;
1191 }
1192
1193 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1194                            char *oui)
1195 {
1196         int i;
1197
1198         for (i = 0; i < MAX_MGMT_OUI; i++)
1199                 /* Is there matching OUI for this vendor class ? */
1200                 if (!memcmp(vendor_class->oui[i], oui, 3))
1201                         return i;
1202
1203         return -1;
1204 }
1205
1206 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1207 {
1208         int i;
1209
1210         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1211                 if (vendor->vendor_class[i])
1212                         return 1;
1213
1214         return 0;
1215 }
1216
1217 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1218                                      struct ib_mad_agent_private *agent)
1219 {
1220         int i;
1221
1222         /* Remove any methods for this mad agent */
1223         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1224                 if (method->agent[i] == agent) {
1225                         method->agent[i] = NULL;
1226                 }
1227         }
1228 }
1229
1230 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1231                               struct ib_mad_agent_private *agent_priv,
1232                               u8 mgmt_class)
1233 {
1234         struct ib_mad_port_private *port_priv;
1235         struct ib_mad_mgmt_class_table **class;
1236         struct ib_mad_mgmt_method_table **method;
1237         int i, ret;
1238
1239         port_priv = agent_priv->qp_info->port_priv;
1240         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1241         if (!*class) {
1242                 /* Allocate management class table for "new" class version */
1243                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1244                 if (!*class) {
1245                         printk(KERN_ERR PFX "No memory for "
1246                                "ib_mad_mgmt_class_table\n");
1247                         ret = -ENOMEM;
1248                         goto error1;
1249                 }
1250
1251                 /* Allocate method table for this management class */
1252                 method = &(*class)->method_table[mgmt_class];
1253                 if ((ret = allocate_method_table(method)))
1254                         goto error2;
1255         } else {
1256                 method = &(*class)->method_table[mgmt_class];
1257                 if (!*method) {
1258                         /* Allocate method table for this management class */
1259                         if ((ret = allocate_method_table(method)))
1260                                 goto error1;
1261                 }
1262         }
1263
1264         /* Now, make sure methods are not already in use */
1265         if (method_in_use(method, mad_reg_req))
1266                 goto error3;
1267
1268         /* Finally, add in methods being registered */
1269         for (i = find_first_bit(mad_reg_req->method_mask,
1270                                 IB_MGMT_MAX_METHODS);
1271              i < IB_MGMT_MAX_METHODS;
1272              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1273                                1+i)) {
1274                 (*method)->agent[i] = agent_priv;
1275         }
1276         return 0;
1277
1278 error3:
1279         /* Remove any methods for this mad agent */
1280         remove_methods_mad_agent(*method, agent_priv);
1281         /* Now, check to see if there are any methods in use */
1282         if (!check_method_table(*method)) {
1283                 /* If not, release management method table */
1284                 kfree(*method);
1285                 *method = NULL;
1286         }
1287         ret = -EINVAL;
1288         goto error1;
1289 error2:
1290         kfree(*class);
1291         *class = NULL;
1292 error1:
1293         return ret;
1294 }
1295
1296 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1297                            struct ib_mad_agent_private *agent_priv)
1298 {
1299         struct ib_mad_port_private *port_priv;
1300         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1301         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1302         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1303         struct ib_mad_mgmt_method_table **method;
1304         int i, ret = -ENOMEM;
1305         u8 vclass;
1306
1307         /* "New" vendor (with OUI) class */
1308         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1309         port_priv = agent_priv->qp_info->port_priv;
1310         vendor_table = &port_priv->version[
1311                                 mad_reg_req->mgmt_class_version].vendor;
1312         if (!*vendor_table) {
1313                 /* Allocate mgmt vendor class table for "new" class version */
1314                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1315                 if (!vendor) {
1316                         printk(KERN_ERR PFX "No memory for "
1317                                "ib_mad_mgmt_vendor_class_table\n");
1318                         goto error1;
1319                 }
1320
1321                 *vendor_table = vendor;
1322         }
1323         if (!(*vendor_table)->vendor_class[vclass]) {
1324                 /* Allocate table for this management vendor class */
1325                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1326                 if (!vendor_class) {
1327                         printk(KERN_ERR PFX "No memory for "
1328                                "ib_mad_mgmt_vendor_class\n");
1329                         goto error2;
1330                 }
1331
1332                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1333         }
1334         for (i = 0; i < MAX_MGMT_OUI; i++) {
1335                 /* Is there matching OUI for this vendor class ? */
1336                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1337                             mad_reg_req->oui, 3)) {
1338                         method = &(*vendor_table)->vendor_class[
1339                                                 vclass]->method_table[i];
1340                         BUG_ON(!*method);
1341                         goto check_in_use;
1342                 }
1343         }
1344         for (i = 0; i < MAX_MGMT_OUI; i++) {
1345                 /* OUI slot available ? */
1346                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1347                                 vclass]->oui[i])) {
1348                         method = &(*vendor_table)->vendor_class[
1349                                 vclass]->method_table[i];
1350                         BUG_ON(*method);
1351                         /* Allocate method table for this OUI */
1352                         if ((ret = allocate_method_table(method)))
1353                                 goto error3;
1354                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1355                                mad_reg_req->oui, 3);
1356                         goto check_in_use;
1357                 }
1358         }
1359         printk(KERN_ERR PFX "All OUI slots in use\n");
1360         goto error3;
1361
1362 check_in_use:
1363         /* Now, make sure methods are not already in use */
1364         if (method_in_use(method, mad_reg_req))
1365                 goto error4;
1366
1367         /* Finally, add in methods being registered */
1368         for (i = find_first_bit(mad_reg_req->method_mask,
1369                                 IB_MGMT_MAX_METHODS);
1370              i < IB_MGMT_MAX_METHODS;
1371              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1372                                1+i)) {
1373                 (*method)->agent[i] = agent_priv;
1374         }
1375         return 0;
1376
1377 error4:
1378         /* Remove any methods for this mad agent */
1379         remove_methods_mad_agent(*method, agent_priv);
1380         /* Now, check to see if there are any methods in use */
1381         if (!check_method_table(*method)) {
1382                 /* If not, release management method table */
1383                 kfree(*method);
1384                 *method = NULL;
1385         }
1386         ret = -EINVAL;
1387 error3:
1388         if (vendor_class) {
1389                 (*vendor_table)->vendor_class[vclass] = NULL;
1390                 kfree(vendor_class);
1391         }
1392 error2:
1393         if (vendor) {
1394                 *vendor_table = NULL;
1395                 kfree(vendor);
1396         }
1397 error1:
1398         return ret;
1399 }
1400
1401 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1402 {
1403         struct ib_mad_port_private *port_priv;
1404         struct ib_mad_mgmt_class_table *class;
1405         struct ib_mad_mgmt_method_table *method;
1406         struct ib_mad_mgmt_vendor_class_table *vendor;
1407         struct ib_mad_mgmt_vendor_class *vendor_class;
1408         int index;
1409         u8 mgmt_class;
1410
1411         /*
1412          * Was MAD registration request supplied
1413          * with original registration ?
