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 2817 2005-07-07 11:29:26Z halr $
  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_dr_smp(smp, device, port_num);
 683         if (!ret || !device->process_mad)
 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_buf_length(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                 if (pad == seg_size)
 776                         pad = 0;
 777         } else
 778                 pad = seg_size;
 779         return hdr_len + data_len + pad;
 780 }
 781
 782 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 783                                             u32 remote_qpn, u16 pkey_index,
 784                                             int rmpp_active,
 785                                             int hdr_len, int data_len,
 786                                             gfp_t gfp_mask)
 787 {
 788         struct ib_mad_agent_private *mad_agent_priv;
 789         struct ib_mad_send_wr_private *mad_send_wr;
 790         int buf_size;
 791         void *buf;
 792
 793         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 794                                       agent);
 795         buf_size = get_buf_length(hdr_len, data_len);
 796
 797         if ((!mad_agent->rmpp_version &&
 798              (rmpp_active || buf_size > sizeof(struct ib_mad))) ||
 799             (!rmpp_active && buf_size > sizeof(struct ib_mad)))
 800                 return ERR_PTR(-EINVAL);
 801
 802         buf = kzalloc(sizeof *mad_send_wr + buf_size, gfp_mask);
 803         if (!buf)
 804                 return ERR_PTR(-ENOMEM);
 805
 806         mad_send_wr = buf + buf_size;
 807         mad_send_wr->send_buf.mad = buf;
 808
 809         mad_send_wr->mad_agent_priv = mad_agent_priv;
 810         mad_send_wr->sg_list[0].length = buf_size;
 811         mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
 812
 813         mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
 814         mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 815         mad_send_wr->send_wr.num_sge = 1;
 816         mad_send_wr->send_wr.opcode = IB_WR_SEND;
 817         mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
 818         mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
 819         mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 820         mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
 821
 822         if (rmpp_active) {
 823                 struct ib_rmpp_mad *rmpp_mad = mad_send_wr->send_buf.mad;
 824                 rmpp_mad->rmpp_hdr.paylen_newwin = cpu_to_be32(hdr_len -
 825                                                    IB_MGMT_RMPP_HDR + data_len);
 826                 rmpp_mad->rmpp_hdr.rmpp_version = mad_agent->rmpp_version;
 827                 rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 828                 ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr,
 829                                   IB_MGMT_RMPP_FLAG_ACTIVE);
 830         }
 831
 832         mad_send_wr->send_buf.mad_agent = mad_agent;
 833         atomic_inc(&mad_agent_priv->refcount);
 834         return &mad_send_wr->send_buf;
 835 }
 836 EXPORT_SYMBOL(ib_create_send_mad);
 837
 838 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
 839 {
 840         struct ib_mad_agent_private *mad_agent_priv;
 841
 842         mad_agent_priv = container_of(send_buf->mad_agent,
 843                                       struct ib_mad_agent_private, agent);
 844         kfree(send_buf->mad);
 845
 846         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 847                 wake_up(&mad_agent_priv->wait);
 848 }
 849 EXPORT_SYMBOL(ib_free_send_mad);
 850
 851 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 852 {
 853         struct ib_mad_qp_info *qp_info;
 854         struct list_head *list;
 855         struct ib_send_wr *bad_send_wr;
 856         struct ib_mad_agent *mad_agent;
 857         struct ib_sge *sge;
 858         unsigned long flags;
 859         int ret;
 860
 861         /* Set WR ID to find mad_send_wr upon completion */
 862         qp_info = mad_send_wr->mad_agent_priv->qp_info;
 863         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
 864         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
 865
 866         mad_agent = mad_send_wr->send_buf.mad_agent;
 867         sge = mad_send_wr->sg_list;
 868         sge->addr = dma_map_single(mad_agent->device->dma_device,
 869                                    mad_send_wr->send_buf.mad, sge->length,
 870                                    DMA_TO_DEVICE);
 871         pci_unmap_addr_set(mad_send_wr, mapping, sge->addr);
 872
 873         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
 874         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
 875                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
 876                                    &bad_send_wr);
 877                 list = &qp_info->send_queue.list;
 878         } else {
 879                 ret = 0;
 880                 list = &qp_info->overflow_list;
 881         }
 882
 883         if (!ret) {
 884                 qp_info->send_queue.count++;
 885                 list_add_tail(&mad_send_wr->mad_list.list, list);
 886         }
 887         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
 888         if (ret)
 889                 dma_unmap_single(mad_agent->device->dma_device,
 890                                  pci_unmap_addr(mad_send_wr, mapping),
 891                                  sge->length, DMA_TO_DEVICE);
 892
 893         return ret;
 894 }
 895
 896 /*
 897  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
 898  *  with the registered client
 899  */
 900 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
 901                      struct ib_mad_send_buf **bad_send_buf)
 902 {
 903         struct ib_mad_agent_private *mad_agent_priv;
 904         struct ib_mad_send_buf *next_send_buf;
 905         struct ib_mad_send_wr_private *mad_send_wr;
 906         unsigned long flags;
 907         int ret = -EINVAL;
 908
 909         /* Walk list of send WRs and post each on send list */
 910         for (; send_buf; send_buf = next_send_buf) {
 911
 912                 mad_send_wr = container_of(send_buf,
 913                                            struct ib_mad_send_wr_private,
 914                                            send_buf);
 915                 mad_agent_priv = mad_send_wr->mad_agent_priv;
 916
 917                 if (!send_buf->mad_agent->send_handler ||
 918                     (send_buf->timeout_ms &&
 919                      !send_buf->mad_agent->recv_handler)) {
 920                         ret = -EINVAL;
 921                         goto error;
 922                 }
 923
 924                 /*
 925                  * Save pointer to next work request to post in case the
 926                  * current one completes, and the user modifies the work
 927                  * request associated with the completion
 928                  */
 929                 next_send_buf = send_buf->next;
 930                 mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
 931
 932                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
 933                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 934                         ret = handle_outgoing_dr_smp(mad_agent_priv,
 935                                                      mad_send_wr);
 936                         if (ret < 0)            /* error */
 937                                 goto error;
 938                         else if (ret == 1)      /* locally consumed */
 939                                 continue;
 940                 }
 941
 942                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
 943                 /* Timeout will be updated after send completes */
 944                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
 945                 mad_send_wr->retries = send_buf->retries;
 946                 /* Reference for work request to QP + response */
 947                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
 948                 mad_send_wr->status = IB_WC_SUCCESS;
 949
 950                 /* Reference MAD agent until send completes */
 951                 atomic_inc(&mad_agent_priv->refcount);
 952                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
 953                 list_add_tail(&mad_send_wr->agent_list,
 954                               &mad_agent_priv->send_list);
 955                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 956
 957                 if (mad_agent_priv->agent.rmpp_version) {
 958                         ret = ib_send_rmpp_mad(mad_send_wr);
 959                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
 960                                 ret = ib_send_mad(mad_send_wr);
 961                 } else
 962                         ret = ib_send_mad(mad_send_wr);
 963                 if (ret < 0) {
 964                         /* Fail send request */
 965                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 966                         list_del(&mad_send_wr->agent_list);
 967                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 968                         atomic_dec(&mad_agent_priv->refcount);
 969                         goto error;
 970                 }
 971         }
 972         return 0;
 973 error:
 974         if (bad_send_buf)
 975                 *bad_send_buf = send_buf;
 976         return ret;
 977 }
 978 EXPORT_SYMBOL(ib_post_send_mad);
 979
 980 /*
 981  * ib_free_recv_mad - Returns data buffers used to receive
 982  *  a MAD to the access layer
 983  */
 984 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
 985 {
 986         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
 987         struct ib_mad_private_header *mad_priv_hdr;
 988         struct ib_mad_private *priv;
 989         struct list_head free_list;
 990
 991         INIT_LIST_HEAD(&free_list);
 992         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
 993
 994         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
 995                                         &free_list, list) {
 996                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
 997                                            recv_buf);
 998                 mad_priv_hdr = container_of(mad_recv_wc,
 999                                             struct ib_mad_private_header,
1000                                             recv_wc);
1001                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1002                                     header);
1003                 kmem_cache_free(ib_mad_cache, priv);
1004         }
1005 }
1006 EXPORT_SYMBOL(ib_free_recv_mad);
1007
1008 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1009                                         u8 rmpp_version,
1010                                         ib_mad_send_handler send_handler,
1011                                         ib_mad_recv_handler recv_handler,
1012                                         void *context)
1013 {
1014         return ERR_PTR(-EINVAL);        /* XXX: for now */
1015 }
1016 EXPORT_SYMBOL(ib_redirect_mad_qp);
1017
