git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/core/mad.c

   1 /*
   2  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
   3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  *
  34  * $Id: mad.c 5596 2006-03-03 01:00:07Z sean.hefty $
  35  */
  36 #include <linux/dma-mapping.h>
  37 #include <rdma/ib_cache.h>
  38
  39 #include "mad_priv.h"
  40 #include "mad_rmpp.h"
  41 #include "smi.h"
  42 #include "agent.h"
  43
  44 MODULE_LICENSE("Dual BSD/GPL");
  45 MODULE_DESCRIPTION("kernel IB MAD API");
  46 MODULE_AUTHOR("Hal Rosenstock");
  47 MODULE_AUTHOR("Sean Hefty");
  48
  49 static struct kmem_cache *ib_mad_cache;
  50
  51 static struct list_head ib_mad_port_list;
  52 static u32 ib_mad_client_id = 0;
  53
  54 /* Port list lock */
  55 static spinlock_t ib_mad_port_list_lock;
  56
  57
  58 /* Forward declarations */
  59 static int method_in_use(struct ib_mad_mgmt_method_table **method,
  60                          struct ib_mad_reg_req *mad_reg_req);
  61 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
  62 static struct ib_mad_agent_private *find_mad_agent(
  63                                         struct ib_mad_port_private *port_priv,
  64                                         struct ib_mad *mad);
  65 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
  66                                     struct ib_mad_private *mad);
  67 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
  68 static void timeout_sends(struct work_struct *work);
  69 static void local_completions(struct work_struct *work);
  70 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  71                               struct ib_mad_agent_private *agent_priv,
  72                               u8 mgmt_class);
  73 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  74                            struct ib_mad_agent_private *agent_priv);
  75
  76 /*
  77  * Returns a ib_mad_port_private structure or NULL for a device/port
  78  * Assumes ib_mad_port_list_lock is being held
  79  */
  80 static inline struct ib_mad_port_private *
  81 __ib_get_mad_port(struct ib_device *device, int port_num)
  82 {
  83         struct ib_mad_port_private *entry;
  84
  85         list_for_each_entry(entry, &ib_mad_port_list, port_list) {
  86                 if (entry->device == device && entry->port_num == port_num)
  87                         return entry;
  88         }
  89         return NULL;
  90 }
  91
  92 /*
  93  * Wrapper function to return a ib_mad_port_private structure or NULL
  94  * for a device/port
  95  */
  96 static inline struct ib_mad_port_private *
  97 ib_get_mad_port(struct ib_device *device, int port_num)
  98 {
  99         struct ib_mad_port_private *entry;
 100         unsigned long flags;
 101
 102         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
 103         entry = __ib_get_mad_port(device, port_num);
 104         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
 105
 106         return entry;
 107 }
 108
 109 static inline u8 convert_mgmt_class(u8 mgmt_class)
 110 {
 111         /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
 112         return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
 113                 0 : mgmt_class;
 114 }
 115
 116 static int get_spl_qp_index(enum ib_qp_type qp_type)
 117 {
 118         switch (qp_type)
 119         {
 120         case IB_QPT_SMI:
 121                 return 0;
 122         case IB_QPT_GSI:
 123                 return 1;
 124         default:
 125                 return -1;
 126         }
 127 }
 128
 129 static int vendor_class_index(u8 mgmt_class)
 130 {
 131         return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
 132 }
 133
 134 static int is_vendor_class(u8 mgmt_class)
 135 {
 136         if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
 137             (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
 138                 return 0;
 139         return 1;
 140 }
 141
 142 static int is_vendor_oui(char *oui)
 143 {
 144         if (oui[0] || oui[1] || oui[2])
 145                 return 1;
 146         return 0;
 147 }
 148
 149 static int is_vendor_method_in_use(
 150                 struct ib_mad_mgmt_vendor_class *vendor_class,
 151                 struct ib_mad_reg_req *mad_reg_req)
 152 {
 153         struct ib_mad_mgmt_method_table *method;
 154         int i;
 155
 156         for (i = 0; i < MAX_MGMT_OUI; i++) {
 157                 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
 158                         method = vendor_class->method_table[i];
 159                         if (method) {
 160                                 if (method_in_use(&method, mad_reg_req))
 161                                         return 1;
 162                                 else
 163                                         break;
 164                         }
 165                 }
 166         }
 167         return 0;
 168 }
 169
 170 int ib_response_mad(struct ib_mad *mad)
 171 {
 172         return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
 173                 (mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
 174                 ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
 175                  (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
 176 }
 177 EXPORT_SYMBOL(ib_response_mad);
 178
 179 /*
 180  * ib_register_mad_agent - Register to send/receive MADs
 181  */
 182 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 183                                            u8 port_num,
 184                                            enum ib_qp_type qp_type,
 185                                            struct ib_mad_reg_req *mad_reg_req,
 186                                            u8 rmpp_version,
 187                                            ib_mad_send_handler send_handler,
 188                                            ib_mad_recv_handler recv_handler,
 189                                            void *context)
 190 {
 191         struct ib_mad_port_private *port_priv;
 192         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
 193         struct ib_mad_agent_private *mad_agent_priv;
 194         struct ib_mad_reg_req *reg_req = NULL;
 195         struct ib_mad_mgmt_class_table *class;
 196         struct ib_mad_mgmt_vendor_class_table *vendor;
 197         struct ib_mad_mgmt_vendor_class *vendor_class;
 198         struct ib_mad_mgmt_method_table *method;
 199         int ret2, qpn;
 200         unsigned long flags;
 201         u8 mgmt_class, vclass;
 202
 203         /* Validate parameters */
 204         qpn = get_spl_qp_index(qp_type);
 205         if (qpn == -1)
 206                 goto error1;
 207
 208         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
 209                 goto error1;
 210
 211         /* Validate MAD registration request if supplied */
 212         if (mad_reg_req) {
 213                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
 214                         goto error1;
 215                 if (!recv_handler)
 216                         goto error1;
 217                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
 218                         /*
 219                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
 220                          * one in this range currently allowed
 221                          */
 222                         if (mad_reg_req->mgmt_class !=
 223                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 224                                 goto error1;
 225                 } else if (mad_reg_req->mgmt_class == 0) {
 226                         /*
 227                          * Class 0 is reserved in IBA and is used for
 228                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
 229                          */
 230                         goto error1;
 231                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
 232                         /*
 233                          * If class is in "new" vendor range,
 234                          * ensure supplied OUI is not zero
 235                          */
 236                         if (!is_vendor_oui(mad_reg_req->oui))
 237                                 goto error1;
 238                 }
 239                 /* Make sure class supplied is consistent with RMPP */
 240                 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
 241                         if (rmpp_version)
 242                                 goto error1;
 243                 }
 244                 /* Make sure class supplied is consistent with QP type */
 245                 if (qp_type == IB_QPT_SMI) {
 246                         if ((mad_reg_req->mgmt_class !=
 247                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
 248                             (mad_reg_req->mgmt_class !=
 249                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 250                                 goto error1;
 251                 } else {
 252                         if ((mad_reg_req->mgmt_class ==
 253                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
 254                             (mad_reg_req->mgmt_class ==
 255                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 256                                 goto error1;
 257                 }
 258         } else {
 259                 /* No registration request supplied */
 260                 if (!send_handler)
 261                         goto error1;
 262         }
 263
 264         /* Validate device and port */
 265         port_priv = ib_get_mad_port(device, port_num);
 266         if (!port_priv) {
 267                 ret = ERR_PTR(-ENODEV);
 268                 goto error1;
 269         }
 270
 271         /* Allocate structures */
 272         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
 273         if (!mad_agent_priv) {
 274                 ret = ERR_PTR(-ENOMEM);
 275                 goto error1;
 276         }
 277
 278         mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
 279                                                  IB_ACCESS_LOCAL_WRITE);
 280         if (IS_ERR(mad_agent_priv->agent.mr)) {
 281                 ret = ERR_PTR(-ENOMEM);
 282                 goto error2;
 283         }
 284
 285         if (mad_reg_req) {
 286                 reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
 287                 if (!reg_req) {
 288                         ret = ERR_PTR(-ENOMEM);
 289                         goto error3;
 290                 }
 291                 /* Make a copy of the MAD registration request */
 292                 memcpy(reg_req, mad_reg_req, sizeof *reg_req);
 293         }
 294
 295         /* Now, fill in the various structures */
 296         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 297         mad_agent_priv->reg_req = reg_req;
 298         mad_agent_priv->agent.rmpp_version = rmpp_version;
 299         mad_agent_priv->agent.device = device;
 300         mad_agent_priv->agent.recv_handler = recv_handler;
 301         mad_agent_priv->agent.send_handler = send_handler;
 302         mad_agent_priv->agent.context = context;
 303         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 304         mad_agent_priv->agent.port_num = port_num;
 305
 306         spin_lock_irqsave(&port_priv->reg_lock, flags);
 307         mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
 308
 309         /*
 310          * Make sure MAD registration (if supplied)
 311          * is non overlapping with any existing ones
 312          */
 313         if (mad_reg_req) {
 314                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 315                 if (!is_vendor_class(mgmt_class)) {
 316                         class = port_priv->version[mad_reg_req->
 317                                                    mgmt_class_version].class;
 318                         if (class) {
 319                                 method = class->method_table[mgmt_class];
 320                                 if (method) {
 321                                         if (method_in_use(&method,
 322                                                            mad_reg_req))
 323                                                 goto error4;
 324                                 }
 325                         }
 326                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 327                                                   mgmt_class);
 328                 } else {
 329                         /* "New" vendor class range */
 330                         vendor = port_priv->version[mad_reg_req->
 331                                                     mgmt_class_version].vendor;
 332                         if (vendor) {
 333                                 vclass = vendor_class_index(mgmt_class);
 334                                 vendor_class = vendor->vendor_class[vclass];
 335                                 if (vendor_class) {
 336                                         if (is_vendor_method_in_use(
 337                                                         vendor_class,
 338                                                         mad_reg_req))
 339                                                 goto error4;
 340                                 }
 341                         }
 342                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 343                 }
 344                 if (ret2) {
 345                         ret = ERR_PTR(ret2);
 346                         goto error4;
 347                 }
 348         }
 349
 350         /* Add mad agent into port's agent list */