1414          */
1415         if (!agent_priv->reg_req) {
1416                 goto out;
1417         }
1418
1419         port_priv = agent_priv->qp_info->port_priv;
1420         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1421         class = port_priv->version[
1422                         agent_priv->reg_req->mgmt_class_version].class;
1423         if (!class)
1424                 goto vendor_check;
1425
1426         method = class->method_table[mgmt_class];
1427         if (method) {
1428                 /* Remove any methods for this mad agent */
1429                 remove_methods_mad_agent(method, agent_priv);
1430                 /* Now, check to see if there are any methods still in use */
1431                 if (!check_method_table(method)) {
1432                         /* If not, release management method table */
1433                          kfree(method);
1434                          class->method_table[mgmt_class] = NULL;
1435                          /* Any management classes left ? */
1436                         if (!check_class_table(class)) {
1437                                 /* If not, release management class table */
1438                                 kfree(class);
1439                                 port_priv->version[
1440                                         agent_priv->reg_req->
1441                                         mgmt_class_version].class = NULL;
1442                         }
1443                 }
1444         }
1445
1446 vendor_check:
1447         if (!is_vendor_class(mgmt_class))
1448                 goto out;
1449
1450         /* normalize mgmt_class to vendor range 2 */
1451         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1452         vendor = port_priv->version[
1453                         agent_priv->reg_req->mgmt_class_version].vendor;
1454
1455         if (!vendor)
1456                 goto out;
1457
1458         vendor_class = vendor->vendor_class[mgmt_class];
1459         if (vendor_class) {
1460                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1461                 if (index < 0)
1462                         goto out;
1463                 method = vendor_class->method_table[index];
1464                 if (method) {
1465                         /* Remove any methods for this mad agent */
1466                         remove_methods_mad_agent(method, agent_priv);
1467                         /*
1468                          * Now, check to see if there are
1469                          * any methods still in use
1470                          */
1471                         if (!check_method_table(method)) {
1472                                 /* If not, release management method table */
1473                                 kfree(method);
1474                                 vendor_class->method_table[index] = NULL;
1475                                 memset(vendor_class->oui[index], 0, 3);
1476                                 /* Any OUIs left ? */
1477                                 if (!check_vendor_class(vendor_class)) {
1478                                         /* If not, release vendor class table */
1479                                         kfree(vendor_class);
1480                                         vendor->vendor_class[mgmt_class] = NULL;
1481                                         /* Any other vendor classes left ? */
1482                                         if (!check_vendor_table(vendor)) {
1483                                                 kfree(vendor);
1484                                                 port_priv->version[
1485                                                         agent_priv->reg_req->
1486                                                         mgmt_class_version].
1487                                                         vendor = NULL;
1488                                         }
1489                                 }
1490                         }
1491                 }
1492         }
1493
1494 out:
1495         return;
1496 }
1497
1498 static struct ib_mad_agent_private *
1499 find_mad_agent(struct ib_mad_port_private *port_priv,
1500                struct ib_mad *mad)
1501 {
1502         struct ib_mad_agent_private *mad_agent = NULL;
1503         unsigned long flags;
1504
1505         spin_lock_irqsave(&port_priv->reg_lock, flags);
1506         if (response_mad(mad)) {
1507                 u32 hi_tid;
1508                 struct ib_mad_agent_private *entry;
1509
1510                 /*
1511                  * Routing is based on high 32 bits of transaction ID
1512                  * of MAD.
1513                  */
1514                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1515                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1516                         if (entry->agent.hi_tid == hi_tid) {
1517                                 mad_agent = entry;
1518                                 break;
1519                         }
1520                 }
1521         } else {
1522                 struct ib_mad_mgmt_class_table *class;
1523                 struct ib_mad_mgmt_method_table *method;
1524                 struct ib_mad_mgmt_vendor_class_table *vendor;
1525                 struct ib_mad_mgmt_vendor_class *vendor_class;
1526                 struct ib_vendor_mad *vendor_mad;
1527                 int index;
1528
1529                 /*
1530                  * Routing is based on version, class, and method
1531                  * For "newer" vendor MADs, also based on OUI
1532                  */
1533                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1534                         goto out;
1535                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1536                         class = port_priv->version[
1537                                         mad->mad_hdr.class_version].class;
1538                         if (!class)
1539                                 goto out;
1540                         method = class->method_table[convert_mgmt_class(
1541                                                         mad->mad_hdr.mgmt_class)];
1542                         if (method)
1543                                 mad_agent = method->agent[mad->mad_hdr.method &
1544                                                           ~IB_MGMT_METHOD_RESP];
1545                 } else {
1546                         vendor = port_priv->version[
1547                                         mad->mad_hdr.class_version].vendor;
1548                         if (!vendor)
1549                                 goto out;
1550                         vendor_class = vendor->vendor_class[vendor_class_index(
1551                                                 mad->mad_hdr.mgmt_class)];
1552                         if (!vendor_class)
1553                                 goto out;
1554                         /* Find matching OUI */
1555                         vendor_mad = (struct ib_vendor_mad *)mad;
1556                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1557                         if (index == -1)
1558                                 goto out;
1559                         method = vendor_class->method_table[index];
1560                         if (method) {
1561                                 mad_agent = method->agent[mad->mad_hdr.method &
1562                                                           ~IB_MGMT_METHOD_RESP];
1563                         }
1564                 }
1565         }
1566
1567         if (mad_agent) {
1568                 if (mad_agent->agent.recv_handler)
1569                         atomic_inc(&mad_agent->refcount);
1570                 else {
1571                         printk(KERN_NOTICE PFX "No receive handler for client "
1572                                "%p on port %d\n",
1573                                &mad_agent->agent, port_priv->port_num);
1574                         mad_agent = NULL;
1575                 }
1576         }
1577 out:
1578         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1579
1580         return mad_agent;
1581 }
1582
1583 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1584 {
1585         int valid = 0;
1586
1587         /* Make sure MAD base version is understood */
1588         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1589                 printk(KERN_ERR PFX "MAD received with unsupported base "
1590                        "version %d\n", mad->mad_hdr.base_version);
1591                 goto out;
1592         }
1593
1594         /* Filter SMI packets sent to other than QP0 */
1595         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1596             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1597                 if (qp_num == 0)
1598                         valid = 1;
1599         } else {
1600                 /* Filter GSI packets sent to QP0 */
1601                 if (qp_num != 0)
1602                         valid = 1;
1603         }
1604
1605 out:
1606         return valid;
1607 }
1608
1609 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1610                        struct ib_mad_hdr *mad_hdr)
1611 {
1612         struct ib_rmpp_mad *rmpp_mad;
1613
1614         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1615         return !mad_agent_priv->agent.rmpp_version ||
1616                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1617                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1618                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1619 }
1620
1621 struct ib_mad_send_wr_private*
1622 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv, __be64 tid)
1623 {
1624         struct ib_mad_send_wr_private *mad_send_wr;
1625
1626         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
1627                             agent_list) {
1628                 if (mad_send_wr->tid == tid)
1629                         return mad_send_wr;
1630         }
1631
1632         /*
1633          * It's possible to receive the response before we've
1634          * been notified that the send has completed
1635          */
1636         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
1637                             agent_list) {
1638                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
1639                     mad_send_wr->tid == tid && mad_send_wr->timeout) {
1640                         /* Verify request has not been canceled */
1641                         return (mad_send_wr->status == IB_WC_SUCCESS) ?
