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

   1 /*
   2  * Copyright (c) 2004-2007 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  */
  35 #include <linux/dma-mapping.h>
  36 #include <rdma/ib_cache.h>
  37
  38 #include "mad_priv.h"
  39 #include "mad_rmpp.h"
  40 #include "smi.h"
  41 #include "agent.h"
  42
  43 MODULE_LICENSE("Dual BSD/GPL");
  44 MODULE_DESCRIPTION("kernel IB MAD API");
  45 MODULE_AUTHOR("Hal Rosenstock");
  46 MODULE_AUTHOR("Sean Hefty");
  47
  48 static struct kmem_cache *ib_mad_cache;
  49
  50 static struct list_head ib_mad_port_list;
  51 static u32 ib_mad_client_id = 0;
  52
  53 /* Port list lock */
  54 static spinlock_t ib_mad_port_list_lock;
  55
  56
  57 /* Forward declarations */
  58 static int method_in_use(struct ib_mad_mgmt_method_table **method,
  59                          struct ib_mad_reg_req *mad_reg_req);
  60 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
  61 static struct ib_mad_agent_private *find_mad_agent(
  62                                         struct ib_mad_port_private *port_priv,
  63                                         struct ib_mad *mad);
  64 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
  65                                     struct ib_mad_private *mad);
  66 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
  67 static void timeout_sends(struct work_struct *work);
  68 static void local_completions(struct work_struct *work);
  69 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  70                               struct ib_mad_agent_private *agent_priv,
  71                               u8 mgmt_class);
  72 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  73                            struct ib_mad_agent_private *agent_priv);
  74
  75 /*
  76  * Returns a ib_mad_port_private structure or NULL for a device/port
  77  * Assumes ib_mad_port_list_lock is being held
  78  */
  79 static inline struct ib_mad_port_private *
  80 __ib_get_mad_port(struct ib_device *device, int port_num)
  81 {
  82         struct ib_mad_port_private *entry;
  83
  84         list_for_each_entry(entry, &ib_mad_port_list, port_list) {
  85                 if (entry->device == device && entry->port_num == port_num)
  86                         return entry;
  87         }
  88         return NULL;
  89 }
  90
  91 /*
  92  * Wrapper function to return a ib_mad_port_private structure or NULL
  93  * for a device/port
  94  */
  95 static inline struct ib_mad_port_private *
  96 ib_get_mad_port(struct ib_device *device, int port_num)
  97 {
  98         struct ib_mad_port_private *entry;
  99         unsigned long flags;
 100
 101         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
 102         entry = __ib_get_mad_port(device, port_num);
 103         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
 104
 105         return entry;
 106 }
 107
 108 static inline u8 convert_mgmt_class(u8 mgmt_class)
 109 {
 110         /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
 111         return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
 112                 0 : mgmt_class;
 113 }
 114
 115 static int get_spl_qp_index(enum ib_qp_type qp_type)
 116 {
 117         switch (qp_type)
 118         {
 119         case IB_QPT_SMI:
 120                 return 0;
 121         case IB_QPT_GSI:
 122                 return 1;
 123         default:
 124                 return -1;
 125         }
 126 }
 127
 128 static int vendor_class_index(u8 mgmt_class)
 129 {
 130         return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
 131 }
 132
 133 static int is_vendor_class(u8 mgmt_class)
 134 {
 135         if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
 136             (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
 137                 return 0;
 138         return 1;
 139 }
 140
 141 static int is_vendor_oui(char *oui)
 142 {
 143         if (oui[0] || oui[1] || oui[2])
 144                 return 1;
 145         return 0;
 146 }
 147
 148 static int is_vendor_method_in_use(
 149                 struct ib_mad_mgmt_vendor_class *vendor_class,
 150                 struct ib_mad_reg_req *mad_reg_req)
 151 {
 152         struct ib_mad_mgmt_method_table *method;
 153         int i;
 154
 155         for (i = 0; i < MAX_MGMT_OUI; i++) {
 156                 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
 157                         method = vendor_class->method_table[i];
 158                         if (method) {
 159                                 if (method_in_use(&method, mad_reg_req))
 160                                         return 1;
 161                                 else
 162                                         break;
 163                         }
 164                 }
 165         }
 166         return 0;
 167 }
 168
 169 int ib_response_mad(struct ib_mad *mad)
 170 {
 171         return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
 172                 (mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
 173                 ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
 174                  (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
 175 }
 176 EXPORT_SYMBOL(ib_response_mad);
 177
 178 /*
 179  * ib_register_mad_agent - Register to send/receive MADs
 180  */
 181 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 182                                            u8 port_num,
 183                                            enum ib_qp_type qp_type,
 184                                            struct ib_mad_reg_req *mad_reg_req,
 185                                            u8 rmpp_version,
 186                                            ib_mad_send_handler send_handler,
 187                                            ib_mad_recv_handler recv_handler,
 188                                            void *context)
 189 {
 190         struct ib_mad_port_private *port_priv;
 191         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
 192         struct ib_mad_agent_private *mad_agent_priv;
 193         struct ib_mad_reg_req *reg_req = NULL;
 194         struct ib_mad_mgmt_class_table *class;
 195         struct ib_mad_mgmt_vendor_class_table *vendor;
 196         struct ib_mad_mgmt_vendor_class *vendor_class;
 197         struct ib_mad_mgmt_method_table *method;
 198         int ret2, qpn;
 199         unsigned long flags;
 200         u8 mgmt_class, vclass;
 201
 202         /* Validate parameters */
 203         qpn = get_spl_qp_index(qp_type);
 204         if (qpn == -1)
 205                 goto error1;
 206
 207         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
 208                 goto error1;
 209
 210         /* Validate MAD registration request if supplied */
 211         if (mad_reg_req) {
 212                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
 213                         goto error1;
 214                 if (!recv_handler)
 215                         goto error1;
 216                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
 217                         /*
 218                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
 219                          * one in this range currently allowed
 220                          */
 221                         if (mad_reg_req->mgmt_class !=
 222                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 223                                 goto error1;
 224                 } else if (mad_reg_req->mgmt_class == 0) {
 225                         /*
 226                          * Class 0 is reserved in IBA and is used for
 227                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
 228                          */
 229                         goto error1;
 230                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
 231                         /*
 232                          * If class is in "new" vendor range,
 233                          * ensure supplied OUI is not zero
 234                          */
 235                         if (!is_vendor_oui(mad_reg_req->oui))
 236                                 goto error1;
 237                 }
 238                 /* Make sure class supplied is consistent with RMPP */
 239                 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
 240                         if (rmpp_version)
 241                                 goto error1;
 242                 }
 243                 /* Make sure class supplied is consistent with QP type */
 244                 if (qp_type == IB_QPT_SMI) {
 245                         if ((mad_reg_req->mgmt_class !=
 246                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
 247                             (mad_reg_req->mgmt_class !=
 248                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 249                                 goto error1;
 250                 } else {
 251                         if ((mad_reg_req->mgmt_class ==
 252                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
 253                             (mad_reg_req->mgmt_class ==
 254                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 255                                 goto error1;
 256                 }
 257         } else {
 258                 /* No registration request supplied */
 259                 if (!send_handler)
 260                         goto error1;
 261         }
 262
 263         /* Validate device and port */
 264         port_priv = ib_get_mad_port(device, port_num);
 265         if (!port_priv) {
 266                 ret = ERR_PTR(-ENODEV);
 267                 goto error1;
 268         }
 269
 270         /* Allocate structures */
 271         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
 272         if (!mad_agent_priv) {
 273                 ret = ERR_PTR(-ENOMEM);
 274                 goto error1;
 275         }
 276
 277         mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
 278                                                  IB_ACCESS_LOCAL_WRITE);
 279         if (IS_ERR(mad_agent_priv->agent.mr)) {
 280                 ret = ERR_PTR(-ENOMEM);
 281                 goto error2;
 282         }
 283
 284         if (mad_reg_req) {
 285                 reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
 286                 if (!reg_req) {
 287                         ret = ERR_PTR(-ENOMEM);
 288                         goto error3;
 289                 }
 290                 /* Make a copy of the MAD registration request */
 291                 memcpy(reg_req, mad_reg_req, sizeof *reg_req);
 292         }
 293
 294         /* Now, fill in the various structures */
 295         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 296         mad_agent_priv->reg_req = reg_req;
 297         mad_agent_priv->agent.rmpp_version = rmpp_version;
 298         mad_agent_priv->agent.device = device;
 299         mad_agent_priv->agent.recv_handler = recv_handler;
 300         mad_agent_priv->agent.send_handler = send_handler;
 301         mad_agent_priv->agent.context = context;
 302         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 303         mad_agent_priv->agent.port_num = port_num;
 304
 305         spin_lock_irqsave(&port_priv->reg_lock, flags);
 306         mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
 307
 308         /*
 309          * Make sure MAD registration (if supplied)
 310          * is non overlapping with any existing ones
 311          */
 312         if (mad_reg_req) {
 313                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 314                 if (!is_vendor_class(mgmt_class)) {
 315                         class = port_priv->version[mad_reg_req->
 316                                                    mgmt_class_version].class;
 317                         if (class) {
 318                                 method = class->method_table[mgmt_class];
 319                                 if (method) {
 320                                         if (method_in_use(&method,
 321                                                            mad_reg_req))
 322                                                 goto error4;
 323                                 }
 324                         }
 325                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 326                                                   mgmt_class);
 327                 } else {
 328                         /* "New" vendor class range */
 329                         vendor = port_priv->version[mad_reg_req->
 330                                                     mgmt_class_version].vendor;
 331                         if (vendor) {
 332                                 vclass = vendor_class_index(mgmt_class);
 333                                 vendor_class = vendor->vendor_class[vclass];
 334                                 if (vendor_class) {
 335                                         if (is_vendor_method_in_use(
 336                                                         vendor_class,
 337                                                         mad_reg_req))
 338                                                 goto error4;
 339                                 }
 340                         }
 341                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 342                 }
 343                 if (ret2) {
 344                         ret = ERR_PTR(ret2);
 345                         goto error4;
 346                 }
 347         }
 348
 349         /* Add mad agent into port's agent list */
 350         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
 351         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 352
 353         spin_lock_init(&mad_agent_priv->lock);
 354         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 355         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 356         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 357         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 358         INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
 359         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 360         INIT_WORK(&mad_agent_priv->local_work, local_completions);
 361         atomic_set(&mad_agent_priv->refcount, 1);
 362         init_completion(&mad_agent_priv->comp);
 363
 364         return &mad_agent_priv->agent;
 365
 366 error4:
 367         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 368         kfree(reg_req);
 369 error3:
 370         ib_dereg_mr(mad_agent_priv->agent.mr);
 371 error2:
 372         kfree(mad_agent_priv);
 373 error1:
 374         return ret;
 375 }
 376 EXPORT_SYMBOL(ib_register_mad_agent);
 377
 378 static inline int is_snooping_sends(int mad_snoop_flags)
 379 {
 380         return (mad_snoop_flags &
 381                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 382                  IB_MAD_SNOOP_RMPP_SENDS |*/
 383                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 384                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 385 }
 386
 387 static inline int is_snooping_recvs(int mad_snoop_flags)
 388 {
 389         return (mad_snoop_flags &
 390                 (IB_MAD_SNOOP_RECVS /*|
 391                  IB_MAD_SNOOP_RMPP_RECVS*/));
 392 }
 393
 394 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 395                                 struct ib_mad_snoop_private *mad_snoop_priv)
 396 {
 397         struct ib_mad_snoop_private **new_snoop_table;
 398         unsigned long flags;
