drivers/infiniband/core/mad.c

   1 /*
   2  * Copyright (c) 2004-2007 Voltaire, Inc. All rights reserved.
   3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
   5  * Copyright (c) 2009 HNR Consulting. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  *
  35  */
  36 #include <linux/dma-mapping.h>
  37 #include <linux/slab.h>
  38 #include <rdma/ib_cache.h>
  39
  40 #include "mad_priv.h"
  41 #include "mad_rmpp.h"
  42 #include "smi.h"
  43 #include "agent.h"
  44
  45 MODULE_LICENSE("Dual BSD/GPL");
  46 MODULE_DESCRIPTION("kernel IB MAD API");
  47 MODULE_AUTHOR("Hal Rosenstock");
  48 MODULE_AUTHOR("Sean Hefty");
  49
  50 int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
  51 int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
  52
  53 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
  54 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
  55 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
  56 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
  57
  58 static struct kmem_cache *ib_mad_cache;
  59
  60 static struct list_head ib_mad_port_list;
  61 static u32 ib_mad_client_id = 0;
  62
  63 /* Port list lock */
  64 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
  65
  66 /* Forward declarations */
  67 static int method_in_use(struct ib_mad_mgmt_method_table **method,
  68                          struct ib_mad_reg_req *mad_reg_req);
  69 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
  70 static struct ib_mad_agent_private *find_mad_agent(
  71                                         struct ib_mad_port_private *port_priv,
  72                                         struct ib_mad *mad);
  73 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
  74                                     struct ib_mad_private *mad);
  75 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
  76 static void timeout_sends(struct work_struct *work);
  77 static void local_completions(struct work_struct *work);
  78 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  79                               struct ib_mad_agent_private *agent_priv,
  80                               u8 mgmt_class);
  81 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  82                            struct ib_mad_agent_private *agent_priv);
  83
  84 /*
  85  * Returns a ib_mad_port_private structure or NULL for a device/port
  86  * Assumes ib_mad_port_list_lock is being held
  87  */
  88 static inline struct ib_mad_port_private *
  89 __ib_get_mad_port(struct ib_device *device, int port_num)
  90 {
  91         struct ib_mad_port_private *entry;
  92
  93         list_for_each_entry(entry, &ib_mad_port_list, port_list) {
  94                 if (entry->device == device && entry->port_num == port_num)
  95                         return entry;
  96         }
  97         return NULL;
  98 }
  99
 100 /*
 101  * Wrapper function to return a ib_mad_port_private structure or NULL
 102  * for a device/port
 103  */
 104 static inline struct ib_mad_port_private *
 105 ib_get_mad_port(struct ib_device *device, int port_num)
 106 {
 107         struct ib_mad_port_private *entry;
 108         unsigned long flags;
 109
 110         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
 111         entry = __ib_get_mad_port(device, port_num);
 112         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
 113
 114         return entry;
 115 }
 116
 117 static inline u8 convert_mgmt_class(u8 mgmt_class)
 118 {
 119         /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
 120         return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
 121                 0 : mgmt_class;
 122 }
 123
 124 static int get_spl_qp_index(enum ib_qp_type qp_type)
 125 {
 126         switch (qp_type)
 127         {
 128         case IB_QPT_SMI:
 129                 return 0;
 130         case IB_QPT_GSI:
 131                 return 1;
 132         default:
 133                 return -1;
 134         }
 135 }
 136
 137 static int vendor_class_index(u8 mgmt_class)
 138 {
 139         return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
 140 }
 141
 142 static int is_vendor_class(u8 mgmt_class)
 143 {
 144         if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
 145             (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
 146                 return 0;
 147         return 1;
 148 }
 149
 150 static int is_vendor_oui(char *oui)
 151 {
 152         if (oui[0] || oui[1] || oui[2])
 153                 return 1;
 154         return 0;
 155 }
 156
 157 static int is_vendor_method_in_use(
 158                 struct ib_mad_mgmt_vendor_class *vendor_class,
 159                 struct ib_mad_reg_req *mad_reg_req)
 160 {
 161         struct ib_mad_mgmt_method_table *method;
 162         int i;
 163
 164         for (i = 0; i < MAX_MGMT_OUI; i++) {
 165                 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
 166                         method = vendor_class->method_table[i];
 167                         if (method) {
 168                                 if (method_in_use(&method, mad_reg_req))
 169                                         return 1;
 170                                 else
 171                                         break;
 172                         }
 173                 }
 174         }
 175         return 0;
 176 }
 177
 178 int ib_response_mad(struct ib_mad *mad)
 179 {
 180         return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
 181                 (mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
 182                 ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
 183                  (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
 184 }
 185 EXPORT_SYMBOL(ib_response_mad);
 186
 187 /*
 188  * ib_register_mad_agent - Register to send/receive MADs
 189  */
 190 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 191                                            u8 port_num,
 192                                            enum ib_qp_type qp_type,
 193                                            struct ib_mad_reg_req *mad_reg_req,
 194                                            u8 rmpp_version,
 195                                            ib_mad_send_handler send_handler,
 196                                            ib_mad_recv_handler recv_handler,
 197                                            void *context)
 198 {
 199         struct ib_mad_port_private *port_priv;
 200         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
 201         struct ib_mad_agent_private *mad_agent_priv;
 202         struct ib_mad_reg_req *reg_req = NULL;
 203         struct ib_mad_mgmt_class_table *class;
 204         struct ib_mad_mgmt_vendor_class_table *vendor;
 205         struct ib_mad_mgmt_vendor_class *vendor_class;
 206         struct ib_mad_mgmt_method_table *method;
 207         int ret2, qpn;
 208         unsigned long flags;
 209         u8 mgmt_class, vclass;
 210
 211         /* Validate parameters */
 212         qpn = get_spl_qp_index(qp_type);
 213         if (qpn == -1)
 214                 goto error1;
 215
 216         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
 217                 goto error1;
 218
 219         /* Validate MAD registration request if supplied */
 220         if (mad_reg_req) {
 221                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
 222                         goto error1;
 223                 if (!recv_handler)
 224                         goto error1;
 225                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
 226                         /*
 227                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
 228                          * one in this range currently allowed
 229                          */
 230                         if (mad_reg_req->mgmt_class !=
 231                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 232                                 goto error1;
 233                 } else if (mad_reg_req->mgmt_class == 0) {
 234                         /*
 235                          * Class 0 is reserved in IBA and is used for
 236                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
 237                          */
 238                         goto error1;
 239                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
 240                         /*
 241                          * If class is in "new" vendor range,
 242                          * ensure supplied OUI is not zero
 243                          */
 244                         if (!is_vendor_oui(mad_reg_req->oui))
 245                                 goto error1;
 246                 }
 247                 /* Make sure class supplied is consistent with RMPP */
 248                 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
 249                         if (rmpp_version)
 250                                 goto error1;
 251                 }
 252                 /* Make sure class supplied is consistent with QP type */
 253                 if (qp_type == IB_QPT_SMI) {
 254                         if ((mad_reg_req->mgmt_class !=
 255                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
 256                             (mad_reg_req->mgmt_class !=
 257                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 258                                 goto error1;
 259                 } else {
 260                         if ((mad_reg_req->mgmt_class ==
 261                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
 262                             (mad_reg_req->mgmt_class ==
 263                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 264                                 goto error1;
 265                 }
 266         } else {
 267                 /* No registration request supplied */
 268                 if (!send_handler)
 269                         goto error1;
 270         }
 271
 272         /* Validate device and port */
 273         port_priv = ib_get_mad_port(device, port_num);
 274         if (!port_priv) {
 275                 ret = ERR_PTR(-ENODEV);
 276                 goto error1;
 277         }
 278
 279         /* Allocate structures */
 280         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
 281         if (!mad_agent_priv) {
 282                 ret = ERR_PTR(-ENOMEM);
 283                 goto error1;
 284         }
 285
 286         mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
 287                                                  IB_ACCESS_LOCAL_WRITE);
 288         if (IS_ERR(mad_agent_priv->agent.mr)) {
 289                 ret = ERR_PTR(-ENOMEM);
 290                 goto error2;
 291         }
 292
 293         if (mad_reg_req) {
 294                 reg_req = kmalloc(sizeof *reg_req, GFP_KERNEL);
 295                 if (!reg_req) {
 296                         ret = ERR_PTR(-ENOMEM);
 297                         goto error3;
 298                 }
 299                 /* Make a copy of the MAD registration request */
 300                 memcpy(reg_req, mad_reg_req, sizeof *reg_req);
 301         }
 302
 303         /* Now, fill in the various structures */
 304         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 305         mad_agent_priv->reg_req = reg_req;
 306         mad_agent_priv->agent.rmpp_version = rmpp_version;
 307         mad_agent_priv->agent.device = device;
 308         mad_agent_priv->agent.recv_handler = recv_handler;
 309         mad_agent_priv->agent.send_handler = send_handler;
 310         mad_agent_priv->agent.context = context;
 311         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 312         mad_agent_priv->agent.port_num = port_num;
 313         spin_lock_init(&mad_agent_priv->lock);
 314         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 315         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 316         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 317         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 318         INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
 319         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 320         INIT_WORK(&mad_agent_priv->local_work, local_completions);
 321         atomic_set(&mad_agent_priv->refcount, 1);
 322         init_completion(&mad_agent_priv->comp);
 323
 324         spin_lock_irqsave(&port_priv->reg_lock, flags);
 325         mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
 326
 327         /*
 328          * Make sure MAD registration (if supplied)
 329          * is non overlapping with any existing ones
 330          */
 331         if (mad_reg_req) {
 332                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 333                 if (!is_vendor_class(mgmt_class)) {
 334                         class = port_priv->version[mad_reg_req->
 335                                                    mgmt_class_version].class;
 336                         if (class) {
 337                                 method = class->method_table[mgmt_class];
 338                                 if (method) {
 339                                         if (method_in_use(&method,
 340                                                            mad_reg_req))
 341                                                 goto error4;
 342                                 }
 343                         }
 344                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 345                                                   mgmt_class);
 346                 } else {
 347                         /* "New" vendor class range */
 348                         vendor = port_priv->version[mad_reg_req->
 349                                                     mgmt_class_version].vendor;
 350                         if (vendor) {
 351                                 vclass = vendor_class_index(mgmt_class);
 352                                 vendor_class = vendor->vendor_class[vclass];
 353                                 if (vendor_class) {
 354                                         if (is_vendor_method_in_use(
 355                                                         vendor_class,
 356                                                         mad_reg_req))
 357                                                 goto error4;
 358                                 }
 359                         }
 360                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 361                 }
 362                 if (ret2) {
 363                         ret = ERR_PTR(ret2);
 364                         goto error4;
 365                 }
 366         }
 367
 368         /* Add mad agent into port's agent list */
 369         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
 370         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 371
 372         return &mad_agent_priv->agent;
 373
 374 error4:
 375         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 376         kfree(reg_req);
 377 error3:
 378         ib_dereg_mr(mad_agent_priv->agent.mr);
 379 error2:
 380         kfree(mad_agent_priv);
 381 error1:
 382         return ret;
 383 }
 384 EXPORT_SYMBOL(ib_register_mad_agent);
 385
 386 static inline int is_snooping_sends(int mad_snoop_flags)
 387 {
 388         return (mad_snoop_flags &
 389                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 390                  IB_MAD_SNOOP_RMPP_SENDS |*/
 391                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 392                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 393 }
 394
 395 static inline int is_snooping_recvs(int mad_snoop_flags)
 396 {
 397         return (mad_snoop_flags &
