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_smp(smp, device);
 683         if (!ret)
 684                 goto out;
 685
 686         local = kmalloc(sizeof *local, GFP_ATOMIC);
 687         if (!local) {
 688                 ret = -ENOMEM;
 689                 printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
 690                 goto out;
 691         }
 692         local->mad_priv = NULL;
 693         local->recv_mad_agent = NULL;
 694         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
 695         if (!mad_priv) {
 696                 ret = -ENOMEM;
 697                 printk(KERN_ERR PFX "No memory for local response MAD\n");
 698                 kfree(local);
 699                 goto out;
 700         }
 701
 702         build_smp_wc(send_wr->wr_id, be16_to_cpu(smp->dr_slid),
 703                      send_wr->wr.ud.pkey_index,
 704                      send_wr->wr.ud.port_num, &mad_wc);
 705
 706         /* No GRH for DR SMP */
 707         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 708                                   (struct ib_mad *)smp,
 709                                   (struct ib_mad *)&mad_priv->mad);
 710         switch (ret)
 711         {
 712         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 713                 if (response_mad(&mad_priv->mad.mad) &&
 714                     mad_agent_priv->agent.recv_handler) {
 715                         local->mad_priv = mad_priv;
 716                         local->recv_mad_agent = mad_agent_priv;
 717                         /*
 718                          * Reference MAD agent until receive
 719                          * side of local completion handled
 720                          */
 721                         atomic_inc(&mad_agent_priv->refcount);
 722                 } else
 723                         kmem_cache_free(ib_mad_cache, mad_priv);
 724                 break;
 725         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 726                 kmem_cache_free(ib_mad_cache, mad_priv);
 727                 break;
 728         case IB_MAD_RESULT_SUCCESS:
 729                 /* Treat like an incoming receive MAD */
 730                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 731                                             mad_agent_priv->agent.port_num);
 732                 if (port_priv) {
 733                         mad_priv->mad.mad.mad_hdr.tid =
 734                                 ((struct ib_mad *)smp)->mad_hdr.tid;
 735                         recv_mad_agent = find_mad_agent(port_priv,
 736                                                         &mad_priv->mad.mad);
 737                 }
 738                 if (!port_priv || !recv_mad_agent) {
 739                         kmem_cache_free(ib_mad_cache, mad_priv);
 740                         kfree(local);
 741                         ret = 0;
 742                         goto out;
 743                 }
 744                 local->mad_priv = mad_priv;
 745                 local->recv_mad_agent = recv_mad_agent;
 746                 break;
 747         default:
 748                 kmem_cache_free(ib_mad_cache, mad_priv);
 749                 kfree(local);
 750                 ret = -EINVAL;
 751                 goto out;
 752         }
 753
 754         local->mad_send_wr = mad_send_wr;
 755         /* Reference MAD agent until send side of local completion handled */
 756         atomic_inc(&mad_agent_priv->refcount);
 757         /* Queue local completion to local list */
 758         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 759         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 760         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 761         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 762                    &mad_agent_priv->local_work);
 763         ret = 1;
 764 out:
 765         return ret;
 766 }
 767
 768 static int get_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_smp(&recv->mad.smp, port_priv->device))
1665                         goto out;
1666         }
1667
1668 local:
1669         /* Give driver "right of first refusal" on incoming MAD */
1670         if (port_priv->device->process_mad) {
1671                 int ret;
1672
1673                 if (!response) {
1674                         printk(KERN_ERR PFX "No memory for response MAD\n");
1675                         /*
1676                          * Is it better to assume that
1677                          * it wouldn't be processed ?