 399         int i;
 400
 401         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 402         /* Check for empty slot in array. */
 403         for (i = 0; i < qp_info->snoop_table_size; i++)
 404                 if (!qp_info->snoop_table[i])
 405                         break;
 406
 407         if (i == qp_info->snoop_table_size) {
 408                 /* Grow table. */
 409                 new_snoop_table = kmalloc(sizeof mad_snoop_priv *
 410                                           qp_info->snoop_table_size + 1,
 411                                           GFP_ATOMIC);
 412                 if (!new_snoop_table) {
 413                         i = -ENOMEM;
 414                         goto out;
 415                 }
 416                 if (qp_info->snoop_table) {
 417                         memcpy(new_snoop_table, qp_info->snoop_table,
 418                                sizeof mad_snoop_priv *
 419                                qp_info->snoop_table_size);
 420                         kfree(qp_info->snoop_table);
 421                 }
 422                 qp_info->snoop_table = new_snoop_table;
 423                 qp_info->snoop_table_size++;
 424         }
 425         qp_info->snoop_table[i] = mad_snoop_priv;
 426         atomic_inc(&qp_info->snoop_count);
 427 out:
 428         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 429         return i;
 430 }
 431
 432 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 433                                            u8 port_num,
 434                                            enum ib_qp_type qp_type,
 435                                            int mad_snoop_flags,
 436                                            ib_mad_snoop_handler snoop_handler,
 437                                            ib_mad_recv_handler recv_handler,
 438                                            void *context)
 439 {
 440         struct ib_mad_port_private *port_priv;
 441         struct ib_mad_agent *ret;
 442         struct ib_mad_snoop_private *mad_snoop_priv;
 443         int qpn;
 444
 445         /* Validate parameters */
 446         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 447             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 448                 ret = ERR_PTR(-EINVAL);
 449                 goto error1;
 450         }
 451         qpn = get_spl_qp_index(qp_type);
 452         if (qpn == -1) {
 453                 ret = ERR_PTR(-EINVAL);
 454                 goto error1;
 455         }
 456         port_priv = ib_get_mad_port(device, port_num);
 457         if (!port_priv) {
 458                 ret = ERR_PTR(-ENODEV);
 459                 goto error1;
 460         }
 461         /* Allocate structures */
 462         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 463         if (!mad_snoop_priv) {
 464                 ret = ERR_PTR(-ENOMEM);
 465                 goto error1;
 466         }
 467
 468         /* Now, fill in the various structures */
 469         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 470         mad_snoop_priv->agent.device = device;
 471         mad_snoop_priv->agent.recv_handler = recv_handler;
 472         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 473         mad_snoop_priv->agent.context = context;
 474         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 475         mad_snoop_priv->agent.port_num = port_num;
 476         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 477         init_completion(&mad_snoop_priv->comp);
 478         mad_snoop_priv->snoop_index = register_snoop_agent(
 479                                                 &port_priv->qp_info[qpn],
 480                                                 mad_snoop_priv);
 481         if (mad_snoop_priv->snoop_index < 0) {
 482                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 483                 goto error2;
 484         }
 485
 486         atomic_set(&mad_snoop_priv->refcount, 1);
 487         return &mad_snoop_priv->agent;
 488
 489 error2:
 490         kfree(mad_snoop_priv);
 491 error1:
 492         return ret;
 493 }
 494 EXPORT_SYMBOL(ib_register_mad_snoop);
 495
 496 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 497 {
 498         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 499                 complete(&mad_agent_priv->comp);
 500 }
 501
 502 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
 503 {
 504         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 505                 complete(&mad_snoop_priv->comp);
 506 }
 507
 508 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 509 {
 510         struct ib_mad_port_private *port_priv;
 511         unsigned long flags;
 512
 513         /* Note that we could still be handling received MADs */
 514
 515         /*
 516          * Canceling all sends results in dropping received response
 517          * MADs, preventing us from queuing additional work
 518          */
 519         cancel_mads(mad_agent_priv);
 520         port_priv = mad_agent_priv->qp_info->port_priv;
 521         cancel_delayed_work(&mad_agent_priv->timed_work);
 522
 523         spin_lock_irqsave(&port_priv->reg_lock, flags);
 524         remove_mad_reg_req(mad_agent_priv);
 525         list_del(&mad_agent_priv->agent_list);
 526         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 527
 528         flush_workqueue(port_priv->wq);
 529         ib_cancel_rmpp_recvs(mad_agent_priv);
 530
 531         deref_mad_agent(mad_agent_priv);
 532         wait_for_completion(&mad_agent_priv->comp);
 533
 534         kfree(mad_agent_priv->reg_req);
 535         ib_dereg_mr(mad_agent_priv->agent.mr);
 536         kfree(mad_agent_priv);
 537 }
 538
 539 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 540 {
 541         struct ib_mad_qp_info *qp_info;
 542         unsigned long flags;
 543
 544         qp_info = mad_snoop_priv->qp_info;
 545         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 546         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 547         atomic_dec(&qp_info->snoop_count);
 548         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 549
 550         deref_snoop_agent(mad_snoop_priv);
 551         wait_for_completion(&mad_snoop_priv->comp);
 552
 553         kfree(mad_snoop_priv);
 554 }
 555
 556 /*
 557  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 558  */
 559 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 560 {
 561         struct ib_mad_agent_private *mad_agent_priv;
 562         struct ib_mad_snoop_private *mad_snoop_priv;
 563
 564         /* If the TID is zero, the agent can only snoop. */
 565         if (mad_agent->hi_tid) {
 566                 mad_agent_priv = container_of(mad_agent,
 567                                               struct ib_mad_agent_private,
 568                                               agent);
 569                 unregister_mad_agent(mad_agent_priv);
 570         } else {
 571                 mad_snoop_priv = container_of(mad_agent,
 572                                               struct ib_mad_snoop_private,
 573                                               agent);
 574                 unregister_mad_snoop(mad_snoop_priv);
 575         }
 576         return 0;
 577 }
 578 EXPORT_SYMBOL(ib_unregister_mad_agent);
 579
 580 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 581 {
 582         struct ib_mad_queue *mad_queue;
 583         unsigned long flags;
 584
 585         BUG_ON(!mad_list->mad_queue);
 586         mad_queue = mad_list->mad_queue;
 587         spin_lock_irqsave(&mad_queue->lock, flags);
 588         list_del(&mad_list->list);
 589         mad_queue->count--;
 590         spin_unlock_irqrestore(&mad_queue->lock, flags);
 591 }
 592
 593 static void snoop_send(struct ib_mad_qp_info *qp_info,
 594                        struct ib_mad_send_buf *send_buf,
 595                        struct ib_mad_send_wc *mad_send_wc,
 596                        int mad_snoop_flags)
 597 {
 598         struct ib_mad_snoop_private *mad_snoop_priv;
 599         unsigned long flags;
 600         int i;
 601
 602         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 603         for (i = 0; i < qp_info->snoop_table_size; i++) {
 604                 mad_snoop_priv = qp_info->snoop_table[i];
 605                 if (!mad_snoop_priv ||
 606                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 607                         continue;
 608
 609                 atomic_inc(&mad_snoop_priv->refcount);
 610                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 611                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 612                                                     send_buf, mad_send_wc);
 613                 deref_snoop_agent(mad_snoop_priv);
 614                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 615         }
 616         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 617 }
 618
 619 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 620                        struct ib_mad_recv_wc *mad_recv_wc,
 621                        int mad_snoop_flags)
 622 {
 623         struct ib_mad_snoop_private *mad_snoop_priv;
 624         unsigned long flags;
 625         int i;
 626
 627         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 628         for (i = 0; i < qp_info->snoop_table_size; i++) {
 629                 mad_snoop_priv = qp_info->snoop_table[i];
 630                 if (!mad_snoop_priv ||
 631                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 632                         continue;
 633
 634                 atomic_inc(&mad_snoop_priv->refcount);
 635                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 636                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
 637                                                    mad_recv_wc);
 638                 deref_snoop_agent(mad_snoop_priv);
 639                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 640         }
 641         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 642 }
 643
 644 static void build_smp_wc(struct ib_qp *qp,
 645                          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 = qp;
 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 = 0;
 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              IB_SMI_DISCARD) {
 692                 ret = -EINVAL;
 693                 printk(KERN_ERR PFX "Invalid directed route\n");
 694                 goto out;
 695         }
 696
 697         /* Check to post send on QP or process locally */
 698         if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD)
 699                 goto out;
 700
 701         local = kmalloc(sizeof *local, GFP_ATOMIC);
 702         if (!local) {
 703                 ret = -ENOMEM;
 704                 printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
 705                 goto out;
 706         }
 707         local->mad_priv = NULL;
 708         local->recv_mad_agent = NULL;
 709         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
 710         if (!mad_priv) {
 711                 ret = -ENOMEM;
 712                 printk(KERN_ERR PFX "No memory for local response MAD\n");
 713                 kfree(local);
 714                 goto out;
 715         }
 716
 717         build_smp_wc(mad_agent_priv->agent.qp,
 718                      send_wr->wr_id, be16_to_cpu(smp->dr_slid),
 719                      send_wr->wr.ud.pkey_index,
 720                      send_wr->wr.ud.port_num, &mad_wc);
 721
 722         /* No GRH for DR SMP */
 723         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 724                                   (struct ib_mad *)smp,
 725                                   (struct ib_mad *)&mad_priv->mad);
 726         switch (ret)
 727         {
 728         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 729                 if (ib_response_mad(&mad_priv->mad.mad) &&
 730                     mad_agent_priv->agent.recv_handler) {
 731                         local->mad_priv = mad_priv;
 732                         local->recv_mad_agent = mad_agent_priv;
 733                         /*
 734                          * Reference MAD agent until receive
 735                          * side of local completion handled
 736                          */
 737                         atomic_inc(&mad_agent_priv->refcount);
 738                 } else
 739                         kmem_cache_free(ib_mad_cache, mad_priv);
 740                 break;
 741         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 742                 kmem_cache_free(ib_mad_cache, mad_priv);
 743                 break;
 744         case IB_MAD_RESULT_SUCCESS:
 745                 /* Treat like an incoming receive MAD */
 746                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 747                                             mad_agent_priv->agent.port_num);
 748                 if (port_priv) {
 749                         mad_priv->mad.mad.mad_hdr.tid =
 750                                 ((struct ib_mad *)smp)->mad_hdr.tid;
 751                         recv_mad_agent = find_mad_agent(port_priv,
 752                                                         &mad_priv->mad.mad);
 753                 }
 754                 if (!port_priv || !recv_mad_agent) {
 755                         kmem_cache_free(ib_mad_cache, mad_priv);
 756                         kfree(local);
 757                         ret = 0;
 758                         goto out;
 759                 }
 760                 local->mad_priv = mad_priv;
 761                 local->recv_mad_agent = recv_mad_agent;
 762                 break;
 763         default:
 764                 kmem_cache_free(ib_mad_cache, mad_priv);
 765                 kfree(local);
 766                 ret = -EINVAL;
 767                 goto out;
 768         }
 769
 770         local->mad_send_wr = mad_send_wr;
 771         /* Reference MAD agent until send side of local completion handled */
 772         atomic_inc(&mad_agent_priv->refcount);
 773         /* Queue local completion to local list */
 774         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 775         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 776         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 777         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 778                    &mad_agent_priv->local_work);
 779         ret = 1;
 780 out:
 781         return ret;
 782 }
 783
 784 static int get_pad_size(int hdr_len, int data_len)
 785 {
 786         int seg_size, pad;
 787
 788         seg_size = sizeof(struct ib_mad) - hdr_len;
 789         if (data_len && seg_size) {
 790                 pad = seg_size - data_len % seg_size;
 791                 return pad == seg_size ? 0 : pad;
 792         } else
 793                 return seg_size;
 794 }
 795
 796 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 797 {
 798         struct ib_rmpp_segment *s, *t;
 799
 800         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 801                 list_del(&s->list);
 802                 kfree(s);
 803         }
 804 }
 805
 806 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 807                                 gfp_t gfp_mask)
 808 {
 809         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 810         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 811         struct ib_rmpp_segment *seg = NULL;
 812         int left, seg_size, pad;
 813
