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

   1 /*
   2  * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved.
   3  *
   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 ib_send_mad(struct ib_mad_agent_private *mad_agent_priv,
 767                        struct ib_mad_send_wr_private *mad_send_wr)
 768 {
 769         struct ib_mad_qp_info *qp_info;
 770         struct ib_send_wr *bad_send_wr;
 771         unsigned long flags;
 772         int ret;
 773
 774         /* Replace user's WR ID with our own to find WR upon completion */
 775         qp_info = mad_agent_priv->qp_info;
 776         mad_send_wr->wr_id = mad_send_wr->send_wr.wr_id;
 777         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
 778         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
 779
 780         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
 781         if (qp_info->send_queue.count++ < qp_info->send_queue.max_active) {
 782                 list_add_tail(&mad_send_wr->mad_list.list,
 783                               &qp_info->send_queue.list);
 784                 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
 785                 ret = ib_post_send(mad_agent_priv->agent.qp,
 786                                    &mad_send_wr->send_wr, &bad_send_wr);
 787                 if (ret) {
 788                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
 789                         dequeue_mad(&mad_send_wr->mad_list);
 790                 }
 791         } else {
 792                 list_add_tail(&mad_send_wr->mad_list.list,
 793                               &qp_info->overflow_list);
 794                 spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
 795                 ret = 0;
 796         }
 797         return ret;
 798 }
 799
 800 /*
 801  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
 802  *  with the registered client
 803  */
 804 int ib_post_send_mad(struct ib_mad_agent *mad_agent,
 805                      struct ib_send_wr *send_wr,
 806                      struct ib_send_wr **bad_send_wr)
 807 {
 808         int ret = -EINVAL;
 809         struct ib_mad_agent_private *mad_agent_priv;
 810
 811         /* Validate supplied parameters */
 812         if (!bad_send_wr)
 813                 goto error1;
 814
 815         if (!mad_agent || !send_wr)
 816                 goto error2;
 817
 818         if (!mad_agent->send_handler)
 819                 goto error2;
 820
 821         mad_agent_priv = container_of(mad_agent,
 822                                       struct ib_mad_agent_private,
 823                                       agent);
 824
 825         /* Walk list of send WRs and post each on send list */
 826         while (send_wr) {
 827                 unsigned long                   flags;
 828                 struct ib_send_wr               *next_send_wr;
 829                 struct ib_mad_send_wr_private   *mad_send_wr;
 830                 struct ib_smp                   *smp;
 831
 832                 /* Validate more parameters */
 833                 if (send_wr->num_sge > IB_MAD_SEND_REQ_MAX_SG)
 834                         goto error2;
 835
 836                 if (send_wr->wr.ud.timeout_ms && !mad_agent->recv_handler)
 837                         goto error2;
 838
 839                 if (!send_wr->wr.ud.mad_hdr) {
 840                         printk(KERN_ERR PFX "MAD header must be supplied "
 841                                "in WR %p\n", send_wr);
 842                         goto error2;
 843                 }
 844
 845                 /*
 846                  * Save pointer to next work request to post in case the
 847                  * current one completes, and the user modifies the work
 848                  * request associated with the completion
 849                  */
 850                 next_send_wr = (struct ib_send_wr *)send_wr->next;
 851
 852                 smp = (struct ib_smp *)send_wr->wr.ud.mad_hdr;
 853                 if (smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 854                         ret = handle_outgoing_dr_smp(mad_agent_priv, smp,
 855                                                      send_wr);
 856                         if (ret < 0)            /* error */
 857                                 goto error2;
 858                         else if (ret == 1)      /* locally consumed */
 859                                 goto next;
 860                 }
 861
 862                 /* Allocate MAD send WR tracking structure */
 863                 mad_send_wr = kmalloc(sizeof *mad_send_wr, GFP_ATOMIC);
 864                 if (!mad_send_wr) {
 865                         printk(KERN_ERR PFX "No memory for "
 866                                "ib_mad_send_wr_private\n");
 867                         ret = -ENOMEM;
 868                         goto error2;
 869                 }
 870
 871                 mad_send_wr->send_wr = *send_wr;
 872                 mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 873                 memcpy(mad_send_wr->sg_list, send_wr->sg_list,
 874                        sizeof *send_wr->sg_list * send_wr->num_sge);
 875                 mad_send_wr->send_wr.next = NULL;
 876                 mad_send_wr->tid = send_wr->wr.ud.mad_hdr->tid;
 877                 mad_send_wr->agent = mad_agent;
 878                 /* Timeout will be updated after send completes */
 879                 mad_send_wr->timeout = msecs_to_jiffies(send_wr->wr.
 880                                                         ud.timeout_ms);
 881                 mad_send_wr->retry = 0;
 882                 /* One reference for each work request to QP + response */
 883                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
 884                 mad_send_wr->status = IB_WC_SUCCESS;
 885
 886                 /* Reference MAD agent until send completes */
 887                 atomic_inc(&mad_agent_priv->refcount);
 888                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
 889                 list_add_tail(&mad_send_wr->agent_list,
 890                               &mad_agent_priv->send_list);
 891                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 892
 893                 ret = ib_send_mad(mad_agent_priv, mad_send_wr);
 894                 if (ret) {
 895                         /* Fail send request */
 896                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 897                         list_del(&mad_send_wr->agent_list);
 898                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 899                         atomic_dec(&mad_agent_priv->refcount);
 900                         goto error2;
 901                 }
 902 next:
 903                 send_wr = next_send_wr;
 904         }
 905         return 0;
 906
 907 error2:
 908         *bad_send_wr = send_wr;
 909 error1:
 910         return ret;
 911 }
 912 EXPORT_SYMBOL(ib_post_send_mad);
 913
 914 /*
 915  * ib_free_recv_mad - Returns data buffers used to receive
 916  *  a MAD to the access layer
 917  */
 918 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
 919 {
 920         struct ib_mad_recv_buf *entry;
 921         struct ib_mad_private_header *mad_priv_hdr;
 922         struct ib_mad_private *priv;
 923
 924         mad_priv_hdr = container_of(mad_recv_wc,
 925                                     struct ib_mad_private_header,
 926                                     recv_wc);
 927         priv = container_of(mad_priv_hdr, struct ib_mad_private, header);
 928
 929         /*
 930          * Walk receive buffer list associated with this WC
 931          * No need to remove them from list of receive buffers
 932          */
 933         list_for_each_entry(entry, &mad_recv_wc->recv_buf.list, list) {
 934                 /* Free previous receive buffer */
 935                 kmem_cache_free(ib_mad_cache, priv);
 936                 mad_priv_hdr = container_of(mad_recv_wc,
 937                                             struct ib_mad_private_header,
 938                                             recv_wc);
 939                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
 940                                     header);
 941         }
 942
 943         /* Free last buffer */
 944         kmem_cache_free(ib_mad_cache, priv);
 945 }
 946 EXPORT_SYMBOL(ib_free_recv_mad);
 947
 948 void ib_coalesce_recv_mad(struct ib_mad_recv_wc *mad_recv_wc,
 949                           void *buf)
 950 {
 951         printk(KERN_ERR PFX "ib_coalesce_recv_mad() not implemented yet\n");
 952 }
 953 EXPORT_SYMBOL(ib_coalesce_recv_mad);
 954
 955 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
 956                                         u8 rmpp_version,
 957                                         ib_mad_send_handler send_handler,
 958                                         ib_mad_recv_handler recv_handler,
 959                                         void *context)
 960 {
 961         return ERR_PTR(-EINVAL);        /* XXX: for now */
 962 }
 963 EXPORT_SYMBOL(ib_redirect_mad_qp);
 964
 965 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
 966                       struct ib_wc *wc)
 967 {
 968         printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
 969         return 0;
 970 }
 971 EXPORT_SYMBOL(ib_process_mad_wc);
 972
 973 static int method_in_use(struct ib_mad_mgmt_method_table **method,
 974                          struct ib_mad_reg_req *mad_reg_req)
 975 {
 976         int i;
 977
 978         for (i = find_first_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS);
 979              i < IB_MGMT_MAX_METHODS;
 980              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
 981                                1+i)) {
 982                 if ((*method)->agent[i]) {
 983                         printk(KERN_ERR PFX "Method %d already in use\n", i);
 984                         return -EINVAL;
 985                 }
 986         }
 987         return 0;
 988 }
 989
 990 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
 991 {
 992         /* Allocate management method table */
 993         *method = kmalloc(sizeof **method, GFP_ATOMIC);
 994         if (!*method) {
 995                 printk(KERN_ERR PFX "No memory for "
 996                        "ib_mad_mgmt_method_table\n");
 997                 return -ENOMEM;
 998         }
 999         /* Clear management method table */
