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 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
  51 static 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 = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
 295                 if (!reg_req) {
 296                         ret = ERR_PTR(-ENOMEM);
 297                         goto error3;
 298                 }
 299         }
 300
 301         /* Now, fill in the various structures */
 302         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 303         mad_agent_priv->reg_req = reg_req;
 304         mad_agent_priv->agent.rmpp_version = rmpp_version;
 305         mad_agent_priv->agent.device = device;
 306         mad_agent_priv->agent.recv_handler = recv_handler;
 307         mad_agent_priv->agent.send_handler = send_handler;
 308         mad_agent_priv->agent.context = context;
 309         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 310         mad_agent_priv->agent.port_num = port_num;
 311         spin_lock_init(&mad_agent_priv->lock);
 312         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 313         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 314         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 315         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 316         INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
 317         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 318         INIT_WORK(&mad_agent_priv->local_work, local_completions);
 319         atomic_set(&mad_agent_priv->refcount, 1);
 320         init_completion(&mad_agent_priv->comp);
 321
 322         spin_lock_irqsave(&port_priv->reg_lock, flags);
 323         mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
 324
 325         /*
 326          * Make sure MAD registration (if supplied)
 327          * is non overlapping with any existing ones
 328          */
 329         if (mad_reg_req) {
 330                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 331                 if (!is_vendor_class(mgmt_class)) {
 332                         class = port_priv->version[mad_reg_req->
 333                                                    mgmt_class_version].class;
 334                         if (class) {
 335                                 method = class->method_table[mgmt_class];
 336                                 if (method) {
 337                                         if (method_in_use(&method,
 338                                                            mad_reg_req))
 339                                                 goto error4;
 340                                 }
 341                         }
 342                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 343                                                   mgmt_class);
 344                 } else {
 345                         /* "New" vendor class range */
 346                         vendor = port_priv->version[mad_reg_req->
 347                                                     mgmt_class_version].vendor;
 348                         if (vendor) {
 349                                 vclass = vendor_class_index(mgmt_class);
 350                                 vendor_class = vendor->vendor_class[vclass];
 351                                 if (vendor_class) {
 352                                         if (is_vendor_method_in_use(
 353                                                         vendor_class,
 354                                                         mad_reg_req))
 355                                                 goto error4;
 356                                 }
 357                         }
 358                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 359                 }
 360                 if (ret2) {
 361                         ret = ERR_PTR(ret2);
 362                         goto error4;
 363                 }
 364         }
 365
 366         /* Add mad agent into port's agent list */
 367         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
 368         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 369
 370         return &mad_agent_priv->agent;
 371
 372 error4:
 373         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 374         kfree(reg_req);
 375 error3:
 376         ib_dereg_mr(mad_agent_priv->agent.mr);
 377 error2:
 378         kfree(mad_agent_priv);
 379 error1:
 380         return ret;
 381 }
 382 EXPORT_SYMBOL(ib_register_mad_agent);
 383
 384 static inline int is_snooping_sends(int mad_snoop_flags)
 385 {
 386         return (mad_snoop_flags &
 387                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 388                  IB_MAD_SNOOP_RMPP_SENDS |*/
 389                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 390                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 391 }
 392
 393 static inline int is_snooping_recvs(int mad_snoop_flags)
 394 {
 395         return (mad_snoop_flags &
 396                 (IB_MAD_SNOOP_RECVS /*|
 397                  IB_MAD_SNOOP_RMPP_RECVS*/));
 398 }
 399
 400 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 401                                 struct ib_mad_snoop_private *mad_snoop_priv)
 402 {
 403         struct ib_mad_snoop_private **new_snoop_table;
 404         unsigned long flags;
 405         int i;
 406
 407         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 408         /* Check for empty slot in array. */
 409         for (i = 0; i < qp_info->snoop_table_size; i++)
 410                 if (!qp_info->snoop_table[i])
 411                         break;
 412
 413         if (i == qp_info->snoop_table_size) {
 414                 /* Grow table. */
 415                 new_snoop_table = krealloc(qp_info->snoop_table,
 416                                            sizeof mad_snoop_priv *
 417                                            (qp_info->snoop_table_size + 1),
 418                                            GFP_ATOMIC);
 419                 if (!new_snoop_table) {
 420                         i = -ENOMEM;
 421                         goto out;
 422                 }
 423
 424                 qp_info->snoop_table = new_snoop_table;
 425                 qp_info->snoop_table_size++;
 426         }
 427         qp_info->snoop_table[i] = mad_snoop_priv;
 428         atomic_inc(&qp_info->snoop_count);
 429 out:
 430         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 431         return i;
 432 }
 433
 434 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 435                                            u8 port_num,
 436                                            enum ib_qp_type qp_type,
 437                                            int mad_snoop_flags,
 438                                            ib_mad_snoop_handler snoop_handler,
 439                                            ib_mad_recv_handler recv_handler,
 440                                            void *context)
 441 {
 442         struct ib_mad_port_private *port_priv;
 443         struct ib_mad_agent *ret;
 444         struct ib_mad_snoop_private *mad_snoop_priv;
 445         int qpn;
 446
 447         /* Validate parameters */
 448         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 449             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 450                 ret = ERR_PTR(-EINVAL);
 451                 goto error1;
 452         }
 453         qpn = get_spl_qp_index(qp_type);
 454         if (qpn == -1) {
 455                 ret = ERR_PTR(-EINVAL);
 456                 goto error1;
 457         }
 458         port_priv = ib_get_mad_port(device, port_num);
 459         if (!port_priv) {
 460                 ret = ERR_PTR(-ENODEV);
 461                 goto error1;
 462         }
 463         /* Allocate structures */
 464         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 465         if (!mad_snoop_priv) {
 466                 ret = ERR_PTR(-ENOMEM);
 467                 goto error1;
 468         }
 469
 470         /* Now, fill in the various structures */
 471         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 472         mad_snoop_priv->agent.device = device;
 473         mad_snoop_priv->agent.recv_handler = recv_handler;
 474         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 475         mad_snoop_priv->agent.context = context;
 476         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 477         mad_snoop_priv->agent.port_num = port_num;
 478         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 479         init_completion(&mad_snoop_priv->comp);
 480         mad_snoop_priv->snoop_index = register_snoop_agent(
 481                                                 &port_priv->qp_info[qpn],
 482                                                 mad_snoop_priv);
 483         if (mad_snoop_priv->snoop_index < 0) {
 484                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 485                 goto error2;
 486         }
 487
 488         atomic_set(&mad_snoop_priv->refcount, 1);
 489         return &mad_snoop_priv->agent;
 490
 491 error2:
 492         kfree(mad_snoop_priv);
 493 error1:
 494         return ret;
 495 }
 496 EXPORT_SYMBOL(ib_register_mad_snoop);
 497
 498 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 499 {
 500         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 501                 complete(&mad_agent_priv->comp);
 502 }
 503
 504 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
 505 {
 506         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 507                 complete(&mad_snoop_priv->comp);
 508 }
 509
 510 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 511 {
 512         struct ib_mad_port_private *port_priv;
 513         unsigned long flags;
 514
 515         /* Note that we could still be handling received MADs */
 516
 517         /*
 518          * Canceling all sends results in dropping received response
 519          * MADs, preventing us from queuing additional work
 520          */
 521         cancel_mads(mad_agent_priv);
 522         port_priv = mad_agent_priv->qp_info->port_priv;
 523         cancel_delayed_work(&mad_agent_priv->timed_work);
 524
 525         spin_lock_irqsave(&port_priv->reg_lock, flags);
 526         remove_mad_reg_req(mad_agent_priv);
 527         list_del(&mad_agent_priv->agent_list);
 528         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 529
 530         flush_workqueue(port_priv->wq);
 531         ib_cancel_rmpp_recvs(mad_agent_priv);
 532
 533         deref_mad_agent(mad_agent_priv);
 534         wait_for_completion(&mad_agent_priv->comp);
 535
 536         kfree(mad_agent_priv->reg_req);
 537         ib_dereg_mr(mad_agent_priv->agent.mr);
 538         kfree(mad_agent_priv);
 539 }
 540
 541 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 542 {
 543         struct ib_mad_qp_info *qp_info;
 544         unsigned long flags;
 545
 546         qp_info = mad_snoop_priv->qp_info;
 547         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 548         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 549         atomic_dec(&qp_info->snoop_count);
 550         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 551
 552         deref_snoop_agent(mad_snoop_priv);
 553         wait_for_completion(&mad_snoop_priv->comp);
 554
 555         kfree(mad_snoop_priv);
 556 }
 557
 558 /*
 559  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 560  */
 561 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 562 {
 563         struct ib_mad_agent_private *mad_agent_priv;
 564         struct ib_mad_snoop_private *mad_snoop_priv;
 565
 566         /* If the TID is zero, the agent can only snoop. */
 567         if (mad_agent->hi_tid) {
 568                 mad_agent_priv = container_of(mad_agent,
 569                                               struct ib_mad_agent_private,
 570                                               agent);
 571                 unregister_mad_agent(mad_agent_priv);
 572         } else {
 573                 mad_snoop_priv = container_of(mad_agent,
 574                                               struct ib_mad_snoop_private,
 575                                               agent);
 576                 unregister_mad_snoop(mad_snoop_priv);
 577         }
 578         return 0;
 579 }
 580 EXPORT_SYMBOL(ib_unregister_mad_agent);
 581
 582 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 583 {
 584         struct ib_mad_queue *mad_queue;
 585         unsigned long flags;
 586
 587         BUG_ON(!mad_list->mad_queue);
 588         mad_queue = mad_list->mad_queue;
 589         spin_lock_irqsave(&mad_queue->lock, flags);
 590         list_del(&mad_list->list);
 591         mad_queue->count--;
 592         spin_unlock_irqrestore(&mad_queue->lock, flags);
 593 }
 594
 595 static void snoop_send(struct ib_mad_qp_info *qp_info,
 596                        struct ib_mad_send_buf *send_buf,
 597                        struct ib_mad_send_wc *mad_send_wc,
 598                        int mad_snoop_flags)
 599 {
 600         struct ib_mad_snoop_private *mad_snoop_priv;
 601         unsigned long flags;
 602         int i;
 603
 604         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 605         for (i = 0; i < qp_info->snoop_table_size; i++) {
 606                 mad_snoop_priv = qp_info->snoop_table[i];
 607                 if (!mad_snoop_priv ||
 608                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 609                         continue;
 610
 611                 atomic_inc(&mad_snoop_priv->refcount);
 612                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 613                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 614                                                     send_buf, mad_send_wc);
 615                 deref_snoop_agent(mad_snoop_priv);
 616                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 617         }
 618         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 619 }
 620
 621 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 622                        struct ib_mad_recv_wc *mad_recv_wc,
 623                        int mad_snoop_flags)
 624 {
 625         struct ib_mad_snoop_private *mad_snoop_priv;
 626         unsigned long flags;
 627         int i;
 628
 629         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 630         for (i = 0; i < qp_info->snoop_table_size; i++) {
 631                 mad_snoop_priv = qp_info->snoop_table[i];
 632                 if (!mad_snoop_priv ||
 633                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 634                         continue;
 635
 636                 atomic_inc(&mad_snoop_priv->refcount);
 637                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 638                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
 639                                                    mad_recv_wc);
 640                 deref_snoop_agent(mad_snoop_priv);
 641                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 642         }
 643         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 644 }
 645
 646 static void build_smp_wc(struct ib_qp *qp,
 647                          u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
 648                          struct ib_wc *wc)
 649 {
 650         memset(wc, 0, sizeof *wc);
 651         wc->wr_id = wr_id;
 652         wc->status = IB_WC_SUCCESS;
 653         wc->opcode = IB_WC_RECV;
 654         wc->pkey_index = pkey_index;
 655         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
 656         wc->src_qp = IB_QP0;
 657         wc->qp = qp;
 658         wc->slid = slid;
 659         wc->sl = 0;
 660         wc->dlid_path_bits = 0;
 661         wc->port_num = port_num;
 662 }
 663
 664 /*
 665  * Return 0 if SMP is to be sent
 666  * Return 1 if SMP was consumed locally (whether or not solicited)
 667  * Return < 0 if error
 668  */
 669 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 670                                   struct ib_mad_send_wr_private *mad_send_wr)
 671 {
 672         int ret = 0;
 673         struct ib_smp *smp = mad_send_wr->send_buf.mad;
 674         unsigned long flags;
 675         struct ib_mad_local_private *local;
 676         struct ib_mad_private *mad_priv;
 677         struct ib_mad_port_private *port_priv;
 678         struct ib_mad_agent_private *recv_mad_agent = NULL;
 679         struct ib_device *device = mad_agent_priv->agent.device;
 680         u8 port_num;
 681         struct ib_wc mad_wc;
 682         struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
 683
 684         if (device->node_type == RDMA_NODE_IB_SWITCH &&
 685             smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 686                 port_num = send_wr->wr.ud.port_num;
 687         else
 688                 port_num = mad_agent_priv->agent.port_num;
 689
 690         /*
 691          * Directed route handling starts if the initial LID routed part of
 692          * a request or the ending LID routed part of a response is empty.