1642                                 mad_send_wr : NULL;
1643                 }
1644         }
1645         return NULL;
1646 }
1647
1648 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1649 {
1650         mad_send_wr->timeout = 0;
1651         if (mad_send_wr->refcount == 1) {
1652                 list_del(&mad_send_wr->agent_list);
1653                 list_add_tail(&mad_send_wr->agent_list,
1654                               &mad_send_wr->mad_agent_priv->done_list);
1655         }
1656 }
1657
1658 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1659                                  struct ib_mad_recv_wc *mad_recv_wc)
1660 {
1661         struct ib_mad_send_wr_private *mad_send_wr;
1662         struct ib_mad_send_wc mad_send_wc;
1663         unsigned long flags;
1664         __be64 tid;
1665
1666         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1667         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1668         if (mad_agent_priv->agent.rmpp_version) {
1669                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1670                                                       mad_recv_wc);
1671                 if (!mad_recv_wc) {
1672                         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1673                                 wake_up(&mad_agent_priv->wait);
1674                         return;
1675                 }
1676         }
1677
1678         /* Complete corresponding request */
1679         if (response_mad(mad_recv_wc->recv_buf.mad)) {
1680                 tid = mad_recv_wc->recv_buf.mad->mad_hdr.tid;
1681                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1682                 mad_send_wr = ib_find_send_mad(mad_agent_priv, tid);
1683                 if (!mad_send_wr) {
1684                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1685                         ib_free_recv_mad(mad_recv_wc);
1686                         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1687                                 wake_up(&mad_agent_priv->wait);
1688                         return;
1689                 }
1690                 ib_mark_mad_done(mad_send_wr);
1691                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1692
1693                 /* Defined behavior is to complete response before request */
1694                 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1695                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1696                                                    mad_recv_wc);
1697                 atomic_dec(&mad_agent_priv->refcount);
1698
1699                 mad_send_wc.status = IB_WC_SUCCESS;
1700                 mad_send_wc.vendor_err = 0;
1701                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1702                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1703         } else {
1704                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1705                                                    mad_recv_wc);
1706                 if (atomic_dec_and_test(&mad_agent_priv->refcount))
1707                         wake_up(&mad_agent_priv->wait);
1708         }
1709 }
1710
1711 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1712                                      struct ib_wc *wc)
1713 {
1714         struct ib_mad_qp_info *qp_info;
1715         struct ib_mad_private_header *mad_priv_hdr;
1716         struct ib_mad_private *recv, *response;
1717         struct ib_mad_list_head *mad_list;
1718         struct ib_mad_agent_private *mad_agent;
1719
1720         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1721         if (!response)
1722                 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1723                        "for response buffer\n");
1724
1725         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1726         qp_info = mad_list->mad_queue->qp_info;
1727         dequeue_mad(mad_list);
1728
1729         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1730                                     mad_list);
1731         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1732         dma_unmap_single(port_priv->device->dma_device,
1733                          pci_unmap_addr(&recv->header, mapping),
1734                          sizeof(struct ib_mad_private) -
1735                          sizeof(struct ib_mad_private_header),
1736                          DMA_FROM_DEVICE);
1737
1738         /* Setup MAD receive work completion from "normal" work completion */
1739         recv->header.wc = *wc;
1740         recv->header.recv_wc.wc = &recv->header.wc;
1741         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1742         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1743         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1744
1745         if (atomic_read(&qp_info->snoop_count))
1746                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1747
1748         /* Validate MAD */
1749         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1750                 goto out;
1751
1752         if (recv->mad.mad.mad_hdr.mgmt_class ==
1753             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1754                 if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1755                                             port_priv->device->node_type,
1756                                             port_priv->port_num,
1757                                             port_priv->device->phys_port_cnt))
1758                         goto out;
1759                 if (!smi_check_forward_dr_smp(&recv->mad.smp))
1760                         goto local;
1761                 if (!smi_handle_dr_smp_send(&recv->mad.smp,
1762                                             port_priv->device->node_type,
1763                                             port_priv->port_num))
1764                         goto out;
1765                 if (!smi_check_local_smp(&recv->mad.smp, port_priv->device))
1766                         goto out;
1767         }
1768
1769 local:
1770         /* Give driver "right of first refusal" on incoming MAD */
1771         if (port_priv->device->process_mad) {
1772                 int ret;
1773
1774                 if (!response) {
1775                         printk(KERN_ERR PFX "No memory for response MAD\n");
1776                         /*
1777                          * Is it better to assume that
1778                          * it wouldn't be processed ?