1000         memset(*method, 0, sizeof **method);
1001
1002         return 0;
1003 }
1004
1005 /*
1006  * Check to see if there are any methods still in use
1007  */
1008 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1009 {
1010         int i;
1011
1012         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1013                 if (method->agent[i])
1014                         return 1;
1015         return 0;
1016 }
1017
1018 /*
1019  * Check to see if there are any method tables for this class still in use
1020  */
1021 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1022 {
1023         int i;
1024
1025         for (i = 0; i < MAX_MGMT_CLASS; i++)
1026                 if (class->method_table[i])
1027                         return 1;
1028         return 0;
1029 }
1030
1031 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1032 {
1033         int i;
1034
1035         for (i = 0; i < MAX_MGMT_OUI; i++)
1036                 if (vendor_class->method_table[i])
1037                         return 1;
1038         return 0;
1039 }
1040
1041 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1042                            char *oui)
1043 {
1044         int i;
1045
1046         for (i = 0; i < MAX_MGMT_OUI; i++)
1047                 /* Is there matching OUI for this vendor class ? */
1048                 if (!memcmp(vendor_class->oui[i], oui, 3))
1049                         return i;
1050
1051         return -1;
1052 }
1053
1054 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1055 {
1056         int i;
1057
1058         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1059                 if (vendor->vendor_class[i])
1060                         return 1;
1061
1062         return 0;
1063 }
1064
1065 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1066                                      struct ib_mad_agent_private *agent)
1067 {
1068         int i;
1069
1070         /* Remove any methods for this mad agent */
1071         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1072                 if (method->agent[i] == agent) {
1073                         method->agent[i] = NULL;
1074                 }
1075         }
1076 }
1077
1078 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1079                               struct ib_mad_agent_private *agent_priv,
1080                               u8 mgmt_class)
1081 {
1082         struct ib_mad_port_private *port_priv;
1083         struct ib_mad_mgmt_class_table **class;
1084         struct ib_mad_mgmt_method_table **method;
1085         int i, ret;
1086
1087         port_priv = agent_priv->qp_info->port_priv;
1088         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1089         if (!*class) {
1090                 /* Allocate management class table for "new" class version */
1091                 *class = kmalloc(sizeof **class, GFP_ATOMIC);
1092                 if (!*class) {
1093                         printk(KERN_ERR PFX "No memory for "
1094                                "ib_mad_mgmt_class_table\n");
1095                         ret = -ENOMEM;
1096                         goto error1;
1097                 }
1098                 /* Clear management class table */
1099                 memset(*class, 0, sizeof(**class));
1100                 /* Allocate method table for this management class */
1101                 method = &(*class)->method_table[mgmt_class];
1102                 if ((ret = allocate_method_table(method)))
1103                         goto error2;
1104         } else {
1105                 method = &(*class)->method_table[mgmt_class];
1106                 if (!*method) {
1107                         /* Allocate method table for this management class */
1108                         if ((ret = allocate_method_table(method)))
1109                                 goto error1;
1110                 }
1111         }
1112
1113         /* Now, make sure methods are not already in use */
1114         if (method_in_use(method, mad_reg_req))
1115                 goto error3;
1116
1117         /* Finally, add in methods being registered */
1118         for (i = find_first_bit(mad_reg_req->method_mask,
1119                                 IB_MGMT_MAX_METHODS);
1120              i < IB_MGMT_MAX_METHODS;
1121              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1122                                1+i)) {
1123                 (*method)->agent[i] = agent_priv;
1124         }
1125         return 0;
1126
1127 error3:
1128         /* Remove any methods for this mad agent */
1129         remove_methods_mad_agent(*method, agent_priv);
1130         /* Now, check to see if there are any methods in use */
1131         if (!check_method_table(*method)) {
1132                 /* If not, release management method table */
1133                 kfree(*method);
1134                 *method = NULL;
1135         }
1136         ret = -EINVAL;
1137         goto error1;
1138 error2:
1139         kfree(*class);
1140         *class = NULL;
1141 error1:
1142         return ret;
1143 }
1144
1145 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1146                            struct ib_mad_agent_private *agent_priv)
1147 {
1148         struct ib_mad_port_private *port_priv;
1149         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1150         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1151         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1152         struct ib_mad_mgmt_method_table **method;
1153         int i, ret = -ENOMEM;
1154         u8 vclass;
1155
1156         /* "New" vendor (with OUI) class */
1157         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1158         port_priv = agent_priv->qp_info->port_priv;
1159         vendor_table = &port_priv->version[
1160                                 mad_reg_req->mgmt_class_version].vendor;
1161         if (!*vendor_table) {
1162                 /* Allocate mgmt vendor class table for "new" class version */
1163                 vendor = kmalloc(sizeof *vendor, GFP_ATOMIC);
1164                 if (!vendor) {
1165                         printk(KERN_ERR PFX "No memory for "
1166                                "ib_mad_mgmt_vendor_class_table\n");
1167                         goto error1;
1168                 }
1169                 /* Clear management vendor class table */
1170                 memset(vendor, 0, sizeof(*vendor));
1171                 *vendor_table = vendor;
1172         }
1173         if (!(*vendor_table)->vendor_class[vclass]) {
1174                 /* Allocate table for this management vendor class */
1175                 vendor_class = kmalloc(sizeof *vendor_class, GFP_ATOMIC);
1176                 if (!vendor_class) {
1177                         printk(KERN_ERR PFX "No memory for "
1178                                "ib_mad_mgmt_vendor_class\n");
1179                         goto error2;
1180                 }
1181                 memset(vendor_class, 0, sizeof(*vendor_class));
1182                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1183         }
1184         for (i = 0; i < MAX_MGMT_OUI; i++) {
1185                 /* Is there matching OUI for this vendor class ? */
1186                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1187                             mad_reg_req->oui, 3)) {
1188                         method = &(*vendor_table)->vendor_class[
1189                                                 vclass]->method_table[i];
1190                         BUG_ON(!*method);
1191                         goto check_in_use;
1192                 }
1193         }
1194         for (i = 0; i < MAX_MGMT_OUI; i++) {
1195                 /* OUI slot available ? */
1196                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1197                                 vclass]->oui[i])) {
1198                         method = &(*vendor_table)->vendor_class[
1199                                 vclass]->method_table[i];
1200                         BUG_ON(*method);
1201                         /* Allocate method table for this OUI */
1202                         if ((ret = allocate_method_table(method)))
1203                                 goto error3;
1204                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1205                                mad_reg_req->oui, 3);
1206                         goto check_in_use;
1207                 }
1208         }
1209         printk(KERN_ERR PFX "All OUI slots in use\n");
1210         goto error3;
1211
1212 check_in_use:
1213         /* Now, make sure methods are not already in use */
1214         if (method_in_use(method, mad_reg_req))
1215                 goto error4;
1216
1217         /* Finally, add in methods being registered */
1218         for (i = find_first_bit(mad_reg_req->method_mask,
1219                                 IB_MGMT_MAX_METHODS);
1220              i < IB_MGMT_MAX_METHODS;
1221              i = find_next_bit(mad_reg_req->method_mask, IB_MGMT_MAX_METHODS,
1222                                1+i)) {
1223                 (*method)->agent[i] = agent_priv;
1224         }
1225         return 0;
1226
1227 error4:
1228         /* Remove any methods for this mad agent */
1229         remove_methods_mad_agent(*method, agent_priv);
1230         /* Now, check to see if there are any methods in use */
1231         if (!check_method_table(*method)) {
1232                 /* If not, release management method table */
1233                 kfree(*method);
1234                 *method = NULL;
1235         }
1236         ret = -EINVAL;
1237 error3:
1238         if (vendor_class) {
1239                 (*vendor_table)->vendor_class[vclass] = NULL;
1240                 kfree(vendor_class);
1241         }
1242 error2:
1243         if (vendor) {
1244                 *vendor_table = NULL;
1245                 kfree(vendor);
1246         }
1247 error1:
1248         return ret;
1249 }
1250
1251 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1252 {
1253         struct ib_mad_port_private *port_priv;
1254         struct ib_mad_mgmt_class_table *class;
1255         struct ib_mad_mgmt_method_table *method;
1256         struct ib_mad_mgmt_vendor_class_table *vendor;
1257         struct ib_mad_mgmt_vendor_class *vendor_class;
1258         int index;
1259         u8 mgmt_class;
1260
1261         /*
1262          * Was MAD registration request supplied
1263          * with original registration ?