 398                 (IB_MAD_SNOOP_RECVS /*|
 399                  IB_MAD_SNOOP_RMPP_RECVS*/));
 400 }
 401
 402 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 403                                 struct ib_mad_snoop_private *mad_snoop_priv)
 404 {
 405         struct ib_mad_snoop_private **new_snoop_table;
 406         unsigned long flags;
 407         int i;
 408
 409         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 410         /* Check for empty slot in array. */
 411         for (i = 0; i < qp_info->snoop_table_size; i++)
 412                 if (!qp_info->snoop_table[i])
 413                         break;
 414
 415         if (i == qp_info->snoop_table_size) {
 416                 /* Grow table. */
 417                 new_snoop_table = krealloc(qp_info->snoop_table,
 418                                            sizeof mad_snoop_priv *
 419                                            (qp_info->snoop_table_size + 1),
 420                                            GFP_ATOMIC);
 421                 if (!new_snoop_table) {
 422                         i = -ENOMEM;
 423                         goto out;
 424                 }
 425
 426                 qp_info->snoop_table = new_snoop_table;
 427                 qp_info->snoop_table_size++;
 428         }
 429         qp_info->snoop_table[i] = mad_snoop_priv;
 430         atomic_inc(&qp_info->snoop_count);
 431 out:
 432         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 433         return i;
 434 }
 435
 436 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 437                                            u8 port_num,
 438                                            enum ib_qp_type qp_type,
 439                                            int mad_snoop_flags,
 440                                            ib_mad_snoop_handler snoop_handler,
 441                                            ib_mad_recv_handler recv_handler,
 442                                            void *context)
 443 {
 444         struct ib_mad_port_private *port_priv;
 445         struct ib_mad_agent *ret;
 446         struct ib_mad_snoop_private *mad_snoop_priv;
 447         int qpn;
 448
 449         /* Validate parameters */
 450         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 451             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 452                 ret = ERR_PTR(-EINVAL);
 453                 goto error1;
 454         }
 455         qpn = get_spl_qp_index(qp_type);
 456         if (qpn == -1) {
 457                 ret = ERR_PTR(-EINVAL);
 458                 goto error1;
 459         }
 460         port_priv = ib_get_mad_port(device, port_num);
 461         if (!port_priv) {
 462                 ret = ERR_PTR(-ENODEV);
 463                 goto error1;
 464         }
 465         /* Allocate structures */
 466         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 467         if (!mad_snoop_priv) {
 468                 ret = ERR_PTR(-ENOMEM);
 469                 goto error1;
 470         }
 471
 472         /* Now, fill in the various structures */
 473         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 474         mad_snoop_priv->agent.device = device;
 475         mad_snoop_priv->agent.recv_handler = recv_handler;
 476         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 477         mad_snoop_priv->agent.context = context;
 478         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 479         mad_snoop_priv->agent.port_num = port_num;
 480         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 481         init_completion(&mad_snoop_priv->comp);
 482         mad_snoop_priv->snoop_index = register_snoop_agent(
 483                                                 &port_priv->qp_info[qpn],
 484                                                 mad_snoop_priv);
 485         if (mad_snoop_priv->snoop_index < 0) {
 486                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 487                 goto error2;
 488         }
 489
 490         atomic_set(&mad_snoop_priv->refcount, 1);
 491         return &mad_snoop_priv->agent;
 492
 493 error2:
 494         kfree(mad_snoop_priv);
 495 error1:
 496         return ret;
 497 }
 498 EXPORT_SYMBOL(ib_register_mad_snoop);
 499
 500 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 501 {
 502         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 503                 complete(&mad_agent_priv->comp);
 504 }
 505
 506 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
 507 {
 508         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 509                 complete(&mad_snoop_priv->comp);
 510 }
 511
 512 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 513 {
 514         struct ib_mad_port_private *port_priv;
 515         unsigned long flags;
 516
 517         /* Note that we could still be handling received MADs */
 518
 519         /*
 520          * Canceling all sends results in dropping received response
 521          * MADs, preventing us from queuing additional work
 522          */
 523         cancel_mads(mad_agent_priv);
 524         port_priv = mad_agent_priv->qp_info->port_priv;
 525         cancel_delayed_work(&mad_agent_priv->timed_work);
 526
 527         spin_lock_irqsave(&port_priv->reg_lock, flags);
 528         remove_mad_reg_req(mad_agent_priv);
 529         list_del(&mad_agent_priv->agent_list);
 530         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 531
 532         flush_workqueue(port_priv->wq);
 533         ib_cancel_rmpp_recvs(mad_agent_priv);
 534
 535         deref_mad_agent(mad_agent_priv);
 536         wait_for_completion(&mad_agent_priv->comp);
 537
 538         kfree(mad_agent_priv->reg_req);
 539         ib_dereg_mr(mad_agent_priv->agent.mr);
 540         kfree(mad_agent_priv);
 541 }
 542
 543 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 544 {
 545         struct ib_mad_qp_info *qp_info;
 546         unsigned long flags;
 547
 548         qp_info = mad_snoop_priv->qp_info;
 549         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 550         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 551         atomic_dec(&qp_info->snoop_count);
 552         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 553
 554         deref_snoop_agent(mad_snoop_priv);
 555         wait_for_completion(&mad_snoop_priv->comp);
 556
 557         kfree(mad_snoop_priv);
 558 }
 559
 560 /*
 561  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 562  */
 563 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 564 {
 565         struct ib_mad_agent_private *mad_agent_priv;
 566         struct ib_mad_snoop_private *mad_snoop_priv;
 567
 568         /* If the TID is zero, the agent can only snoop. */
 569         if (mad_agent->hi_tid) {
 570                 mad_agent_priv = container_of(mad_agent,
 571                                               struct ib_mad_agent_private,
 572                                               agent);
 573                 unregister_mad_agent(mad_agent_priv);
 574         } else {
 575                 mad_snoop_priv = container_of(mad_agent,
 576                                               struct ib_mad_snoop_private,
 577                                               agent);
 578                 unregister_mad_snoop(mad_snoop_priv);
 579         }
 580         return 0;
 581 }
 582 EXPORT_SYMBOL(ib_unregister_mad_agent);
 583
 584 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 585 {
 586         struct ib_mad_queue *mad_queue;
 587         unsigned long flags;
 588
 589         BUG_ON(!mad_list->mad_queue);
 590         mad_queue = mad_list->mad_queue;
 591         spin_lock_irqsave(&mad_queue->lock, flags);
 592         list_del(&mad_list->list);
 593         mad_queue->count--;
 594         spin_unlock_irqrestore(&mad_queue->lock, flags);
 595 }
 596
 597 static void snoop_send(struct ib_mad_qp_info *qp_info,
 598                        struct ib_mad_send_buf *send_buf,
 599                        struct ib_mad_send_wc *mad_send_wc,
 600                        int mad_snoop_flags)
 601 {
 602         struct ib_mad_snoop_private *mad_snoop_priv;
 603         unsigned long flags;
 604         int i;
 605
 606         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 607         for (i = 0; i < qp_info->snoop_table_size; i++) {
 608                 mad_snoop_priv = qp_info->snoop_table[i];
 609                 if (!mad_snoop_priv ||
 610                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 611                         continue;
 612
 613                 atomic_inc(&mad_snoop_priv->refcount);
 614                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 615                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 616                                                     send_buf, mad_send_wc);
 617                 deref_snoop_agent(mad_snoop_priv);
 618                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 619         }
 620         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 621 }
 622
 623 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 624                        struct ib_mad_recv_wc *mad_recv_wc,
 625                        int mad_snoop_flags)
 626 {
 627         struct ib_mad_snoop_private *mad_snoop_priv;
 628         unsigned long flags;
 629         int i;
 630
 631         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 632         for (i = 0; i < qp_info->snoop_table_size; i++) {
 633                 mad_snoop_priv = qp_info->snoop_table[i];
 634                 if (!mad_snoop_priv ||
 635                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 636                         continue;
 637
 638                 atomic_inc(&mad_snoop_priv->refcount);
 639                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 640                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
 641                                                    mad_recv_wc);
 642                 deref_snoop_agent(mad_snoop_priv);
 643                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 644         }
 645         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 646 }
 647
 648 static void build_smp_wc(struct ib_qp *qp,
 649                          u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
 650                          struct ib_wc *wc)
 651 {
 652         memset(wc, 0, sizeof *wc);
 653         wc->wr_id = wr_id;
 654         wc->status = IB_WC_SUCCESS;
 655         wc->opcode = IB_WC_RECV;
 656         wc->pkey_index = pkey_index;
 657         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
 658         wc->src_qp = IB_QP0;
 659         wc->qp = qp;
 660         wc->slid = slid;
 661         wc->sl = 0;
 662         wc->dlid_path_bits = 0;
 663         wc->port_num = port_num;
 664 }
 665
 666 /*
 667  * Return 0 if SMP is to be sent
 668  * Return 1 if SMP was consumed locally (whether or not solicited)
 669  * Return < 0 if error
 670  */
 671 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 672                                   struct ib_mad_send_wr_private *mad_send_wr)
 673 {
 674         int ret = 0;
 675         struct ib_smp *smp = mad_send_wr->send_buf.mad;
 676         unsigned long flags;
 677         struct ib_mad_local_private *local;
 678         struct ib_mad_private *mad_priv;
 679         struct ib_mad_port_private *port_priv;
 680         struct ib_mad_agent_private *recv_mad_agent = NULL;
 681         struct ib_device *device = mad_agent_priv->agent.device;
 682         u8 port_num;
 683         struct ib_wc mad_wc;
 684         struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
 685
 686         if (device->node_type == RDMA_NODE_IB_SWITCH &&
 687             smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 688                 port_num = send_wr->wr.ud.port_num;
 689         else
 690                 port_num = mad_agent_priv->agent.port_num;
 691
 692         /*
 693          * Directed route handling starts if the initial LID routed part of
 694          * a request or the ending LID routed part of a response is empty.
 695          * If we are at the start of the LID routed part, don't update the
 696          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
 697          */
 698         if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
 699              IB_LID_PERMISSIVE &&
 700              smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
 701              IB_SMI_DISCARD) {
 702                 ret = -EINVAL;
 703                 printk(KERN_ERR PFX "Invalid directed route\n");
 704                 goto out;
 705         }
 706
 707         /* Check to post send on QP or process locally */
 708         if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
 709             smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
 710                 goto out;
 711
 712         local = kmalloc(sizeof *local, GFP_ATOMIC);
 713         if (!local) {
 714                 ret = -ENOMEM;
 715                 printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
 716                 goto out;
 717         }
 718         local->mad_priv = NULL;
 719         local->recv_mad_agent = NULL;
 720         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
 721         if (!mad_priv) {
 722                 ret = -ENOMEM;
 723                 printk(KERN_ERR PFX "No memory for local response MAD\n");
 724                 kfree(local);
 725                 goto out;
 726         }
 727
 728         build_smp_wc(mad_agent_priv->agent.qp,
 729                      send_wr->wr_id, be16_to_cpu(smp->dr_slid),
 730                      send_wr->wr.ud.pkey_index,
 731                      send_wr->wr.ud.port_num, &mad_wc);
 732
 733         /* No GRH for DR SMP */
 734         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 735                                   (struct ib_mad *)smp,
 736                                   (struct ib_mad *)&mad_priv->mad);
 737         switch (ret)
 738         {
 739         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 740                 if (ib_response_mad(&mad_priv->mad.mad) &&
 741                     mad_agent_priv->agent.recv_handler) {
 742                         local->mad_priv = mad_priv;
 743                         local->recv_mad_agent = mad_agent_priv;
 744                         /*
 745                          * Reference MAD agent until receive
 746                          * side of local completion handled
 747                          */
 748                         atomic_inc(&mad_agent_priv->refcount);
 749                 } else
 750                         kmem_cache_free(ib_mad_cache, mad_priv);
 751                 break;
 752         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 753                 kmem_cache_free(ib_mad_cache, mad_priv);
 754                 break;
 755         case IB_MAD_RESULT_SUCCESS:
 756                 /* Treat like an incoming receive MAD */
 757                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 758                                             mad_agent_priv->agent.port_num);
 759                 if (port_priv) {
 760                         memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
 761                         recv_mad_agent = find_mad_agent(port_priv,
 762                                                         &mad_priv->mad.mad);
 763                 }
 764                 if (!port_priv || !recv_mad_agent) {
 765                         /*
 766                          * No receiving agent so drop packet and
 767                          * generate send completion.