1678                          */
1679                         goto out;
1680                 }
1681
1682                 ret = port_priv->device->process_mad(port_priv->device, 0,
1683                                                      port_priv->port_num,
1684                                                      wc, &recv->grh,
1685                                                      &recv->mad.mad,
1686                                                      &response->mad.mad);
1687                 if (ret & IB_MAD_RESULT_SUCCESS) {
1688                         if (ret & IB_MAD_RESULT_CONSUMED)
1689                                 goto out;
1690                         if (ret & IB_MAD_RESULT_REPLY) {
1691                                 agent_send_response(&response->mad.mad,
1692                                                     &recv->grh, wc,
1693                                                     port_priv->device,
1694                                                     port_priv->port_num,
1695                                                     qp_info->qp->qp_num);
1696                                 goto out;
1697                         }
1698                 }
1699         }
1700
1701         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1702         if (mad_agent) {
1703                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1704                 /*
1705                  * recv is freed up in error cases in ib_mad_complete_recv
1706                  * or via recv_handler in ib_mad_complete_recv()
1707                  */
1708                 recv = NULL;
1709         }
1710
1711 out:
1712         /* Post another receive request for this QP */
1713         if (response) {
1714                 ib_mad_post_receive_mads(qp_info, response);
1715                 if (recv)
1716                         kmem_cache_free(ib_mad_cache, recv);
1717         } else
1718                 ib_mad_post_receive_mads(qp_info, recv);
1719 }
1720
1721 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1722 {
1723         struct ib_mad_send_wr_private *mad_send_wr;
1724         unsigned long delay;
1725
1726         if (list_empty(&mad_agent_priv->wait_list)) {
1727                 cancel_delayed_work(&mad_agent_priv->timed_work);
1728         } else {
1729                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1730                                          struct ib_mad_send_wr_private,
1731                                          agent_list);
1732
1733                 if (time_after(mad_agent_priv->timeout,
1734                                mad_send_wr->timeout)) {
1735                         mad_agent_priv->timeout = mad_send_wr->timeout;
1736                         cancel_delayed_work(&mad_agent_priv->timed_work);
1737                         delay = mad_send_wr->timeout - jiffies;
1738                         if ((long)delay <= 0)
1739                                 delay = 1;
1740                         queue_delayed_work(mad_agent_priv->qp_info->
1741                                            port_priv->wq,
1742                                            &mad_agent_priv->timed_work, delay);
1743                 }
1744         }
1745 }
1746
1747 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1748 {
1749         struct ib_mad_agent_private *mad_agent_priv;
1750         struct ib_mad_send_wr_private *temp_mad_send_wr;
1751         struct list_head *list_item;
1752         unsigned long delay;
1753
1754         mad_agent_priv = mad_send_wr->mad_agent_priv;
1755         list_del(&mad_send_wr->agent_list);
1756
1757         delay = mad_send_wr->timeout;
1758         mad_send_wr->timeout += jiffies;
1759
1760         if (delay) {
1761                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1762                         temp_mad_send_wr = list_entry(list_item,
1763                                                 struct ib_mad_send_wr_private,
1764                                                 agent_list);
1765                         if (time_after(mad_send_wr->timeout,
1766                                        temp_mad_send_wr->timeout))
1767                                 break;
1768                 }
1769         }
1770         else
1771                 list_item = &mad_agent_priv->wait_list;
1772         list_add(&mad_send_wr->agent_list, list_item);
1773
1774         /* Reschedule a work item if we have a shorter timeout */
1775         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1776                 cancel_delayed_work(&mad_agent_priv->timed_work);
1777                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1778                                    &mad_agent_priv->timed_work, delay);
1779         }
1780 }
1781
1782 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
1783                           int timeout_ms)
1784 {
1785         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
1786         wait_for_response(mad_send_wr);
1787 }
1788
1789 /*
1790  * Process a send work completion
1791  */
1792 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
1793                              struct ib_mad_send_wc *mad_send_wc)
1794 {
1795         struct ib_mad_agent_private     *mad_agent_priv;
1796         unsigned long                   flags;
1797         int                             ret;
1798
1799         mad_agent_priv = mad_send_wr->mad_agent_priv;
1800         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1801         if (mad_agent_priv->agent.rmpp_version) {
1802                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
1803                 if (ret == IB_RMPP_RESULT_CONSUMED)
1804                         goto done;
1805         } else
1806                 ret = IB_RMPP_RESULT_UNHANDLED;
1807
1808         if (mad_send_wc->status != IB_WC_SUCCESS &&
1809             mad_send_wr->status == IB_WC_SUCCESS) {
1810                 mad_send_wr->status = mad_send_wc->status;
1811                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
1812         }
1813
1814         if (--mad_send_wr->refcount > 0) {
1815                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
1816                     mad_send_wr->status == IB_WC_SUCCESS) {
1817                         wait_for_response(mad_send_wr);
1818                 }
1819                 goto done;
1820         }
1821
1822         /* Remove send from MAD agent and notify client of completion */
1823         list_del(&mad_send_wr->agent_list);
1824         adjust_timeout(mad_agent_priv);
1825         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1826