1018 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1019                       struct ib_wc *wc)
1020 {
1021         printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1022         return 0;
1023 }
1024 EXPORT_SYMBOL(ib_process_mad_wc);
1025
1026 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1027                          struct ib_mad_reg_req *mad_reg_req)
1028 {
1029         int i;
1030
1031         for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1032              i < IB_MGMT_MAX_METHODS;
1033              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1034                                1+i)) {
1035                 if ((*method)->agent[i]) {
1036                         printk(KERN_ERR PFX "Method %d already in use\n", i);
1037                         return -EINVAL;
1038                 }
1039         }
1040         return 0;
1041 }
1042
1043 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1044 {
1045         /* Allocate management method table */
1046         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1047         if (!*method) {
1048                 printk(KERN_ERR PFX "No memory for "
1049                        "ib_mad_mgmt_method_table\n");
1050                 return -ENOMEM;
1051         }
1052
1053         return 0;
1054 }
1055
1056 /*
1057  * Check to see if there are any methods still in use
1058  */
1059 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1060 {
1061         int i;
1062
1063         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1064                 if (method->agent[i])
1065                         return 1;
1066         return 0;
1067 }
1068
1069 /*
1070  * Check to see if there are any method tables for this class still in use
1071  */
1072 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1073 {
1074         int i;
1075
1076         for (i = 0; i < MAX_MGMT_CLASS; i++)
1077                 if (class->method_table[i])
1078                         return 1;
1079         return 0;
1080 }
1081
1082 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1083 {
1084         int i;
1085
1086         for (i = 0; i < MAX_MGMT_OUI; i++)
1087                 if (vendor_class->method_table[i])
1088                         return 1;
1089         return 0;
1090 }
1091
1092 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1093                            char *oui)
1094 {
1095         int i;
1096
1097         for (i = 0; i < MAX_MGMT_OUI; i++)
1098                 /* Is there matching OUI for this vendor class ? */
1099                 if (!memcmp(vendor_class->oui[i], oui, 3))
1100                         return i;
1101
1102         return -1;
1103 }
1104
1105 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1106 {
1107         int i;
1108
1109         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1110                 if (vendor->vendor_class[i])
1111                         return 1;
1112
1113         return 0;
1114 }
1115
1116 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1117                                      struct ib_mad_agent_private *agent)
1118 {
1119         int i;
1120
1121         /* Remove any methods for this mad agent */
1122         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1123                 if (method->agent[i] == agent) {
1124                         method->agent[i] = NULL;
1125                 }
1126         }
1127 }
1128
1129 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1130                               struct ib_mad_agent_private *agent_priv,
1131                               u8 mgmt_class)
1132 {
1133         struct ib_mad_port_private *port_priv;
1134         struct ib_mad_mgmt_class_table **class;
1135         struct ib_mad_mgmt_method_table **method;
1136         int i, ret;
1137
1138         port_priv = agent_priv->qp_info->port_priv;
1139         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1140         if (!*class) {
1141                 /* Allocate management class table for "new" class version */
1142                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1143                 if (!*class) {
1144                         printk(KERN_ERR PFX "No memory for "
1145                                "ib_mad_mgmt_class_table\n");
1146                         ret = -ENOMEM;
1147                         goto error1;
1148                 }
1149
1150                 /* Allocate method table for this management class */
1151                 method = &(*class)->method_table[mgmt_class];
1152                 if ((ret = allocate_method_table(method)))
1153                         goto error2;
1154         } else {
1155                 method = &(*class)->method_table[mgmt_class];
1156                 if (!*method) {
1157                         /* Allocate method table for this management class */
1158                         if ((ret = allocate_method_table(method)))
1159                                 goto error1;
1160                 }
1161         }
1162
1163         /* Now, make sure methods are not already in use */
1164         if (method_in_use(method, mad_reg_req))
1165                 goto error3;
1166
1167         /* Finally, add in methods being registered */
1168         for (i = find_first_bit(mad_reg_req->method_mask,
1169                                 IB_MGMT_MAX_METHODS);
1170              i < IB_MGMT_MAX_METHODS;
1171              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1172                                1+i)) {
1173                 (*method)->agent[i] = agent_priv;
1174         }
1175         return 0;
1176
1177 error3:
1178         /* Remove any methods for this mad agent */
1179         remove_methods_mad_agent(*method, agent_priv);
1180         /* Now, check to see if there are any methods in use */
1181         if (!check_method_table(*method)) {
1182                 /* If not, release management method table */
1183                 kfree(*method);
1184                 *method = NULL;
1185         }
1186         ret = -EINVAL;
1187         goto error1;
1188 error2:
1189         kfree(*class);
1190         *class = NULL;
1191 error1:
1192         return ret;
1193 }
1194
1195 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1196                            struct ib_mad_agent_private *agent_priv)
1197 {
1198         struct ib_mad_port_private *port_priv;
1199         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1200         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1201         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1202         struct ib_mad_mgmt_method_table **method;
1203         int i, ret = -ENOMEM;
1204         u8 vclass;
1205
1206         /* "New" vendor (with OUI) class */
1207         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1208         port_priv = agent_priv->qp_info->port_priv;
1209         vendor_table = &port_priv->version[
1210                                 mad_reg_req->mgmt_class_version].vendor;
1211         if (!*vendor_table) {
1212                 /* Allocate mgmt vendor class table for "new" class version */
1213                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1214                 if (!vendor) {
1215                         printk(KERN_ERR PFX "No memory for "
1216                                "ib_mad_mgmt_vendor_class_table\n");
1217                         goto error1;
1218                 }
1219
1220                 *vendor_table = vendor;
1221         }
1222         if (!(*vendor_table)->vendor_class[vclass]) {
1223                 /* Allocate table for this management vendor class */
1224                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1225                 if (!vendor_class) {
1226                         printk(KERN_ERR PFX "No memory for "
1227                                "ib_mad_mgmt_vendor_class\n");
1228                         goto error2;
1229                 }
1230
1231                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1232         }
1233         for (i = 0; i < MAX_MGMT_OUI; i++) {
1234                 /* Is there matching OUI for this vendor class ? */
1235                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1236                             mad_reg_req->oui, 3)) {
1237                         method = &(*vendor_table)->vendor_class[
1238                                                 vclass]->method_table[i];
1239                         BUG_ON(!*method);
1240                         goto check_in_use;
1241                 }
1242         }
1243         for (i = 0; i < MAX_MGMT_OUI; i++) {
1244                 /* OUI slot available ? */
1245                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1246                                 vclass]->oui[i])) {
1247                         method = &(*vendor_table)->vendor_class[
1248                                 vclass]->method_table[i];
1249                         BUG_ON(*method);
1250                         /* Allocate method table for this OUI */
1251                         if ((ret = allocate_method_table(method)))
1252                                 goto error3;
1253                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1254                                mad_reg_req->oui, 3);
1255                         goto check_in_use;
1256                 }
1257         }
1258         printk(KERN_ERR PFX "All OUI slots in use\n");
1259         goto error3;
1260
1261 check_in_use:
1262         /* Now, make sure methods are not already in use */
1263         if (method_in_use(method, mad_reg_req))
1264                 goto error4;
1265
1266         /* Finally, add in methods being registered */
1267         for (i = find_first_bit(mad_reg_req->method_mask,
1268                                 IB_MGMT_MAX_METHODS);
1269              i < IB_MGMT_MAX_METHODS;
1270              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1271                                1+i)) {
1272                 (*method)->agent[i] = agent_priv;
1273         }
1274         return 0;
1275
1276 error4:
1277         /* Remove any methods for this mad agent */
1278         remove_methods_mad_agent(*method, agent_priv);
1279         /* Now, check to see if there are any methods in use */
1280         if (!check_method_table(*method)) {
1281                 /* If not, release management method table */
1282                 kfree(*method);
1283                 *method = NULL;
1284         }
1285         ret = -EINVAL;
1286 error3:
1287         if (vendor_class) {
1288                 (*vendor_table)->vendor_class[vclass] = NULL;
1289                 kfree(vendor_class);
1290         }
1291 error2:
1292         if (vendor) {
1293                 *vendor_table = NULL;
1294                 kfree(vendor);
1295         }
1296 error1:
1297         return ret;
1298 }
1299
1300 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1301 {
1302         struct ib_mad_port_private *port_priv;
1303         struct ib_mad_mgmt_class_table *class;
1304         struct ib_mad_mgmt_method_table *method;
1305         struct ib_mad_mgmt_vendor_class_table *vendor;
1306         struct ib_mad_mgmt_vendor_class *vendor_class;
1307         int index;
1308         u8 mgmt_class;
1309
1310         /*
1311          * Was MAD registration request supplied
1312          * with original registration ?