 693          * If we are at the start of the LID routed part, don't update the
 694          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
 695          */
 696         if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
 697              IB_LID_PERMISSIVE &&
 698              smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
 699              IB_SMI_DISCARD) {
 700                 ret = -EINVAL;
 701                 printk(KERN_ERR PFX "Invalid directed route\n");
 702                 goto out;
 703         }
 704
 705         /* Check to post send on QP or process locally */
 706         if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
 707             smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
 708                 goto out;
 709
 710         local = kmalloc(sizeof *local, GFP_ATOMIC);
 711         if (!local) {
 712                 ret = -ENOMEM;
 713                 printk(KERN_ERR PFX "No memory for ib_mad_local_private\n");
 714                 goto out;
 715         }
 716         local->mad_priv = NULL;
 717         local->recv_mad_agent = NULL;
 718         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
 719         if (!mad_priv) {
 720                 ret = -ENOMEM;
 721                 printk(KERN_ERR PFX "No memory for local response MAD\n");
 722                 kfree(local);
 723                 goto out;
 724         }
 725
 726         build_smp_wc(mad_agent_priv->agent.qp,
 727                      send_wr->wr_id, be16_to_cpu(smp->dr_slid),
 728                      send_wr->wr.ud.pkey_index,
 729                      send_wr->wr.ud.port_num, &mad_wc);
 730
 731         /* No GRH for DR SMP */
 732         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 733                                   (struct ib_mad *)smp,
 734                                   (struct ib_mad *)&mad_priv->mad);
 735         switch (ret)
 736         {
 737         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 738                 if (ib_response_mad(&mad_priv->mad.mad) &&
 739                     mad_agent_priv->agent.recv_handler) {
 740                         local->mad_priv = mad_priv;
 741                         local->recv_mad_agent = mad_agent_priv;
 742                         /*
 743                          * Reference MAD agent until receive
 744                          * side of local completion handled
 745                          */
 746                         atomic_inc(&mad_agent_priv->refcount);
 747                 } else
 748                         kmem_cache_free(ib_mad_cache, mad_priv);
 749                 break;
 750         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 751                 kmem_cache_free(ib_mad_cache, mad_priv);
 752                 break;
 753         case IB_MAD_RESULT_SUCCESS:
 754                 /* Treat like an incoming receive MAD */
 755                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 756                                             mad_agent_priv->agent.port_num);
 757                 if (port_priv) {
 758                         memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
 759                         recv_mad_agent = find_mad_agent(port_priv,
 760                                                         &mad_priv->mad.mad);
 761                 }
 762                 if (!port_priv || !recv_mad_agent) {
 763                         /*
 764                          * No receiving agent so drop packet and
 765                          * generate send completion.
 766                          */
 767                         kmem_cache_free(ib_mad_cache, mad_priv);
 768                         break;
 769                 }
 770                 local->mad_priv = mad_priv;
 771                 local->recv_mad_agent = recv_mad_agent;
 772                 break;
 773         default:
 774                 kmem_cache_free(ib_mad_cache, mad_priv);
 775                 kfree(local);
 776                 ret = -EINVAL;
 777                 goto out;
 778         }
 779
 780         local->mad_send_wr = mad_send_wr;
 781         /* Reference MAD agent until send side of local completion handled */
 782         atomic_inc(&mad_agent_priv->refcount);
 783         /* Queue local completion to local list */
 784         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 785         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 786         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 787         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 788                    &mad_agent_priv->local_work);
 789         ret = 1;
 790 out:
 791         return ret;
 792 }
 793
 794 static int get_pad_size(int hdr_len, int data_len)
 795 {
 796         int seg_size, pad;
 797
 798         seg_size = sizeof(struct ib_mad) - hdr_len;
 799         if (data_len && seg_size) {
 800                 pad = seg_size - data_len % seg_size;
 801                 return pad == seg_size ? 0 : pad;
 802         } else
 803                 return seg_size;
 804 }
 805
 806 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 807 {
 808         struct ib_rmpp_segment *s, *t;
 809
 810         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 811                 list_del(&s->list);
 812                 kfree(s);
 813         }
 814 }
 815
 816 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 817                                 gfp_t gfp_mask)
 818 {
 819         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 820         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 821         struct ib_rmpp_segment *seg = NULL;
 822         int left, seg_size, pad;
 823
 824         send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
 825         seg_size = send_buf->seg_size;
 826         pad = send_wr->pad;
 827
 828         /* Allocate data segments. */
 829         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 830                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 831                 if (!seg) {
 832                         printk(KERN_ERR "alloc_send_rmpp_segs: RMPP mem "
 833                                "alloc failed for len %zd, gfp %#x\n",
 834                                sizeof (*seg) + seg_size, gfp_mask);
 835                         free_send_rmpp_list(send_wr);
 836                         return -ENOMEM;
 837                 }
 838                 seg->num = ++send_buf->seg_count;
 839                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 840         }
 841
 842         /* Zero any padding */
 843         if (pad)
 844                 memset(seg->data + seg_size - pad, 0, pad);
 845
 846         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
 847                                           agent.rmpp_version;
 848         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 849         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
 850
 851         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
 852                                         struct ib_rmpp_segment, list);
 853         send_wr->last_ack_seg = send_wr->cur_seg;
 854         return 0;
 855 }
 856
 857 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 858                                             u32 remote_qpn, u16 pkey_index,
 859                                             int rmpp_active,
 860                                             int hdr_len, int data_len,
 861                                             gfp_t gfp_mask)
 862 {
 863         struct ib_mad_agent_private *mad_agent_priv;
 864         struct ib_mad_send_wr_private *mad_send_wr;
 865         int pad, message_size, ret, size;
 866         void *buf;
 867
 868         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 869                                       agent);
 870         pad = get_pad_size(hdr_len, data_len);
 871         message_size = hdr_len + data_len + pad;
 872
 873         if ((!mad_agent->rmpp_version &&
 874              (rmpp_active || message_size > sizeof(struct ib_mad))) ||
 875             (!rmpp_active && message_size > sizeof(struct ib_mad)))
 876                 return ERR_PTR(-EINVAL);
 877
 878         size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
 879         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
 880         if (!buf)
 881                 return ERR_PTR(-ENOMEM);
 882
 883         mad_send_wr = buf + size;
 884         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
 885         mad_send_wr->send_buf.mad = buf;
 886         mad_send_wr->send_buf.hdr_len = hdr_len;
 887         mad_send_wr->send_buf.data_len = data_len;
 888         mad_send_wr->pad = pad;
 889
 890         mad_send_wr->mad_agent_priv = mad_agent_priv;
 891         mad_send_wr->sg_list[0].length = hdr_len;
 892         mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
 893         mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
 894         mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
 895
 896         mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
 897         mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 898         mad_send_wr->send_wr.num_sge = 2;
 899         mad_send_wr->send_wr.opcode = IB_WR_SEND;
 900         mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
 901         mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
 902         mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 903         mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
 904
 905         if (rmpp_active) {
 906                 ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
 907                 if (ret) {
 908                         kfree(buf);
 909                         return ERR_PTR(ret);
 910                 }
 911         }
 912
 913         mad_send_wr->send_buf.mad_agent = mad_agent;
 914         atomic_inc(&mad_agent_priv->refcount);
 915         return &mad_send_wr->send_buf;
 916 }
 917 EXPORT_SYMBOL(ib_create_send_mad);
 918
 919 int ib_get_mad_data_offset(u8 mgmt_class)
 920 {
 921         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
 922                 return IB_MGMT_SA_HDR;
 923         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 924                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 925                  (mgmt_class == IB_MGMT_CLASS_BIS))
 926                 return IB_MGMT_DEVICE_HDR;
 927         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 928                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
 929                 return IB_MGMT_VENDOR_HDR;
 930         else
 931                 return IB_MGMT_MAD_HDR;
 932 }
 933 EXPORT_SYMBOL(ib_get_mad_data_offset);
 934
 935 int ib_is_mad_class_rmpp(u8 mgmt_class)
 936 {
 937         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
 938             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 939             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 940             (mgmt_class == IB_MGMT_CLASS_BIS) ||
 941             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 942              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
 943                 return 1;
 944         return 0;
 945 }
 946 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
 947
 948 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
 949 {
 950         struct ib_mad_send_wr_private *mad_send_wr;
 951         struct list_head *list;
 952
 953         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 954                                    send_buf);
 955         list = &mad_send_wr->cur_seg->list;
 956
 957         if (mad_send_wr->cur_seg->num < seg_num) {
 958                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
 959                         if (mad_send_wr->cur_seg->num == seg_num)
 960                                 break;
 961         } else if (mad_send_wr->cur_seg->num > seg_num) {
 962                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
 963                         if (mad_send_wr->cur_seg->num == seg_num)
 964                                 break;
 965         }
 966         return mad_send_wr->cur_seg->data;
 967 }
 968 EXPORT_SYMBOL(ib_get_rmpp_segment);
 969
 970 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
 971 {
 972         if (mad_send_wr->send_buf.seg_count)
 973                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
 974                                            mad_send_wr->seg_num);
 975         else
 976                 return mad_send_wr->send_buf.mad +
 977                        mad_send_wr->send_buf.hdr_len;
 978 }
 979
 980 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
 981 {
 982         struct ib_mad_agent_private *mad_agent_priv;
 983         struct ib_mad_send_wr_private *mad_send_wr;
 984
 985         mad_agent_priv = container_of(send_buf->mad_agent,
 986                                       struct ib_mad_agent_private, agent);
 987         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 988                                    send_buf);
 989
 990         free_send_rmpp_list(mad_send_wr);
 991         kfree(send_buf->mad);
 992         deref_mad_agent(mad_agent_priv);
 993 }
 994 EXPORT_SYMBOL(ib_free_send_mad);
 995
 996 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
 997 {
 998         struct ib_mad_qp_info *qp_info;
 999         struct list_head *list;
1000         struct ib_send_wr *bad_send_wr;
1001         struct ib_mad_agent *mad_agent;
1002         struct ib_sge *sge;
1003         unsigned long flags;
1004         int ret;
1005
1006         /* Set WR ID to find mad_send_wr upon completion */
1007         qp_info = mad_send_wr->mad_agent_priv->qp_info;
1008         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1009         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1010
1011         mad_agent = mad_send_wr->send_buf.mad_agent;
1012         sge = mad_send_wr->sg_list;
1013         sge[0].addr = ib_dma_map_single(mad_agent->device,
1014                                         mad_send_wr->send_buf.mad,
1015                                         sge[0].length,
1016                                         DMA_TO_DEVICE);
1017         mad_send_wr->header_mapping = sge[0].addr;
1018
1019         sge[1].addr = ib_dma_map_single(mad_agent->device,
1020                                         ib_get_payload(mad_send_wr),
1021                                         sge[1].length,
1022                                         DMA_TO_DEVICE);
1023         mad_send_wr->payload_mapping = sge[1].addr;
1024
1025         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1026         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1027                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1028                                    &bad_send_wr);
1029                 list = &qp_info->send_queue.list;
1030         } else {
1031                 ret = 0;
1032                 list = &qp_info->overflow_list;
1033         }
1034
1035         if (!ret) {
1036                 qp_info->send_queue.count++;