1827         if (mad_send_wr->status != IB_WC_SUCCESS )
1828                 mad_send_wc->status = mad_send_wr->status;
1829         if (ret == IB_RMPP_RESULT_INTERNAL)
1830                 ib_rmpp_send_handler(mad_send_wc);
1831         else
1832                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
1833                                                    mad_send_wc);
1834
1835         /* Release reference on agent taken when sending */
1836         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1837                 wake_up(&mad_agent_priv->wait);
1838         return;
1839 done:
1840         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1841 }
1842
1843 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
1844                                      struct ib_wc *wc)
1845 {
1846         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
1847         struct ib_mad_list_head         *mad_list;
1848         struct ib_mad_qp_info           *qp_info;
1849         struct ib_mad_queue             *send_queue;
1850         struct ib_send_wr               *bad_send_wr;
1851         struct ib_mad_send_wc           mad_send_wc;
1852         unsigned long flags;
1853         int ret;
1854
1855         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1856         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1857                                    mad_list);
1858         send_queue = mad_list->mad_queue;
1859         qp_info = send_queue->qp_info;
1860
1861 retry:
1862         dma_unmap_single(mad_send_wr->send_buf.mad_agent->device->dma_device,
1863                          pci_unmap_addr(mad_send_wr, mapping),
1864                          mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
1865         queued_send_wr = NULL;
1866         spin_lock_irqsave(&send_queue->lock, flags);
1867         list_del(&mad_list->list);
1868
1869         /* Move queued send to the send queue */
1870         if (send_queue->count-- > send_queue->max_active) {
1871                 mad_list = container_of(qp_info->overflow_list.next,
1872                                         struct ib_mad_list_head, list);
1873                 queued_send_wr = container_of(mad_list,
1874                                         struct ib_mad_send_wr_private,
1875                                         mad_list);
1876                 list_del(&mad_list->list);
1877                 list_add_tail(&mad_list->list, &send_queue->list);
1878         }
1879         spin_unlock_irqrestore(&send_queue->lock, flags);
1880
1881         mad_send_wc.send_buf = &mad_send_wr->send_buf;
1882         mad_send_wc.status = wc->status;
1883         mad_send_wc.vendor_err = wc->vendor_err;
1884         if (atomic_read(&qp_info->snoop_count))
1885                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
1886                            IB_MAD_SNOOP_SEND_COMPLETIONS);
1887         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1888
1889         if (queued_send_wr) {
1890                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
1891                                    &bad_send_wr);
1892                 if (ret) {
1893                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
1894                         mad_send_wr = queued_send_wr;
1895                         wc->status = IB_WC_LOC_QP_OP_ERR;
1896                         goto retry;
1897                 }
1898         }
1899 }
1900
1901 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
1902 {
1903         struct ib_mad_send_wr_private *mad_send_wr;
1904         struct ib_mad_list_head *mad_list;
1905         unsigned long flags;
1906
1907         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1908         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
1909                 mad_send_wr = container_of(mad_list,
1910                                            struct ib_mad_send_wr_private,
1911                                            mad_list);
1912                 mad_send_wr->retry = 1;
1913         }
1914         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1915 }
1916
1917 static void mad_error_handler(struct ib_mad_port_private *port_priv,
1918                               struct ib_wc *wc)
1919 {
1920         struct ib_mad_list_head *mad_list;
1921         struct ib_mad_qp_info *qp_info;
1922         struct ib_mad_send_wr_private *mad_send_wr;
1923         int ret;
1924
1925         /* Determine if failure was a send or receive */
1926         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1927         qp_info = mad_list->mad_queue->qp_info;
1928         if (mad_list->mad_queue == &qp_info->recv_queue)
1929                 /*
1930                  * Receive errors indicate that the QP has entered the error
1931                  * state - error handling/shutdown code will cleanup
1932                  */
1933                 return;
1934
1935         /*
1936          * Send errors will transition the QP to SQE - move
1937          * QP to RTS and repost flushed work requests
1938          */
1939         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1940                                    mad_list);
1941         if (wc->status == IB_WC_WR_FLUSH_ERR) {
1942                 if (mad_send_wr->retry) {
1943                         /* Repost send */
1944                         struct ib_send_wr *bad_send_wr;
1945
1946                         mad_send_wr->retry = 0;
1947                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
1948                                         &bad_send_wr);
1949                         if (ret)
1950                                 ib_mad_send_done_handler(port_priv, wc);
1951                 } else
1952                         ib_mad_send_done_handler(port_priv, wc);
1953         } else {
1954                 struct ib_qp_attr *attr;
1955
1956                 /* Transition QP to RTS and fail offending send */
1957                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
1958                 if (attr) {
1959                         attr->qp_state = IB_QPS_RTS;
1960                         attr->cur_qp_state = IB_QPS_SQE;
1961                         ret = ib_modify_qp(qp_info->qp, attr,
1962                                            IB_QP_STATE | IB_QP_CUR_STATE);
1963                         kfree(attr);
1964                         if (ret)
1965                                 printk(KERN_ERR PFX "mad_error_handler - "
1966                                        "ib_modify_qp to RTS : %d\n", ret);
1967                         else