 768                          */
 769                         kmem_cache_free(ib_mad_cache, mad_priv);
 770                         break;
 771                 }
 772                 local->mad_priv = mad_priv;
 773                 local->recv_mad_agent = recv_mad_agent;
 774                 break;
 775         default:
 776                 kmem_cache_free(ib_mad_cache, mad_priv);
 777                 kfree(local);
 778                 ret = -EINVAL;
 779                 goto out;
 780         }
 781
 782         local->mad_send_wr = mad_send_wr;
 783         /* Reference MAD agent until send side of local completion handled */
 784         atomic_inc(&mad_agent_priv->refcount);
 785         /* Queue local completion to local list */
 786         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 787         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 788         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 789         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 790                    &mad_agent_priv->local_work);
 791         ret = 1;
 792 out:
 793         return ret;
 794 }
 795
 796 static int get_pad_size(int hdr_len, int data_len)
 797 {
 798         int seg_size, pad;
 799
 800         seg_size = sizeof(struct ib_mad) - hdr_len;
 801         if (data_len && seg_size) {
 802                 pad = seg_size - data_len % seg_size;
 803                 return pad == seg_size ? 0 : pad;
 804         } else
 805                 return seg_size;
 806 }
 807
 808 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 809 {
 810         struct ib_rmpp_segment *s, *t;
 811
 812         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 813                 list_del(&s->list);
 814                 kfree(s);
 815         }
 816 }
 817
 818 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 819                                 gfp_t gfp_mask)
 820 {
 821         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 822         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 823         struct ib_rmpp_segment *seg = NULL;
 824         int left, seg_size, pad;
 825
 826         send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
 827         seg_size = send_buf->seg_size;
 828         pad = send_wr->pad;
 829
 830         /* Allocate data segments. */
 831         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 832                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 833                 if (!seg) {
 834                         printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
 835                                "alloc failed for len %zd, gfp %#x\n",
 836                                sizeof (*seg) + seg_size, gfp_mask);
 837                         free_send_rmpp_list(send_wr);
 838                         return -ENOMEM;
 839                 }
 840                 seg->num = ++send_buf->seg_count;
 841                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 842         }
 843
 844         /* Zero any padding */
 845         if (pad)
 846                 memset(seg->data + seg_size - pad, 0, pad);
 847
 848         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
 849                                           agent.rmpp_version;
 850         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 851         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
 852
 853         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
 854                                         struct ib_rmpp_segment, list);
 855         send_wr->last_ack_seg = send_wr->cur_seg;
 856         return 0;
 857 }
 858
 859 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 860                                             u32 remote_qpn, u16 pkey_index,
 861                                             int rmpp_active,
 862                                             int hdr_len, int data_len,
 863                                             gfp_t gfp_mask)
 864 {
 865         struct ib_mad_agent_private *mad_agent_priv;
 866         struct ib_mad_send_wr_private *mad_send_wr;
 867         int pad, message_size, ret, size;
 868         void *buf;
 869
 870         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 871                                       agent);
 872         pad = get_pad_size(hdr_len, data_len);
 873         message_size = hdr_len + data_len + pad;
 874
 875         if ((!mad_agent->rmpp_version &&
 876              (rmpp_active || message_size > sizeof(struct ib_mad))) ||
 877             (!rmpp_active && message_size > sizeof(struct ib_mad)))
 878                 return ERR_PTR(-EINVAL);
 879
 880         size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
 881         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
 882         if (!buf)
 883                 return ERR_PTR(-ENOMEM);
 884
 885         mad_send_wr = buf + size;
 886         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
 887         mad_send_wr->send_buf.mad = buf;
 888         mad_send_wr->send_buf.hdr_len = hdr_len;
 889         mad_send_wr->send_buf.data_len = data_len;
 890         mad_send_wr->pad = pad;
 891
 892         mad_send_wr->mad_agent_priv = mad_agent_priv;
 893         mad_send_wr->sg_list[0].length = hdr_len;
 894         mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
 895         mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
 896         mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
 897
 898         mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
 899         mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 900         mad_send_wr->send_wr.num_sge = 2;
 901         mad_send_wr->send_wr.opcode = IB_WR_SEND;
 902         mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
 903         mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
 904         mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 905         mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
 906
 907         if (rmpp_active) {
 908                 ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
 909                 if (ret) {
 910                         kfree(buf);
 911                         return ERR_PTR(ret);
 912                 }
 913         }
 914
 915         mad_send_wr->send_buf.mad_agent = mad_agent;
 916         atomic_inc(&mad_agent_priv->refcount);
 917         return &mad_send_wr->send_buf;
 918 }
 919 EXPORT_SYMBOL(ib_create_send_mad);
 920
 921 int ib_get_mad_data_offset(u8 mgmt_class)
 922 {
 923         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
 924                 return IB_MGMT_SA_HDR;
 925         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 926                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 927                  (mgmt_class == IB_MGMT_CLASS_BIS))
 928                 return IB_MGMT_DEVICE_HDR;
 929         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 930                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
 931                 return IB_MGMT_VENDOR_HDR;
 932         else
 933                 return IB_MGMT_MAD_HDR;
 934 }
 935 EXPORT_SYMBOL(ib_get_mad_data_offset);
 936
 937 int ib_is_mad_class_rmpp(u8 mgmt_class)
 938 {
 939         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
 940             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 941             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 942             (mgmt_class == IB_MGMT_CLASS_BIS) ||
 943             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 944              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
 945                 return 1;
 946         return 0;
 947 }
 948 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
 949
 950 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
 951 {
 952         struct ib_mad_send_wr_private *mad_send_wr;
 953         struct list_head *list;
 954
 955         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 956                                    send_buf);
 957         list = &mad_send_wr->cur_seg->list;
 958
 959         if (mad_send_wr->cur_seg->num < seg_num) {
 960                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
 961                         if (mad_send_wr->cur_seg->num == seg_num)
 962                                 break;
 963         } else if (mad_send_wr->cur_seg->num > seg_num) {
 964                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
 965                         if (mad_send_wr->cur_seg->num == seg_num)
 966                                 break;
 967         }
 968         return mad_send_wr->cur_seg->data;
 969 }
 970 EXPORT_SYMBOL(ib_get_rmpp_segment);
 971
 972 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
 973 {
 974         if (mad_send_wr->send_buf.seg_count)
 975                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
 976                                            mad_send_wr->seg_num);
 977         else
 978                 return mad_send_wr->send_buf.mad +
 979                        mad_send_wr->send_buf.hdr_len;
 980 }
 981
 982 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
 983 {
 984         struct ib_mad_agent_private *mad_agent_priv;
 985         struct ib_mad_send_wr_private *mad_send_wr;
 986
 987         mad_agent_priv = container_of(send_buf->mad_agent,
 988                                       struct ib_mad_agent_private, agent);
 989         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 990                                    send_buf);
 991
 992         free_send_rmpp_list(mad_send_wr);
 993         kfree(send_buf->mad);
 994         deref_mad_agent(mad_agent_priv);
 995 }
 996 EXPORT_SYMBOL(ib_free_send_mad);
 997
 998 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 999 {
1000         struct ib_mad_qp_info *qp_info;
1001         struct list_head *list;
1002         struct ib_send_wr *bad_send_wr;
1003         struct ib_mad_agent *mad_agent;
1004         struct ib_sge *sge;
1005         unsigned long flags;
1006         int ret;
1007
1008         /* Set WR ID to find mad_send_wr upon completion */
1009         qp_info = mad_send_wr->mad_agent_priv->qp_info;
1010         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1011         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1012
1013         mad_agent = mad_send_wr->send_buf.mad_agent;
1014         sge = mad_send_wr->sg_list;
1015         sge[0].addr = ib_dma_map_single(mad_agent->device,
1016                                         mad_send_wr->send_buf.mad,
1017                                         sge[0].length,
1018                                         DMA_TO_DEVICE);
1019         mad_send_wr->header_mapping = sge[0].addr;
1020
1021         sge[1].addr = ib_dma_map_single(mad_agent->device,
1022                                         ib_get_payload(mad_send_wr),
1023                                         sge[1].length,
1024                                         DMA_TO_DEVICE);
1025         mad_send_wr->payload_mapping = sge[1].addr;
1026
1027         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1028         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1029                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1030                                    &bad_send_wr);
1031                 list = &qp_info->send_queue.list;
1032         } else {
1033                 ret = 0;
1034                 list = &qp_info->overflow_list;
1035         }
1036
1037         if (!ret) {
1038                 qp_info->send_queue.count++;
1039                 list_add_tail(&mad_send_wr->mad_list.list, list);
1040         }
1041         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1042         if (ret) {
1043                 ib_dma_unmap_single(mad_agent->device,
1044                                     mad_send_wr->header_mapping,
1045                                     sge[0].length, DMA_TO_DEVICE);
1046                 ib_dma_unmap_single(mad_agent->device,
1047                                     mad_send_wr->payload_mapping,
1048                                     sge[1].length, DMA_TO_DEVICE);
1049         }
1050         return ret;
1051 }
1052
1053 /*
1054  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1055  *  with the registered client
1056  */
1057 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1058                      struct ib_mad_send_buf **bad_send_buf)
1059 {
1060         struct ib_mad_agent_private *mad_agent_priv;
1061         struct ib_mad_send_buf *next_send_buf;
1062         struct ib_mad_send_wr_private *mad_send_wr;
1063         unsigned long flags;
1064         int ret = -EINVAL;
1065
1066         /* Walk list of send WRs and post each on send list */
1067         for (; send_buf; send_buf = next_send_buf) {
1068
1069                 mad_send_wr = container_of(send_buf,
1070                                            struct ib_mad_send_wr_private,
1071                                            send_buf);
1072                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1073
1074                 if (!send_buf->mad_agent->send_handler ||
1075                     (send_buf->timeout_ms &&
1076                      !send_buf->mad_agent->recv_handler)) {
1077                         ret = -EINVAL;
1078                         goto error;
1079                 }
1080
1081                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1082                         if (mad_agent_priv->agent.rmpp_version) {
1083                                 ret = -EINVAL;
1084                                 goto error;
1085                         }
1086                 }
1087
1088                 /*
1089                  * Save pointer to next work request to post in case the
1090                  * current one completes, and the user modifies the work
1091                  * request associated with the completion
1092                  */
1093                 next_send_buf = send_buf->next;
1094                 mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1095
1096                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1097                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1098                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1099                                                      mad_send_wr);
1100                         if (ret < 0)            /* error */
1101                                 goto error;
