drivers/infiniband/core/mad.c

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