1968                                 mark_sends_for_retry(qp_info);
1969                 }
1970                 ib_mad_send_done_handler(port_priv, wc);
1971         }
1972 }
1973
1974 /*
1975  * IB MAD completion callback
1976  */
1977 static void ib_mad_completion_handler(void *data)
1978 {
1979         struct ib_mad_port_private *port_priv;
1980         struct ib_wc wc;
1981
1982         port_priv = (struct ib_mad_port_private *)data;
1983         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
1984
1985         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
1986                 if (wc.status == IB_WC_SUCCESS) {
1987                         switch (wc.opcode) {
1988                         case IB_WC_SEND:
1989                                 ib_mad_send_done_handler(port_priv, &wc);
1990                                 break;
1991                         case IB_WC_RECV:
1992                                 ib_mad_recv_done_handler(port_priv, &wc);
1993                                 break;
1994                         default:
1995                                 BUG_ON(1);
1996                                 break;
1997                         }
1998                 } else
1999                         mad_error_handler(port_priv, &wc);
2000         }
2001 }
2002
2003 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2004 {
2005         unsigned long flags;
2006         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2007         struct ib_mad_send_wc mad_send_wc;
2008         struct list_head cancel_list;
2009
2010         INIT_LIST_HEAD(&cancel_list);
2011
2012         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2013         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2014                                  &mad_agent_priv->send_list, agent_list) {
2015                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2016                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2017                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2018                 }
2019         }
2020
2021         /* Empty wait list to prevent receives from finding a request */
2022         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2023         /* Empty local completion list as well */
2024         list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2025         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2026
2027         /* Report all cancelled requests */
2028         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2029         mad_send_wc.vendor_err = 0;
2030
2031         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2032                                  &cancel_list, agent_list) {
2033                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2034                 list_del(&mad_send_wr->agent_list);
2035                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2036                                                    &mad_send_wc);
2037                 atomic_dec(&mad_agent_priv->refcount);
2038         }
2039 }
2040
2041 static struct ib_mad_send_wr_private*
2042 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2043              struct ib_mad_send_buf *send_buf)
2044 {
2045         struct ib_mad_send_wr_private *mad_send_wr;
2046
2047         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2048                             agent_list) {
2049                 if (&mad_send_wr->send_buf == send_buf)
2050                         return mad_send_wr;
2051         }
2052
2053         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2054                             agent_list) {
2055                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2056                     &mad_send_wr->send_buf == send_buf)
2057                         return mad_send_wr;
2058         }
2059         return NULL;
2060 }
2061
2062 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2063                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2064 {
2065         struct ib_mad_agent_private *mad_agent_priv;
2066         struct ib_mad_send_wr_private *mad_send_wr;
2067         unsigned long flags;
2068         int active;
2069
2070         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2071                                       agent);
2072         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2073         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2074         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2075                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2076                 return -EINVAL;
2077         }
2078
2079         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2080         if (!timeout_ms) {
2081                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2082                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2083         }
2084
2085         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2086         if (active)
2087                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2088         else
2089                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2090
2091         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2092         return 0;
2093 }
2094 EXPORT_SYMBOL(ib_modify_mad);
2095
2096 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2097                    struct ib_mad_send_buf *send_buf)
2098 {
2099         ib_modify_mad(mad_agent, send_buf, 0);
2100 }
2101 EXPORT_SYMBOL(ib_cancel_mad);
2102
2103 static void local_completions(void *data)
2104 {
2105         struct ib_mad_agent_private *mad_agent_priv;
2106         struct ib_mad_local_private *local;
2107         struct ib_mad_agent_private *recv_mad_agent;
2108         unsigned long flags;
2109         int recv = 0;
2110         struct ib_wc wc;
2111         struct ib_mad_send_wc mad_send_wc;
2112
2113         mad_agent_priv = (struct ib_mad_agent_private *)data;
2114
2115         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2116         while (!list_empty(&mad_agent_priv->local_list)) {
2117                 local = list_entry(mad_agent_priv->local_list.next,
2118                                    struct ib_mad_local_private,
2119                                    completion_list);
2120                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2121                 if (local->mad_priv) {
2122                         recv_mad_agent = local->recv_mad_agent;
2123                         if (!recv_mad_agent) {
