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