1264          */
1265         if (!agent_priv->reg_req) {
1266                 goto out;
1267         }
1268
1269         port_priv = agent_priv->qp_info->port_priv;
1270         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1271         class = port_priv->version[
1272                         agent_priv->reg_req->mgmt_class_version].class;
1273         if (!class)
1274                 goto vendor_check;
1275
1276         method = class->method_table[mgmt_class];
1277         if (method) {
1278                 /* Remove any methods for this mad agent */
1279                 remove_methods_mad_agent(method, agent_priv);
1280                 /* Now, check to see if there are any methods still in use */
1281                 if (!check_method_table(method)) {
1282                         /* If not, release management method table */
1283                          kfree(method);
1284                          class->method_table[mgmt_class] = NULL;
1285                          /* Any management classes left ? */
1286                         if (!check_class_table(class)) {
1287                                 /* If not, release management class table */
1288                                 kfree(class);
1289                                 port_priv->version[
1290                                         agent_priv->reg_req->
1291                                         mgmt_class_version].class = NULL;
1292                         }
1293                 }
1294         }
1295
1296 vendor_check:
1297         if (!is_vendor_class(mgmt_class))
1298                 goto out;
1299
1300         /* normalize mgmt_class to vendor range 2 */
1301         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1302         vendor = port_priv->version[
1303                         agent_priv->reg_req->mgmt_class_version].vendor;
1304
1305         if (!vendor)
1306                 goto out;
1307
1308         vendor_class = vendor->vendor_class[mgmt_class];
1309         if (vendor_class) {
1310                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1311                 if (index < 0)
1312                         goto out;
1313                 method = vendor_class->method_table[index];
1314                 if (method) {
1315                         /* Remove any methods for this mad agent */
1316                         remove_methods_mad_agent(method, agent_priv);
1317                         /*
1318                          * Now, check to see if there are
1319                          * any methods still in use
1320                          */
1321                         if (!check_method_table(method)) {
1322                                 /* If not, release management method table */
1323                                 kfree(method);
1324                                 vendor_class->method_table[index] = NULL;
1325                                 memset(vendor_class->oui[index], 0, 3);
1326                                 /* Any OUIs left ? */
1327                                 if (!check_vendor_class(vendor_class)) {
1328                                         /* If not, release vendor class table */
1329                                         kfree(vendor_class);
1330                                         vendor->vendor_class[mgmt_class] = NULL;
1331                                         /* Any other vendor classes left ? */
1332                                         if (!check_vendor_table(vendor)) {
1333                                                 kfree(vendor);
1334                                                 port_priv->version[
1335                                                         agent_priv->reg_req->
1336                                                         mgmt_class_version].
1337                                                         vendor = NULL;
1338                                         }
1339                                 }
1340                         }
1341                 }
1342         }
1343
1344 out:
1345         return;
1346 }
1347
1348 static int response_mad(struct ib_mad *mad)
1349 {
1350         /* Trap represses are responses although response bit is reset */
1351         return ((mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
1352                 (mad->mad_hdr.method & IB_MGMT_METHOD_RESP));
1353 }
1354
1355 static int solicited_mad(struct ib_mad *mad)
1356 {
1357         /* CM MADs are never solicited */
1358         if (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_CM) {
1359                 return 0;
1360         }
1361
1362         /* XXX: Determine whether MAD is using RMPP */
1363
1364         /* Not using RMPP */
1365         /* Is this MAD a response to a previous MAD ? */
1366         return response_mad(mad);
1367 }
1368
1369 static struct ib_mad_agent_private *
1370 find_mad_agent(struct ib_mad_port_private *port_priv,
1371                struct ib_mad *mad,
1372                int solicited)
1373 {
1374         struct ib_mad_agent_private *mad_agent = NULL;
1375         unsigned long flags;
1376
1377         spin_lock_irqsave(&port_priv->reg_lock, flags);
1378
1379         /*
1380          * Whether MAD was solicited determines type of routing to
1381          * MAD client.
1382          */
1383         if (solicited) {
1384                 u32 hi_tid;
1385                 struct ib_mad_agent_private *entry;
1386
1387                 /*
1388                  * Routing is based on high 32 bits of transaction ID
1389                  * of MAD.
1390                  */
1391                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1392                 list_for_each_entry(entry, &port_priv->agent_list,
1393                                     agent_list) {
1394                         if (entry->agent.hi_tid == hi_tid) {
1395                                 mad_agent = entry;
1396                                 break;
1397                         }
1398                 }
1399         } else {
1400                 struct ib_mad_mgmt_class_table *class;
1401                 struct ib_mad_mgmt_method_table *method;
1402                 struct ib_mad_mgmt_vendor_class_table *vendor;
1403                 struct ib_mad_mgmt_vendor_class *vendor_class;
1404                 struct ib_vendor_mad *vendor_mad;
1405                 int index;
1406
1407                 /*
1408                  * Routing is based on version, class, and method
1409                  * For "newer" vendor MADs, also based on OUI
1410                  */
1411                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1412                         goto out;
1413                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1414                         class = port_priv->version[
1415                                         mad->mad_hdr.class_version].class;
1416                         if (!class)
1417                                 goto out;
1418                         method = class->method_table[convert_mgmt_class(
1419                                                         mad->mad_hdr.mgmt_class)];
1420                         if (method)
1421                                 mad_agent = method->agent[mad->mad_hdr.method &
1422                                                           ~IB_MGMT_METHOD_RESP];
1423                 } else {
1424                         vendor = port_priv->version[
1425                                         mad->mad_hdr.class_version].vendor;
1426                         if (!vendor)
1427                                 goto out;
1428                         vendor_class = vendor->vendor_class[vendor_class_index(
1429                                                 mad->mad_hdr.mgmt_class)];
1430                         if (!vendor_class)
1431                                 goto out;
1432                         /* Find matching OUI */
1433                         vendor_mad = (struct ib_vendor_mad *)mad;
1434                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1435                         if (index == -1)
1436                                 goto out;
1437                         method = vendor_class->method_table[index];
1438                         if (method) {
1439                                 mad_agent = method->agent[mad->mad_hdr.method &
1440                                                           ~IB_MGMT_METHOD_RESP];
1441                         }
1442                 }
1443         }
1444
1445         if (mad_agent) {
1446                 if (mad_agent->agent.recv_handler)
1447                         atomic_inc(&mad_agent->refcount);
1448                 else {
1449                         printk(KERN_NOTICE PFX "No receive handler for client "
1450                                "%p on port %d\n",
1451                                &mad_agent->agent, port_priv->port_num);
1452                         mad_agent = NULL;
1453                 }
1454         }
1455 out:
1456         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1457
1458         return mad_agent;
1459 }
1460
1461 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1462 {
1463         int valid = 0;
1464
1465         /* Make sure MAD base version is understood */
1466         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1467                 printk(KERN_ERR PFX "MAD received with unsupported base "
1468                        "version %d\n", mad->mad_hdr.base_version);
1469                 goto out;
1470         }
1471
1472         /* Filter SMI packets sent to other than QP0 */
1473         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1474             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1475                 if (qp_num == 0)
1476                         valid = 1;
1477         } else {
1478                 /* Filter GSI packets sent to QP0 */
1479                 if (qp_num != 0)
1480                         valid = 1;
1481         }
1482
1483 out:
1484         return valid;
1485 }
1486
1487 /*
1488  * Return start of fully reassembled MAD, or NULL, if MAD isn't assembled yet
1489  */
1490 static struct ib_mad_private *
1491 reassemble_recv(struct ib_mad_agent_private *mad_agent_priv,
1492                 struct ib_mad_private *recv)
1493 {
1494         /* Until we have RMPP, all receives are reassembled!... */
1495         INIT_LIST_HEAD(&recv->header.recv_wc.recv_buf.list);
1496         return recv;
1497 }
1498
1499 static struct ib_mad_send_wr_private*
1500 find_send_req(struct ib_mad_agent_private *mad_agent_priv,
1501               u64 tid)
1502 {
1503         struct ib_mad_send_wr_private *mad_send_wr;
1504
1505         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
1506                             agent_list) {
1507                 if (mad_send_wr->tid == tid)
1508                         return mad_send_wr;
1509         }
1510
1511         /*
1512          * It's possible to receive the response before we've
1513          * been notified that the send has completed
1514          */
1515         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
1516                             agent_list) {
1517                 if (mad_send_wr->tid == tid && mad_send_wr->timeout) {
1518                         /* Verify request has not been canceled */
1519                         return (mad_send_wr->status == IB_WC_SUCCESS) ?