 814         send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
 815         seg_size = send_buf->seg_size;
 816         pad = send_wr->pad;
 817
 818         /* Allocate data segments. */
 819         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 820                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 821                 if (!seg) {
 822                         printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
 823                                "alloc failed for len %zd, gfp %#x\n",
 824                                sizeof (*seg) + seg_size, gfp_mask);
 825                         free_send_rmpp_list(send_wr);
 826                         return -ENOMEM;
 827                 }
 828                 seg->num = ++send_buf->seg_count;
 829                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 830         }
 831
 832         /* Zero any padding */
 833         if (pad)
 834                 memset(seg->data + seg_size - pad, 0, pad);
 835
 836         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
 837                                           agent.rmpp_version;
 838         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 839         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
 840
 841         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
 842                                         struct ib_rmpp_segment, list);
 843         send_wr->last_ack_seg = send_wr->cur_seg;
 844         return 0;
 845 }
 846
 847 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 848                                             u32 remote_qpn, u16 pkey_index,
 849                                             int rmpp_active,
 850                                             int hdr_len, int data_len,
 851                                             gfp_t gfp_mask)
 852 {
 853         struct ib_mad_agent_private *mad_agent_priv;
 854         struct ib_mad_send_wr_private *mad_send_wr;
 855         int pad, message_size, ret, size;
 856         void *buf;
 857
 858         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 859                                       agent);
 860         pad = get_pad_size(hdr_len, data_len);
 861         message_size = hdr_len + data_len + pad;
 862
 863         if ((!mad_agent->rmpp_version &&
 864              (rmpp_active || message_size > sizeof(struct ib_mad))) ||
 865             (!rmpp_active && message_size > sizeof(struct ib_mad)))
 866                 return ERR_PTR(-EINVAL);
 867
 868         size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
 869         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
 870         if (!buf)
 871                 return ERR_PTR(-ENOMEM);
 872
 873         mad_send_wr = buf + size;
 874         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
 875         mad_send_wr->send_buf.mad = buf;
 876         mad_send_wr->send_buf.hdr_len = hdr_len;
 877         mad_send_wr->send_buf.data_len = data_len;
 878         mad_send_wr->pad = pad;
 879
 880         mad_send_wr->mad_agent_priv = mad_agent_priv;
 881         mad_send_wr->sg_list[0].length = hdr_len;
 882         mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
 883         mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
 884         mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
 885
 886         mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
 887         mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 888         mad_send_wr->send_wr.num_sge = 2;
 889         mad_send_wr->send_wr.opcode = IB_WR_SEND;
 890         mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
 891         mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
 892         mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 893         mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
 894
 895         if (rmpp_active) {
 896                 ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
 897                 if (ret) {
 898                         kfree(buf);
 899                         return ERR_PTR(ret);
 900                 }
 901         }
 902
 903         mad_send_wr->send_buf.mad_agent = mad_agent;
 904         atomic_inc(&mad_agent_priv->refcount);
 905         return &mad_send_wr->send_buf;
 906 }
 907 EXPORT_SYMBOL(ib_create_send_mad);
 908
 909 int ib_get_mad_data_offset(u8 mgmt_class)
 910 {
 911         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
 912                 return IB_MGMT_SA_HDR;
 913         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 914                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 915                  (mgmt_class == IB_MGMT_CLASS_BIS))
 916                 return IB_MGMT_DEVICE_HDR;
 917         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 918                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
 919                 return IB_MGMT_VENDOR_HDR;
 920         else
 921                 return IB_MGMT_MAD_HDR;
 922 }
 923 EXPORT_SYMBOL(ib_get_mad_data_offset);
 924
 925 int ib_is_mad_class_rmpp(u8 mgmt_class)
 926 {
 927         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
 928             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 929             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 930             (mgmt_class == IB_MGMT_CLASS_BIS) ||
 931             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 932              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
 933                 return 1;
 934         return 0;
 935 }
 936 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
 937
 938 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
 939 {
 940         struct ib_mad_send_wr_private *mad_send_wr;
 941         struct list_head *list;
 942
 943         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 944                                    send_buf);
 945         list = &mad_send_wr->cur_seg->list;
 946
 947         if (mad_send_wr->cur_seg->num < seg_num) {
 948                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
 949                         if (mad_send_wr->cur_seg->num == seg_num)
 950                                 break;
 951         } else if (mad_send_wr->cur_seg->num > seg_num) {
 952                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
 953                         if (mad_send_wr->cur_seg->num == seg_num)
 954                                 break;
 955         }
 956         return mad_send_wr->cur_seg->data;
 957 }
 958 EXPORT_SYMBOL(ib_get_rmpp_segment);
 959
 960 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
 961 {
 962         if (mad_send_wr->send_buf.seg_count)
 963                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
 964                                            mad_send_wr->seg_num);
 965         else
 966                 return mad_send_wr->send_buf.mad +
 967                        mad_send_wr->send_buf.hdr_len;
 968 }
 969
 970 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
 971 {
 972         struct ib_mad_agent_private *mad_agent_priv;
 973         struct ib_mad_send_wr_private *mad_send_wr;
 974
 975         mad_agent_priv = container_of(send_buf->mad_agent,
 976                                       struct ib_mad_agent_private, agent);
 977         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 978                                    send_buf);
 979
 980         free_send_rmpp_list(mad_send_wr);
 981         kfree(send_buf->mad);
 982         deref_mad_agent(mad_agent_priv);
 983 }
 984 EXPORT_SYMBOL(ib_free_send_mad);
 985
 986 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 987 {
 988         struct ib_mad_qp_info *qp_info;
 989         struct list_head *list;
 990         struct ib_send_wr *bad_send_wr;
 991         struct ib_mad_agent *mad_agent;
 992         struct ib_sge *sge;
 993         unsigned long flags;
 994         int ret;
 995
 996         /* Set WR ID to find mad_send_wr upon completion */
 997         qp_info = mad_send_wr->mad_agent_priv->qp_info;
 998         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
 999         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1000
1001         mad_agent = mad_send_wr->send_buf.mad_agent;
1002         sge = mad_send_wr->sg_list;
1003         sge[0].addr = ib_dma_map_single(mad_agent->device,
1004                                         mad_send_wr->send_buf.mad,
1005                                         sge[0].length,
1006                                         DMA_TO_DEVICE);
1007         mad_send_wr->header_mapping = sge[0].addr;
1008
1009         sge[1].addr = ib_dma_map_single(mad_agent->device,
1010                                         ib_get_payload(mad_send_wr),
1011                                         sge[1].length,
1012                                         DMA_TO_DEVICE);
1013         mad_send_wr->payload_mapping = sge[1].addr;
1014
1015         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1016         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1017                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1018                                    &bad_send_wr);
1019                 list = &qp_info->send_queue.list;
1020         } else {
1021                 ret = 0;
1022                 list = &qp_info->overflow_list;
1023         }
1024
1025         if (!ret) {
1026                 qp_info->send_queue.count++;
1027                 list_add_tail(&mad_send_wr->mad_list.list, list);
1028         }
1029         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1030         if (ret) {
1031                 ib_dma_unmap_single(mad_agent->device,
1032                                     mad_send_wr->header_mapping,
1033                                     sge[0].length, DMA_TO_DEVICE);
1034                 ib_dma_unmap_single(mad_agent->device,
1035                                     mad_send_wr->payload_mapping,
1036                                     sge[1].length, DMA_TO_DEVICE);
1037         }
1038         return ret;
1039 }
1040
1041 /*
1042  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1043  *  with the registered client
1044  */
1045 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1046                      struct ib_mad_send_buf **bad_send_buf)
1047 {
1048         struct ib_mad_agent_private *mad_agent_priv;
1049         struct ib_mad_send_buf *next_send_buf;
1050         struct ib_mad_send_wr_private *mad_send_wr;
1051         unsigned long flags;
1052         int ret = -EINVAL;
1053
1054         /* Walk list of send WRs and post each on send list */
1055         for (; send_buf; send_buf = next_send_buf) {
1056
1057                 mad_send_wr = container_of(send_buf,
1058                                            struct ib_mad_send_wr_private,
1059                                            send_buf);
1060                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1061
1062                 if (!send_buf->mad_agent->send_handler ||
1063                     (send_buf->timeout_ms &&
1064                      !send_buf->mad_agent->recv_handler)) {
1065                         ret = -EINVAL;
1066                         goto error;
1067                 }
1068
1069                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1070                         if (mad_agent_priv->agent.rmpp_version) {
1071                                 ret = -EINVAL;
1072                                 goto error;
1073                         }
1074                 }
1075
1076                 /*
1077                  * Save pointer to next work request to post in case the
1078                  * current one completes, and the user modifies the work
1079                  * request associated with the completion
1080                  */
1081                 next_send_buf = send_buf->next;
1082                 mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1083
1084                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1085                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1086                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1087                                                      mad_send_wr);
1088                         if (ret < 0)            /* error */
1089                                 goto error;
1090                         else if (ret == 1)      /* locally consumed */
1091                                 continue;
1092                 }
1093
1094                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1095                 /* Timeout will be updated after send completes */
1096                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1097                 mad_send_wr->retries = send_buf->retries;
1098                 /* Reference for work request to QP + response */
1099                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1100                 mad_send_wr->status = IB_WC_SUCCESS;
1101
1102                 /* Reference MAD agent until send completes */
1103                 atomic_inc(&mad_agent_priv->refcount);
1104                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1105                 list_add_tail(&mad_send_wr->agent_list,
1106                               &mad_agent_priv->send_list);
1107                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1108
1109                 if (mad_agent_priv->agent.rmpp_version) {
1110                         ret = ib_send_rmpp_mad(mad_send_wr);
1111                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1112                                 ret = ib_send_mad(mad_send_wr);
1113                 } else
1114                         ret = ib_send_mad(mad_send_wr);
1115                 if (ret < 0) {
1116                         /* Fail send request */
1117                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1118                         list_del(&mad_send_wr->agent_list);
1119                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1120                         atomic_dec(&mad_agent_priv->refcount);
1121                         goto error;
1122                 }
1123         }
1124         return 0;
1125 error:
1126         if (bad_send_buf)
1127                 *bad_send_buf = send_buf;
1128         return ret;
1129 }
1130 EXPORT_SYMBOL(ib_post_send_mad);
1131
1132 /*
1133  * ib_free_recv_mad - Returns data buffers used to receive
1134  *  a MAD to the access layer
1135  */
1136 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1137 {
1138         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1139         struct ib_mad_private_header *mad_priv_hdr;
1140         struct ib_mad_private *priv;
1141         struct list_head free_list;
1142
1143         INIT_LIST_HEAD(&free_list);
1144         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1145
1146         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1147                                         &free_list, list) {
1148                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1149                                            recv_buf);
1150                 mad_priv_hdr = container_of(mad_recv_wc,
1151                                             struct ib_mad_private_header,
1152                                             recv_wc);
1153                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1154                                     header);
1155                 kmem_cache_free(ib_mad_cache, priv);
1156         }
1157 }
1158 EXPORT_SYMBOL(ib_free_recv_mad);