1102                         else if (ret == 1)      /* locally consumed */
1103                                 continue;
1104                 }
1105
1106                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1107                 /* Timeout will be updated after send completes */
1108                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1109                 mad_send_wr->max_retries = send_buf->retries;
1110                 mad_send_wr->retries_left = send_buf->retries;
1111                 send_buf->retries = 0;
1112                 /* Reference for work request to QP + response */
1113                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1114                 mad_send_wr->status = IB_WC_SUCCESS;
1115
1116                 /* Reference MAD agent until send completes */
1117                 atomic_inc(&mad_agent_priv->refcount);
1118                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1119                 list_add_tail(&mad_send_wr->agent_list,
1120                               &mad_agent_priv->send_list);
1121                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1122
1123                 if (mad_agent_priv->agent.rmpp_version) {
1124                         ret = ib_send_rmpp_mad(mad_send_wr);
1125                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1126                                 ret = ib_send_mad(mad_send_wr);
1127                 } else
1128                         ret = ib_send_mad(mad_send_wr);
1129                 if (ret < 0) {
1130                         /* Fail send request */
1131                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1132                         list_del(&mad_send_wr->agent_list);
1133                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1134                         atomic_dec(&mad_agent_priv->refcount);
1135                         goto error;
1136                 }
1137         }
1138         return 0;
1139 error:
1140         if (bad_send_buf)
1141                 *bad_send_buf = send_buf;
1142         return ret;
1143 }
1144 EXPORT_SYMBOL(ib_post_send_mad);
1145
1146 /*
1147  * ib_free_recv_mad - Returns data buffers used to receive
1148  *  a MAD to the access layer
1149  */
1150 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1151 {
1152         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1153         struct ib_mad_private_header *mad_priv_hdr;
1154         struct ib_mad_private *priv;
1155         struct list_head free_list;
1156
1157         INIT_LIST_HEAD(&free_list);
1158         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1159
1160         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1161                                         &free_list, list) {
1162                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1163                                            recv_buf);
1164                 mad_priv_hdr = container_of(mad_recv_wc,
1165                                             struct ib_mad_private_header,
1166                                             recv_wc);
1167                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1168                                     header);
1169                 kmem_cache_free(ib_mad_cache, priv);
1170         }
1171 }
1172 EXPORT_SYMBOL(ib_free_recv_mad);
1173
1174 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1175                                         u8 rmpp_version,
1176                                         ib_mad_send_handler send_handler,
1177                                         ib_mad_recv_handler recv_handler,
1178                                         void *context)
1179 {
1180         return ERR_PTR(-EINVAL);        /* XXX: for now */
1181 }
1182 EXPORT_SYMBOL(ib_redirect_mad_qp);
1183
1184 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1185                       struct ib_wc *wc)
1186 {
1187         printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1188         return 0;
1189 }
1190 EXPORT_SYMBOL(ib_process_mad_wc);
1191
1192 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1193                          struct ib_mad_reg_req *mad_reg_req)
1194 {
1195         int i;
1196
1197         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1198                 if ((*method)->agent[i]) {
1199                         printk(KERN_ERR PFX "Method %d already in use\n", i);
1200                         return -EINVAL;
1201                 }
1202         }
1203         return 0;
1204 }
1205
1206 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1207 {
1208         /* Allocate management method table */
1209         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1210         if (!*method) {
1211                 printk(KERN_ERR PFX "No memory for "
1212                        "ib_mad_mgmt_method_table\n");
1213                 return -ENOMEM;
1214         }
1215
1216         return 0;
1217 }
1218
1219 /*
1220  * Check to see if there are any methods still in use
1221  */
1222 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1223 {
1224         int i;
1225
1226         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1227                 if (method->agent[i])
1228                         return 1;
1229         return 0;
1230 }
1231
1232 /*
1233  * Check to see if there are any method tables for this class still in use
1234  */
1235 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1236 {
1237         int i;
1238
1239         for (i = 0; i < MAX_MGMT_CLASS; i++)
1240                 if (class->method_table[i])
1241                         return 1;
1242         return 0;
1243 }
1244
1245 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1246 {
1247         int i;
1248
1249         for (i = 0; i < MAX_MGMT_OUI; i++)
1250                 if (vendor_class->method_table[i])
1251                         return 1;
1252         return 0;
1253 }
1254
1255 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1256                            char *oui)
1257 {
1258         int i;
1259
1260         for (i = 0; i < MAX_MGMT_OUI; i++)
1261                 /* Is there matching OUI for this vendor class ? */
1262                 if (!memcmp(vendor_class->oui[i], oui, 3))
1263                         return i;
1264
1265         return -1;
1266 }
1267
1268 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1269 {
1270         int i;
1271
1272         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1273                 if (vendor->vendor_class[i])
1274                         return 1;
1275
1276         return 0;
1277 }
1278
1279 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1280                                      struct ib_mad_agent_private *agent)
1281 {
1282         int i;
1283
1284         /* Remove any methods for this mad agent */
1285         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1286                 if (method->agent[i] == agent) {
1287                         method->agent[i] = NULL;
1288                 }
1289         }
1290 }
1291
1292 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1293                               struct ib_mad_agent_private *agent_priv,
1294                               u8 mgmt_class)
1295 {
1296         struct ib_mad_port_private *port_priv;
1297         struct ib_mad_mgmt_class_table **class;
1298         struct ib_mad_mgmt_method_table **method;
1299         int i, ret;
1300
1301         port_priv = agent_priv->qp_info->port_priv;
1302         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1303         if (!*class) {
1304                 /* Allocate management class table for "new" class version */
1305                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1306                 if (!*class) {
1307                         printk(KERN_ERR PFX "No memory for "
1308                                "ib_mad_mgmt_class_table\n");
1309                         ret = -ENOMEM;
1310                         goto error1;
1311                 }
1312
1313                 /* Allocate method table for this management class */
1314                 method = &(*class)->method_table[mgmt_class];
1315                 if ((ret = allocate_method_table(method)))
1316                         goto error2;
1317         } else {
1318                 method = &(*class)->method_table[mgmt_class];
1319                 if (!*method) {
1320                         /* Allocate method table for this management class */
1321                         if ((ret = allocate_method_table(method)))
1322                                 goto error1;
1323                 }
1324         }
1325
1326         /* Now, make sure methods are not already in use */
1327         if (method_in_use(method, mad_reg_req))
1328                 goto error3;
1329
1330         /* Finally, add in methods being registered */
1331         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1332                 (*method)->agent[i] = agent_priv;
1333
1334         return 0;
1335
1336 error3:
1337         /* Remove any methods for this mad agent */
1338         remove_methods_mad_agent(*method, agent_priv);
1339         /* Now, check to see if there are any methods in use */
1340         if (!check_method_table(*method)) {
1341                 /* If not, release management method table */
1342                 kfree(*method);
1343                 *method = NULL;
1344         }
1345         ret = -EINVAL;
1346         goto error1;
1347 error2:
1348         kfree(*class);
1349         *class = NULL;
1350 error1:
1351         return ret;
1352 }
1353
1354 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1355                            struct ib_mad_agent_private *agent_priv)
1356 {
1357         struct ib_mad_port_private *port_priv;
1358         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1359         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1360         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1361         struct ib_mad_mgmt_method_table **method;
1362         int i, ret = -ENOMEM;
1363         u8 vclass;
1364
1365         /* "New" vendor (with OUI) class */
1366         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1367         port_priv = agent_priv->qp_info->port_priv;
1368         vendor_table = &port_priv->version[
1369                                 mad_reg_req->mgmt_class_version].vendor;
1370         if (!*vendor_table) {
1371                 /* Allocate mgmt vendor class table for "new" class version */
1372                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1373                 if (!vendor) {
1374                         printk(KERN_ERR PFX "No memory for "
1375                                "ib_mad_mgmt_vendor_class_table\n");
1376                         goto error1;
1377                 }
1378
1379                 *vendor_table = vendor;
1380         }
1381         if (!(*vendor_table)->vendor_class[vclass]) {
1382                 /* Allocate table for this management vendor class */
1383                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1384                 if (!vendor_class) {
1385                         printk(KERN_ERR PFX "No memory for "
1386                                "ib_mad_mgmt_vendor_class\n");
1387                         goto error2;
1388                 }
1389
1390                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1391         }
1392         for (i = 0; i < MAX_MGMT_OUI; i++) {
1393                 /* Is there matching OUI for this vendor class ? */
1394                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1395                             mad_reg_req->oui, 3)) {
1396                         method = &(*vendor_table)->vendor_class[
1397                                                 vclass]->method_table[i];
1398                         BUG_ON(!*method);
1399                         goto check_in_use;
1400                 }
1401         }
1402         for (i = 0; i < MAX_MGMT_OUI; i++) {
1403                 /* OUI slot available ? */
1404                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1405                                 vclass]->oui[i])) {
1406                         method = &(*vendor_table)->vendor_class[
1407                                 vclass]->method_table[i];
1408                         BUG_ON(*method);
1409                         /* Allocate method table for this OUI */
1410                         if ((ret = allocate_method_table(method)))
1411                                 goto error3;
1412                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1413                                mad_reg_req->oui, 3);
1414                         goto check_in_use;
1415                 }
1416         }
1417         printk(KERN_ERR PFX "All OUI slots in use\n");
1418         goto error3;
1419
1420 check_in_use:
1421         /* Now, make sure methods are not already in use */
1422         if (method_in_use(method, mad_reg_req))
1423                 goto error4;
1424
1425         /* Finally, add in methods being registered */
1426         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1427                 (*method)->agent[i] = agent_priv;
1428
1429         return 0;
1430
1431 error4:
1432         /* Remove any methods for this mad agent */
1433         remove_methods_mad_agent(*method, agent_priv);
1434         /* Now, check to see if there are any methods in use */
1435         if (!check_method_table(*method)) {
1436                 /* If not, release management method table */
1437                 kfree(*method);
1438                 *method = NULL;
1439         }
1440         ret = -EINVAL;
1441 error3:
1442         if (vendor_class) {
1443                 (*vendor_table)->vendor_class[vclass] = NULL;
1444                 kfree(vendor_class);
1445         }
1446 error2:
1447         if (vendor) {
1448                 *vendor_table = NULL;
1449                 kfree(vendor);
1450         }
1451 error1:
1452         return ret;
1453 }
1454
1455 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1456 {
1457         struct ib_mad_port_private *port_priv;
1458         struct ib_mad_mgmt_class_table *class;
1459         struct ib_mad_mgmt_method_table *method;
1460         struct ib_mad_mgmt_vendor_class_table *vendor;
1461         struct ib_mad_mgmt_vendor_class *vendor_class;
1462         int index;
1463         u8 mgmt_class;
1464
1465         /*
1466          * Was MAD registration request supplied
1467          * with original registration ?