1520                                 mad_send_wr : NULL;
1521                 }
1522         }
1523         return NULL;
1524 }
1525
1526 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1527                                  struct ib_mad_private *recv,
1528                                  int solicited)
1529 {
1530         struct ib_mad_send_wr_private *mad_send_wr;
1531         struct ib_mad_send_wc mad_send_wc;
1532         unsigned long flags;
1533
1534         /* Fully reassemble receive before processing */
1535         recv = reassemble_recv(mad_agent_priv, recv);
1536         if (!recv) {
1537                 if (atomic_dec_and_test(&mad_agent_priv->refcount))
1538                         wake_up(&mad_agent_priv->wait);
1539                 return;
1540         }
1541
1542         /* Complete corresponding request */
1543         if (solicited) {
1544                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1545                 mad_send_wr = find_send_req(mad_agent_priv,
1546                                             recv->mad.mad.mad_hdr.tid);
1547                 if (!mad_send_wr) {
1548                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1549                         ib_free_recv_mad(&recv->header.recv_wc);
1550                         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1551                                 wake_up(&mad_agent_priv->wait);
1552                         return;
1553                 }
1554                 /* Timeout = 0 means that we won't wait for a response */
1555                 mad_send_wr->timeout = 0;
1556                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1557
1558                 /* Defined behavior is to complete response before request */
1559                 recv->header.recv_wc.wc->wr_id = mad_send_wr->wr_id;
1560                 mad_agent_priv->agent.recv_handler(
1561                                                 &mad_agent_priv->agent,
1562                                                 &recv->header.recv_wc);
1563                 atomic_dec(&mad_agent_priv->refcount);
1564
1565                 mad_send_wc.status = IB_WC_SUCCESS;
1566                 mad_send_wc.vendor_err = 0;
1567                 mad_send_wc.wr_id = mad_send_wr->wr_id;
1568                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1569         } else {
1570                 mad_agent_priv->agent.recv_handler(
1571                                                 &mad_agent_priv->agent,
1572                                                 &recv->header.recv_wc);
1573                 if (atomic_dec_and_test(&mad_agent_priv->refcount))
1574                         wake_up(&mad_agent_priv->wait);
1575         }
1576 }
1577
1578 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1579                                      struct ib_wc *wc)
1580 {
1581         struct ib_mad_qp_info *qp_info;
1582         struct ib_mad_private_header *mad_priv_hdr;
1583         struct ib_mad_private *recv, *response;
1584         struct ib_mad_list_head *mad_list;
1585         struct ib_mad_agent_private *mad_agent;
1586         int solicited;
1587
1588         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1589         if (!response)
1590                 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1591                        "for response buffer\n");
1592
1593         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1594         qp_info = mad_list->mad_queue->qp_info;
1595         dequeue_mad(mad_list);
1596
1597         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1598                                     mad_list);
1599         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1600         dma_unmap_single(port_priv->device->dma_device,
1601                          pci_unmap_addr(&recv->header, mapping),
1602                          sizeof(struct ib_mad_private) -
1603                          sizeof(struct ib_mad_private_header),
1604                          DMA_FROM_DEVICE);
1605
1606         /* Setup MAD receive work completion from "normal" work completion */
1607         recv->header.wc = *wc;
1608         recv->header.recv_wc.wc = &recv->header.wc;
1609         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1610         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1611         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1612
1613         if (atomic_read(&qp_info->snoop_count))
1614                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1615
1616         /* Validate MAD */
1617         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1618                 goto out;
1619
1620         if (recv->mad.mad.mad_hdr.mgmt_class ==
1621             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1622                 if (!smi_handle_dr_smp_recv(&recv->mad.smp,
1623                                             port_priv->device->node_type,
1624                                             port_priv->port_num,
1625                                             port_priv->device->phys_port_cnt))
1626                         goto out;
1627                 if (!smi_check_forward_dr_smp(&recv->mad.smp))
1628                         goto local;
1629                 if (!smi_handle_dr_smp_send(&recv->mad.smp,
1630                                             port_priv->device->node_type,
1631                                             port_priv->port_num))
1632                         goto out;
1633                 if (!smi_check_local_dr_smp(&recv->mad.smp,
1634                                             port_priv->device,
1635                                             port_priv->port_num))
1636                         goto out;
1637         }
1638
1639 local:
1640         /* Give driver "right of first refusal" on incoming MAD */
1641         if (port_priv->device->process_mad) {
1642                 int ret;
1643
1644                 if (!response) {
1645                         printk(KERN_ERR PFX "No memory for response MAD\n");
1646                         /*
1647                          * Is it better to assume that
1648                          * it wouldn't be processed ?
1649                          */
1650                         goto out;
1651                 }
1652
1653                 ret = port_priv->device->process_mad(port_priv->device, 0,
1654                                                      port_priv->port_num,
1655                                                      wc, &recv->grh,
1656                                                      &recv->mad.mad,
1657                                                      &response->mad.mad);
1658                 if (ret & IB_MAD_RESULT_SUCCESS) {
1659                         if (ret & IB_MAD_RESULT_CONSUMED)
1660                                 goto out;
1661                         if (ret & IB_MAD_RESULT_REPLY) {
1662                                 /* Send response */
1663                                 if (!agent_send(response, &recv->grh, wc,
1664                                                 port_priv->device,
1665                                                 port_priv->port_num))
1666                                         response = NULL;
1667                                 goto out;
1668                         }
1669                 }
1670         }
1671
1672         /* Determine corresponding MAD agent for incoming receive MAD */
1673         solicited = solicited_mad(&recv->mad.mad);
1674         mad_agent = find_mad_agent(port_priv, &recv->mad.mad, solicited);
1675         if (mad_agent) {
1676                 ib_mad_complete_recv(mad_agent, recv, solicited);
1677                 /*
1678                  * recv is freed up in error cases in ib_mad_complete_recv
1679                  * or via recv_handler in ib_mad_complete_recv()
1680                  */
1681                 recv = NULL;
1682         }
1683
1684 out:
1685         /* Post another receive request for this QP */
1686         if (response) {
1687                 ib_mad_post_receive_mads(qp_info, response);
1688                 if (recv)
1689                         kmem_cache_free(ib_mad_cache, recv);
1690         } else
1691                 ib_mad_post_receive_mads(qp_info, recv);
1692 }
1693
1694 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1695 {
1696         struct ib_mad_send_wr_private *mad_send_wr;
1697         unsigned long delay;
1698
1699         if (list_empty(&mad_agent_priv->wait_list)) {
1700                 cancel_delayed_work(&mad_agent_priv->timed_work);
1701         } else {
1702                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1703                                          struct ib_mad_send_wr_private,
1704                                          agent_list);
1705
1706                 if (time_after(mad_agent_priv->timeout,
1707                                mad_send_wr->timeout)) {
1708                         mad_agent_priv->timeout = mad_send_wr->timeout;
1709                         cancel_delayed_work(&mad_agent_priv->timed_work);
1710                         delay = mad_send_wr->timeout - jiffies;
1711                         if ((long)delay <= 0)
1712                                 delay = 1;
1713                         queue_delayed_work(mad_agent_priv->qp_info->
1714                                            port_priv->wq,
1715                                            &mad_agent_priv->timed_work, delay);
1716                 }
1717         }
1718 }
1719
1720 static void wait_for_response(struct ib_mad_agent_private *mad_agent_priv,
1721                               struct ib_mad_send_wr_private *mad_send_wr )
1722 {
1723         struct ib_mad_send_wr_private *temp_mad_send_wr;
1724         struct list_head *list_item;
1725         unsigned long delay;
1726
1727         list_del(&mad_send_wr->agent_list);
1728
1729         delay = mad_send_wr->timeout;
1730         mad_send_wr->timeout += jiffies;
1731
1732         list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
1733                 temp_mad_send_wr = list_entry(list_item,
1734                                               struct ib_mad_send_wr_private,
1735                                               agent_list);
1736                 if (time_after(mad_send_wr->timeout,
1737                                temp_mad_send_wr->timeout))