1159
1160 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1161                                         u8 rmpp_version,
1162                                         ib_mad_send_handler send_handler,
1163                                         ib_mad_recv_handler recv_handler,
1164                                         void *context)
1165 {
1166         return ERR_PTR(-EINVAL);        /* XXX: for now */
1167 }
1168 EXPORT_SYMBOL(ib_redirect_mad_qp);
1169
1170 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1171                       struct ib_wc *wc)
1172 {
1173         printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1174         return 0;
1175 }
1176 EXPORT_SYMBOL(ib_process_mad_wc);
1177
1178 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1179                          struct ib_mad_reg_req *mad_reg_req)
1180 {
1181         int i;
1182
1183         for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
1184              i < IB_MGMT_MAX_METHODS;
1185              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1186                                1+i)) {
1187                 if ((*method)->agent[i]) {
1188                         printk(KERN_ERR PFX "Method %d already in use\n", i);
1189                         return -EINVAL;
1190                 }
1191         }
1192         return 0;
1193 }
1194
1195 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1196 {
1197         /* Allocate management method table */
1198         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1199         if (!*method) {
1200                 printk(KERN_ERR PFX "No memory for "
1201                        "ib_mad_mgmt_method_table\n");
1202                 return -ENOMEM;
1203         }
1204
1205         return 0;
1206 }
1207
1208 /*
1209  * Check to see if there are any methods still in use
1210  */
1211 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1212 {
1213         int i;
1214
1215         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1216                 if (method->agent[i])
1217                         return 1;
1218         return 0;
1219 }
1220
1221 /*
1222  * Check to see if there are any method tables for this class still in use
1223  */
1224 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1225 {
1226         int i;
1227
1228         for (i = 0; i < MAX_MGMT_CLASS; i++)
1229                 if (class->method_table[i])
1230                         return 1;
1231         return 0;
1232 }
1233
1234 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1235 {
1236         int i;
1237
1238         for (i = 0; i < MAX_MGMT_OUI; i++)
1239                 if (vendor_class->method_table[i])
1240                         return 1;
1241         return 0;
1242 }
1243
1244 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1245                            char *oui)
1246 {
1247         int i;
1248
1249         for (i = 0; i < MAX_MGMT_OUI; i++)
1250                 /* Is there matching OUI for this vendor class ? */
1251                 if (!memcmp(vendor_class->oui[i], oui, 3))
1252                         return i;
1253
1254         return -1;
1255 }
1256
1257 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1258 {
1259         int i;
1260
1261         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1262                 if (vendor->vendor_class[i])
1263                         return 1;
1264
1265         return 0;
1266 }
1267
1268 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1269                                      struct ib_mad_agent_private *agent)
1270 {
1271         int i;
1272
1273         /* Remove any methods for this mad agent */
1274         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1275                 if (method->agent[i] == agent) {
1276                         method->agent[i] = NULL;
1277                 }
1278         }
1279 }
1280
1281 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1282                               struct ib_mad_agent_private *agent_priv,
1283                               u8 mgmt_class)
1284 {
1285         struct ib_mad_port_private *port_priv;
1286         struct ib_mad_mgmt_class_table **class;
1287         struct ib_mad_mgmt_method_table **method;
1288         int i, ret;
1289
1290         port_priv = agent_priv->qp_info->port_priv;
1291         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1292         if (!*class) {
1293                 /* Allocate management class table for "new" class version */
1294                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1295                 if (!*class) {
1296                         printk(KERN_ERR PFX "No memory for "
1297                                "ib_mad_mgmt_class_table\n");
1298                         ret = -ENOMEM;
1299                         goto error1;
1300                 }
1301
1302                 /* Allocate method table for this management class */
1303                 method = &(*class)->method_table[mgmt_class];
1304                 if ((ret = allocate_method_table(method)))
1305                         goto error2;
1306         } else {
1307                 method = &(*class)->method_table[mgmt_class];
1308                 if (!*method) {
1309                         /* Allocate method table for this management class */
1310                         if ((ret = allocate_method_table(method)))
1311                                 goto error1;
1312                 }
1313         }
1314
1315         /* Now, make sure methods are not already in use */
1316         if (method_in_use(method, mad_reg_req))
1317                 goto error3;
1318
1319         /* Finally, add in methods being registered */
1320         for (i = find_first_bit(mad_reg_req->method_mask,
1321                                 IB_MGMT_MAX_METHODS);
1322              i < IB_MGMT_MAX_METHODS;
1323              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1324                                1+i)) {
1325                 (*method)->agent[i] = agent_priv;
1326         }
1327         return 0;
1328
1329 error3:
1330         /* Remove any methods for this mad agent */
1331         remove_methods_mad_agent(*method, agent_priv);
1332         /* Now, check to see if there are any methods in use */
1333         if (!check_method_table(*method)) {
1334                 /* If not, release management method table */
1335                 kfree(*method);
1336                 *method = NULL;
1337         }
1338         ret = -EINVAL;
1339         goto error1;
1340 error2:
1341         kfree(*class);
1342         *class = NULL;
1343 error1:
1344         return ret;
1345 }
1346
1347 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1348                            struct ib_mad_agent_private *agent_priv)
1349 {
1350         struct ib_mad_port_private *port_priv;
1351         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1352         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1353         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1354         struct ib_mad_mgmt_method_table **method;
1355         int i, ret = -ENOMEM;
1356         u8 vclass;
1357
1358         /* "New" vendor (with OUI) class */
1359         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1360         port_priv = agent_priv->qp_info->port_priv;
1361         vendor_table = &port_priv->version[
1362                                 mad_reg_req->mgmt_class_version].vendor;
1363         if (!*vendor_table) {
1364                 /* Allocate mgmt vendor class table for "new" class version */
1365                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1366                 if (!vendor) {
1367                         printk(KERN_ERR PFX "No memory for "
1368                                "ib_mad_mgmt_vendor_class_table\n");
1369                         goto error1;
1370                 }
1371
1372                 *vendor_table = vendor;
1373         }
1374         if (!(*vendor_table)->vendor_class[vclass]) {
1375                 /* Allocate table for this management vendor class */
1376                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1377                 if (!vendor_class) {
1378                         printk(KERN_ERR PFX "No memory for "
1379                                "ib_mad_mgmt_vendor_class\n");
1380                         goto error2;
1381                 }
1382
1383                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1384         }
1385         for (i = 0; i < MAX_MGMT_OUI; i++) {
1386                 /* Is there matching OUI for this vendor class ? */
1387                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1388                             mad_reg_req->oui, 3)) {
1389                         method = &(*vendor_table)->vendor_class[
1390                                                 vclass]->method_table[i];
1391                         BUG_ON(!*method);
1392                         goto check_in_use;
1393                 }
1394         }
1395         for (i = 0; i < MAX_MGMT_OUI; i++) {
1396                 /* OUI slot available ? */
1397                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1398                                 vclass]->oui[i])) {
1399                         method = &(*vendor_table)->vendor_class[
1400                                 vclass]->method_table[i];
1401                         BUG_ON(*method);
1402                         /* Allocate method table for this OUI */
1403                         if ((ret = allocate_method_table(method)))
1404                                 goto error3;
1405                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1406                                mad_reg_req->oui, 3);
1407                         goto check_in_use;
1408                 }
1409         }
1410         printk(KERN_ERR PFX "All OUI slots in use\n");
1411         goto error3;
1412
1413 check_in_use:
1414         /* Now, make sure methods are not already in use */
1415         if (method_in_use(method, mad_reg_req))
1416                 goto error4;
1417
1418         /* Finally, add in methods being registered */
1419         for (i = find_first_bit(mad_reg_req->method_mask,
1420                                 IB_MGMT_MAX_METHODS);
1421              i < IB_MGMT_MAX_METHODS;
1422              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1423                                1+i)) {
1424                 (*method)->agent[i] = agent_priv;
1425         }
1426         return 0;
1427
1428 error4:
1429         /* Remove any methods for this mad agent */
1430         remove_methods_mad_agent(*method, agent_priv);
1431         /* Now, check to see if there are any methods in use */
1432         if (!check_method_table(*method)) {
1433                 /* If not, release management method table */
1434                 kfree(*method);
1435                 *method = NULL;
1436         }
1437         ret = -EINVAL;
1438 error3:
1439         if (vendor_class) {
1440                 (*vendor_table)->vendor_class[vclass] = NULL;
1441                 kfree(vendor_class);
1442         }
1443 error2:
1444         if (vendor) {
1445                 *vendor_table = NULL;
1446                 kfree(vendor);
1447         }
1448 error1:
1449         return ret;
1450 }
1451
1452 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1453 {
1454         struct ib_mad_port_private *port_priv;
1455         struct ib_mad_mgmt_class_table *class;
1456         struct ib_mad_mgmt_method_table *method;
1457         struct ib_mad_mgmt_vendor_class_table *vendor;
1458         struct ib_mad_mgmt_vendor_class *vendor_class;
1459         int index;
1460         u8 mgmt_class;
1461
1462         /*
1463          * Was MAD registration request supplied
1464          * with original registration ?
1465          */
1466         if (!agent_priv->reg_req) {
1467                 goto out;
1468         }
1469
1470         port_priv = agent_priv->qp_info->port_priv;
1471         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1472         class = port_priv->version[
1473                         agent_priv->reg_req->mgmt_class_version].class;
1474         if (!class)
1475                 goto vendor_check;
1476
1477         method = class->method_table[mgmt_class];
1478         if (method) {
1479                 /* Remove any methods for this mad agent */
1480                 remove_methods_mad_agent(method, agent_priv);
1481                 /* Now, check to see if there are any methods still in use */
1482                 if (!check_method_table(method)) {
1483                         /* If not, release management method table */
1484                          kfree(method);
1485                          class->method_table[mgmt_class] = NULL;
1486                          /* Any management classes left ? */
1487                         if (!check_class_table(class)) {
1488                                 /* If not, release management class table */
1489                                 kfree(class);
1490                                 port_priv->version[
1491                                         agent_priv->reg_req->
1492                                         mgmt_class_version].class = NULL;
1493                         }
1494                 }
1495         }
1496
1497 vendor_check:
1498         if (!is_vendor_class(mgmt_class))
1499                 goto out;
1500
1501         /* normalize mgmt_class to vendor range 2 */
1502         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1503         vendor = port_priv->version[
1504                         agent_priv->reg_req->mgmt_class_version].vendor;
1505
1506         if (!vendor)
1507                 goto out;
1508
1509         vendor_class = vendor->vendor_class[mgmt_class];
1510         if (vendor_class) {
1511                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1512                 if (index < 0)
1513                         goto out;
1514                 method = vendor_class->method_table[index];
1515                 if (method) {
1516                         /* Remove any methods for this mad agent */
1517                         remove_methods_mad_agent(method, agent_priv);
1518                         /*
1519                          * Now, check to see if there are
1520                          * any methods still in use
1521                          */
1522                         if (!check_method_table(method)) {
1523                                 /* If not, release management method table */
1524                                 kfree(method);
1525                                 vendor_class->method_table[index] = NULL;
1526                                 memset(vendor_class->oui[index], 0, 3);
1527                                 /* Any OUIs left ? */
1528                                 if (!check_vendor_class(vendor_class)) {
1529                                         /* If not, release vendor class table */
1530                                         kfree(vendor_class);
1531                                         vendor->vendor_class[mgmt_class] = NULL;
1532                                         /* Any other vendor classes left ? */
1533                                         if (!check_vendor_table(vendor)) {
1534                                                 kfree(vendor);
1535                                                 port_priv->version[
1536                                                         agent_priv->reg_req->
1537                                                         mgmt_class_version].