1037                 list_add_tail(&mad_send_wr->mad_list.list, list);
1038         }
1039         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1040         if (ret) {
1041                 ib_dma_unmap_single(mad_agent->device,
1042                                     mad_send_wr->header_mapping,
1043                                     sge[0].length, DMA_TO_DEVICE);
1044                 ib_dma_unmap_single(mad_agent->device,
1045                                     mad_send_wr->payload_mapping,
1046                                     sge[1].length, DMA_TO_DEVICE);
1047         }
1048         return ret;
1049 }
1050
1051 /*
1052  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1053  *  with the registered client
1054  */
1055 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1056                      struct ib_mad_send_buf **bad_send_buf)
1057 {
1058         struct ib_mad_agent_private *mad_agent_priv;
1059         struct ib_mad_send_buf *next_send_buf;
1060         struct ib_mad_send_wr_private *mad_send_wr;
1061         unsigned long flags;
1062         int ret = -EINVAL;
1063
1064         /* Walk list of send WRs and post each on send list */
1065         for (; send_buf; send_buf = next_send_buf) {
1066
1067                 mad_send_wr = container_of(send_buf,
1068                                            struct ib_mad_send_wr_private,
1069                                            send_buf);
1070                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1071
1072                 if (!send_buf->mad_agent->send_handler ||
1073                     (send_buf->timeout_ms &&
1074                      !send_buf->mad_agent->recv_handler)) {
1075                         ret = -EINVAL;
1076                         goto error;
1077                 }
1078
1079                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1080                         if (mad_agent_priv->agent.rmpp_version) {
1081                                 ret = -EINVAL;
1082                                 goto error;
1083                         }
1084                 }
1085
1086                 /*
1087                  * Save pointer to next work request to post in case the
1088                  * current one completes, and the user modifies the work
1089                  * request associated with the completion
1090                  */
1091                 next_send_buf = send_buf->next;
1092                 mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1093
1094                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1095                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1096                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1097                                                      mad_send_wr);
1098                         if (ret < 0)            /* error */
1099                                 goto error;
1100                         else if (ret == 1)      /* locally consumed */
1101                                 continue;
1102                 }
1103
1104                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1105                 /* Timeout will be updated after send completes */
1106                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1107                 mad_send_wr->max_retries = send_buf->retries;
1108                 mad_send_wr->retries_left = send_buf->retries;
1109                 send_buf->retries = 0;
1110                 /* Reference for work request to QP + response */
1111                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1112                 mad_send_wr->status = IB_WC_SUCCESS;
1113
1114                 /* Reference MAD agent until send completes */
1115                 atomic_inc(&mad_agent_priv->refcount);
1116                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1117                 list_add_tail(&mad_send_wr->agent_list,
1118                               &mad_agent_priv->send_list);
1119                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1120
1121                 if (mad_agent_priv->agent.rmpp_version) {
1122                         ret = ib_send_rmpp_mad(mad_send_wr);
1123                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1124                                 ret = ib_send_mad(mad_send_wr);
1125                 } else
1126                         ret = ib_send_mad(mad_send_wr);
1127                 if (ret < 0) {
1128                         /* Fail send request */
1129                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1130                         list_del(&mad_send_wr->agent_list);
1131                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1132                         atomic_dec(&mad_agent_priv->refcount);
1133                         goto error;
1134                 }
1135         }
1136         return 0;
1137 error:
1138         if (bad_send_buf)
1139                 *bad_send_buf = send_buf;
1140         return ret;
1141 }
1142 EXPORT_SYMBOL(ib_post_send_mad);
1143
1144 /*
1145  * ib_free_recv_mad - Returns data buffers used to receive
1146  *  a MAD to the access layer
1147  */
1148 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1149 {
1150         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1151         struct ib_mad_private_header *mad_priv_hdr;
1152         struct ib_mad_private *priv;
1153         struct list_head free_list;
1154
1155         INIT_LIST_HEAD(&free_list);
1156         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1157
1158         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1159                                         &free_list, list) {
1160                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1161                                            recv_buf);
1162                 mad_priv_hdr = container_of(mad_recv_wc,
1163                                             struct ib_mad_private_header,
1164                                             recv_wc);
1165                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1166                                     header);
1167                 kmem_cache_free(ib_mad_cache, priv);
1168         }
1169 }
1170 EXPORT_SYMBOL(ib_free_recv_mad);
1171
1172 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1173                                         u8 rmpp_version,
1174                                         ib_mad_send_handler send_handler,
1175                                         ib_mad_recv_handler recv_handler,
1176                                         void *context)
1177 {
1178         return ERR_PTR(-EINVAL);        /* XXX: for now */
1179 }
1180 EXPORT_SYMBOL(ib_redirect_mad_qp);
1181
1182 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1183                       struct ib_wc *wc)
1184 {
1185         printk(KERN_ERR PFX "ib_process_mad_wc() not implemented yet\n");
1186         return 0;
1187 }
1188 EXPORT_SYMBOL(ib_process_mad_wc);
1189
1190 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1191                          struct ib_mad_reg_req *mad_reg_req)
1192 {
1193         int i;
1194
1195         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1196                 if ((*method)->agent[i]) {
1197                         printk(KERN_ERR PFX "Method %d already in use\n", i);
1198                         return -EINVAL;
1199                 }
1200         }
1201         return 0;
1202 }
1203
1204 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1205 {
1206         /* Allocate management method table */
1207         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1208         if (!*method) {
1209                 printk(KERN_ERR PFX "No memory for "
1210                        "ib_mad_mgmt_method_table\n");
1211                 return -ENOMEM;
1212         }
1213
1214         return 0;
1215 }
1216
1217 /*
1218  * Check to see if there are any methods still in use
1219  */
1220 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1221 {
1222         int i;
1223
1224         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1225                 if (method->agent[i])
1226                         return 1;
1227         return 0;
1228 }
1229
1230 /*
1231  * Check to see if there are any method tables for this class still in use
1232  */
1233 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1234 {
1235         int i;
1236
1237         for (i = 0; i < MAX_MGMT_CLASS; i++)
1238                 if (class->method_table[i])
1239                         return 1;
1240         return 0;
1241 }
1242
1243 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1244 {
1245         int i;
1246
1247         for (i = 0; i < MAX_MGMT_OUI; i++)
1248                 if (vendor_class->method_table[i])
1249                         return 1;
1250         return 0;
1251 }
1252
1253 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1254                            char *oui)
1255 {
1256         int i;
1257
1258         for (i = 0; i < MAX_MGMT_OUI; i++)
1259                 /* Is there matching OUI for this vendor class ? */
1260                 if (!memcmp(vendor_class->oui[i], oui, 3))
1261                         return i;
1262
1263         return -1;
1264 }
1265
1266 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1267 {
1268         int i;
1269
1270         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1271                 if (vendor->vendor_class[i])
1272                         return 1;
1273
1274         return 0;
1275 }
1276
1277 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1278                                      struct ib_mad_agent_private *agent)
1279 {
1280         int i;
1281
1282         /* Remove any methods for this mad agent */
1283         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1284                 if (method->agent[i] == agent) {
1285                         method->agent[i] = NULL;
1286                 }
1287         }
1288 }
1289
1290 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1291                               struct ib_mad_agent_private *agent_priv,
1292                               u8 mgmt_class)
1293 {
1294         struct ib_mad_port_private *port_priv;
1295         struct ib_mad_mgmt_class_table **class;
1296         struct ib_mad_mgmt_method_table **method;
1297         int i, ret;
1298
1299         port_priv = agent_priv->qp_info->port_priv;
1300         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1301         if (!*class) {
1302                 /* Allocate management class table for "new" class version */
1303                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1304                 if (!*class) {
1305                         printk(KERN_ERR PFX "No memory for "
1306                                "ib_mad_mgmt_class_table\n");
1307                         ret = -ENOMEM;
1308                         goto error1;
1309                 }
1310
1311                 /* Allocate method table for this management class */
1312                 method = &(*class)->method_table[mgmt_class];
1313                 if ((ret = allocate_method_table(method)))
1314                         goto error2;
1315         } else {
1316                 method = &(*class)->method_table[mgmt_class];
1317                 if (!*method) {
1318                         /* Allocate method table for this management class */
1319                         if ((ret = allocate_method_table(method)))
1320                                 goto error1;
1321                 }
1322         }
1323
1324         /* Now, make sure methods are not already in use */
1325         if (method_in_use(method, mad_reg_req))
1326                 goto error3;
1327
1328         /* Finally, add in methods being registered */
1329         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1330                 (*method)->agent[i] = agent_priv;
1331
1332         return 0;
1333
1334 error3:
1335         /* Remove any methods for this mad agent */
1336         remove_methods_mad_agent(*method, agent_priv);
1337         /* Now, check to see if there are any methods in use */
1338         if (!check_method_table(*method)) {
1339                 /* If not, release management method table */
1340                 kfree(*method);
1341                 *method = NULL;
1342         }
1343         ret = -EINVAL;
1344         goto error1;
1345 error2:
1346         kfree(*class);
1347         *class = NULL;
1348 error1:
1349         return ret;
1350 }
1351
1352 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1353                            struct ib_mad_agent_private *agent_priv)
1354 {
1355         struct ib_mad_port_private *port_priv;
1356         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1357         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1358         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1359         struct ib_mad_mgmt_method_table **method;
1360         int i, ret = -ENOMEM;
1361         u8 vclass;
1362
1363         /* "New" vendor (with OUI) class */
1364         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1365         port_priv = agent_priv->qp_info->port_priv;
1366         vendor_table = &port_priv->version[
1367                                 mad_reg_req->mgmt_class_version].vendor;
1368         if (!*vendor_table) {
1369                 /* Allocate mgmt vendor class table for "new" class version */
1370                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1371                 if (!vendor) {
1372                         printk(KERN_ERR PFX "No memory for "
1373                                "ib_mad_mgmt_vendor_class_table\n");
1374                         goto error1;
1375                 }
1376
1377                 *vendor_table = vendor;
1378         }
1379         if (!(*vendor_table)->vendor_class[vclass]) {
1380                 /* Allocate table for this management vendor class */
1381                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1382                 if (!vendor_class) {
1383                         printk(KERN_ERR PFX "No memory for "
1384                                "ib_mad_mgmt_vendor_class\n");
1385                         goto error2;
1386                 }
1387
1388                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1389         }
1390         for (i = 0; i < MAX_MGMT_OUI; i++) {
1391                 /* Is there matching OUI for this vendor class ? */
1392                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1393                             mad_reg_req->oui, 3)) {
1394                         method = &(*vendor_table)->vendor_class[
1395                                                 vclass]->method_table[i];
1396                         BUG_ON(!*method);
1397                         goto check_in_use;
1398                 }
1399         }
1400         for (i = 0; i < MAX_MGMT_OUI; i++) {
1401                 /* OUI slot available ? */
1402                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1403                                 vclass]->oui[i])) {
1404                         method = &(*vendor_table)->vendor_class[