1738                         break;
1739         }
1740         list_add(&mad_send_wr->agent_list, list_item);
1741
1742         /* Reschedule a work item if we have a shorter timeout */
1743         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
1744                 cancel_delayed_work(&mad_agent_priv->timed_work);
1745                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
1746                                    &mad_agent_priv->timed_work, delay);
1747         }
1748 }
1749
1750 /*
1751  * Process a send work completion
1752  */
1753 static void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
1754                                     struct ib_mad_send_wc *mad_send_wc)
1755 {
1756         struct ib_mad_agent_private     *mad_agent_priv;
1757         unsigned long                   flags;
1758
1759         mad_agent_priv = container_of(mad_send_wr->agent,
1760                                       struct ib_mad_agent_private, agent);
1761
1762         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1763         if (mad_send_wc->status != IB_WC_SUCCESS &&
1764             mad_send_wr->status == IB_WC_SUCCESS) {
1765                 mad_send_wr->status = mad_send_wc->status;
1766                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
1767         }
1768
1769         if (--mad_send_wr->refcount > 0) {
1770                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
1771                     mad_send_wr->status == IB_WC_SUCCESS) {
1772                         wait_for_response(mad_agent_priv, mad_send_wr);
1773                 }
1774                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1775                 return;
1776         }
1777
1778         /* Remove send from MAD agent and notify client of completion */
1779         list_del(&mad_send_wr->agent_list);
1780         adjust_timeout(mad_agent_priv);
1781         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1782
1783         if (mad_send_wr->status != IB_WC_SUCCESS )
1784                 mad_send_wc->status = mad_send_wr->status;
1785         mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
1786                                             mad_send_wc);
1787
1788         /* Release reference on agent taken when sending */
1789         if (atomic_dec_and_test(&mad_agent_priv->refcount))
1790                 wake_up(&mad_agent_priv->wait);
1791
1792         kfree(mad_send_wr);
1793 }
1794
1795 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
1796                                      struct ib_wc *wc)
1797 {
1798         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
1799         struct ib_mad_list_head         *mad_list;
1800         struct ib_mad_qp_info           *qp_info;
1801         struct ib_mad_queue             *send_queue;
1802         struct ib_send_wr               *bad_send_wr;
1803         unsigned long flags;
1804         int ret;
1805
1806         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1807         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1808                                    mad_list);
1809         send_queue = mad_list->mad_queue;
1810         qp_info = send_queue->qp_info;
1811
1812 retry:
1813         queued_send_wr = NULL;
1814         spin_lock_irqsave(&send_queue->lock, flags);
1815         list_del(&mad_list->list);
1816
1817         /* Move queued send to the send queue */
1818         if (send_queue->count-- > send_queue->max_active) {
1819                 mad_list = container_of(qp_info->overflow_list.next,
1820                                         struct ib_mad_list_head, list);
1821                 queued_send_wr = container_of(mad_list,
1822                                         struct ib_mad_send_wr_private,
1823                                         mad_list);
1824                 list_del(&mad_list->list);
1825                 list_add_tail(&mad_list->list, &send_queue->list);
1826         }
1827         spin_unlock_irqrestore(&send_queue->lock, flags);
1828
1829         /* Restore client wr_id in WC and complete send */
1830         wc->wr_id = mad_send_wr->wr_id;
1831         if (atomic_read(&qp_info->snoop_count))
1832                 snoop_send(qp_info, &mad_send_wr->send_wr,
1833                            (struct ib_mad_send_wc *)wc,
1834                            IB_MAD_SNOOP_SEND_COMPLETIONS);
1835         ib_mad_complete_send_wr(mad_send_wr, (struct ib_mad_send_wc *)wc);
1836
1837         if (queued_send_wr) {
1838                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
1839                                 &bad_send_wr);
1840                 if (ret) {
1841                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
1842                         mad_send_wr = queued_send_wr;
1843                         wc->status = IB_WC_LOC_QP_OP_ERR;
1844                         goto retry;
1845                 }
1846         }
1847 }
1848
1849 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
1850 {
1851         struct ib_mad_send_wr_private *mad_send_wr;
1852         struct ib_mad_list_head *mad_list;
1853         unsigned long flags;
1854
1855         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1856         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
1857                 mad_send_wr = container_of(mad_list,
1858                                            struct ib_mad_send_wr_private,
1859                                            mad_list);
1860                 mad_send_wr->retry = 1;
1861         }
1862         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1863 }
1864
1865 static void mad_error_handler(struct ib_mad_port_private *port_priv,
1866                               struct ib_wc *wc)
1867 {
1868         struct ib_mad_list_head *mad_list;
1869         struct ib_mad_qp_info *qp_info;
1870         struct ib_mad_send_wr_private *mad_send_wr;
1871         int ret;
1872
1873         /* Determine if failure was a send or receive */
1874         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1875         qp_info = mad_list->mad_queue->qp_info;
1876         if (mad_list->mad_queue == &qp_info->recv_queue)
1877                 /*
1878                  * Receive errors indicate that the QP has entered the error
1879                  * state - error handling/shutdown code will cleanup
1880                  */
1881                 return;
1882
1883         /*
1884          * Send errors will transition the QP to SQE - move
1885          * QP to RTS and repost flushed work requests
1886          */
1887         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
1888                                    mad_list);
1889         if (wc->status == IB_WC_WR_FLUSH_ERR) {
1890                 if (mad_send_wr->retry) {
1891                         /* Repost send */
1892                         struct ib_send_wr *bad_send_wr;
1893
1894                         mad_send_wr->retry = 0;
1895                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
1896                                         &bad_send_wr);
1897                         if (ret)
1898                                 ib_mad_send_done_handler(port_priv, wc);
1899                 } else
1900                         ib_mad_send_done_handler(port_priv, wc);
1901         } else {
1902                 struct ib_qp_attr *attr;
1903
1904                 /* Transition QP to RTS and fail offending send */
1905                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
1906                 if (attr) {
1907                         attr->qp_state = IB_QPS_RTS;
1908                         attr->cur_qp_state = IB_QPS_SQE;
1909                         ret = ib_modify_qp(qp_info->qp, attr,
1910                                            IB_QP_STATE | IB_QP_CUR_STATE);
1911                         kfree(attr);
1912                         if (ret)
1913                                 printk(KERN_ERR PFX "mad_error_handler - "
1914                                        "ib_modify_qp to RTS : %d\n", ret);
1915                         else
1916                                 mark_sends_for_retry(qp_info);
1917                 }
1918                 ib_mad_send_done_handler(port_priv, wc);
1919         }
1920 }
1921
1922 /*
1923  * IB MAD completion callback
1924  */
1925 static void ib_mad_completion_handler(void *data)
1926 {
1927         struct ib_mad_port_private *port_priv;
1928         struct ib_wc wc;
1929
1930         port_priv = (struct ib_mad_port_private *)data;
1931         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
1932
1933         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
1934                 if (wc.status == IB_WC_SUCCESS) {
1935                         switch (wc.opcode) {
1936                         case IB_WC_SEND:
1937                                 ib_mad_send_done_handler(port_priv, &wc);
1938                                 break;
1939                         case IB_WC_RECV:
1940                                 ib_mad_recv_done_handler(port_priv, &wc);
1941                                 break;
1942                         default:
1943                                 BUG_ON(1);
1944                                 break;
1945                         }
1946                 } else
1947                         mad_error_handler(port_priv, &wc);
1948         }
1949 }
1950
1951 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
1952 {
1953         unsigned long flags;
1954         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
1955         struct ib_mad_send_wc mad_send_wc;
1956         struct list_head cancel_list;
1957
1958         INIT_LIST_HEAD(&cancel_list);
1959
1960         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1961         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
1962                                  &mad_agent_priv->send_list, agent_list) {
1963                 if (mad_send_wr->status == IB_WC_SUCCESS) {
1964                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
1965                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
1966                 }
1967         }
1968
1969         /* Empty wait list to prevent receives from finding a request */
1970         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
1971         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1972