1538                                                         vendor = NULL;
1539                                         }
1540                                 }
1541                         }
1542                 }
1543         }
1544
1545 out:
1546         return;
1547 }
1548
1549 static struct ib_mad_agent_private *
1550 find_mad_agent(struct ib_mad_port_private *port_priv,
1551                struct ib_mad *mad)
1552 {
1553         struct ib_mad_agent_private *mad_agent = NULL;
1554         unsigned long flags;
1555
1556         spin_lock_irqsave(&port_priv->reg_lock, flags);
1557         if (ib_response_mad(mad)) {
1558                 u32 hi_tid;
1559                 struct ib_mad_agent_private *entry;
1560
1561                 /*
1562                  * Routing is based on high 32 bits of transaction ID
1563                  * of MAD.
1564                  */
1565                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1566                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1567                         if (entry->agent.hi_tid == hi_tid) {
1568                                 mad_agent = entry;
1569                                 break;
1570                         }
1571                 }
1572         } else {
1573                 struct ib_mad_mgmt_class_table *class;
1574                 struct ib_mad_mgmt_method_table *method;
1575                 struct ib_mad_mgmt_vendor_class_table *vendor;
1576                 struct ib_mad_mgmt_vendor_class *vendor_class;
1577                 struct ib_vendor_mad *vendor_mad;
1578                 int index;
1579
1580                 /*
1581                  * Routing is based on version, class, and method
1582                  * For "newer" vendor MADs, also based on OUI
1583                  */
1584                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1585                         goto out;
1586                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1587                         class = port_priv->version[
1588                                         mad->mad_hdr.class_version].class;
1589                         if (!class)
1590                                 goto out;
1591                         method = class->method_table[convert_mgmt_class(
1592                                                         mad->mad_hdr.mgmt_class)];
1593                         if (method)
1594                                 mad_agent = method->agent[mad->mad_hdr.method &
1595                                                           ~IB_MGMT_METHOD_RESP];
1596                 } else {
1597                         vendor = port_priv->version[
1598                                         mad->mad_hdr.class_version].vendor;
1599                         if (!vendor)
1600                                 goto out;
1601                         vendor_class = vendor->vendor_class[vendor_class_index(
1602                                                 mad->mad_hdr.mgmt_class)];
1603                         if (!vendor_class)
1604                                 goto out;
1605                         /* Find matching OUI */
1606                         vendor_mad = (struct ib_vendor_mad *)mad;
1607                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1608                         if (index == -1)
1609                                 goto out;
1610                         method = vendor_class->method_table[index];
1611                         if (method) {
1612                                 mad_agent = method->agent[mad->mad_hdr.method &
1613                                                           ~IB_MGMT_METHOD_RESP];
1614                         }
1615                 }
1616         }
1617
1618         if (mad_agent) {
1619                 if (mad_agent->agent.recv_handler)
1620                         atomic_inc(&mad_agent->refcount);
1621                 else {
1622                         printk(KERN_NOTICE PFX "No receive handler for client "
1623                                "%p on port %d\n",
1624                                &mad_agent->agent, port_priv->port_num);
1625                         mad_agent = NULL;
1626                 }
1627         }
1628 out:
1629         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1630
1631         return mad_agent;
1632 }
1633
1634 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1635 {
1636         int valid = 0;
1637
1638         /* Make sure MAD base version is understood */
1639         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1640                 printk(KERN_ERR PFX "MAD received with unsupported base "
1641                        "version %d\n", mad->mad_hdr.base_version);
1642                 goto out;
1643         }
1644
1645         /* Filter SMI packets sent to other than QP0 */
1646         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1647             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1648                 if (qp_num == 0)
1649                         valid = 1;
1650         } else {
1651                 /* Filter GSI packets sent to QP0 */
1652                 if (qp_num != 0)
1653                         valid = 1;
1654         }
1655
1656 out:
1657         return valid;
1658 }
1659
1660 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1661                        struct ib_mad_hdr *mad_hdr)
1662 {
1663         struct ib_rmpp_mad *rmpp_mad;
1664
1665         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1666         return !mad_agent_priv->agent.rmpp_version ||
1667                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1668                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1669                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1670 }
1671
1672 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1673                                      struct ib_mad_recv_wc *rwc)
1674 {
1675         return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1676                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1677 }
1678
1679 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1680                                    struct ib_mad_send_wr_private *wr,
1681                                    struct ib_mad_recv_wc *rwc )
1682 {
1683         struct ib_ah_attr attr;
1684         u8 send_resp, rcv_resp;
1685         union ib_gid sgid;
1686         struct ib_device *device = mad_agent_priv->agent.device;
1687         u8 port_num = mad_agent_priv->agent.port_num;
1688         u8 lmc;
1689
1690         send_resp = ((struct ib_mad *)(wr->send_buf.mad))->
1691                      mad_hdr.method & IB_MGMT_METHOD_RESP;
1692         rcv_resp = rwc->recv_buf.mad->mad_hdr.method & IB_MGMT_METHOD_RESP;
1693
1694         if (send_resp == rcv_resp)
1695                 /* both requests, or both responses. GIDs different */
1696                 return 0;
1697
1698         if (ib_query_ah(wr->send_buf.ah, &attr))
1699                 /* Assume not equal, to avoid false positives. */
1700                 return 0;
1701
1702         if (!!(attr.ah_flags & IB_AH_GRH) !=
1703             !!(rwc->wc->wc_flags & IB_WC_GRH))
1704                 /* one has GID, other does not.  Assume different */
1705                 return 0;
1706
1707         if (!send_resp && rcv_resp) {
1708                 /* is request/response. */
1709                 if (!(attr.ah_flags & IB_AH_GRH)) {
1710                         if (ib_get_cached_lmc(device, port_num, &lmc))
1711                                 return 0;
1712                         return (!lmc || !((attr.src_path_bits ^
1713                                            rwc->wc->dlid_path_bits) &
1714                                           ((1 << lmc) - 1)));
1715                 } else {
1716                         if (ib_get_cached_gid(device, port_num,
1717                                               attr.grh.sgid_index, &sgid))
1718                                 return 0;
1719                         return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1720                                        16);
1721                 }
1722         }
1723
1724         if (!(attr.ah_flags & IB_AH_GRH))
1725                 return attr.dlid == rwc->wc->slid;
1726         else
1727                 return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1728                                16);
1729 }
1730
1731 static inline int is_direct(u8 class)
1732 {
1733         return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1734 }
1735
1736 struct ib_mad_send_wr_private*
1737 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1738                  struct ib_mad_recv_wc *wc)
1739 {
1740         struct ib_mad_send_wr_private *wr;
1741         struct ib_mad *mad;
1742
1743         mad = (struct ib_mad *)wc->recv_buf.mad;
1744
1745         list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1746                 if ((wr->tid == mad->mad_hdr.tid) &&
1747                     rcv_has_same_class(wr, wc) &&
1748                     /*
1749                      * Don't check GID for direct routed MADs.
1750                      * These might have permissive LIDs.
1751                      */
1752                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1753                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1754                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1755         }
1756
1757         /*
1758          * It's possible to receive the response before we've
1759          * been notified that the send has completed
1760          */
1761         list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1762                 if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1763                     wr->tid == mad->mad_hdr.tid &&
1764                     wr->timeout &&
1765                     rcv_has_same_class(wr, wc) &&
1766                     /*
1767                      * Don't check GID for direct routed MADs.
1768                      * These might have permissive LIDs.
1769                      */
1770                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1771                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1772                         /* Verify request has not been canceled */
1773                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1774         }
1775         return NULL;
1776 }
1777
1778 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1779 {
1780         mad_send_wr->timeout = 0;
1781         if (mad_send_wr->refcount == 1)
1782                 list_move_tail(&mad_send_wr->agent_list,
1783                               &mad_send_wr->mad_agent_priv->done_list);
1784 }
1785
1786 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1787                                  struct ib_mad_recv_wc *mad_recv_wc)
1788 {
1789         struct ib_mad_send_wr_private *mad_send_wr;
1790         struct ib_mad_send_wc mad_send_wc;
1791         unsigned long flags;
1792
1793         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1794         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1795         if (mad_agent_priv->agent.rmpp_version) {
1796                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1797                                                       mad_recv_wc);
1798                 if (!mad_recv_wc) {
1799                         deref_mad_agent(mad_agent_priv);
1800                         return;
1801                 }
1802         }
1803
1804         /* Complete corresponding request */
1805         if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1806                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1807                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1808                 if (!mad_send_wr) {
1809                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1810                         ib_free_recv_mad(mad_recv_wc);
1811                         deref_mad_agent(mad_agent_priv);
1812                         return;
1813                 }
1814                 ib_mark_mad_done(mad_send_wr);
1815                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1816
1817                 /* Defined behavior is to complete response before request */
1818                 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1819                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1820                                                    mad_recv_wc);
1821                 atomic_dec(&mad_agent_priv->refcount);
1822
1823                 mad_send_wc.status = IB_WC_SUCCESS;
1824                 mad_send_wc.vendor_err = 0;
1825                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1826                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1827         } else {
1828                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1829                                                    mad_recv_wc);
1830                 deref_mad_agent(mad_agent_priv);
1831         }
1832 }
1833
1834 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1835                                      struct ib_wc *wc)
1836 {
1837         struct ib_mad_qp_info *qp_info;
1838         struct ib_mad_private_header *mad_priv_hdr;
1839         struct ib_mad_private *recv, *response;
1840         struct ib_mad_list_head *mad_list;
1841         struct ib_mad_agent_private *mad_agent;
1842
1843         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1844         if (!response)
1845                 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1846                        "for response buffer\n");
1847
1848         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1849         qp_info = mad_list->mad_queue->qp_info;
1850         dequeue_mad(mad_list);
1851
1852         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1853                                     mad_list);
1854         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1855         ib_dma_unmap_single(port_priv->device,
1856                             recv->header.mapping,
1857                             sizeof(struct ib_mad_private) -
1858                               sizeof(struct ib_mad_private_header),
1859                             DMA_FROM_DEVICE);
1860
1861         /* Setup MAD receive work completion from "normal" work completion */
1862         recv->header.wc = *wc;
1863         recv->header.recv_wc.wc = &recv->header.wc;
1864         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1865         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1866         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1867
1868         if (atomic_read(&qp_info->snoop_count))
1869                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1870
1871         /* Validate MAD */
1872         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1873                 goto out;
1874
1875         if (recv->mad.mad.mad_hdr.mgmt_class ==
1876             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1877                 if (smi_handle_dr_smp_recv(&recv->mad.smp,
1878                                            port_priv->device->node_type,
1879                                            port_priv->port_num,
1880                                            port_priv->device->phys_port_cnt) ==
1881                                            IB_SMI_DISCARD)
1882                         goto out;
1883
1884                 if (smi_check_forward_dr_smp(&recv->mad.smp) == IB_SMI_LOCAL)
1885                         goto local;
1886
1887                 if (smi_handle_dr_smp_send(&recv->mad.smp,
1888                                            port_priv->device->node_type,
1889                                            port_priv->port_num) == IB_SMI_DISCARD)
1890                         goto out;
1891
1892                 if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1893                         goto out;
1894         }
1895
1896 local:
1897         /* Give driver "right of first refusal" on incoming MAD */
1898         if (port_priv->device->process_mad) {
1899                 int ret;
1900
1901                 if (!response) {
1902                         printk(KERN_ERR PFX "No memory for response MAD\n");
1903                         /*
1904                          * Is it better to assume that
1905                          * it wouldn't be processed ?