 351         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
 352         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 353
 354         spin_lock_init(&mad_agent_priv->lock);
 355         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 356         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 357         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 358         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 359         INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
 360         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 361         INIT_WORK(&mad_agent_priv->local_work, local_completions);
 362         atomic_set(&mad_agent_priv->refcount, 1);
 363         init_completion(&mad_agent_priv->comp);
 364
 365         return &mad_agent_priv->agent;
 366
 367 error4:
 368         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 369         kfree(reg_req);
 370 error3:
 371         ib_dereg_mr(mad_agent_priv->agent.mr);
 372 error2:
 373         kfree(mad_agent_priv);
 374 error1:
 375         return ret;
 376 }
 377 EXPORT_SYMBOL(ib_register_mad_agent);
 378
 379 static inline int is_snooping_sends(int mad_snoop_flags)
 380 {
 381         return (mad_snoop_flags &
 382                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 383                  IB_MAD_SNOOP_RMPP_SENDS |*/
 384                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 385                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 386 }
 387
 388 static inline int is_snooping_recvs(int mad_snoop_flags)
 389 {
 390         return (mad_snoop_flags &
 391                 (IB_MAD_SNOOP_RECVS /*|
 392                  IB_MAD_SNOOP_RMPP_RECVS*/));
 393 }
 394
 395 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 396                                 struct ib_mad_snoop_private *mad_snoop_priv)
 397 {
 398         struct ib_mad_snoop_private **new_snoop_table;
 399         unsigned long flags;
 400         int i;
 401
 402         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 403         /* Check for empty slot in array. */
 404         for (i = 0; i < qp_info->snoop_table_size; i++)
 405                 if (!qp_info->snoop_table[i])
 406                         break;
 407
 408         if (i == qp_info->snoop_table_size) {
 409                 /* Grow table. */
 410                 new_snoop_table = kmalloc(sizeof mad_snoop_priv *
 411                                           qp_info->snoop_table_size + 1,
 412                                           GFP_ATOMIC);
 413                 if (!new_snoop_table) {
 414                         i = -ENOMEM;
 415                         goto out;
 416                 }
 417                 if (qp_info->snoop_table) {
 418                         memcpy(new_snoop_table, qp_info->snoop_table,
 419                                sizeof mad_snoop_priv *
 420                                qp_info->snoop_table_size);
 421                         kfree(qp_info->snoop_table);
 422                 }
 423                 qp_info->snoop_table = new_snoop_table;
 424                 qp_info->snoop_table_size++;
 425         }
 426         qp_info->snoop_table[i] = mad_snoop_priv;
 427         atomic_inc(&qp_info->snoop_count);
 428 out:
 429         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 430         return i;
 431 }
 432
 433 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 434                                            u8 port_num,
 435                                            enum ib_qp_type qp_type,
 436                                            int mad_snoop_flags,
 437                                            ib_mad_snoop_handler snoop_handler,
 438                                            ib_mad_recv_handler recv_handler,
 439                                            void *context)
 440 {
 441         struct ib_mad_port_private *port_priv;
 442         struct ib_mad_agent *ret;
 443         struct ib_mad_snoop_private *mad_snoop_priv;
 444         int qpn;
 445
 446         /* Validate parameters */
 447         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 448             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 449                 ret = ERR_PTR(-EINVAL);
 450                 goto error1;
 451         }
 452         qpn = get_spl_qp_index(qp_type);
 453         if (qpn == -1) {
 454                 ret = ERR_PTR(-EINVAL);
 455                 goto error1;
 456         }
 457         port_priv = ib_get_mad_port(device, port_num);
 458         if (!port_priv) {
 459                 ret = ERR_PTR(-ENODEV);
 460                 goto error1;
 461         }
 462         /* Allocate structures */
 463         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 464         if (!mad_snoop_priv) {
 465                 ret = ERR_PTR(-ENOMEM);
 466                 goto error1;
 467         }
 468
 469         /* Now, fill in the various structures */
 470         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 471         mad_snoop_priv->agent.device = device;
 472         mad_snoop_priv->agent.recv_handler = recv_handler;
 473         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 474         mad_snoop_priv->agent.context = context;
 475         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 476         mad_snoop_priv->agent.port_num = port_num;
 477         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 478         init_completion(&mad_snoop_priv->comp);
 479         mad_snoop_priv->snoop_index = register_snoop_agent(
 480                                                 &port_priv->qp_info[qpn],
 481                                                 mad_snoop_priv);
 482         if (mad_snoop_priv->snoop_index < 0) {
 483                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 484                 goto error2;
 485         }
 486
 487         atomic_set(&mad_snoop_priv->refcount, 1);
 488         return &mad_snoop_priv->agent;
 489
 490 error2:
 491         kfree(mad_snoop_priv);
 492 error1:
 493         return ret;
 494 }
 495 EXPORT_SYMBOL(ib_register_mad_snoop);
 496
 497 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 498 {
 499         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 500                 complete(&mad_agent_priv->comp);
 501 }
 502
 503 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
 504 {
 505         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 506                 complete(&mad_snoop_priv->comp);
 507 }
 508
 509 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 510 {
 511         struct ib_mad_port_private *port_priv;
 512         unsigned long flags;
 513
 514         /* Note that we could still be handling received MADs */
 515
 516         /*
 517          * Canceling all sends results in dropping received response
 518          * MADs, preventing us from queuing additional work
 519          */
 520         cancel_mads(mad_agent_priv);
 521         port_priv = mad_agent_priv->qp_info->port_priv;
 522         cancel_delayed_work(&mad_agent_priv->timed_work);
 523
 524         spin_lock_irqsave(&port_priv->reg_lock, flags);
 525         remove_mad_reg_req(mad_agent_priv);
 526         list_del(&mad_agent_priv->agent_list);
 527         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 528
 529         flush_workqueue(port_priv->wq);
 530         ib_cancel_rmpp_recvs(mad_agent_priv);
 531
 532         deref_mad_agent(mad_agent_priv);
 533         wait_for_completion(&mad_agent_priv->comp);
 534
 535         kfree(mad_agent_priv->reg_req);
 536         ib_dereg_mr(mad_agent_priv->agent.mr);
 537         kfree(mad_agent_priv);
 538 }
 539
 540 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 541 {
 542         struct ib_mad_qp_info *qp_info;
 543         unsigned long flags;
 544
 545         qp_info = mad_snoop_priv->qp_info;
 546         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 547         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 548         atomic_dec(&qp_info->snoop_count);
 549         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 550
 551         deref_snoop_agent(mad_snoop_priv);
 552         wait_for_completion(&mad_snoop_priv->comp);
 553
 554         kfree(mad_snoop_priv);
 555 }
 556
 557 /*
 558  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 559  */
 560 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 561 {
 562         struct ib_mad_agent_private *mad_agent_priv;
 563         struct ib_mad_snoop_private *mad_snoop_priv;
 564
 565         /* If the TID is zero, the agent can only snoop. */
 566         if (mad_agent->hi_tid) {
 567                 mad_agent_priv = container_of(mad_agent,
 568                                               struct ib_mad_agent_private,
 569                                               agent);
 570                 unregister_mad_agent(mad_agent_priv);
 571         } else {
 572                 mad_snoop_priv = container_of(mad_agent,
 573                                               struct ib_mad_snoop_private,
 574                                               agent);
 575                 unregister_mad_snoop(mad_snoop_priv);
 576         }
 577         return 0;
 578 }
 579 EXPORT_SYMBOL(ib_unregister_mad_agent);
 580
 581 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 582 {
 583         struct ib_mad_queue *mad_queue;
 584         unsigned long flags;
 585
 586         BUG_ON(!mad_list->mad_queue);
 587         mad_queue = mad_list->mad_queue;
 588         spin_lock_irqsave(&mad_queue->lock, flags);
 589         list_del(&mad_list->list);
 590         mad_queue->count--;
 591         spin_unlock_irqrestore(&mad_queue->lock, flags);
 592 }
 593
 594 static void snoop_send(struct ib_mad_qp_info *qp_info,
 595                        struct ib_mad_send_buf *send_buf,
 596                        struct ib_mad_send_wc *mad_send_wc,
 597                        int mad_snoop_flags)
 598 {
 599         struct ib_mad_snoop_private *mad_snoop_priv;
 600         unsigned long flags;
 601         int i;
 602
 603         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 604         for (i = 0; i < qp_info->snoop_table_size; i++) {
 605                 mad_snoop_priv = qp_info->snoop_table[i];
 606                 if (!mad_snoop_priv ||
 607                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 608                         continue;
 609
 610                 atomic_inc(&mad_snoop_priv->refcount);
 611                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 612                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 613                                                     send_buf, mad_send_wc);
 614                 deref_snoop_agent(mad_snoop_priv);
 615                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 616         }
 617         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 618 }
 619
 620 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 621                        struct ib_mad_recv_wc *mad_recv_wc,
 622                        int mad_snoop_flags)
 623 {
 624         struct ib_mad_snoop_private *mad_snoop_priv;
 625         unsigned long flags;
 626         int i;
 627
 628         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 629         for (i = 0; i < qp_info->snoop_table_size; i++) {
 630                 mad_snoop_priv = qp_info->snoop_table[i];
 631                 if (!mad_snoop_priv ||
 632                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 633                         continue;
 634
 635                 atomic_inc(&mad_snoop_priv->refcount);
 636                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 637                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
 638                                                    mad_recv_wc);
 639                 deref_snoop_agent(mad_snoop_priv);
 640                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 641         }
 642         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 643 }
 644
 645 static void build_smp_wc(u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
 646                          struct ib_wc *wc)
 647 {
 648         memset(wc, 0, sizeof *wc);
 649         wc->wr_id = wr_id;
 650         wc->status = IB_WC_SUCCESS;
 651         wc->opcode = IB_WC_RECV;
 652         wc->pkey_index = pkey_index;
 653         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
 654         wc->src_qp = IB_QP0;
 655         wc->qp_num = IB_QP0;
 656         wc->slid = slid;
 657         wc->sl = 0;
 658         wc->dlid_path_bits = 0;
 659         wc->port_num = port_num;
 660 }
 661
 662 /*
 663  * Return 0 if SMP is to be sent
 664  * Return 1 if SMP was consumed locally (whether or not solicited)
 665  * Return < 0 if error
 666  */
 667 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 668                                   struct ib_mad_send_wr_private *mad_send_wr)
 669 {
 670         int ret;
 671         struct ib_smp *smp = mad_send_wr->send_buf.mad;
 672         unsigned long flags;
 673         struct ib_mad_local_private *local;
 674         struct ib_mad_private *mad_priv;
 675         struct ib_mad_port_private *port_priv;
 676         struct ib_mad_agent_private *recv_mad_agent = NULL;
 677         struct ib_device *device = mad_agent_priv->agent.device;
 678         u8 port_num = mad_agent_priv->agent.port_num;
 679         struct ib_wc mad_wc;
 680         struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
 681
 682         /*
 683          * Directed route handling starts if the initial LID routed part of
 684          * a request or the ending LID routed part of a response is empty.