2124                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2125                                 goto local_send_completion;
2126                         }
2127
2128                         recv = 1;
2129                         /*
2130                          * Defined behavior is to complete response
2131                          * before request
2132                          */
2133                         build_smp_wc((unsigned long) local->mad_send_wr,
2134                                      be16_to_cpu(IB_LID_PERMISSIVE),
2135                                      0, recv_mad_agent->agent.port_num, &wc);
2136
2137                         local->mad_priv->header.recv_wc.wc = &wc;
2138                         local->mad_priv->header.recv_wc.mad_len =
2139                                                 sizeof(struct ib_mad);
2140                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2141                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2142                                  &local->mad_priv->header.recv_wc.rmpp_list);
2143                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2144                         local->mad_priv->header.recv_wc.recv_buf.mad =
2145                                                 &local->mad_priv->mad.mad;
2146                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2147                                 snoop_recv(recv_mad_agent->qp_info,
2148                                           &local->mad_priv->header.recv_wc,
2149                                            IB_MAD_SNOOP_RECVS);
2150                         recv_mad_agent->agent.recv_handler(
2151                                                 &recv_mad_agent->agent,
2152                                                 &local->mad_priv->header.recv_wc);
2153                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2154                         atomic_dec(&recv_mad_agent->refcount);
2155                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2156                 }
2157
2158 local_send_completion:
2159                 /* Complete send */
2160                 mad_send_wc.status = IB_WC_SUCCESS;
2161                 mad_send_wc.vendor_err = 0;
2162                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2163                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2164                         snoop_send(mad_agent_priv->qp_info,
2165                                    &local->mad_send_wr->send_buf,
2166                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2167                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2168                                                    &mad_send_wc);
2169
2170                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2171                 list_del(&local->completion_list);
2172                 atomic_dec(&mad_agent_priv->refcount);
2173                 if (!recv)
2174                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2175                 kfree(local);
2176         }
2177         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2178 }
2179
2180 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2181 {
2182         int ret;
2183
2184         if (!mad_send_wr->retries--)
2185                 return -ETIMEDOUT;
2186
2187         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2188
2189         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2190                 ret = ib_retry_rmpp(mad_send_wr);
2191                 switch (ret) {
2192                 case IB_RMPP_RESULT_UNHANDLED:
2193                         ret = ib_send_mad(mad_send_wr);
2194                         break;
2195                 case IB_RMPP_RESULT_CONSUMED:
2196                         ret = 0;
2197                         break;
2198                 default:
2199                         ret = -ECOMM;
2200                         break;
2201                 }
2202         } else
2203                 ret = ib_send_mad(mad_send_wr);
2204
2205         if (!ret) {
2206                 mad_send_wr->refcount++;
2207                 list_add_tail(&mad_send_wr->agent_list,
2208                               &mad_send_wr->mad_agent_priv->send_list);
2209         }
2210         return ret;
2211 }
2212
2213 static void timeout_sends(void *data)
2214 {
2215         struct ib_mad_agent_private *mad_agent_priv;
2216         struct ib_mad_send_wr_private *mad_send_wr;
2217         struct ib_mad_send_wc mad_send_wc;
2218         unsigned long flags, delay;
2219
2220         mad_agent_priv = (struct ib_mad_agent_private *)data;
2221         mad_send_wc.vendor_err = 0;
2222
2223         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2224         while (!list_empty(&mad_agent_priv->wait_list)) {
2225                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2226                                          struct ib_mad_send_wr_private,
2227                                          agent_list);
2228
2229                 if (time_after(mad_send_wr->timeout, jiffies)) {
2230                         delay = mad_send_wr->timeout - jiffies;
2231                         if ((long)delay <= 0)
2232                                 delay = 1;
2233                         queue_delayed_work(mad_agent_priv->qp_info->
2234                                            port_priv->wq,
2235                                            &mad_agent_priv->timed_work, delay);
2236                         break;
2237                 }
2238
2239                 list_del(&mad_send_wr->agent_list);
2240                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2241                     !retry_send(mad_send_wr))
2242                         continue;
2243
2244                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2245
2246                 if (mad_send_wr->status == IB_WC_SUCCESS)
2247                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2248                 else
2249                         mad_send_wc.status = mad_send_wr->status;
2250                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2251                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2252                                                    &mad_send_wc);
2253
2254                 atomic_dec(&mad_agent_priv->refcount);
2255                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2256         }
2257         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2258 }
2259
2260 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2261 {
2262         struct ib_mad_port_private *port_priv = cq->cq_context;