1906                          */
1907                         goto out;
1908                 }
1909
1910                 ret = port_priv->device->process_mad(port_priv->device, 0,
1911                                                      port_priv->port_num,
1912                                                      wc, &recv->grh,
1913                                                      &recv->mad.mad,
1914                                                      &response->mad.mad);
1915                 if (ret & IB_MAD_RESULT_SUCCESS) {
1916                         if (ret & IB_MAD_RESULT_CONSUMED)
1917                                 goto out;
1918                         if (ret & IB_MAD_RESULT_REPLY) {
1919                                 agent_send_response(&response->mad.mad,
1920                                                     &recv->grh, wc,
1921                                                     port_priv->device,
1922                                                     port_priv->port_num,
1923                                                     qp_info->qp->qp_num);
1924                                 goto out;
1925                         }
1926                 }
1927         }
1928
1929         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1930         if (mad_agent) {
1931                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1932                 /*
1933                  * recv is freed up in error cases in ib_mad_complete_recv
1934                  * or via recv_handler in ib_mad_complete_recv()
1935                  */
1936                 recv = NULL;
1937         }
1938
1939 out:
1940         /* Post another receive request for this QP */
1941         if (response) {
1942                 ib_mad_post_receive_mads(qp_info, response);
1943                 if (recv)
1944                         kmem_cache_free(ib_mad_cache, recv);
1945         } else
1946                 ib_mad_post_receive_mads(qp_info, recv);
1947 }
1948
1949 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1950 {
1951         struct ib_mad_send_wr_private *mad_send_wr;
1952         unsigned long delay;
1953
1954         if (list_empty(&mad_agent_priv->wait_list)) {
1955                 cancel_delayed_work(&mad_agent_priv->timed_work);
1956         } else {
1957                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1958                                          struct ib_mad_send_wr_private,
1959                                          agent_list);
1960
1961                 if (time_after(mad_agent_priv->timeout,
1962                                mad_send_wr->timeout)) {
1963                         mad_agent_priv->timeout = mad_send_wr->timeout;
1964                         cancel_delayed_work(&mad_agent_priv->timed_work);
1965                         delay = mad_send_wr->timeout - jiffies;
1966                         if ((long)delay <= 0)
1967                                 delay = 1;
1968                         queue_delayed_work(mad_agent_priv->qp_info->
1969                                            port_priv->wq,
1970                                            &mad_agent_priv->timed_work, delay);
1971                 }
1972         }
1973 }
1974
1975 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1976 {
1977         struct ib_mad_agent_private *mad_agent_priv;
1978         struct ib_mad_send_wr_private *temp_mad_send_wr;
1979         struct list_head *list_item;
1980         unsigned long delay;
1981
1982         mad_agent_priv = mad_send_wr->mad_agent_priv;
1983         list_del(&mad_send_wr->agent_list);
1984
1985         delay = mad_send_wr->timeout;
1986         mad_send_wr->timeout += jiffies;
1987
1988         if (delay) {
1989                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1990                         temp_mad_send_wr = list_entry(list_item,
1991                                                 struct ib_mad_send_wr_private,
1992                                                 agent_list);
1993                         if (time_after(mad_send_wr->timeout,
1994                                        temp_mad_send_wr->timeout))
1995                                 break;
1996                 }
1997         }
1998         else
1999                 list_item = &mad_agent_priv->wait_list;
2000         list_add(&mad_send_wr->agent_list, list_item);
2001
2002         /* Reschedule a work item if we have a shorter timeout */
2003         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
2004                 cancel_delayed_work(&mad_agent_priv->timed_work);
2005                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2006                                    &mad_agent_priv->timed_work, delay);
2007         }
2008 }
2009
2010 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2011                           int timeout_ms)
2012 {
2013         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2014         wait_for_response(mad_send_wr);
2015 }
2016
2017 /*
2018  * Process a send work completion
2019  */
2020 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2021                              struct ib_mad_send_wc *mad_send_wc)
2022 {
2023         struct ib_mad_agent_private     *mad_agent_priv;
2024         unsigned long                   flags;
2025         int                             ret;
2026
2027         mad_agent_priv = mad_send_wr->mad_agent_priv;
2028         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2029         if (mad_agent_priv->agent.rmpp_version) {
2030                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2031                 if (ret == IB_RMPP_RESULT_CONSUMED)
2032                         goto done;
2033         } else
2034                 ret = IB_RMPP_RESULT_UNHANDLED;
2035
2036         if (mad_send_wc->status != IB_WC_SUCCESS &&
2037             mad_send_wr->status == IB_WC_SUCCESS) {
2038                 mad_send_wr->status = mad_send_wc->status;
2039                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2040         }
2041
2042         if (--mad_send_wr->refcount > 0) {
2043                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2044                     mad_send_wr->status == IB_WC_SUCCESS) {
2045                         wait_for_response(mad_send_wr);
2046                 }
2047                 goto done;
2048         }
2049
2050         /* Remove send from MAD agent and notify client of completion */
2051         list_del(&mad_send_wr->agent_list);
2052         adjust_timeout(mad_agent_priv);
2053         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2054
2055         if (mad_send_wr->status != IB_WC_SUCCESS )
2056                 mad_send_wc->status = mad_send_wr->status;
2057         if (ret == IB_RMPP_RESULT_INTERNAL)
2058                 ib_rmpp_send_handler(mad_send_wc);
2059         else
2060                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2061                                                    mad_send_wc);
2062
2063         /* Release reference on agent taken when sending */
2064         deref_mad_agent(mad_agent_priv);
2065         return;
2066 done:
2067         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2068 }
2069
2070 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2071                                      struct ib_wc *wc)
2072 {
2073         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2074         struct ib_mad_list_head         *mad_list;
2075         struct ib_mad_qp_info           *qp_info;
2076         struct ib_mad_queue             *send_queue;
2077         struct ib_send_wr               *bad_send_wr;
2078         struct ib_mad_send_wc           mad_send_wc;
2079         unsigned long flags;
2080         int ret;
2081
2082         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2083         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2084                                    mad_list);
2085         send_queue = mad_list->mad_queue;
2086         qp_info = send_queue->qp_info;
2087
2088 retry:
2089         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2090                             mad_send_wr->header_mapping,
2091                             mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2092         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2093                             mad_send_wr->payload_mapping,
2094                             mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2095         queued_send_wr = NULL;
2096         spin_lock_irqsave(&send_queue->lock, flags);
2097         list_del(&mad_list->list);
2098
2099         /* Move queued send to the send queue */
2100         if (send_queue->count-- > send_queue->max_active) {
2101                 mad_list = container_of(qp_info->overflow_list.next,
2102                                         struct ib_mad_list_head, list);
2103                 queued_send_wr = container_of(mad_list,
2104                                         struct ib_mad_send_wr_private,
2105                                         mad_list);
2106                 list_move_tail(&mad_list->list, &send_queue->list);
2107         }
2108         spin_unlock_irqrestore(&send_queue->lock, flags);
2109
2110         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2111         mad_send_wc.status = wc->status;
2112         mad_send_wc.vendor_err = wc->vendor_err;
2113         if (atomic_read(&qp_info->snoop_count))
2114                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2115                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2116         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2117
2118         if (queued_send_wr) {
2119                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2120                                    &bad_send_wr);
2121                 if (ret) {
2122                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2123                         mad_send_wr = queued_send_wr;
2124                         wc->status = IB_WC_LOC_QP_OP_ERR;
2125                         goto retry;
2126                 }
2127         }
2128 }
2129
2130 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2131 {
2132         struct ib_mad_send_wr_private *mad_send_wr;
2133         struct ib_mad_list_head *mad_list;
2134         unsigned long flags;
2135
2136         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2137         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2138                 mad_send_wr = container_of(mad_list,
2139                                            struct ib_mad_send_wr_private,
2140                                            mad_list);
2141                 mad_send_wr->retry = 1;
2142         }
2143         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2144 }
2145
2146 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2147                               struct ib_wc *wc)
2148 {
2149         struct ib_mad_list_head *mad_list;
2150         struct ib_mad_qp_info *qp_info;
2151         struct ib_mad_send_wr_private *mad_send_wr;
2152         int ret;
2153
2154         /* Determine if failure was a send or receive */
2155         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2156         qp_info = mad_list->mad_queue->qp_info;
2157         if (mad_list->mad_queue == &qp_info->recv_queue)
2158                 /*
2159                  * Receive errors indicate that the QP has entered the error
2160                  * state - error handling/shutdown code will cleanup
2161                  */
2162                 return;
2163
2164         /*
2165          * Send errors will transition the QP to SQE - move
2166          * QP to RTS and repost flushed work requests
2167          */
2168         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2169                                    mad_list);
2170         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2171                 if (mad_send_wr->retry) {
2172                         /* Repost send */
2173                         struct ib_send_wr *bad_send_wr;
2174
2175                         mad_send_wr->retry = 0;
2176                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2177                                         &bad_send_wr);
2178                         if (ret)
2179                                 ib_mad_send_done_handler(port_priv, wc);
2180                 } else
2181                         ib_mad_send_done_handler(port_priv, wc);
2182         } else {
2183                 struct ib_qp_attr *attr;
2184
2185                 /* Transition QP to RTS and fail offending send */
2186                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2187                 if (attr) {
2188                         attr->qp_state = IB_QPS_RTS;
2189                         attr->cur_qp_state = IB_QPS_SQE;
2190                         ret = ib_modify_qp(qp_info->qp, attr,
2191                                            IB_QP_STATE | IB_QP_CUR_STATE);
2192                         kfree(attr);
2193                         if (ret)
2194                                 printk(KERN_ERR PFX "mad_error_handler - "
2195                                        "ib_modify_qp to RTS : %d\n", ret);
2196                         else
2197                                 mark_sends_for_retry(qp_info);
2198                 }
2199                 ib_mad_send_done_handler(port_priv, wc);
2200         }
2201 }
2202
2203 /*
2204  * IB MAD completion callback
2205  */
2206 static void ib_mad_completion_handler(struct work_struct *work)
2207 {
2208         struct ib_mad_port_private *port_priv;
2209         struct ib_wc wc;
2210
2211         port_priv = container_of(work, struct ib_mad_port_private, work);
2212         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2213
2214         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2215                 if (wc.status == IB_WC_SUCCESS) {
2216                         switch (wc.opcode) {
2217                         case IB_WC_SEND:
2218                                 ib_mad_send_done_handler(port_priv, &wc);
2219                                 break;
2220                         case IB_WC_RECV:
2221                                 ib_mad_recv_done_handler(port_priv, &wc);
2222                                 break;
2223                         default:
2224                                 BUG_ON(1);
2225                                 break;
2226                         }
2227                 } else
2228                         mad_error_handler(port_priv, &wc);
2229         }
2230 }
2231
2232 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2233 {
2234         unsigned long flags;
2235         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2236         struct ib_mad_send_wc mad_send_wc;
2237         struct list_head cancel_list;
2238
2239         INIT_LIST_HEAD(&cancel_list);
2240
2241         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2242         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2243                                  &mad_agent_priv->send_list, agent_list) {
2244                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2245                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2246                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2247                 }
2248         }
2249
2250         /* Empty wait list to prevent receives from finding a request */
2251         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2252         /* Empty local completion list as well */