 685          * If we are at the start of the LID routed part, don't update the
 686          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
 687          */
 688         if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
 689              IB_LID_PERMISSIVE &&
 690             !smi_handle_dr_smp_send(smp, device->node_type, port_num)) {
 691                 ret = -EINVAL;
 692                 printk(KERN_ERR PFX "Invalid directed route\n");
 693                 goto out;
 694         }
 695         /* Check to post send on QP or process locally */
 696         ret = smi_check_local_smp(smp, device);
 697         if (!ret)
 698                 goto out;
 699
 700         local = kmalloc(sizeof *local, GFP_ATOMIC);
 701         if (!local) {
 702                 ret = -ENOMEM;
 703                 printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
 704                 goto out;
 705         }
 706         local->mad_priv = NULL;
 707         local->recv_mad_agent = NULL;
 708         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
 709         if (!mad_priv) {
 710                 ret = -ENOMEM;
 711                 printk(KERN_ERR PFX "No memory for local response MAD\n");
 712                 kfree(local);
 713                 goto out;
 714         }
 715
 716         build_smp_wc(send_wr->wr_id, be16_to_cpu(smp->dr_slid),
 717                      send_wr->wr.ud.pkey_index,
 718                      send_wr->wr.ud.port_num, &mad_wc);
 719
 720         /* No GRH for DR SMP */
 721         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 722                                   (struct ib_mad *)smp,
 723                                   (struct ib_mad *)&mad_priv->mad);
 724         switch (ret)
 725         {
 726         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 727                 if (ib_response_mad(&mad_priv->mad.mad) &&
 728                     mad_agent_priv->agent.recv_handler) {
 729                         local->mad_priv = mad_priv;
 730                         local->recv_mad_agent = mad_agent_priv;
 731                         /*
 732                          * Reference MAD agent until receive
 733                          * side of local completion handled
 734                          */
 735                         atomic_inc(&mad_agent_priv->refcount);
 736                 } else
 737                         kmem_cache_free(ib_mad_cache, mad_priv);
 738                 break;
 739         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 740                 kmem_cache_free(ib_mad_cache, mad_priv);
 741                 break;
 742         case IB_MAD_RESULT_SUCCESS:
 743                 /* Treat like an incoming receive MAD */
 744                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 745                                             mad_agent_priv->agent.port_num);
 746                 if (port_priv) {
 747                         mad_priv->mad.mad.mad_hdr.tid =
 748                                 ((struct ib_mad *)smp)->mad_hdr.tid;
 749                         recv_mad_agent = find_mad_agent(port_priv,
 750                                                         &mad_priv->mad.mad);
 751                 }
 752                 if (!port_priv || !recv_mad_agent) {
 753                         kmem_cache_free(ib_mad_cache, mad_priv);
 754                         kfree(local);
 755                         ret = 0;
 756                         goto out;
 757                 }
 758                 local->mad_priv = mad_priv;
 759                 local->recv_mad_agent = recv_mad_agent;
 760                 break;
 761         default:
 762                 kmem_cache_free(ib_mad_cache, mad_priv);
 763                 kfree(local);
 764                 ret = -EINVAL;
 765                 goto out;
 766         }
 767
 768         local->mad_send_wr = mad_send_wr;
 769         /* Reference MAD agent until send side of local completion handled */
 770         atomic_inc(&mad_agent_priv->refcount);
 771         /* Queue local completion to local list */
 772         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 773         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 774         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 775         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 776                    &mad_agent_priv->local_work);
 777         ret = 1;
 778 out:
 779         return ret;
 780 }
 781
 782 static int get_pad_size(int hdr_len, int data_len)
 783 {
 784         int seg_size, pad;
 785
 786         seg_size = sizeof(struct ib_mad) - hdr_len;
 787         if (data_len && seg_size) {
 788                 pad = seg_size - data_len % seg_size;
 789                 return pad == seg_size ? 0 : pad;
 790         } else
 791                 return seg_size;
 792 }
 793
 794 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 795 {
 796         struct ib_rmpp_segment *s, *t;
 797
 798         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 799                 list_del(&s->list);
 800                 kfree(s);
 801         }
 802 }
 803
 804 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 805                                 gfp_t gfp_mask)
 806 {
 807         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 808         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 809         struct ib_rmpp_segment *seg = NULL;
 810         int left, seg_size, pad;
 811
 812         send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
 813         seg_size = send_buf->seg_size;
 814         pad = send_wr->pad;
 815
 816         /* Allocate data segments. */
 817         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 818                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 819                 if (!seg) {
 820                         printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
 821                                "alloc failed for len %zd, gfp %#x\n",
 822                                sizeof (*seg) + seg_size, gfp_mask);
 823                         free_send_rmpp_list(send_wr);
 824                         return -ENOMEM;
 825                 }
 826                 seg->num = ++send_buf->seg_count;
 827                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 828         }
 829
 830         /* Zero any padding */
 831         if (pad)
 832                 memset(seg->data + seg_size - pad, 0, pad);
 833
 834         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
 835                                           agent.rmpp_version;
 836         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 837         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
 838
 839         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
 840                                         struct ib_rmpp_segment, list);
 841         send_wr->last_ack_seg = send_wr->cur_seg;
 842         return 0;
 843 }
 844
 845 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 846                                             u32 remote_qpn, u16 pkey_index,
 847                                             int rmpp_active,
 848                                             int hdr_len, int data_len,
 849                                             gfp_t gfp_mask)
 850 {
 851         struct ib_mad_agent_private *mad_agent_priv;
 852         struct ib_mad_send_wr_private *mad_send_wr;
 853         int pad, message_size, ret, size;
 854         void *buf;
 855
 856         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 857                                       agent);
 858         pad = get_pad_size(hdr_len, data_len);
 859         message_size = hdr_len + data_len + pad;
 860
 861         if ((!mad_agent->rmpp_version &&
 862              (rmpp_active || message_size > sizeof(struct ib_mad))) ||
 863             (!rmpp_active && message_size > sizeof(struct ib_mad)))
 864                 return ERR_PTR(-EINVAL);
 865
 866         size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
 867         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
 868         if (!buf)
 869                 return ERR_PTR(-ENOMEM);
 870
 871         mad_send_wr = buf + size;
 872         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
 873         mad_send_wr->send_buf.mad = buf;
 874         mad_send_wr->send_buf.hdr_len = hdr_len;
 875         mad_send_wr->send_buf.data_len = data_len;
 876         mad_send_wr->pad = pad;
 877
 878         mad_send_wr->mad_agent_priv = mad_agent_priv;
 879         mad_send_wr->sg_list[0].length = hdr_len;
 880         mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
 881         mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
 882         mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
 883
 884         mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
 885         mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 886         mad_send_wr->send_wr.num_sge = 2;
 887         mad_send_wr->send_wr.opcode = IB_WR_SEND;
 888         mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
 889         mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
 890         mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 891         mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
 892
 893         if (rmpp_active) {
 894                 ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
 895                 if (ret) {
 896                         kfree(buf);
 897                         return ERR_PTR(ret);
 898                 }
 899         }
 900
 901         mad_send_wr->send_buf.mad_agent = mad_agent;
 902         atomic_inc(&mad_agent_priv->refcount);
 903         return &mad_send_wr->send_buf;
 904 }
 905 EXPORT_SYMBOL(ib_create_send_mad);
 906
 907 int ib_get_mad_data_offset(u8 mgmt_class)
 908 {
 909         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
 910                 return IB_MGMT_SA_HDR;
 911         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 912                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 913                  (mgmt_class == IB_MGMT_CLASS_BIS))
 914                 return IB_MGMT_DEVICE_HDR;
 915         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 916                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
 917                 return IB_MGMT_VENDOR_HDR;
 918         else
 919                 return IB_MGMT_MAD_HDR;
 920 }
 921 EXPORT_SYMBOL(ib_get_mad_data_offset);
 922
 923 int ib_is_mad_class_rmpp(u8 mgmt_class)
 924 {
 925         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
 926             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 927             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 928             (mgmt_class == IB_MGMT_CLASS_BIS) ||
 929             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 930              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
 931                 return 1;
 932         return 0;
 933 }
 934 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
 935
 936 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
 937 {
 938         struct ib_mad_send_wr_private *mad_send_wr;
 939         struct list_head *list;
 940
 941         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 942                                    send_buf);
 943         list = &mad_send_wr->cur_seg->list;
 944
 945         if (mad_send_wr->cur_seg->num < seg_num) {
 946                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
 947                         if (mad_send_wr->cur_seg->num == seg_num)
 948                                 break;
 949         } else if (mad_send_wr->cur_seg->num > seg_num) {
 950                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
 951                         if (mad_send_wr->cur_seg->num == seg_num)
 952                                 break;
 953         }
 954         return mad_send_wr->cur_seg->data;
 955 }
 956 EXPORT_SYMBOL(ib_get_rmpp_segment);
 957
 958 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
 959 {
 960         if (mad_send_wr->send_buf.seg_count)
 961                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
 962                                            mad_send_wr->seg_num);
 963         else
 964                 return mad_send_wr->send_buf.mad +
 965                        mad_send_wr->send_buf.hdr_len;
 966 }
 967
 968 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
 969 {
 970         struct ib_mad_agent_private *mad_agent_priv;
 971         struct ib_mad_send_wr_private *mad_send_wr;
 972
 973         mad_agent_priv = container_of(send_buf->mad_agent,
 974                                       struct ib_mad_agent_private, agent);
 975         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 976                                    send_buf);
 977
 978         free_send_rmpp_list(mad_send_wr);
 979         kfree(send_buf->mad);
 980         deref_mad_agent(mad_agent_priv);
 981 }
 982 EXPORT_SYMBOL(ib_free_send_mad);
 983
 984 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 985 {
 986         struct ib_mad_qp_info *qp_info;
 987         struct list_head *list;
 988         struct ib_send_wr *bad_send_wr;
 989         struct ib_mad_agent *mad_agent;
 990         struct ib_sge *sge;
 991         unsigned long flags;
 992         int ret;
 993
 994         /* Set WR ID to find mad_send_wr upon completion */
 995         qp_info = mad_send_wr->mad_agent_priv->qp_info;
 996         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
 997         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
 998
 999         mad_agent = mad_send_wr->send_buf.mad_agent;
1000         sge = mad_send_wr->sg_list;
1001         sge[0].addr = ib_dma_map_single(mad_agent->device,
1002                                         mad_send_wr->send_buf.mad,
1003                                         sge[0].length,
1004                                         DMA_TO_DEVICE);
1005         mad_send_wr->header_mapping = sge[0].addr;
1006
1007         sge[1].addr = ib_dma_map_single(mad_agent->device,
1008                                         ib_get_payload(mad_send_wr),
1009                                         sge[1].length,
1010                                         DMA_TO_DEVICE);
1011         mad_send_wr->payload_mapping = sge[1].addr;
1012
1013         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1014         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1015                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1016                                    &bad_send_wr);
1017                 list = &qp_info->send_queue.list;
1018         } else {
1019                 ret = 0;
1020                 list = &qp_info->overflow_list;
1021         }
1022
1023         if (!ret) {
1024                 qp_info->send_queue.count++;
1025                 list_add_tail(&mad_send_wr->mad_list.list, list);
1026         }
1027         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1028         if (ret) {
1029                 ib_dma_unmap_single(mad_agent->device,
1030                                     mad_send_wr->header_mapping,
1031                                     sge[0].length, DMA_TO_DEVICE);
1032                 ib_dma_unmap_single(mad_agent->device,
1033                                     mad_send_wr->payload_mapping,
1034                                     sge[1].length, DMA_TO_DEVICE);
1035         }
1036         return ret;
1037 }
1038
1039 /*
1040  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1041  *  with the registered client
1042  */
1043 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1044                      struct ib_mad_send_buf **bad_send_buf)
1045 {
1046         struct ib_mad_agent_private *mad_agent_priv;
1047         struct ib_mad_send_buf *next_send_buf;
1048         struct ib_mad_send_wr_private *mad_send_wr;
1049         unsigned long flags;
1050         int ret = -EINVAL;
1051
1052         /* Walk list of send WRs and post each on send list */
1053         for (; send_buf; send_buf = next_send_buf) {
1054
1055                 mad_send_wr = container_of(send_buf,
1056                                            struct ib_mad_send_wr_private,
1057                                            send_buf);
1058                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1059