1779                          */
1780                         goto out;
1781                 }
1782
1783                 ret = port_priv->device->process_mad(port_priv->device, 0,
1784                                                      port_priv->port_num,
1785                                                      wc, &recv->grh,
1786                                                      &recv->mad.mad,
1787                                                      &response->mad.mad);
1788                 if (ret & IB_MAD_RESULT_SUCCESS) {
1789                         if (ret & IB_MAD_RESULT_CONSUMED)
1790                                 goto out;
1791                         if (ret & IB_MAD_RESULT_REPLY) {
1792                                 agent_send_response(&response->mad.mad,
1793                                                     &recv->grh, wc,
1794                                                     port_priv->device,
1795                                                     port_priv->port_num,
1796                                                     qp_info->qp->qp_num);
1797                                 goto out;
1798                         }
1799                 }
1800         }
1801
1802         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1803         if (mad_agent) {
1804                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1805                 /*
1806                  * recv is freed up in error cases in ib_mad_complete_recv
1807                  * or via recv_handler in ib_mad_complete_recv()
1808                  */
1809                 recv = NULL;
1810         }
1811
1812 out:
1813         /* Post another receive request for this QP */
1814         if (response) {
1815                 ib_mad_post_receive_mads(qp_info, response);
1816                 if (recv)
1817                         kmem_cache_free(ib_mad_cache, recv);
1818         } else
1819                 ib_mad_post_receive_mads(qp_info, recv);
1820 }
1821
1822 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1823 {
1824         struct ib_mad_send_wr_private *mad_send_wr;
1825         unsigned long delay;
1826
1827         if (list_empty(&mad_agent_priv->wait_list)) {
1828                 cancel_delayed_work(&mad_agent_priv->timed_work);
1829         } else {
1830                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1831                                          struct ib_mad_send_wr_private,
1832                                          agent_list);
1833
1834                 if (time_after(mad_agent_priv->timeout,
1835                                mad_send_wr->timeout)) {
1836                         mad_agent_priv->timeout = mad_send_wr->timeout;
1837                         cancel_delayed_work(&mad_agent_priv->timed_work);
1838                         delay = mad_send_wr->timeout - jiffies;
1839                         if ((long)delay <= 0)
1840                                 delay = 1;
1841                         queue_delayed_work(mad_agent_priv->qp_info->
1842                                            port_priv->wq,
1843                                            &mad_agent_priv->timed_work, delay);
1844                 }
1845         }
1846 }
1847
1848 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1849 {
1850         struct ib_mad_agent_private *mad_agent_priv;
1851         struct ib_mad_send_wr_private *temp_mad_send_wr;
1852         struct list_head *list_item;
1853         unsigned long delay;
1854
1855         mad_agent_priv = mad_send_wr->mad_agent_priv;
1856         list_del(&mad_send_wr->agent_list);
1857
1858         delay = mad_send_wr->timeout;
1859         mad_send_wr->timeout += jiffies;
1860
1861         if (delay) {
1862                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1863                         temp_mad_send_wr = list_entry(list_item,
1864                                                 struct ib_mad_send_wr_private,
1865                                                 agent_list);
1866                         if (time_after(mad_send_wr->timeout,
1867                                        temp_mad_send_wr->timeout))
1868                                 break;
1869                 }
1870         }
1871         else
1872                 list_item = &mad_agent_priv->wait_list;
1873         list_add(&mad_send_wr->agent_list, list_item);
1874
1875         /* Reschedule a work item if we have a shorter timeout */
1876         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1877                 cancel_delayed_work(&mad_agent_priv->timed_work);
1878                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1879                                    &mad_agent_priv->timed_work, delay);
1880         }
1881 }
1882
1883 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
1884                           int timeout_ms)
1885 {
1886         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
1887         wait_for_response(mad_send_wr);
1888 }
1889
1890 /*
1891  * Process a send work completion
1892  */
1893 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
1894                              struct ib_mad_send_wc *mad_send_wc)
1895 {
1896         struct ib_mad_agent_private     *mad_agent_priv;
1897         unsigned long                   flags;
1898         int                             ret;
1899
1900         mad_agent_priv = mad_send_wr->mad_agent_priv;
1901         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1902         if (mad_agent_priv->agent.rmpp_version) {
1903                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
1904                 if (ret == IB_RMPP_RESULT_CONSUMED)
1905                         goto done;
1906         } else
1907                 ret = IB_RMPP_RESULT_UNHANDLED;
1908
1909         if (mad_send_wc->status != IB_WC_SUCCESS &&
1910             mad_send_wr->status == IB_WC_SUCCESS) {
1911                 mad_send_wr->status = mad_send_wc->status;
1912                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
1913         }
1914
1915         if (--mad_send_wr->refcount > 0) {
1916                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
1917                     mad_send_wr->status == IB_WC_SUCCESS) {
1918                         wait_for_response(mad_send_wr);
1919                 }
1920                 goto done;
1921         }
1922
1923         /* Remove send from MAD agent and notify client of completion */
1924         list_del(&mad_send_wr->agent_list);
1925         adjust_timeout(mad_agent_priv);
1926         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1927
1928         if (mad_send_wr->status != IB_WC_SUCCESS )
1929                 mad_send_wc->status = mad_send_wr->status;
1930         if (ret == IB_RMPP_RESULT_INTERNAL)
1931                 ib_rmpp_send_handler(mad_send_wc);
1932         else
1933                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
1934                                                    mad_send_wc);
1935
1936         /* Release reference on agent taken when sending */
1937         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1938                 wake_up(&mad_agent_priv->wait);
1939         return;
1940 done:
1941         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1942 }
1943
1944 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
1945                                      struct ib_wc *wc)
1946 {
1947         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
1948         struct ib_mad_list_head         *mad_list;
1949         struct ib_mad_qp_info           *qp_info;
1950         struct ib_mad_queue             *send_queue;
1951         struct ib_send_wr               *bad_send_wr;
1952         struct ib_mad_send_wc           mad_send_wc;
1953         unsigned long flags;
1954         int ret;
1955
1956         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1957         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1958                                    mad_list);
1959         send_queue = mad_list->mad_queue;
1960         qp_info = send_queue->qp_info;
1961
1962 retry:
1963         dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
1964                          pci_unmap_addr(mad_send_wr, header_mapping),
1965                          mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
1966         dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
1967                          pci_unmap_addr(mad_send_wr, payload_mapping),
1968                          mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
1969         queued_send_wr = NULL;
1970         spin_lock_irqsave(&send_queue->lock, flags);
1971         list_del(&mad_list->list);
1972
1973         /* Move queued send to the send queue */
1974         if (send_queue->count-- > send_queue->max_active) {
1975                 mad_list = container_of(qp_info->overflow_list.next,
1976                                         struct ib_mad_list_head, list);
1977                 queued_send_wr = container_of(mad_list,
1978                                         struct ib_mad_send_wr_private,
1979                                         mad_list);
1980                 list_del(&mad_list->list);
1981                 list_add_tail(&mad_list->list, &send_queue->list);
1982         }