1313          */
1314         if (!agent_priv->reg_req) {
1315                 goto out;
1316         }
1317
1318         port_priv = agent_priv->qp_info->port_priv;
1319         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1320         class = port_priv->version[
1321                         agent_priv->reg_req->mgmt_class_version].class;
1322         if (!class)
1323                 goto vendor_check;
1324
1325         method = class->method_table[mgmt_class];
1326         if (method) {
1327                 /* Remove any methods for this mad agent */
1328                 remove_methods_mad_agent(method, agent_priv);
1329                 /* Now, check to see if there are any methods still in use */
1330                 if (!check_method_table(method)) {
1331                         /* If not, release management method table */
1332                          kfree(method);
1333                          class->method_table[mgmt_class] = NULL;
1334                          /* Any management classes left ? */
1335                         if (!check_class_table(class)) {
1336                                 /* If not, release management class table */
1337                                 kfree(class);
1338                                 port_priv->version[
1339                                         agent_priv->reg_req->
1340                                         mgmt_class_version].class = NULL;
1341                         }
1342                 }
1343         }
1344
1345 vendor_check:
1346         if (!is_vendor_class(mgmt_class))
1347                 goto out;
1348
1349         /* normalize mgmt_class to vendor range 2 */
1350         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1351         vendor = port_priv->version[
1352                         agent_priv->reg_req->mgmt_class_version].vendor;
1353
1354         if (!vendor)
1355                 goto out;
1356
1357         vendor_class = vendor->vendor_class[mgmt_class];
1358         if (vendor_class) {
1359                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1360                 if (index < 0)
1361                         goto out;
1362                 method = vendor_class->method_table[index];
1363                 if (method) {
1364                         /* Remove any methods for this mad agent */
1365                         remove_methods_mad_agent(method, agent_priv);
1366                         /*
1367                          * Now, check to see if there are
1368                          * any methods still in use
1369                          */
1370                         if (!check_method_table(method)) {
1371                                 /* If not, release management method table */
1372                                 kfree(method);
1373                                 vendor_class->method_table[index] = NULL;
1374                                 memset(vendor_class->oui[index], 0, 3);
1375                                 /* Any OUIs left ? */
1376                                 if (!check_vendor_class(vendor_class)) {
1377                                         /* If not, release vendor class table */
1378                                         kfree(vendor_class);
1379                                         vendor->vendor_class[mgmt_class] = NULL;
1380                                         /* Any other vendor classes left ? */
1381                                         if (!check_vendor_table(vendor)) {
1382                                                 kfree(vendor);
1383                                                 port_priv->version[
1384                                                         agent_priv->reg_req->
1385                                                         mgmt_class_version].
1386                                                         vendor = NULL;
1387                                         }
1388                                 }
1389                         }
1390                 }
1391         }
1392
1393 out:
1394         return;
1395 }
1396
1397 static struct ib_mad_agent_private *
1398 find_mad_agent(struct ib_mad_port_private *port_priv,
1399                struct ib_mad *mad)
1400 {
1401         struct ib_mad_agent_private *mad_agent = NULL;
1402         unsigned long flags;
1403
1404         spin_lock_irqsave(&port_priv->reg_lock, flags);
1405         if (response_mad(mad)) {
1406                 u32 hi_tid;
1407                 struct ib_mad_agent_private *entry;
1408
1409                 /*
1410                  * Routing is based on high 32 bits of transaction ID
1411                  * of MAD.
1412                  */
1413                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1414                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1415                         if (entry->agent.hi_tid == hi_tid) {
1416                                 mad_agent = entry;
1417                                 break;
1418                         }
1419                 }
1420         } else {
1421                 struct ib_mad_mgmt_class_table *class;
1422                 struct ib_mad_mgmt_method_table *method;
1423                 struct ib_mad_mgmt_vendor_class_table *vendor;
1424                 struct ib_mad_mgmt_vendor_class *vendor_class;
1425                 struct ib_vendor_mad *vendor_mad;
1426                 int index;
1427
1428                 /*
1429                  * Routing is based on version, class, and method
1430                  * For "newer" vendor MADs, also based on OUI
1431                  */
1432                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1433                         goto out;
1434                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1435                         class = port_priv->version[
1436                                         mad->mad_hdr.class_version].class;
1437                         if (!class)
1438                                 goto out;
1439                         method = class->method_table[convert_mgmt_class(
1440                                                         mad->mad_hdr.mgmt_class)];
1441                         if (method)
1442                                 mad_agent = method->agent[mad->mad_hdr.method &
1443                                                           ~IB_MGMT_METHOD_RESP];
1444                 } else {
1445                         vendor = port_priv->version[
1446                                         mad->mad_hdr.class_version].vendor;
1447                         if (!vendor)
1448                                 goto out;
1449                         vendor_class = vendor->vendor_class[vendor_class_index(
1450                                                 mad->mad_hdr.mgmt_class)];
1451                         if (!vendor_class)
1452                                 goto out;
1453                         /* Find matching OUI */
1454                         vendor_mad = (struct ib_vendor_mad *)mad;
1455                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1456                         if (index == -1)
1457                                 goto out;
1458                         method = vendor_class->method_table[index];
1459                         if (method) {
1460                                 mad_agent = method->agent[mad->mad_hdr.method &
1461                                                           ~IB_MGMT_METHOD_RESP];
1462                         }
1463                 }
1464         }
1465
1466         if (mad_agent) {
1467                 if (mad_agent->agent.recv_handler)
1468                         atomic_inc(&mad_agent->refcount);
1469                 else {
1470                         printk(KERN_NOTICE PFX "No receive handler for client "
1471                                "%p on port %d\n",
1472                                &mad_agent->agent, port_priv->port_num);
1473                         mad_agent = NULL;
1474                 }
1475         }
1476 out:
1477         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1478
1479         return mad_agent;
1480 }
1481
1482 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1483 {
1484         int valid = 0;
1485
1486         /* Make sure MAD base version is understood */
1487         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1488                 printk(KERN_ERR PFX "MAD received with unsupported base "
1489                        "version %d\n", mad->mad_hdr.base_version);
1490                 goto out;
1491         }
1492
1493         /* Filter SMI packets sent to other than QP0 */
1494         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1495             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1496                 if (qp_num == 0)
1497                         valid = 1;
1498         } else {
1499                 /* Filter GSI packets sent to QP0 */
1500                 if (qp_num != 0)
1501                         valid = 1;
1502         }
1503
1504 out:
1505         return valid;
1506 }
1507
1508 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1509                        struct ib_mad_hdr *mad_hdr)
1510 {
1511         struct ib_rmpp_mad *rmpp_mad;
1512
1513         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1514         return !mad_agent_priv->agent.rmpp_version ||
1515                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1516                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1517                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1518 }
1519
1520 struct ib_mad_send_wr_private*
1521 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv, __be64 tid)
1522 {
1523         struct ib_mad_send_wr_private *mad_send_wr;
1524
1525         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
1526                             agent_list) {
1527                 if (mad_send_wr->tid == tid)
1528                         return mad_send_wr;
1529         }
1530
1531         /*
1532          * It's possible to receive the response before we've
1533          * been notified that the send has completed
1534          */
1535         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
1536                             agent_list) {
1537                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
1538                     mad_send_wr->tid == tid && mad_send_wr->timeout) {
1539                         /* Verify request has not been canceled */
1540                         return (mad_send_wr->status == IB_WC_SUCCESS) ?
