err.no Git - linux-2.6/blob - drivers/infiniband/core/mad.c

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