1983         spin_unlock_irqrestore(&send_queue->lock, flags);
1984
1985         mad_send_wc.send_buf = &mad_send_wr->send_buf;
1986         mad_send_wc.status = wc->status;
1987         mad_send_wc.vendor_err = wc->vendor_err;
1988         if (atomic_read(&qp_info->snoop_count))
1989                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
1990                            IB_MAD_SNOOP_SEND_COMPLETIONS);
1991         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1992
1993         if (queued_send_wr) {
1994                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
1995                                    &bad_send_wr);
1996                 if (ret) {
1997                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
1998                         mad_send_wr = queued_send_wr;
1999                         wc->status = IB_WC_LOC_QP_OP_ERR;
2000                         goto retry;
2001                 }
2002         }
2003 }
2004
2005 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2006 {
2007         struct ib_mad_send_wr_private *mad_send_wr;
2008         struct ib_mad_list_head *mad_list;
2009         unsigned long flags;
2010
2011         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2012         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2013                 mad_send_wr = container_of(mad_list,
2014                                            struct ib_mad_send_wr_private,
2015                                            mad_list);
2016                 mad_send_wr->retry = 1;
2017         }
2018         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2019 }
2020
2021 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2022                               struct ib_wc *wc)
2023 {
2024         struct ib_mad_list_head *mad_list;
2025         struct ib_mad_qp_info *qp_info;
2026         struct ib_mad_send_wr_private *mad_send_wr;
2027         int ret;
2028
2029         /* Determine if failure was a send or receive */
2030         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2031         qp_info = mad_list->mad_queue->qp_info;
2032         if (mad_list->mad_queue == &qp_info->recv_queue)
2033                 /*
2034                  * Receive errors indicate that the QP has entered the error
2035                  * state - error handling/shutdown code will cleanup
2036                  */
2037                 return;
2038
2039         /*
2040          * Send errors will transition the QP to SQE - move
2041          * QP to RTS and repost flushed work requests
2042          */
2043         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2044                                    mad_list);
2045         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2046                 if (mad_send_wr->retry) {
2047                         /* Repost send */
2048                         struct ib_send_wr *bad_send_wr;
2049
2050                         mad_send_wr->retry = 0;
2051                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2052                                         &bad_send_wr);
2053                         if (ret)
2054                                 ib_mad_send_done_handler(port_priv, wc);
2055                 } else
2056                         ib_mad_send_done_handler(port_priv, wc);
2057         } else {
2058                 struct ib_qp_attr *attr;
2059
2060                 /* Transition QP to RTS and fail offending send */
2061                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2062                 if (attr) {
2063                         attr->qp_state = IB_QPS_RTS;
2064                         attr->cur_qp_state = IB_QPS_SQE;
2065                         ret = ib_modify_qp(qp_info->qp, attr,
2066                                            IB_QP_STATE | IB_QP_CUR_STATE);
2067                         kfree(attr);
2068                         if (ret)
2069                                 printk(KERN_ERR PFX "mad_error_handler - "
2070                                        "ib_modify_qp to RTS : %d\n", ret);
2071                         else
2072                                 mark_sends_for_retry(qp_info);
2073                 }
2074                 ib_mad_send_done_handler(port_priv, wc);
2075         }
2076 }
2077
2078 /*
2079  * IB MAD completion callback
2080  */
2081 static void ib_mad_completion_handler(void *data)
2082 {
2083         struct ib_mad_port_private *port_priv;
2084         struct ib_wc wc;
2085
2086         port_priv = (struct ib_mad_port_private *)data;
2087         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2088
2089         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2090                 if (wc.status == IB_WC_SUCCESS) {
2091                         switch (wc.opcode) {
2092                         case IB_WC_SEND:
2093                                 ib_mad_send_done_handler(port_priv, &wc);
2094                                 break;
2095                         case IB_WC_RECV:
2096                                 ib_mad_recv_done_handler(port_priv, &wc);
2097                                 break;
2098                         default:
2099                                 BUG_ON(1);
2100                                 break;
2101                         }
2102                 } else
2103                         mad_error_handler(port_priv, &wc);
2104         }
2105 }
2106
2107 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2108 {
2109         unsigned long flags;
2110         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2111         struct ib_mad_send_wc mad_send_wc;
2112         struct list_head cancel_list;
2113
2114         INIT_LIST_HEAD(&cancel_list);
2115
2116         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2117         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2118                                  &mad_agent_priv->send_list, agent_list) {
2119                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2120                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2121                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2122                 }
2123         }
2124
2125         /* Empty wait list to prevent receives from finding a request */
2126         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2127         /* Empty local completion list as well */
2128         list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2129         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2130
2131         /* Report all cancelled requests */
2132         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2133         mad_send_wc.vendor_err = 0;
2134
2135         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2136                                  &cancel_list, agent_list) {
2137                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2138                 list_del(&mad_send_wr->agent_list);
2139                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2140                                                    &mad_send_wc);
2141                 atomic_dec(&mad_agent_priv->refcount);
2142         }
2143 }
2144
2145 static struct ib_mad_send_wr_private*
2146 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2147              struct ib_mad_send_buf *send_buf)
2148 {
2149         struct ib_mad_send_wr_private *mad_send_wr;
2150
2151         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2152                             agent_list) {
2153                 if (&mad_send_wr->send_buf == send_buf)
2154                         return mad_send_wr;
2155         }
2156
2157         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2158                             agent_list) {
2159                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2160                     &mad_send_wr->send_buf == send_buf)
2161                         return mad_send_wr;
2162         }
2163         return NULL;
2164 }
2165
2166 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2167                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2168 {
2169         struct ib_mad_agent_private *mad_agent_priv;
2170         struct ib_mad_send_wr_private *mad_send_wr;
2171         unsigned long flags;
2172         int active;
2173
2174         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2175                                       agent);
2176         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2177         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2178         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2179                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2180                 return -EINVAL;
2181         }
2182
2183         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2184         if (!timeout_ms) {
2185                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2186                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2187         }
2188
2189         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2190         if (active)
2191                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2192         else
2193                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2194
2195         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2196         return 0;
2197 }