1541                                 mad_send_wr : NULL;
1542                 }
1543         }
1544         return NULL;
1545 }
1546
1547 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1548 {
1549         mad_send_wr->timeout = 0;
1550         if (mad_send_wr->refcount == 1) {
1551                 list_del(&mad_send_wr->agent_list);
1552                 list_add_tail(&mad_send_wr->agent_list,
1553                               &mad_send_wr->mad_agent_priv->done_list);
1554         }
1555 }
1556
1557 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1558                                  struct ib_mad_recv_wc *mad_recv_wc)
1559 {
1560         struct ib_mad_send_wr_private *mad_send_wr;
1561         struct ib_mad_send_wc mad_send_wc;
1562         unsigned long flags;
1563         __be64 tid;
1564
1565         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1566         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1567         if (mad_agent_priv->agent.rmpp_version) {
1568                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1569                                                       mad_recv_wc);
1570                 if (!mad_recv_wc) {
1571                         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1572                                 wake_up(&mad_agent_priv->wait);
1573                         return;
1574                 }
1575         }
1576
1577         /* Complete corresponding request */
1578         if (response_mad(mad_recv_wc->recv_buf.mad)) {
1579                 tid = mad_recv_wc->recv_buf.mad->mad_hdr.tid;
1580                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1581                 mad_send_wr = ib_find_send_mad(mad_agent_priv, tid);
1582                 if (!mad_send_wr) {
1583                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1584                         ib_free_recv_mad(mad_recv_wc);
1585                         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1586                                 wake_up(&mad_agent_priv->wait);
1587                         return;
1588                 }
1589                 ib_mark_mad_done(mad_send_wr);
1590                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1591
1592                 /* Defined behavior is to complete response before request */
1593                 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1594                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1595                                                    mad_recv_wc);
1596                 atomic_dec(&mad_agent_priv->refcount);
1597
1598                 mad_send_wc.status = IB_WC_SUCCESS;
1599                 mad_send_wc.vendor_err = 0;
1600                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1601                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1602         } else {
1603                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1604                                                    mad_recv_wc);
1605                 if (atomic_dec_and_test(&mad_agent_priv->refcount))
1606                         wake_up(&mad_agent_priv->wait);
1607         }
1608 }
1609
1610 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1611                                      struct ib_wc *wc)
1612 {
1613         struct ib_mad_qp_info *qp_info;
1614         struct ib_mad_private_header *mad_priv_hdr;
1615         struct ib_mad_private *recv, *response;
1616         struct ib_mad_list_head *mad_list;
1617         struct ib_mad_agent_private *mad_agent;
1618
1619         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1620         if (!response)
1621                 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1622                        "for response buffer\n");
1623
1624         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1625         qp_info = mad_list->mad_queue->qp_info;
1626         dequeue_mad(mad_list);
1627
1628         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1629                                     mad_list);
1630         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1631         dma_unmap_single(port_priv->device->dma_device,
1632                          pci_unmap_addr(&recv->header, mapping),
1633                          sizeof(struct ib_mad_private) -
1634                          sizeof(struct ib_mad_private_header),
1635                          DMA_FROM_DEVICE);
1636
1637         /* Setup MAD receive work completion from "normal" work completion */
1638         recv->header.wc = *wc;
1639         recv->header.recv_wc.wc = &recv->header.wc;
1640         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1641         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1642         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1643
1644         if (atomic_read(&qp_info->snoop_count))
1645                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1646
1647         /* Validate MAD */
1648         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1649                 goto out;
1650
1651         if (recv->mad.mad.mad_hdr.mgmt_class ==
1652             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1653                 if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1654                                             port_priv->device->node_type,
1655                                             port_priv->port_num,
1656                                             port_priv->device->phys_port_cnt))
1657                         goto out;
1658                 if (!smi_check_forward_dr_smp(&recv->mad.smp))
1659                         goto local;
1660                 if (!smi_handle_dr_smp_send(&recv->mad.smp,
1661                                             port_priv->device->node_type,
1662                                             port_priv->port_num))
1663                         goto out;
1664                 if (!smi_check_local_dr_smp(&recv->mad.smp,
1665                                             port_priv->device,
1666                                             port_priv->port_num))
1667                         goto out;
1668         }
1669
1670 local:
1671         /* Give driver "right of first refusal" on incoming MAD */
1672         if (port_priv->device->process_mad) {
1673                 int ret;
1674
1675                 if (!response) {
1676                         printk(KERN_ERR PFX "No memory for response MAD\n");
1677                         /*
1678                          * Is it better to assume that
1679                          * it wouldn't be processed ?