2253         list_splice_init(&mad_agent_priv->local_list, &cancel_list);
2254         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2255
2256         /* Report all cancelled requests */
2257         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2258         mad_send_wc.vendor_err = 0;
2259
2260         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2261                                  &cancel_list, agent_list) {
2262                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2263                 list_del(&mad_send_wr->agent_list);
2264                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2265                                                    &mad_send_wc);
2266                 atomic_dec(&mad_agent_priv->refcount);
2267         }
2268 }
2269
2270 static struct ib_mad_send_wr_private*
2271 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2272              struct ib_mad_send_buf *send_buf)
2273 {
2274         struct ib_mad_send_wr_private *mad_send_wr;
2275
2276         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2277                             agent_list) {
2278                 if (&mad_send_wr->send_buf == send_buf)
2279                         return mad_send_wr;
2280         }
2281
2282         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2283                             agent_list) {
2284                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2285                     &mad_send_wr->send_buf == send_buf)
2286                         return mad_send_wr;
2287         }
2288         return NULL;
2289 }
2290
2291 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2292                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2293 {
2294         struct ib_mad_agent_private *mad_agent_priv;
2295         struct ib_mad_send_wr_private *mad_send_wr;
2296         unsigned long flags;
2297         int active;
2298
2299         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2300                                       agent);
2301         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2302         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2303         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2304                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2305                 return -EINVAL;
2306         }
2307
2308         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2309         if (!timeout_ms) {
2310                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2311                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2312         }
2313
2314         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2315         if (active)
2316                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2317         else
2318                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2319
2320         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2321         return 0;
2322 }
2323 EXPORT_SYMBOL(ib_modify_mad);
2324
2325 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2326                    struct ib_mad_send_buf *send_buf)
2327 {
2328         ib_modify_mad(mad_agent, send_buf, 0);
2329 }
2330 EXPORT_SYMBOL(ib_cancel_mad);
2331
2332 static void local_completions(struct work_struct *work)
2333 {
2334         struct ib_mad_agent_private *mad_agent_priv;
2335         struct ib_mad_local_private *local;
2336         struct ib_mad_agent_private *recv_mad_agent;
2337         unsigned long flags;
2338         int recv = 0;
2339         struct ib_wc wc;
2340         struct ib_mad_send_wc mad_send_wc;
2341
2342         mad_agent_priv =
2343                 container_of(work, struct ib_mad_agent_private, local_work);
2344
2345         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2346         while (!list_empty(&mad_agent_priv->local_list)) {
2347                 local = list_entry(mad_agent_priv->local_list.next,
2348                                    struct ib_mad_local_private,
2349                                    completion_list);
2350                 list_del(&local->completion_list);
2351                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2352                 if (local->mad_priv) {
2353                         recv_mad_agent = local->recv_mad_agent;
2354                         if (!recv_mad_agent) {
2355                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2356                                 goto local_send_completion;
2357                         }
2358
2359                         recv = 1;
2360                         /*
2361                          * Defined behavior is to complete response
2362                          * before request
2363                          */
2364                         build_smp_wc(recv_mad_agent->agent.qp,
2365                                      (unsigned long) local->mad_send_wr,
2366                                      be16_to_cpu(IB_LID_PERMISSIVE),
2367                                      0, recv_mad_agent->agent.port_num, &wc);
2368
2369                         local->mad_priv->header.recv_wc.wc = &wc;
2370                         local->mad_priv->header.recv_wc.mad_len =
2371                                                 sizeof(struct ib_mad);
2372                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2373                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2374                                  &local->mad_priv->header.recv_wc.rmpp_list);
2375                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2376                         local->mad_priv->header.recv_wc.recv_buf.mad =
2377                                                 &local->mad_priv->mad.mad;
2378                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2379                                 snoop_recv(recv_mad_agent->qp_info,
2380                                           &local->mad_priv->header.recv_wc,
2381                                            IB_MAD_SNOOP_RECVS);
2382                         recv_mad_agent->agent.recv_handler(
2383                                                 &recv_mad_agent->agent,
2384                                                 &local->mad_priv->header.recv_wc);
2385                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2386                         atomic_dec(&recv_mad_agent->refcount);
2387                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2388                 }
2389
2390 local_send_completion:
2391                 /* Complete send */
2392                 mad_send_wc.status = IB_WC_SUCCESS;
2393                 mad_send_wc.vendor_err = 0;
2394                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2395                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2396                         snoop_send(mad_agent_priv->qp_info,
2397                                    &local->mad_send_wr->send_buf,
2398                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2399                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2400                                                    &mad_send_wc);
2401
2402                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2403                 atomic_dec(&mad_agent_priv->refcount);
2404                 if (!recv)
2405                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2406                 kfree(local);
2407         }
2408         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2409 }
2410
2411 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2412 {
2413         int ret;
2414
2415         if (!mad_send_wr->retries--)
2416                 return -ETIMEDOUT;
2417
2418         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2419
2420         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2421                 ret = ib_retry_rmpp(mad_send_wr);
2422                 switch (ret) {
2423                 case IB_RMPP_RESULT_UNHANDLED:
2424                         ret = ib_send_mad(mad_send_wr);
2425                         break;
2426                 case IB_RMPP_RESULT_CONSUMED:
2427                         ret = 0;
2428                         break;
2429                 default:
2430                         ret = -ECOMM;
2431                         break;
2432                 }
2433         } else
2434                 ret = ib_send_mad(mad_send_wr);
2435
2436         if (!ret) {
2437                 mad_send_wr->refcount++;
2438                 list_add_tail(&mad_send_wr->agent_list,
2439                               &mad_send_wr->mad_agent_priv->send_list);
2440         }
2441         return ret;
2442 }
2443
2444 static void timeout_sends(struct work_struct *work)
2445 {
2446         struct ib_mad_agent_private *mad_agent_priv;
2447         struct ib_mad_send_wr_private *mad_send_wr;
2448         struct ib_mad_send_wc mad_send_wc;
2449         unsigned long flags, delay;
2450
2451         mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2452                                       timed_work.work);
2453         mad_send_wc.vendor_err = 0;
2454
2455         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2456         while (!list_empty(&mad_agent_priv->wait_list)) {
2457                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2458                                          struct ib_mad_send_wr_private,
2459                                          agent_list);
2460
2461                 if (time_after(mad_send_wr->timeout, jiffies)) {
2462                         delay = mad_send_wr->timeout - jiffies;
2463                         if ((long)delay <= 0)
2464                                 delay = 1;
2465                         queue_delayed_work(mad_agent_priv->qp_info->
2466                                            port_priv->wq,
2467                                            &mad_agent_priv->timed_work, delay);
2468                         break;
2469                 }
2470
2471                 list_del(&mad_send_wr->agent_list);
2472                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2473                     !retry_send(mad_send_wr))
2474                         continue;
2475
2476                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2477
2478                 if (mad_send_wr->status == IB_WC_SUCCESS)
2479                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2480                 else
2481                         mad_send_wc.status = mad_send_wr->status;
2482                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2483                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2484                                                    &mad_send_wc);
2485
2486                 atomic_dec(&mad_agent_priv->refcount);
2487                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2488         }
2489         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2490 }
2491
2492 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2493 {
2494         struct ib_mad_port_private *port_priv = cq->cq_context;
2495         unsigned long flags;
2496
2497         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2498         if (!list_empty(&port_priv->port_list))
2499                 queue_work(port_priv->wq, &port_priv->work);
2500         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2501 }
2502
2503 /*
2504  * Allocate receive MADs and post receive WRs for them
2505  */
2506 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2507                                     struct ib_mad_private *mad)
2508 {
2509         unsigned long flags;
2510         int post, ret;
2511         struct ib_mad_private *mad_priv;
2512         struct ib_sge sg_list;
2513         struct ib_recv_wr recv_wr, *bad_recv_wr;
2514         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2515
2516         /* Initialize common scatter list fields */
2517         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2518         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2519
2520         /* Initialize common receive WR fields */
2521         recv_wr.next = NULL;
2522         recv_wr.sg_list = &sg_list;
2523         recv_wr.num_sge = 1;
2524
2525         do {
2526                 /* Allocate and map receive buffer */
2527                 if (mad) {
2528                         mad_priv = mad;
2529                         mad = NULL;
2530                 } else {
2531                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2532                         if (!mad_priv) {
2533                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2534                                 ret = -ENOMEM;
2535                                 break;
2536                         }
2537                 }
2538                 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2539                                                  &mad_priv->grh,
2540                                                  sizeof *mad_priv -
2541                                                    sizeof mad_priv->header,
2542                                                  DMA_FROM_DEVICE);
2543                 mad_priv->header.mapping = sg_list.addr;
2544                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2545                 mad_priv->header.mad_list.mad_queue = recv_queue;
2546
2547                 /* Post receive WR */
2548                 spin_lock_irqsave(&recv_queue->lock, flags);
2549                 post = (++recv_queue->count < recv_queue->max_active);
2550                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2551                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2552                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2553                 if (ret) {
2554                         spin_lock_irqsave(&recv_queue->lock, flags);
2555                         list_del(&mad_priv->header.mad_list.list);
2556                         recv_queue->count--;
2557                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2558                         ib_dma_unmap_single(qp_info->port_priv->device,
2559                                             mad_priv->header.mapping,
2560                                             sizeof *mad_priv -
2561                                               sizeof mad_priv->header,
2562                                             DMA_FROM_DEVICE);
2563                         kmem_cache_free(ib_mad_cache, mad_priv);
2564                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2565                         break;
2566                 }
2567         } while (post);
2568
2569         return ret;
2570 }
2571
2572 /*
2573  * Return all the posted receive MADs
2574  */
2575 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2576 {
2577         struct ib_mad_private_header *mad_priv_hdr;
2578         struct ib_mad_private *recv;
2579         struct ib_mad_list_head *mad_list;
2580
2581         while (!list_empty(&qp_info->recv_queue.list)) {
2582
2583                 mad_list = list_entry(qp_info->recv_queue.list.next,
2584                                       struct ib_mad_list_head, list);
2585                 mad_priv_hdr = container_of(mad_list,
2586                                             struct ib_mad_private_header,
2587                                             mad_list);
2588                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2589                                     header);
2590
