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