1405                                 vclass]->method_table[i];
1406                         BUG_ON(*method);
1407                         /* Allocate method table for this OUI */
1408                         if ((ret = allocate_method_table(method)))
1409                                 goto error3;
1410                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1411                                mad_reg_req->oui, 3);
1412                         goto check_in_use;
1413                 }
1414         }
1415         printk(KERN_ERR PFX "All OUI slots in use\n");
1416         goto error3;
1417
1418 check_in_use:
1419         /* Now, make sure methods are not already in use */
1420         if (method_in_use(method, mad_reg_req))
1421                 goto error4;
1422
1423         /* Finally, add in methods being registered */
1424         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1425                 (*method)->agent[i] = agent_priv;
1426
1427         return 0;
1428
1429 error4:
1430         /* Remove any methods for this mad agent */
1431         remove_methods_mad_agent(*method, agent_priv);
1432         /* Now, check to see if there are any methods in use */
1433         if (!check_method_table(*method)) {
1434                 /* If not, release management method table */
1435                 kfree(*method);
1436                 *method = NULL;
1437         }
1438         ret = -EINVAL;
1439 error3:
1440         if (vendor_class) {
1441                 (*vendor_table)->vendor_class[vclass] = NULL;
1442                 kfree(vendor_class);
1443         }
1444 error2:
1445         if (vendor) {
1446                 *vendor_table = NULL;
1447                 kfree(vendor);
1448         }
1449 error1:
1450         return ret;
1451 }
1452
1453 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1454 {
1455         struct ib_mad_port_private *port_priv;
1456         struct ib_mad_mgmt_class_table *class;
1457         struct ib_mad_mgmt_method_table *method;
1458         struct ib_mad_mgmt_vendor_class_table *vendor;
1459         struct ib_mad_mgmt_vendor_class *vendor_class;
1460         int index;
1461         u8 mgmt_class;
1462
1463         /*
1464          * Was MAD registration request supplied
1465          * with original registration ?
1466          */
1467         if (!agent_priv->reg_req) {
1468                 goto out;
1469         }
1470
1471         port_priv = agent_priv->qp_info->port_priv;
1472         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1473         class = port_priv->version[
1474                         agent_priv->reg_req->mgmt_class_version].class;
1475         if (!class)
1476                 goto vendor_check;
1477
1478         method = class->method_table[mgmt_class];
1479         if (method) {
1480                 /* Remove any methods for this mad agent */
1481                 remove_methods_mad_agent(method, agent_priv);
1482                 /* Now, check to see if there are any methods still in use */
1483                 if (!check_method_table(method)) {
1484                         /* If not, release management method table */
1485                          kfree(method);
1486                          class->method_table[mgmt_class] = NULL;
1487                          /* Any management classes left ? */
1488                         if (!check_class_table(class)) {
1489                                 /* If not, release management class table */
1490                                 kfree(class);
1491                                 port_priv->version[
1492                                         agent_priv->reg_req->
1493                                         mgmt_class_version].class = NULL;
1494                         }
1495                 }
1496         }
1497
1498 vendor_check:
1499         if (!is_vendor_class(mgmt_class))
1500                 goto out;
1501
1502         /* normalize mgmt_class to vendor range 2 */
1503         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1504         vendor = port_priv->version[
1505                         agent_priv->reg_req->mgmt_class_version].vendor;
1506
1507         if (!vendor)
1508                 goto out;
1509
1510         vendor_class = vendor->vendor_class[mgmt_class];
1511         if (vendor_class) {
1512                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1513                 if (index < 0)
1514                         goto out;
1515                 method = vendor_class->method_table[index];
1516                 if (method) {
1517                         /* Remove any methods for this mad agent */
1518                         remove_methods_mad_agent(method, agent_priv);
1519                         /*
1520                          * Now, check to see if there are
1521                          * any methods still in use
1522                          */
1523                         if (!check_method_table(method)) {
1524                                 /* If not, release management method table */
1525                                 kfree(method);
1526                                 vendor_class->method_table[index] = NULL;
1527                                 memset(vendor_class->oui[index], 0, 3);
1528                                 /* Any OUIs left ? */
1529                                 if (!check_vendor_class(vendor_class)) {
1530                                         /* If not, release vendor class table */
1531                                         kfree(vendor_class);
1532                                         vendor->vendor_class[mgmt_class] = NULL;
1533                                         /* Any other vendor classes left ? */
1534                                         if (!check_vendor_table(vendor)) {
1535                                                 kfree(vendor);
1536                                                 port_priv->version[
1537                                                         agent_priv->reg_req->
1538                                                         mgmt_class_version].
1539                                                         vendor = NULL;
1540                                         }
1541                                 }
1542                         }
1543                 }
1544         }
1545
1546 out:
1547         return;
1548 }
1549
1550 static struct ib_mad_agent_private *
1551 find_mad_agent(struct ib_mad_port_private *port_priv,
1552                struct ib_mad *mad)
1553 {
1554         struct ib_mad_agent_private *mad_agent = NULL;
1555         unsigned long flags;
1556
1557         spin_lock_irqsave(&port_priv->reg_lock, flags);
1558         if (ib_response_mad(mad)) {
1559                 u32 hi_tid;
1560                 struct ib_mad_agent_private *entry;
1561
1562                 /*
1563                  * Routing is based on high 32 bits of transaction ID
1564                  * of MAD.
1565                  */
1566                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1567                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1568                         if (entry->agent.hi_tid == hi_tid) {
1569                                 mad_agent = entry;
1570                                 break;
1571                         }
1572                 }
1573         } else {
1574                 struct ib_mad_mgmt_class_table *class;
1575                 struct ib_mad_mgmt_method_table *method;
1576                 struct ib_mad_mgmt_vendor_class_table *vendor;
1577                 struct ib_mad_mgmt_vendor_class *vendor_class;
1578                 struct ib_vendor_mad *vendor_mad;
1579                 int index;
1580
1581                 /*
1582                  * Routing is based on version, class, and method
1583                  * For "newer" vendor MADs, also based on OUI
1584                  */
1585                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1586                         goto out;
1587                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1588                         class = port_priv->version[
1589                                         mad->mad_hdr.class_version].class;
1590                         if (!class)
1591                                 goto out;
1592                         method = class->method_table[convert_mgmt_class(
1593                                                         mad->mad_hdr.mgmt_class)];
1594                         if (method)
1595                                 mad_agent = method->agent[mad->mad_hdr.method &
1596                                                           ~IB_MGMT_METHOD_RESP];
1597                 } else {
1598                         vendor = port_priv->version[
1599                                         mad->mad_hdr.class_version].vendor;
1600                         if (!vendor)
1601                                 goto out;
1602                         vendor_class = vendor->vendor_class[vendor_class_index(
1603                                                 mad->mad_hdr.mgmt_class)];
1604                         if (!vendor_class)
1605                                 goto out;
1606                         /* Find matching OUI */
1607                         vendor_mad = (struct ib_vendor_mad *)mad;
1608                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1609                         if (index == -1)
1610                                 goto out;
1611                         method = vendor_class->method_table[index];
1612                         if (method) {
1613                                 mad_agent = method->agent[mad->mad_hdr.method &
1614                                                           ~IB_MGMT_METHOD_RESP];
1615                         }
1616                 }
1617         }
1618
1619         if (mad_agent) {
1620                 if (mad_agent->agent.recv_handler)
1621                         atomic_inc(&mad_agent->refcount);
1622                 else {
1623                         printk(KERN_NOTICE PFX "No receive handler for client "
1624                                "%p on port %d\n",
1625                                &mad_agent->agent, port_priv->port_num);
1626                         mad_agent = NULL;
1627                 }
1628         }
1629 out:
1630         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1631
1632         return mad_agent;
1633 }
1634
1635 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1636 {
1637         int valid = 0;
1638
1639         /* Make sure MAD base version is understood */
1640         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1641                 printk(KERN_ERR PFX "MAD received with unsupported base "
1642                        "version %d\n", mad->mad_hdr.base_version);
1643                 goto out;
1644         }
1645
1646         /* Filter SMI packets sent to other than QP0 */
1647         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1648             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1649                 if (qp_num == 0)
1650                         valid = 1;
1651         } else {
1652                 /* Filter GSI packets sent to QP0 */
1653                 if (qp_num != 0)
1654                         valid = 1;
1655         }
1656
1657 out:
1658         return valid;
1659 }
1660
1661 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1662                        struct ib_mad_hdr *mad_hdr)
1663 {
1664         struct ib_rmpp_mad *rmpp_mad;
1665
1666         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1667         return !mad_agent_priv->agent.rmpp_version ||
1668                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1669                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1670                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1671 }
1672
1673 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1674                                      struct ib_mad_recv_wc *rwc)
1675 {
1676         return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1677                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1678 }
1679
1680 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1681                                    struct ib_mad_send_wr_private *wr,
1682                                    struct ib_mad_recv_wc *rwc )
1683 {
1684         struct ib_ah_attr attr;
1685         u8 send_resp, rcv_resp;
1686         union ib_gid sgid;
1687         struct ib_device *device = mad_agent_priv->agent.device;
1688         u8 port_num = mad_agent_priv->agent.port_num;
1689         u8 lmc;
1690
1691         send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1692         rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1693
1694         if (send_resp == rcv_resp)
1695                 /* both requests, or both responses. GIDs different */
1696                 return 0;
1697
1698         if (ib_query_ah(wr->send_buf.ah, &attr))
1699                 /* Assume not equal, to avoid false positives. */
1700                 return 0;
1701
1702         if (!!(attr.ah_flags & IB_AH_GRH) !=
1703             !!(rwc->wc->wc_flags & IB_WC_GRH))
1704                 /* one has GID, other does not.  Assume different */
1705                 return 0;
1706
1707         if (!send_resp && rcv_resp) {
1708                 /* is request/response. */
1709                 if (!(attr.ah_flags & IB_AH_GRH)) {
1710                         if (ib_get_cached_lmc(device, port_num, &lmc))
1711                                 return 0;
1712                         return (!lmc || !((attr.src_path_bits ^
1713                                            rwc->wc->dlid_path_bits) &
1714                                           ((1 << lmc) - 1)));
1715                 } else {
1716                         if (ib_get_cached_gid(device, port_num,
1717                                               attr.grh.sgid_index, &sgid))
1718                                 return 0;
1719                         return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1720                                        16);
1721                 }
1722         }
1723
1724         if (!(attr.ah_flags & IB_AH_GRH))
1725                 return attr.dlid == rwc->wc->slid;
1726         else
1727                 return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1728                                16);
1729 }
1730
1731 static inline int is_direct(u8 class)
1732 {
1733         return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1734 }
1735
1736 struct ib_mad_send_wr_private*
1737 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1738                  struct ib_mad_recv_wc *wc)
1739 {
1740         struct ib_mad_send_wr_private *wr;
1741         struct ib_mad *mad;
1742
1743         mad = (struct ib_mad *)wc->recv_buf.mad;
1744
1745         list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1746                 if ((wr->tid == mad->mad_hdr.tid) &&
1747                     rcv_has_same_class(wr, wc) &&
1748                     /*
1749                      * Don't check GID for direct routed MADs.