1060                 if (!send_buf->mad_agent->send_handler ||
1061                     (send_buf->timeout_ms &&
1062                      !send_buf->mad_agent->recv_handler)) {
1063                         ret = -EINVAL;
1064                         goto error;
1065                 }
1066
1067                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1068                         if (mad_agent_priv->agent.rmpp_version) {
1069                                 ret = -EINVAL;
1070                                 goto error;
1071                         }
1072                 }
1073
1074                 /*
1075                  * Save pointer to next work request to post in case the
1076                  * current one completes, and the user modifies the work
1077                  * request associated with the completion
1078                  */
1079                 next_send_buf = send_buf->next;
1080                 mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1081
1082                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1083                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1084                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1085                                                      mad_send_wr);
1086                         if (ret < 0)            /* error */
1087                                 goto error;
1088                         else if (ret == 1)      /* locally consumed */
1089                                 continue;
1090                 }
1091
1092                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1093                 /* Timeout will be updated after send completes */
1094                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1095                 mad_send_wr->retries = send_buf->retries;
1096                 /* Reference for work request to QP + response */
1097                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1098                 mad_send_wr->status = IB_WC_SUCCESS;
1099
1100                 /* Reference MAD agent until send completes */
1101                 atomic_inc(&mad_agent_priv->refcount);
1102                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1103                 list_add_tail(&mad_send_wr->agent_list,
1104                               &mad_agent_priv->send_list);
1105                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1106
1107                 if (mad_agent_priv->agent.rmpp_version) {
1108                         ret = ib_send_rmpp_mad(mad_send_wr);
1109                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1110                                 ret = ib_send_mad(mad_send_wr);
1111                 } else
1112                         ret = ib_send_mad(mad_send_wr);
1113                 if (ret < 0) {
1114                         /* Fail send request */
1115                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1116                         list_del(&mad_send_wr->agent_list);
1117                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1118                         atomic_dec(&mad_agent_priv->refcount);
1119                         goto error;
1120                 }
1121         }
1122         return 0;
1123 error:
1124         if (bad_send_buf)
1125                 *bad_send_buf = send_buf;
1126         return ret;
1127 }
1128 EXPORT_SYMBOL(ib_post_send_mad);
1129
1130 /*
1131  * ib_free_recv_mad - Returns data buffers used to receive
1132  *  a MAD to the access layer
1133  */
1134 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1135 {
1136         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1137         struct ib_mad_private_header *mad_priv_hdr;
1138         struct ib_mad_private *priv;
1139         struct list_head free_list;
1140
1141         INIT_LIST_HEAD(&free_list);
1142         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1143
1144         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1145                                         &free_list, list) {
1146                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1147                                            recv_buf);
1148                 mad_priv_hdr = container_of(mad_recv_wc,
1149                                             struct ib_mad_private_header,
1150                                             recv_wc);
1151                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1152                                     header);
1153                 kmem_cache_free(ib_mad_cache, priv);
1154         }
1155 }
1156 EXPORT_SYMBOL(ib_free_recv_mad);
1157
1158 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1159                                         u8 rmpp_version,
1160                                         ib_mad_send_handler send_handler,
1161                                         ib_mad_recv_handler recv_handler,
1162                                         void *context)
1163 {
1164         return ERR_PTR(-EINVAL);        /* XXX: for now */
1165 }
1166 EXPORT_SYMBOL(ib_redirect_mad_qp);
1167
1168 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1169                       struct ib_wc *wc)
1170 {
1171         printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1172         return 0;
1173 }
1174 EXPORT_SYMBOL(ib_process_mad_wc);
1175
1176 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1177                          struct ib_mad_reg_req *mad_reg_req)
1178 {
1179         int i;
1180
1181         for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1182              i < IB_MGMT_MAX_METHODS;
1183              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1184                                1+i)) {
1185                 if ((*method)->agent[i]) {
1186                         printk(KERN_ERR PFX "Method %d already in use\n", i);
1187                         return -EINVAL;
1188                 }
1189         }
1190         return 0;
1191 }
1192
1193 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1194 {
1195         /* Allocate management method table */
1196         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1197         if (!*method) {
1198                 printk(KERN_ERR PFX "No memory for "
1199                        "ib_mad_mgmt_method_table\n");
1200                 return -ENOMEM;
1201         }
1202
1203         return 0;
1204 }
1205
1206 /*
1207  * Check to see if there are any methods still in use
1208  */
1209 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1210 {
1211         int i;
1212
1213         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1214                 if (method->agent[i])
1215                         return 1;
1216         return 0;
1217 }
1218
1219 /*
1220  * Check to see if there are any method tables for this class still in use
1221  */
1222 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1223 {
1224         int i;
1225
1226         for (i = 0; i < MAX_MGMT_CLASS; i++)
1227                 if (class->method_table[i])
1228                         return 1;
1229         return 0;
1230 }
1231
1232 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1233 {
1234         int i;
1235
1236         for (i = 0; i < MAX_MGMT_OUI; i++)
1237                 if (vendor_class->method_table[i])
1238                         return 1;
1239         return 0;
1240 }
1241
1242 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1243                            char *oui)
1244 {
1245         int i;
1246
1247         for (i = 0; i < MAX_MGMT_OUI; i++)
1248                 /* Is there matching OUI for this vendor class ? */
1249                 if (!memcmp(vendor_class->oui[i], oui, 3))
1250                         return i;
1251
1252         return -1;
1253 }
1254
1255 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1256 {
1257         int i;
1258
1259         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1260                 if (vendor->vendor_class[i])
1261                         return 1;
1262
1263         return 0;
1264 }
1265
1266 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1267                                      struct ib_mad_agent_private *agent)
1268 {
1269         int i;
1270
1271         /* Remove any methods for this mad agent */
1272         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1273                 if (method->agent[i] == agent) {
1274                         method->agent[i] = NULL;
1275                 }
1276         }
1277 }
1278
1279 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1280                               struct ib_mad_agent_private *agent_priv,
1281                               u8 mgmt_class)
1282 {
1283         struct ib_mad_port_private *port_priv;
1284         struct ib_mad_mgmt_class_table **class;
1285         struct ib_mad_mgmt_method_table **method;
1286         int i, ret;
1287
1288         port_priv = agent_priv->qp_info->port_priv;
1289         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1290         if (!*class) {
1291                 /* Allocate management class table for "new" class version */
1292                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1293                 if (!*class) {
1294                         printk(KERN_ERR PFX "No memory for "
1295                                "ib_mad_mgmt_class_table\n");
1296                         ret = -ENOMEM;
1297                         goto error1;
1298                 }
1299
1300                 /* Allocate method table for this management class */
1301                 method = &(*class)->method_table[mgmt_class];
1302                 if ((ret = allocate_method_table(method)))
1303                         goto error2;
1304         } else {
1305                 method = &(*class)->method_table[mgmt_class];
1306                 if (!*method) {
1307                         /* Allocate method table for this management class */
1308                         if ((ret = allocate_method_table(method)))
1309                                 goto error1;
1310                 }
1311         }
1312
1313         /* Now, make sure methods are not already in use */
1314         if (method_in_use(method, mad_reg_req))
1315                 goto error3;
1316
1317         /* Finally, add in methods being registered */
1318         for (i = find_first_bit(mad_reg_req->method_mask,
1319                                 IB_MGMT_MAX_METHODS);
1320              i < IB_MGMT_MAX_METHODS;
1321              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1322                                1+i)) {
1323                 (*method)->agent[i] = agent_priv;
1324         }
1325         return 0;
1326
1327 error3:
1328         /* Remove any methods for this mad agent */
1329         remove_methods_mad_agent(*method, agent_priv);
1330         /* Now, check to see if there are any methods in use */
1331         if (!check_method_table(*method)) {
1332                 /* If not, release management method table */
1333                 kfree(*method);
1334                 *method = NULL;
1335         }
1336         ret = -EINVAL;
1337         goto error1;
1338 error2:
1339         kfree(*class);
1340         *class = NULL;
1341 error1:
1342         return ret;
1343 }
1344
1345 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1346                            struct ib_mad_agent_private *agent_priv)
1347 {
1348         struct ib_mad_port_private *port_priv;
1349         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1350         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1351         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1352         struct ib_mad_mgmt_method_table **method;
1353         int i, ret = -ENOMEM;
1354         u8 vclass;
1355
1356         /* "New" vendor (with OUI) class */
1357         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1358         port_priv = agent_priv->qp_info->port_priv;
1359         vendor_table = &port_priv->version[
1360                                 mad_reg_req->mgmt_class_version].vendor;
1361         if (!*vendor_table) {
1362                 /* Allocate mgmt vendor class table for "new" class version */
1363                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1364                 if (!vendor) {
1365                         printk(KERN_ERR PFX "No memory for "
1366                                "ib_mad_mgmt_vendor_class_table\n");
1367                         goto error1;
1368                 }
1369
1370                 *vendor_table = vendor;
1371         }
1372         if (!(*vendor_table)->vendor_class[vclass]) {
1373                 /* Allocate table for this management vendor class */
1374                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1375                 if (!vendor_class) {
1376                         printk(KERN_ERR PFX "No memory for "
1377                                "ib_mad_mgmt_vendor_class\n");
1378                         goto error2;
1379                 }
1380
1381                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1382         }
1383         for (i = 0; i < MAX_MGMT_OUI; i++) {
1384                 /* Is there matching OUI for this vendor class ? */
1385                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1386                             mad_reg_req->oui, 3)) {
1387                         method = &(*vendor_table)->vendor_class[
1388                                                 vclass]->method_table[i];
1389                         BUG_ON(!*method);
1390                         goto check_in_use;
1391                 }
1392         }
1393         for (i = 0; i < MAX_MGMT_OUI; i++) {
1394                 /* OUI slot available ? */
1395                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1396                                 vclass]->oui[i])) {
1397                         method = &(*vendor_table)->vendor_class[
1398                                 vclass]->method_table[i];
1399                         BUG_ON(*method);
1400                         /* Allocate method table for this OUI */
1401                         if ((ret = allocate_method_table(method)))
1402                                 goto error3;
1403                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1404                                mad_reg_req->oui, 3);
1405                         goto check_in_use;
1406                 }
1407         }
1408         printk(KERN_ERR PFX "All OUI slots in use\n");
1409         goto error3;
1410
1411 check_in_use:
1412         /* Now, make sure methods are not already in use */
1413         if (method_in_use(method, mad_reg_req))
1414                 goto error4;
1415
1416         /* Finally, add in methods being registered */
1417         for (i = find_first_bit(mad_reg_req->method_mask,
1418                                 IB_MGMT_MAX_METHODS);
1419              i < IB_MGMT_MAX_METHODS;
1420              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1421                                1+i)) {
1422                 (*method)->agent[i] = agent_priv;
1423         }
1424         return 0;
1425
1426 error4:
1427         /* Remove any methods for this mad agent */
1428         remove_methods_mad_agent(*method, agent_priv);
1429         /* Now, check to see if there are any methods in use */
1430         if (!check_method_table(*method)) {
1431                 /* If not, release management method table */
1432                 kfree(*method);
1433                 *method = NULL;
1434         }
1435         ret = -EINVAL;
1436 error3:
1437         if (vendor_class) {
1438                 (*vendor_table)->vendor_class[vclass] = NULL;
1439                 kfree(vendor_class);
1440         }
1441 error2:
1442         if (vendor) {
1443                 *vendor_table = NULL;
1444                 kfree(vendor);
1445         }
1446 error1:
1447         return ret;
1448 }
1449
1450 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1451 {
1452         struct ib_mad_port_private *port_priv;
1453         struct ib_mad_mgmt_class_table *class;
1454         struct ib_mad_mgmt_method_table *method;
1455         struct ib_mad_mgmt_vendor_class_table *vendor;
1456         struct ib_mad_mgmt_vendor_class *vendor_class;
1457         int index;
1458         u8 mgmt_class;
1459
1460         /*
1461          * Was MAD registration request supplied
1462          * with original registration ?