2198 EXPORT_SYMBOL(ib_modify_mad);
2199
2200 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2201                    struct ib_mad_send_buf *send_buf)
2202 {
2203         ib_modify_mad(mad_agent, send_buf, 0);
2204 }
2205 EXPORT_SYMBOL(ib_cancel_mad);
2206
2207 static void local_completions(void *data)
2208 {
2209         struct ib_mad_agent_private *mad_agent_priv;
2210         struct ib_mad_local_private *local;
2211         struct ib_mad_agent_private *recv_mad_agent;
2212         unsigned long flags;
2213         int recv = 0;
2214         struct ib_wc wc;
2215         struct ib_mad_send_wc mad_send_wc;
2216
2217         mad_agent_priv = (struct ib_mad_agent_private *)data;
2218
2219         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2220         while (!list_empty(&mad_agent_priv->local_list)) {
2221                 local = list_entry(mad_agent_priv->local_list.next,
2222                                    struct ib_mad_local_private,
2223                                    completion_list);
2224                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2225                 if (local->mad_priv) {
2226                         recv_mad_agent = local->recv_mad_agent;
2227                         if (!recv_mad_agent) {
2228                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2229                                 goto local_send_completion;
2230                         }
2231
2232                         recv = 1;
2233                         /*
2234                          * Defined behavior is to complete response
2235                          * before request
2236                          */
2237                         build_smp_wc((unsigned long) local->mad_send_wr,
2238                                      be16_to_cpu(IB_LID_PERMISSIVE),
2239                                      0, recv_mad_agent->agent.port_num, &wc);
2240
2241                         local->mad_priv->header.recv_wc.wc = &wc;
2242                         local->mad_priv->header.recv_wc.mad_len =
2243                                                 sizeof(struct ib_mad);
2244                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2245                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2246                                  &local->mad_priv->header.recv_wc.rmpp_list);
2247                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2248                         local->mad_priv->header.recv_wc.recv_buf.mad =
2249                                                 &local->mad_priv->mad.mad;
2250                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2251                                 snoop_recv(recv_mad_agent->qp_info,
2252                                           &local->mad_priv->header.recv_wc,
2253                                            IB_MAD_SNOOP_RECVS);
2254                         recv_mad_agent->agent.recv_handler(
2255                                                 &recv_mad_agent->agent,
2256                                                 &local->mad_priv->header.recv_wc);
2257                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2258                         atomic_dec(&recv_mad_agent->refcount);
2259                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2260                 }
2261
2262 local_send_completion:
2263                 /* Complete send */
2264                 mad_send_wc.status = IB_WC_SUCCESS;
2265                 mad_send_wc.vendor_err = 0;
2266                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2267                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2268                         snoop_send(mad_agent_priv->qp_info,
2269                                    &local->mad_send_wr->send_buf,
2270                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2271                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2272                                                    &mad_send_wc);
2273
2274                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2275                 list_del(&local->completion_list);
2276                 atomic_dec(&mad_agent_priv->refcount);
2277                 if (!recv)
2278                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2279                 kfree(local);
2280         }
2281         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2282 }
2283
2284 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2285 {
2286         int ret;
2287
2288         if (!mad_send_wr->retries--)
2289                 return -ETIMEDOUT;
2290
2291         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2292
2293         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2294                 ret = ib_retry_rmpp(mad_send_wr);
2295                 switch (ret) {
2296                 case IB_RMPP_RESULT_UNHANDLED:
2297                         ret = ib_send_mad(mad_send_wr);
2298                         break;
2299                 case IB_RMPP_RESULT_CONSUMED:
2300                         ret = 0;
2301                         break;
2302                 default:
2303                         ret = -ECOMM;
2304                         break;
2305                 }
2306         } else
2307                 ret = ib_send_mad(mad_send_wr);
2308
2309         if (!ret) {
2310                 mad_send_wr->refcount++;
2311                 list_add_tail(&mad_send_wr->agent_list,
2312                               &mad_send_wr->mad_agent_priv->send_list);
2313         }
2314         return ret;
2315 }
2316
2317 static void timeout_sends(void *data)
2318 {
2319         struct ib_mad_agent_private *mad_agent_priv;
2320         struct ib_mad_send_wr_private *mad_send_wr;
2321         struct ib_mad_send_wc mad_send_wc;
2322         unsigned long flags, delay;
2323
2324         mad_agent_priv = (struct ib_mad_agent_private *)data;
2325         mad_send_wc.vendor_err = 0;
2326
2327         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2328         while (!list_empty(&mad_agent_priv->wait_list)) {
2329                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2330                                          struct ib_mad_send_wr_private,
2331                                          agent_list);
2332
2333                 if (time_after(mad_send_wr->timeout, jiffies)) {
2334                         delay = mad_send_wr->timeout - jiffies;
2335                         if ((long)delay <= 0)
2336                                 delay = 1;
2337                         queue_delayed_work(mad_agent_priv->qp_info->
2338                                            port_priv->wq,
2339                                            &mad_agent_priv->timed_work, delay);
2340                         break;
2341                 }
2342
2343                 list_del(&mad_send_wr->agent_list);
2344                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2345                     !retry_send(mad_send_wr))
2346                         continue;
2347
2348                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2349
2350                 if (mad_send_wr->status == IB_WC_SUCCESS)
2351                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2352                 else
2353                         mad_send_wc.status = mad_send_wr->status;
2354                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2355                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2356                                                    &mad_send_wc);
2357
2358                 atomic_dec(&mad_agent_priv->refcount);
2359                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2360         }
2361         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2362 }
2363
2364 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2365 {
2366         struct ib_mad_port_private *port_priv = cq->cq_context;
2367         unsigned long flags;
2368
2369         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2370         if (!list_empty(&port_priv->port_list))
2371                 queue_work(port_priv->wq, &port_priv->work);
2372         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2373 }
2374
2375 /*
2376  * Allocate receive MADs and post receive WRs for them
2377  */
2378 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2379                                     struct ib_mad_private *mad)
2380 {
2381         unsigned long flags;
2382         int post, ret;
2383         struct ib_mad_private *mad_priv;
2384         struct ib_sge sg_list;
2385         struct ib_recv_wr recv_wr, *bad_recv_wr;
2386         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2387
2388         /* Initialize common scatter list fields */
2389         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2390         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2391
2392         /* Initialize common receive WR fields */
2393         recv_wr.next = NULL;
2394         recv_wr.sg_list = &sg_list;
2395         recv_wr.num_sge = 1;
2396
2397         do {
2398                 /* Allocate and map receive buffer */
2399                 if (mad) {
2400                         mad_priv = mad;
2401                         mad = NULL;
2402                 } else {
2403                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2404                         if (!mad_priv) {