1750                      * These might have permissive LIDs.
1751                      */
1752                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1753                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1754                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1755         }
1756
1757         /*
1758          * It's possible to receive the response before we've
1759          * been notified that the send has completed
1760          */
1761         list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1762                 if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1763                     wr->tid == mad->mad_hdr.tid &&
1764                     wr->timeout &&
1765                     rcv_has_same_class(wr, wc) &&
1766                     /*
1767                      * Don't check GID for direct routed MADs.
1768                      * These might have permissive LIDs.
1769                      */
1770                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1771                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1772                         /* Verify request has not been canceled */
1773                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1774         }
1775         return NULL;
1776 }
1777
1778 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1779 {
1780         mad_send_wr->timeout = 0;
1781         if (mad_send_wr->refcount == 1)
1782                 list_move_tail(&mad_send_wr->agent_list,
1783                               &mad_send_wr->mad_agent_priv->done_list);
1784 }
1785
1786 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1787                                  struct ib_mad_recv_wc *mad_recv_wc)
1788 {
1789         struct ib_mad_send_wr_private *mad_send_wr;
1790         struct ib_mad_send_wc mad_send_wc;
1791         unsigned long flags;
1792
1793         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1794         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1795         if (mad_agent_priv->agent.rmpp_version) {
1796                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1797                                                       mad_recv_wc);
1798                 if (!mad_recv_wc) {
1799                         deref_mad_agent(mad_agent_priv);
1800                         return;
1801                 }
1802         }
1803
1804         /* Complete corresponding request */
1805         if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1806                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1807                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1808                 if (!mad_send_wr) {
1809                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1810                         ib_free_recv_mad(mad_recv_wc);
1811                         deref_mad_agent(mad_agent_priv);
1812                         return;
1813                 }
1814                 ib_mark_mad_done(mad_send_wr);
1815                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1816
1817                 /* Defined behavior is to complete response before request */
1818                 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1819                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1820                                                    mad_recv_wc);
1821                 atomic_dec(&mad_agent_priv->refcount);
1822
1823                 mad_send_wc.status = IB_WC_SUCCESS;
1824                 mad_send_wc.vendor_err = 0;
1825                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1826                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1827         } else {
1828                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1829                                                    mad_recv_wc);
1830                 deref_mad_agent(mad_agent_priv);
1831         }
1832 }
1833
1834 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1835                                      struct ib_wc *wc)
1836 {
1837         struct ib_mad_qp_info *qp_info;
1838         struct ib_mad_private_header *mad_priv_hdr;
1839         struct ib_mad_private *recv, *response = NULL;
1840         struct ib_mad_list_head *mad_list;
1841         struct ib_mad_agent_private *mad_agent;
1842         int port_num;
1843
1844         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1845         qp_info = mad_list->mad_queue->qp_info;
1846         dequeue_mad(mad_list);
1847
1848         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1849                                     mad_list);
1850         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1851         ib_dma_unmap_single(port_priv->device,
1852                             recv->header.mapping,
1853                             sizeof(struct ib_mad_private) -
1854                               sizeof(struct ib_mad_private_header),
1855                             DMA_FROM_DEVICE);
1856
1857         /* Setup MAD receive work completion from "normal" work completion */
1858         recv->header.wc = *wc;
1859         recv->header.recv_wc.wc = &recv->header.wc;
1860         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1861         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1862         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1863
1864         if (atomic_read(&qp_info->snoop_count))
1865                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1866
1867         /* Validate MAD */
1868         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1869                 goto out;
1870
1871         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1872         if (!response) {
1873                 printk(KERN_ERR PFX "ib_mad_recv_done_handler no memory "
1874                        "for response buffer\n");
1875                 goto out;
1876         }
1877
1878         if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1879                 port_num = wc->port_num;
1880         else
1881                 port_num = port_priv->port_num;
1882
1883         if (recv->mad.mad.mad_hdr.mgmt_class ==
1884             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1885                 enum smi_forward_action retsmi;
1886
1887                 if (smi_handle_dr_smp_recv(&recv->mad.smp,
1888                                            port_priv->device->node_type,
1889                                            port_num,
1890                                            port_priv->device->phys_port_cnt) ==
1891                                            IB_SMI_DISCARD)
1892                         goto out;
1893
1894                 retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
1895                 if (retsmi == IB_SMI_LOCAL)
1896                         goto local;
1897
1898                 if (retsmi == IB_SMI_SEND) { /* don't forward */
1899                         if (smi_handle_dr_smp_send(&recv->mad.smp,
1900                                                    port_priv->device->node_type,
1901                                                    port_num) == IB_SMI_DISCARD)
1902                                 goto out;
1903
1904                         if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1905                                 goto out;
1906                 } else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
1907                         /* forward case for switches */
1908                         memcpy(response, recv, sizeof(*response));
1909                         response->header.recv_wc.wc = &response->header.wc;
1910                         response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1911                         response->header.recv_wc.recv_buf.grh = &response->grh;
1912
1913                         agent_send_response(&response->mad.mad,
1914                                             &response->grh, wc,
1915                                             port_priv->device,
1916                                             smi_get_fwd_port(&recv->mad.smp),
1917                                             qp_info->qp->qp_num);
1918
1919                         goto out;
1920                 }
1921         }
1922
1923 local:
1924         /* Give driver "right of first refusal" on incoming MAD */
1925         if (port_priv->device->process_mad) {
1926                 int ret;
1927
1928                 ret = port_priv->device->process_mad(port_priv->device, 0,
1929                                                      port_priv->port_num,
1930                                                      wc, &recv->grh,
1931                                                      &recv->mad.mad,
1932                                                      &response->mad.mad);
1933                 if (ret & IB_MAD_RESULT_SUCCESS) {
1934                         if (ret & IB_MAD_RESULT_CONSUMED)
1935                                 goto out;
1936                         if (ret & IB_MAD_RESULT_REPLY) {
1937                                 agent_send_response(&response->mad.mad,
1938                                                     &recv->grh, wc,
1939                                                     port_priv->device,
1940                                                     port_num,
1941                                                     qp_info->qp->qp_num);
1942                                 goto out;
1943                         }
1944                 }
1945         }
1946
1947         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1948         if (mad_agent) {
1949                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1950                 /*
1951                  * recv is freed up in error cases in ib_mad_complete_recv
1952                  * or via recv_handler in ib_mad_complete_recv()
1953                  */
1954                 recv = NULL;
1955         }
1956
1957 out:
1958         /* Post another receive request for this QP */
1959         if (response) {
1960                 ib_mad_post_receive_mads(qp_info, response);
1961                 if (recv)
1962                         kmem_cache_free(ib_mad_cache, recv);
1963         } else
1964                 ib_mad_post_receive_mads(qp_info, recv);
1965 }
1966
1967 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
1968 {
1969         struct ib_mad_send_wr_private *mad_send_wr;
1970         unsigned long delay;
1971
1972         if (list_empty(&mad_agent_priv->wait_list)) {
1973                 __cancel_delayed_work(&mad_agent_priv->timed_work);
1974         } else {
1975                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
1976                                          struct ib_mad_send_wr_private,
1977                                          agent_list);
1978
1979                 if (time_after(mad_agent_priv->timeout,
1980                                mad_send_wr->timeout)) {
1981                         mad_agent_priv->timeout = mad_send_wr->timeout;
1982                         __cancel_delayed_work(&mad_agent_priv->timed_work);
1983                         delay = mad_send_wr->timeout - jiffies;
1984                         if ((long)delay <= 0)
1985                                 delay = 1;
1986                         queue_delayed_work(mad_agent_priv->qp_info->
1987                                            port_priv->wq,
1988                                            &mad_agent_priv->timed_work, delay);
1989                 }
1990         }
1991 }
1992
1993 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
1994 {
1995         struct ib_mad_agent_private *mad_agent_priv;
1996         struct ib_mad_send_wr_private *temp_mad_send_wr;
1997         struct list_head *list_item;
1998         unsigned long delay;
1999
2000         mad_agent_priv = mad_send_wr->mad_agent_priv;
2001         list_del(&mad_send_wr->agent_list);
2002
2003         delay = mad_send_wr->timeout;
2004         mad_send_wr->timeout += jiffies;
2005
2006         if (delay) {
2007                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2008                         temp_mad_send_wr = list_entry(list_item,
2009                                                 struct ib_mad_send_wr_private,
2010                                                 agent_list);
2011                         if (time_after(mad_send_wr->timeout,
2012                                        temp_mad_send_wr->timeout))
2013                                 break;
2014                 }
2015         }
2016         else
2017                 list_item = &mad_agent_priv->wait_list;
2018         list_add(&mad_send_wr->agent_list, list_item);
2019
2020         /* Reschedule a work item if we have a shorter timeout */
2021         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list) {
2022                 __cancel_delayed_work(&mad_agent_priv->timed_work);
2023                 queue_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2024                                    &mad_agent_priv->timed_work, delay);
2025         }
2026 }
2027
2028 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2029                           int timeout_ms)
2030 {
2031         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2032         wait_for_response(mad_send_wr);
2033 }
2034
2035 /*
2036  * Process a send work completion
2037  */
2038 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2039                              struct ib_mad_send_wc *mad_send_wc)
2040 {
2041         struct ib_mad_agent_private     *mad_agent_priv;
2042         unsigned long                   flags;
2043         int                             ret;
2044
2045         mad_agent_priv = mad_send_wr->mad_agent_priv;
2046         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2047         if (mad_agent_priv->agent.rmpp_version) {
2048                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2049                 if (ret == IB_RMPP_RESULT_CONSUMED)