1463          */
1464         if (!agent_priv->reg_req) {
1465                 goto out;
1466         }
1467
1468         port_priv = agent_priv->qp_info->port_priv;
1469         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1470         class = port_priv->version[
1471                         agent_priv->reg_req->mgmt_class_version].class;
1472         if (!class)
1473                 goto vendor_check;
1474
1475         method = class->method_table[mgmt_class];
1476         if (method) {
1477                 /* Remove any methods for this mad agent */
1478                 remove_methods_mad_agent(method, agent_priv);
1479                 /* Now, check to see if there are any methods still in use */
1480                 if (!check_method_table(method)) {
1481                         /* If not, release management method table */
1482                          kfree(method);
1483                          class->method_table[mgmt_class] = NULL;
1484                          /* Any management classes left ? */
1485                         if (!check_class_table(class)) {
1486                                 /* If not, release management class table */
1487                                 kfree(class);
1488                                 port_priv->version[
1489                                         agent_priv->reg_req->
1490                                         mgmt_class_version].class = NULL;
1491                         }
1492                 }
1493         }
1494
1495 vendor_check:
1496         if (!is_vendor_class(mgmt_class))
1497                 goto out;
1498
1499         /* normalize mgmt_class to vendor range 2 */
1500         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1501         vendor = port_priv->version[
1502                         agent_priv->reg_req->mgmt_class_version].vendor;
1503
1504         if (!vendor)
1505                 goto out;
1506
1507         vendor_class = vendor->vendor_class[mgmt_class];
1508         if (vendor_class) {
1509                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1510                 if (index < 0)
1511                         goto out;
1512                 method = vendor_class->method_table[index];
1513                 if (method) {
1514                         /* Remove any methods for this mad agent */
1515                         remove_methods_mad_agent(method, agent_priv);
1516                         /*
1517                          * Now, check to see if there are
1518                          * any methods still in use
1519                          */
1520                         if (!check_method_table(method)) {
1521                                 /* If not, release management method table */
1522                                 kfree(method);
1523                                 vendor_class->method_table[index] = NULL;
1524                                 memset(vendor_class->oui[index], 0, 3);
1525                                 /* Any OUIs left ? */
1526                                 if (!check_vendor_class(vendor_class)) {
1527                                         /* If not, release vendor class table */
1528                                         kfree(vendor_class);
1529                                         vendor->vendor_class[mgmt_class] = NULL;
1530                                         /* Any other vendor classes left ? */
1531                                         if (!check_vendor_table(vendor)) {
1532                                                 kfree(vendor);
1533                                                 port_priv->version[
1534                                                         agent_priv->reg_req->
1535                                                         mgmt_class_version].
1536                                                         vendor = NULL;
1537                                         }
1538                                 }
1539                         }
1540                 }
1541         }
1542
1543 out:
1544         return;
1545 }
1546
1547 static struct ib_mad_agent_private *
1548 find_mad_agent(struct ib_mad_port_private *port_priv,
1549                struct ib_mad *mad)
1550 {
1551         struct ib_mad_agent_private *mad_agent = NULL;
1552         unsigned long flags;
1553
1554         spin_lock_irqsave(&port_priv->reg_lock, flags);
1555         if (ib_response_mad(mad)) {
1556                 u32 hi_tid;
1557                 struct ib_mad_agent_private *entry;
1558
1559                 /*
1560                  * Routing is based on high 32 bits of transaction ID
1561                  * of MAD.
1562                  */
1563                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1564                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1565                         if (entry->agent.hi_tid == hi_tid) {
1566                                 mad_agent = entry;
1567                                 break;
1568                         }
1569                 }
1570         } else {
1571                 struct ib_mad_mgmt_class_table *class;
1572                 struct ib_mad_mgmt_method_table *method;
1573                 struct ib_mad_mgmt_vendor_class_table *vendor;
1574                 struct ib_mad_mgmt_vendor_class *vendor_class;
1575                 struct ib_vendor_mad *vendor_mad;
1576                 int index;
1577
1578                 /*
1579                  * Routing is based on version, class, and method
1580                  * For "newer" vendor MADs, also based on OUI
1581                  */
1582                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1583                         goto out;
1584                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1585                         class = port_priv->version[
1586                                         mad->mad_hdr.class_version].class;
1587                         if (!class)
1588                                 goto out;
1589                         method = class->method_table[convert_mgmt_class(
1590                                                         mad->mad_hdr.mgmt_class)];
1591                         if (method)
1592                                 mad_agent = method->agent[mad->mad_hdr.method &
1593                                                           ~IB_MGMT_METHOD_RESP];
1594                 } else {
1595                         vendor = port_priv->version[
1596                                         mad->mad_hdr.class_version].vendor;
1597                         if (!vendor)
1598                                 goto out;
1599                         vendor_class = vendor->vendor_class[vendor_class_index(
1600                                                 mad->mad_hdr.mgmt_class)];
1601                         if (!vendor_class)
1602                                 goto out;
1603                         /* Find matching OUI */
1604                         vendor_mad = (struct ib_vendor_mad *)mad;
1605                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1606                         if (index == -1)
1607                                 goto out;
1608                         method = vendor_class->method_table[index];
1609                         if (method) {
1610                                 mad_agent = method->agent[mad->mad_hdr.method &
1611                                                           ~IB_MGMT_METHOD_RESP];
1612                         }
1613                 }
1614         }
1615
1616         if (mad_agent) {
1617                 if (mad_agent->agent.recv_handler)
1618                         atomic_inc(&mad_agent->refcount);
1619                 else {
1620                         printk(KERN_NOTICE PFX "No receive handler for client "
1621                                "%p on port %d\n",
1622                                &mad_agent->agent, port_priv->port_num);
1623                         mad_agent = NULL;
1624                 }
1625         }
1626 out:
1627         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1628
1629         return mad_agent;
1630 }
1631
1632 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1633 {
1634         int valid = 0;
1635
1636         /* Make sure MAD base version is understood */
1637         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1638                 printk(KERN_ERR PFX "MAD received with unsupported base "
1639                        "version %d\n", mad->mad_hdr.base_version);
1640                 goto out;
1641         }
1642
1643         /* Filter SMI packets sent to other than QP0 */
1644         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1645             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1646                 if (qp_num == 0)
1647                         valid = 1;
1648         } else {
1649                 /* Filter GSI packets sent to QP0 */
1650                 if (qp_num != 0)
1651                         valid = 1;
1652         }
1653
1654 out:
1655         return valid;
1656 }
1657
1658 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1659                        struct ib_mad_hdr *mad_hdr)
1660 {
1661         struct ib_rmpp_mad *rmpp_mad;
1662
1663         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1664         return !mad_agent_priv->agent.rmpp_version ||
1665                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1666                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1667                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1668 }
1669
1670 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1671                                      struct ib_mad_recv_wc *rwc)
1672 {
1673         return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1674                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1675 }
1676
1677 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1678                                    struct ib_mad_send_wr_private *wr,
1679                                    struct ib_mad_recv_wc *rwc )
1680 {
1681         struct ib_ah_attr attr;
1682         u8 send_resp, rcv_resp;
1683         union ib_gid sgid;
1684         struct ib_device *device = mad_agent_priv->agent.device;
1685         u8 port_num = mad_agent_priv->agent.port_num;
1686         u8 lmc;
1687
1688         send_resp = ((struct ib_mad *)(wr->send_buf.mad))->
1689                      mad_hdr.method & IB_MGMT_METHOD_RESP;
1690         rcv_resp = rwc->recv_buf.mad->mad_hdr.method & IB_MGMT_METHOD_RESP;
1691
1692         if (send_resp == rcv_resp)
1693                 /* both requests, or both responses. GIDs different */
1694                 return 0;
1695
1696         if (ib_query_ah(wr->send_buf.ah, &attr))
1697                 /* Assume not equal, to avoid false positives. */
1698                 return 0;
1699
1700         if (!!(attr.ah_flags & IB_AH_GRH) !=
1701             !!(rwc->wc->wc_flags & IB_WC_GRH))
1702                 /* one has GID, other does not.  Assume different */
1703                 return 0;
1704
1705         if (!send_resp && rcv_resp) {
1706                 /* is request/response. */
1707                 if (!(attr.ah_flags & IB_AH_GRH)) {
1708                         if (ib_get_cached_lmc(device, port_num, &lmc))
1709                                 return 0;
1710                         return (!lmc || !((attr.src_path_bits ^
1711                                            rwc->wc->dlid_path_bits) &
1712                                           ((1 << lmc) - 1)));
1713                 } else {
1714                         if (ib_get_cached_gid(device, port_num,
1715                                               attr.grh.sgid_index, &sgid))
1716                                 return 0;
1717                         return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1718                                        16);
1719                 }
1720         }
1721
1722         if (!(attr.ah_flags & IB_AH_GRH))
1723                 return attr.dlid == rwc->wc->slid;
1724         else
1725                 return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1726                                16);
1727 }
1728
1729 static inline int is_direct(u8 class)
1730 {
1731         return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1732 }
1733
1734 struct ib_mad_send_wr_private*
1735 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1736                  struct ib_mad_recv_wc *wc)
1737 {
1738         struct ib_mad_send_wr_private *wr;
1739         struct ib_mad *mad;
1740
1741         mad = (struct ib_mad *)wc->recv_buf.mad;
1742
1743         list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1744                 if ((wr->tid == mad->mad_hdr.tid) &&
1745                     rcv_has_same_class(wr, wc) &&
1746                     /*
1747                      * Don't check GID for direct routed MADs.
1748                      * These might have permissive LIDs.
1749                      */
1750                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1751                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1752                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1753         }
1754
1755         /*
1756          * It's possible to receive the response before we've
1757          * been notified that the send has completed
1758          */
1759         list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1760                 if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1761                     wr->tid == mad->mad_hdr.tid &&
1762                     wr->timeout &&
1763                     rcv_has_same_class(wr, wc) &&
1764                     /*
1765                      * Don't check GID for direct routed MADs.
1766                      * These might have permissive LIDs.
1767                      */
1768                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1769                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1770                         /* Verify request has not been canceled */
1771                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1772         }
1773         return NULL;
1774 }
1775
1776 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1777 {
1778         mad_send_wr->timeout = 0;
1779         if (mad_send_wr->refcount == 1)
1780                 list_move_tail(&mad_send_wr->agent_list,
1781                               &mad_send_wr->mad_agent_priv->done_list);
1782 }
1783
1784 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1785                                  struct ib_mad_recv_wc *mad_recv_wc)
1786 {
1787         struct ib_mad_send_wr_private *mad_send_wr;
1788         struct ib_mad_send_wc mad_send_wc;
1789         unsigned long flags;
1790
1791         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1792         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1793         if (mad_agent_priv->agent.rmpp_version) {
1794                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1795                                                       mad_recv_wc);
1796                 if (!mad_recv_wc) {
1797                         deref_mad_agent(mad_agent_priv);
1798                         return;
1799                 }
1800         }
1801
1802         /* Complete corresponding request */
1803         if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1804                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1805                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1806                 if (!mad_send_wr) {
1807                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1808                         ib_free_recv_mad(mad_recv_wc);
1809                         deref_mad_agent(mad_agent_priv);
1810                         return;
1811                 }
1812                 ib_mark_mad_done(mad_send_wr);
1813                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1814
1815                 /* Defined behavior is to complete response before request */
1816                 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1817                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1818                                                    mad_recv_wc);
1819                 atomic_dec(&mad_agent_priv->refcount);
1820
1821                 mad_send_wc.status = IB_WC_SUCCESS;
1822                 mad_send_wc.vendor_err = 0;
1823                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1824                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1825         } else {
1826                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1827                                                    mad_recv_wc);
1828                 deref_mad_agent(mad_agent_priv);
1829         }
1830 }
1831
1832 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1833                                      struct ib_wc *wc)
1834 {
1835         struct ib_mad_qp_info *qp_info;
1836         struct ib_mad_private_header *mad_priv_hdr;
1837         struct ib_mad_private *recv, *response;
1838         struct ib_mad_list_head *mad_list;
1839         struct ib_mad_agent_private *mad_agent;
1840
1841         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1842         if (!response)
1843                 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1844                        "for response buffer\n");
1845
1846         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1847         qp_info = mad_list->mad_queue->qp_info;
1848         dequeue_mad(mad_list);
1849
1850         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1851                                     mad_list);
1852         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1853         ib_dma_unmap_single(port_priv->device,
1854                             recv->header.mapping,
1855                             sizeof(struct ib_mad_private) -
1856                               sizeof(struct ib_mad_private_header),
1857                             DMA_FROM_DEVICE);
1858
1859         /* Setup MAD receive work completion from "normal" work completion */
1860         recv->header.wc = *wc;
1861         recv->header.recv_wc.wc = &recv->header.wc;
1862         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1863         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1864         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1865
1866         if (atomic_read(&qp_info->snoop_count))
1867                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1868
1869         /* Validate MAD */
1870         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1871                 goto out;
1872
1873         if (recv->mad.mad.mad_hdr.mgmt_class ==
1874             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1875                 if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1876                                             port_priv->device->node_type,
1877                                             port_priv->port_num,
1878                                             port_priv->device->phys_port_cnt))
1879                         goto out;
1880                 if (!smi_check_forward_dr_smp(&recv->mad.smp))
1881                         goto local;
1882                 if (!smi_handle_dr_smp_send(&recv->mad.smp,
1883                                             port_priv->device->node_type,
1884                                             port_priv->port_num))
1885                         goto out;
1886                 if (!smi_check_local_smp(&recv->mad.smp, port_priv->device))
1887                         goto out;
1888         }
1889
1890 local:
1891         /* Give driver "right of first refusal" on incoming MAD */
1892         if (port_priv->device->process_mad) {
1893                 int ret;
1894
1895                 if (!response) {
1896                         printk(KERN_ERR PFX "No memory for response MAD\n");
1897                         /*
1898                          * Is it better to assume that
1899                          * it wouldn't be processed ?