1973         /* Report all cancelled requests */
1974         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
1975         mad_send_wc.vendor_err = 0;
1976
1977         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
1978                                  &cancel_list, agent_list) {
1979                 mad_send_wc.wr_id = mad_send_wr->wr_id;
1980                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
1981                                                    &mad_send_wc);
1982
1983                 list_del(&mad_send_wr->agent_list);
1984                 kfree(mad_send_wr);
1985                 atomic_dec(&mad_agent_priv->refcount);
1986         }
1987 }
1988
1989 static struct ib_mad_send_wr_private*
1990 find_send_by_wr_id(struct ib_mad_agent_private *mad_agent_priv,
1991                    u64 wr_id)
1992 {
1993         struct ib_mad_send_wr_private *mad_send_wr;
1994
1995         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
1996                             agent_list) {
1997                 if (mad_send_wr->wr_id == wr_id)
1998                         return mad_send_wr;
1999         }
2000
2001         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2002                             agent_list) {
2003                 if (mad_send_wr->wr_id == wr_id)
2004                         return mad_send_wr;
2005         }
2006         return NULL;
2007 }
2008
2009 void cancel_sends(void *data)
2010 {
2011         struct ib_mad_agent_private *mad_agent_priv;
2012         struct ib_mad_send_wr_private *mad_send_wr;
2013         struct ib_mad_send_wc mad_send_wc;
2014         unsigned long flags;
2015
2016         mad_agent_priv = data;
2017
2018         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2019         mad_send_wc.vendor_err = 0;
2020
2021         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2022         while (!list_empty(&mad_agent_priv->canceled_list)) {
2023                 mad_send_wr = list_entry(mad_agent_priv->canceled_list.next,
2024                                          struct ib_mad_send_wr_private,
2025                                          agent_list);
2026
2027                 list_del(&mad_send_wr->agent_list);
2028                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2029
2030                 mad_send_wc.wr_id = mad_send_wr->wr_id;
2031                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2032                                                    &mad_send_wc);
2033
2034                 kfree(mad_send_wr);
2035                 if (atomic_dec_and_test(&mad_agent_priv->refcount))
2036                         wake_up(&mad_agent_priv->wait);
2037                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2038         }
2039         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2040 }
2041
2042 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2043                   u64 wr_id)
2044 {
2045         struct ib_mad_agent_private *mad_agent_priv;
2046         struct ib_mad_send_wr_private *mad_send_wr;
2047         unsigned long flags;
2048
2049         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2050                                       agent);
2051         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2052         mad_send_wr = find_send_by_wr_id(mad_agent_priv, wr_id);
2053         if (!mad_send_wr) {
2054                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2055                 goto out;
2056         }
2057
2058         if (mad_send_wr->status == IB_WC_SUCCESS)
2059                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2060
2061         if (mad_send_wr->refcount != 0) {
2062                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2063                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2064                 goto out;
2065         }
2066
2067         list_del(&mad_send_wr->agent_list);
2068         list_add_tail(&mad_send_wr->agent_list, &mad_agent_priv->canceled_list);
2069         adjust_timeout(mad_agent_priv);
2070         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2071
2072         queue_work(mad_agent_priv->qp_info->port_priv->wq,
2073                    &mad_agent_priv->canceled_work);
2074 out:
2075         return;
2076 }
2077 EXPORT_SYMBOL(ib_cancel_mad);
2078
2079 static void local_completions(void *data)
2080 {
2081         struct ib_mad_agent_private *mad_agent_priv;
2082         struct ib_mad_local_private *local;
2083         struct ib_mad_agent_private *recv_mad_agent;
2084         unsigned long flags;
2085         struct ib_wc wc;
2086         struct ib_mad_send_wc mad_send_wc;
2087
2088         mad_agent_priv = (struct ib_mad_agent_private *)data;
2089
2090         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2091         while (!list_empty(&mad_agent_priv->local_list)) {
2092                 local = list_entry(mad_agent_priv->local_list.next,
2093                                    struct ib_mad_local_private,
2094                                    completion_list);
2095                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2096                 if (local->mad_priv) {
2097                         recv_mad_agent = local->recv_mad_agent;
2098                         if (!recv_mad_agent) {
2099                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2100                                 kmem_cache_free(ib_mad_cache, local->mad_priv);
2101                                 goto local_send_completion;
2102                         }
2103
2104                         /*
2105                          * Defined behavior is to complete response
2106                          * before request
2107                          */
2108                         build_smp_wc(local->wr_id, IB_LID_PERMISSIVE,
2109                                      0 /* pkey index */,
2110                                      recv_mad_agent->agent.port_num, &wc);
2111
2112                         local->mad_priv->header.recv_wc.wc = &wc;
2113                         local->mad_priv->header.recv_wc.mad_len =
2114                                                 sizeof(struct ib_mad);
2115                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.recv_buf.list);
2116                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2117                         local->mad_priv->header.recv_wc.recv_buf.mad =
2118                                                 &local->mad_priv->mad.mad;
2119                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2120                                 snoop_recv(recv_mad_agent->qp_info,
2121                                           &local->mad_priv->header.recv_wc,
2122                                            IB_MAD_SNOOP_RECVS);
2123                         recv_mad_agent->agent.recv_handler(
2124                                                 &recv_mad_agent->agent,
2125                                                 &local->mad_priv->header.recv_wc);
2126                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2127                         atomic_dec(&recv_mad_agent->refcount);
2128                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2129                 }
2130
2131 local_send_completion:
2132                 /* Complete send */
2133                 mad_send_wc.status = IB_WC_SUCCESS;
2134                 mad_send_wc.vendor_err = 0;
2135                 mad_send_wc.wr_id = local->wr_id;
2136                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2137                         snoop_send(mad_agent_priv->qp_info, &local->send_wr,
2138                                   &mad_send_wc,
2139                                    IB_MAD_SNOOP_SEND_COMPLETIONS);
2140                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2141                                                    &mad_send_wc);
2142
2143                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2144                 list_del(&local->completion_list);
2145                 atomic_dec(&mad_agent_priv->refcount);
2146                 kfree(local);
2147         }
2148         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2149 }
2150
2151 static void timeout_sends(void *data)
2152 {
2153         struct ib_mad_agent_private *mad_agent_priv;
2154         struct ib_mad_send_wr_private *mad_send_wr;
2155         struct ib_mad_send_wc mad_send_wc;
2156         unsigned long flags, delay;
2157
2158         mad_agent_priv = (struct ib_mad_agent_private *)data;
2159
2160         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2161         mad_send_wc.vendor_err = 0;
2162
2163         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2164         while (!list_empty(&mad_agent_priv->wait_list)) {
2165                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2166                                          struct ib_mad_send_wr_private,
2167                                          agent_list);
2168
2169                 if (time_after(mad_send_wr->timeout, jiffies)) {
2170                         delay = mad_send_wr->timeout - jiffies;
2171                         if ((long)delay <= 0)
2172                                 delay = 1;
2173                         queue_delayed_work(mad_agent_priv->qp_info->
2174                                            port_priv->wq,
2175                                            &mad_agent_priv->timed_work, delay);
2176                         break;
2177                 }
2178
2179                 list_del(&mad_send_wr->agent_list);
2180                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2181
2182                 mad_send_wc.wr_id = mad_send_wr->wr_id;
2183                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2184                                                    &mad_send_wc);
2185
2186                 kfree(mad_send_wr);
2187                 atomic_dec(&mad_agent_priv->refcount);
2188                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2189         }
2190         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2191 }
2192
2193 static void ib_mad_thread_completion_handler(struct ib_cq *cq)
2194 {
2195         struct ib_mad_port_private *port_priv = cq->cq_context;
2196
2197         queue_work(port_priv->wq, &port_priv->work);
2198 }
2199
2200 /*