2050                         goto done;
2051         } else
2052                 ret = IB_RMPP_RESULT_UNHANDLED;
2053
2054         if (mad_send_wc->status != IB_WC_SUCCESS &&
2055             mad_send_wr->status == IB_WC_SUCCESS) {
2056                 mad_send_wr->status = mad_send_wc->status;
2057                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2058         }
2059
2060         if (--mad_send_wr->refcount > 0) {
2061                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2062                     mad_send_wr->status == IB_WC_SUCCESS) {
2063                         wait_for_response(mad_send_wr);
2064                 }
2065                 goto done;
2066         }
2067
2068         /* Remove send from MAD agent and notify client of completion */
2069         list_del(&mad_send_wr->agent_list);
2070         adjust_timeout(mad_agent_priv);
2071         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2072
2073         if (mad_send_wr->status != IB_WC_SUCCESS )
2074                 mad_send_wc->status = mad_send_wr->status;
2075         if (ret == IB_RMPP_RESULT_INTERNAL)
2076                 ib_rmpp_send_handler(mad_send_wc);
2077         else
2078                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2079                                                    mad_send_wc);
2080
2081         /* Release reference on agent taken when sending */
2082         deref_mad_agent(mad_agent_priv);
2083         return;
2084 done:
2085         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2086 }
2087
2088 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2089                                      struct ib_wc *wc)
2090 {
2091         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2092         struct ib_mad_list_head         *mad_list;
2093         struct ib_mad_qp_info           *qp_info;
2094         struct ib_mad_queue             *send_queue;
2095         struct ib_send_wr               *bad_send_wr;
2096         struct ib_mad_send_wc           mad_send_wc;
2097         unsigned long flags;
2098         int ret;
2099
2100         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2101         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2102                                    mad_list);
2103         send_queue = mad_list->mad_queue;
2104         qp_info = send_queue->qp_info;
2105
2106 retry:
2107         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2108                             mad_send_wr->header_mapping,
2109                             mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2110         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2111                             mad_send_wr->payload_mapping,
2112                             mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2113         queued_send_wr = NULL;
2114         spin_lock_irqsave(&send_queue->lock, flags);
2115         list_del(&mad_list->list);
2116
2117         /* Move queued send to the send queue */
2118         if (send_queue->count-- > send_queue->max_active) {
2119                 mad_list = container_of(qp_info->overflow_list.next,
2120                                         struct ib_mad_list_head, list);
2121                 queued_send_wr = container_of(mad_list,
2122                                         struct ib_mad_send_wr_private,
2123                                         mad_list);
2124                 list_move_tail(&mad_list->list, &send_queue->list);
2125         }
2126         spin_unlock_irqrestore(&send_queue->lock, flags);
2127
2128         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2129         mad_send_wc.status = wc->status;
2130         mad_send_wc.vendor_err = wc->vendor_err;
2131         if (atomic_read(&qp_info->snoop_count))
2132                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2133                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2134         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2135
2136         if (queued_send_wr) {
2137                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2138                                    &bad_send_wr);
2139                 if (ret) {
2140                         printk(KERN_ERR PFX "ib_post_send failed: %d\n", ret);
2141                         mad_send_wr = queued_send_wr;
2142                         wc->status = IB_WC_LOC_QP_OP_ERR;
2143                         goto retry;
2144                 }
2145         }
2146 }
2147
2148 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2149 {
2150         struct ib_mad_send_wr_private *mad_send_wr;
2151         struct ib_mad_list_head *mad_list;
2152         unsigned long flags;
2153
2154         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2155         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2156                 mad_send_wr = container_of(mad_list,
2157                                            struct ib_mad_send_wr_private,
2158                                            mad_list);
2159                 mad_send_wr->retry = 1;
2160         }
2161         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2162 }
2163
2164 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2165                               struct ib_wc *wc)
2166 {
2167         struct ib_mad_list_head *mad_list;
2168         struct ib_mad_qp_info *qp_info;
2169         struct ib_mad_send_wr_private *mad_send_wr;
2170         int ret;
2171
2172         /* Determine if failure was a send or receive */
2173         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2174         qp_info = mad_list->mad_queue->qp_info;
2175         if (mad_list->mad_queue == &qp_info->recv_queue)
2176                 /*
2177                  * Receive errors indicate that the QP has entered the error
2178                  * state - error handling/shutdown code will cleanup
2179                  */
2180                 return;
2181
2182         /*
2183          * Send errors will transition the QP to SQE - move
2184          * QP to RTS and repost flushed work requests
2185          */
2186         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2187                                    mad_list);
2188         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2189                 if (mad_send_wr->retry) {
2190                         /* Repost send */
2191                         struct ib_send_wr *bad_send_wr;
2192
2193                         mad_send_wr->retry = 0;
2194                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2195                                         &bad_send_wr);
2196                         if (ret)
2197                                 ib_mad_send_done_handler(port_priv, wc);
2198                 } else
2199                         ib_mad_send_done_handler(port_priv, wc);
2200         } else {
2201                 struct ib_qp_attr *attr;
2202
2203                 /* Transition QP to RTS and fail offending send */
2204                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2205                 if (attr) {
2206                         attr->qp_state = IB_QPS_RTS;
2207                         attr->cur_qp_state = IB_QPS_SQE;
2208                         ret = ib_modify_qp(qp_info->qp, attr,
2209                                            IB_QP_STATE | IB_QP_CUR_STATE);
2210                         kfree(attr);
2211                         if (ret)
2212                                 printk(KERN_ERR PFX "mad_error_handler - "
2213                                        "ib_modify_qp to RTS : %d\n", ret);
2214                         else
2215                                 mark_sends_for_retry(qp_info);
2216                 }
2217                 ib_mad_send_done_handler(port_priv, wc);
2218         }
2219 }
2220
2221 /*
2222  * IB MAD completion callback
2223  */
2224 static void ib_mad_completion_handler(struct work_struct *work)
2225 {
2226         struct ib_mad_port_private *port_priv;
2227         struct ib_wc wc;
2228
2229         port_priv = container_of(work, struct ib_mad_port_private, work);
2230         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2231
2232         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2233                 if (wc.status == IB_WC_SUCCESS) {
2234                         switch (wc.opcode) {
2235                         case IB_WC_SEND:
2236                                 ib_mad_send_done_handler(port_priv, &wc);
2237                                 break;
2238                         case IB_WC_RECV:
2239                                 ib_mad_recv_done_handler(port_priv, &wc);
2240                                 break;
2241                         default:
2242                                 BUG_ON(1);
2243                                 break;
2244                         }
2245                 } else
2246                         mad_error_handler(port_priv, &wc);
2247         }
2248 }
2249
2250 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2251 {
2252         unsigned long flags;
2253         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2254         struct ib_mad_send_wc mad_send_wc;
2255         struct list_head cancel_list;
2256
2257         INIT_LIST_HEAD(&cancel_list);
2258
2259         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2260         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2261                                  &mad_agent_priv->send_list, agent_list) {
2262                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2263                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2264                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2265                 }
2266         }
2267
2268         /* Empty wait list to prevent receives from finding a request */
2269         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2270         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2271
2272         /* Report all cancelled requests */
2273         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2274         mad_send_wc.vendor_err = 0;
2275
2276         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2277                                  &cancel_list, agent_list) {
2278                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2279                 list_del(&mad_send_wr->agent_list);
2280                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2281                                                    &mad_send_wc);
2282                 atomic_dec(&mad_agent_priv->refcount);
2283         }
2284 }
2285
2286 static struct ib_mad_send_wr_private*
2287 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2288              struct ib_mad_send_buf *send_buf)
2289 {
2290         struct ib_mad_send_wr_private *mad_send_wr;
2291
2292         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2293                             agent_list) {
2294                 if (&mad_send_wr->send_buf == send_buf)
2295                         return mad_send_wr;
2296         }
2297
2298         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2299                             agent_list) {
2300                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2301                     &mad_send_wr->send_buf == send_buf)
2302                         return mad_send_wr;
2303         }
2304         return NULL;
2305 }
2306
2307 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2308                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2309 {
2310         struct ib_mad_agent_private *mad_agent_priv;
2311         struct ib_mad_send_wr_private *mad_send_wr;
2312         unsigned long flags;
2313         int active;
2314
2315         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2316                                       agent);
2317         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2318         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2319         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2320                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2321                 return -EINVAL;
2322         }
2323
2324         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2325         if (!timeout_ms) {
2326                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2327                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2328         }
2329
2330         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2331         if (active)
2332                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2333         else
2334                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2335
2336         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2337         return 0;
2338 }
2339 EXPORT_SYMBOL(ib_modify_mad);
2340
2341 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2342                    struct ib_mad_send_buf *send_buf)
2343 {
2344         ib_modify_mad(mad_agent, send_buf, 0);
2345 }
2346 EXPORT_SYMBOL(ib_cancel_mad);
2347
2348 static void local_completions(struct work_struct *work)
2349 {
2350         struct ib_mad_agent_private *mad_agent_priv;
2351         struct ib_mad_local_private *local;
2352         struct ib_mad_agent_private *recv_mad_agent;
2353         unsigned long flags;
2354         int free_mad;
2355         struct ib_wc wc;
2356         struct ib_mad_send_wc mad_send_wc;
2357
2358         mad_agent_priv =
2359                 container_of(work, struct ib_mad_agent_private, local_work);
2360
2361         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2362         while (!list_empty(&mad_agent_priv->local_list)) {
2363                 local = list_entry(mad_agent_priv->local_list.next,
2364                                    struct ib_mad_local_private,
2365                                    completion_list);
2366                 list_del(&local->completion_list);