1468          */
1469         if (!agent_priv->reg_req) {
1470                 goto out;
1471         }
1472
1473         port_priv = agent_priv->qp_info->port_priv;
1474         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1475         class = port_priv->version[
1476                         agent_priv->reg_req->mgmt_class_version].class;
1477         if (!class)
1478                 goto vendor_check;
1479
1480         method = class->method_table[mgmt_class];
1481         if (method) {
1482                 /* Remove any methods for this mad agent */
1483                 remove_methods_mad_agent(method, agent_priv);
1484                 /* Now, check to see if there are any methods still in use */
1485                 if (!check_method_table(method)) {
1486                         /* If not, release management method table */
1487                          kfree(method);
1488                          class->method_table[mgmt_class] = NULL;
1489                          /* Any management classes left ? */
1490                         if (!check_class_table(class)) {
1491                                 /* If not, release management class table */
1492                                 kfree(class);
1493                                 port_priv->version[
1494                                         agent_priv->reg_req->
1495                                         mgmt_class_version].class = NULL;
1496                         }
1497                 }
1498         }
1499
1500 vendor_check:
1501         if (!is_vendor_class(mgmt_class))
1502                 goto out;
1503
1504         /* normalize mgmt_class to vendor range 2 */
1505         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1506         vendor = port_priv->version[
1507                         agent_priv->reg_req->mgmt_class_version].vendor;
1508
1509         if (!vendor)
1510                 goto out;
1511
1512         vendor_class = vendor->vendor_class[mgmt_class];
1513         if (vendor_class) {
1514                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1515                 if (index < 0)
1516                         goto out;
1517                 method = vendor_class->method_table[index];
1518                 if (method) {
1519                         /* Remove any methods for this mad agent */
1520                         remove_methods_mad_agent(method, agent_priv);
1521                         /*
1522                          * Now, check to see if there are
1523                          * any methods still in use
1524                          */
1525                         if (!check_method_table(method)) {
1526                                 /* If not, release management method table */
1527                                 kfree(method);
1528                                 vendor_class->method_table[index] = NULL;
1529                                 memset(vendor_class->oui[index], 0, 3);
1530                                 /* Any OUIs left ? */
1531                                 if (!check_vendor_class(vendor_class)) {
1532                                         /* If not, release vendor class table */
1533                                         kfree(vendor_class);
1534                                         vendor->vendor_class[mgmt_class] = NULL;
1535                                         /* Any other vendor classes left ? */
1536                                         if (!check_vendor_table(vendor)) {
1537                                                 kfree(vendor);
1538                                                 port_priv->version[
1539                                                         agent_priv->reg_req->
1540                                                         mgmt_class_version].
1541                                                         vendor = NULL;
1542                                         }
1543                                 }
1544                         }
1545                 }
1546         }
1547
1548 out:
1549         return;
1550 }
1551
1552 static struct ib_mad_agent_private *
1553 find_mad_agent(struct ib_mad_port_private *port_priv,
1554                struct ib_mad *mad)
1555 {
1556         struct ib_mad_agent_private *mad_agent = NULL;
1557         unsigned long flags;
1558
1559         spin_lock_irqsave(&port_priv->reg_lock, flags);
1560         if (ib_response_mad(mad)) {
1561                 u32 hi_tid;
1562                 struct ib_mad_agent_private *entry;
1563
1564                 /*
1565                  * Routing is based on high 32 bits of transaction ID
1566                  * of MAD.
1567                  */
1568                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1569                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1570                         if (entry->agent.hi_tid == hi_tid) {
1571                                 mad_agent = entry;
1572                                 break;
1573                         }
1574                 }
1575         } else {
1576                 struct ib_mad_mgmt_class_table *class;
1577                 struct ib_mad_mgmt_method_table *method;
1578                 struct ib_mad_mgmt_vendor_class_table *vendor;
1579                 struct ib_mad_mgmt_vendor_class *vendor_class;
1580                 struct ib_vendor_mad *vendor_mad;
1581                 int index;
1582
1583                 /*
1584                  * Routing is based on version, class, and method
1585                  * For "newer" vendor MADs, also based on OUI
1586                  */
1587                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1588                         goto out;
1589                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1590                         class = port_priv->version[
1591                                         mad->mad_hdr.class_version].class;
1592                         if (!class)
1593                                 goto out;
1594                         method = class->method_table[convert_mgmt_class(
1595                                                         mad->mad_hdr.mgmt_class)];
1596                         if (method)
1597                                 mad_agent = method->agent[mad->mad_hdr.method &
1598                                                           ~IB_MGMT_METHOD_RESP];
1599                 } else {
1600                         vendor = port_priv->version[
1601                                         mad->mad_hdr.class_version].vendor;
1602                         if (!vendor)
1603                                 goto out;
1604                         vendor_class = vendor->vendor_class[vendor_class_index(
1605                                                 mad->mad_hdr.mgmt_class)];
1606                         if (!vendor_class)
1607                                 goto out;
1608                         /* Find matching OUI */
1609                         vendor_mad = (struct ib_vendor_mad *)mad;
1610                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1611                         if (index == -1)
1612                                 goto out;
1613                         method = vendor_class->method_table[index];
1614                         if (method) {
1615                                 mad_agent = method->agent[mad->mad_hdr.method &
1616                                                           ~IB_MGMT_METHOD_RESP];
1617                         }
1618                 }
1619         }
1620
1621         if (mad_agent) {
1622                 if (mad_agent->agent.recv_handler)
1623                         atomic_inc(&mad_agent->refcount);
1624                 else {
1625                         printk(KERN_NOTICE PFX "No receive handler for client "
1626                                "%p on port %d\n",
1627                                &mad_agent->agent, port_priv->port_num);
1628                         mad_agent = NULL;
1629                 }
1630         }
1631 out:
1632         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1633
1634         return mad_agent;
1635 }
1636
1637 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1638 {
1639         int valid = 0;
1640
1641         /* Make sure MAD base version is understood */
1642         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1643                 printk(KERN_ERR PFX "MAD received with unsupported base "
1644                        "version %d\n", mad->mad_hdr.base_version);
1645                 goto out;
1646         }
1647
1648         /* Filter SMI packets sent to other than QP0 */
1649         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1650             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1651                 if (qp_num == 0)
1652                         valid = 1;
1653         } else {
1654                 /* Filter GSI packets sent to QP0 */
1655                 if (qp_num != 0)
1656                         valid = 1;
1657         }
1658
1659 out:
1660         return valid;
1661 }
1662
1663 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1664                        struct ib_mad_hdr *mad_hdr)
1665 {
1666         struct ib_rmpp_mad *rmpp_mad;
1667
1668         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1669         return !mad_agent_priv->agent.rmpp_version ||
1670                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1671                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1672                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1673 }
1674
1675 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1676                                      struct ib_mad_recv_wc *rwc)
1677 {
1678         return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1679                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1680 }
1681
1682 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1683                                    struct ib_mad_send_wr_private *wr,
1684                                    struct ib_mad_recv_wc *rwc )
1685 {
1686         struct ib_ah_attr attr;
1687         u8 send_resp, rcv_resp;
1688         union ib_gid sgid;
1689         struct ib_device *device = mad_agent_priv->agent.device;
1690         u8 port_num = mad_agent_priv->agent.port_num;
1691         u8 lmc;
1692
1693         send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1694         rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1695
1696         if (send_resp == rcv_resp)
1697                 /* both requests, or both responses. GIDs different */
1698                 return 0;
1699
1700         if (ib_query_ah(wr->send_buf.ah, &attr))
1701                 /* Assume not equal, to avoid false positives. */
1702                 return 0;
1703
1704         if (!!(attr.ah_flags & IB_AH_GRH) !=
1705             !!(rwc->wc->wc_flags & IB_WC_GRH))
1706                 /* one has GID, other does not.  Assume different */
1707                 return 0;
1708
1709         if (!send_resp && rcv_resp) {
1710                 /* is request/response. */
1711                 if (!(attr.ah_flags & IB_AH_GRH)) {
1712                         if (ib_get_cached_lmc(device, port_num, &lmc))
1713                                 return 0;
1714                         return (!lmc || !((attr.src_path_bits ^
1715                                            rwc->wc->dlid_path_bits) &
1716                                           ((1 << lmc) - 1)));
1717                 } else {
1718                         if (ib_get_cached_gid(device, port_num,
1719                                               attr.grh.sgid_index, &sgid))
1720                                 return 0;
1721                         return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1722                                        16);
1723                 }
1724         }
1725
1726         if (!(attr.ah_flags & IB_AH_GRH))
1727                 return attr.dlid == rwc->wc->slid;
1728         else
1729                 return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1730                                16);
1731 }
1732
1733 static inline int is_direct(u8 class)
1734 {
1735         return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1736 }
1737
1738 struct ib_mad_send_wr_private*
1739 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1740                  struct ib_mad_recv_wc *wc)
1741 {
1742         struct ib_mad_send_wr_private *wr;
1743         struct ib_mad *mad;
1744
1745         mad = (struct ib_mad *)wc->recv_buf.mad;
1746
1747         list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1748                 if ((wr->tid == mad->mad_hdr.tid) &&
1749                     rcv_has_same_class(wr, wc) &&
1750                     /*
1751                      * Don't check GID for direct routed MADs.
1752                      * These might have permissive LIDs.
1753                      */
1754                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1755                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1756                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1757         }
1758
1759         /*
1760          * It's possible to receive the response before we've
1761          * been notified that the send has completed
1762          */
1763         list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1764                 if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1765                     wr->tid == mad->mad_hdr.tid &&
1766                     wr->timeout &&
1767                     rcv_has_same_class(wr, wc) &&
1768                     /*
1769                      * Don't check GID for direct routed MADs.
1770                      * These might have permissive LIDs.