1680                          */
1681                         goto out;
1682                 }
1683
1684                 ret = port_priv->device->process_mad(port_priv->device, 0,
1685                                                      port_priv->port_num,
1686                                                      wc, &recv->grh,
1687                                                      &recv->mad.mad,
1688                                                      &response->mad.mad);
1689                 if (ret & IB_MAD_RESULT_SUCCESS) {
1690                         if (ret & IB_MAD_RESULT_CONSUMED)
1691                                 goto out;
1692                         if (ret & IB_MAD_RESULT_REPLY) {
1693                                 agent_send_response(&response->mad.mad,
1694                                                     &recv->grh, wc,
1695                                                     port_priv->device,
1696                                                     port_priv->port_num,
1697                                                     qp_info->qp->qp_num);
1698                                 goto out;
1699                         }
1700                 }
1701         }
1702
1703         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1704         if (mad_agent) {
1705                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1706                 /*
1707                  * recv is freed up in error cases in ib_mad_complete_recv
1708                  * or via recv_handler in ib_mad_complete_recv()
1709                  */
1710                 recv = NULL;
1711         }
1712
1713 out:
1714         /* Post another receive request for this QP */
1715         if (response) {
1716                 ib_mad_post_receive_mads(qp_info, response);
1717                 if (recv)
1718                         kmem_cache_free(ib_mad_cache, recv);
1719         } else
1720                 ib_mad_post_receive_mads(qp_info, recv);
1721 }
1722
1723 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1724 {
1725         struct ib_mad_send_wr_private *mad_send_wr;
1726         unsigned long delay;
1727
1728         if (list_empty(&mad_agent_priv->wait_list)) {
1729                 cancel_delayed_work(&mad_agent_priv->timed_work);
1730         } else {
1731                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1732                                          struct ib_mad_send_wr_private,
1733                                          agent_list);
1734
1735                 if (time_after(mad_agent_priv->timeout,
1736                                mad_send_wr->timeout)) {
1737                         mad_agent_priv->timeout = mad_send_wr->timeout;
1738                         cancel_delayed_work(&mad_agent_priv->timed_work);
1739                         delay = mad_send_wr->timeout - jiffies;
1740                         if ((long)delay <= 0)
1741                                 delay = 1;
1742                         queue_delayed_work(mad_agent_priv->qp_info->
1743                                            port_priv->wq,
1744                                            &mad_agent_priv->timed_work, delay);
1745                 }
1746         }
1747 }
1748
1749 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1750 {
1751         struct ib_mad_agent_private *mad_agent_priv;
1752         struct ib_mad_send_wr_private *temp_mad_send_wr;
1753         struct list_head *list_item;
1754         unsigned long delay;
1755
1756         mad_agent_priv = mad_send_wr->mad_agent_priv;
1757         list_del(&mad_send_wr->agent_list);
1758
1759         delay = mad_send_wr->timeout;
1760         mad_send_wr->timeout += jiffies;
1761
1762         if (delay) {
1763                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1764                         temp_mad_send_wr = list_entry(list_item,
1765                                                 struct ib_mad_send_wr_private,
1766                                                 agent_list);
1767                         if (time_after(mad_send_wr->timeout,
1768                                        temp_mad_send_wr->timeout))
1769                                 break;
1770                 }
1771         }
1772         else
1773                 list_item = &mad_agent_priv->wait_list;
1774         list_add(&mad_send_wr->agent_list, list_item);
1775
1776         /* Reschedule a work item if we have a shorter timeout */
1777         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1778                 cancel_delayed_work(&mad_agent_priv->timed_work);
1779                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1780                                    &mad_agent_priv->timed_work, delay);
1781         }
1782 }
1783
1784 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
1785                           int timeout_ms)
1786 {
1787         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
1788         wait_for_response(mad_send_wr);
1789 }
1790
1791 /*
1792  * Process a send work completion
1793  */
1794 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
1795                              struct ib_mad_send_wc *mad_send_wc)
1796 {
1797         struct ib_mad_agent_private     *mad_agent_priv;
1798         unsigned long                   flags;
1799         int                             ret;
1800
1801         mad_agent_priv = mad_send_wr->mad_agent_priv;
1802         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1803         if (mad_agent_priv->agent.rmpp_version) {
1804                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
1805                 if (ret == IB_RMPP_RESULT_CONSUMED)
1806                         goto done;
1807         } else
1808                 ret = IB_RMPP_RESULT_UNHANDLED;
1809
1810         if (mad_send_wc->status != IB_WC_SUCCESS &&
1811             mad_send_wr->status == IB_WC_SUCCESS) {
1812                 mad_send_wr->status = mad_send_wc->status;
1813                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
1814         }
1815
1816         if (--mad_send_wr->refcount > 0) {
1817                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
1818                     mad_send_wr->status == IB_WC_SUCCESS) {
1819                         wait_for_response(mad_send_wr);
1820                 }
1821                 goto done;
1822         }
1823
1824         /* Remove send from MAD agent and notify client of completion */
1825         list_del(&mad_send_wr->agent_list);
1826         adjust_timeout(mad_agent_priv);
1827         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1828
1829         if (mad_send_wr->status != IB_WC_SUCCESS )
1830                 mad_send_wc->status = mad_send_wr->status;
1831         if (ret == IB_RMPP_RESULT_INTERNAL)
1832                 ib_rmpp_send_handler(mad_send_wc);
1833         else
1834                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
1835                                                    mad_send_wc);
1836
1837         /* Release reference on agent taken when sending */
1838         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1839                 wake_up(&mad_agent_priv->wait);
1840         return;
1841 done:
1842         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1843 }
1844
1845 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
1846                                      struct ib_wc *wc)
1847 {
1848         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
1849         struct ib_mad_list_head         *mad_list;
1850         struct ib_mad_qp_info           *qp_info;
1851         struct ib_mad_queue             *send_queue;
1852         struct ib_send_wr               *bad_send_wr;
1853         struct ib_mad_send_wc           mad_send_wc;
1854         unsigned long flags;
1855         int ret;
1856
1857         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1858         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1859                                    mad_list);
1860         send_queue = mad_list->mad_queue;
1861         qp_info = send_queue->qp_info;
1862
1863 retry:
1864         dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
1865                          pci_unmap_addr(mad_send_wr, mapping),
1866                          mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
1867         queued_send_wr = NULL;
1868         spin_lock_irqsave(&send_queue->lock, flags);
1869         list_del(&mad_list->list);
1870
1871         /* Move queued send to the send queue */
1872         if (send_queue->count-- > send_queue->max_active) {
1873                 mad_list = container_of(qp_info->overflow_list.next,
1874                                         struct ib_mad_list_head, list);
1875                 queued_send_wr = container_of(mad_list,
1876                                         struct ib_mad_send_wr_private,
1877                                         mad_list);
1878                 list_del(&mad_list->list);
1879                 list_add_tail(&mad_list->list, &send_queue->list);
1880         }
1881         spin_unlock_irqrestore(&send_queue->lock, flags);
1882
1883         mad_send_wc.send_buf = &mad_send_wr->send_buf;
1884         mad_send_wc.status = wc->status;
1885         mad_send_wc.vendor_err = wc->vendor_err;
1886         if (atomic_read(&qp_info->snoop_count))
1887                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
1888                            IB_MAD_SNOOP_SEND_COMPLETIONS);
1889         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1890
1891         if (queued_send_wr) {
1892                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
1893                                    &bad_send_wr);
1894                 if (ret) {
1895                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
1896                         mad_send_wr = queued_send_wr;
1897                         wc->status = IB_WC_LOC_QP_OP_ERR;
1898                         goto retry;
1899                 }
1900         }
1901 }
1902
1903 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
1904 {
1905         struct ib_mad_send_wr_private *mad_send_wr;
1906         struct ib_mad_list_head *mad_list;
1907         unsigned long flags;
1908
1909         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1910         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
1911                 mad_send_wr = container_of(mad_list,
1912                                            struct ib_mad_send_wr_private,
1913                                            mad_list);
1914                 mad_send_wr->retry = 1;
1915         }
1916         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1917 }
1918
1919 static void mad_error_handler(struct ib_mad_port_private *port_priv,
1920                               struct ib_wc *wc)
1921 {
1922         struct ib_mad_list_head *mad_list;
1923         struct ib_mad_qp_info *qp_info;
1924         struct ib_mad_send_wr_private *mad_send_wr;
1925         int ret;
1926
1927         /* Determine if failure was a send or receive */
1928         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1929         qp_info = mad_list->mad_queue->qp_info;
1930         if (mad_list->mad_queue == &qp_info->recv_queue)
1931                 /*
1932                  * Receive errors indicate that the QP has entered the error
1933                  * state - error handling/shutdown code will cleanup
1934                  */
1935                 return;
1936
1937         /*