2367                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2368                 free_mad = 0;
2369                 if (local->mad_priv) {
2370                         recv_mad_agent = local->recv_mad_agent;
2371                         if (!recv_mad_agent) {
2372                                 printk(KERN_ERR PFX "No receive MAD agent for local completion\n");
2373                                 free_mad = 1;
2374                                 goto local_send_completion;
2375                         }
2376
2377                         /*
2378                          * Defined behavior is to complete response
2379                          * before request
2380                          */
2381                         build_smp_wc(recv_mad_agent->agent.qp,
2382                                      (unsigned long) local->mad_send_wr,
2383                                      be16_to_cpu(IB_LID_PERMISSIVE),
2384                                      0, recv_mad_agent->agent.port_num, &wc);
2385
2386                         local->mad_priv->header.recv_wc.wc = &wc;
2387                         local->mad_priv->header.recv_wc.mad_len =
2388                                                 sizeof(struct ib_mad);
2389                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2390                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2391                                  &local->mad_priv->header.recv_wc.rmpp_list);
2392                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2393                         local->mad_priv->header.recv_wc.recv_buf.mad =
2394                                                 &local->mad_priv->mad.mad;
2395                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2396                                 snoop_recv(recv_mad_agent->qp_info,
2397                                           &local->mad_priv->header.recv_wc,
2398                                            IB_MAD_SNOOP_RECVS);
2399                         recv_mad_agent->agent.recv_handler(
2400                                                 &recv_mad_agent->agent,
2401                                                 &local->mad_priv->header.recv_wc);
2402                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2403                         atomic_dec(&recv_mad_agent->refcount);
2404                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2405                 }
2406
2407 local_send_completion:
2408                 /* Complete send */
2409                 mad_send_wc.status = IB_WC_SUCCESS;
2410                 mad_send_wc.vendor_err = 0;
2411                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2412                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2413                         snoop_send(mad_agent_priv->qp_info,
2414                                    &local->mad_send_wr->send_buf,
2415                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2416                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2417                                                    &mad_send_wc);
2418
2419                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2420                 atomic_dec(&mad_agent_priv->refcount);
2421                 if (free_mad)
2422                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2423                 kfree(local);
2424         }
2425         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2426 }
2427
2428 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2429 {
2430         int ret;
2431
2432         if (!mad_send_wr->retries_left)
2433                 return -ETIMEDOUT;
2434
2435         mad_send_wr->retries_left--;
2436         mad_send_wr->send_buf.retries++;
2437
2438         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2439
2440         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2441                 ret = ib_retry_rmpp(mad_send_wr);
2442                 switch (ret) {
2443                 case IB_RMPP_RESULT_UNHANDLED:
2444                         ret = ib_send_mad(mad_send_wr);
2445                         break;
2446                 case IB_RMPP_RESULT_CONSUMED:
2447                         ret = 0;
2448                         break;
2449                 default:
2450                         ret = -ECOMM;
2451                         break;
2452                 }
2453         } else
2454                 ret = ib_send_mad(mad_send_wr);
2455
2456         if (!ret) {
2457                 mad_send_wr->refcount++;
2458                 list_add_tail(&mad_send_wr->agent_list,
2459                               &mad_send_wr->mad_agent_priv->send_list);
2460         }
2461         return ret;
2462 }
2463
2464 static void timeout_sends(struct work_struct *work)
2465 {
2466         struct ib_mad_agent_private *mad_agent_priv;
2467         struct ib_mad_send_wr_private *mad_send_wr;
2468         struct ib_mad_send_wc mad_send_wc;
2469         unsigned long flags, delay;
2470
2471         mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2472                                       timed_work.work);
2473         mad_send_wc.vendor_err = 0;
2474
2475         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2476         while (!list_empty(&mad_agent_priv->wait_list)) {
2477                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2478                                          struct ib_mad_send_wr_private,
2479                                          agent_list);
2480
2481                 if (time_after(mad_send_wr->timeout, jiffies)) {
2482                         delay = mad_send_wr->timeout - jiffies;
2483                         if ((long)delay <= 0)
2484                                 delay = 1;
2485                         queue_delayed_work(mad_agent_priv->qp_info->
2486                                            port_priv->wq,
2487                                            &mad_agent_priv->timed_work, delay);
2488                         break;
2489                 }
2490
2491                 list_del(&mad_send_wr->agent_list);
2492                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2493                     !retry_send(mad_send_wr))
2494                         continue;
2495
2496                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2497
2498                 if (mad_send_wr->status == IB_WC_SUCCESS)
2499                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2500                 else
2501                         mad_send_wc.status = mad_send_wr->status;
2502                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2503                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2504                                                    &mad_send_wc);
2505
2506                 atomic_dec(&mad_agent_priv->refcount);
2507                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2508         }
2509         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2510 }
2511
2512 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2513 {
2514         struct ib_mad_port_private *port_priv = cq->cq_context;
2515         unsigned long flags;
2516
2517         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2518         if (!list_empty(&port_priv->port_list))
2519                 queue_work(port_priv->wq, &port_priv->work);
2520         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2521 }
2522
2523 /*
2524  * Allocate receive MADs and post receive WRs for them
2525  */
2526 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2527                                     struct ib_mad_private *mad)
2528 {
2529         unsigned long flags;
2530         int post, ret;
2531         struct ib_mad_private *mad_priv;
2532         struct ib_sge sg_list;
2533         struct ib_recv_wr recv_wr, *bad_recv_wr;
2534         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2535
2536         /* Initialize common scatter list fields */
2537         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2538         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2539
2540         /* Initialize common receive WR fields */
2541         recv_wr.next = NULL;
2542         recv_wr.sg_list = &sg_list;
2543         recv_wr.num_sge = 1;
2544
2545         do {
2546                 /* Allocate and map receive buffer */
2547                 if (mad) {
2548                         mad_priv = mad;
2549                         mad = NULL;
2550                 } else {
2551                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2552                         if (!mad_priv) {
2553                                 printk(KERN_ERR PFX "No memory for receive buffer\n");
2554                                 ret = -ENOMEM;
2555                                 break;
2556                         }
2557                 }
2558                 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2559                                                  &mad_priv->grh,
2560                                                  sizeof *mad_priv -
2561                                                    sizeof mad_priv->header,
2562                                                  DMA_FROM_DEVICE);
2563                 mad_priv->header.mapping = sg_list.addr;
2564                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2565                 mad_priv->header.mad_list.mad_queue = recv_queue;
2566
2567                 /* Post receive WR */
2568                 spin_lock_irqsave(&recv_queue->lock, flags);
2569                 post = (++recv_queue->count < recv_queue->max_active);
2570                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2571                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2572                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2573                 if (ret) {
2574                         spin_lock_irqsave(&recv_queue->lock, flags);
2575                         list_del(&mad_priv->header.mad_list.list);
2576                         recv_queue->count--;
2577                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2578                         ib_dma_unmap_single(qp_info->port_priv->device,
2579                                             mad_priv->header.mapping,
2580                                             sizeof *mad_priv -
2581                                               sizeof mad_priv->header,
2582                                             DMA_FROM_DEVICE);
2583                         kmem_cache_free(ib_mad_cache, mad_priv);
2584                         printk(KERN_ERR PFX "ib_post_recv failed: %d\n", ret);
2585                         break;
2586                 }
2587         } while (post);
2588
2589         return ret;
2590 }
2591
2592 /*
2593  * Return all the posted receive MADs
2594  */
2595 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2596 {
2597         struct ib_mad_private_header *mad_priv_hdr;
2598         struct ib_mad_private *recv;
2599         struct ib_mad_list_head *mad_list;
2600
2601         if (!qp_info->qp)
2602                 return;
2603
2604         while (!list_empty(&qp_info->recv_queue.list)) {
2605
2606                 mad_list = list_entry(qp_info->recv_queue.list.next,
2607                                       struct ib_mad_list_head, list);
2608                 mad_priv_hdr = container_of(mad_list,
2609                                             struct ib_mad_private_header,
2610                                             mad_list);
2611                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2612                                     header);
2613
2614                 /* Remove from posted receive MAD list */
2615                 list_del(&mad_list->list);
2616
2617                 ib_dma_unmap_single(qp_info->port_priv->device,
2618                                     recv->header.mapping,
2619                                     sizeof(struct ib_mad_private) -
2620                                       sizeof(struct ib_mad_private_header),
2621                                     DMA_FROM_DEVICE);
2622                 kmem_cache_free(ib_mad_cache, recv);
2623         }
2624
2625         qp_info->recv_queue.count = 0;
2626 }
2627
2628 /*
2629  * Start the port
2630  */
2631 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2632 {
2633         int ret, i;
2634         struct ib_qp_attr *attr;
2635         struct ib_qp *qp;
2636
2637         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2638         if (!attr) {
2639                 printk(KERN_ERR PFX "Couldn't kmalloc ib_qp_attr\n");
2640                 return -ENOMEM;
2641         }
2642
2643         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2644                 qp = port_priv->qp_info[i].qp;
2645                 if (!qp)
2646                         continue;
2647
2648                 /*
2649                  * PKey index for QP1 is irrelevant but
2650                  * one is needed for the Reset to Init transition
2651                  */
2652                 attr->qp_state = IB_QPS_INIT;
2653                 attr->pkey_index = 0;
2654                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2655                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2656                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2657                 if (ret) {
2658                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2659                                "INIT: %d\n", i, ret);
2660                         goto out;
2661                 }
2662
2663                 attr->qp_state = IB_QPS_RTR;
2664                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2665                 if (ret) {
2666                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2667                                "RTR: %d\n", i, ret);
2668                         goto out;
2669                 }
2670
2671                 attr->qp_state = IB_QPS_RTS;