1771                      */
1772                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1773                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1774                         /* Verify request has not been canceled */
1775                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1776         }
1777         return NULL;
1778 }
1779
1780 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1781 {
1782         mad_send_wr->timeout = 0;
1783         if (mad_send_wr->refcount == 1)
1784                 list_move_tail(&mad_send_wr->agent_list,
1785                               &mad_send_wr->mad_agent_priv->done_list);
1786 }
1787
1788 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1789                                  struct ib_mad_recv_wc *mad_recv_wc)
1790 {
1791         struct ib_mad_send_wr_private *mad_send_wr;
1792         struct ib_mad_send_wc mad_send_wc;
1793         unsigned long flags;
1794
1795         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1796         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1797         if (mad_agent_priv->agent.rmpp_version) {
1798                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1799                                                       mad_recv_wc);
1800                 if (!mad_recv_wc) {
1801                         deref_mad_agent(mad_agent_priv);
1802                         return;
1803                 }
1804         }
1805
1806         /* Complete corresponding request */
1807         if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1808                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1809                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1810                 if (!mad_send_wr) {
1811                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1812                         ib_free_recv_mad(mad_recv_wc);
1813                         deref_mad_agent(mad_agent_priv);
1814                         return;
1815                 }
1816                 ib_mark_mad_done(mad_send_wr);
1817                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1818
1819                 /* Defined behavior is to complete response before request */
1820                 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1821                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1822                                                    mad_recv_wc);
1823                 atomic_dec(&mad_agent_priv->refcount);
1824
1825                 mad_send_wc.status = IB_WC_SUCCESS;
1826                 mad_send_wc.vendor_err = 0;
1827                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1828                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1829         } else {
1830                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1831                                                    mad_recv_wc);
1832                 deref_mad_agent(mad_agent_priv);
1833         }
1834 }
1835
1836 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1837                                      struct ib_wc *wc)
1838 {
1839         struct ib_mad_qp_info *qp_info;
1840         struct ib_mad_private_header *mad_priv_hdr;
1841         struct ib_mad_private *recv, *response = NULL;
1842         struct ib_mad_list_head *mad_list;
1843         struct ib_mad_agent_private *mad_agent;
1844         int port_num;
1845
1846         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1847         qp_info = mad_list->mad_queue->qp_info;
1848         dequeue_mad(mad_list);
1849
1850         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1851                                     mad_list);
1852         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1853         ib_dma_unmap_single(port_priv->device,
1854                             recv->header.mapping,
1855                             sizeof(struct ib_mad_private) -
1856                               sizeof(struct ib_mad_private_header),
1857                             DMA_FROM_DEVICE);
1858
1859         /* Setup MAD receive work completion from "normal" work completion */
1860         recv->header.wc = *wc;
1861         recv->header.recv_wc.wc = &recv->header.wc;
1862         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1863         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1864         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1865
1866         if (atomic_read(&qp_info->snoop_count))
1867                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1868
1869         /* Validate MAD */
1870         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1871                 goto out;
1872
1873         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1874         if (!response) {
1875                 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1876                        "for response buffer\n");
1877                 goto out;
1878         }
1879
1880         if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1881                 port_num = wc->port_num;
1882         else
1883                 port_num = port_priv->port_num;
1884
1885         if (recv->mad.mad.mad_hdr.mgmt_class ==
1886             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1887                 enum smi_forward_action retsmi;
1888
1889                 if (smi_handle_dr_smp_recv(&recv->mad.smp,
1890                                            port_priv->device->node_type,
1891                                            port_num,
1892                                            port_priv->device->phys_port_cnt) ==
1893                                            IB_SMI_DISCARD)
1894                         goto out;
1895
1896                 retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
1897                 if (retsmi == IB_SMI_LOCAL)
1898                         goto local;
1899
1900                 if (retsmi == IB_SMI_SEND) { /* don't forward */
1901                         if (smi_handle_dr_smp_send(&recv->mad.smp,
1902                                                    port_priv->device->node_type,
1903                                                    port_num) == IB_SMI_DISCARD)
1904                                 goto out;
1905
1906                         if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1907                                 goto out;
1908                 } else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
1909                         /* forward case for switches */
1910                         memcpy(response, recv, sizeof(*response));
1911                         response->header.recv_wc.wc = &response->header.wc;
1912                         response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1913                         response->header.recv_wc.recv_buf.grh = &response->grh;
1914
1915                         agent_send_response(&response->mad.mad,
1916                                             &response->grh, wc,
1917                                             port_priv->device,
1918                                             smi_get_fwd_port(&recv->mad.smp),
1919                                             qp_info->qp->qp_num);
1920
1921                         goto out;
1922                 }
1923         }
1924
1925 local:
1926         /* Give driver "right of first refusal" on incoming MAD */
1927         if (port_priv->device->process_mad) {
1928                 int ret;
1929
1930                 ret = port_priv->device->process_mad(port_priv->device, 0,
1931                                                      port_priv->port_num,
1932                                                      wc, &recv->grh,
1933                                                      &recv->mad.mad,
1934                                                      &response->mad.mad);
1935                 if (ret & IB_MAD_RESULT_SUCCESS) {
1936                         if (ret & IB_MAD_RESULT_CONSUMED)
1937                                 goto out;
1938                         if (ret & IB_MAD_RESULT_REPLY) {
1939                                 agent_send_response(&response->mad.mad,
1940                                                     &recv->grh, wc,
1941                                                     port_priv->device,
1942                                                     port_num,
1943                                                     qp_info->qp->qp_num);
1944                                 goto out;
1945                         }
1946                 }
1947         }
1948
1949         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1950         if (mad_agent) {
1951                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1952                 /*
1953                  * recv is freed up in error cases in ib_mad_complete_recv
1954                  * or via recv_handler in ib_mad_complete_recv()
1955                  */
1956                 recv = NULL;
1957         }
1958
1959 out:
1960         /* Post another receive request for this QP */
1961         if (response) {
1962                 ib_mad_post_receive_mads(qp_info, response);
1963                 if (recv)
1964                         kmem_cache_free(ib_mad_cache, recv);
1965         } else
1966                 ib_mad_post_receive_mads(qp_info, recv);
1967 }
1968
1969 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1970 {
1971         struct ib_mad_send_wr_private *mad_send_wr;
1972         unsigned long delay;
1973
1974         if (list_empty(&mad_agent_priv->wait_list)) {
1975                 __cancel_delayed_work(&mad_agent_priv->timed_work);
1976         } else {
1977                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1978                                          struct ib_mad_send_wr_private,
1979                                          agent_list);
1980
1981                 if (time_after(mad_agent_priv->timeout,
1982                                mad_send_wr->timeout)) {
1983                         mad_agent_priv->timeout = mad_send_wr->timeout;
1984                         __cancel_delayed_work(&mad_agent_priv->timed_work);
1985                         delay = mad_send_wr->timeout - jiffies;
1986                         if ((long)delay <= 0)
1987                                 delay = 1;
1988                         queue_delayed_work(mad_agent_priv->qp_info->
1989                                            port_priv->wq,
1990                                            &mad_agent_priv->timed_work, delay);
1991                 }
1992         }
1993 }
1994
1995 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1996 {
1997         struct ib_mad_agent_private *mad_agent_priv;
1998         struct ib_mad_send_wr_private *temp_mad_send_wr;
1999         struct list_head *list_item;
2000         unsigned long delay;
2001
2002         mad_agent_priv = mad_send_wr->mad_agent_priv;
2003         list_del(&mad_send_wr->agent_list);
2004
2005         delay = mad_send_wr->timeout;
2006         mad_send_wr->timeout += jiffies;
2007
2008         if (delay) {
2009                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2010                         temp_mad_send_wr = list_entry(list_item,
2011                                                 struct ib_mad_send_wr_private,
2012                                                 agent_list);
2013                         if (time_after(mad_send_wr->timeout,
2014                                        temp_mad_send_wr->timeout))
2015                                 break;
2016                 }
2017         }
2018         else
2019                 list_item = &mad_agent_priv->wait_list;
2020         list_add(&mad_send_wr->agent_list, list_item);
2021
2022         /* Reschedule a work item if we have a shorter timeout */
2023         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
2024                 __cancel_delayed_work(&mad_agent_priv->timed_work);
2025                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2026                                    &mad_agent_priv->timed_work, delay);
2027         }
2028 }
2029
2030 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2031                           int timeout_ms)
2032 {
2033         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2034         wait_for_response(mad_send_wr);
2035 }
2036
2037 /*
2038  * Process a send work completion
2039  */
2040 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2041                              struct ib_mad_send_wc *mad_send_wc)
2042 {
2043         struct ib_mad_agent_private     *mad_agent_priv;
2044         unsigned long                   flags;
2045         int                             ret;
2046
2047         mad_agent_priv = mad_send_wr->mad_agent_priv;
2048         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2049         if (mad_agent_priv->agent.rmpp_version) {
2050                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2051                 if (ret == IB_RMPP_RESULT_CONSUMED)
2052                         goto done;
2053         } else
2054                 ret = IB_RMPP_RESULT_UNHANDLED;
2055
2056         if (mad_send_wc->status != IB_WC_SUCCESS &&
2057             mad_send_wr->status == IB_WC_SUCCESS) {
2058                 mad_send_wr->status = mad_send_wc->status;
2059                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2060         }
2061
2062         if (--mad_send_wr->refcount > 0) {
2063                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2064                     mad_send_wr->status == IB_WC_SUCCESS) {
2065                         wait_for_response(mad_send_wr);
2066                 }
2067                 goto done;
2068         }
2069
2070         /* Remove send from MAD agent and notify client of completion */
2071         list_del(&mad_send_wr->agent_list);
2072         adjust_timeout(mad_agent_priv);
2073         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2074
2075         if (mad_send_wr->status != IB_WC_SUCCESS )
2076                 mad_send_wc->status = mad_send_wr->status;
2077         if (ret == IB_RMPP_RESULT_INTERNAL)
2078                 ib_rmpp_send_handler(mad_send_wc);
2079         else
2080                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2081                                                    mad_send_wc);
2082
2083         /* Release reference on agent taken when sending */
2084         deref_mad_agent(mad_agent_priv);
2085         return;
2086 done:
2087         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2088 }
2089
2090 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2091                                      struct ib_wc *wc)
2092 {
2093         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2094         struct ib_mad_list_head         *mad_list;
2095         struct ib_mad_qp_info           *qp_info;
2096         struct ib_mad_queue             *send_queue;
2097         struct ib_send_wr               *bad_send_wr;
2098         struct ib_mad_send_wc           mad_send_wc;
2099         unsigned long flags;
2100         int ret;
2101
2102         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2103         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2104                                    mad_list);
2105         send_queue = mad_list->mad_queue;
2106         qp_info = send_queue->qp_info;
2107
2108 retry:
2109         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2110                             mad_send_wr->header_mapping,
2111                             mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2112         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2113                             mad_send_wr->payload_mapping,
2114                             mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2115         queued_send_wr = NULL;
2116         spin_lock_irqsave(&send_queue->lock, flags);
2117         list_del(&mad_list->list);
2118
2119         /* Move queued send to the send queue */
2120         if (send_queue->count-- > send_queue->max_active) {
2121                 mad_list = container_of(qp_info->overflow_list.next,
2122                                         struct ib_mad_list_head, list);
2123                 queued_send_wr = container_of(mad_list,
2124                                         struct ib_mad_send_wr_private,
2125                                         mad_list);
2126                 list_move_tail(&mad_list->list, &send_queue->list);
2127         }
2128         spin_unlock_irqrestore(&send_queue->lock, flags);
2129
2130         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2131         mad_send_wc.status = wc->status;
2132         mad_send_wc.vendor_err = wc->vendor_err;
2133         if (atomic_read(&qp_info->snoop_count))
2134                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2135                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2136         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2137
2138         if (queued_send_wr) {
2139                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2140                                    &bad_send_wr);
2141                 if (ret) {
2142                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2143                         mad_send_wr = queued_send_wr;
2144                         wc->status = IB_WC_LOC_QP_OP_ERR;
2145                         goto retry;
2146                 }
2147         }
2148 }
2149
2150 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2151 {
2152         struct ib_mad_send_wr_private *mad_send_wr;
2153         struct ib_mad_list_head *mad_list;
2154         unsigned long flags;
2155
2156         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2157         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2158                 mad_send_wr = container_of(mad_list,
2159                                            struct ib_mad_send_wr_private,
2160                                            mad_list);
2161                 mad_send_wr->retry = 1;