2405                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2406                                 ret = -ENOMEM;
2407                                 break;
2408                         }
2409                 }
2410                 sg_list.addr = dma_map_single(qp_info->port_priv->
2411                                                 device->dma_device,
2412                                         &mad_priv->grh,
2413                                         sizeof *mad_priv -
2414                                                 sizeof mad_priv->header,
2415                                         DMA_FROM_DEVICE);
2416                 pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2417                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2418                 mad_priv->header.mad_list.mad_queue = recv_queue;
2419
2420                 /* Post receive WR */
2421                 spin_lock_irqsave(&recv_queue->lock, flags);
2422                 post = (++recv_queue->count < recv_queue->max_active);
2423                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2424                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2425                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2426                 if (ret) {
2427                         spin_lock_irqsave(&recv_queue->lock, flags);
2428                         list_del(&mad_priv->header.mad_list.list);
2429                         recv_queue->count--;
2430                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2431                         dma_unmap_single(qp_info->port_priv->device->dma_device,
2432                                          pci_unmap_addr(&mad_priv->header,
2433                                                         mapping),
2434                                          sizeof *mad_priv -
2435                                            sizeof mad_priv->header,
2436                                          DMA_FROM_DEVICE);
2437                         kmem_cache_free(ib_mad_cache, mad_priv);
2438                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2439                         break;
2440                 }
2441         } while (post);
2442
2443         return ret;
2444 }
2445
2446 /*
2447  * Return all the posted receive MADs
2448  */
2449 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2450 {
2451         struct ib_mad_private_header *mad_priv_hdr;
2452         struct ib_mad_private *recv;
2453         struct ib_mad_list_head *mad_list;
2454
2455         while (!list_empty(&qp_info->recv_queue.list)) {
2456
2457                 mad_list = list_entry(qp_info->recv_queue.list.next,
2458                                       struct ib_mad_list_head, list);
2459                 mad_priv_hdr = container_of(mad_list,
2460                                             struct ib_mad_private_header,
2461                                             mad_list);
2462                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2463                                     header);
2464
2465                 /* Remove from posted receive MAD list */
2466                 list_del(&mad_list->list);
2467
2468                 dma_unmap_single(qp_info->port_priv->device->dma_device,
2469                                  pci_unmap_addr(&recv->header, mapping),
2470                                  sizeof(struct ib_mad_private) -
2471                                  sizeof(struct ib_mad_private_header),
2472                                  DMA_FROM_DEVICE);
2473                 kmem_cache_free(ib_mad_cache, recv);
2474         }
2475
2476         qp_info->recv_queue.count = 0;
2477 }
2478
2479 /*
2480  * Start the port
2481  */
2482 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2483 {
2484         int ret, i;
2485         struct ib_qp_attr *attr;
2486         struct ib_qp *qp;
2487
2488         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2489         if (!attr) {
2490                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2491                 return -ENOMEM;
2492         }
2493
2494         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2495                 qp = port_priv->qp_info[i].qp;
2496                 /*
2497                  * PKey index for QP1 is irrelevant but
2498                  * one is needed for the Reset to Init transition
2499                  */
2500                 attr->qp_state = IB_QPS_INIT;
2501                 attr->pkey_index = 0;
2502                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2503                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2504                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2505                 if (ret) {
2506                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2507                                "INIT: %d\n", i, ret);
2508                         goto out;
2509                 }
2510
2511                 attr->qp_state = IB_QPS_RTR;
2512                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2513                 if (ret) {
2514                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2515                                "RTR: %d\n", i, ret);
2516                         goto out;
2517                 }
2518
2519                 attr->qp_state = IB_QPS_RTS;
2520                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2521                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2522                 if (ret) {
2523                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2524                                "RTS: %d\n", i, ret);
2525                         goto out;
2526                 }
2527         }
2528
2529         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2530         if (ret) {
2531                 printk(KERN_ERR PFX "Failed to request completion "
2532                        "notification: %d\n", ret);
2533                 goto out;
2534         }
2535
2536         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2537                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2538                 if (ret) {
2539                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2540                         goto out;
2541                 }
2542         }
2543 out:
2544         kfree(attr);
2545         return ret;
2546 }
2547
2548 static void qp_event_handler(struct ib_event *event, void *qp_context)
2549 {
2550         struct ib_mad_qp_info   *qp_info = qp_context;
2551
2552         /* It's worse than that! He's dead, Jim! */
2553         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2554                 event->event, qp_info->qp->qp_num);
2555 }
2556
2557 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2558                            struct ib_mad_queue *mad_queue)
2559 {
2560         mad_queue->qp_info = qp_info;
2561         mad_queue->count = 0;
2562         spin_lock_init(&mad_queue->lock);
2563         INIT_LIST_HEAD(&mad_queue->list);
2564 }
2565
2566 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2567                         struct ib_mad_qp_info *qp_info)
2568 {
2569         qp_info->port_priv = port_priv;
2570         init_mad_queue(qp_info, &qp_info->send_queue);
2571         init_mad_queue(qp_info, &qp_info->recv_queue);
2572         INIT_LIST_HEAD(&qp_info->overflow_list);
2573         spin_lock_init(&qp_info->snoop_lock);
2574         qp_info->snoop_table = NULL;
2575         qp_info->snoop_table_size = 0;
2576         atomic_set(&qp_info->snoop_count, 0);
2577 }
2578
2579 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2580                          enum ib_qp_type qp_type)
2581 {
2582         struct ib_qp_init_attr  qp_init_attr;
2583         int ret;
2584
2585         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2586         qp_init_attr.send_cq = qp_info->port_priv->cq;
2587         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2588         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2589         qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2590         qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2591         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2592         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2593         qp_init_attr.qp_type = qp_type;
2594         qp_init_attr.port_num = qp_info->port_priv->port_num;
2595         qp_init_attr.qp_context = qp_info;
2596         qp_init_attr.event_handler = qp_event_handler;
2597         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2598         if (IS_ERR(qp_info->qp)) {
2599                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2600                        get_spl_qp_index(qp_type));
2601                 ret = PTR_ERR(qp_info->qp);
2602                 goto error;
2603         }
2604         /* Use minimum queue sizes unless the CQ is resized */
2605         qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2606         qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2607         return 0;
2608
2609 error:
2610         return ret;
2611 }
2612
2613 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2614 {
2615         ib_destroy_qp(qp_info->qp);
2616         kfree(qp_info->snoop_table);
2617 }