2672                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2673                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2674                 if (ret) {
2675                         printk(KERN_ERR PFX "Couldn't change QP%d state to "
2676                                "RTS: %d\n", i, ret);
2677                         goto out;
2678                 }
2679         }
2680
2681         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2682         if (ret) {
2683                 printk(KERN_ERR PFX "Failed to request completion "
2684                        "notification: %d\n", ret);
2685                 goto out;
2686         }
2687
2688         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2689                 if (!port_priv->qp_info[i].qp)
2690                         continue;
2691
2692                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2693                 if (ret) {
2694                         printk(KERN_ERR PFX "Couldn't post receive WRs\n");
2695                         goto out;
2696                 }
2697         }
2698 out:
2699         kfree(attr);
2700         return ret;
2701 }
2702
2703 static void qp_event_handler(struct ib_event *event, void *qp_context)
2704 {
2705         struct ib_mad_qp_info   *qp_info = qp_context;
2706
2707         /* It's worse than that! He's dead, Jim! */
2708         printk(KERN_ERR PFX "Fatal error (%d) on MAD QP (%d)\n",
2709                 event->event, qp_info->qp->qp_num);
2710 }
2711
2712 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2713                            struct ib_mad_queue *mad_queue)
2714 {
2715         mad_queue->qp_info = qp_info;
2716         mad_queue->count = 0;
2717         spin_lock_init(&mad_queue->lock);
2718         INIT_LIST_HEAD(&mad_queue->list);
2719 }
2720
2721 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2722                         struct ib_mad_qp_info *qp_info)
2723 {
2724         qp_info->port_priv = port_priv;
2725         init_mad_queue(qp_info, &qp_info->send_queue);
2726         init_mad_queue(qp_info, &qp_info->recv_queue);
2727         INIT_LIST_HEAD(&qp_info->overflow_list);
2728         spin_lock_init(&qp_info->snoop_lock);
2729         qp_info->snoop_table = NULL;
2730         qp_info->snoop_table_size = 0;
2731         atomic_set(&qp_info->snoop_count, 0);
2732 }
2733
2734 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2735                          enum ib_qp_type qp_type)
2736 {
2737         struct ib_qp_init_attr  qp_init_attr;
2738         int ret;
2739
2740         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2741         qp_init_attr.send_cq = qp_info->port_priv->cq;
2742         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2743         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2744         qp_init_attr.cap.max_send_wr = mad_sendq_size;
2745         qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2746         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2747         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2748         qp_init_attr.qp_type = qp_type;
2749         qp_init_attr.port_num = qp_info->port_priv->port_num;
2750         qp_init_attr.qp_context = qp_info;
2751         qp_init_attr.event_handler = qp_event_handler;
2752         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2753         if (IS_ERR(qp_info->qp)) {
2754                 printk(KERN_ERR PFX "Couldn't create ib_mad QP%d\n",
2755                        get_spl_qp_index(qp_type));
2756                 ret = PTR_ERR(qp_info->qp);
2757                 goto error;
2758         }
2759         /* Use minimum queue sizes unless the CQ is resized */
2760         qp_info->send_queue.max_active = mad_sendq_size;
2761         qp_info->recv_queue.max_active = mad_recvq_size;
2762         return 0;
2763
2764 error:
2765         return ret;
2766 }
2767
2768 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2769 {
2770         if (!qp_info->qp)
2771                 return;
2772
2773         ib_destroy_qp(qp_info->qp);
2774         kfree(qp_info->snoop_table);
2775 }
2776
2777 /*
2778  * Open the port
2779  * Create the QP, PD, MR, and CQ if needed
2780  */
2781 static int ib_mad_port_open(struct ib_device *device,
2782                             int port_num)
2783 {
2784         int ret, cq_size;
2785         struct ib_mad_port_private *port_priv;
2786         unsigned long flags;
2787         char name[sizeof "ib_mad123"];
2788         int has_smi;
2789
2790         /* Create new device info */
2791         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2792         if (!port_priv) {
2793                 printk(KERN_ERR PFX "No memory for ib_mad_port_private\n");
2794                 return -ENOMEM;
2795         }
2796
2797         port_priv->device = device;
2798         port_priv->port_num = port_num;
2799         spin_lock_init(&port_priv->reg_lock);
2800         INIT_LIST_HEAD(&port_priv->agent_list);
2801         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2802         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2803
2804         cq_size = mad_sendq_size + mad_recvq_size;
2805         has_smi = rdma_port_get_link_layer(device, port_num) == IB_LINK_LAYER_INFINIBAND;
2806         if (has_smi)
2807                 cq_size *= 2;
2808
2809         port_priv->cq = ib_create_cq(port_priv->device,
2810                                      ib_mad_thread_completion_handler,
2811                                      NULL, port_priv, cq_size, 0);
2812         if (IS_ERR(port_priv->cq)) {
2813                 printk(KERN_ERR PFX "Couldn't create ib_mad CQ\n");
2814                 ret = PTR_ERR(port_priv->cq);
2815                 goto error3;
2816         }
2817
2818         port_priv->pd = ib_alloc_pd(device);
2819         if (IS_ERR(port_priv->pd)) {
2820                 printk(KERN_ERR PFX "Couldn't create ib_mad PD\n");
2821                 ret = PTR_ERR(port_priv->pd);
2822                 goto error4;
2823         }
2824
2825         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2826         if (IS_ERR(port_priv->mr)) {
2827                 printk(KERN_ERR PFX "Couldn't get ib_mad DMA MR\n");
2828                 ret = PTR_ERR(port_priv->mr);
2829                 goto error5;
2830         }
2831
2832         if (has_smi) {
2833                 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2834                 if (ret)
2835                         goto error6;
2836         }
2837         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2838         if (ret)
2839                 goto error7;
2840
2841         snprintf(name, sizeof name, "ib_mad%d", port_num);
2842         port_priv->wq = create_singlethread_workqueue(name);
2843         if (!port_priv->wq) {
2844                 ret = -ENOMEM;
2845                 goto error8;
2846         }
2847         INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2848
2849         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2850         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2851         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2852
2853         ret = ib_mad_port_start(port_priv);
2854         if (ret) {
2855                 printk(KERN_ERR PFX "Couldn't start port\n");
2856                 goto error9;
2857         }
2858
2859         return 0;
2860
2861 error9:
2862         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2863         list_del_init(&port_priv->port_list);
2864         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2865
2866         destroy_workqueue(port_priv->wq);
2867 error8:
2868         destroy_mad_qp(&port_priv->qp_info[1]);
2869 error7:
2870         destroy_mad_qp(&port_priv->qp_info[0]);
2871 error6:
2872         ib_dereg_mr(port_priv->mr);
2873 error5:
2874         ib_dealloc_pd(port_priv->pd);
2875 error4:
2876         ib_destroy_cq(port_priv->cq);
2877         cleanup_recv_queue(&port_priv->qp_info[1]);
2878         cleanup_recv_queue(&port_priv->qp_info[0]);
2879 error3:
2880         kfree(port_priv);
2881
2882         return ret;
2883 }
2884
2885 /*
2886  * Close the port
2887  * If there are no classes using the port, free the port
2888  * resources (CQ, MR, PD, QP) and remove the port's info structure
2889  */
2890 static int ib_mad_port_close(struct ib_device *device, int port_num)
2891 {
2892         struct ib_mad_port_private *port_priv;
2893         unsigned long flags;
2894
2895         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2896         port_priv = __ib_get_mad_port(device, port_num);
2897         if (port_priv == NULL) {
2898                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2899                 printk(KERN_ERR PFX "Port %d not found\n", port_num);
2900                 return -ENODEV;
2901         }
2902         list_del_init(&port_priv->port_list);
2903         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2904
2905         destroy_workqueue(port_priv->wq);
2906         destroy_mad_qp(&port_priv->qp_info[1]);
2907         destroy_mad_qp(&port_priv->qp_info[0]);
2908         ib_dereg_mr(port_priv->mr);
2909         ib_dealloc_pd(port_priv->pd);
2910         ib_destroy_cq(port_priv->cq);
2911         cleanup_recv_queue(&port_priv->qp_info[1]);
2912         cleanup_recv_queue(&port_priv->qp_info[0]);
2913         /* XXX: Handle deallocation of MAD registration tables */
2914
2915         kfree(port_priv);
2916
2917         return 0;
2918 }
2919
2920 static void ib_mad_init_device(struct ib_device *device)
2921 {
2922         int start, end, i;
2923
2924         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2925                 return;
2926
2927         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2928                 start = 0;
2929                 end   = 0;
2930         } else {
2931                 start = 1;
2932                 end   = device->phys_port_cnt;
2933         }
2934
2935         for (i = start; i <= end; i++) {
2936                 if (ib_mad_port_open(device, i)) {
2937                         printk(KERN_ERR PFX "Couldn't open %s port %d\n",
2938                                device->name, i);
2939                         goto error;
2940                 }
2941                 if (ib_agent_port_open(device, i)) {
2942                         printk(KERN_ERR PFX "Couldn't open %s port %d "
2943                                "for agents\n",
2944                                device->name, i);
2945                         goto error_agent;
2946                 }
2947         }
2948         return;
2949
2950 error_agent:
2951         if (ib_mad_port_close(device, i))
2952                 printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2953                        device->name, i);
2954
2955 error:
2956         i--;
2957
2958         while (i >= start) {
2959                 if (ib_agent_port_close(device, i))
2960                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2961                                "for agents\n",
2962                                device->name, i);
2963                 if (ib_mad_port_close(device, i))
2964                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2965                                device->name, i);
2966                 i--;
2967         }
2968 }
2969
2970 static void ib_mad_remove_device(struct ib_device *device)
2971 {
2972         int i, num_ports, cur_port;
2973
2974         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2975                 return;
2976
2977         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2978                 num_ports = 1;
2979                 cur_port = 0;
2980         } else {
2981                 num_ports = device->phys_port_cnt;
2982                 cur_port = 1;
2983         }
2984         for (i = 0; i < num_ports; i++, cur_port++) {
2985                 if (ib_agent_port_close(device, cur_port))
2986                         printk(KERN_ERR PFX "Couldn't close %s port %d "
2987                                "for agents\n",
2988                                device->name, cur_port);
2989                 if (ib_mad_port_close(device, cur_port))
2990                         printk(KERN_ERR PFX "Couldn't close %s port %d\n",
2991                                device->name, cur_port);
2992         }
2993 }
2994
2995 static struct ib_client mad_client = {
2996         .name   = "mad",
2997         .add = ib_mad_init_device,
2998         .remove = ib_mad_remove_device
2999 };
3000
3001 static int __init ib_mad_init_module(void)
3002 {
3003         int ret;
3004
3005         mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3006         mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3007
3008         mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3009         mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3010
3011         ib_mad_cache = kmem_cache_create("ib_mad",
3012                                          sizeof(struct ib_mad_private),
3013                                          0,
3014                                          SLAB_HWCACHE_ALIGN,
3015                                          NULL);
3016         if (!ib_mad_cache) {
3017                 printk(KERN_ERR PFX "Couldn't create ib_mad cache\n");
3018                 ret = -ENOMEM;
3019                 goto error1;
3020         }
3021
3022         INIT_LIST_HEAD(&ib_mad_port_list);
3023
3024         if (ib_register_client(&mad_client)) {
3025                 printk(KERN_ERR PFX "Couldn't register ib_mad client\n");
3026                 ret = -EINVAL;
3027                 goto error2;
3028         }
3029
3030         return 0;
3031
3032 error2:
3033         kmem_cache_destroy(ib_mad_cache);
3034 error1:
3035         return ret;
3036 }
3037
3038 static void __exit ib_mad_cleanup_module(void)
3039 {
3040         ib_unregister_client(&mad_client);
3041         kmem_cache_destroy(ib_mad_cache);
3042 }
3043
3044 module_init(ib_mad_init_module);
3045 module_exit(ib_mad_cleanup_module);