2201  * Allocate receive MADs and post receive WRs for them
2202  */
2203 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2204                                     struct ib_mad_private *mad)
2205 {
2206         unsigned long flags;
2207         int post, ret;
2208         struct ib_mad_private *mad_priv;
2209         struct ib_sge sg_list;
2210         struct ib_recv_wr recv_wr, *bad_recv_wr;
2211         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2212
2213         /* Initialize common scatter list fields */
2214         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2215         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2216
2217         /* Initialize common receive WR fields */
2218         recv_wr.next = NULL;
2219         recv_wr.sg_list = &sg_list;
2220         recv_wr.num_sge = 1;
2221
2222         do {
2223                 /* Allocate and map receive buffer */
2224                 if (mad) {
2225                         mad_priv = mad;
2226                         mad = NULL;
2227                 } else {
2228                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2229                         if (!mad_priv) {
2230                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2231                                 ret = -ENOMEM;
2232                                 break;
2233                         }
2234                 }
2235                 sg_list.addr = dma_map_single(qp_info->port_priv->
2236                                                 device->dma_device,
2237                                         &mad_priv->grh,
2238                                         sizeof *mad_priv -
2239                                                 sizeof mad_priv->header,
2240                                         DMA_FROM_DEVICE);
2241                 pci_unmap_addr_set(&mad_priv->header, mapping, sg_list.addr);
2242                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2243                 mad_priv->header.mad_list.mad_queue = recv_queue;
2244
2245                 /* Post receive WR */
2246                 spin_lock_irqsave(&recv_queue->lock, flags);
2247                 post = (++recv_queue->count < recv_queue->max_active);
2248                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2249                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2250                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2251                 if (ret) {
2252                         spin_lock_irqsave(&recv_queue->lock, flags);
2253                         list_del(&mad_priv->header.mad_list.list);
2254                         recv_queue->count--;
2255                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2256                         dma_unmap_single(qp_info->port_priv->device->dma_device,
2257                                          pci_unmap_addr(&mad_priv->header,
2258                                                         mapping),
2259                                          sizeof *mad_priv -
2260                                            sizeof mad_priv->header,
2261                                          DMA_FROM_DEVICE);
2262                         kmem_cache_free(ib_mad_cache, mad_priv);
2263                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2264                         break;
2265                 }
2266         } while (post);
2267
2268         return ret;
2269 }
2270
2271 /*
2272  * Return all the posted receive MADs
2273  */
2274 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2275 {
2276         struct ib_mad_private_header *mad_priv_hdr;
2277         struct ib_mad_private *recv;
2278         struct ib_mad_list_head *mad_list;
2279
2280         while (!list_empty(&qp_info->recv_queue.list)) {
2281
2282                 mad_list = list_entry(qp_info->recv_queue.list.next,
2283                                       struct ib_mad_list_head, list);
2284                 mad_priv_hdr = container_of(mad_list,
2285                                             struct ib_mad_private_header,
2286                                             mad_list);
2287                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2288                                     header);
2289
2290                 /* Remove from posted receive MAD list */
2291                 list_del(&mad_list->list);
2292
2293                 dma_unmap_single(qp_info->port_priv->device->dma_device,
2294                                  pci_unmap_addr(&recv->header, mapping),
2295                                  sizeof(struct ib_mad_private) -
2296                                  sizeof(struct ib_mad_private_header),
2297                                  DMA_FROM_DEVICE);
2298                 kmem_cache_free(ib_mad_cache, recv);
2299         }
2300
2301         qp_info->recv_queue.count = 0;
2302 }
2303
2304 /*
2305  * Start the port
2306  */
2307 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2308 {
2309         int ret, i;
2310         struct ib_qp_attr *attr;
2311         struct ib_qp *qp;
2312
2313         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2314         if (!attr) {
2315                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2316                 return -ENOMEM;
2317         }
2318
2319         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2320                 qp = port_priv->qp_info[i].qp;
2321                 /*
2322                  * PKey index for QP1 is irrelevant but
2323                  * one is needed for the Reset to Init transition
2324                  */
2325                 attr->qp_state = IB_QPS_INIT;
2326                 attr->pkey_index = 0;
2327                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2328                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2329                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2330                 if (ret) {
2331                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2332                                "INIT: %d\n", i, ret);
2333                         goto out;
2334                 }
2335
2336                 attr->qp_state = IB_QPS_RTR;
2337                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2338                 if (ret) {
2339                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2340                                "RTR: %d\n", i, ret);
2341                         goto out;
2342                 }
2343
2344                 attr->qp_state = IB_QPS_RTS;
2345                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2346                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2347                 if (ret) {
2348                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2349                                "RTS: %d\n", i, ret);
2350                         goto out;
2351                 }
2352         }
2353
2354         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2355         if (ret) {
2356                 printk(KERN_ERR PFX "Failed to request completion "
2357                        "notification: %d\n", ret);
2358                 goto out;
2359         }
2360
2361         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2362                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2363                 if (ret) {
2364                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2365                         goto out;
2366                 }
2367         }
2368 out:
2369         kfree(attr);
2370         return ret;
2371 }
2372
2373 static void qp_event_handler(struct ib_event *event, void *qp_context)
2374 {
2375         struct ib_mad_qp_info   *qp_info = qp_context;
2376
2377         /* It's worse than that! He's dead, Jim! */
2378         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2379                 event->event, qp_info->qp->qp_num);
2380 }
2381
2382 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2383                            struct ib_mad_queue *mad_queue)
2384 {
2385         mad_queue->qp_info = qp_info;
2386         mad_queue->count = 0;
2387         spin_lock_init(&mad_queue->lock);
2388         INIT_LIST_HEAD(&mad_queue->list);
2389 }
2390
2391 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2392                         struct ib_mad_qp_info *qp_info)
2393 {
2394         qp_info->port_priv = port_priv;
2395         init_mad_queue(qp_info, &qp_info->send_queue);
2396         init_mad_queue(qp_info, &qp_info->recv_queue);
2397         INIT_LIST_HEAD(&qp_info->overflow_list);
2398         spin_lock_init(&qp_info->snoop_lock);
2399         qp_info->snoop_table = NULL;
2400         qp_info->snoop_table_size = 0;
2401         atomic_set(&qp_info->snoop_count, 0);
2402 }
2403
2404 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2405                          enum ib_qp_type qp_type)
2406 {
2407         struct ib_qp_init_attr  qp_init_attr;
2408         int ret;
2409
2410         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2411         qp_init_attr.send_cq = qp_info->port_priv->cq;
2412         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2413         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2414         qp_init_attr.cap.max_send_wr = IB_MAD_QP_SEND_SIZE;
2415         qp_init_attr.cap.max_recv_wr = IB_MAD_QP_RECV_SIZE;
2416         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2417         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2418         qp_init_attr.qp_type = qp_type;
2419         qp_init_attr.port_num = qp_info->port_priv->port_num;
2420         qp_init_attr.qp_context = qp_info;
2421         qp_init_attr.event_handler = qp_event_handler;
2422         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2423         if (IS_ERR(qp_info->qp)) {
2424                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2425                        get_spl_qp_index(qp_type));
2426                 ret = PTR_ERR(qp_info->qp);
2427                 goto error;
2428         }
2429         /* Use minimum queue sizes unless the CQ is resized */
2430         qp_info->send_queue.max_active = IB_MAD_QP_SEND_SIZE;
2431         qp_info->recv_queue.max_active = IB_MAD_QP_RECV_SIZE;
2432         return 0;
2433
2434 error:
2435         return ret;
2436 }
2437
2438 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2439 {