1900                          */
1901                         goto out;
1902                 }
1903
1904                 ret = port_priv->device->process_mad(port_priv->device, 0,
1905                                                      port_priv->port_num,
1906                                                      wc, &recv->grh,
1907                                                      &recv->mad.mad,
1908                                                      &response->mad.mad);
1909                 if (ret & IB_MAD_RESULT_SUCCESS) {
1910                         if (ret & IB_MAD_RESULT_CONSUMED)
1911                                 goto out;
1912                         if (ret & IB_MAD_RESULT_REPLY) {
1913                                 agent_send_response(&response->mad.mad,
1914                                                     &recv->grh, wc,
1915                                                     port_priv->device,
1916                                                     port_priv->port_num,
1917                                                     qp_info->qp->qp_num);
1918                                 goto out;
1919                         }
1920                 }
1921         }
1922
1923         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1924         if (mad_agent) {
1925                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1926                 /*
1927                  * recv is freed up in error cases in ib_mad_complete_recv
1928                  * or via recv_handler in ib_mad_complete_recv()
1929                  */
1930                 recv = NULL;
1931         }
1932
1933 out:
1934         /* Post another receive request for this QP */
1935         if (response) {
1936                 ib_mad_post_receive_mads(qp_info, response);
1937                 if (recv)
1938                         kmem_cache_free(ib_mad_cache, recv);
1939         } else
1940                 ib_mad_post_receive_mads(qp_info, recv);
1941 }
1942
1943 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1944 {
1945         struct ib_mad_send_wr_private *mad_send_wr;
1946         unsigned long delay;
1947
1948         if (list_empty(&mad_agent_priv->wait_list)) {
1949                 cancel_delayed_work(&mad_agent_priv->timed_work);
1950         } else {
1951                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1952                                          struct ib_mad_send_wr_private,
1953                                          agent_list);
1954
1955                 if (time_after(mad_agent_priv->timeout,
1956                                mad_send_wr->timeout)) {
1957                         mad_agent_priv->timeout = mad_send_wr->timeout;
1958                         cancel_delayed_work(&mad_agent_priv->timed_work);
1959                         delay = mad_send_wr->timeout - jiffies;
1960                         if ((long)delay <= 0)
1961                                 delay = 1;
1962                         queue_delayed_work(mad_agent_priv->qp_info->
1963                                            port_priv->wq,
1964                                            &mad_agent_priv->timed_work, delay);
1965                 }
1966         }
1967 }
1968
1969 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1970 {
1971         struct ib_mad_agent_private *mad_agent_priv;
1972         struct ib_mad_send_wr_private *temp_mad_send_wr;
1973         struct list_head *list_item;
1974         unsigned long delay;
1975
1976         mad_agent_priv = mad_send_wr->mad_agent_priv;
1977         list_del(&mad_send_wr->agent_list);
1978
1979         delay = mad_send_wr->timeout;
1980         mad_send_wr->timeout += jiffies;
1981
1982         if (delay) {
1983                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1984                         temp_mad_send_wr = list_entry(list_item,
1985                                                 struct ib_mad_send_wr_private,
1986                                                 agent_list);
1987                         if (time_after(mad_send_wr->timeout,
1988                                        temp_mad_send_wr->timeout))
1989                                 break;
1990                 }
1991         }
1992         else
1993                 list_item = &mad_agent_priv->wait_list;
1994         list_add(&mad_send_wr->agent_list, list_item);
1995
1996         /* Reschedule a work item if we have a shorter timeout */
1997         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1998                 cancel_delayed_work(&mad_agent_priv->timed_work);
1999                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2000                                    &mad_agent_priv->timed_work, delay);
2001         }
2002 }
2003
2004 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2005                           int timeout_ms)
2006 {
2007         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2008         wait_for_response(mad_send_wr);
2009 }
2010
2011 /*
2012  * Process a send work completion
2013  */
2014 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2015                              struct ib_mad_send_wc *mad_send_wc)
2016 {
2017         struct ib_mad_agent_private     *mad_agent_priv;
2018         unsigned long                   flags;
2019         int                             ret;
2020
2021         mad_agent_priv = mad_send_wr->mad_agent_priv;
2022         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2023         if (mad_agent_priv->agent.rmpp_version) {
2024                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2025                 if (ret == IB_RMPP_RESULT_CONSUMED)
2026                         goto done;
2027         } else
2028                 ret = IB_RMPP_RESULT_UNHANDLED;
2029
2030         if (mad_send_wc->status != IB_WC_SUCCESS &&
2031             mad_send_wr->status == IB_WC_SUCCESS) {
2032                 mad_send_wr->status = mad_send_wc->status;
2033                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2034         }
2035
2036         if (--mad_send_wr->refcount > 0) {
2037                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2038                     mad_send_wr->status == IB_WC_SUCCESS) {
2039                         wait_for_response(mad_send_wr);
2040                 }
2041                 goto done;
2042         }
2043
2044         /* Remove send from MAD agent and notify client of completion */
2045         list_del(&mad_send_wr->agent_list);
2046         adjust_timeout(mad_agent_priv);
2047         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2048
2049         if (mad_send_wr->status != IB_WC_SUCCESS )
2050                 mad_send_wc->status = mad_send_wr->status;
2051         if (ret == IB_RMPP_RESULT_INTERNAL)
2052                 ib_rmpp_send_handler(mad_send_wc);
2053         else
2054                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2055                                                    mad_send_wc);
2056
2057         /* Release reference on agent taken when sending */
2058         deref_mad_agent(mad_agent_priv);
2059         return;
2060 done:
2061         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2062 }
2063
2064 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2065                                      struct ib_wc *wc)
2066 {
2067         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2068         struct ib_mad_list_head         *mad_list;
2069         struct ib_mad_qp_info           *qp_info;
2070         struct ib_mad_queue             *send_queue;
2071         struct ib_send_wr               *bad_send_wr;
2072         struct ib_mad_send_wc           mad_send_wc;
2073         unsigned long flags;
2074         int ret;
2075
2076         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2077         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2078                                    mad_list);
2079         send_queue = mad_list->mad_queue;
2080         qp_info = send_queue->qp_info;
2081
2082 retry:
2083         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2084                             mad_send_wr->header_mapping,
2085                             mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2086         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2087                             mad_send_wr->payload_mapping,
2088                             mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2089         queued_send_wr = NULL;
2090         spin_lock_irqsave(&send_queue->lock, flags);
2091         list_del(&mad_list->list);
2092
2093         /* Move queued send to the send queue */
2094         if (send_queue->count-- > send_queue->max_active) {
2095                 mad_list = container_of(qp_info->overflow_list.next,
2096                                         struct ib_mad_list_head, list);
2097                 queued_send_wr = container_of(mad_list,
2098                                         struct ib_mad_send_wr_private,
2099                                         mad_list);
2100                 list_move_tail(&mad_list->list, &send_queue->list);
2101         }
2102         spin_unlock_irqrestore(&send_queue->lock, flags);
2103
2104         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2105         mad_send_wc.status = wc->status;
2106         mad_send_wc.vendor_err = wc->vendor_err;
2107         if (atomic_read(&qp_info->snoop_count))
2108                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2109                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2110         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2111
2112         if (queued_send_wr) {
2113                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2114                                    &bad_send_wr);
2115                 if (ret) {
2116                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2117                         mad_send_wr = queued_send_wr;
2118                         wc->status = IB_WC_LOC_QP_OP_ERR;
2119                         goto retry;
2120                 }
2121         }
2122 }
2123
2124 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2125 {
2126         struct ib_mad_send_wr_private *mad_send_wr;
2127         struct ib_mad_list_head *mad_list;
2128         unsigned long flags;
2129
2130         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2131         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2132                 mad_send_wr = container_of(mad_list,
2133                                            struct ib_mad_send_wr_private,
2134                                            mad_list);
2135                 mad_send_wr->retry = 1;
2136         }
2137         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2138 }
2139
2140 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2141                               struct ib_wc *wc)
2142 {
2143         struct ib_mad_list_head *mad_list;
2144         struct ib_mad_qp_info *qp_info;
2145         struct ib_mad_send_wr_private *mad_send_wr;
2146         int ret;
2147
2148         /* Determine if failure was a send or receive */
2149         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2150         qp_info = mad_list->mad_queue->qp_info;
2151         if (mad_list->mad_queue == &qp_info->recv_queue)
2152                 /*
2153                  * Receive errors indicate that the QP has entered the error
2154                  * state - error handling/shutdown code will cleanup
2155                  */
2156                 return;
2157
2158         /*
2159          * Send errors will transition the QP to SQE - move
2160          * QP to RTS and repost flushed work requests
2161          */
2162         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2163                                    mad_list);
2164         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2165                 if (mad_send_wr->retry) {
2166                         /* Repost send */
2167                         struct ib_send_wr *bad_send_wr;
2168
2169                         mad_send_wr->retry = 0;
2170                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2171                                         &bad_send_wr);
2172                         if (ret)
2173                                 ib_mad_send_done_handler(port_priv, wc);
2174                 } else
2175                         ib_mad_send_done_handler(port_priv, wc);
2176         } else {
2177                 struct ib_qp_attr *attr;
2178
2179                 /* Transition QP to RTS and fail offending send */
2180                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2181                 if (attr) {
2182                         attr->qp_state = IB_QPS_RTS;
2183                         attr->cur_qp_state = IB_QPS_SQE;
2184                         ret = ib_modify_qp(qp_info->qp, attr,
2185                                            IB_QP_STATE | IB_QP_CUR_STATE);
2186                         kfree(attr);
2187                         if (ret)
2188                                 printk(KERN_ERR PFX "mad_error_handler - "
2189                                        "ib_modify_qp to RTS : %d\n", ret);
2190                         else
2191                                 mark_sends_for_retry(qp_info);
2192                 }
2193                 ib_mad_send_done_handler(port_priv, wc);
2194         }
2195 }
2196
2197 /*
2198  * IB MAD completion callback
2199  */
2200 static void ib_mad_completion_handler(struct work_struct *work)
2201 {
2202         struct ib_mad_port_private *port_priv;
2203         struct ib_wc wc;
2204
2205         port_priv = container_of(work, struct ib_mad_port_private, work);
2206         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2207
2208         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2209                 if (wc.status == IB_WC_SUCCESS) {
2210                         switch (wc.opcode) {
2211                         case IB_WC_SEND:
2212                                 ib_mad_send_done_handler(port_priv, &wc);
2213                                 break;
2214                         case IB_WC_RECV:
2215                                 ib_mad_recv_done_handler(port_priv, &wc);
2216                                 break;
2217                         default:
2218                                 BUG_ON(1);
2219                                 break;
2220                         }
2221                 } else
2222                         mad_error_handler(port_priv, &wc);
2223         }
2224 }
2225
2226 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2227 {
2228         unsigned long flags;
2229         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2230         struct ib_mad_send_wc mad_send_wc;
2231         struct list_head cancel_list;
2232
2233         INIT_LIST_HEAD(&cancel_list);
2234
2235         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2236         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2237                                  &mad_agent_priv->send_list, agent_list) {
2238                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2239                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2240                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2241                 }
2242         }
2243
2244         /* Empty wait list to prevent receives from finding a request */
2245         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2246         /* Empty local completion list as well */