2263
2264         queue_work(port_priv->wq, &port_priv->work);
2265 }
2266
2267 /*
2268  * Allocate receive MADs and post receive WRs for them
2269  */
2270 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2271                                     struct ib_mad_private *mad)
2272 {
2273         unsigned long flags;
2274         int post, ret;
2275         struct ib_mad_private *mad_priv;
2276         struct ib_sge sg_list;
2277         struct ib_recv_wr recv_wr, *bad_recv_wr;
2278         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2279
2280         /* Initialize common scatter list fields */
2281         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2282         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2283
2284         /* Initialize common receive WR fields */
2285         recv_wr.next = NULL;
2286         recv_wr.sg_list = &sg_list;
2287         recv_wr.num_sge = 1;
2288
2289         do {
2290                 /* Allocate and map receive buffer */
2291                 if (mad) {
2292                         mad_priv = mad;
2293                         mad = NULL;
2294                 } else {
2295                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2296                         if (!mad_priv) {
2297                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2298                                 ret = -ENOMEM;
2299                                 break;
2300                         }
2301                 }
2302                 sg_list.addr = dma_map_single(qp_info->port_priv->
2303                                                 device->dma_device,
2304                                         &mad_priv->grh,
2305                                         sizeof *mad_priv -
2306                                                 sizeof mad_priv->header,
2307                                         DMA_FROM_DEVICE);
2308                 pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2309                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2310                 mad_priv->header.mad_list.mad_queue = recv_queue;
2311
2312                 /* Post receive WR */
2313                 spin_lock_irqsave(&recv_queue->lock, flags);
2314                 post = (++recv_queue->count < recv_queue->max_active);
2315                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2316                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2317                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2318                 if (ret) {
2319                         spin_lock_irqsave(&recv_queue->lock, flags);
2320                         list_del(&mad_priv->header.mad_list.list);
2321                         recv_queue->count--;
2322                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2323                         dma_unmap_single(qp_info->port_priv->device->dma_device,
2324                                          pci_unmap_addr(&mad_priv->header,
2325                                                         mapping),
2326                                          sizeof *mad_priv -
2327                                            sizeof mad_priv->header,
2328                                          DMA_FROM_DEVICE);
2329                         kmem_cache_free(ib_mad_cache, mad_priv);
2330                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2331                         break;
2332                 }
2333         } while (post);
2334
2335         return ret;
2336 }
2337
2338 /*
2339  * Return all the posted receive MADs
2340  */
2341 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2342 {
2343         struct ib_mad_private_header *mad_priv_hdr;
2344         struct ib_mad_private *recv;
2345         struct ib_mad_list_head *mad_list;
2346
2347         while (!list_empty(&qp_info->recv_queue.list)) {
2348
2349                 mad_list = list_entry(qp_info->recv_queue.list.next,
2350                                       struct ib_mad_list_head, list);
2351                 mad_priv_hdr = container_of(mad_list,
2352                                             struct ib_mad_private_header,
2353                                             mad_list);
2354                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2355                                     header);
2356
2357                 /* Remove from posted receive MAD list */
2358                 list_del(&mad_list->list);
2359
2360                 dma_unmap_single(qp_info->port_priv->device->dma_device,
2361                                  pci_unmap_addr(&recv->header, mapping),
2362                                  sizeof(struct ib_mad_private) -
2363                                  sizeof(struct ib_mad_private_header),
2364                                  DMA_FROM_DEVICE);
2365                 kmem_cache_free(ib_mad_cache, recv);
2366         }
2367
2368         qp_info->recv_queue.count = 0;
2369 }
2370
2371 /*
2372  * Start the port
2373  */
2374 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2375 {
2376         int ret, i;
2377         struct ib_qp_attr *attr;
2378         struct ib_qp *qp;
2379
2380         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2381         if (!attr) {
2382                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2383                 return -ENOMEM;
2384         }
2385
2386         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2387                 qp = port_priv->qp_info[i].qp;
2388                 /*
2389                  * PKey index for QP1 is irrelevant but
2390                  * one is needed for the Reset to Init transition
2391                  */
2392                 attr->qp_state = IB_QPS_INIT;
2393                 attr->pkey_index = 0;
2394                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2395                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2396                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2397                 if (ret) {
2398                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2399                                "INIT: %d\n", i, ret);
2400                         goto out;
2401                 }
2402
2403                 attr->qp_state = IB_QPS_RTR;
2404                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2405                 if (ret) {
2406                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2407                                "RTR: %d\n", i, ret);
2408                         goto out;
2409                 }
2410
2411                 attr->qp_state = IB_QPS_RTS;