2440         ib_destroy_qp(qp_info->qp);
2441         if (qp_info->snoop_table)
2442                 kfree(qp_info->snoop_table);
2443 }
2444
2445 /*
2446  * Open the port
2447  * Create the QP, PD, MR, and CQ if needed
2448  */
2449 static int ib_mad_port_open(struct ib_device *device,
2450                             int port_num)
2451 {
2452         int ret, cq_size;
2453         struct ib_mad_port_private *port_priv;
2454         unsigned long flags;
2455         char name[sizeof "ib_mad123"];
2456
2457         /* First, check if port already open at MAD layer */
2458         port_priv = ib_get_mad_port(device, port_num);
2459         if (port_priv) {
2460                 printk(KERN_DEBUG PFX "%s port %d already open\n",
2461                        device->name, port_num);
2462                 return 0;
2463         }
2464
2465         /* Create new device info */
2466         port_priv = kmalloc(sizeof *port_priv, GFP_KERNEL);
2467         if (!port_priv) {
2468                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2469                 return -ENOMEM;
2470         }
2471         memset(port_priv, 0, sizeof *port_priv);
2472         port_priv->device = device;
2473         port_priv->port_num = port_num;
2474         spin_lock_init(&port_priv->reg_lock);
2475         INIT_LIST_HEAD(&port_priv->agent_list);
2476         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2477         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2478
2479         cq_size = (IB_MAD_QP_SEND_SIZE + IB_MAD_QP_RECV_SIZE) * 2;
2480         port_priv->cq = ib_create_cq(port_priv->device,
2481                                      (ib_comp_handler)
2482                                         ib_mad_thread_completion_handler,
2483                                      NULL, port_priv, cq_size);
2484         if (IS_ERR(port_priv->cq)) {
2485                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2486                 ret = PTR_ERR(port_priv->cq);
2487                 goto error3;
2488         }
2489
2490         port_priv->pd = ib_alloc_pd(device);
2491         if (IS_ERR(port_priv->pd)) {
2492                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2493                 ret = PTR_ERR(port_priv->pd);
2494                 goto error4;
2495         }
2496
2497         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2498         if (IS_ERR(port_priv->mr)) {
2499                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2500                 ret = PTR_ERR(port_priv->mr);
2501                 goto error5;
2502         }
2503
2504         ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2505         if (ret)
2506                 goto error6;
2507         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2508         if (ret)
2509                 goto error7;
2510
2511         snprintf(name, sizeof name, "ib_mad%d", port_num);
2512         port_priv->wq = create_singlethread_workqueue(name);
2513         if (!port_priv->wq) {
2514                 ret = -ENOMEM;
2515                 goto error8;
2516         }
2517         INIT_WORK(&port_priv->work, ib_mad_completion_handler, port_priv);
2518
2519         ret = ib_mad_port_start(port_priv);
2520         if (ret) {
2521                 printk(KERN_ERR PFX "Couldn't start port\n");
2522                 goto error9;
2523         }
2524
2525         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2526         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2527         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2528         return 0;
2529
2530 error9:
2531         destroy_workqueue(port_priv->wq);
2532 error8:
2533         destroy_mad_qp(&port_priv->qp_info[1]);
2534 error7:
2535         destroy_mad_qp(&port_priv->qp_info[0]);
2536 error6:
2537         ib_dereg_mr(port_priv->mr);
2538 error5:
2539         ib_dealloc_pd(port_priv->pd);
2540 error4:
2541         ib_destroy_cq(port_priv->cq);
2542         cleanup_recv_queue(&port_priv->qp_info[1]);
2543         cleanup_recv_queue(&port_priv->qp_info[0]);
2544 error3:
2545         kfree(port_priv);
2546
2547         return ret;
2548 }
2549
2550 /*
2551  * Close the port
2552  * If there are no classes using the port, free the port
2553  * resources (CQ, MR, PD, QP) and remove the port's info structure
2554  */
2555 static int ib_mad_port_close(struct ib_device *device, int port_num)
2556 {
2557         struct ib_mad_port_private *port_priv;
2558         unsigned long flags;
2559
2560         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2561         port_priv = __ib_get_mad_port(device, port_num);
2562         if (port_priv == NULL) {
2563                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2564                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2565                 return -ENODEV;
2566         }
2567         list_del(&port_priv->port_list);
2568         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2569
2570         /* Stop processing completions. */
2571         flush_workqueue(port_priv->wq);
2572         destroy_workqueue(port_priv->wq);
2573         destroy_mad_qp(&port_priv->qp_info[1]);
2574         destroy_mad_qp(&port_priv->qp_info[0]);
2575         ib_dereg_mr(port_priv->mr);
2576         ib_dealloc_pd(port_priv->pd);
2577         ib_destroy_cq(port_priv->cq);
2578         cleanup_recv_queue(&port_priv->qp_info[1]);
2579         cleanup_recv_queue(&port_priv->qp_info[0]);
2580         /* XXX: Handle deallocation of MAD registration tables */
2581
2582         kfree(port_priv);
2583
2584         return 0;
2585 }
2586
2587 static void ib_mad_init_device(struct ib_device *device)
2588 {
2589         int ret, num_ports, cur_port, i, ret2;
2590
2591         if (device->node_type == IB_NODE_SWITCH) {
2592                 num_ports = 1;
2593                 cur_port = 0;
2594         } else {
2595                 num_ports = device->phys_port_cnt;
2596                 cur_port = 1;
2597         }
2598         for (i = 0; i < num_ports; i++, cur_port++) {
2599                 ret = ib_mad_port_open(device, cur_port);
2600                 if (ret) {
2601                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2602                                device->name, cur_port);
2603                         goto error_device_open;
2604                 }
2605                 ret = ib_agent_port_open(device, cur_port);
2606                 if (ret) {
2607                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2608                                "for agents\n",
2609                                device->name, cur_port);
2610                         goto error_device_open;
2611                 }
2612         }
2613
2614         goto error_device_query;
2615
2616 error_device_open:
2617         while (i > 0) {
2618                 cur_port--;
2619                 ret2 = ib_agent_port_close(device, cur_port);
2620                 if (ret2) {
2621                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2622                                "for agents\n",
2623                                device->name, cur_port);
2624                 }
2625                 ret2 = ib_mad_port_close(device, cur_port);
2626                 if (ret2) {
2627                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2628                                device->name, cur_port);
2629                 }
2630                 i--;
2631         }
2632
2633 error_device_query:
2634         return;
2635 }
2636
2637 static void ib_mad_remove_device(struct ib_device *device)
2638 {
2639         int ret = 0, i, num_ports, cur_port, ret2;
2640
2641         if (device->node_type == IB_NODE_SWITCH) {
2642                 num_ports = 1;
2643                 cur_port = 0;
2644         } else {
2645                 num_ports = device->phys_port_cnt;
2646                 cur_port = 1;
2647         }
2648         for (i = 0; i < num_ports; i++, cur_port++) {
2649                 ret2 = ib_agent_port_close(device, cur_port);
2650                 if (ret2) {
2651                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2652                                "for agents\n",
2653                                device->name, cur_port);
2654                         if (!ret)
2655                                 ret = ret2;
2656                 }
2657                 ret2 = ib_mad_port_close(device, cur_port);
2658                 if (ret2) {
2659                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2660                                device->name, cur_port);
2661                         if (!ret)
2662                                 ret = ret2;
2663                 }
2664         }
2665 }
2666
2667 static struct ib_client mad_client = {
2668         .name   = "mad",
2669         .add = ib_mad_init_device,
2670         .remove = ib_mad_remove_device
2671 };
2672
2673 static int __init ib_mad_init_module(void)
2674 {
2675         int ret;
2676
2677         spin_lock_init(&ib_mad_port_list_lock);
2678         spin_lock_init(&ib_agent_port_list_lock);
2679
2680         ib_mad_cache = kmem_cache_create("ib_mad",
2681                                          sizeof(struct ib_mad_private),
2682                                          0,
2683                                          SLAB_HWCACHE_ALIGN,
2684                                          NULL,
2685                                          NULL);
2686         if (!ib_mad_cache) {
2687                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
2688                 ret = -ENOMEM;
2689                 goto error1;
2690         }
2691
2692         INIT_LIST_HEAD(&ib_mad_port_list);
2693
2694         if (ib_register_client(&mad_client)) {
2695                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
2696                 ret = -EINVAL;
2697                 goto error2;
2698         }
2699
2700         return 0;
2701
2702 error2:
2703         kmem_cache_destroy(ib_mad_cache);
2704 error1:
2705         return ret;
2706 }
2707
2708 static void __exit ib_mad_cleanup_module(void)
2709 {
2710         ib_unregister_client(&mad_client);
2711
2712         if (kmem_cache_destroy(ib_mad_cache)) {
2713                 printk(KERN_DEBUG PFX "Failed to destroy ib_mad cache\n");
2714         }
2715 }
2716
2717 module_init(ib_mad_init_module);
2718 module_exit(ib_mad_cleanup_module);