1938          * Send errors will transition the QP to SQE - move
1939          * QP to RTS and repost flushed work requests
1940          */
1941         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1942                                    mad_list);
1943         if (wc->status == IB_WC_WR_FLUSH_ERR) {
1944                 if (mad_send_wr->retry) {
1945                         /* Repost send */
1946                         struct ib_send_wr *bad_send_wr;
1947
1948                         mad_send_wr->retry = 0;
1949                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
1950                                         &bad_send_wr);
1951                         if (ret)
1952                                 ib_mad_send_done_handler(port_priv, wc);
1953                 } else
1954                         ib_mad_send_done_handler(port_priv, wc);
1955         } else {
1956                 struct ib_qp_attr *attr;
1957
1958                 /* Transition QP to RTS and fail offending send */
1959                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
1960                 if (attr) {
1961                         attr->qp_state = IB_QPS_RTS;
1962                         attr->cur_qp_state = IB_QPS_SQE;
1963                         ret = ib_modify_qp(qp_info->qp, attr,
1964                                            IB_QP_STATE | IB_QP_CUR_STATE);
1965                         kfree(attr);
1966                         if (ret)
1967                                 printk(KERN_ERR PFX "mad_error_handler - "
1968                                        "ib_modify_qp to RTS : %d\n", ret);
1969                         else
1970                                 mark_sends_for_retry(qp_info);
1971                 }
1972                 ib_mad_send_done_handler(port_priv, wc);
1973         }
1974 }
1975
1976 /*
1977  * IB MAD completion callback
1978  */
1979 static void ib_mad_completion_handler(void *data)
1980 {
1981         struct ib_mad_port_private *port_priv;
1982         struct ib_wc wc;
1983
1984         port_priv = (struct ib_mad_port_private *)data;
1985         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
1986
1987         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
1988                 if (wc.status == IB_WC_SUCCESS) {
1989                         switch (wc.opcode) {
1990                         case IB_WC_SEND:
1991                                 ib_mad_send_done_handler(port_priv, &wc);
1992                                 break;
1993                         case IB_WC_RECV:
1994                                 ib_mad_recv_done_handler(port_priv, &wc);
1995                                 break;
1996                         default:
1997                                 BUG_ON(1);
1998                                 break;
1999                         }
2000                 } else
2001                         mad_error_handler(port_priv, &wc);
2002         }
2003 }
2004
2005 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2006 {
2007         unsigned long flags;
2008         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2009         struct ib_mad_send_wc mad_send_wc;
2010         struct list_head cancel_list;
2011
2012         INIT_LIST_HEAD(&cancel_list);
2013
2014         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2015         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2016                                  &mad_agent_priv->send_list, agent_list) {
2017                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2018                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2019                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2020                 }
2021         }
2022
2023         /* Empty wait list to prevent receives from finding a request */
2024         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2025         /* Empty local completion list as well */
2026         list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2027         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2028
2029         /* Report all cancelled requests */
2030         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2031         mad_send_wc.vendor_err = 0;
2032
2033         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2034                                  &cancel_list, agent_list) {
2035                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2036                 list_del(&mad_send_wr->agent_list);
2037                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2038                                                    &mad_send_wc);
2039                 atomic_dec(&mad_agent_priv->refcount);
2040         }
2041 }
2042
2043 static struct ib_mad_send_wr_private*
2044 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2045              struct ib_mad_send_buf *send_buf)
2046 {
2047         struct ib_mad_send_wr_private *mad_send_wr;
2048
2049         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2050                             agent_list) {
2051                 if (&mad_send_wr->send_buf == send_buf)
2052                         return mad_send_wr;
2053         }
2054
2055         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2056                             agent_list) {
2057                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2058                     &mad_send_wr->send_buf == send_buf)
2059                         return mad_send_wr;
2060         }
2061         return NULL;
2062 }
2063
2064 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2065                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2066 {
2067         struct ib_mad_agent_private *mad_agent_priv;
2068         struct ib_mad_send_wr_private *mad_send_wr;
2069         unsigned long flags;
2070         int active;
2071
2072         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2073                                       agent);
2074         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2075         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2076         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2077                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2078                 return -EINVAL;
2079         }
2080
2081         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2082         if (!timeout_ms) {
2083                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2084                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2085         }
2086
2087         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2088         if (active)
2089                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2090         else
2091                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2092
2093         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2094         return 0;
2095 }
2096 EXPORT_SYMBOL(ib_modify_mad);
2097
2098 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2099                    struct ib_mad_send_buf *send_buf)
2100 {
2101         ib_modify_mad(mad_agent, send_buf, 0);
2102 }
2103 EXPORT_SYMBOL(ib_cancel_mad);
2104
2105 static void local_completions(void *data)
2106 {
2107         struct ib_mad_agent_private *mad_agent_priv;
2108         struct ib_mad_local_private *local;
2109         struct ib_mad_agent_private *recv_mad_agent;
2110         unsigned long flags;
2111         int recv = 0;
2112         struct ib_wc wc;
2113         struct ib_mad_send_wc mad_send_wc;
2114
2115         mad_agent_priv = (struct ib_mad_agent_private *)data;
2116
2117         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2118         while (!list_empty(&mad_agent_priv->local_list)) {
2119                 local = list_entry(mad_agent_priv->local_list.next,
2120                                    struct ib_mad_local_private,
2121                                    completion_list);
2122                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2123                 if (local->mad_priv) {
2124                         recv_mad_agent = local->recv_mad_agent;
2125                         if (!recv_mad_agent) {
2126                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2127                                 goto local_send_completion;
2128                         }
2129
2130                         recv = 1;
2131                         /*
2132                          * Defined behavior is to complete response
2133                          * before request
2134                          */
2135                         build_smp_wc((unsigned long) local->mad_send_wr,
2136                                      be16_to_cpu(IB_LID_PERMISSIVE),
2137                                      0, recv_mad_agent->agent.port_num, &wc);
2138
2139                         local->mad_priv->header.recv_wc.wc = &wc;
2140                         local->mad_priv->header.recv_wc.mad_len =
2141                                                 sizeof(struct ib_mad);
2142                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2143                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2144                                  &local->mad_priv->header.recv_wc.rmpp_list);
2145                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2146                         local->mad_priv->header.recv_wc.recv_buf.mad =
2147                                                 &local->mad_priv->mad.mad;
2148                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2149                                 snoop_recv(recv_mad_agent->qp_info,
2150                                           &local->mad_priv->header.recv_wc,
2151                                            IB_MAD_SNOOP_RECVS);
2152                         recv_mad_agent->agent.recv_handler(
2153                                                 &recv_mad_agent->agent,
2154                                                 &local->mad_priv->header.recv_wc);
2155                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2156                         atomic_dec(&recv_mad_agent->refcount);
2157                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2158                 }
2159
2160 local_send_completion:
2161                 /* Complete send */
2162                 mad_send_wc.status = IB_WC_SUCCESS;
2163                 mad_send_wc.vendor_err = 0;
2164                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2165                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2166                         snoop_send(mad_agent_priv->qp_info,
2167                                    &local->mad_send_wr->send_buf,
2168                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2169                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2170                                                    &mad_send_wc);
2171
2172                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2173                 list_del(&local->completion_list);
2174                 atomic_dec(&mad_agent_priv->refcount);
2175                 if (!recv)
2176                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2177                 kfree(local);
2178         }
2179         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2180 }
2181
2182 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2183 {
2184         int ret;
2185
2186         if (!mad_send_wr->retries--)
2187                 return -ETIMEDOUT;
2188