2412                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2413                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2414                 if (ret) {
2415                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2416                                "RTS: %d\n", i, ret);
2417                         goto out;
2418                 }
2419         }
2420
2421         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2422         if (ret) {
2423                 printk(KERN_ERR PFX "Failed to request completion "
2424                        "notification: %d\n", ret);
2425                 goto out;
2426         }
2427
2428         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2429                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2430                 if (ret) {
2431                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2432                         goto out;
2433                 }
2434         }
2435 out:
2436         kfree(attr);
2437         return ret;
2438 }
2439
2440 static void qp_event_handler(struct ib_event *event, void *qp_context)
2441 {
2442         struct ib_mad_qp_info   *qp_info = qp_context;
2443
2444         /* It's worse than that! He's dead, Jim! */
2445         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2446                 event->event, qp_info->qp->qp_num);
2447 }
2448
2449 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2450                            struct ib_mad_queue *mad_queue)
2451 {
2452         mad_queue->qp_info = qp_info;
2453         mad_queue->count = 0;
2454         spin_lock_init(&mad_queue->lock);
2455         INIT_LIST_HEAD(&mad_queue->list);
2456 }
2457
2458 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2459                         struct ib_mad_qp_info *qp_info)
2460 {
2461         qp_info->port_priv = port_priv;
2462         init_mad_queue(qp_info, &qp_info->send_queue);
2463         init_mad_queue(qp_info, &qp_info->recv_queue);
2464         INIT_LIST_HEAD(&qp_info->overflow_list);
2465         spin_lock_init(&qp_info->snoop_lock);
2466         qp_info->snoop_table = NULL;
2467         qp_info->snoop_table_size = 0;
2468         atomic_set(&qp_info->snoop_count, 0);
2469 }
2470
2471 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2472                          enum ib_qp_type qp_type)
2473 {
2474         struct ib_qp_init_attr  qp_init_attr;
2475         int ret;
2476
2477         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2478         qp_init_attr.send_cq = qp_info->port_priv->cq;
2479         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2480         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2481         qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2482         qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2483         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2484         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2485         qp_init_attr.qp_type = qp_type;
2486         qp_init_attr.port_num = qp_info->port_priv->port_num;
2487         qp_init_attr.qp_context = qp_info;
2488         qp_init_attr.event_handler = qp_event_handler;
2489         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2490         if (IS_ERR(qp_info->qp)) {
2491                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2492                        get_spl_qp_index(qp_type));
2493                 ret = PTR_ERR(qp_info->qp);
2494                 goto error;
2495         }
2496         /* Use minimum queue sizes unless the CQ is resized */
2497         qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2498         qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2499         return 0;
2500
2501 error:
2502         return ret;
2503 }
2504
2505 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2506 {
2507         ib_destroy_qp(qp_info->qp);
2508         kfree(qp_info->snoop_table);
2509 }
2510
2511 /*
2512  * Open the port
2513  * Create the QP, PD, MR, and CQ if needed
2514  */
2515 static int ib_mad_port_open(struct ib_device *device,
2516                             int port_num)
2517 {
2518         int ret, cq_size;
2519         struct ib_mad_port_private *port_priv;
2520         unsigned long flags;
2521         char name[sizeof "ib_mad123"];
2522
2523         /* Create new device info */
2524         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2525         if (!port_priv) {
2526                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2527                 return -ENOMEM;
2528         }
2529
2530         port_priv->device = device;
2531         port_priv->port_num = port_num;
2532         spin_lock_init(&port_priv->reg_lock);
2533         INIT_LIST_HEAD(&port_priv->agent_list);
2534         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2535         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2536
2537         cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2538         port_priv->cq = ib_create_cq(port_priv->device,
2539                                      ib_mad_thread_completion_handler,
2540                                      NULL, port_priv, cq_size);
2541         if (IS_ERR(port_priv->cq)) {
2542                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2543                 ret = PTR_ERR(port_priv->cq);
2544                 goto error3;
2545         }
2546
2547         port_priv->pd = ib_alloc_pd(device);
2548         if (IS_ERR(port_priv->pd)) {
2549                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2550                 ret = PTR_ERR(port_priv->pd);
2551                 goto error4;
2552         }
2553
2554         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2555         if (IS_ERR(port_priv->mr)) {
2556                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2557                 ret = PTR_ERR(port_priv->mr);
2558                 goto error5;
2559         }
2560
2561         ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2562         if (ret)
2563                 goto error6;
2564         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2565         if (ret)
2566                 goto error7;
2567
2568         snprintf(name, sizeof name, "ib_mad%d", port_num);
2569         port_priv->wq = create_singlethread_workqueue(name);
2570         if (!port_priv->wq) {
2571                 ret = -ENOMEM;
2572                 goto error8;
2573         }
2574         INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2575
2576         ret = ib_mad_port_start(port_priv);
2577         if (ret) {