2162         }
2163         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2164 }
2165
2166 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2167                               struct ib_wc *wc)
2168 {
2169         struct ib_mad_list_head *mad_list;
2170         struct ib_mad_qp_info *qp_info;
2171         struct ib_mad_send_wr_private *mad_send_wr;
2172         int ret;
2173
2174         /* Determine if failure was a send or receive */
2175         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2176         qp_info = mad_list->mad_queue->qp_info;
2177         if (mad_list->mad_queue == &qp_info->recv_queue)
2178                 /*
2179                  * Receive errors indicate that the QP has entered the error
2180                  * state - error handling/shutdown code will cleanup
2181                  */
2182                 return;
2183
2184         /*
2185          * Send errors will transition the QP to SQE - move
2186          * QP to RTS and repost flushed work requests
2187          */
2188         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2189                                    mad_list);
2190         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2191                 if (mad_send_wr->retry) {
2192                         /* Repost send */
2193                         struct ib_send_wr *bad_send_wr;
2194
2195                         mad_send_wr->retry = 0;
2196                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2197                                         &bad_send_wr);
2198                         if (ret)
2199                                 ib_mad_send_done_handler(port_priv, wc);
2200                 } else
2201                         ib_mad_send_done_handler(port_priv, wc);
2202         } else {
2203                 struct ib_qp_attr *attr;
2204
2205                 /* Transition QP to RTS and fail offending send */
2206                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2207                 if (attr) {
2208                         attr->qp_state = IB_QPS_RTS;
2209                         attr->cur_qp_state = IB_QPS_SQE;
2210                         ret = ib_modify_qp(qp_info->qp, attr,
2211                                            IB_QP_STATE | IB_QP_CUR_STATE);
2212                         kfree(attr);
2213                         if (ret)
2214                                 printk(KERN_ERR PFX "mad_error_handler - "
2215                                        "ib_modify_qp to RTS : %d\n", ret);
2216                         else
2217                                 mark_sends_for_retry(qp_info);
2218                 }
2219                 ib_mad_send_done_handler(port_priv, wc);
2220         }
2221 }
2222
2223 /*
2224  * IB MAD completion callback
2225  */
2226 static void ib_mad_completion_handler(struct work_struct *work)
2227 {
2228         struct ib_mad_port_private *port_priv;
2229         struct ib_wc wc;
2230
2231         port_priv = container_of(work, struct ib_mad_port_private, work);
2232         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2233
2234         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2235                 if (wc.status == IB_WC_SUCCESS) {
2236                         switch (wc.opcode) {
2237                         case IB_WC_SEND:
2238                                 ib_mad_send_done_handler(port_priv, &wc);
2239                                 break;
2240                         case IB_WC_RECV:
2241                                 ib_mad_recv_done_handler(port_priv, &wc);
2242                                 break;
2243                         default:
2244                                 BUG_ON(1);
2245                                 break;
2246                         }
2247                 } else
2248                         mad_error_handler(port_priv, &wc);
2249         }
2250 }
2251
2252 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2253 {
2254         unsigned long flags;
2255         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2256         struct ib_mad_send_wc mad_send_wc;
2257         struct list_head cancel_list;
2258
2259         INIT_LIST_HEAD(&cancel_list);
2260
2261         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2262         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2263                                  &mad_agent_priv->send_list, agent_list) {
2264                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2265                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2266                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2267                 }
2268         }
2269
2270         /* Empty wait list to prevent receives from finding a request */
2271         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2272         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2273
2274         /* Report all cancelled requests */
2275         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2276         mad_send_wc.vendor_err = 0;
2277
2278         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2279                                  &cancel_list, agent_list) {
2280                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2281                 list_del(&mad_send_wr->agent_list);
2282                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2283                                                    &mad_send_wc);
2284                 atomic_dec(&mad_agent_priv->refcount);
2285         }
2286 }
2287
2288 static struct ib_mad_send_wr_private*
2289 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2290              struct ib_mad_send_buf *send_buf)
2291 {
2292         struct ib_mad_send_wr_private *mad_send_wr;
2293
2294         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2295                             agent_list) {
2296                 if (&mad_send_wr->send_buf == send_buf)
2297                         return mad_send_wr;
2298         }
2299
2300         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2301                             agent_list) {
2302                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2303                     &mad_send_wr->send_buf == send_buf)
2304                         return mad_send_wr;
2305         }
2306         return NULL;
2307 }
2308
2309 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2310                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2311 {
2312         struct ib_mad_agent_private *mad_agent_priv;
2313         struct ib_mad_send_wr_private *mad_send_wr;
2314         unsigned long flags;
2315         int active;
2316
2317         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2318                                       agent);
2319         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2320         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2321         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2322                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2323                 return -EINVAL;
2324         }
2325
2326         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2327         if (!timeout_ms) {
2328                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2329                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2330         }
2331
2332         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2333         if (active)
2334                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2335         else
2336                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2337
2338         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2339         return 0;
2340 }
2341 EXPORT_SYMBOL(ib_modify_mad);
2342
2343 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2344                    struct ib_mad_send_buf *send_buf)
2345 {
2346         ib_modify_mad(mad_agent, send_buf, 0);
2347 }
2348 EXPORT_SYMBOL(ib_cancel_mad);
2349
2350 static void local_completions(struct work_struct *work)
2351 {
2352         struct ib_mad_agent_private *mad_agent_priv;
2353         struct ib_mad_local_private *local;
2354         struct ib_mad_agent_private *recv_mad_agent;
2355         unsigned long flags;
2356         int free_mad;
2357         struct ib_wc wc;
2358         struct ib_mad_send_wc mad_send_wc;
2359
2360         mad_agent_priv =
2361                 container_of(work, struct ib_mad_agent_private, local_work);
2362
2363         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2364         while (!list_empty(&mad_agent_priv->local_list)) {
2365                 local = list_entry(mad_agent_priv->local_list.next,
2366                                    struct ib_mad_local_private,
2367                                    completion_list);
2368                 list_del(&local->completion_list);
2369                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2370                 free_mad = 0;
2371                 if (local->mad_priv) {
2372                         recv_mad_agent = local->recv_mad_agent;
2373                         if (!recv_mad_agent) {
2374                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2375                                 free_mad = 1;
2376                                 goto local_send_completion;
2377                         }
2378
2379                         /*
2380                          * Defined behavior is to complete response
2381                          * before request
2382                          */
2383                         build_smp_wc(recv_mad_agent->agent.qp,
2384                                      (unsigned long) local->mad_send_wr,
2385                                      be16_to_cpu(IB_LID_PERMISSIVE),
2386                                      0, recv_mad_agent->agent.port_num, &wc);
2387
2388                         local->mad_priv->header.recv_wc.wc = &wc;
2389                         local->mad_priv->header.recv_wc.mad_len =
2390                                                 sizeof(struct ib_mad);
2391                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2392                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2393                                  &local->mad_priv->header.recv_wc.rmpp_list);
2394                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2395                         local->mad_priv->header.recv_wc.recv_buf.mad =
2396                                                 &local->mad_priv->mad.mad;
2397                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2398                                 snoop_recv(recv_mad_agent->qp_info,
2399                                           &local->mad_priv->header.recv_wc,
2400                                            IB_MAD_SNOOP_RECVS);
2401                         recv_mad_agent->agent.recv_handler(
2402                                                 &recv_mad_agent->agent,
2403                                                 &local->mad_priv->header.recv_wc);
2404                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2405                         atomic_dec(&recv_mad_agent->refcount);
2406                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2407                 }
2408
2409 local_send_completion:
2410                 /* Complete send */
2411                 mad_send_wc.status = IB_WC_SUCCESS;
2412                 mad_send_wc.vendor_err = 0;
2413                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2414                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2415                         snoop_send(mad_agent_priv->qp_info,
2416                                    &local->mad_send_wr->send_buf,
2417                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2418                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2419                                                    &mad_send_wc);
2420
2421                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2422                 atomic_dec(&mad_agent_priv->refcount);
2423                 if (free_mad)
2424                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2425                 kfree(local);
2426         }
2427         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2428 }
2429
2430 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2431 {
2432         int ret;
2433
2434         if (!mad_send_wr->retries_left)
2435                 return -ETIMEDOUT;
2436
2437         mad_send_wr->retries_left--;
2438         mad_send_wr->send_buf.retries++;
2439
2440         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2441
2442         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2443                 ret = ib_retry_rmpp(mad_send_wr);
2444                 switch (ret) {
2445                 case IB_RMPP_RESULT_UNHANDLED:
2446                         ret = ib_send_mad(mad_send_wr);
2447                         break;
2448                 case IB_RMPP_RESULT_CONSUMED:
2449                         ret = 0;
2450                         break;
2451                 default:
2452                         ret = -ECOMM;
2453                         break;
2454                 }
2455         } else
2456                 ret = ib_send_mad(mad_send_wr);
2457
2458         if (!ret) {
2459                 mad_send_wr->refcount++;
2460                 list_add_tail(&mad_send_wr->agent_list,
2461                               &mad_send_wr->mad_agent_priv->send_list);
2462         }
2463         return ret;
2464 }
2465
2466 static void timeout_sends(struct work_struct *work)
2467 {
2468         struct ib_mad_agent_private *mad_agent_priv;
2469         struct ib_mad_send_wr_private *mad_send_wr;
2470         struct ib_mad_send_wc mad_send_wc;
2471         unsigned long flags, delay;
2472
2473         mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2474                                       timed_work.work);
2475         mad_send_wc.vendor_err = 0;
2476
2477         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2478         while (!list_empty(&mad_agent_priv->wait_list)) {
2479                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2480                                          struct ib_mad_send_wr_private,
2481                                          agent_list);
2482
2483                 if (time_after(mad_send_wr->timeout, jiffies)) {
2484                         delay = mad_send_wr->timeout - jiffies;
2485                         if ((long)delay <= 0)
2486                                 delay = 1;
2487                         queue_delayed_work(mad_agent_priv->qp_info->
2488                                            port_priv->wq,
2489                                            &mad_agent_priv->timed_work, delay);
2490                         break;
2491                 }
2492
2493                 list_del(&mad_send_wr->agent_list);
2494                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2495                     !retry_send(mad_send_wr))
2496                         continue;
2497
2498                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2499
2500                 if (mad_send_wr->status == IB_WC_SUCCESS)
2501                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2502                 else
2503                         mad_send_wc.status = mad_send_wr->status;
2504                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2505                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2506                                                    &mad_send_wc);
2507
2508                 atomic_dec(&mad_agent_priv->refcount);
2509                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2510         }
2511         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2512 }
2513
2514 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2515 {
2516         struct ib_mad_port_private *port_priv = cq->cq_context;
2517         unsigned long flags;
2518
2519         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2520         if (!list_empty(&port_priv->port_list))
2521                 queue_work(port_priv->wq, &port_priv->work);
2522         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2523 }
2524
2525 /*
2526  * Allocate receive MADs and post receive WRs for them
2527  */
2528 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2529                                     struct ib_mad_private *mad)
2530 {
2531         unsigned long flags;
2532         int post, ret;
2533         struct ib_mad_private *mad_priv;
2534         struct ib_sge sg_list;
2535         struct ib_recv_wr recv_wr, *bad_recv_wr;
2536         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2537
2538         /* Initialize common scatter list fields */
2539         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2540         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2541
2542         /* Initialize common receive WR fields */
2543         recv_wr.next = NULL;
2544         recv_wr.sg_list = &sg_list;
2545         recv_wr.num_sge = 1;
2546
2547         do {
2548                 /* Allocate and map receive buffer */
2549                 if (mad) {
2550                         mad_priv = mad;
2551                         mad = NULL;
2552                 } else {
2553                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2554                         if (!mad_priv) {
2555                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2556                                 ret = -ENOMEM;
2557                                 break;
2558                         }
2559                 }
2560                 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2561                                                  &mad_priv->grh,
2562                                                  sizeof *mad_priv -
2563                                                    sizeof mad_priv->header,
2564                                                  DMA_FROM_DEVICE);
2565                 mad_priv->header.mapping = sg_list.addr;