2189         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2190
2191         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2192                 ret = ib_retry_rmpp(mad_send_wr);
2193                 switch (ret) {
2194                 case IB_RMPP_RESULT_UNHANDLED:
2195                         ret = ib_send_mad(mad_send_wr);
2196                         break;
2197                 case IB_RMPP_RESULT_CONSUMED:
2198                         ret = 0;
2199                         break;
2200                 default:
2201                         ret = -ECOMM;
2202                         break;
2203                 }
2204         } else
2205                 ret = ib_send_mad(mad_send_wr);
2206
2207         if (!ret) {
2208                 mad_send_wr->refcount++;
2209                 list_add_tail(&mad_send_wr->agent_list,
2210                               &mad_send_wr->mad_agent_priv->send_list);
2211         }
2212         return ret;
2213 }
2214
2215 static void timeout_sends(void *data)
2216 {
2217         struct ib_mad_agent_private *mad_agent_priv;
2218         struct ib_mad_send_wr_private *mad_send_wr;
2219         struct ib_mad_send_wc mad_send_wc;
2220         unsigned long flags, delay;
2221
2222         mad_agent_priv = (struct ib_mad_agent_private *)data;
2223         mad_send_wc.vendor_err = 0;
2224
2225         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2226         while (!list_empty(&mad_agent_priv->wait_list)) {
2227                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2228                                          struct ib_mad_send_wr_private,
2229                                          agent_list);
2230
2231                 if (time_after(mad_send_wr->timeout, jiffies)) {
2232                         delay = mad_send_wr->timeout - jiffies;
2233                         if ((long)delay <= 0)
2234                                 delay = 1;
2235                         queue_delayed_work(mad_agent_priv->qp_info->
2236                                            port_priv->wq,
2237                                            &mad_agent_priv->timed_work, delay);
2238                         break;
2239                 }
2240
2241                 list_del(&mad_send_wr->agent_list);
2242                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2243                     !retry_send(mad_send_wr))
2244                         continue;
2245
2246                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2247
2248                 if (mad_send_wr->status == IB_WC_SUCCESS)
2249                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2250                 else
2251                         mad_send_wc.status = mad_send_wr->status;
2252                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2253                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2254                                                    &mad_send_wc);
2255
2256                 atomic_dec(&mad_agent_priv->refcount);
2257                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2258         }
2259         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2260 }
2261
2262 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2263 {
2264         struct ib_mad_port_private *port_priv = cq->cq_context;
2265
2266         queue_work(port_priv->wq, &port_priv->work);
2267 }
2268
2269 /*
2270  * Allocate receive MADs and post receive WRs for them
2271  */
2272 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2273                                     struct ib_mad_private *mad)
2274 {
2275         unsigned long flags;
2276         int post, ret;
2277         struct ib_mad_private *mad_priv;
2278         struct ib_sge sg_list;
2279         struct ib_recv_wr recv_wr, *bad_recv_wr;
2280         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2281
2282         /* Initialize common scatter list fields */
2283         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2284         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2285
2286         /* Initialize common receive WR fields */
2287         recv_wr.next = NULL;
2288         recv_wr.sg_list = &sg_list;
2289         recv_wr.num_sge = 1;
2290
2291         do {
2292                 /* Allocate and map receive buffer */
2293                 if (mad) {
2294                         mad_priv = mad;
2295                         mad = NULL;
2296                 } else {
2297                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2298                         if (!mad_priv) {
2299                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2300                                 ret = -ENOMEM;
2301                                 break;
2302                         }
2303                 }
2304                 sg_list.addr = dma_map_single(qp_info->port_priv->
2305                                                 device->dma_device,
2306                                         &mad_priv->grh,
2307                                         sizeof *mad_priv -
2308                                                 sizeof mad_priv->header,
2309                                         DMA_FROM_DEVICE);
2310                 pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2311                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2312                 mad_priv->header.mad_list.mad_queue = recv_queue;
2313
2314                 /* Post receive WR */
2315                 spin_lock_irqsave(&recv_queue->lock, flags);
2316                 post = (++recv_queue->count < recv_queue->max_active);
2317                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2318                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2319                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2320                 if (ret) {
2321                         spin_lock_irqsave(&recv_queue->lock, flags);
2322                         list_del(&mad_priv->header.mad_list.list);
2323                         recv_queue->count--;
2324                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2325                         dma_unmap_single(qp_info->port_priv->device->dma_device,
2326                                          pci_unmap_addr(&mad_priv->header,
2327                                                         mapping),
2328                                          sizeof *mad_priv -
2329                                            sizeof mad_priv->header,
2330                                          DMA_FROM_DEVICE);
2331                         kmem_cache_free(ib_mad_cache, mad_priv);
2332                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2333                         break;
2334                 }
2335         } while (post);
2336
2337         return ret;
2338 }
2339
2340 /*
2341  * Return all the posted receive MADs
2342  */
2343 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2344 {
2345         struct ib_mad_private_header *mad_priv_hdr;
2346         struct ib_mad_private *recv;
2347         struct ib_mad_list_head *mad_list;
2348
2349         while (!list_empty(&qp_info->recv_queue.list)) {
2350
2351                 mad_list = list_entry(qp_info->recv_queue.list.next,
2352                                       struct ib_mad_list_head, list);
2353                 mad_priv_hdr = container_of(mad_list,
2354                                             struct ib_mad_private_header,
2355                                             mad_list);
2356                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2357                                     header);
2358
2359                 /* Remove from posted receive MAD list */
2360                 list_del(&mad_list->list);
2361
2362                 dma_unmap_single(qp_info->port_priv->device->dma_device,
2363                                  pci_unmap_addr(&recv->header, mapping),
2364                                  sizeof(struct ib_mad_private) -
2365                                  sizeof(struct ib_mad_private_header),
2366                                  DMA_FROM_DEVICE);
2367                 kmem_cache_free(ib_mad_cache, recv);
2368         }
2369
2370         qp_info->recv_queue.count = 0;
2371 }
2372
2373 /*
2374  * Start the port
2375  */
2376 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2377 {
2378         int ret, i;
2379         struct ib_qp_attr *attr;
2380         struct ib_qp *qp;
2381
2382         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2383         if (!attr) {
2384                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2385                 return -ENOMEM;
2386         }
2387
2388         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2389                 qp = port_priv->qp_info[i].qp;
2390                 /*
2391                  * PKey index for QP1 is irrelevant but
2392                  * one is needed for the Reset to Init transition
2393                  */
2394                 attr->qp_state = IB_QPS_INIT;
2395                 attr->pkey_index = 0;
2396                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2397                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2398                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2399                 if (ret) {
2400                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2401                                "INIT: %d\n", i, ret);
2402                         goto out;
2403                 }
2404
2405                 attr->qp_state = IB_QPS_RTR;
2406                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2407                 if (ret) {
2408                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2409                                "RTR: %d\n", i, ret);
2410                         goto out;
2411                 }
2412
2413                 attr->qp_state = IB_QPS_RTS;
2414                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2415                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2416                 if (ret) {
2417                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2418                                "RTS: %d\n", i, ret);
2419                         goto out;
2420                 }
2421         }
2422
2423         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2424         if (ret) {
2425                 printk(KERN_ERR PFX "Failed to request completion "
2426                        "notification: %d\n", ret);
2427                 goto out;
2428         }
2429
2430         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2431                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2432                 if (ret) {
2433                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2434                         goto out;
2435                 }
2436         }
2437 out:
2438         kfree(attr);
2439         return ret;
2440 }
2441
2442 static void qp_event_handler(struct ib_event *event, void *qp_context)
2443 {
2444         struct ib_mad_qp_info   *qp_info = qp_context;
2445
2446         /* It's worse than that! He's dead, Jim! */
2447         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2448                 event->event, qp_info->qp->qp_num);
2449 }
2450
2451 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2452                            struct ib_mad_queue *mad_queue)
2453 {
2454         mad_queue->qp_info = qp_info;
2455         mad_queue->count = 0;
2456         spin_lock_init(&mad_queue->lock);