2247         list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2248         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2249
2250         /* Report all cancelled requests */
2251         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2252         mad_send_wc.vendor_err = 0;
2253
2254         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2255                                  &cancel_list, agent_list) {
2256                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2257                 list_del(&mad_send_wr->agent_list);
2258                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2259                                                    &mad_send_wc);
2260                 atomic_dec(&mad_agent_priv->refcount);
2261         }
2262 }
2263
2264 static struct ib_mad_send_wr_private*
2265 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2266              struct ib_mad_send_buf *send_buf)
2267 {
2268         struct ib_mad_send_wr_private *mad_send_wr;
2269
2270         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2271                             agent_list) {
2272                 if (&mad_send_wr->send_buf == send_buf)
2273                         return mad_send_wr;
2274         }
2275
2276         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2277                             agent_list) {
2278                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2279                     &mad_send_wr->send_buf == send_buf)
2280                         return mad_send_wr;
2281         }
2282         return NULL;
2283 }
2284
2285 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2286                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2287 {
2288         struct ib_mad_agent_private *mad_agent_priv;
2289         struct ib_mad_send_wr_private *mad_send_wr;
2290         unsigned long flags;
2291         int active;
2292
2293         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2294                                       agent);
2295         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2296         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2297         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2298                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2299                 return -EINVAL;
2300         }
2301
2302         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2303         if (!timeout_ms) {
2304                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2305                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2306         }
2307
2308         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2309         if (active)
2310                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2311         else
2312                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2313
2314         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2315         return 0;
2316 }
2317 EXPORT_SYMBOL(ib_modify_mad);
2318
2319 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2320                    struct ib_mad_send_buf *send_buf)
2321 {
2322         ib_modify_mad(mad_agent, send_buf, 0);
2323 }
2324 EXPORT_SYMBOL(ib_cancel_mad);
2325
2326 static void local_completions(struct work_struct *work)
2327 {
2328         struct ib_mad_agent_private *mad_agent_priv;
2329         struct ib_mad_local_private *local;
2330         struct ib_mad_agent_private *recv_mad_agent;
2331         unsigned long flags;
2332         int recv = 0;
2333         struct ib_wc wc;
2334         struct ib_mad_send_wc mad_send_wc;
2335
2336         mad_agent_priv =
2337                 container_of(work, struct ib_mad_agent_private, local_work);
2338
2339         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2340         while (!list_empty(&mad_agent_priv->local_list)) {
2341                 local = list_entry(mad_agent_priv->local_list.next,
2342                                    struct ib_mad_local_private,
2343                                    completion_list);
2344                 list_del(&local->completion_list);
2345                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2346                 if (local->mad_priv) {
2347                         recv_mad_agent = local->recv_mad_agent;
2348                         if (!recv_mad_agent) {
2349                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2350                                 goto local_send_completion;
2351                         }
2352
2353                         recv = 1;
2354                         /*
2355                          * Defined behavior is to complete response
2356                          * before request
2357                          */
2358                         build_smp_wc((unsigned long) local->mad_send_wr,
2359                                      be16_to_cpu(IB_LID_PERMISSIVE),
2360                                      0, recv_mad_agent->agent.port_num, &wc);
2361
2362                         local->mad_priv->header.recv_wc.wc = &wc;
2363                         local->mad_priv->header.recv_wc.mad_len =
2364                                                 sizeof(struct ib_mad);
2365                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2366                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2367                                  &local->mad_priv->header.recv_wc.rmpp_list);
2368                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2369                         local->mad_priv->header.recv_wc.recv_buf.mad =
2370                                                 &local->mad_priv->mad.mad;
2371                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2372                                 snoop_recv(recv_mad_agent->qp_info,
2373                                           &local->mad_priv->header.recv_wc,
2374                                            IB_MAD_SNOOP_RECVS);
2375                         recv_mad_agent->agent.recv_handler(
2376                                                 &recv_mad_agent->agent,
2377                                                 &local->mad_priv->header.recv_wc);
2378                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2379                         atomic_dec(&recv_mad_agent->refcount);
2380                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2381                 }
2382
2383 local_send_completion:
2384                 /* Complete send */
2385                 mad_send_wc.status = IB_WC_SUCCESS;
2386                 mad_send_wc.vendor_err = 0;
2387                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2388                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2389                         snoop_send(mad_agent_priv->qp_info,
2390                                    &local->mad_send_wr->send_buf,
2391                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2392                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2393                                                    &mad_send_wc);
2394
2395                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2396                 atomic_dec(&mad_agent_priv->refcount);
2397                 if (!recv)
2398                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2399                 kfree(local);
2400         }
2401         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2402 }
2403
2404 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2405 {
2406         int ret;
2407
2408         if (!mad_send_wr->retries--)
2409                 return -ETIMEDOUT;
2410
2411         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2412
2413         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2414                 ret = ib_retry_rmpp(mad_send_wr);
2415                 switch (ret) {
2416                 case IB_RMPP_RESULT_UNHANDLED:
2417                         ret = ib_send_mad(mad_send_wr);
2418                         break;
2419                 case IB_RMPP_RESULT_CONSUMED:
2420                         ret = 0;
2421                         break;
2422                 default:
2423                         ret = -ECOMM;
2424                         break;
2425                 }
2426         } else
2427                 ret = ib_send_mad(mad_send_wr);
2428
2429         if (!ret) {
2430                 mad_send_wr->refcount++;
2431                 list_add_tail(&mad_send_wr->agent_list,
2432                               &mad_send_wr->mad_agent_priv->send_list);
2433         }
2434         return ret;
2435 }
2436
2437 static void timeout_sends(struct work_struct *work)
2438 {
2439         struct ib_mad_agent_private *mad_agent_priv;
2440         struct ib_mad_send_wr_private *mad_send_wr;
2441         struct ib_mad_send_wc mad_send_wc;
2442         unsigned long flags, delay;
2443
2444         mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2445                                       timed_work.work);
2446         mad_send_wc.vendor_err = 0;
2447
2448         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2449         while (!list_empty(&mad_agent_priv->wait_list)) {
2450                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2451                                          struct ib_mad_send_wr_private,
2452                                          agent_list);
2453
2454                 if (time_after(mad_send_wr->timeout, jiffies)) {
2455                         delay = mad_send_wr->timeout - jiffies;
2456                         if ((long)delay <= 0)
2457                                 delay = 1;
2458                         queue_delayed_work(mad_agent_priv->qp_info->
2459                                            port_priv->wq,
2460                                            &mad_agent_priv->timed_work, delay);
2461                         break;
2462                 }
2463
2464                 list_del(&mad_send_wr->agent_list);
2465                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2466                     !retry_send(mad_send_wr))
2467                         continue;
2468
2469                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2470
2471                 if (mad_send_wr->status == IB_WC_SUCCESS)
2472                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2473                 else
2474                         mad_send_wc.status = mad_send_wr->status;
2475                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2476                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2477                                                    &mad_send_wc);
2478
2479                 atomic_dec(&mad_agent_priv->refcount);
2480                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2481         }
2482         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2483 }
2484
2485 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2486 {
2487         struct ib_mad_port_private *port_priv = cq->cq_context;
2488         unsigned long flags;
2489
2490         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2491         if (!list_empty(&port_priv->port_list))
2492                 queue_work(port_priv->wq, &port_priv->work);
2493         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2494 }
2495
2496 /*
2497  * Allocate receive MADs and post receive WRs for them
2498  */
2499 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2500                                     struct ib_mad_private *mad)
2501 {
2502         unsigned long flags;
2503         int post, ret;
2504         struct ib_mad_private *mad_priv;
2505         struct ib_sge sg_list;
2506         struct ib_recv_wr recv_wr, *bad_recv_wr;
2507         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2508
2509         /* Initialize common scatter list fields */
2510         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2511         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2512
2513         /* Initialize common receive WR fields */
2514         recv_wr.next = NULL;
2515         recv_wr.sg_list = &sg_list;
2516         recv_wr.num_sge = 1;
2517
2518         do {
2519                 /* Allocate and map receive buffer */
2520                 if (mad) {
2521                         mad_priv = mad;
2522                         mad = NULL;
2523                 } else {
2524                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2525                         if (!mad_priv) {
2526                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2527                                 ret = -ENOMEM;
2528                                 break;
2529                         }
2530                 }
2531                 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2532                                                  &mad_priv->grh,
2533                                                  sizeof *mad_priv -
2534                                                    sizeof mad_priv->header,
2535                                                  DMA_FROM_DEVICE);
2536                 mad_priv->header.mapping = sg_list.addr;
2537                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2538                 mad_priv->header.mad_list.mad_queue = recv_queue;
2539
2540                 /* Post receive WR */
2541                 spin_lock_irqsave(&recv_queue->lock, flags);
2542                 post = (++recv_queue->count < recv_queue->max_active);
2543                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2544                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2545                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2546                 if (ret) {
2547                         spin_lock_irqsave(&recv_queue->lock, flags);
2548                         list_del(&mad_priv->header.mad_list.list);
2549                         recv_queue->count--;
2550                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2551                         ib_dma_unmap_single(qp_info->port_priv->device,
2552                                             mad_priv->header.mapping,
2553                                             sizeof *mad_priv -
2554                                               sizeof mad_priv->header,
2555                                             DMA_FROM_DEVICE);
2556                         kmem_cache_free(ib_mad_cache, mad_priv);
2557                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2558                         break;
2559                 }
2560         } while (post);
2561
2562         return ret;
2563 }
2564
2565 /*
2566  * Return all the posted receive MADs
2567  */
2568 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2569 {
2570         struct ib_mad_private_header *mad_priv_hdr;
2571         struct ib_mad_private *recv;
2572         struct ib_mad_list_head *mad_list;
2573
2574         while (!list_empty(&qp_info->recv_queue.list)) {
2575
2576                 mad_list = list_entry(qp_info->recv_queue.list.next,
2577                                       struct ib_mad_list_head, list);
2578                 mad_priv_hdr = container_of(mad_list,
2579                                             struct ib_mad_private_header,
2580                                             mad_list);
2581                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2582                                     header);
2583