2566                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2567                 mad_priv->header.mad_list.mad_queue = recv_queue;
2568
2569                 /* Post receive WR */
2570                 spin_lock_irqsave(&recv_queue->lock, flags);
2571                 post = (++recv_queue->count < recv_queue->max_active);
2572                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2573                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2574                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2575                 if (ret) {
2576                         spin_lock_irqsave(&recv_queue->lock, flags);
2577                         list_del(&mad_priv->header.mad_list.list);
2578                         recv_queue->count--;
2579                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2580                         ib_dma_unmap_single(qp_info->port_priv->device,
2581                                             mad_priv->header.mapping,
2582                                             sizeof *mad_priv -
2583                                               sizeof mad_priv->header,
2584                                             DMA_FROM_DEVICE);
2585                         kmem_cache_free(ib_mad_cache, mad_priv);
2586                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2587                         break;
2588                 }
2589         } while (post);
2590
2591         return ret;
2592 }
2593
2594 /*
2595  * Return all the posted receive MADs
2596  */
2597 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2598 {
2599         struct ib_mad_private_header *mad_priv_hdr;
2600         struct ib_mad_private *recv;
2601         struct ib_mad_list_head *mad_list;
2602
2603         while (!list_empty(&qp_info->recv_queue.list)) {
2604
2605                 mad_list = list_entry(qp_info->recv_queue.list.next,
2606                                       struct ib_mad_list_head, list);
2607                 mad_priv_hdr = container_of(mad_list,
2608                                             struct ib_mad_private_header,
2609                                             mad_list);
2610                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2611                                     header);
2612
2613                 /* Remove from posted receive MAD list */
2614                 list_del(&mad_list->list);
2615
2616                 ib_dma_unmap_single(qp_info->port_priv->device,
2617                                     recv->header.mapping,
2618                                     sizeof(struct ib_mad_private) -
2619                                       sizeof(struct ib_mad_private_header),
2620                                     DMA_FROM_DEVICE);
2621                 kmem_cache_free(ib_mad_cache, recv);
2622         }
2623
2624         qp_info->recv_queue.count = 0;
2625 }
2626
2627 /*
2628  * Start the port
2629  */
2630 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2631 {
2632         int ret, i;
2633         struct ib_qp_attr *attr;
2634         struct ib_qp *qp;
2635
2636         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2637         if (!attr) {
2638                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2639                 return -ENOMEM;
2640         }
2641
2642         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2643                 qp = port_priv->qp_info[i].qp;
2644                 /*
2645                  * PKey index for QP1 is irrelevant but
2646                  * one is needed for the Reset to Init transition
2647                  */
2648                 attr->qp_state = IB_QPS_INIT;
2649                 attr->pkey_index = 0;
2650                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2651                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2652                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2653                 if (ret) {
2654                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2655                                "INIT: %d\n", i, ret);
2656                         goto out;
2657                 }
2658
2659                 attr->qp_state = IB_QPS_RTR;
2660                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2661                 if (ret) {
2662                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2663                                "RTR: %d\n", i, ret);
2664                         goto out;
2665                 }
2666
2667                 attr->qp_state = IB_QPS_RTS;
2668                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2669                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2670                 if (ret) {
2671                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2672                                "RTS: %d\n", i, ret);
2673                         goto out;
2674                 }
2675         }
2676
2677         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2678         if (ret) {
2679                 printk(KERN_ERR PFX "Failed to request completion "
2680                        "notification: %d\n", ret);
2681                 goto out;
2682         }
2683
2684         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2685                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2686                 if (ret) {
2687                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2688                         goto out;
2689                 }
2690         }
2691 out:
2692         kfree(attr);
2693         return ret;
2694 }
2695
2696 static void qp_event_handler(struct ib_event *event, void *qp_context)
2697 {
2698         struct ib_mad_qp_info   *qp_info = qp_context;
2699
2700         /* It's worse than that! He's dead, Jim! */
2701         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2702                 event->event, qp_info->qp->qp_num);
2703 }
2704
2705 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2706                            struct ib_mad_queue *mad_queue)
2707 {
2708         mad_queue->qp_info = qp_info;
2709         mad_queue->count = 0;
2710         spin_lock_init(&mad_queue->lock);
2711         INIT_LIST_HEAD(&mad_queue->list);
2712 }
2713
2714 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2715                         struct ib_mad_qp_info *qp_info)
2716 {
2717         qp_info->port_priv = port_priv;
2718         init_mad_queue(qp_info, &qp_info->send_queue);
2719         init_mad_queue(qp_info, &qp_info->recv_queue);
2720         INIT_LIST_HEAD(&qp_info->overflow_list);
2721         spin_lock_init(&qp_info->snoop_lock);
2722         qp_info->snoop_table = NULL;
2723         qp_info->snoop_table_size = 0;
2724         atomic_set(&qp_info->snoop_count, 0);
2725 }
2726
2727 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2728                          enum ib_qp_type qp_type)
2729 {
2730         struct ib_qp_init_attr  qp_init_attr;
2731         int ret;
2732
2733         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2734         qp_init_attr.send_cq = qp_info->port_priv->cq;
2735         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2736         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2737         qp_init_attr.cap.max_send_wr = mad_sendq_size;
2738         qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2739         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2740         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2741         qp_init_attr.qp_type = qp_type;
2742         qp_init_attr.port_num = qp_info->port_priv->port_num;
2743         qp_init_attr.qp_context = qp_info;
2744         qp_init_attr.event_handler = qp_event_handler;
2745         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2746         if (IS_ERR(qp_info->qp)) {
2747                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2748                        get_spl_qp_index(qp_type));
2749                 ret = PTR_ERR(qp_info->qp);
2750                 goto error;
2751         }
2752         /* Use minimum queue sizes unless the CQ is resized */
2753         qp_info->send_queue.max_active = mad_sendq_size;
2754         qp_info->recv_queue.max_active = mad_recvq_size;
2755         return 0;
2756
2757 error:
2758         return ret;
2759 }
2760
2761 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2762 {
2763         ib_destroy_qp(qp_info->qp);
2764         kfree(qp_info->snoop_table);
2765 }
2766
2767 /*
2768  * Open the port
2769  * Create the QP, PD, MR, and CQ if needed
2770  */
2771 static int ib_mad_port_open(struct ib_device *device,
2772                             int port_num)
2773 {
2774         int ret, cq_size;
2775         struct ib_mad_port_private *port_priv;
2776         unsigned long flags;
2777         char name[sizeof "ib_mad123"];
2778
2779         /* Create new device info */
2780         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2781         if (!port_priv) {
2782                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2783                 return -ENOMEM;
2784         }
2785
2786         port_priv->device = device;
2787         port_priv->port_num = port_num;
2788         spin_lock_init(&port_priv->reg_lock);
2789         INIT_LIST_HEAD(&port_priv->agent_list);
2790         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2791         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2792
2793         cq_size = (mad_sendq_size + mad_recvq_size) * 2;
2794         port_priv->cq = ib_create_cq(port_priv->device,
2795                                      ib_mad_thread_completion_handler,
2796                                      NULL, port_priv, cq_size, 0);
2797         if (IS_ERR(port_priv->cq)) {
2798                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2799                 ret = PTR_ERR(port_priv->cq);
2800                 goto error3;
2801         }
2802
2803         port_priv->pd = ib_alloc_pd(device);
2804         if (IS_ERR(port_priv->pd)) {
2805                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2806                 ret = PTR_ERR(port_priv->pd);
2807                 goto error4;
2808         }
2809
2810         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2811         if (IS_ERR(port_priv->mr)) {
2812                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2813                 ret = PTR_ERR(port_priv->mr);
2814                 goto error5;
2815         }
2816
2817         ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2818         if (ret)
2819                 goto error6;
2820         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2821         if (ret)
2822                 goto error7;
2823
2824         snprintf(name, sizeof name, "ib_mad%d", port_num);
2825         port_priv->wq = create_singlethread_workqueue(name);
2826         if (!port_priv->wq) {
2827                 ret = -ENOMEM;
2828                 goto error8;
2829         }
2830         INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2831
2832         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2833         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2834         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2835
2836         ret = ib_mad_port_start(port_priv);
2837         if (ret) {
2838                 printk(KERN_ERR PFX "Couldn't start port\n");
2839                 goto error9;
2840         }
2841
2842         return 0;
2843
2844 error9:
2845         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2846         list_del_init(&port_priv->port_list);
2847         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2848
2849         destroy_workqueue(port_priv->wq);
2850 error8:
2851         destroy_mad_qp(&port_priv->qp_info[1]);
2852 error7:
2853         destroy_mad_qp(&port_priv->qp_info[0]);
2854 error6:
2855         ib_dereg_mr(port_priv->mr);
2856 error5:
2857         ib_dealloc_pd(port_priv->pd);
2858 error4:
2859         ib_destroy_cq(port_priv->cq);
2860         cleanup_recv_queue(&port_priv->qp_info[1]);
2861         cleanup_recv_queue(&port_priv->qp_info[0]);
2862 error3:
2863         kfree(port_priv);
2864
2865         return ret;
2866 }
2867
2868 /*
2869  * Close the port
2870  * If there are no classes using the port, free the port
2871  * resources (CQ, MR, PD, QP) and remove the port's info structure
2872  */
2873 static int ib_mad_port_close(struct ib_device *device, int port_num)
2874 {
2875         struct ib_mad_port_private *port_priv;
2876         unsigned long flags;
2877
2878         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2879         port_priv = __ib_get_mad_port(device, port_num);
2880         if (port_priv == NULL) {
2881                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2882                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2883                 return -ENODEV;
2884         }
2885         list_del_init(&port_priv->port_list);
2886         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2887
2888         destroy_workqueue(port_priv->wq);
2889         destroy_mad_qp(&port_priv->qp_info[1]);
2890         destroy_mad_qp(&port_priv->qp_info[0]);
2891         ib_dereg_mr(port_priv->mr);
2892         ib_dealloc_pd(port_priv->pd);
2893         ib_destroy_cq(port_priv->cq);
2894         cleanup_recv_queue(&port_priv->qp_info[1]);
2895         cleanup_recv_queue(&port_priv->qp_info[0]);
2896         /* XXX: Handle deallocation of MAD registration tables */
2897
2898         kfree(port_priv);
2899
2900         return 0;
2901 }
2902
2903 static void ib_mad_init_device(struct ib_device *device)
2904 {
2905         int start, end, i;
2906
2907         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2908                 return;
2909
2910         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2911                 start = 0;
2912                 end   = 0;
2913         } else {
2914                 start = 1;
2915                 end   = device->phys_port_cnt;
2916         }
2917
2918         for (i = start; i <= end; i++) {
2919                 if (ib_mad_port_open(device, i)) {
2920                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2921                                device->name, i);
2922                         goto error;
2923                 }
2924                 if (ib_agent_port_open(device, i)) {
2925                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2926                                "for agents\n",
2927                                device->name, i);
2928                         goto error_agent;
2929                 }
2930         }
2931         return;
2932
2933 error_agent:
2934         if (ib_mad_port_close(device, i))
2935                 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2936                        device->name, i);
2937
2938 error:
2939         i--;
2940
2941         while (i >= start) {
2942                 if (ib_agent_port_close(device, i))
2943                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2944                                "for agents\n",
2945                                device->name, i);
2946                 if (ib_mad_port_close(device, i))
2947                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2948                                device->name, i);
2949                 i--;
2950         }
2951 }
2952
2953 static void ib_mad_remove_device(struct ib_device *device)
2954 {
2955         int i, num_ports, cur_port;
2956
2957         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2958                 return;
2959
2960         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2961                 num_ports = 1;
2962                 cur_port = 0;
2963         } else {
2964                 num_ports = device->phys_port_cnt;
2965                 cur_port = 1;
2966         }
2967         for (i = 0; i < num_ports; i++, cur_port++) {
2968                 if (ib_agent_port_close(device, cur_port))
2969                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2970                                "for agents\n",
2971                                device->name, cur_port);
2972                 if (ib_mad_port_close(device, cur_port))
2973                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2974                                device->name, cur_port);
2975         }
2976 }
2977
2978 static struct ib_client mad_client = {
2979         .name   = "mad",
2980         .add = ib_mad_init_device,
2981         .remove = ib_mad_remove_device
2982 };
2983
2984 static int __init ib_mad_init_module(void)
2985 {
2986         int ret;
2987
2988         mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
2989         mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
2990
2991         mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
2992         mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
2993
2994         ib_mad_cache = kmem_cache_create("ib_mad",
2995                                          sizeof(struct ib_mad_private),
2996                                          0,
2997                                          SLAB_HWCACHE_ALIGN,
2998                                          NULL);
2999         if (!ib_mad_cache) {
3000                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
3001                 ret = -ENOMEM;
3002                 goto error1;
3003         }
3004
3005         INIT_LIST_HEAD(&ib_mad_port_list);
3006
3007         if (ib_register_client(&mad_client)) {
3008                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
3009                 ret = -EINVAL;
3010                 goto error2;
3011         }
3012
3013         return 0;
3014
3015 error2:
3016         kmem_cache_destroy(ib_mad_cache);
3017 error1:
3018         return ret;
3019 }
3020
3021 static void __exit ib_mad_cleanup_module(void)
3022 {
3023         ib_unregister_client(&mad_client);
3024         kmem_cache_destroy(ib_mad_cache);
3025 }
3026
3027 module_init(ib_mad_init_module);
3028 module_exit(ib_mad_cleanup_module);