2591                 /* Remove from posted receive MAD list */
2592                 list_del(&mad_list->list);
2593
2594                 ib_dma_unmap_single(qp_info->port_priv->device,
2595                                     recv->header.mapping,
2596                                     sizeof(struct ib_mad_private) -
2597                                       sizeof(struct ib_mad_private_header),
2598                                     DMA_FROM_DEVICE);
2599                 kmem_cache_free(ib_mad_cache, recv);
2600         }
2601
2602         qp_info->recv_queue.count = 0;
2603 }
2604
2605 /*
2606  * Start the port
2607  */
2608 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2609 {
2610         int ret, i;
2611         struct ib_qp_attr *attr;
2612         struct ib_qp *qp;
2613
2614         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2615         if (!attr) {
2616                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2617                 return -ENOMEM;
2618         }
2619
2620         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2621                 qp = port_priv->qp_info[i].qp;
2622                 /*
2623                  * PKey index for QP1 is irrelevant but
2624                  * one is needed for the Reset to Init transition
2625                  */
2626                 attr->qp_state = IB_QPS_INIT;
2627                 attr->pkey_index = 0;
2628                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2629                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2630                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2631                 if (ret) {
2632                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2633                                "INIT: %d\n", i, ret);
2634                         goto out;
2635                 }
2636
2637                 attr->qp_state = IB_QPS_RTR;
2638                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2639                 if (ret) {
2640                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2641                                "RTR: %d\n", i, ret);
2642                         goto out;
2643                 }
2644
2645                 attr->qp_state = IB_QPS_RTS;
2646                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2647                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2648                 if (ret) {
2649                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2650                                "RTS: %d\n", i, ret);
2651                         goto out;
2652                 }
2653         }
2654
2655         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2656         if (ret) {
2657                 printk(KERN_ERR PFX "Failed to request completion "
2658                        "notification: %d\n", ret);
2659                 goto out;
2660         }
2661
2662         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2663                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2664                 if (ret) {
2665                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2666                         goto out;
2667                 }
2668         }
2669 out:
2670         kfree(attr);
2671         return ret;
2672 }
2673
2674 static void qp_event_handler(struct ib_event *event, void *qp_context)
2675 {
2676         struct ib_mad_qp_info   *qp_info = qp_context;
2677
2678         /* It's worse than that! He's dead, Jim! */
2679         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2680                 event->event, qp_info->qp->qp_num);
2681 }
2682
2683 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2684                            struct ib_mad_queue *mad_queue)
2685 {
2686         mad_queue->qp_info = qp_info;
2687         mad_queue->count = 0;
2688         spin_lock_init(&mad_queue->lock);
2689         INIT_LIST_HEAD(&mad_queue->list);
2690 }
2691
2692 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2693                         struct ib_mad_qp_info *qp_info)
2694 {
2695         qp_info->port_priv = port_priv;
2696         init_mad_queue(qp_info, &qp_info->send_queue);
2697         init_mad_queue(qp_info, &qp_info->recv_queue);
2698         INIT_LIST_HEAD(&qp_info->overflow_list);
2699         spin_lock_init(&qp_info->snoop_lock);
2700         qp_info->snoop_table = NULL;
2701         qp_info->snoop_table_size = 0;
2702         atomic_set(&qp_info->snoop_count, 0);
2703 }
2704
2705 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2706                          enum ib_qp_type qp_type)
2707 {
2708         struct ib_qp_init_attr  qp_init_attr;
2709         int ret;
2710
2711         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2712         qp_init_attr.send_cq = qp_info->port_priv->cq;
2713         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2714         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2715         qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2716         qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2717         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2718         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2719         qp_init_attr.qp_type = qp_type;
2720         qp_init_attr.port_num = qp_info->port_priv->port_num;
2721         qp_init_attr.qp_context = qp_info;
2722         qp_init_attr.event_handler = qp_event_handler;
2723         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2724         if (IS_ERR(qp_info->qp)) {
2725                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2726                        get_spl_qp_index(qp_type));
2727                 ret = PTR_ERR(qp_info->qp);
2728                 goto error;
2729         }
2730         /* Use minimum queue sizes unless the CQ is resized */
2731         qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2732         qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2733         return 0;
2734
2735 error:
2736         return ret;
2737 }
2738
2739 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2740 {
2741         ib_destroy_qp(qp_info->qp);
2742         kfree(qp_info->snoop_table);
2743 }
2744
2745 /*
2746  * Open the port
2747  * Create the QP, PD, MR, and CQ if needed
2748  */
2749 static int ib_mad_port_open(struct ib_device *device,
2750                             int port_num)
2751 {
2752         int ret, cq_size;
2753         struct ib_mad_port_private *port_priv;
2754         unsigned long flags;
2755         char name[sizeof "ib_mad123"];
2756
2757         /* Create new device info */
2758         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2759         if (!port_priv) {
2760                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2761                 return -ENOMEM;
2762         }
2763
2764         port_priv->device = device;
2765         port_priv->port_num = port_num;
2766         spin_lock_init(&port_priv->reg_lock);
2767         INIT_LIST_HEAD(&port_priv->agent_list);
2768         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2769         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2770
2771         cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2772         port_priv->cq = ib_create_cq(port_priv->device,
2773                                      ib_mad_thread_completion_handler,
2774                                      NULL, port_priv, cq_size, 0);
2775         if (IS_ERR(port_priv->cq)) {
2776                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2777                 ret = PTR_ERR(port_priv->cq);
2778                 goto error3;
2779         }
2780
2781         port_priv->pd = ib_alloc_pd(device);
2782         if (IS_ERR(port_priv->pd)) {
2783                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2784                 ret = PTR_ERR(port_priv->pd);
2785                 goto error4;
2786         }
2787
2788         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2789         if (IS_ERR(port_priv->mr)) {
2790                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2791                 ret = PTR_ERR(port_priv->mr);
2792                 goto error5;
2793         }
2794
2795         ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2796         if (ret)
2797                 goto error6;
2798         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2799         if (ret)
2800                 goto error7;
2801
2802         snprintf(name, sizeof name, "ib_mad%d", port_num);
2803         port_priv->wq = create_singlethread_workqueue(name);
2804         if (!port_priv->wq) {
2805                 ret = -ENOMEM;
2806                 goto error8;
2807         }
2808         INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2809
2810         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2811         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2812         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2813
2814         ret = ib_mad_port_start(port_priv);
2815         if (ret) {
2816                 printk(KERN_ERR PFX "Couldn't start port\n");
2817                 goto error9;
2818         }
2819
2820         return 0;
2821
2822 error9:
2823         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2824         list_del_init(&port_priv->port_list);
2825         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2826
2827         destroy_workqueue(port_priv->wq);
2828 error8:
2829         destroy_mad_qp(&port_priv->qp_info[1]);
2830 error7:
2831         destroy_mad_qp(&port_priv->qp_info[0]);
2832 error6:
2833         ib_dereg_mr(port_priv->mr);
2834 error5:
2835         ib_dealloc_pd(port_priv->pd);
2836 error4:
2837         ib_destroy_cq(port_priv->cq);
2838         cleanup_recv_queue(&port_priv->qp_info[1]);
2839         cleanup_recv_queue(&port_priv->qp_info[0]);
2840 error3:
2841         kfree(port_priv);
2842
2843         return ret;
2844 }
2845
2846 /*
2847  * Close the port
2848  * If there are no classes using the port, free the port
2849  * resources (CQ, MR, PD, QP) and remove the port's info structure
2850  */
2851 static int ib_mad_port_close(struct ib_device *device, int port_num)
2852 {
2853         struct ib_mad_port_private *port_priv;
2854         unsigned long flags;
2855
2856         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2857         port_priv = __ib_get_mad_port(device, port_num);
2858         if (port_priv == NULL) {
2859                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2860                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2861                 return -ENODEV;
2862         }
2863         list_del_init(&port_priv->port_list);
2864         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2865
2866         destroy_workqueue(port_priv->wq);
2867         destroy_mad_qp(&port_priv->qp_info[1]);
2868         destroy_mad_qp(&port_priv->qp_info[0]);
2869         ib_dereg_mr(port_priv->mr);
2870         ib_dealloc_pd(port_priv->pd);
2871         ib_destroy_cq(port_priv->cq);
2872         cleanup_recv_queue(&port_priv->qp_info[1]);
2873         cleanup_recv_queue(&port_priv->qp_info[0]);
2874         /* XXX: Handle deallocation of MAD registration tables */
2875
2876         kfree(port_priv);
2877
2878         return 0;
2879 }
2880
2881 static void ib_mad_init_device(struct ib_device *device)
2882 {
2883         int start, end, i;
2884
2885         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2886                 return;
2887
2888         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2889                 start = 0;
2890                 end   = 0;
2891         } else {
2892                 start = 1;
2893                 end   = device->phys_port_cnt;
2894         }
2895
2896         for (i = start; i <= end; i++) {
2897                 if (ib_mad_port_open(device, i)) {
2898                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2899                                device->name, i);
2900                         goto error;
2901                 }
2902                 if (ib_agent_port_open(device, i)) {
2903                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2904                                "for agents\n",
2905                                device->name, i);
2906                         goto error_agent;
2907                 }
2908         }
2909         return;
2910
2911 error_agent:
2912         if (ib_mad_port_close(device, i))
2913                 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2914                        device->name, i);
2915
2916 error:
2917         i--;
2918
2919         while (i >= start) {
2920                 if (ib_agent_port_close(device, i))
2921                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2922                                "for agents\n",
2923                                device->name, i);
2924                 if (ib_mad_port_close(device, i))
2925                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2926                                device->name, i);
2927                 i--;
2928         }
2929 }
2930
2931 static void ib_mad_remove_device(struct ib_device *device)
2932 {
2933         int i, num_ports, cur_port;
2934
2935         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2936                 num_ports = 1;
2937                 cur_port = 0;
2938         } else {
2939                 num_ports = device->phys_port_cnt;
2940                 cur_port = 1;
2941         }
2942         for (i = 0; i < num_ports; i++, cur_port++) {
2943                 if (ib_agent_port_close(device, cur_port))
2944                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2945                                "for agents\n",
2946                                device->name, cur_port);
2947                 if (ib_mad_port_close(device, cur_port))
2948                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2949                                device->name, cur_port);
2950         }
2951 }
2952
2953 static struct ib_client mad_client = {
2954         .name   = "mad",
2955         .add = ib_mad_init_device,
2956         .remove = ib_mad_remove_device
2957 };
2958
2959 static int __init ib_mad_init_module(void)
2960 {
2961         int ret;
2962
2963         spin_lock_init(&ib_mad_port_list_lock);
2964
2965         ib_mad_cache = kmem_cache_create("ib_mad",
2966                                          sizeof(struct ib_mad_private),
2967                                          0,
2968                                          SLAB_HWCACHE_ALIGN,
2969                                          NULL,
2970                                          NULL);
2971         if (!ib_mad_cache) {
2972                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2973                 ret = -ENOMEM;
2974                 goto error1;
2975         }
2976
2977         INIT_LIST_HEAD(&ib_mad_port_list);
2978
2979         if (ib_register_client(&mad_client)) {
2980                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2981                 ret = -EINVAL;
2982                 goto error2;
2983         }
2984
2985         return 0;
2986
2987 error2:
2988         kmem_cache_destroy(ib_mad_cache);
2989 error1:
2990         return ret;
2991 }
2992
2993 static void __exit ib_mad_cleanup_module(void)
2994 {
2995         ib_unregister_client(&mad_client);
2996         kmem_cache_destroy(ib_mad_cache);
2997 }
2998
2999 module_init(ib_mad_init_module);
3000 module_exit(ib_mad_cleanup_module);
3001