2584                 /* Remove from posted receive MAD list */
2585                 list_del(&mad_list->list);
2586
2587                 ib_dma_unmap_single(qp_info->port_priv->device,
2588                                     recv->header.mapping,
2589                                     sizeof(struct ib_mad_private) -
2590                                       sizeof(struct ib_mad_private_header),
2591                                     DMA_FROM_DEVICE);
2592                 kmem_cache_free(ib_mad_cache, recv);
2593         }
2594
2595         qp_info->recv_queue.count = 0;
2596 }
2597
2598 /*
2599  * Start the port
2600  */
2601 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2602 {
2603         int ret, i;
2604         struct ib_qp_attr *attr;
2605         struct ib_qp *qp;
2606
2607         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2608         if (!attr) {
2609                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2610                 return -ENOMEM;
2611         }
2612
2613         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2614                 qp = port_priv->qp_info[i].qp;
2615                 /*
2616                  * PKey index for QP1 is irrelevant but
2617                  * one is needed for the Reset to Init transition
2618                  */
2619                 attr->qp_state = IB_QPS_INIT;
2620                 attr->pkey_index = 0;
2621                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2622                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2623                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2624                 if (ret) {
2625                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2626                                "INIT: %d\n", i, ret);
2627                         goto out;
2628                 }
2629
2630                 attr->qp_state = IB_QPS_RTR;
2631                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2632                 if (ret) {
2633                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2634                                "RTR: %d\n", i, ret);
2635                         goto out;
2636                 }
2637
2638                 attr->qp_state = IB_QPS_RTS;
2639                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2640                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2641                 if (ret) {
2642                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2643                                "RTS: %d\n", i, ret);
2644                         goto out;
2645                 }
2646         }
2647
2648         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2649         if (ret) {
2650                 printk(KERN_ERR PFX "Failed to request completion "
2651                        "notification: %d\n", ret);
2652                 goto out;
2653         }
2654
2655         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2656                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2657                 if (ret) {
2658                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2659                         goto out;
2660                 }
2661         }
2662 out:
2663         kfree(attr);
2664         return ret;
2665 }
2666
2667 static void qp_event_handler(struct ib_event *event, void *qp_context)
2668 {
2669         struct ib_mad_qp_info   *qp_info = qp_context;
2670
2671         /* It's worse than that! He's dead, Jim! */
2672         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2673                 event->event, qp_info->qp->qp_num);
2674 }
2675
2676 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2677                            struct ib_mad_queue *mad_queue)
2678 {
2679         mad_queue->qp_info = qp_info;
2680         mad_queue->count = 0;
2681         spin_lock_init(&mad_queue->lock);
2682         INIT_LIST_HEAD(&mad_queue->list);
2683 }
2684
2685 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2686                         struct ib_mad_qp_info *qp_info)
2687 {
2688         qp_info->port_priv = port_priv;
2689         init_mad_queue(qp_info, &qp_info->send_queue);
2690         init_mad_queue(qp_info, &qp_info->recv_queue);
2691         INIT_LIST_HEAD(&qp_info->overflow_list);
2692         spin_lock_init(&qp_info->snoop_lock);
2693         qp_info->snoop_table = NULL;
2694         qp_info->snoop_table_size = 0;
2695         atomic_set(&qp_info->snoop_count, 0);
2696 }
2697
2698 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2699                          enum ib_qp_type qp_type)
2700 {
2701         struct ib_qp_init_attr  qp_init_attr;
2702         int ret;
2703
2704         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2705         qp_init_attr.send_cq = qp_info->port_priv->cq;
2706         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2707         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2708         qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2709         qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2710         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2711         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2712         qp_init_attr.qp_type = qp_type;
2713         qp_init_attr.port_num = qp_info->port_priv->port_num;
2714         qp_init_attr.qp_context = qp_info;
2715         qp_init_attr.event_handler = qp_event_handler;
2716         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2717         if (IS_ERR(qp_info->qp)) {
2718                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2719                        get_spl_qp_index(qp_type));
2720                 ret = PTR_ERR(qp_info->qp);
2721                 goto error;
2722         }
2723         /* Use minimum queue sizes unless the CQ is resized */
2724         qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2725         qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2726         return 0;
2727
2728 error:
2729         return ret;
2730 }
2731
2732 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2733 {
2734         ib_destroy_qp(qp_info->qp);
2735         kfree(qp_info->snoop_table);
2736 }
2737
2738 /*
2739  * Open the port
2740  * Create the QP, PD, MR, and CQ if needed
2741  */
2742 static int ib_mad_port_open(struct ib_device *device,
2743                             int port_num)
2744 {
2745         int ret, cq_size;
2746         struct ib_mad_port_private *port_priv;
2747         unsigned long flags;
2748         char name[sizeof "ib_mad123"];
2749
2750         /* Create new device info */
2751         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2752         if (!port_priv) {
2753                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2754                 return -ENOMEM;
2755         }
2756
2757         port_priv->device = device;
2758         port_priv->port_num = port_num;
2759         spin_lock_init(&port_priv->reg_lock);
2760         INIT_LIST_HEAD(&port_priv->agent_list);
2761         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2762         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2763
2764         cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2765         port_priv->cq = ib_create_cq(port_priv->device,
2766                                      ib_mad_thread_completion_handler,
2767                                      NULL, port_priv, cq_size);
2768         if (IS_ERR(port_priv->cq)) {
2769                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2770                 ret = PTR_ERR(port_priv->cq);
2771                 goto error3;
2772         }
2773
2774         port_priv->pd = ib_alloc_pd(device);
2775         if (IS_ERR(port_priv->pd)) {
2776                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2777                 ret = PTR_ERR(port_priv->pd);
2778                 goto error4;
2779         }
2780
2781         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2782         if (IS_ERR(port_priv->mr)) {
2783                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2784                 ret = PTR_ERR(port_priv->mr);
2785                 goto error5;
2786         }
2787
2788         ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2789         if (ret)
2790                 goto error6;
2791         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2792         if (ret)
2793                 goto error7;
2794
2795         snprintf(name, sizeof name, "ib_mad%d", port_num);
2796         port_priv->wq = create_singlethread_workqueue(name);
2797         if (!port_priv->wq) {
2798                 ret = -ENOMEM;
2799                 goto error8;
2800         }
2801         INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2802
2803         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2804         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2805         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2806
2807         ret = ib_mad_port_start(port_priv);
2808         if (ret) {
2809                 printk(KERN_ERR PFX "Couldn't start port\n");
2810                 goto error9;
2811         }
2812
2813         return 0;
2814
2815 error9:
2816         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2817         list_del_init(&port_priv->port_list);
2818         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2819
2820         destroy_workqueue(port_priv->wq);
2821 error8:
2822         destroy_mad_qp(&port_priv->qp_info[1]);
2823 error7:
2824         destroy_mad_qp(&port_priv->qp_info[0]);
2825 error6:
2826         ib_dereg_mr(port_priv->mr);
2827 error5:
2828         ib_dealloc_pd(port_priv->pd);
2829 error4:
2830         ib_destroy_cq(port_priv->cq);
2831         cleanup_recv_queue(&port_priv->qp_info[1]);
2832         cleanup_recv_queue(&port_priv->qp_info[0]);
2833 error3:
2834         kfree(port_priv);
2835
2836         return ret;
2837 }
2838
2839 /*
2840  * Close the port
2841  * If there are no classes using the port, free the port
2842  * resources (CQ, MR, PD, QP) and remove the port's info structure
2843  */
2844 static int ib_mad_port_close(struct ib_device *device, int port_num)
2845 {
2846         struct ib_mad_port_private *port_priv;
2847         unsigned long flags;
2848
2849         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2850         port_priv = __ib_get_mad_port(device, port_num);
2851         if (port_priv == NULL) {
2852                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2853                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2854                 return -ENODEV;
2855         }
2856         list_del_init(&port_priv->port_list);
2857         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2858
2859         destroy_workqueue(port_priv->wq);
2860         destroy_mad_qp(&port_priv->qp_info[1]);
2861         destroy_mad_qp(&port_priv->qp_info[0]);
2862         ib_dereg_mr(port_priv->mr);
2863         ib_dealloc_pd(port_priv->pd);
2864         ib_destroy_cq(port_priv->cq);
2865         cleanup_recv_queue(&port_priv->qp_info[1]);
2866         cleanup_recv_queue(&port_priv->qp_info[0]);
2867         /* XXX: Handle deallocation of MAD registration tables */
2868
2869         kfree(port_priv);
2870
2871         return 0;
2872 }
2873
2874 static void ib_mad_init_device(struct ib_device *device)
2875 {
2876         int start, end, i;
2877
2878         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2879                 return;
2880
2881         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2882                 start = 0;
2883                 end   = 0;
2884         } else {
2885                 start = 1;
2886                 end   = device->phys_port_cnt;
2887         }
2888
2889         for (i = start; i <= end; i++) {
2890                 if (ib_mad_port_open(device, i)) {
2891                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2892                                device->name, i);
2893                         goto error;
2894                 }
2895                 if (ib_agent_port_open(device, i)) {
2896                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2897                                "for agents\n",
2898                                device->name, i);
2899                         goto error_agent;
2900                 }
2901         }
2902         return;
2903
2904 error_agent:
2905         if (ib_mad_port_close(device, i))
2906                 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2907                        device->name, i);
2908
2909 error:
2910         i--;
2911
2912         while (i >= start) {
2913                 if (ib_agent_port_close(device, i))
2914                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2915                                "for agents\n",
2916                                device->name, i);
2917                 if (ib_mad_port_close(device, i))
2918                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2919                                device->name, i);
2920                 i--;
2921         }
2922 }
2923
2924 static void ib_mad_remove_device(struct ib_device *device)
2925 {
2926         int i, num_ports, cur_port;
2927
2928         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2929                 num_ports = 1;
2930                 cur_port = 0;
2931         } else {
2932                 num_ports = device->phys_port_cnt;
2933                 cur_port = 1;
2934         }
2935         for (i = 0; i < num_ports; i++, cur_port++) {
2936                 if (ib_agent_port_close(device, cur_port))
2937                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2938                                "for agents\n",
2939                                device->name, cur_port);
2940                 if (ib_mad_port_close(device, cur_port))
2941                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2942                                device->name, cur_port);
2943         }
2944 }
2945
2946 static struct ib_client mad_client = {
2947         .name   = "mad",
2948         .add = ib_mad_init_device,
2949         .remove = ib_mad_remove_device
2950 };
2951
2952 static int __init ib_mad_init_module(void)
2953 {
2954         int ret;
2955
2956         spin_lock_init(&ib_mad_port_list_lock);
2957
2958         ib_mad_cache = kmem_cache_create("ib_mad",
2959                                          sizeof(struct ib_mad_private),
2960                                          0,
2961                                          SLAB_HWCACHE_ALIGN,
2962                                          NULL,
2963                                          NULL);
2964         if (!ib_mad_cache) {
2965                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2966                 ret = -ENOMEM;
2967                 goto error1;
2968         }
2969
2970         INIT_LIST_HEAD(&ib_mad_port_list);
2971
2972         if (ib_register_client(&mad_client)) {
2973                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2974                 ret = -EINVAL;
2975                 goto error2;
2976         }
2977
2978         return 0;
2979
2980 error2:
2981         kmem_cache_destroy(ib_mad_cache);
2982 error1:
2983         return ret;
2984 }
2985
2986 static void __exit ib_mad_cleanup_module(void)
2987 {
2988         ib_unregister_client(&mad_client);
2989         kmem_cache_destroy(ib_mad_cache);
2990 }
2991
2992 module_init(ib_mad_init_module);
2993 module_exit(ib_mad_cleanup_module);
2994