2578                 printk(KERN_ERR PFX "Couldn't start port\n");
2579                 goto error9;
2580         }
2581
2582         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2583         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2584         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2585         return 0;
2586
2587 error9:
2588         destroy_workqueue(port_priv->wq);
2589 error8:
2590         destroy_mad_qp(&port_priv->qp_info[1]);
2591 error7:
2592         destroy_mad_qp(&port_priv->qp_info[0]);
2593 error6:
2594         ib_dereg_mr(port_priv->mr);
2595 error5:
2596         ib_dealloc_pd(port_priv->pd);
2597 error4:
2598         ib_destroy_cq(port_priv->cq);
2599         cleanup_recv_queue(&port_priv->qp_info[1]);
2600         cleanup_recv_queue(&port_priv->qp_info[0]);
2601 error3:
2602         kfree(port_priv);
2603
2604         return ret;
2605 }
2606
2607 /*
2608  * Close the port
2609  * If there are no classes using the port, free the port
2610  * resources (CQ, MR, PD, QP) and remove the port's info structure
2611  */
2612 static int ib_mad_port_close(struct ib_device *device, int port_num)
2613 {
2614         struct ib_mad_port_private *port_priv;
2615         unsigned long flags;
2616
2617         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2618         port_priv = __ib_get_mad_port(device, port_num);
2619         if (port_priv == NULL) {
2620                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2621                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2622                 return -ENODEV;
2623         }
2624         list_del(&port_priv->port_list);
2625         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2626
2627         /* Stop processing completions. */
2628         flush_workqueue(port_priv->wq);
2629         destroy_workqueue(port_priv->wq);
2630         destroy_mad_qp(&port_priv->qp_info[1]);
2631         destroy_mad_qp(&port_priv->qp_info[0]);
2632         ib_dereg_mr(port_priv->mr);
2633         ib_dealloc_pd(port_priv->pd);
2634         ib_destroy_cq(port_priv->cq);
2635         cleanup_recv_queue(&port_priv->qp_info[1]);
2636         cleanup_recv_queue(&port_priv->qp_info[0]);
2637         /* XXX: Handle deallocation of MAD registration tables */
2638
2639         kfree(port_priv);
2640
2641         return 0;
2642 }
2643
2644 static void ib_mad_init_device(struct ib_device *device)
2645 {
2646         int start, end, i;
2647
2648         if (device->node_type == IB_NODE_SWITCH) {
2649                 start = 0;
2650                 end   = 0;
2651         } else {
2652                 start = 1;
2653                 end   = device->phys_port_cnt;
2654         }
2655
2656         for (i = start; i <= end; i++) {
2657                 if (ib_mad_port_open(device, i)) {
2658                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2659                                device->name, i);
2660                         goto error;
2661                 }
2662                 if (ib_agent_port_open(device, i)) {
2663                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2664                                "for agents\n",
2665                                device->name, i);
2666                         goto error_agent;
2667                 }
2668         }
2669         return;
2670
2671 error_agent:
2672         if (ib_mad_port_close(device, i))
2673                 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2674                        device->name, i);
2675
2676 error:
2677         i--;
2678
2679         while (i >= start) {
2680                 if (ib_agent_port_close(device, i))
2681                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2682                                "for agents\n",
2683                                device->name, i);
2684                 if (ib_mad_port_close(device, i))
2685                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2686                                device->name, i);
2687                 i--;
2688         }
2689 }
2690
2691 static void ib_mad_remove_device(struct ib_device *device)
2692 {
2693         int i, num_ports, cur_port;
2694
2695         if (device->node_type == IB_NODE_SWITCH) {
2696                 num_ports = 1;
2697                 cur_port = 0;
2698         } else {
2699                 num_ports = device->phys_port_cnt;
2700                 cur_port = 1;
2701         }
2702         for (i = 0; i < num_ports; i++, cur_port++) {
2703                 if (ib_agent_port_close(device, cur_port))
2704                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2705                                "for agents\n",
2706                                device->name, cur_port);
2707                 if (ib_mad_port_close(device, cur_port))
2708                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2709                                device->name, cur_port);
2710         }
2711 }
2712
2713 static struct ib_client mad_client = {
2714         .name   = "mad",
2715         .add = ib_mad_init_device,
2716         .remove = ib_mad_remove_device
2717 };
2718
2719 static int __init ib_mad_init_module(void)
2720 {
2721         int ret;
2722
2723         spin_lock_init(&ib_mad_port_list_lock);
2724
2725         ib_mad_cache = kmem_cache_create("ib_mad",
2726                                          sizeof(struct ib_mad_private),
2727                                          0,
2728                                          SLAB_HWCACHE_ALIGN,
2729                                          NULL,
2730                                          NULL);
2731         if (!ib_mad_cache) {
2732                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2733                 ret = -ENOMEM;
2734                 goto error1;
2735         }
2736
2737         INIT_LIST_HEAD(&ib_mad_port_list);
2738
2739         if (ib_register_client(&mad_client)) {
2740                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2741                 ret = -EINVAL;
2742                 goto error2;
2743         }
2744
2745         return 0;
2746
2747 error2:
2748         kmem_cache_destroy(ib_mad_cache);
2749 error1:
2750         return ret;
2751 }
2752
2753 static void __exit ib_mad_cleanup_module(void)
2754 {
2755         ib_unregister_client(&mad_client);
2756
2757         if (kmem_cache_destroy(ib_mad_cache)) {
2758                 printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2759         }
2760 }
2761
2762 module_init(ib_mad_init_module);
2763 module_exit(ib_mad_cleanup_module);
2764