2457         INIT_LIST_HEAD(&mad_queue->list);
2458 }
2459
2460 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2461                         struct ib_mad_qp_info *qp_info)
2462 {
2463         qp_info->port_priv = port_priv;
2464         init_mad_queue(qp_info, &qp_info->send_queue);
2465         init_mad_queue(qp_info, &qp_info->recv_queue);
2466         INIT_LIST_HEAD(&qp_info->overflow_list);
2467         spin_lock_init(&qp_info->snoop_lock);
2468         qp_info->snoop_table = NULL;
2469         qp_info->snoop_table_size = 0;
2470         atomic_set(&qp_info->snoop_count, 0);
2471 }
2472
2473 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2474                          enum ib_qp_type qp_type)
2475 {
2476         struct ib_qp_init_attr  qp_init_attr;
2477         int ret;
2478
2479         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2480         qp_init_attr.send_cq = qp_info->port_priv->cq;
2481         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2482         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2483         qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2484         qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2485         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2486         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2487         qp_init_attr.qp_type = qp_type;
2488         qp_init_attr.port_num = qp_info->port_priv->port_num;
2489         qp_init_attr.qp_context = qp_info;
2490         qp_init_attr.event_handler = qp_event_handler;
2491         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2492         if (IS_ERR(qp_info->qp)) {
2493                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2494                        get_spl_qp_index(qp_type));
2495                 ret = PTR_ERR(qp_info->qp);
2496                 goto error;
2497         }
2498         /* Use minimum queue sizes unless the CQ is resized */
2499         qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2500         qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2501         return 0;
2502
2503 error:
2504         return ret;
2505 }
2506
2507 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2508 {
2509         ib_destroy_qp(qp_info->qp);
2510         kfree(qp_info->snoop_table);
2511 }
2512
2513 /*
2514  * Open the port
2515  * Create the QP, PD, MR, and CQ if needed
2516  */
2517 static int ib_mad_port_open(struct ib_device *device,
2518                             int port_num)
2519 {
2520         int ret, cq_size;
2521         struct ib_mad_port_private *port_priv;
2522         unsigned long flags;
2523         char name[sizeof "ib_mad123"];
2524
2525         /* Create new device info */
2526         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2527         if (!port_priv) {
2528                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2529                 return -ENOMEM;
2530         }
2531
2532         port_priv->device = device;
2533         port_priv->port_num = port_num;
2534         spin_lock_init(&port_priv->reg_lock);
2535         INIT_LIST_HEAD(&port_priv->agent_list);
2536         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2537         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2538
2539         cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2540         port_priv->cq = ib_create_cq(port_priv->device,
2541                                      ib_mad_thread_completion_handler,
2542                                      NULL, port_priv, cq_size);
2543         if (IS_ERR(port_priv->cq)) {
2544                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2545                 ret = PTR_ERR(port_priv->cq);
2546                 goto error3;
2547         }
2548
2549         port_priv->pd = ib_alloc_pd(device);
2550         if (IS_ERR(port_priv->pd)) {
2551                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2552                 ret = PTR_ERR(port_priv->pd);
2553                 goto error4;
2554         }
2555
2556         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2557         if (IS_ERR(port_priv->mr)) {
2558                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2559                 ret = PTR_ERR(port_priv->mr);
2560                 goto error5;
2561         }
2562
2563         ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2564         if (ret)
2565                 goto error6;
2566         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2567         if (ret)
2568                 goto error7;
2569
2570         snprintf(name, sizeof name, "ib_mad%d", port_num);
2571         port_priv->wq = create_singlethread_workqueue(name);
2572         if (!port_priv->wq) {
2573                 ret = -ENOMEM;
2574                 goto error8;
2575         }
2576         INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2577
2578         ret = ib_mad_port_start(port_priv);
2579         if (ret) {
2580                 printk(KERN_ERR PFX "Couldn't start port\n");
2581                 goto error9;
2582         }
2583
2584         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2585         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2586         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2587         return 0;
2588
2589 error9:
2590         destroy_workqueue(port_priv->wq);
2591 error8:
2592         destroy_mad_qp(&port_priv->qp_info[1]);
2593 error7:
2594         destroy_mad_qp(&port_priv->qp_info[0]);
2595 error6:
2596         ib_dereg_mr(port_priv->mr);
2597 error5:
2598         ib_dealloc_pd(port_priv->pd);
2599 error4:
2600         ib_destroy_cq(port_priv->cq);
2601         cleanup_recv_queue(&port_priv->qp_info[1]);
2602         cleanup_recv_queue(&port_priv->qp_info[0]);
2603 error3:
2604         kfree(port_priv);
2605
2606         return ret;
2607 }
2608
2609 /*
2610  * Close the port
2611  * If there are no classes using the port, free the port
2612  * resources (CQ, MR, PD, QP) and remove the port's info structure
2613  */
2614 static int ib_mad_port_close(struct ib_device *device, int port_num)
2615 {
2616         struct ib_mad_port_private *port_priv;
2617         unsigned long flags;
2618
2619         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2620         port_priv = __ib_get_mad_port(device, port_num);
2621         if (port_priv == NULL) {
2622                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2623                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2624                 return -ENODEV;
2625         }
2626         list_del(&port_priv->port_list);
2627         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2628
2629         /* Stop processing completions. */
2630         flush_workqueue(port_priv->wq);
2631         destroy_workqueue(port_priv->wq);
2632         destroy_mad_qp(&port_priv->qp_info[1]);
2633         destroy_mad_qp(&port_priv->qp_info[0]);
2634         ib_dereg_mr(port_priv->mr);
2635         ib_dealloc_pd(port_priv->pd);
2636         ib_destroy_cq(port_priv->cq);
2637         cleanup_recv_queue(&port_priv->qp_info[1]);
2638         cleanup_recv_queue(&port_priv->qp_info[0]);
2639         /* XXX: Handle deallocation of MAD registration tables */
2640
2641         kfree(port_priv);
2642
2643         return 0;
2644 }
2645
2646 static void ib_mad_init_device(struct ib_device *device)
2647 {
2648         int start, end, i;
2649
2650         if (device->node_type == IB_NODE_SWITCH) {
2651                 start = 0;
2652                 end   = 0;
2653         } else {
2654                 start = 1;
2655                 end   = device->phys_port_cnt;
2656         }
2657
2658         for (i = start; i <= end; i++) {
2659                 if (ib_mad_port_open(device, i)) {
2660                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2661                                device->name, i);
2662                         goto error;
2663                 }
2664                 if (ib_agent_port_open(device, i)) {
2665                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2666                                "for agents\n",
2667                                device->name, i);
2668                         goto error_agent;
2669                 }
2670         }
2671         return;
2672
2673 error_agent:
2674         if (ib_mad_port_close(device, i))
2675                 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2676                        device->name, i);
2677
2678 error:
2679         i--;
2680
2681         while (i >= start) {
2682                 if (ib_agent_port_close(device, i))
2683                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2684                                "for agents\n",
2685                                device->name, i);
2686                 if (ib_mad_port_close(device, i))
2687                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2688                                device->name, i);
2689                 i--;
2690         }
2691 }
2692
2693 static void ib_mad_remove_device(struct ib_device *device)
2694 {
2695         int i, num_ports, cur_port;
2696
2697         if (device->node_type == IB_NODE_SWITCH) {
2698                 num_ports = 1;
2699                 cur_port = 0;
2700         } else {
2701                 num_ports = device->phys_port_cnt;
2702                 cur_port = 1;
2703         }
2704         for (i = 0; i < num_ports; i++, cur_port++) {
2705                 if (ib_agent_port_close(device, cur_port))
2706                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2707                                "for agents\n",
2708                                device->name, cur_port);
2709                 if (ib_mad_port_close(device, cur_port))
2710                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2711                                device->name, cur_port);
2712         }
2713 }
2714
2715 static struct ib_client mad_client = {
2716         .name   = "mad",
2717         .add = ib_mad_init_device,
2718         .remove = ib_mad_remove_device
2719 };
2720
2721 static int __init ib_mad_init_module(void)
2722 {
2723         int ret;
2724
2725         spin_lock_init(&ib_mad_port_list_lock);
2726
2727         ib_mad_cache = kmem_cache_create("ib_mad",
2728                                          sizeof(struct ib_mad_private),
2729                                          0,
2730                                          SLAB_HWCACHE_ALIGN,
2731                                          NULL,
2732                                          NULL);
2733         if (!ib_mad_cache) {
2734                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2735                 ret = -ENOMEM;
2736                 goto error1;
2737         }
2738
2739         INIT_LIST_HEAD(&ib_mad_port_list);
2740
2741         if (ib_register_client(&mad_client)) {
2742                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2743                 ret = -EINVAL;
2744                 goto error2;
2745         }
2746
2747         return 0;
2748
2749 error2:
2750         kmem_cache_destroy(ib_mad_cache);
2751 error1:
2752         return ret;
2753 }
2754
2755 static void __exit ib_mad_cleanup_module(void)
2756 {
2757         ib_unregister_client(&mad_client);
2758
2759         if (kmem_cache_destroy(ib_mad_cache)) {
2760                 printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2761         }
2762 }
2763
2764 module_init(ib_mad_init_module);
2765 module_exit(ib_mad_cleanup_module);
2766