drivers/infiniband/core/mad.c

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