2618
2619 /*
2620  * Open the port
2621  * Create the QP, PD, MR, and CQ if needed
2622  */
2623 static int ib_mad_port_open(struct ib_device *device,
2624                             int port_num)
2625 {
2626         int ret, cq_size;
2627         struct ib_mad_port_private *port_priv;
2628         unsigned long flags;
2629         char name[sizeof "ib_mad123"];
2630
2631         /* Create new device info */
2632         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2633         if (!port_priv) {
2634                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2635                 return -ENOMEM;
2636         }
2637
2638         port_priv->device = device;
2639         port_priv->port_num = port_num;
2640         spin_lock_init(&port_priv->reg_lock);
2641         INIT_LIST_HEAD(&port_priv->agent_list);
2642         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2643         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2644
2645         cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2646         port_priv->cq = ib_create_cq(port_priv->device,
2647                                      ib_mad_thread_completion_handler,
2648                                      NULL, port_priv, cq_size);
2649         if (IS_ERR(port_priv->cq)) {
2650                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2651                 ret = PTR_ERR(port_priv->cq);
2652                 goto error3;
2653         }
2654
2655         port_priv->pd = ib_alloc_pd(device);
2656         if (IS_ERR(port_priv->pd)) {
2657                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2658                 ret = PTR_ERR(port_priv->pd);
2659                 goto error4;
2660         }
2661
2662         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2663         if (IS_ERR(port_priv->mr)) {
2664                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2665                 ret = PTR_ERR(port_priv->mr);
2666                 goto error5;
2667         }
2668
2669         ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2670         if (ret)
2671                 goto error6;
2672         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2673         if (ret)
2674                 goto error7;
2675
2676         snprintf(name, sizeof name, "ib_mad%d", port_num);
2677         port_priv->wq = create_singlethread_workqueue(name);
2678         if (!port_priv->wq) {
2679                 ret = -ENOMEM;
2680                 goto error8;
2681         }
2682         INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2683
2684         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2685         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2686         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2687
2688         ret = ib_mad_port_start(port_priv);
2689         if (ret) {
2690                 printk(KERN_ERR PFX "Couldn't start port\n");
2691                 goto error9;
2692         }
2693
2694         return 0;
2695
2696 error9:
2697         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2698         list_del_init(&port_priv->port_list);
2699         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2700
2701         destroy_workqueue(port_priv->wq);
2702 error8:
2703         destroy_mad_qp(&port_priv->qp_info[1]);
2704 error7:
2705         destroy_mad_qp(&port_priv->qp_info[0]);
2706 error6:
2707         ib_dereg_mr(port_priv->mr);
2708 error5:
2709         ib_dealloc_pd(port_priv->pd);
2710 error4:
2711         ib_destroy_cq(port_priv->cq);
2712         cleanup_recv_queue(&port_priv->qp_info[1]);
2713         cleanup_recv_queue(&port_priv->qp_info[0]);
2714 error3:
2715         kfree(port_priv);
2716
2717         return ret;
2718 }
2719
2720 /*
2721  * Close the port
2722  * If there are no classes using the port, free the port
2723  * resources (CQ, MR, PD, QP) and remove the port's info structure
2724  */
2725 static int ib_mad_port_close(struct ib_device *device, int port_num)
2726 {
2727         struct ib_mad_port_private *port_priv;
2728         unsigned long flags;
2729
2730         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2731         port_priv = __ib_get_mad_port(device, port_num);
2732         if (port_priv == NULL) {
2733                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2734                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2735                 return -ENODEV;
2736         }
2737         list_del_init(&port_priv->port_list);
2738         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2739
2740         destroy_workqueue(port_priv->wq);
2741         destroy_mad_qp(&port_priv->qp_info[1]);
2742         destroy_mad_qp(&port_priv->qp_info[0]);
2743         ib_dereg_mr(port_priv->mr);
2744         ib_dealloc_pd(port_priv->pd);
2745         ib_destroy_cq(port_priv->cq);
2746         cleanup_recv_queue(&port_priv->qp_info[1]);
2747         cleanup_recv_queue(&port_priv->qp_info[0]);
2748         /* XXX: Handle deallocation of MAD registration tables */
2749
2750         kfree(port_priv);
2751
2752         return 0;
2753 }
2754
2755 static void ib_mad_init_device(struct ib_device *device)
2756 {
2757         int start, end, i;
2758
2759         if (device->node_type == IB_NODE_SWITCH) {
2760                 start = 0;
2761                 end   = 0;
2762         } else {
2763                 start = 1;
2764                 end   = device->phys_port_cnt;
2765         }
2766
2767         for (i = start; i <= end; i++) {
2768                 if (ib_mad_port_open(device, i)) {
2769                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2770                                device->name, i);
2771                         goto error;
2772                 }
2773                 if (ib_agent_port_open(device, i)) {
2774                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2775                                "for agents\n",
2776                                device->name, i);
2777                         goto error_agent;
2778                 }
2779         }
2780         return;
2781
2782 error_agent:
2783         if (ib_mad_port_close(device, i))
2784                 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2785                        device->name, i);
2786
2787 error:
2788         i--;
2789
2790         while (i >= start) {
2791                 if (ib_agent_port_close(device, i))
2792                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2793                                "for agents\n",
2794                                device->name, i);
2795                 if (ib_mad_port_close(device, i))
2796                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2797                                device->name, i);
2798                 i--;
2799         }
2800 }
2801
2802 static void ib_mad_remove_device(struct ib_device *device)
2803 {
2804         int i, num_ports, cur_port;
2805
2806         if (device->node_type == IB_NODE_SWITCH) {
2807                 num_ports = 1;
2808                 cur_port = 0;
2809         } else {
2810                 num_ports = device->phys_port_cnt;
2811                 cur_port = 1;
2812         }
2813         for (i = 0; i < num_ports; i++, cur_port++) {
2814                 if (ib_agent_port_close(device, cur_port))
2815                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2816                                "for agents\n",
2817                                device->name, cur_port);
2818                 if (ib_mad_port_close(device, cur_port))
2819                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2820                                device->name, cur_port);
2821         }
2822 }
2823
2824 static struct ib_client mad_client = {
2825         .name   = "mad",
2826         .add = ib_mad_init_device,
2827         .remove = ib_mad_remove_device
2828 };
2829
2830 static int __init ib_mad_init_module(void)
2831 {
2832         int ret;
2833
2834         spin_lock_init(&ib_mad_port_list_lock);
2835
2836         ib_mad_cache = kmem_cache_create("ib_mad",
2837                                          sizeof(struct ib_mad_private),
2838                                          0,
2839                                          SLAB_HWCACHE_ALIGN,
2840                                          NULL,
2841                                          NULL);
2842         if (!ib_mad_cache) {
2843                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2844                 ret = -ENOMEM;
2845                 goto error1;
2846         }
2847
2848         INIT_LIST_HEAD(&ib_mad_port_list);
2849
2850         if (ib_register_client(&mad_client)) {
2851                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2852                 ret = -EINVAL;
2853                 goto error2;
2854         }
2855
2856         return 0;
2857
2858 error2:
2859         kmem_cache_destroy(ib_mad_cache);
2860 error1:
2861         return ret;
2862 }
2863
2864 static void __exit ib_mad_cleanup_module(void)
2865 {
2866         ib_unregister_client(&mad_client);
2867
2868         if (kmem_cache_destroy(ib_mad_cache)) {
2869                 printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2870         }
2871 }
2872
2873 module_init(ib_mad_init_module);
2874 module_exit(ib_mad_cleanup_module);
2875